From ca9c9d5a9a2fff94e176bd8cc6460d786b6b1563 Mon Sep 17 00:00:00 2001 From: abb128 <65567823+abb128@users.noreply.github.com> Date: Sat, 25 Nov 2023 09:39:04 +0200 Subject: [PATCH] ggml backend v2 --- native/jni/src/ggml/finetune.cpp | 52 +- native/jni/src/ggml/ggml-alloc.c | 637 +++-- native/jni/src/ggml/ggml-alloc.h | 80 +- native/jni/src/ggml/ggml-backend-impl.h | 87 + native/jni/src/ggml/ggml-backend.c | 683 ++++- native/jni/src/ggml/ggml-backend.h | 185 +- native/jni/src/ggml/ggml-impl.h | 22 +- native/jni/src/ggml/ggml-quants.c | 256 +- native/jni/src/ggml/ggml.c | 2774 +++++++------------ native/jni/src/ggml/ggml.h | 3368 +++++++++++----------- native/jni/src/ggml/llama.cpp | 3389 +++++++++++++---------- native/jni/src/ggml/llama.h | 1239 +++++---- native/jni/src/ggml/train.cpp | 13 + native/jni/src/ggml/train.h | 2 + 14 files changed, 6852 insertions(+), 5935 deletions(-) create mode 100644 native/jni/src/ggml/ggml-backend-impl.h diff --git a/native/jni/src/ggml/finetune.cpp b/native/jni/src/ggml/finetune.cpp index 92786e6fa..51a877e20 100644 --- a/native/jni/src/ggml/finetune.cpp +++ b/native/jni/src/ggml/finetune.cpp @@ -550,35 +550,35 @@ static void randomize_lora(struct my_llama_lora * lora, int seed, float mean, fl struct random_normal_distribution * rnd = init_random_normal_distribution(seed, mean, std, min, max); randomize_tensor_normal(lora->tok_embeddings_a, rnd); - randomize_tensor_normal(lora->tok_embeddings_b, rnd); + ggml_set_zero(lora->tok_embeddings_b); randomize_tensor_normal(lora->norm_a, rnd); - randomize_tensor_normal(lora->norm_b, rnd); + ggml_set_zero(lora->norm_b); randomize_tensor_normal(lora->output_a, rnd); - randomize_tensor_normal(lora->output_b, rnd); + ggml_set_zero(lora->output_b); for (uint32_t i = 0; i < n_layer; ++i) { auto & layer = lora->layers[i]; randomize_tensor_normal(layer.attention_norm_a, rnd); - randomize_tensor_normal(layer.attention_norm_b, rnd); + ggml_set_zero(layer.attention_norm_b); randomize_tensor_normal(layer.wq_a, rnd); - randomize_tensor_normal(layer.wq_b, rnd); + ggml_set_zero(layer.wq_b); randomize_tensor_normal(layer.wk_a, rnd); - randomize_tensor_normal(layer.wk_b, rnd); + ggml_set_zero(layer.wk_b); randomize_tensor_normal(layer.wv_a, rnd); - randomize_tensor_normal(layer.wv_b, rnd); + ggml_set_zero(layer.wv_b); randomize_tensor_normal(layer.wo_a, rnd); - randomize_tensor_normal(layer.wo_b, rnd); + ggml_set_zero(layer.wo_b); randomize_tensor_normal(layer.ffn_norm_a, rnd); - randomize_tensor_normal(layer.ffn_norm_b, rnd); + ggml_set_zero(layer.ffn_norm_b); randomize_tensor_normal(layer.w1_a, rnd); - randomize_tensor_normal(layer.w1_b, rnd); + ggml_set_zero(layer.w1_b); randomize_tensor_normal(layer.w2_a, rnd); - randomize_tensor_normal(layer.w2_b, rnd); + ggml_set_zero(layer.w2_b); randomize_tensor_normal(layer.w3_a, rnd); - randomize_tensor_normal(layer.w3_b, rnd); + ggml_set_zero(layer.w3_b); } free_random_normal_distribution(rnd); @@ -644,8 +644,9 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs( const int rope_mode = 0; return ggml_rope_custom(ctx, - t, KQ_pos, n_rot, rope_mode, n_ctx, - rope_freq_base, rope_freq_scale); + t, KQ_pos, n_rot, rope_mode, n_ctx, 0, + rope_freq_base, rope_freq_scale, 0.0f, 1.0f, 0.0f, 0.0f + ); }; set_name(tokens_input, "tokens_input"); @@ -773,7 +774,7 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs( if (enable_checkpointing) { ggml_build_backward_gradient_checkpointing(ctx, gf, gb, gb_tmp, checkpoints.data(), (int) checkpoints.size()); } else { - *gb = *gf; + ggml_graph_cpy(gf, gb); ggml_build_backward_expand(ctx, gf, gb, true); } @@ -1308,6 +1309,7 @@ int finetune_train(struct train_params params) { opt->params = ggml_opt_default_params(GGML_OPT_ADAM); opt->params.print_forward_graph = false; opt->params.print_backward_graph = false; + opt->params.graph_size = LLAMA_TRAIN_MAX_NODES; opt->params.n_threads = params.common.n_threads; opt->params.past = params.common.opt_past; opt->params.delta = params.common.opt_delta; @@ -1434,11 +1436,9 @@ int finetune_train(struct train_params params) { ggml_allocr_free(alloc); // context for compute tensors without their data - size_t estimated_compute_size_wo_data = ( - ggml_tensor_overhead()*GGML_MAX_NODES*2 - + (GGML_OBJECT_SIZE+GGML_GRAPH_SIZE)*( - params.common.use_checkpointing ? 3 : 2 - ) + const size_t estimated_compute_size_wo_data = ( + 2*LLAMA_TRAIN_MAX_NODES*ggml_tensor_overhead() + + (params.common.use_checkpointing ? 3 : 2)*(GGML_OBJECT_SIZE+ggml_graph_overhead_custom(LLAMA_TRAIN_MAX_NODES, true)) ); struct ggml_init_params ctx_compute_params = { estimated_compute_size_wo_data, // mem_size @@ -1461,11 +1461,11 @@ int finetune_train(struct train_params params) { for (unsigned order = 0; order < (unsigned) GGML_CGRAPH_EVAL_ORDER_COUNT; ++order) { ctx_compute = ggml_init(ctx_compute_params); alloc = ggml_allocr_new_measure(tensor_alignment); - gf = ggml_new_graph(ctx_compute); + gf = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true); gf->order = (enum ggml_cgraph_eval_order) order; - gb = ggml_new_graph(ctx_compute); + gb = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true); gb_tmp = params.common.use_checkpointing - ? ggml_new_graph(ctx_compute) + ? ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true) : NULL; loss = llama_build_lora_finetune_graphs( &model, &lora, alloc, ctx_compute, @@ -1494,11 +1494,11 @@ int finetune_train(struct train_params params) { mem_compute_data.resize(max_compute_size); ctx_compute = ggml_init(ctx_compute_params); alloc = ggml_allocr_new(mem_compute_data.data(), mem_compute_data.size(), tensor_alignment); - gf = ggml_new_graph(ctx_compute); + gf = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true); gf->order = best_order; - gb = ggml_new_graph(ctx_compute); + gb = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true); gb_tmp = params.common.use_checkpointing - ? ggml_new_graph(ctx_compute) + ? ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true) : NULL; loss = llama_build_lora_finetune_graphs( &model, &lora, alloc, ctx_compute, diff --git a/native/jni/src/ggml/ggml-alloc.c b/native/jni/src/ggml/ggml-alloc.c index 34eba3f83..137eff185 100644 --- a/native/jni/src/ggml/ggml-alloc.c +++ b/native/jni/src/ggml/ggml-alloc.c @@ -1,51 +1,21 @@ #include "ggml-alloc.h" -#include "ggml-backend.h" +#include "ggml-backend-impl.h" #include "ggml.h" +#include "ggml-impl.h" #include +#include #include #include #include #include - -#define UNUSED(x) (void)(x) #define MAX(a, b) ((a) > (b) ? (a) : (b)) -#define GGML_MAX_CONCUR (2*GGML_MAX_NODES) +#define MAX_FREE_BLOCKS 256 //#define GGML_ALLOCATOR_DEBUG -//#define AT_PRINTF printf -#define AT_PRINTF(...) ((void)0) - -struct hash_node { - struct ggml_tensor * t; - int n_children; - int n_views; -}; - -static size_t hash(void * p) { - return (size_t)p % GGML_GRAPH_HASHTABLE_SIZE; -} - -static struct hash_node * hash_get(struct hash_node hash_table[], struct ggml_tensor * t) { - size_t h = hash(t); - - // linear probing - size_t i = h; - while (hash_table[i].t != NULL) { - if (hash_table[i].t == t) { - return &hash_table[i]; - } - i = (i + 1) % GGML_GRAPH_HASHTABLE_SIZE; - if (i == h) { - // hash table is full - GGML_ASSERT(false); - } - } - - hash_table[i].t = t; - return &hash_table[i]; -} +//#define AT_PRINTF(...) fprintf(stderr, __VA_ARGS__) +#define AT_PRINTF(...) // TODO: GGML_PAD ? static size_t aligned_offset(const void * buffer, size_t offset, size_t alignment) { @@ -59,20 +29,18 @@ struct free_block { size_t size; }; -#define MAX_FREE_BLOCKS 256 - -struct ggml_allocr { +struct ggml_tallocr { struct ggml_backend_buffer * buffer; bool buffer_owned; - void * data; + void * base; size_t alignment; + int n_free_blocks; struct free_block free_blocks[MAX_FREE_BLOCKS]; - struct hash_node hash_table[GGML_GRAPH_HASHTABLE_SIZE]; + size_t max_size; + bool measure; - int parse_seq[GGML_MAX_CONCUR]; - int parse_seq_len; #ifdef GGML_ALLOCATOR_DEBUG struct ggml_tensor * allocated_tensors[1024]; @@ -80,7 +48,7 @@ struct ggml_allocr { }; #ifdef GGML_ALLOCATOR_DEBUG -static void add_allocated_tensor(struct ggml_allocr * alloc, struct ggml_tensor * tensor) { +static void add_allocated_tensor(ggml_tallocr_t alloc, struct ggml_tensor * tensor) { for (int i = 0; i < 1024; i++) { if (alloc->allocated_tensors[i] == NULL) { alloc->allocated_tensors[i] = tensor; @@ -89,7 +57,7 @@ static void add_allocated_tensor(struct ggml_allocr * alloc, struct ggml_tensor } GGML_ASSERT(!"out of allocated_tensors"); } -static void remove_allocated_tensor(struct ggml_allocr * alloc, struct ggml_tensor * tensor) { +static void remove_allocated_tensor(ggml_tallocr_t alloc, struct ggml_tensor * tensor) { for (int i = 0; i < 1024; i++) { if (alloc->allocated_tensors[i] == tensor || (alloc->allocated_tensors[i] != NULL && alloc->allocated_tensors[i]->data == tensor->data)) { @@ -103,7 +71,7 @@ static void remove_allocated_tensor(struct ggml_allocr * alloc, struct ggml_tens #endif // check if a tensor is allocated by this buffer -static bool ggml_allocr_is_own(struct ggml_allocr * alloc, const struct ggml_tensor * tensor) { +static bool ggml_tallocr_is_own(ggml_tallocr_t alloc, const struct ggml_tensor * tensor) { return tensor->buffer == alloc->buffer; } @@ -111,7 +79,7 @@ static bool ggml_is_view(struct ggml_tensor * t) { return t->view_src != NULL; } -void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor) { +void ggml_tallocr_alloc(ggml_tallocr_t alloc, struct ggml_tensor * tensor) { GGML_ASSERT(!ggml_is_view(tensor)); // views generally get data pointer from one of their sources GGML_ASSERT(tensor->data == NULL); // avoid allocating tensor which already has memory allocated @@ -162,9 +130,10 @@ void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor) } tensor->data = addr; - AT_PRINTF("%s: allocated data at %p\n", __func__, tensor->data); tensor->buffer = alloc->buffer; - ggml_backend_buffer_init_tensor(alloc->buffer, tensor); + if (!alloc->measure) { + ggml_backend_buffer_init_tensor(alloc->buffer, tensor); + } #ifdef GGML_ALLOCATOR_DEBUG add_allocated_tensor(alloc, tensor); @@ -180,16 +149,16 @@ void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor) } #endif - alloc->max_size = MAX(alloc->max_size, (char*)addr - (char*)alloc->data + size); + alloc->max_size = MAX(alloc->max_size, (char*)addr - (char*)alloc->base + size); } // this is a very naive implementation, but for our case the number of free blocks should be very small -static void ggml_allocr_free_tensor(struct ggml_allocr * alloc, struct ggml_tensor * tensor) { - if (ggml_allocr_is_own(alloc, tensor) == false) { +static void ggml_tallocr_free_tensor(ggml_tallocr_t alloc, struct ggml_tensor * tensor) { + if (ggml_tallocr_is_own(alloc, tensor) == false) { // the tensor was not allocated in this buffer // this can happen because the graph allocator will try to free weights and other tensors from different buffers // the easiest way to deal with this is just to ignore it - AT_PRINTF("ignoring %s (their buffer: %p, our buffer: %p)\n", tensor->name, (void *)tensor->buffer, (void *)alloc->buffer); + // AT_PRINTF("ignoring %s (their buffer: %p, our buffer: %p)\n", tensor->name, (void *)tensor->buffer, (void *)alloc->buffer); return; } @@ -199,7 +168,9 @@ static void ggml_allocr_free_tensor(struct ggml_allocr * alloc, struct ggml_tens size = aligned_offset(NULL, size, alloc->alignment); AT_PRINTF("%s: freeing %s at %p (%zu bytes) - n_free_blocks = %d\n", __func__, tensor->name, ptr, size, alloc->n_free_blocks); - ggml_backend_buffer_free_tensor(alloc->buffer, tensor); + if (!alloc->measure) { + ggml_backend_buffer_free_tensor(alloc->buffer, tensor); + } #ifdef GGML_ALLOCATOR_DEBUG remove_allocated_tensor(alloc, tensor); @@ -253,91 +224,180 @@ static void ggml_allocr_free_tensor(struct ggml_allocr * alloc, struct ggml_tens alloc->n_free_blocks++; } -void ggml_allocr_set_parse_seq(struct ggml_allocr * alloc, const int * list, int n) { - for (int i = 0; i < n; i++) { - alloc->parse_seq[i] = list[i]; - } - alloc->parse_seq_len = n; -} - -void ggml_allocr_reset(struct ggml_allocr * alloc) { +void ggml_tallocr_reset(ggml_tallocr_t alloc) { alloc->n_free_blocks = 1; - size_t align_offset = aligned_offset(alloc->data, 0, alloc->alignment); - alloc->free_blocks[0].addr = (char *)alloc->data + align_offset; - alloc->free_blocks[0].size = ggml_backend_buffer_get_size(alloc->buffer) - align_offset; + size_t align_offset = aligned_offset(alloc->base, 0, alloc->alignment); + alloc->free_blocks[0].addr = (char *)alloc->base + align_offset; + + if (alloc->measure) { + alloc->free_blocks[0].size = SIZE_MAX/2; // restrict maximum size of a measure allocator to half size_t max to avoid overflows + } else { + alloc->free_blocks[0].size = ggml_backend_buffer_get_size(alloc->buffer) - align_offset; + } } -struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment) { +ggml_tallocr_t ggml_tallocr_new(void * data, size_t size, size_t alignment) { struct ggml_backend_buffer * buffer = ggml_backend_cpu_buffer_from_ptr(NULL, data, size); - struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr)); + ggml_tallocr_t alloc = (ggml_tallocr_t)malloc(sizeof(struct ggml_tallocr)); - *alloc = (struct ggml_allocr){ - /*.buffer = */ buffer, - /*.buffer_owned = */ true, - /*.base = */ ggml_backend_buffer_get_base(buffer), - /*.alignment = */ alignment, - /*.n_free_blocks = */ 0, - /*.free_blocks = */ {{0}}, - /*.hash_table = */ {{0}}, - /*.max_size = */ 0, - /*.measure = */ false, - /*.parse_seq = */ {0}, - /*.parse_seq_len = */ 0, + *alloc = (struct ggml_tallocr) { + /*.buffer = */ buffer, + /*.buffer_owned = */ true, + /*.base = */ ggml_backend_buffer_get_base(buffer), + /*.alignment = */ alignment, + /*.n_free_blocks = */ 0, + /*.free_blocks = */ {{0}}, + /*.max_size = */ 0, + /*.measure = */ false, #ifdef GGML_ALLOCATOR_DEBUG - /*.allocated_tensors = */ {0}, + /*.allocated_tensors = */ {0}, #endif }; - ggml_allocr_reset(alloc); + ggml_tallocr_reset(alloc); return alloc; } -struct ggml_allocr * ggml_allocr_new_measure(size_t alignment) { - struct ggml_allocr * alloc = ggml_allocr_new((void *)0x1000, (size_t)-0x1001, alignment); +ggml_tallocr_t ggml_tallocr_new_measure(size_t alignment) { + ggml_tallocr_t alloc = ggml_tallocr_new((void *)0x1000, SIZE_MAX/2, alignment); alloc->measure = true; return alloc; } -struct ggml_allocr * ggml_allocr_new_from_buffer(struct ggml_backend_buffer * buffer) { - struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr)); +ggml_tallocr_t ggml_tallocr_new_measure_from_backend(struct ggml_backend * backend) { + // create a backend buffer to get the correct tensor allocation sizes + ggml_backend_buffer_t buffer = ggml_backend_alloc_buffer(backend, 1); - *alloc = (struct ggml_allocr){ - /*.buffer = */ buffer, - /*.buffer_owned = */ false, - /*.base = */ ggml_backend_buffer_get_base(buffer), - /*.alignment = */ ggml_backend_buffer_get_alignment(buffer), - /*.n_free_blocks = */ 0, - /*.free_blocks = */ {{0}}, - /*.hash_table = */ {{0}}, - /*.max_size = */ 0, - /*.measure = */ false, - /*.parse_seq = */ {0}, - /*.parse_seq_len = */ 0, + // TODO: move alloc initialization to a common ggml_tallocr_new_impl function + ggml_tallocr_t alloc = ggml_tallocr_new_from_buffer(buffer); + alloc->buffer_owned = true; + alloc->measure = true; + ggml_tallocr_reset(alloc); + return alloc; +} + +ggml_tallocr_t ggml_tallocr_new_from_backend(struct ggml_backend * backend, size_t size) { + ggml_backend_buffer_t buffer = ggml_backend_alloc_buffer(backend, size); + ggml_tallocr_t alloc = ggml_tallocr_new_from_buffer(buffer); + alloc->buffer_owned = true; + return alloc; +} + +ggml_tallocr_t ggml_tallocr_new_from_buffer(struct ggml_backend_buffer * buffer) { + ggml_tallocr_t alloc = (ggml_tallocr_t)malloc(sizeof(struct ggml_tallocr)); + + *alloc = (struct ggml_tallocr) { + /*.buffer = */ buffer, + /*.buffer_owned = */ false, + /*.base = */ ggml_backend_buffer_get_base(buffer), + /*.alignment = */ ggml_backend_buffer_get_alignment(buffer), + /*.n_free_blocks = */ 0, + /*.free_blocks = */ {{0}}, + /*.max_size = */ 0, + /*.measure = */ false, #ifdef GGML_ALLOCATOR_DEBUG - /*.allocated_tensors = */ {0}, + /*.allocated_tensors = */ {0}, #endif }; - ggml_allocr_reset(alloc); + ggml_tallocr_reset(alloc); return alloc; } -void ggml_allocr_free(struct ggml_allocr * alloc) { +struct ggml_backend_buffer * ggml_tallocr_get_buffer(ggml_tallocr_t alloc) { + return alloc->buffer; +} + +void ggml_tallocr_free(ggml_tallocr_t alloc) { + if (alloc == NULL) { + return; + } + if (alloc->buffer_owned) { ggml_backend_buffer_free(alloc->buffer); } free(alloc); } -bool ggml_allocr_is_measure(struct ggml_allocr * alloc) { +bool ggml_tallocr_is_measure(ggml_tallocr_t alloc) { return alloc->measure; } -//////////// compute graph allocator +size_t ggml_tallocr_max_size(ggml_tallocr_t alloc) { + return alloc->max_size; +} + +// graph allocator + +struct hash_node { + int n_children; + int n_views; +}; + +struct ggml_gallocr { + ggml_tallocr_t talloc; + struct ggml_hash_set hash_set; + struct hash_node * hash_values; + size_t hash_values_size; + ggml_tallocr_t * hash_allocs; + int * parse_seq; + int parse_seq_len; +}; + +ggml_gallocr_t ggml_gallocr_new(void) { + ggml_gallocr_t galloc = (ggml_gallocr_t)malloc(sizeof(struct ggml_gallocr)); + + *galloc = (struct ggml_gallocr) { + /*.talloc = */ NULL, + /*.hash_set = */ {0}, + /*.hash_values = */ NULL, + /*.hash_values_size = */ 0, + /*.hash_allocs = */ NULL, + /*.parse_seq = */ NULL, + /*.parse_seq_len = */ 0, + }; + + return galloc; +} + +void ggml_gallocr_free(ggml_gallocr_t galloc) { + if (galloc == NULL) { + return; + } + + if (galloc->hash_set.keys != NULL) { + free(galloc->hash_set.keys); + } + if (galloc->hash_values != NULL) { + free(galloc->hash_values); + } + if (galloc->hash_allocs != NULL) { + free(galloc->hash_allocs); + } + if (galloc->parse_seq != NULL) { + free(galloc->parse_seq); + } + free(galloc); +} + +void ggml_gallocr_set_parse_seq(ggml_gallocr_t galloc, const int * list, int n) { + free(galloc->parse_seq); + galloc->parse_seq = malloc(sizeof(int) * n); + + for (int i = 0; i < n; i++) { + galloc->parse_seq[i] = list[i]; + } + galloc->parse_seq_len = n; +} + +static struct hash_node * hash_get(ggml_gallocr_t galloc, struct ggml_tensor * t) { + size_t i = ggml_hash_find_or_insert(galloc->hash_set, t); + return &galloc->hash_values[i]; +} static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) { if (a->type != b->type) { @@ -378,23 +438,40 @@ static bool ggml_op_can_inplace(enum ggml_op op) { } } -static void init_view(struct ggml_allocr * alloc, struct ggml_tensor * view) { - assert(view->view_src != NULL && view->view_src->data != NULL); - view->backend = view->view_src->backend; +static ggml_tallocr_t node_tallocr(ggml_gallocr_t galloc, struct ggml_tensor * node) { + if (galloc->talloc != NULL) { + return galloc->talloc; + } + + return galloc->hash_allocs[ggml_hash_find_or_insert(galloc->hash_set, node)]; +} + +static void init_view(ggml_gallocr_t galloc, struct ggml_tensor * view, bool update_backend) { + ggml_tallocr_t alloc = node_tallocr(galloc, view); + + //printf("init_view: %s from src %s\n", view->name, view->view_src->name); + GGML_ASSERT(view->view_src != NULL && view->view_src->data != NULL); + if (update_backend) { + view->backend = view->view_src->backend; + } view->buffer = view->view_src->buffer; view->data = (char *)view->view_src->data + view->view_offs; // FIXME: the view should be initialized by the owning buffer, but currently this breaks the CUDA backend // due to the ggml_tensor_extra_gpu ring buffer overwriting the KV cache extras - assert(ggml_allocr_is_measure(alloc) || !view->buffer || view->buffer->backend == alloc->buffer->backend); - ggml_backend_buffer_init_tensor(alloc->buffer, view); + assert(ggml_tallocr_is_measure(alloc) || !view->buffer || view->buffer->backend == alloc->buffer->backend); + + if (!alloc->measure) { + ggml_backend_buffer_init_tensor(alloc->buffer, view); + } } -static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node) { - struct hash_node * ht = alloc->hash_table; +static void allocate_node(ggml_gallocr_t galloc, struct ggml_tensor * node) { + ggml_tallocr_t alloc = node_tallocr(galloc, node); + if (node->data == NULL) { if (ggml_is_view(node)) { - init_view(alloc, node); + init_view(galloc, node, true); } else { // see if we can reuse a parent's buffer (inplace) if (ggml_op_can_inplace(node->op)) { @@ -405,16 +482,16 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node) } // if the node's data is external, then we cannot re-use it - if (ggml_allocr_is_own(alloc, parent) == false) { + if (ggml_tallocr_is_own(alloc, parent) == false) { AT_PRINTF("not reusing parent %s for %s as %p is external\n", parent->name, node->name, parent->data); continue; } - struct hash_node * p_hn = hash_get(ht, parent); + struct hash_node * p_hn = hash_get(galloc, parent); if (parent->data != NULL && p_hn->n_children == 1 && p_hn->n_views == 0 && ggml_are_same_layout(node, parent)) { if (ggml_is_view(parent)) { struct ggml_tensor * view_src = parent->view_src; - struct hash_node * view_src_hn = hash_get(ht, view_src); + struct hash_node * view_src_hn = hash_get(galloc, view_src); if (view_src_hn->n_views == 1 && view_src_hn->n_children == 0 && view_src->data == parent->data) { // TODO: the offset of the view parent must be kept to ensure that the op doesn't overwrite // the parent's data that it will need later (same layout requirement). the problem is that then @@ -424,171 +501,267 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node) AT_PRINTF("reusing view parent %s (%s) for %s\n", parent->name, view_src->name, node->name); node->view_src = view_src; view_src_hn->n_views += 1; - init_view(alloc, node); + init_view(galloc, node, false); return; } - } - else { + } else { AT_PRINTF("reusing parent %s for %s\n", parent->name, node->name); node->view_src = parent; p_hn->n_views += 1; - init_view(alloc, node); + init_view(galloc, node, false); return; } } } } - ggml_allocr_alloc(alloc, node); + ggml_tallocr_alloc(alloc, node); } } } -size_t ggml_allocr_alloc_graph_n( - struct ggml_allocr * alloc, - struct ggml_cgraph ** graphs, int n_graphs, - struct ggml_tensor *** inputs, struct ggml_tensor *** outputs) { +static void free_node(ggml_gallocr_t galloc, struct ggml_tensor * node) { + ggml_tallocr_t alloc = node_tallocr(galloc, node); - // reset hash table - struct hash_node * ht = alloc->hash_table; - memset(ht, 0, sizeof(struct hash_node) * GGML_GRAPH_HASHTABLE_SIZE); + ggml_tallocr_free_tensor(alloc, node); +} + +static void ggml_tallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgraph * gf) { + const int * parse_seq = galloc->parse_seq; + int parse_seq_len = galloc->parse_seq_len; // count number of children and views - for (int g = 0; g < n_graphs; g++) { - struct ggml_cgraph * gf = graphs[g]; - for (int i = 0; i < gf->n_nodes; i++) { - struct ggml_tensor * node = gf->nodes[i]; + for (int i = 0; i < gf->n_nodes; i++) { + struct ggml_tensor * node = gf->nodes[i]; - if (ggml_is_view(node)) { - struct ggml_tensor * view_src = node->view_src; - hash_get(ht, view_src)->n_views += 1; - if (node->buffer == NULL && node->data != NULL) { - // view of a pre-allocated tensor, didn't call init_view() yet - init_view(alloc, node); - } + if (ggml_is_view(node)) { + struct ggml_tensor * view_src = node->view_src; + hash_get(galloc, view_src)->n_views += 1; + if (node->buffer == NULL && node->data != NULL) { + // view of a pre-allocated tensor, didn't call init_view() yet + init_view(galloc, node, true); } + } - for (int j = 0; j < GGML_MAX_SRC; j++) { - struct ggml_tensor * parent = node->src[j]; - if (parent == NULL) { - break; - } - hash_get(ht, parent)->n_children += 1; - if (ggml_is_view(parent) && parent->buffer == NULL && parent->data != NULL) { - init_view(alloc, parent); - } + for (int j = 0; j < GGML_MAX_SRC; j++) { + struct ggml_tensor * parent = node->src[j]; + if (parent == NULL) { + break; + } + hash_get(galloc, parent)->n_children += 1; + if (ggml_is_view(parent) && parent->buffer == NULL && parent->data != NULL) { + init_view(galloc, parent, true); } } } // allocate tensors - for (int g = 0; g < n_graphs; g++) { - struct ggml_cgraph * gf = graphs[g]; - AT_PRINTF("####### graph %d/%d\n", g, n_graphs); - // graph inputs are allocated first to ensure that they are not overwritten by each other - if (inputs != NULL && inputs[g] != NULL) { - for (int i = 0; inputs[g][i] != NULL; i++) { - struct ggml_tensor * input = inputs[g][i]; - AT_PRINTF("input: %s\n", input->name); - allocate_node(alloc, input); + // if we have parse_seq then we allocate nodes following the list, and we only free nodes at barriers + int last_barrier_pos = 0; + int n_nodes = parse_seq_len ? parse_seq_len : gf->n_nodes; + + for (int ind = 0; ind < n_nodes; ind++) { + // allocate a node if there is no parse_seq or this is not a barrier + if (parse_seq_len == 0 || parse_seq[ind] != -1) { + int i = parse_seq_len ? parse_seq[ind] : ind; + struct ggml_tensor * node = gf->nodes[i]; + + // allocate parents (leafs) + for (int j = 0; j < GGML_MAX_SRC; j++) { + struct ggml_tensor * parent = node->src[j]; + if (parent == NULL) { + break; + } + allocate_node(galloc, parent); } + + // allocate node + allocate_node(galloc, node); + + AT_PRINTF("exec: %s (%s) <= ", ggml_op_name(node->op), node->name); + for (int j = 0; j < GGML_MAX_SRC; j++) { + struct ggml_tensor * parent = node->src[j]; + if (parent == NULL) { + break; + } + AT_PRINTF("%s", parent->name); + if (j < GGML_MAX_SRC - 1 && node->src[j + 1] != NULL) { + AT_PRINTF(", "); + } + } + AT_PRINTF("\n"); } - // if we have parse_seq then we allocate nodes following the list, and we only free nodes at barriers - int last_barrier_pos = 0; - int n_nodes = alloc->parse_seq_len ? alloc->parse_seq_len : gf->n_nodes; - for (int ind = 0; ind < n_nodes; ind++) { - // allocate a node if there is no parse_seq or this is not a barrier - if ((alloc->parse_seq_len==0) || alloc->parse_seq[ind] != -1) { - int i = alloc->parse_seq_len ? alloc->parse_seq[ind] : ind; - struct ggml_tensor * node = gf->nodes[i]; + // update parents + // update immediately if there is no parse_seq + // update only at barriers if there is parse_seq + if ((parse_seq_len == 0) || parse_seq[ind] == -1) { + int update_start = parse_seq_len ? last_barrier_pos : ind; + int update_end = parse_seq_len ? ind : ind + 1; + for (int i = update_start; i < update_end; i++) { + int node_i = parse_seq_len ? parse_seq[i] : i; + struct ggml_tensor * node = gf->nodes[node_i]; - // allocate parents (leafs) for (int j = 0; j < GGML_MAX_SRC; j++) { struct ggml_tensor * parent = node->src[j]; if (parent == NULL) { break; } - allocate_node(alloc, parent); - } + struct hash_node * p_hn = hash_get(galloc, parent); + p_hn->n_children -= 1; - // allocate node - allocate_node(alloc, node); + //AT_PRINTF("parent %s: %d children, %d views\n", parent->name, parent->n_children, parent->n_views); - AT_PRINTF("exec: %s (%s) <= ", ggml_op_name(node->op), node->name); - for (int j = 0; j < GGML_MAX_SRC; j++) { - struct ggml_tensor * parent = node->src[j]; - if (parent == NULL) { - break; - } - AT_PRINTF("%s", parent->name); - if (j < GGML_MAX_SRC - 1 && node->src[j + 1] != NULL) { - AT_PRINTF(", "); - } - } - AT_PRINTF("\n"); - } - - // update parents - // update immediately if there is no parse_seq - // update only at barriers if there is parse_seq - if ((alloc->parse_seq_len == 0) || alloc->parse_seq[ind] == -1) { - int update_start = alloc->parse_seq_len ? last_barrier_pos : ind; - int update_end = alloc->parse_seq_len ? ind : ind + 1; - for (int i = update_start; i < update_end; i++) { - int node_i = alloc->parse_seq_len ? alloc->parse_seq[i] : i; - struct ggml_tensor * node = gf->nodes[node_i]; - - for (int j = 0; j < GGML_MAX_SRC; j++) { - struct ggml_tensor * parent = node->src[j]; - if (parent == NULL) { - break; + if (p_hn->n_children == 0 && p_hn->n_views == 0) { + if (ggml_is_view(parent)) { + struct ggml_tensor * view_src = parent->view_src; + struct hash_node * view_src_hn = hash_get(galloc, view_src); + view_src_hn->n_views -= 1; + AT_PRINTF("view_src %s: %d children, %d views\n", view_src->name, view_src_hn->n_children, view_src_hn->n_views); + if (view_src_hn->n_views == 0 && view_src_hn->n_children == 0) { + free_node(galloc, view_src); + } } - struct hash_node * p_hn = hash_get(ht, parent); - p_hn->n_children -= 1; - - //AT_PRINTF("parent %s: %d children, %d views\n", parent->name, parent->n_children, parent->n_views); - - if (p_hn->n_children == 0 && p_hn->n_views == 0) { - if (ggml_is_view(parent)) { - struct ggml_tensor * view_src = parent->view_src; - struct hash_node * view_src_hn = hash_get(ht, view_src); - view_src_hn->n_views -= 1; - AT_PRINTF("view_src %s: %d children, %d views\n", view_src->name, view_src_hn->n_children, view_src_hn->n_views); - if (view_src_hn->n_views == 0 && view_src_hn->n_children == 0 && view_src->data != node->data) { - ggml_allocr_free_tensor(alloc, view_src); - } - } - else { - if (parent->data != node->data) { - ggml_allocr_free_tensor(alloc, parent); - } - } + else { + free_node(galloc, parent); } } } - AT_PRINTF("\n"); - if (alloc->parse_seq_len) { - last_barrier_pos = ind + 1; - } } - } - // free graph outputs here that wouldn't be freed otherwise because they have no children - if (outputs != NULL && outputs[g] != NULL) { - for (int i = 0; outputs[g][i] != NULL; i++) { - struct ggml_tensor * output = outputs[g][i]; - AT_PRINTF("output: %s\n", output->name); - ggml_allocr_free_tensor(alloc, output); + AT_PRINTF("\n"); + if (parse_seq_len) { + last_barrier_pos = ind + 1; } } } - - return alloc->max_size; } -size_t ggml_allocr_alloc_graph(struct ggml_allocr * alloc, struct ggml_cgraph * graph) { - return ggml_allocr_alloc_graph_n(alloc, &graph, 1, NULL, NULL); +size_t ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, ggml_tallocr_t talloc, struct ggml_cgraph * graph) { + size_t hash_size = graph->visited_hash_table.size; + + // check if the hash table is initialized and large enough + if (galloc->hash_set.size < hash_size) { + if (galloc->hash_set.keys != NULL) { + free(galloc->hash_set.keys); + } + if (galloc->hash_values != NULL) { + free(galloc->hash_values); + } + galloc->hash_set.keys = malloc(sizeof(struct ggml_tensor *) * hash_size); + galloc->hash_set.size = hash_size; + galloc->hash_values = malloc(sizeof(struct hash_node) * hash_size); + } + + // reset hash table + memset(galloc->hash_set.keys, 0, sizeof(struct ggml_tensor *) * hash_size); + memset(galloc->hash_values, 0, sizeof(struct hash_node) * hash_size); + + galloc->talloc = talloc; + ggml_tallocr_alloc_graph_impl(galloc, graph); + galloc->talloc = NULL; + + size_t max_size = ggml_tallocr_max_size(talloc); + + return max_size; } -size_t ggml_allocr_max_size(struct ggml_allocr * alloc) { - return alloc->max_size; +void ggml_gallocr_alloc_graph_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, struct ggml_hash_set hash_set, ggml_tallocr_t * hash_node_talloc) { + const size_t hash_size = hash_set.size; + + GGML_ASSERT(hash_size >= (size_t)(graph->n_nodes + graph->n_leafs)); + + galloc->talloc = NULL; + + // alloc hash_values if needed + if (galloc->hash_values == NULL || galloc->hash_values_size < hash_size) { + free(galloc->hash_values); + galloc->hash_values = malloc(sizeof(struct hash_node) * hash_size); + galloc->hash_values_size = hash_size; + } + + // free hash_set.keys if needed + if (galloc->hash_set.keys != NULL) { + free(galloc->hash_set.keys); + } + galloc->hash_set = hash_set; + + // reset hash values + memset(galloc->hash_values, 0, sizeof(struct hash_node) * hash_size); + + galloc->hash_allocs = hash_node_talloc; + + ggml_tallocr_alloc_graph_impl(galloc, graph); + + // remove unowned resources + galloc->hash_set.keys = NULL; + galloc->hash_allocs = NULL; +} + +// legacy API wrapper + +struct ggml_allocr { + ggml_tallocr_t talloc; + ggml_gallocr_t galloc; +}; + +static ggml_allocr_t ggml_allocr_new_impl(ggml_tallocr_t talloc) { + ggml_allocr_t alloc = (ggml_allocr_t)malloc(sizeof(struct ggml_allocr)); + *alloc = (struct ggml_allocr) { + /*.talloc = */ talloc, + /*.galloc = */ ggml_gallocr_new(), + }; + return alloc; +} + +ggml_allocr_t ggml_allocr_new(void * data, size_t size, size_t alignment) { + return ggml_allocr_new_impl(ggml_tallocr_new(data, size, alignment)); +} + +ggml_allocr_t ggml_allocr_new_measure(size_t alignment) { + return ggml_allocr_new_impl(ggml_tallocr_new_measure(alignment)); +} + +ggml_allocr_t ggml_allocr_new_from_buffer(struct ggml_backend_buffer * buffer) { + return ggml_allocr_new_impl(ggml_tallocr_new_from_buffer(buffer)); +} + +ggml_allocr_t ggml_allocr_new_from_backend(struct ggml_backend * backend, size_t size) { + return ggml_allocr_new_impl(ggml_tallocr_new_from_backend(backend, size)); +} + +ggml_allocr_t ggml_allocr_new_measure_from_backend(struct ggml_backend * backend) { + return ggml_allocr_new_impl(ggml_tallocr_new_measure_from_backend(backend)); +} + +struct ggml_backend_buffer * ggml_allocr_get_buffer(ggml_allocr_t alloc) { + return ggml_tallocr_get_buffer(alloc->talloc); +} + +void ggml_allocr_set_parse_seq(ggml_allocr_t alloc, const int * list, int n) { + ggml_gallocr_set_parse_seq(alloc->galloc, list, n); +} + +void ggml_allocr_free(ggml_allocr_t alloc) { + ggml_gallocr_free(alloc->galloc); + ggml_tallocr_free(alloc->talloc); + free(alloc); +} + +bool ggml_allocr_is_measure(ggml_allocr_t alloc) { + return ggml_tallocr_is_measure(alloc->talloc); +} + +void ggml_allocr_reset(ggml_allocr_t alloc) { + ggml_tallocr_reset(alloc->talloc); +} + +void ggml_allocr_alloc(ggml_allocr_t alloc, struct ggml_tensor * tensor) { + ggml_tallocr_alloc(alloc->talloc, tensor); +} + +size_t ggml_allocr_max_size(ggml_allocr_t alloc) { + return ggml_tallocr_max_size(alloc->talloc); +} + +size_t ggml_allocr_alloc_graph(ggml_allocr_t alloc, struct ggml_cgraph * graph) { + return ggml_gallocr_alloc_graph(alloc->galloc, alloc->talloc, graph); } diff --git a/native/jni/src/ggml/ggml-alloc.h b/native/jni/src/ggml/ggml-alloc.h index e38758878..d662f2e6e 100644 --- a/native/jni/src/ggml/ggml-alloc.h +++ b/native/jni/src/ggml/ggml-alloc.h @@ -6,27 +6,79 @@ extern "C" { #endif +struct ggml_backend; struct ggml_backend_buffer; -GGML_API struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment); -GGML_API struct ggml_allocr * ggml_allocr_new_measure(size_t alignment); -GGML_API struct ggml_allocr * ggml_allocr_new_from_buffer(struct ggml_backend_buffer * buffer); +// +// Legacy API +// + +typedef struct ggml_allocr * ggml_allocr_t; + +// initialize allocator for use with CPU backend only +GGML_API ggml_allocr_t ggml_allocr_new(void * data, size_t size, size_t alignment); +GGML_API ggml_allocr_t ggml_allocr_new_measure(size_t alignment); + +// initialize allocator for use with ggml-backend +GGML_API ggml_allocr_t ggml_allocr_new_from_buffer(struct ggml_backend_buffer * buffer); +GGML_API ggml_allocr_t ggml_allocr_new_from_backend(struct ggml_backend * backend, size_t size); // allocates an owned buffer +GGML_API ggml_allocr_t ggml_allocr_new_measure_from_backend(struct ggml_backend * backend); + +GGML_API struct ggml_backend_buffer * ggml_allocr_get_buffer(ggml_allocr_t alloc); // tell the allocator to parse nodes following the order described in the list // you should call this if your graph are optimized to execute out-of-order -GGML_API void ggml_allocr_set_parse_seq(struct ggml_allocr * alloc, const int * list, int n); +GGML_API void ggml_allocr_set_parse_seq(ggml_allocr_t alloc, const int * list, int n); -GGML_API void ggml_allocr_free (struct ggml_allocr * alloc); -GGML_API bool ggml_allocr_is_measure (struct ggml_allocr * alloc); -GGML_API void ggml_allocr_reset (struct ggml_allocr * alloc); -GGML_API void ggml_allocr_alloc (struct ggml_allocr * alloc, struct ggml_tensor * tensor); -GGML_API size_t ggml_allocr_alloc_graph(struct ggml_allocr * alloc, struct ggml_cgraph * graph); -GGML_API size_t ggml_allocr_max_size (struct ggml_allocr * alloc); +GGML_API void ggml_allocr_free (ggml_allocr_t alloc); +GGML_API bool ggml_allocr_is_measure (ggml_allocr_t alloc); +GGML_API void ggml_allocr_reset (ggml_allocr_t alloc); +GGML_API void ggml_allocr_alloc (ggml_allocr_t alloc, struct ggml_tensor * tensor); +GGML_API size_t ggml_allocr_max_size (ggml_allocr_t alloc); -GGML_API size_t ggml_allocr_alloc_graph_n( - struct ggml_allocr * alloc, - struct ggml_cgraph ** graphs, int n_graphs, - struct ggml_tensor *** inputs, struct ggml_tensor *** outputs); +GGML_API size_t ggml_allocr_alloc_graph(ggml_allocr_t alloc, struct ggml_cgraph * graph); + +// +// ggml-backend v2 API +// + +// Seperate tensor and graph allocator objects +// This is necessary for multi-backend allocation because the graph allocator needs to use multiple tensor allocators +// The original API is kept as a wrapper around the new API + +// Tensor allocator +typedef struct ggml_tallocr * ggml_tallocr_t; + +GGML_API ggml_tallocr_t ggml_tallocr_new(void * data, size_t size, size_t alignment); +GGML_API ggml_tallocr_t ggml_tallocr_new_measure(size_t alignment); +GGML_API ggml_tallocr_t ggml_tallocr_new_from_buffer(struct ggml_backend_buffer * buffer); +GGML_API ggml_tallocr_t ggml_tallocr_new_from_backend(struct ggml_backend * backend, size_t size); // allocates an owned buffer +GGML_API ggml_tallocr_t ggml_tallocr_new_measure_from_backend(struct ggml_backend * backend); + +GGML_API struct ggml_backend_buffer * ggml_tallocr_get_buffer(ggml_tallocr_t talloc); + +GGML_API void ggml_tallocr_free (ggml_tallocr_t talloc); +GGML_API bool ggml_tallocr_is_measure (ggml_tallocr_t talloc); +GGML_API void ggml_tallocr_reset (ggml_tallocr_t talloc); +GGML_API void ggml_tallocr_alloc (ggml_tallocr_t talloc, struct ggml_tensor * tensor); +GGML_API size_t ggml_tallocr_max_size (ggml_tallocr_t talloc); + + +// Graph allocator +typedef struct ggml_gallocr * ggml_gallocr_t; + +GGML_API ggml_gallocr_t ggml_gallocr_new(void); +GGML_API void ggml_gallocr_free(ggml_gallocr_t galloc); + +GGML_API void ggml_gallocr_set_parse_seq(ggml_gallocr_t galloc, const int * list, int n); +GGML_API size_t ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, ggml_tallocr_t talloc, struct ggml_cgraph * graph); + +// Allocate tensors from the allocators given by the hash table +GGML_API void ggml_gallocr_alloc_graph_n( + ggml_gallocr_t galloc, + struct ggml_cgraph * graph, + struct ggml_hash_set hash_set, + ggml_tallocr_t * hash_node_talloc); #ifdef __cplusplus } diff --git a/native/jni/src/ggml/ggml-backend-impl.h b/native/jni/src/ggml/ggml-backend-impl.h new file mode 100644 index 000000000..a63fe3919 --- /dev/null +++ b/native/jni/src/ggml/ggml-backend-impl.h @@ -0,0 +1,87 @@ +#pragma once + +// ggml-backend internal header + +#include "ggml-backend.h" + +#ifdef __cplusplus +extern "C" { +#endif + +// +// Backend buffer +// + +typedef void * ggml_backend_buffer_context_t; + +struct ggml_backend_buffer_i { + void (*free_buffer) (ggml_backend_buffer_t buffer); + void * (*get_base) (ggml_backend_buffer_t buffer); // get base pointer + size_t (*get_alloc_size)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-allocation callback + void (*init_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // post-allocation callback + void (*free_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-free callback +}; + +struct ggml_backend_buffer { + struct ggml_backend_buffer_i iface; + + ggml_backend_t backend; + ggml_backend_buffer_context_t context; + + size_t size; +}; + +GGML_API ggml_backend_buffer_t ggml_backend_buffer_init( + struct ggml_backend * backend, + struct ggml_backend_buffer_i iface, + ggml_backend_buffer_context_t context, + size_t size); + +// +// Backend +// + +typedef void * ggml_backend_context_t; + +struct ggml_backend_i { + const char * (*get_name)(ggml_backend_t backend); + + void (*free)(ggml_backend_t backend); + + // buffer allocation + ggml_backend_buffer_t (*alloc_buffer)(ggml_backend_t backend, size_t size); + + // get buffer alignment + size_t (*get_alignment)(ggml_backend_t backend); + + // tensor data access + // these functions can be asynchronous, helper functions are provided for synchronous access that automatically call synchronize + void (*set_tensor_async)(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size); + void (*get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size); + void (*synchronize) (ggml_backend_t backend); + + // (optional) copy tensor between different backends, allow for single-copy tranfers + void (*cpy_tensor_from)(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst); + void (*cpy_tensor_to) (ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst); + + // compute graph with a plan + ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, struct ggml_cgraph * cgraph); + void (*graph_plan_free) (ggml_backend_t backend, ggml_backend_graph_plan_t plan); + void (*graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan); + + // compute graph without a plan + void (*graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph); + + // check if the backend supports an operation + bool (*supports_op)(ggml_backend_t backend, const struct ggml_tensor * op); +}; + +struct ggml_backend { + struct ggml_backend_i iface; + + ggml_backend_context_t context; +}; + +#ifdef __cplusplus +} +#endif diff --git a/native/jni/src/ggml/ggml-backend.c b/native/jni/src/ggml/ggml-backend.c index ca8d83daf..e49ffce1a 100644 --- a/native/jni/src/ggml/ggml-backend.c +++ b/native/jni/src/ggml/ggml-backend.c @@ -1,7 +1,9 @@ -#include "ggml-backend.h" +#include "ggml-backend-impl.h" #include "ggml-alloc.h" +#include "ggml-impl.h" #include +#include #include #include #include @@ -16,23 +18,27 @@ ggml_backend_buffer_t ggml_backend_buffer_init( struct ggml_backend * backend, struct ggml_backend_buffer_i iface, - ggml_backend_buffer_context_t context, - size_t size) { + ggml_backend_buffer_context_t context, + size_t size) { ggml_backend_buffer_t buffer = malloc(sizeof(struct ggml_backend_buffer)); GGML_ASSERT(iface.get_base != NULL); (*buffer) = (struct ggml_backend_buffer) { - /* .interface = */ iface, - /* .backend = */ backend, - /* .context = */ context, - /* .size = */ size, + /* .interface = */ iface, + /* .backend = */ backend, + /* .context = */ context, + /* .size = */ size, }; return buffer; } void ggml_backend_buffer_free(ggml_backend_buffer_t buffer) { + if (buffer == NULL) { + return; + } + if (buffer->iface.free_buffer != NULL) { buffer->iface.free_buffer(buffer); } @@ -43,15 +49,20 @@ size_t ggml_backend_buffer_get_alignment(ggml_backend_buffer_t buffer) { return ggml_backend_get_alignment(buffer->backend); } -void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) { - return buffer->iface.get_base(buffer); -} - size_t ggml_backend_buffer_get_size(ggml_backend_buffer_t buffer) { return buffer->size; } +void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) { + void * base = buffer->iface.get_base(buffer); + + GGML_ASSERT(base != NULL && "backend buffer base cannot be NULL"); + + return base; +} + size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) { + // get_alloc_size is optional, defaults to ggml_nbytes if (buffer->iface.get_alloc_size) { return buffer->iface.get_alloc_size(buffer, tensor); } @@ -59,12 +70,14 @@ size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct g } void ggml_backend_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) { + // init_tensor is optional if (buffer->iface.init_tensor) { buffer->iface.init_tensor(buffer, tensor); } } void ggml_backend_buffer_free_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) { + // free_tensor is optional if (buffer->iface.free_tensor) { buffer->iface.free_tensor(buffer, tensor); } @@ -73,14 +86,21 @@ void ggml_backend_buffer_free_tensor(ggml_backend_buffer_t buffer, struct ggml_t // backend ggml_backend_t ggml_get_backend(const struct ggml_tensor * tensor) { - return tensor->buffer->backend; + return tensor->buffer ? tensor->buffer->backend : NULL; } const char * ggml_backend_name(ggml_backend_t backend) { + if (backend == NULL) { + return "NULL"; + } return backend->iface.get_name(backend); } void ggml_backend_free(ggml_backend_t backend) { + if (backend == NULL) { + return; + } + backend->iface.free(backend); } @@ -101,13 +121,23 @@ void ggml_backend_tensor_get_async(const struct ggml_tensor * tensor, void * dat } void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) { - ggml_get_backend(tensor)->iface.set_tensor_async(ggml_get_backend(tensor), tensor, data, offset, size); - ggml_get_backend(tensor)->iface.synchronize(ggml_get_backend(tensor)); + ggml_backend_t backend = ggml_get_backend(tensor); + + GGML_ASSERT(tensor->data != NULL && "tensor not allocated"); + GGML_ASSERT(backend != NULL && "tensor backend not set"); + + backend->iface.set_tensor_async(backend, tensor, data, offset, size); + backend->iface.synchronize(backend); } void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) { - ggml_get_backend(tensor)->iface.get_tensor_async(ggml_get_backend(tensor), tensor, data, offset, size); - ggml_get_backend(tensor)->iface.synchronize(ggml_get_backend(tensor)); + ggml_backend_t backend = ggml_get_backend(tensor); + + GGML_ASSERT(tensor->data != NULL && "tensor not allocated"); + GGML_ASSERT(backend != NULL && "tensor backend not set"); + + backend->iface.get_tensor_async(backend, tensor, data, offset, size); + backend->iface.synchronize(backend); } void ggml_backend_synchronize(ggml_backend_t backend) { @@ -156,7 +186,7 @@ void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst //printf("dst: %s ne: [%d %d %d %d] nb: [%d %d %d %d]\n", dst->name, (int)dst->ne[0], (int)dst->ne[1], (int)dst->ne[2], (int)dst->ne[3], (int)dst->nb[0], (int)dst->nb[1], (int)dst->nb[2], (int)dst->nb[3]); GGML_ASSERT(ggml_are_same_layout(src, dst) && "cannot copy tensors with different layouts"); - // printf("cpy tensor %s from %s to %s (%lu bytes)\n", src->name, ggml_backend_name(src->backend), ggml_backend_name(dst->backend), ggml_nbytes(src)); + // fprintf(stderr, "cpy tensor %s from %s to %s (%lu bytes)\n", src->name, ggml_backend_name(src->backend), ggml_backend_name(dst->backend), ggml_nbytes(src)); if (src == dst) { return; @@ -170,9 +200,9 @@ void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst ggml_get_backend(src)->iface.cpy_tensor_to(ggml_get_backend(src)->context, src, dst); } else { // shouldn't be hit when copying from/to CPU - #ifndef NDEBUG +#ifndef NDEBUG fprintf(stderr, "ggml_backend_tensor_copy: neither cpy_tensor_from nor cpy_tensor_to are implemented for backends %s and %s, falling back to get/set\n", ggml_backend_name(src->buffer->backend), ggml_backend_name(dst->buffer->backend)); - #endif +#endif size_t nbytes = ggml_nbytes(src); void * data = malloc(nbytes); ggml_backend_tensor_get(src, data, 0, nbytes); @@ -192,7 +222,7 @@ struct ggml_backend_cpu_context { static const char * ggml_backend_cpu_name(ggml_backend_t backend) { return "CPU"; - UNUSED(backend); + UNUSED(backend); } static void ggml_backend_cpu_free(ggml_backend_t backend) { @@ -208,24 +238,24 @@ static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) { static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) { free(buffer->context); - UNUSED(buffer); + UNUSED(buffer); } static struct ggml_backend_buffer_i cpu_backend_buffer_i = { - /* .free_buffer = */ ggml_backend_cpu_buffer_free_buffer, - /* .get_base = */ ggml_backend_cpu_buffer_get_base, - /* .get_alloc_size = */ NULL, // defaults to ggml_nbytes - /* .init_tensor = */ NULL, // no initialization required - /* .free_tensor = */ NULL, // no cleanup required + /* .free_buffer = */ ggml_backend_cpu_buffer_free_buffer, + /* .get_base = */ ggml_backend_cpu_buffer_get_base, + /* .get_alloc_size = */ NULL, // defaults to ggml_nbytes + /* .init_tensor = */ NULL, // no initialization required + /* .free_tensor = */ NULL, // no cleanup required }; // for buffers from ptr, free is not called static struct ggml_backend_buffer_i cpu_backend_buffer_i_from_ptr = { - /* .free_buffer = */ NULL, // ptr is not owned by the buffer, so it does not need to be freed - /* .get_base = */ ggml_backend_cpu_buffer_get_base, - /* .get_alloc_size = */ NULL, // defaults to ggml_nbytes - /* .init_tensor = */ NULL, - /* .free_tensor = */ NULL, + /* .free_buffer = */ NULL, // ptr is not owned by the buffer, so it does not need to be freed + /* .get_base = */ ggml_backend_cpu_buffer_get_base, + /* .get_alloc_size = */ NULL, // defaults to ggml_nbytes + /* .init_tensor = */ NULL, + /* .free_tensor = */ NULL, }; static const size_t TENSOR_ALIGNMENT = 64; // should be enough for AVX 512 @@ -234,12 +264,14 @@ static ggml_backend_buffer_t ggml_backend_cpu_alloc_buffer(ggml_backend_t backen size += TENSOR_ALIGNMENT; // malloc may return an address that is not aligned void * data = malloc(size); // TODO: maybe use GGML_ALIGNED_MALLOC? + GGML_ASSERT(data != NULL && "failed to allocate buffer"); + return ggml_backend_buffer_init(backend, cpu_backend_buffer_i, data, size); } static size_t ggml_backend_cpu_get_alignment(ggml_backend_t backend) { return TENSOR_ALIGNMENT; - UNUSED(backend); + UNUSED(backend); } static void ggml_backend_cpu_set_tensor_async(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) { @@ -248,7 +280,7 @@ static void ggml_backend_cpu_set_tensor_async(ggml_backend_t backend, struct ggm memcpy((char *)tensor->data + offset, data, size); - UNUSED(backend); + UNUSED(backend); } static void ggml_backend_cpu_get_tensor_async(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) { @@ -257,24 +289,23 @@ static void ggml_backend_cpu_get_tensor_async(ggml_backend_t backend, const stru memcpy(data, (const char *)tensor->data + offset, size); - UNUSED(backend); + UNUSED(backend); } static void ggml_backend_cpu_synchronize(ggml_backend_t backend) { - UNUSED(backend); + UNUSED(backend); } static void ggml_backend_cpu_cpy_tensor_from(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst) { ggml_backend_tensor_get(src, dst->data, 0, ggml_nbytes(src)); - UNUSED(backend); + UNUSED(backend); } static void ggml_backend_cpu_cpy_tensor_to(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst) { - // for a backend such as CUDA that can queue async calls, it is ok to do this asynchronously, but it may not be the case for other backends - ggml_backend_tensor_set_async(dst, src->data, 0, ggml_nbytes(src)); + ggml_backend_tensor_set(dst, src->data, 0, ggml_nbytes(src)); - UNUSED(backend); + UNUSED(backend); } struct ggml_backend_plan_cpu { @@ -303,7 +334,7 @@ static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, ggml_backen free(cpu_plan->cplan.work_data); free(cpu_plan); - UNUSED(backend); + UNUSED(backend); } static void ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) { @@ -311,7 +342,7 @@ static void ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_bac ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan); - UNUSED(backend); + UNUSED(backend); } static void ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) { @@ -332,25 +363,25 @@ static void ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_c static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) { return true; - UNUSED(backend); - UNUSED(op); + UNUSED(backend); + UNUSED(op); } static struct ggml_backend_i cpu_backend_i = { - /* .get_name = */ ggml_backend_cpu_name, - /* .free = */ ggml_backend_cpu_free, - /* .alloc_buffer = */ ggml_backend_cpu_alloc_buffer, - /* .get_alignment = */ ggml_backend_cpu_get_alignment, - /* .set_tensor_async = */ ggml_backend_cpu_set_tensor_async, - /* .get_tensor_async = */ ggml_backend_cpu_get_tensor_async, - /* .synchronize = */ ggml_backend_cpu_synchronize, - /* .cpy_tensor_from = */ ggml_backend_cpu_cpy_tensor_from, - /* .cpy_tensor_to = */ ggml_backend_cpu_cpy_tensor_to, - /* .graph_plan_create = */ ggml_backend_cpu_graph_plan_create, - /* .graph_plan_free = */ ggml_backend_cpu_graph_plan_free, - /* .graph_plan_compute = */ ggml_backend_cpu_graph_plan_compute, - /* .graph_compute = */ ggml_backend_cpu_graph_compute, - /* .supports_op = */ ggml_backend_cpu_supports_op, + /* .get_name = */ ggml_backend_cpu_name, + /* .free = */ ggml_backend_cpu_free, + /* .alloc_buffer = */ ggml_backend_cpu_alloc_buffer, + /* .get_alignment = */ ggml_backend_cpu_get_alignment, + /* .set_tensor_async = */ ggml_backend_cpu_set_tensor_async, + /* .get_tensor_async = */ ggml_backend_cpu_get_tensor_async, + /* .synchronize = */ ggml_backend_cpu_synchronize, + /* .cpy_tensor_from = */ ggml_backend_cpu_cpy_tensor_from, + /* .cpy_tensor_to = */ ggml_backend_cpu_cpy_tensor_to, + /* .graph_plan_create = */ ggml_backend_cpu_graph_plan_create, + /* .graph_plan_free = */ ggml_backend_cpu_graph_plan_free, + /* .graph_plan_compute = */ ggml_backend_cpu_graph_plan_compute, + /* .graph_compute = */ ggml_backend_cpu_graph_compute, + /* .supports_op = */ ggml_backend_cpu_supports_op, }; ggml_backend_t ggml_backend_cpu_init(void) { @@ -363,8 +394,8 @@ ggml_backend_t ggml_backend_cpu_init(void) { ggml_backend_t cpu_backend = malloc(sizeof(struct ggml_backend)); *cpu_backend = (struct ggml_backend) { - /* .interface = */ cpu_backend_i, - /* .context = */ ctx + /* .interface = */ cpu_backend_i, + /* .context = */ ctx }; return cpu_backend; } @@ -383,3 +414,537 @@ void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) { ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(ggml_backend_t backend_cpu, void * ptr, size_t size) { return ggml_backend_buffer_init(backend_cpu, cpu_backend_buffer_i_from_ptr, ptr, size); } + +// scheduler + +#define GGML_MAX_BACKENDS 4 +#define GGML_MAX_SPLITS 256 +#define GGML_MAX_SPLIT_INPUTS 16 + +struct ggml_backend_sched_split { + ggml_tallocr_t tallocr; + int i_start; + int i_end; + struct ggml_tensor * inputs[GGML_MAX_SPLIT_INPUTS]; + int n_inputs; + struct ggml_cgraph * graph; +}; + +struct ggml_backend_sched { + int n_backends; + ggml_backend_t backends[GGML_MAX_BACKENDS]; + ggml_tallocr_t tallocs[GGML_MAX_BACKENDS]; + + ggml_gallocr_t galloc; + + struct ggml_hash_set hash_set; + ggml_tallocr_t * node_talloc; // [hash_set.size] + struct ggml_tensor * (* node_copies)[GGML_MAX_BACKENDS]; // [hash_set.size][GGML_MAX_BACKENDS] + + struct ggml_cgraph * graph; + struct ggml_backend_sched_split splits[GGML_MAX_SPLITS]; + int n_splits; + + struct ggml_context * ctx; + + // align context_buffer to GGML_MEM_ALIGN +#ifdef _MSC_VER + __declspec(align(GGML_MEM_ALIGN)) +#else + __attribute__((aligned(GGML_MEM_ALIGN))) +#endif + char context_buffer[GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS*sizeof(struct ggml_tensor) + GGML_MAX_SPLITS*sizeof(struct ggml_cgraph)]; +}; + +#define hash_id(node) ggml_hash_find_or_insert(sched->hash_set, node) +#define node_allocr(node) sched->node_talloc[hash_id(node)] + +static bool ggml_is_view_op(enum ggml_op op) { + return op == GGML_OP_VIEW || op == GGML_OP_RESHAPE || op == GGML_OP_PERMUTE || op == GGML_OP_TRANSPOSE; +} + +// returns the priority of the backend, lower is better +static int sched_backend_prio(ggml_backend_sched_t sched, ggml_backend_t backend) { + for (int i = 0; i < sched->n_backends; i++) { + if (sched->backends[i] == backend) { + return i; + } + } + return INT_MAX; +} + +static int sched_allocr_prio(ggml_backend_sched_t sched, ggml_tallocr_t allocr) { + for (int i = 0; i < sched->n_backends; i++) { + if (sched->tallocs[i] == allocr) { + return i; + } + } + return INT_MAX; +} + +// returns the backend that should be used for the node based on the current locations +char causes[GGML_DEFAULT_GRAPH_SIZE*4 + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS][128]; // debug, remove +static ggml_backend_t sched_backend_from_cur(ggml_backend_sched_t sched, struct ggml_tensor * node) { + // if the dst tensor is already allocated in a buffer, we must assume that it is critical to keep it there + // ie. kv cache updates + // note that this doesn't allow fallback to CPU. need to add output tensors to the splits to copy the data back to the original backend. + // dst + ggml_backend_t cur_backend = ggml_get_backend(node); + if (cur_backend != NULL) { + sprintf(causes[hash_id(node)], "1.dst"); + return cur_backend; + } + + // view_src + if (node->view_src != NULL && ggml_get_backend(node->view_src) != NULL) { + sprintf(causes[hash_id(node)], "1.vsrc"); + return ggml_get_backend(node->view_src); + } + + // src + int cur_prio = INT_MAX; + size_t cur_size = 0; + + for (int i = 0; i < GGML_MAX_SRC; i++) { + const struct ggml_tensor * src = node->src[i]; + if (src == NULL) { + break; + } + ggml_backend_t src_backend = ggml_get_backend(src); + if (src_backend != NULL) { + int src_prio = sched_backend_prio(sched, src_backend); + size_t src_size = ggml_nbytes(src); + if (src_prio < cur_prio && src_size >= cur_size) { + cur_prio = src_prio; + cur_size = src_size; + cur_backend = src_backend; + sprintf(causes[hash_id(node)], "1.src%d", i); + } + } + } + return cur_backend; +} + +static char * fmt_size(size_t size) { + static char buffer[128]; + if (size >= 1024*1024) { + sprintf(buffer, "%zuM", size/1024/1024); + } else { + sprintf(buffer, "%zuK", size/1024); + } + return buffer; +} + +static void sched_print_assignments(ggml_backend_sched_t sched, struct ggml_cgraph * graph) { + int cur_split = 0; + for (int i = 0; i < graph->n_nodes; i++) { + if (cur_split < sched->n_splits && i == sched->splits[cur_split].i_start) { + ggml_backend_t split_backend = ggml_tallocr_get_buffer(sched->splits[cur_split].tallocr)->backend; + fprintf(stderr, "\n## SPLIT #%d: %s # %d inputs: ", cur_split, ggml_backend_name(split_backend), sched->splits[cur_split].n_inputs); + for (int j = 0; j < sched->splits[cur_split].n_inputs; j++) { + fprintf(stderr, "[%s (%5.5s)] ", sched->splits[cur_split].inputs[j]->name, fmt_size(ggml_nbytes(sched->splits[cur_split].inputs[j]))); + } + fprintf(stderr, "\n"); + cur_split++; + } + struct ggml_tensor * node = graph->nodes[i]; + if (ggml_is_view_op(node->op)) { + continue; + } + ggml_tallocr_t node_allocr = node_allocr(node); + ggml_backend_t node_backend = node_allocr ? ggml_tallocr_get_buffer(node_allocr)->backend : NULL; + fprintf(stderr, "node #%3d (%10.10s): %20.20s (%4.4s) [%4.4s %8.8s]:", i, ggml_op_name(node->op), node->name, fmt_size(ggml_nbytes(node)), node_allocr ? ggml_backend_name(node_backend) : "NULL", causes[hash_id(node)]); + for (int j = 0; j < GGML_MAX_SRC; j++) { + struct ggml_tensor * src = node->src[j]; + if (src == NULL) { + break; + } + ggml_tallocr_t src_allocr = node_allocr(src); + ggml_backend_t src_backend = src_allocr ? ggml_tallocr_get_buffer(src_allocr)->backend : NULL; + fprintf(stderr, " %20.20s (%4.4s) [%4.4s %8.8s]", src->name, fmt_size(ggml_nbytes(src)), src_backend ? ggml_backend_name(src_backend) : "NULL", causes[hash_id(src)]); + } + fprintf(stderr, "\n"); + } +} + +// creates a copy of the tensor with the same memory layout +static struct ggml_tensor * ggml_dup_tensor_layout(struct ggml_context * ctx, const struct ggml_tensor * tensor) { + struct ggml_tensor * dup = ggml_dup_tensor(ctx, tensor); + for (int i = 0; i < GGML_MAX_DIMS; i++) { + dup->nb[i] = tensor->nb[i]; + } + return dup; +} + +// assigns backends to ops and splits the graph into subgraphs that can be computed on the same backend +// TODO: merge passes +static void sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) { + // reset state + size_t hash_size = sched->hash_set.size; + memset(sched->hash_set.keys, 0, sizeof(sched->hash_set.keys[0]) * hash_size); + memset(sched->node_talloc, 0, sizeof(sched->node_talloc[0]) * hash_size); + memset(sched->node_copies, 0, sizeof(sched->node_copies[0]) * hash_size); + sched->n_splits = 0; + + struct ggml_init_params params = { + /*.mem_size = */ sizeof(sched->context_buffer), + /*.mem_buffer = */ sched->context_buffer, + /*.no_alloc = */ true + }; + + if (sched->ctx != NULL) { + ggml_free(sched->ctx); + } + + sched->ctx = ggml_init(params); + + // pass 1: assign backends to ops with allocated inputs + for (int i = 0; i < graph->n_leafs; i++) { + struct ggml_tensor * leaf = graph->leafs[i]; + if (node_allocr(leaf) != NULL) { + // do not overwrite user assignments + continue; + } + ggml_backend_t leaf_backend = ggml_get_backend(leaf); + if (leaf_backend == NULL && leaf->view_src != NULL) { + leaf_backend = ggml_get_backend(leaf->view_src); + } + if (leaf_backend != NULL) { + node_allocr(leaf) = ggml_backend_sched_get_tallocr(sched, leaf_backend); + } + } + + for (int i = 0; i < graph->n_nodes; i++) { + struct ggml_tensor * node = graph->nodes[i]; + if (node_allocr(node) != NULL) { + // do not overwrite user assignments + continue; + } + ggml_backend_t node_backend = sched_backend_from_cur(sched, node); + if (node_backend != NULL) { + node_allocr(node) = ggml_backend_sched_get_tallocr(sched, node_backend); + } + } + //printf("PASS 1 ASSIGNMENTS\n"); sched_print_assignments(sched, graph); + + // pass 2: assign backends to ops from current assignments + // TODO: + // - reuse sched_backend_from_cur + for (int i = 0; i < graph->n_nodes; i++) { + struct ggml_tensor * node = graph->nodes[i]; + ggml_tallocr_t node_allocr = node_allocr(node); + if (node_allocr == NULL) { + int cur_prio = INT_MAX; + size_t cur_size = 0; + for (int j = 0; j < GGML_MAX_SRC; j++) { + struct ggml_tensor * src = node->src[j]; + if (src == NULL) { + break; + } + ggml_tallocr_t src_allocr = node_allocr(src); + if (src_allocr != NULL) { + int src_prio = sched_allocr_prio(sched, src_allocr); + size_t src_size = ggml_nbytes(src); + if (src_prio < cur_prio && src_size >= cur_size) { + cur_prio = src_prio; + cur_size = src_size; + node_allocr = src_allocr; + sprintf(causes[hash_id(node)], "2.src%d", j); + } + } + } + if (node_allocr != NULL) { + node_allocr(node) = node_allocr; + } + } + } + //printf("PASS 2 ASSIGNMENTS\n"); sched_print_assignments(sched, graph); + + // pass 3: assign backends to remaining src from dst (should only be leafs) + for (int i = 0; i < graph->n_nodes; i++) { + struct ggml_tensor * node = graph->nodes[i]; + ggml_tallocr_t node_allocr = node_allocr(node); + for (int j = 0; j < GGML_MAX_SRC; j++) { + struct ggml_tensor * src = node->src[j]; + if (src == NULL) { + break; + } + ggml_tallocr_t src_allocr = node_allocr(src); + if (src_allocr == NULL) { + node_allocr(src) = node_allocr; + } + } + } + //printf("PASS 3 ASSIGNMENTS\n"); sched_print_assignments(sched, graph); + + // pass 4: split graph, find tensors that need to be copied + // TODO: + // - when switching from a less preferred backend to a more preferred backend, check if it is possible to move the switch to an earlier point for the same cost + // find first backend + int cur_split = 0; + for (int i = 0; i < graph->n_nodes; i++) { + struct ggml_tensor * node = graph->nodes[i]; + if (node->view_src == NULL) { + sched->splits[0].tallocr = node_allocr(node); + break; + } + } + sched->splits[0].i_start = 0; + sched->splits[0].n_inputs = 0; + memset(sched->splits[0].inputs, 0, sizeof(sched->splits[0].inputs)); //HACK + ggml_tallocr_t cur_allocr = sched->splits[0].tallocr; + size_t cur_backend_id = sched_allocr_prio(sched, cur_allocr); + for (int i = 0; i < graph->n_nodes; i++) { + struct ggml_tensor * node = graph->nodes[i]; + + if (ggml_is_view_op(node->op)) { + continue; + } + + ggml_tallocr_t node_allocr = node_allocr(node); + + if (node_allocr != cur_allocr) { + sched->splits[cur_split].i_end = i; + cur_split++; + GGML_ASSERT(cur_split < GGML_MAX_SPLITS); + sched->splits[cur_split].tallocr = node_allocr; + sched->splits[cur_split].i_start = i; + sched->splits[cur_split].n_inputs = 0; + memset(sched->splits[cur_split].inputs, 0, sizeof(sched->splits[cur_split].inputs)); //HACK + cur_allocr = node_allocr; + cur_backend_id = sched_allocr_prio(sched, cur_allocr); + } + + // find inputs that are not on the same backend + for (int j = 0; j < GGML_MAX_SRC; j++) { + struct ggml_tensor * src = node->src[j]; + if (src == NULL) { + break; + } + ggml_tallocr_t src_allocr = node_allocr(src); + if (src_allocr != node_allocr) { + int n_inputs = sched->splits[cur_split].n_inputs++; + GGML_ASSERT(n_inputs < GGML_MAX_SPLIT_INPUTS); + sched->splits[cur_split].inputs[n_inputs] = (struct ggml_tensor *)src; + + // create copies + size_t id = hash_id(src); + if (sched->node_copies[id][cur_backend_id] == NULL) { + struct ggml_tensor * tensor_copy = ggml_dup_tensor_layout(sched->ctx, src); + sched->node_copies[id][cur_backend_id] = tensor_copy; + node_allocr(tensor_copy) = cur_allocr; + ggml_backend_t backend = ggml_tallocr_get_buffer(cur_allocr)->backend; + ggml_format_name(tensor_copy, "%s#%s", ggml_backend_name(backend), src->name); + } + node->src[j] = sched->node_copies[id][cur_backend_id]; + } + } + } + sched->splits[cur_split].i_end = graph->n_nodes; + sched->n_splits = cur_split + 1; + + //fprintf(stderr, "PASS 4 ASSIGNMENTS\n"); sched_print_assignments(sched, graph); fflush(stdout); + +#if 1 + // sanity check: all sources should have the same backend as the node + for (int i = 0; i < graph->n_nodes; i++) { + struct ggml_tensor * node = graph->nodes[i]; + ggml_tallocr_t node_allocr = node_allocr(node); + if (node_allocr == NULL) { + fprintf(stderr, "!!!!!!! %s has no backend\n", node->name); + } + for (int j = 0; j < GGML_MAX_SRC; j++) { + struct ggml_tensor * src = node->src[j]; + if (src == NULL) { + break; + } + ggml_tallocr_t src_allocr = node_allocr(src); + if (src_allocr != node_allocr /* && src_backend != NULL */) { // ignore nulls for now + fprintf(stderr, "!!!! %s has backend %s, src %d (%s) has backend %s\n", + node->name, node_allocr ? ggml_backend_name(ggml_tallocr_get_buffer(node_allocr)->backend) : "NULL", + j, src->name, src_allocr ? ggml_backend_name(ggml_tallocr_get_buffer(src_allocr)->backend) : "NULL"); + } + } + } +#endif + + // create copies of the graph for each split + // FIXME: avoid this copy, pass split inputs to ggml_gallocr_alloc_graph_n in some other way + struct ggml_cgraph * graph_copy = ggml_new_graph_custom(sched->ctx, graph->n_nodes + sched->n_splits*GGML_MAX_SPLIT_INPUTS, false); + for (int i = 0; i < sched->n_splits; i++) { + struct ggml_backend_sched_split * split = &sched->splits[i]; + split->graph = ggml_graph_view(sched->ctx, graph, split->i_start, split->i_end); + + // add inputs to the graph copy so that they are allocated by ggml-alloc at the start of the split + for (int j = 0; j < split->n_inputs; j++) { + struct ggml_tensor * input = split->inputs[j]; + struct ggml_tensor * input_cpy = sched->node_copies[hash_id(input)][sched_allocr_prio(sched, split->tallocr)]; + input_cpy->src[0] = input; + graph_copy->nodes[graph_copy->n_nodes++] = input_cpy; + } + + for (int j = split->i_start; j < split->i_end; j++) { + graph_copy->nodes[graph_copy->n_nodes++] = graph->nodes[j]; + } + } + sched->graph = graph_copy; +} + +static void sched_alloc_splits(ggml_backend_sched_t sched) { + ggml_gallocr_alloc_graph_n( + sched->galloc, + sched->graph, + sched->hash_set, + sched->node_talloc); +} + +static void sched_compute_splits(ggml_backend_sched_t sched) { + uint64_t copy_us[GGML_MAX_BACKENDS] = {0}; + uint64_t compute_us[GGML_MAX_BACKENDS] = {0}; + + struct ggml_backend_sched_split * splits = sched->splits; + + for (int i = 0; i < sched->n_splits; i++) { + struct ggml_backend_sched_split * split = &splits[i]; + ggml_backend_t split_backend = ggml_tallocr_get_buffer(split->tallocr)->backend; + int split_backend_id = sched_backend_prio(sched, split_backend); + + // copy the input tensors to the split backend + uint64_t copy_start_us = ggml_time_us(); + for (int j = 0; j < split->n_inputs; j++) { + struct ggml_tensor * input_cpy = sched->node_copies[hash_id(split->inputs[j])][sched_backend_prio(sched, split_backend)]; + if (split->inputs[j]->buffer == NULL) { + if (split->inputs[j]->view_src == NULL) { + fprintf(stderr, "input %s has no buffer and no view_src\n", split->inputs[j]->name); + exit(1); + } + struct ggml_tensor * view = split->inputs[j]; + view->backend = view->view_src->backend; + view->buffer = view->view_src->buffer; + view->data = (char *)view->view_src->data + view->view_offs; + ggml_backend_buffer_init_tensor(ggml_backend_sched_get_buffer(sched, view->buffer->backend), view); + } + if (input_cpy->buffer == NULL) { + fprintf(stderr, "input_cpy %s has no buffer\n", input_cpy->name); + exit(1); + } + GGML_ASSERT(split->inputs[j]->buffer->backend != input_cpy->buffer->backend); + GGML_ASSERT(input_cpy->buffer->backend == split_backend); + ggml_backend_tensor_copy(split->inputs[j], input_cpy); + } + // ggml_backend_synchronize(split_backend); + int64_t copy_end_us = ggml_time_us(); + copy_us[split_backend_id] += copy_end_us - copy_start_us; + +#if 0 + char split_filename[GGML_MAX_NAME]; + snprintf(split_filename, GGML_MAX_NAME, "split_%i_%s.dot", i, ggml_backend_name(split_backend)); + ggml_graph_dump_dot(split->graph, NULL, split_filename); +#endif + + uint64_t compute_start_us = ggml_time_us(); + ggml_backend_graph_compute(split_backend, split->graph); + // ggml_backend_synchronize(split_backend); + uint64_t compute_end_us = ggml_time_us(); + compute_us[split_backend_id] += compute_end_us - compute_start_us; + } + +#if 0 + // per-backend timings + fprintf(stderr, "sched_compute_splits times (%d splits):\n", sched->n_splits); + for (int i = 0; i < sched->n_backends; i++) { + if (copy_us[i] > 0 || compute_us[i] > 0) { + fprintf(stderr, "\t%5.5s: %lu us copy, %lu us compute\n", ggml_backend_name(sched->backends[i]), copy_us[i], compute_us[i]); + } + } +#endif +} + +static void sched_reset(ggml_backend_sched_t sched) { + for (int i = 0; i < sched->n_backends; i++) { + ggml_tallocr_reset(sched->tallocs[i]); + } +} + +ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, int n_backends) { + GGML_ASSERT(n_backends <= GGML_MAX_BACKENDS); + + struct ggml_backend_sched * sched = malloc(sizeof(struct ggml_backend_sched)); + memset(sched, 0, sizeof(struct ggml_backend_sched)); + + fprintf(stderr, "ggml_backend_sched size: %lu KB\n", sizeof(struct ggml_backend_sched)/1024); + + sched->n_backends = n_backends; + for (int i = 0; i < n_backends; i++) { + sched->backends[i] = backends[i]; + } + + sched->galloc = ggml_gallocr_new(); + + // init measure allocs for each backend + for (int i = 0; i < n_backends; i++) { + sched->tallocs[i] = ggml_tallocr_new_measure_from_backend(backends[i]); + } + + return sched; +} + +void ggml_backend_sched_free(ggml_backend_sched_t sched) { + if (sched == NULL) { + return; + } + for (int i = 0; i < sched->n_backends; i++) { + ggml_tallocr_free(sched->tallocs[i]); + } + ggml_gallocr_free(sched->galloc); + free(sched->hash_set.keys); + free(sched->node_talloc); + free(sched->node_copies); + free(sched); +} + +void ggml_backend_sched_init_measure(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph) { + // initialize hash tables + size_t hash_size = measure_graph->visited_hash_table.size + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS; + sched->hash_set.size = hash_size; + sched->hash_set.keys = malloc(sizeof(sched->hash_set.keys[0]) * hash_size); + sched->node_talloc = malloc(sizeof(sched->node_talloc[0]) * hash_size); + sched->node_copies = malloc(sizeof(sched->node_copies[0]) * hash_size); + + sched_split_graph(sched, measure_graph); + sched_alloc_splits(sched); + + // allocate buffers and reset allocators + for (int i = 0; i < sched->n_backends; i++) { + size_t size = ggml_tallocr_max_size(sched->tallocs[i]); + ggml_tallocr_free(sched->tallocs[i]); + sched->tallocs[i] = ggml_tallocr_new_from_backend(sched->backends[i], size); + } + + sched_reset(sched); +} + +void ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph) { + GGML_ASSERT(sched->hash_set.size >= graph->visited_hash_table.size + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS); + + sched_split_graph(sched, graph); + sched_alloc_splits(sched); + sched_compute_splits(sched); + sched_reset(sched); +} + +ggml_tallocr_t ggml_backend_sched_get_tallocr(ggml_backend_sched_t sched, ggml_backend_t backend) { + int backend_index = sched_backend_prio(sched, backend); + return sched->tallocs[backend_index]; +} + +ggml_backend_buffer_t ggml_backend_sched_get_buffer(ggml_backend_sched_t sched, ggml_backend_t backend) { + int backend_index = sched_backend_prio(sched, backend); + return ggml_tallocr_get_buffer(sched->tallocs[backend_index]); +} + +void ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend) { + int backend_index = sched_backend_prio(sched, backend); + GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends); + node_allocr(node) = sched->tallocs[backend_index]; +} diff --git a/native/jni/src/ggml/ggml-backend.h b/native/jni/src/ggml/ggml-backend.h index da134b0db..dd28a750a 100644 --- a/native/jni/src/ggml/ggml-backend.h +++ b/native/jni/src/ggml/ggml-backend.h @@ -1,142 +1,135 @@ #pragma once #include "ggml.h" +#include "ggml-alloc.h" #ifdef __cplusplus extern "C" { #endif - struct ggml_backend; - struct ggml_backend_buffer; - // type-erased backend-specific types / wrappers - typedef void * ggml_backend_context_t; - typedef void * ggml_backend_graph_plan_t; - typedef void * ggml_backend_buffer_context_t; +// +// Backend buffer +// - // avoid accessing internals of these types - typedef struct ggml_backend * ggml_backend_t; - typedef struct ggml_backend_buffer * ggml_backend_buffer_t; +struct ggml_backend_buffer; +typedef struct ggml_backend_buffer * ggml_backend_buffer_t; - // - // backend buffer - // +// backend buffer functions +GGML_API void ggml_backend_buffer_free (ggml_backend_buffer_t buffer); +GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer); +GGML_API void * ggml_backend_buffer_get_base (ggml_backend_buffer_t buffer); +GGML_API size_t ggml_backend_buffer_get_size (ggml_backend_buffer_t buffer); +GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); +GGML_API void ggml_backend_buffer_init_tensor (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); +GGML_API void ggml_backend_buffer_free_tensor (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); - struct ggml_backend_buffer_i { - void (*free_buffer) (ggml_backend_buffer_t buffer); - void * (*get_base) (ggml_backend_buffer_t buffer); // get base pointer - size_t (*get_alloc_size)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-allocation callback - void (*init_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // post-allocation callback - void (*free_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-free callback - }; +// +// Backend +// - // TODO: hide behind API - struct ggml_backend_buffer { - struct ggml_backend_buffer_i iface; +struct ggml_backend; +typedef struct ggml_backend * ggml_backend_t; +typedef void * ggml_backend_graph_plan_t; - ggml_backend_t backend; - ggml_backend_buffer_context_t context; +GGML_API ggml_backend_t ggml_get_backend(const struct ggml_tensor * tensor); - size_t size; - }; +GGML_API const char * ggml_backend_name(ggml_backend_t backend); +GGML_API void ggml_backend_free(ggml_backend_t backend); - // backend buffer functions - GGML_API ggml_backend_buffer_t ggml_backend_buffer_init( - struct ggml_backend * backend, - struct ggml_backend_buffer_i iface, - ggml_backend_buffer_context_t context, - size_t size); +GGML_API ggml_backend_buffer_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size); - GGML_API void ggml_backend_buffer_free (ggml_backend_buffer_t buffer); - GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer); - GGML_API void * ggml_backend_buffer_get_base (ggml_backend_buffer_t buffer); - GGML_API size_t ggml_backend_buffer_get_size (ggml_backend_buffer_t buffer); - GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); - GGML_API void ggml_backend_buffer_init_tensor (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); - GGML_API void ggml_backend_buffer_free_tensor (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); +GGML_API size_t ggml_backend_get_alignment(ggml_backend_t backend); - // - // backend - // +GGML_API void ggml_backend_tensor_set_async( struct ggml_tensor * tensor, const void * data, size_t offset, size_t size); +GGML_API void ggml_backend_tensor_get_async(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size); - struct ggml_backend_i { - const char * (*get_name)(ggml_backend_t backend); +GGML_API void ggml_backend_tensor_set( struct ggml_tensor * tensor, const void * data, size_t offset, size_t size); +GGML_API void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size); - void (*free)(ggml_backend_t backend); +GGML_API void ggml_backend_synchronize(ggml_backend_t backend); - // buffer allocation - ggml_backend_buffer_t (*alloc_buffer)(ggml_backend_t backend, size_t size); +GGML_API ggml_backend_graph_plan_t ggml_backend_graph_plan_create (ggml_backend_t backend, struct ggml_cgraph * cgraph); - // get buffer alignment - size_t (*get_alignment)(ggml_backend_t backend); +GGML_API void ggml_backend_graph_plan_free (ggml_backend_t backend, ggml_backend_graph_plan_t plan); +GGML_API void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan); +GGML_API void ggml_backend_graph_compute (ggml_backend_t backend, struct ggml_cgraph * cgraph); +GGML_API bool ggml_backend_supports_op (ggml_backend_t backend, const struct ggml_tensor * op); - // tensor data access - // these functions can be asynchronous, helper functions are provided for synchronous access that automatically call synchronize - void (*set_tensor_async)(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size); - void (*get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size); - void (*synchronize) (ggml_backend_t backend); +// tensor copy between different backends +GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst); - // (optional) copy tensor between different backends, allow for single-copy tranfers - void (*cpy_tensor_from)(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst); - void (*cpy_tensor_to) (ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst); +// +// CPU backend +// - // compute graph with a plan - ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, struct ggml_cgraph * cgraph); - void (*graph_plan_free) (ggml_backend_t backend, ggml_backend_graph_plan_t plan); - void (*graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan); +GGML_API ggml_backend_t ggml_backend_cpu_init(void); - // compute graph without a plan - void (*graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph); +GGML_API bool ggml_backend_is_cpu(ggml_backend_t backend); +GGML_API void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads); - // check if the backend supports an operation - bool (*supports_op)(ggml_backend_t backend, const struct ggml_tensor * op); - }; +// Create a backend buffer from an existing pointer +GGML_API ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(ggml_backend_t backend_cpu, void * ptr, size_t size); - // TODO: hide behind API - struct ggml_backend { - struct ggml_backend_i iface; - ggml_backend_context_t context; - }; +// +// Backend scheduler +// - // backend helper functions - GGML_API ggml_backend_t ggml_get_backend(const struct ggml_tensor * tensor); +// The backend scheduler allows for multiple backends to be used together +// Handles compute buffer allocation, assignment of tensors to backends, and copying of tensors between backends +// The backends are selected based on: +// - the backend that supports the operation +// - the location of the pre-allocated tensors (e.g. the weights) +/* + Example usage: - GGML_API const char * ggml_backend_name(ggml_backend_t backend); - GGML_API void ggml_backend_free(ggml_backend_t backend); + sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, num_backends); + // sched is initialized with measure allocators and cannot be used until allocated with a measure graph - GGML_API ggml_backend_buffer_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size); + // initialize buffers from a measure graph + measure_graph = build_graph(sched); // use the allocr to allocate inputs as needed - GGML_API size_t ggml_backend_get_alignment(ggml_backend_t backend); + // in build_graph: + build_graph(...) { + // allocating tensors in a specific backend (optional, recommended: pre-allocate inputs in a different buffer) + alloc_cpu = ggml_backend_sched_get_allocr(sched, backend_cpu); + ggml_allocr_alloc(alloc_cpu, tensor); - GGML_API void ggml_backend_tensor_set_async( struct ggml_tensor * tensor, const void * data, size_t offset, size_t size); - GGML_API void ggml_backend_tensor_get_async(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size); + // manually assigning nodes to a backend (optional, shouldn't be needed in most cases) + struct ggml_tensor * node = ggml_mul_mat(ctx, ...); + ggml_backend_sched_set_node_backend(sched, node, backend_gpu); + } - GGML_API void ggml_backend_tensor_set( struct ggml_tensor * tensor, const void * data, size_t offset, size_t size); - GGML_API void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size); + // allocate backend buffers from measure graph + ggml_backend_sched_init_measure(sched, measure_graph); - GGML_API void ggml_backend_synchronize(ggml_backend_t backend); + // the scheduler is now ready to compute graphs - GGML_API ggml_backend_graph_plan_t ggml_backend_graph_plan_create (ggml_backend_t backend, struct ggml_cgraph * cgraph); + // compute + graph = build_graph(sched); + ggml_backend_sched_graph_compute(sched, graph); +*/ - GGML_API void ggml_backend_graph_plan_free (ggml_backend_t backend, ggml_backend_graph_plan_t plan); - GGML_API void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan); - GGML_API void ggml_backend_graph_compute (ggml_backend_t backend, struct ggml_cgraph * cgraph); - GGML_API bool ggml_backend_supports_op (ggml_backend_t backend, const struct ggml_tensor * op); +struct ggml_backend_sched; +typedef struct ggml_backend_sched * ggml_backend_sched_t; - // tensor copy between different backends - GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst); +// Initialize a backend scheduler +GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, int n_backends); - // - // CPU backend - // +GGML_API void ggml_backend_sched_free(ggml_backend_sched_t sched); - GGML_API ggml_backend_t ggml_backend_cpu_init(void); +// Initialize backend buffers from a measure graph +GGML_API void ggml_backend_sched_init_measure(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph); - GGML_API bool ggml_backend_is_cpu(ggml_backend_t backend); +GGML_API ggml_tallocr_t ggml_backend_sched_get_tallocr(ggml_backend_sched_t sched, ggml_backend_t backend); +GGML_API ggml_backend_buffer_t ggml_backend_sched_get_buffer (ggml_backend_sched_t sched, ggml_backend_t backend); - GGML_API void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads); +GGML_API void ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend); - GGML_API ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(ggml_backend_t backend_cpu, void * ptr, size_t size); +// Allocate a graph on the backend scheduler +GGML_API void ggml_backend_sched_graph_compute( + ggml_backend_sched_t sched, + struct ggml_cgraph * graph); #ifdef __cplusplus } diff --git a/native/jni/src/ggml/ggml-impl.h b/native/jni/src/ggml/ggml-impl.h index 5ec18a50c..7fa83ee5e 100644 --- a/native/jni/src/ggml/ggml-impl.h +++ b/native/jni/src/ggml/ggml-impl.h @@ -39,12 +39,6 @@ extern "C" { #endif #endif -#undef MIN -#undef MAX - -#define MIN(a, b) ((a) < (b) ? (a) : (b)) -#define MAX(a, b) ((a) > (b) ? (a) : (b)) - // 16-bit float // on Arm, we use __fp16 // on x86, we use uint16_t @@ -173,7 +167,7 @@ static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) { const uint32_t denormalized_cutoff = UINT32_C(1) << 27; const uint32_t result = sign | - (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value)); + (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value)); return fp32_from_bits(result); } @@ -230,7 +224,19 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) { #endif - // TODO: backend v2 PR +#define GGML_HASHTABLE_FULL ((size_t)-1) +#define GGML_HASHTABLE_ALREADY_EXISTS ((size_t)-2) + +bool ggml_hash_contains (const struct ggml_hash_set hash_set, struct ggml_tensor * key); + +// returns GGML_HASHTABLE_FULL if table is full, otherwise the current index of the key or where it should be inserted +size_t ggml_hash_find (const struct ggml_hash_set hash_set, struct ggml_tensor * key); + +// returns GGML_HAHSHTABLE_ALREADY_EXISTS if key already exists, index otherwise, asserts if table is full +size_t ggml_hash_insert ( struct ggml_hash_set hash_set, struct ggml_tensor * key); + +// return index, asserts if table is full +size_t ggml_hash_find_or_insert( struct ggml_hash_set hash_set, struct ggml_tensor * key); #ifdef __cplusplus } diff --git a/native/jni/src/ggml/ggml-quants.c b/native/jni/src/ggml/ggml-quants.c index 1ae02b75c..36079796b 100644 --- a/native/jni/src/ggml/ggml-quants.c +++ b/native/jni/src/ggml/ggml-quants.c @@ -14,32 +14,12 @@ // #include -#if !defined(__aarch64__) -inline static int32_t vaddvq_s16(int16x8_t v) { - return - (int32_t)vgetq_lane_s16(v, 0) + (int32_t)vgetq_lane_s16(v, 1) + - (int32_t)vgetq_lane_s16(v, 2) + (int32_t)vgetq_lane_s16(v, 3) + - (int32_t)vgetq_lane_s16(v, 4) + (int32_t)vgetq_lane_s16(v, 5) + - (int32_t)vgetq_lane_s16(v, 6) + (int32_t)vgetq_lane_s16(v, 7); -} - -inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) { - int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a)); - int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b)); - return vcombine_s16(a0, b0); -} - -inline static int32_t vaddvq_s32(int32x4_t v) { - return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3); -} -#endif - #else #ifdef __wasm_simd128__ #include #else -#ifdef __POWER9_VECTOR__ +#if defined(__POWER9_VECTOR__) || defined(__powerpc64__) #include #undef bool #define bool _Bool @@ -47,13 +27,15 @@ inline static int32_t vaddvq_s32(int32x4_t v) { #if defined(_MSC_VER) || defined(__MINGW32__) #include #else -#if !defined(__riscv) && !defined(__s390__) +#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__) +#if !defined(__riscv) #include #endif #endif #endif #endif #endif +#endif #ifdef __riscv_v_intrinsic #include @@ -61,6 +43,7 @@ inline static int32_t vaddvq_s32(int32x4_t v) { #undef MIN #undef MAX + #define MIN(a, b) ((a) < (b) ? (a) : (b)) #define MAX(a, b) ((a) > (b) ? (a) : (b)) @@ -283,14 +266,34 @@ static inline float hsum_float_4x4(const __m128 a, const __m128 b, const __m128 #endif // defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) #if defined(__ARM_NEON) - #if !defined(__aarch64__) -/* +// 64-bit compatibility + +// vaddvq_s16 +// vpaddq_s16 +// vaddvq_s32 +// vaddvq_f32 +// vmaxvq_f32 +// vcvtnq_s32_f32 + +inline static int32_t vaddvq_s16(int16x8_t v) { + return + (int32_t)vgetq_lane_s16(v, 0) + (int32_t)vgetq_lane_s16(v, 1) + + (int32_t)vgetq_lane_s16(v, 2) + (int32_t)vgetq_lane_s16(v, 3) + + (int32_t)vgetq_lane_s16(v, 4) + (int32_t)vgetq_lane_s16(v, 5) + + (int32_t)vgetq_lane_s16(v, 6) + (int32_t)vgetq_lane_s16(v, 7); +} + +inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) { + int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a)); + int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b)); + return vcombine_s16(a0, b0); +} + inline static int32_t vaddvq_s32(int32x4_t v) { return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3); } -*/ inline static float vaddvq_f32(float32x4_t v) { return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3); @@ -313,6 +316,96 @@ inline static int32x4_t vcvtnq_s32_f32(float32x4_t v) { return res; } +// vld1q_s16_x2 +// vld1q_u8_x2 +// vld1q_u8_x4 +// vld1q_s8_x2 +// vld1q_s8_x4 +// TODO: double-check these work correctly + +typedef struct ggml_int16x8x2_t { + int16x8_t val[2]; +} ggml_int16x8x2_t; + +inline static ggml_int16x8x2_t ggml_vld1q_s16_x2(const int16_t * ptr) { + ggml_int16x8x2_t res; + + res.val[0] = vld1q_s16(ptr + 0); + res.val[1] = vld1q_s16(ptr + 8); + + return res; +} + +typedef struct ggml_uint8x16x2_t { + uint8x16_t val[2]; +} ggml_uint8x16x2_t; + +inline static ggml_uint8x16x2_t ggml_vld1q_u8_x2(const uint8_t * ptr) { + ggml_uint8x16x2_t res; + + res.val[0] = vld1q_u8(ptr + 0); + res.val[1] = vld1q_u8(ptr + 16); + + return res; +} + +typedef struct ggml_uint8x16x4_t { + uint8x16_t val[4]; +} ggml_uint8x16x4_t; + +inline static ggml_uint8x16x4_t ggml_vld1q_u8_x4(const uint8_t * ptr) { + ggml_uint8x16x4_t res; + + res.val[0] = vld1q_u8(ptr + 0); + res.val[1] = vld1q_u8(ptr + 16); + res.val[2] = vld1q_u8(ptr + 32); + res.val[3] = vld1q_u8(ptr + 48); + + return res; +} + +typedef struct ggml_int8x16x2_t { + int8x16_t val[2]; +} ggml_int8x16x2_t; + +inline static ggml_int8x16x2_t ggml_vld1q_s8_x2(const int8_t * ptr) { + ggml_int8x16x2_t res; + + res.val[0] = vld1q_s8(ptr + 0); + res.val[1] = vld1q_s8(ptr + 16); + + return res; +} + +typedef struct ggml_int8x16x4_t { + int8x16_t val[4]; +} ggml_int8x16x4_t; + +inline static ggml_int8x16x4_t ggml_vld1q_s8_x4(const int8_t * ptr) { + ggml_int8x16x4_t res; + + res.val[0] = vld1q_s8(ptr + 0); + res.val[1] = vld1q_s8(ptr + 16); + res.val[2] = vld1q_s8(ptr + 32); + res.val[3] = vld1q_s8(ptr + 48); + + return res; +} + +#else + +#define ggml_int16x8x2_t int16x8x2_t +#define ggml_uint8x16x2_t uint8x16x2_t +#define ggml_uint8x16x4_t uint8x16x4_t +#define ggml_int8x16x2_t int8x16x2_t +#define ggml_int8x16x4_t int8x16x4_t + +#define ggml_vld1q_s16_x2 vld1q_s16_x2 +#define ggml_vld1q_u8_x2 vld1q_u8_x2 +#define ggml_vld1q_u8_x4 vld1q_u8_x4 +#define ggml_vld1q_s8_x2 vld1q_s8_x2 +#define ggml_vld1q_s8_x4 vld1q_s8_x4 + #endif #endif @@ -1226,7 +1319,7 @@ static float make_q3_quants(int n, int nmax, const float * restrict x, int8_t * } static float make_qkx1_quants(int n, int nmax, const float * restrict x, uint8_t * restrict L, float * restrict the_min, - int ntry, float alpha) { + int ntry, float alpha) { float min = x[0]; float max = x[0]; for (int i = 1; i < n; ++i) { @@ -1269,13 +1362,18 @@ static float make_qkx1_quants(int n, int nmax, const float * restrict x, uint8_t } static float make_qkx2_quants(int n, int nmax, const float * restrict x, const float * restrict weights, - uint8_t * restrict L, float * restrict the_min, uint8_t * restrict Laux, - float rmin, float rdelta, int nstep, bool use_mad) { + uint8_t * restrict L, float * restrict the_min, uint8_t * restrict Laux, + float rmin, float rdelta, int nstep, bool use_mad) { float min = x[0]; float max = x[0]; float sum_w = weights[0]; float sum_x = sum_w * x[0]; +#ifdef HAVE_BUGGY_APPLE_LINKER + // use 'volatile' to prevent unroll and work around a bug in Apple ld64 1015.7 + for (volatile int i = 1; i < n; ++i) { +#else for (int i = 1; i < n; ++i) { +#endif if (x[i] < min) min = x[i]; if (x[i] > max) max = x[i]; float w = weights[i]; @@ -3559,7 +3657,7 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri const int32x4_t vzero = vdupq_n_s32(0); #endif - int8x16x2_t q2bytes; + ggml_int8x16x2_t q2bytes; uint8_t aux[16]; float sum = 0; @@ -3578,8 +3676,8 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri vst1q_u8(aux, scales); const uint8x16_t mins = vshrq_n_u8(mins_and_scales, 4); - const int16x8x2_t q8sums = vld1q_s16_x2(y[i].bsums); - const int16x8x2_t mins16 = {vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(mins))), vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(mins)))}; + const ggml_int16x8x2_t q8sums = ggml_vld1q_s16_x2(y[i].bsums); + const ggml_int16x8x2_t mins16 = {vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(mins))), vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(mins)))}; const int32x4_t s0 = vaddq_s32(vmull_s16(vget_low_s16 (mins16.val[0]), vget_low_s16 (q8sums.val[0])), vmull_s16(vget_high_s16(mins16.val[0]), vget_high_s16(q8sums.val[0]))); const int32x4_t s1 = vaddq_s32(vmull_s16(vget_low_s16 (mins16.val[1]), vget_low_s16 (q8sums.val[1])), @@ -3607,7 +3705,7 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri #endif #define SHIFT_MULTIPLY_ACCUM_WITH_SCALE(shift, index)\ - q8bytes = vld1q_s8_x2(q8); q8 += 32;\ + q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32;\ q2bytes.val[0] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits.val[0], (shift)), m3));\ q2bytes.val[1] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits.val[1], (shift)), m3));\ MULTIPLY_ACCUM_WITH_SCALE((index)); @@ -3615,9 +3713,9 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri for (int j = 0; j < QK_K/128; ++j) { - const uint8x16x2_t q2bits = vld1q_u8_x2(q2); q2 += 32; + const ggml_uint8x16x2_t q2bits = ggml_vld1q_u8_x2(q2); q2 += 32; - int8x16x2_t q8bytes = vld1q_s8_x2(q8); q8 += 32; + ggml_int8x16x2_t q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32; q2bytes.val[0] = vreinterpretq_s8_u8(vandq_u8(q2bits.val[0], m3)); q2bytes.val[1] = vreinterpretq_s8_u8(vandq_u8(q2bits.val[1], m3)); MULTIPLY_ACCUM_WITH_SCALE(0); @@ -3951,7 +4049,7 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri const int32x4_t vzero = vdupq_n_s32(0); #endif - int8x16x4_t q2bytes; + ggml_int8x16x4_t q2bytes; uint32_t aux32[2]; const uint8_t * scales = (const uint8_t *)aux32; @@ -3976,7 +4074,7 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri const uint8x16_t q2bits = vld1q_u8(q2); - const int8x16x4_t q8bytes = vld1q_s8_x4(q8); + const ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(q8); q2bytes.val[0] = vreinterpretq_s8_u8(vandq_u8(q2bits, m3)); q2bytes.val[1] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits, 2), m3)); @@ -4240,7 +4338,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri const uint8x16_t m3 = vshlq_n_u8(m0, 3); const int8_t m32 = 32; - int8x16x4_t q3bytes; + ggml_int8x16x4_t q3bytes; float sum = 0; @@ -4252,9 +4350,9 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri const uint8_t * restrict qh = x[i].hmask; const int8_t * restrict q8 = y[i].qs; - uint8x16x2_t qhbits = vld1q_u8_x2(qh); + ggml_uint8x16x2_t qhbits = ggml_vld1q_u8_x2(qh); - uint8x16x4_t q3h; + ggml_uint8x16x4_t q3h; int32_t isum = 0; @@ -4270,9 +4368,9 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri for (int j = 0; j < QK_K/128; ++j) { - const uint8x16x2_t q3bits = vld1q_u8_x2(q3); q3 += 32; - const int8x16x4_t q8bytes_1 = vld1q_s8_x4(q8); q8 += 64; - const int8x16x4_t q8bytes_2 = vld1q_s8_x4(q8); q8 += 64; + const ggml_uint8x16x2_t q3bits = ggml_vld1q_u8_x2(q3); q3 += 32; + const ggml_int8x16x4_t q8bytes_1 = ggml_vld1q_s8_x4(q8); q8 += 64; + const ggml_int8x16x4_t q8bytes_2 = ggml_vld1q_s8_x4(q8); q8 += 64; q3h.val[0] = vshlq_n_u8(vbicq_u8(m0, qhbits.val[0]), 2); q3h.val[1] = vshlq_n_u8(vbicq_u8(m0, qhbits.val[1]), 2); @@ -4774,7 +4872,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri const uint8x16_t m3b = vdupq_n_u8(0x3); const uint8x16_t mh = vdupq_n_u8(4); - int8x16x4_t q3bytes; + ggml_int8x16x4_t q3bytes; uint16_t aux16[2]; int8_t * scales = (int8_t *)aux16; @@ -4783,11 +4881,11 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri for (int i = 0; i < nb; ++i) { - uint8x16x4_t q3h; + ggml_uint8x16x4_t q3h; const uint8x8_t hbits = vld1_u8(x[i].hmask); const uint8x16_t q3bits = vld1q_u8(x[i].qs); - const int8x16x4_t q8bytes = vld1q_s8_x4(y[i].qs); + const ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(y[i].qs); const uint16_t a = *(const uint16_t *)x[i].scales; aux16[0] = a & 0x0f0f; @@ -5136,8 +5234,8 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri const int32x4_t mzero = vdupq_n_s32(0); #endif - int8x16x2_t q4bytes; - int8x16x2_t q8bytes; + ggml_int8x16x2_t q4bytes; + ggml_int8x16x2_t q8bytes; float sumf = 0; @@ -5172,17 +5270,17 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri for (int j = 0; j < QK_K/64; ++j) { - const uint8x16x2_t q4bits = vld1q_u8_x2(q4); q4 += 32; + const ggml_uint8x16x2_t q4bits = ggml_vld1q_u8_x2(q4); q4 += 32; #ifdef __ARM_FEATURE_DOTPROD - q8bytes = vld1q_s8_x2(q8); q8 += 32; + q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32; q4bytes.val[0] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[0], m4b)); q4bytes.val[1] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[1], m4b)); const int32x4_t p1 = vdotq_s32(vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[0]), q4bytes.val[1], q8bytes.val[1]); sumi1 += vaddvq_s32(p1) * scales[2*j+0]; - q8bytes = vld1q_s8_x2(q8); q8 += 32; + q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32; q4bytes.val[0] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[0], 4)); q4bytes.val[1] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[1], 4)); @@ -5190,7 +5288,7 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri sumi2 += vaddvq_s32(p2) * scales[2*j+1]; #else - q8bytes = vld1q_s8_x2(q8); q8 += 32; + q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32; q4bytes.val[0] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[0], m4b)); q4bytes.val[1] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[1], m4b)); const int16x8_t p0 = vaddq_s16(vmull_s8(vget_low_s8 (q4bytes.val[0]), vget_low_s8 (q8bytes.val[0])), @@ -5199,7 +5297,7 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri vmull_s8(vget_high_s8(q4bytes.val[1]), vget_high_s8(q8bytes.val[1]))); sumi1 += vaddvq_s16(vaddq_s16(p0, p1)) * scales[2*j+0]; - q8bytes = vld1q_s8_x2(q8); q8 += 32; + q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32; q4bytes.val[0] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[0], 4)); q4bytes.val[1] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[1], 4)); const int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q4bytes.val[0]), vget_low_s8 (q8bytes.val[0])), @@ -5514,8 +5612,8 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri float sumf = 0; - int8x16x2_t q4bytes; - int8x16x4_t q8bytes; + ggml_int8x16x2_t q4bytes; + ggml_int8x16x4_t q8bytes; float sum_mins = 0.f; @@ -5536,10 +5634,10 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri const float d = y[i].d * (float)x[i].d[0]; - const uint8x16x2_t q4bits = vld1q_u8_x2(q4); + const ggml_uint8x16x2_t q4bits = ggml_vld1q_u8_x2(q4); #ifdef __ARM_FEATURE_DOTPROD - q8bytes = vld1q_s8_x4(q8); + q8bytes = ggml_vld1q_s8_x4(q8); q4bytes.val[0] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[0], m4b)); q4bytes.val[1] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[1], m4b)); @@ -5553,7 +5651,7 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri const int32_t sumi2 = vaddvq_s32(p2) * scales[1]; #else - q8bytes = vld1q_s8_x4(q8); + q8bytes = ggml_vld1q_s8_x4(q8); q4bytes.val[0] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[0], m4b)); q4bytes.val[1] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[1], m4b)); const int16x8_t p0 = vaddq_s16(vmull_s8(vget_low_s8 (q4bytes.val[0]), vget_low_s8 (q8bytes.val[0])), @@ -5787,7 +5885,7 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri const int32x4_t mzero = vdupq_n_s32(0); #endif - int8x16x4_t q5bytes; + ggml_int8x16x4_t q5bytes; float sumf = 0; @@ -5817,16 +5915,16 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri const uint8_t * restrict qh = x[i].qh; const int8_t * restrict q8 = y[i].qs; - uint8x16x2_t qhbits = vld1q_u8_x2(qh); + ggml_uint8x16x2_t qhbits = ggml_vld1q_u8_x2(qh); - uint8x16x4_t q5h; + ggml_uint8x16x4_t q5h; int32_t sumi = 0; for (int j = 0; j < QK_K/64; ++j) { - const uint8x16x2_t q5bits = vld1q_u8_x2(q5); q5 += 32; - const int8x16x4_t q8bytes = vld1q_s8_x4(q8); q8 += 64; + const ggml_uint8x16x2_t q5bits = ggml_vld1q_u8_x2(q5); q5 += 32; + const ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(q8); q8 += 64; q5h.val[0] = vshlq_n_u8(vandq_u8(mone, qhbits.val[0]), 4); q5h.val[1] = vshlq_n_u8(vandq_u8(mone, qhbits.val[1]), 4); @@ -6220,8 +6318,8 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri const int32x4_t mzero = vdupq_n_s32(0); #endif - int8x16x4_t q5bytes; - uint8x16x4_t q5h; + ggml_int8x16x4_t q5bytes; + ggml_uint8x16x4_t q5h; float sumf = 0; @@ -6236,8 +6334,8 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri const uint8x8_t qhbits = vld1_u8(qh); - const uint8x16x2_t q5bits = vld1q_u8_x2(q5); - const int8x16x4_t q8bytes = vld1q_s8_x4(q8); + const ggml_uint8x16x2_t q5bits = ggml_vld1q_u8_x2(q5); + const ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(q8); const uint8x16_t htmp = vcombine_u8(qhbits, vshr_n_u8(qhbits, 1)); q5h.val[0] = vbicq_u8(mh, vshlq_n_u8(htmp, 4)); @@ -6513,8 +6611,8 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri const uint8x16_t mone = vdupq_n_u8(3); - int8x16x4_t q6bytes; - uint8x16x4_t q6h; + ggml_int8x16x4_t q6bytes; + ggml_uint8x16x4_t q6h; for (int i = 0; i < nb; ++i) { @@ -6526,9 +6624,9 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri const int8_t * restrict scale = x[i].scales; - const int16x8x2_t q8sums = vld1q_s16_x2(y[i].bsums); + const ggml_int16x8x2_t q8sums = ggml_vld1q_s16_x2(y[i].bsums); const int8x16_t scales = vld1q_s8(scale); - const int16x8x2_t q6scales = {vmovl_s8(vget_low_s8(scales)), vmovl_s8(vget_high_s8(scales))}; + const ggml_int16x8x2_t q6scales = {vmovl_s8(vget_low_s8(scales)), vmovl_s8(vget_high_s8(scales))}; const int32x4_t prod = vaddq_s32(vaddq_s32(vmull_s16(vget_low_s16 (q8sums.val[0]), vget_low_s16 (q6scales.val[0])), vmull_s16(vget_high_s16(q8sums.val[0]), vget_high_s16(q6scales.val[0]))), @@ -6540,9 +6638,9 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri for (int j = 0; j < QK_K/128; ++j) { - uint8x16x2_t qhbits = vld1q_u8_x2(qh); qh += 32; - uint8x16x4_t q6bits = vld1q_u8_x4(q6); q6 += 64; - int8x16x4_t q8bytes = vld1q_s8_x4(q8); q8 += 64; + ggml_uint8x16x2_t qhbits = ggml_vld1q_u8_x2(qh); qh += 32; + ggml_uint8x16x4_t q6bits = ggml_vld1q_u8_x4(q6); q6 += 64; + ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(q8); q8 += 64; q6h.val[0] = vshlq_n_u8(vandq_u8(mone, qhbits.val[0]), 4); q6h.val[1] = vshlq_n_u8(vandq_u8(mone, qhbits.val[1]), 4); @@ -6585,7 +6683,7 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri scale += 2; #endif - q8bytes = vld1q_s8_x4(q8); q8 += 64; + q8bytes = ggml_vld1q_s8_x4(q8); q8 += 64; shifted = vshrq_n_u8(qhbits.val[0], 4); q6h.val[0] = vshlq_n_u8(vandq_u8(mone, shifted), 4); @@ -6989,8 +7087,8 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri const uint8x16_t mone = vdupq_n_u8(3); - int8x16x4_t q6bytes; - uint8x16x4_t q6h; + ggml_int8x16x4_t q6bytes; + ggml_uint8x16x4_t q6h; for (int i = 0; i < nb; ++i) { @@ -7004,9 +7102,9 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri int32_t isum = 0; - uint8x16_t qhbits = vld1q_u8(qh); - uint8x16x2_t q6bits = vld1q_u8_x2(q6); - int8x16x4_t q8bytes = vld1q_s8_x4(q8); + uint8x16_t qhbits = vld1q_u8(qh); + ggml_uint8x16x2_t q6bits = ggml_vld1q_u8_x2(q6); + ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(q8); q6h.val[0] = vshlq_n_u8(vandq_u8(mone, qhbits), 4); uint8x16_t shifted = vshrq_n_u8(qhbits, 2); diff --git a/native/jni/src/ggml/ggml.c b/native/jni/src/ggml/ggml.c index 80d682255..f92292b39 100644 --- a/native/jni/src/ggml/ggml.c +++ b/native/jni/src/ggml/ggml.c @@ -1,4 +1,5 @@ #define _CRT_SECURE_NO_DEPRECATE // Disables ridiculous "unsafe" warnigns on Windows +#define _USE_MATH_DEFINES // For M_PI on MSVC #include "ggml-impl.h" #include "ggml-quants.h" @@ -99,6 +100,49 @@ typedef void * thread_ret_t; #include #endif +#if defined(__APPLE__) +#include +#endif + +#if (defined(__linux__) || defined(__APPLE__) || defined(__FreeBSD__) || defined(__NetBSD__) || defined(__OpenBSD__)) && \ + (!defined(TARGET_OS_TV) && !defined(TARGET_OS_WATCH)) + +#include + +void ggml_print_backtrace(void) { + /* + #include + #include + + void * trace[100]; + + int nptrs = backtrace(trace, sizeof(trace)/sizeof(trace[0])); + + backtrace_symbols_fd(trace, nptrs, STDERR_FILENO); + */ + + // backtrack_symbols does not show line numbers, use gdb instead + char attach[32]; + snprintf(attach, sizeof(attach), "attach %d", getpid()); + int pid = fork(); + if (pid == 0) { + execlp("gdb", "gdb", "--batch", + "-ex", "set style enabled on", + "-ex", attach, + "-ex", "bt -frame-info source-and-location", + "-ex", "detach", + "-ex", "quit", + NULL); + } else { + waitpid(pid, NULL, 0); + } +} +#else +void ggml_print_backtrace(void) { + // platform not supported +} +#endif + /*#define GGML_PERF*/ #define GGML_DEBUG 0 #define GGML_GELU_FP16 @@ -227,6 +271,12 @@ inline static void * ggml_aligned_malloc(size_t size) { // floating point type used to accumulate sums typedef double ggml_float; +#undef MIN +#undef MAX + +#define MIN(a, b) ((a) < (b) ? (a) : (b)) +#define MAX(a, b) ((a) > (b) ? (a) : (b)) + // // global data // @@ -560,6 +610,18 @@ ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type) { // simd mappings // +#if defined(__ARM_NEON) +#if !defined(__aarch64__) + +// 64-bit compatibility + +inline static float vaddvq_f32(float32x4_t v) { + return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3); +} + +#endif +#endif + // we define a common set of C macros which map to specific intrinsics based on the current architecture // we then implement the fundamental computation operations below using only these macros // adding support for new architectures requires to define the corresponding SIMD macros @@ -1351,6 +1413,7 @@ inline static void ggml_vec_step_f32 (const int n, float * y, const float * x) { inline static void ggml_vec_tanh_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = tanhf(x[i]); } inline static void ggml_vec_elu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : expf(x[i])-1; } inline static void ggml_vec_relu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : 0.f; } +inline static void ggml_vec_leaky_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : 0.1f*x[i]; } static const float GELU_COEF_A = 0.044715f; static const float GELU_QUICK_COEF = -1.702f; @@ -1571,13 +1634,8 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = { "ROPE_BACK", "ALIBI", "CLAMP", - "CONV_1D", - "CONV_1D_STAGE_0", - "CONV_1D_STAGE_1", "CONV_TRANSPOSE_1D", - "CONV_2D", - "CONV_2D_STAGE_0", - "CONV_2D_STAGE_1", + "IM2COL", "CONV_TRANSPOSE_2D", "POOL_1D", "POOL_2D", @@ -1608,7 +1666,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = { "CROSS_ENTROPY_LOSS_BACK", }; -static_assert(GGML_OP_COUNT == 73, "GGML_OP_COUNT != 73"); +static_assert(GGML_OP_COUNT == 68, "GGML_OP_COUNT != 68"); static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "none", @@ -1658,13 +1716,8 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "rope_back(x)", "alibi(x)", "clamp(x)", - "conv_1d(x)", - "conv_1d_stage_0(x)", - "conv_1d_stage_1(x)", "conv_transpose_1d(x)", - "conv_2d(x)", - "conv_2d_stage_0(x)", - "conv_2d_stage_1(x)", + "im2col(x)", "conv_transpose_2d(x)", "pool_1d(x)", "pool_2d(x)", @@ -1695,7 +1748,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "cross_entropy_loss_back(x,y)", }; -static_assert(GGML_OP_COUNT == 73, "GGML_OP_COUNT != 73"); +static_assert(GGML_OP_COUNT == 68, "GGML_OP_COUNT != 68"); static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2"); @@ -1723,13 +1776,7 @@ static void ggml_setup_op_has_task_pass(void) { p[GGML_OP_GET_ROWS_BACK ] = true; p[GGML_OP_DIAG_MASK_INF ] = true; p[GGML_OP_DIAG_MASK_ZERO ] = true; - p[GGML_OP_CONV_1D ] = true; - p[GGML_OP_CONV_1D_STAGE_0 ] = true; - p[GGML_OP_CONV_1D_STAGE_1 ] = true; p[GGML_OP_CONV_TRANSPOSE_1D ] = true; - p[GGML_OP_CONV_2D ] = true; - p[GGML_OP_CONV_2D_STAGE_0 ] = true; - p[GGML_OP_CONV_2D_STAGE_1 ] = true; p[GGML_OP_CONV_TRANSPOSE_2D ] = true; p[GGML_OP_FLASH_ATTN_BACK ] = true; p[GGML_OP_CROSS_ENTROPY_LOSS ] = true; @@ -3768,6 +3815,14 @@ struct ggml_tensor * ggml_relu_inplace( return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_RELU); } +// ggml_leaky + +struct ggml_tensor * ggml_leaky( + struct ggml_context * ctx, + struct ggml_tensor * a) { + return ggml_unary(ctx, a, GGML_UNARY_OP_LEAKY); +} + // ggml_gelu struct ggml_tensor * ggml_gelu( @@ -4845,8 +4900,13 @@ static struct ggml_tensor * ggml_rope_impl( int n_dims, int mode, int n_ctx, + int n_orig_ctx, float freq_base, float freq_scale, + float ext_factor, + float attn_factor, + float beta_fast, + float beta_slow, float xpos_base, bool xpos_down, bool inplace) { @@ -4862,11 +4922,15 @@ static struct ggml_tensor * ggml_rope_impl( struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - int32_t params[8] = { /*n_past*/ 0, n_dims, mode, n_ctx }; - memcpy(params + 4, &freq_base, sizeof(float)); - memcpy(params + 5, &freq_scale, sizeof(float)); - memcpy(params + 6, &xpos_base, sizeof(float)); - memcpy(params + 7, &xpos_down, sizeof(bool)); + int32_t params[13] = { /*n_past*/ 0, n_dims, mode, n_ctx, n_orig_ctx }; + memcpy(params + 5, &freq_base, sizeof(float)); + memcpy(params + 6, &freq_scale, sizeof(float)); + memcpy(params + 7, &ext_factor, sizeof(float)); + memcpy(params + 8, &attn_factor, sizeof(float)); + memcpy(params + 9, &beta_fast, sizeof(float)); + memcpy(params + 10, &beta_slow, sizeof(float)); + memcpy(params + 11, &xpos_base, sizeof(float)); + memcpy(params + 12, &xpos_down, sizeof(bool)); ggml_set_op_params(result, params, sizeof(params)); result->op = GGML_OP_ROPE; @@ -4884,7 +4948,9 @@ struct ggml_tensor * ggml_rope( int n_dims, int mode, int n_ctx) { - return ggml_rope_impl(ctx, a, b, n_dims, mode, n_ctx, 10000.0f, 1.0f, 0.0f, false, false); + return ggml_rope_impl( + ctx, a, b, n_dims, mode, n_ctx, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, 0.0f, false, false + ); } struct ggml_tensor * ggml_rope_inplace( @@ -4894,7 +4960,9 @@ struct ggml_tensor * ggml_rope_inplace( int n_dims, int mode, int n_ctx) { - return ggml_rope_impl(ctx, a, b, n_dims, mode, n_ctx, 10000.0f, 1.0f, 0.0f, false, true); + return ggml_rope_impl( + ctx, a, b, n_dims, mode, n_ctx, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, 0.0f, false, true + ); } struct ggml_tensor * ggml_rope_custom( @@ -4904,9 +4972,17 @@ struct ggml_tensor * ggml_rope_custom( int n_dims, int mode, int n_ctx, + int n_orig_ctx, float freq_base, - float freq_scale) { - return ggml_rope_impl(ctx, a, b, n_dims, mode, n_ctx, freq_base, freq_scale, 0.0f, false, false); + float freq_scale, + float ext_factor, + float attn_factor, + float beta_fast, + float beta_slow) { + return ggml_rope_impl( + ctx, a, b, n_dims, mode, n_ctx, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow, 0.0f, false, false + ); } struct ggml_tensor * ggml_rope_custom_inplace( @@ -4916,9 +4992,17 @@ struct ggml_tensor * ggml_rope_custom_inplace( int n_dims, int mode, int n_ctx, + int n_orig_ctx, float freq_base, - float freq_scale) { - return ggml_rope_impl(ctx, a, b, n_dims, mode, n_ctx, freq_base, freq_scale, 0.0f, false, true); + float freq_scale, + float ext_factor, + float attn_factor, + float beta_fast, + float beta_slow) { + return ggml_rope_impl( + ctx, a, b, n_dims, mode, n_ctx, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow, 0.0f, false, true + ); } struct ggml_tensor * ggml_rope_xpos_inplace( @@ -4928,7 +5012,7 @@ struct ggml_tensor * ggml_rope_xpos_inplace( int n_dims, float base, bool down) { - return ggml_rope_impl(ctx, a, b, n_dims, 0, 0, 10000.0f, 1.0f, base, down, true); + return ggml_rope_impl(ctx, a, b, n_dims, 0, 0, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, base, down, true); } // ggml_rope_back @@ -4940,8 +5024,13 @@ struct ggml_tensor * ggml_rope_back( int n_dims, int mode, int n_ctx, + int n_orig_ctx, float freq_base, float freq_scale, + float ext_factor, + float attn_factor, + float beta_fast, + float beta_slow, float xpos_base, bool xpos_down) { GGML_ASSERT(ggml_is_vector(b)); @@ -4958,11 +5047,15 @@ struct ggml_tensor * ggml_rope_back( struct ggml_tensor * result = ggml_dup_tensor(ctx, a); - int32_t params[8] = { /*n_past*/ 0, n_dims, mode, n_ctx }; - memcpy(params + 4, &freq_base, sizeof(float)); - memcpy(params + 5, &freq_scale, sizeof(float)); - memcpy(params + 6, &xpos_base, sizeof(float)); - memcpy(params + 7, &xpos_down, sizeof(bool)); + int32_t params[13] = { /*n_past*/ 0, n_dims, mode, n_ctx, n_orig_ctx }; + memcpy(params + 5, &freq_base, sizeof(float)); + memcpy(params + 6, &freq_scale, sizeof(float)); + memcpy(params + 7, &ext_factor, sizeof(float)); + memcpy(params + 8, &attn_factor, sizeof(float)); + memcpy(params + 9, &beta_fast, sizeof(float)); + memcpy(params + 10, &beta_slow, sizeof(float)); + memcpy(params + 11, &xpos_base, sizeof(float)); + memcpy(params + 12, &xpos_down, sizeof(bool)); ggml_set_op_params(result, params, sizeof(params)); result->op = GGML_OP_ROPE_BACK; @@ -5037,82 +5130,6 @@ static int64_t ggml_calc_conv_output_size(int64_t ins, int64_t ks, int s, int p, return (ins + 2 * p - d * (ks - 1) - 1) / s + 1; } -// im2col: [N, IC, IL] => [N, OL, IC*K] -// a: [OC，IC, K] -// b: [N, IC, IL] -// result: [N, OL, IC*K] -static struct ggml_tensor * ggml_conv_1d_stage_0( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - int s0, - int p0, - int d0) { - GGML_ASSERT(a->ne[1] == b->ne[1]); - bool is_node = false; - - if (a->grad || b->grad) { - GGML_ASSERT(false); // TODO: implement backward - is_node = true; - } - - const int64_t OL = ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0); - - const int64_t ne[4] = { - a->ne[1] * a->ne[0], - OL, - b->ne[2], - 1, - }; - struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F16, 4, ne); - - int32_t params[] = { s0, p0, d0 }; - ggml_set_op_params(result, params, sizeof(params)); - - result->op = GGML_OP_CONV_1D_STAGE_0; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; - result->src[0] = a; - result->src[1] = b; - - return result; -} - -// ggml_conv_1d_stage_1 - -// gemm: [N, OC, OL] = [OC, IC * K] x [N*OL, IC * K] -// a: [OC, IC, K] -// b: [N, OL, IC * K] -// result: [N, OC, OL] -static struct ggml_tensor * ggml_conv_1d_stage_1( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b) { - - bool is_node = false; - - if (a->grad || b->grad) { - GGML_ASSERT(false); // TODO: implement backward - is_node = true; - } - - const int64_t ne[4] = { - b->ne[1], - a->ne[2], - b->ne[2], - 1, - }; - struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne); - - result->op = GGML_OP_CONV_1D_STAGE_1; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; - result->src[0] = a; - result->src[1] = b; - - return result; -} - -// ggml_conv_1d - GGML_API struct ggml_tensor * ggml_conv_1d( struct ggml_context * ctx, struct ggml_tensor * a, @@ -5120,44 +5137,18 @@ GGML_API struct ggml_tensor * ggml_conv_1d( int s0, int p0, int d0) { - struct ggml_tensor * result = ggml_conv_1d_stage_0(ctx, a, b, s0, p0, d0); - result = ggml_conv_1d_stage_1(ctx, a, result); + struct ggml_tensor * im2col = ggml_im2col(ctx, a, b, s0, 0, p0, 0, d0, 0, false); // [N, OL, IC * K] + + struct ggml_tensor * result = + ggml_mul_mat(ctx, + ggml_reshape_2d(ctx, im2col, im2col->ne[0], (im2col->ne[2] * im2col->ne[1])), // [N, OL, IC * K] => [N*OL, IC * K] + ggml_reshape_2d(ctx, a, (a->ne[0] * a->ne[1]), a->ne[2])); // [OC，IC, K] => [OC, IC * K] + + result = ggml_reshape_3d(ctx, result, im2col->ne[1], a->ne[2], im2col->ne[2]); // [N, OC, OL] + return result; } -// GGML_API struct ggml_tensor * ggml_conv_1d( -// struct ggml_context * ctx, -// struct ggml_tensor * a, -// struct ggml_tensor * b, -// int s0, -// int p0, -// int d0) { -// GGML_ASSERT(ggml_is_matrix(b)); -// GGML_ASSERT(a->ne[1] == b->ne[1]); -// bool is_node = false; - -// if (a->grad || b->grad) { -// GGML_ASSERT(false); // TODO: implement backward -// is_node = true; -// } - -// const int64_t ne[4] = { -// ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0), -// a->ne[2], 1, 1, -// }; -// struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 2, ne); - -// int32_t params[] = { s0, p0, d0 }; -// ggml_set_op_params(result, params, sizeof(params)); - -// result->op = GGML_OP_CONV_1D; -// result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; -// result->src[0] = a; -// result->src[1] = b; - -// return result; -// } - // ggml_conv_1d_ph struct ggml_tensor* ggml_conv_1d_ph( @@ -5219,7 +5210,7 @@ GGML_API struct ggml_tensor * ggml_conv_transpose_1d( // a: [OC，IC, KH, KW] // b: [N, IC, IH, IW] // result: [N, OH, OW, IC*KH*KW] -static struct ggml_tensor * ggml_conv_2d_stage_0( +struct ggml_tensor * ggml_im2col( struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, @@ -5228,9 +5219,14 @@ static struct ggml_tensor * ggml_conv_2d_stage_0( int p0, int p1, int d0, - int d1) { + int d1, + bool is_2D) { - GGML_ASSERT(a->ne[2] == b->ne[2]); + if(is_2D) { + GGML_ASSERT(a->ne[2] == b->ne[2]); + } else { + GGML_ASSERT(a->ne[1] == b->ne[1]); + } bool is_node = false; if (a->grad || b->grad) { @@ -5238,81 +5234,51 @@ static struct ggml_tensor * ggml_conv_2d_stage_0( is_node = true; } - const int64_t OH = ggml_calc_conv_output_size(b->ne[1], a->ne[1], s1, p1, d1); - const int64_t OW = ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0); + const int64_t OH = is_2D ? ggml_calc_conv_output_size(b->ne[1], a->ne[1], s1, p1, d1) : 0; + const int64_t OW = ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0); const int64_t ne[4] = { - a->ne[2] * a->ne[1] * a->ne[0], + is_2D ? (a->ne[2] * a->ne[1] * a->ne[0]) : a->ne[1] * a->ne[0], OW, - OH, - b->ne[3], + is_2D ? OH : b->ne[2], + is_2D ? b->ne[3] : 1, }; - struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F16, 4, ne); - int32_t params[] = { s0, s1, p0, p1, d0, d1 }; + struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F16, 4, ne); + int32_t params[] = { s0, s1, p0, p1, d0, d1, (is_2D ? 1 : 0) }; ggml_set_op_params(result, params, sizeof(params)); - result->op = GGML_OP_CONV_2D_STAGE_0; + result->op = GGML_OP_IM2COL; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; result->src[1] = b; return result; - -} - -// gemm: [N, OC, OH, OW] = [OC, IC * KH * KW] x [N*OH*OW, IC * KH * KW] -// a: [OC, IC, KH, KW] -// b: [N, OH, OW, IC * KH * KW] -// result: [N, OC, OH, OW] -static struct ggml_tensor * ggml_conv_2d_stage_1( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b) { - - bool is_node = false; - - if (a->grad || b->grad) { - GGML_ASSERT(false); // TODO: implement backward - is_node = true; - } - - const int64_t ne[4] = { - b->ne[1], - b->ne[2], - a->ne[3], - b->ne[3], - }; - struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne); - - result->op = GGML_OP_CONV_2D_STAGE_1; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; - result->src[0] = a; - result->src[1] = b; - - return result; - } // a: [OC，IC, KH, KW] // b: [N, IC, IH, IW] // result: [N, OC, OH, OW] struct ggml_tensor * ggml_conv_2d( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - int s0, - int s1, - int p0, - int p1, - int d0, - int d1) { + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + int s0, + int s1, + int p0, + int p1, + int d0, + int d1) { + struct ggml_tensor * im2col = ggml_im2col(ctx, a, b, s0, s1, p0, p1, d0, d1, true); // [N, OH, OW, IC * KH * KW] - struct ggml_tensor * result = ggml_conv_2d_stage_0(ctx, a, b, s0, s1, p0, p1, d0, d1); // [N, OH, OW, IC * KH * KW] - result = ggml_conv_2d_stage_1(ctx, a, result); + struct ggml_tensor * result = + ggml_mul_mat(ctx, + ggml_reshape_2d(ctx, im2col, im2col->ne[0], im2col->ne[3] * im2col->ne[2] * im2col->ne[1]), // [N, OH, OW, IC * KH * KW] => [N*OH*OW, IC * KH * KW] + ggml_reshape_2d(ctx, a, (a->ne[0] * a->ne[1] * a->ne[2]), a->ne[3])); // [OC，IC, KH, KW] => [OC, IC * KH * KW] + + result = ggml_reshape_4d(ctx, result, im2col->ne[1], im2col->ne[2], a->ne[3], im2col->ne[3]); // [N, OC, OH, OW] return result; - } // ggml_conv_2d_sk_p0 @@ -5372,7 +5338,7 @@ struct ggml_tensor * ggml_conv_transpose_2d_p0( // ggml_pool_* -static int64_t ggml_calc_pool_output_size(int64_t ins, int ks, int s, int p) { +static int64_t ggml_calc_pool_output_size(int64_t ins, int ks, int s, float p) { return (ins + 2 * p - ks) / s + 1; } @@ -5419,8 +5385,8 @@ struct ggml_tensor * ggml_pool_2d( int k1, int s0, int s1, - int p0, - int p1) { + float p0, + float p1) { bool is_node = false; @@ -8882,6 +8848,48 @@ static void ggml_compute_forward_silu( } } +// ggml_compute_forward_leaky + +static void ggml_compute_forward_leaky_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + struct ggml_tensor * dst) { + assert(params->ith == 0); + assert(ggml_are_same_shape(src0, dst)); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + assert(dst->nb[0] == sizeof(float)); + assert(src0->nb[0] == sizeof(float)); + + for (int i = 0; i < n; i++) { + ggml_vec_leaky_f32(nc, + (float *) ((char *) dst->data + i*( dst->nb[1])), + (float *) ((char *) src0->data + i*(src0->nb[1]))); + } +} + +static void ggml_compute_forward_leaky( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + struct ggml_tensor * dst) { + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_leaky_f32(params, src0, dst); + } break; + default: + { + GGML_ASSERT(false); + } break; + } +} + // ggml_compute_forward_silu_back static void ggml_compute_forward_silu_back_f32( @@ -9365,6 +9373,8 @@ static bool ggml_compute_forward_mul_mat_use_blas( // TODO: find the optimal values for these if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && + src0->type == GGML_TYPE_F32 && + src1->type == GGML_TYPE_F32 && (ne0 >= 32 && ne1 >= 32 && ne10 >= 32)) { /*printf("BLAS: %d %d %d %d %d\n", ne0, ne1, ne10, ne00, ne01);*/ @@ -9403,7 +9413,7 @@ static void ggml_compute_forward_mul_mat( // we don't support permuted src0 or src1 GGML_ASSERT(nb00 == ggml_type_size(type)); - GGML_ASSERT(nb10 == sizeof(float)); + GGML_ASSERT(nb10 == ggml_type_size(src1->type)); // dst cannot be transposed or permuted GGML_ASSERT(nb0 == sizeof(float)); @@ -9601,10 +9611,12 @@ static void ggml_compute_forward_out_prod_f32( const int ith = params->ith; const int nth = params->nth; + GGML_ASSERT(ne0 == ne00); + GGML_ASSERT(ne1 == ne10); + GGML_ASSERT(ne2 == ne02); GGML_ASSERT(ne02 == ne12); - GGML_ASSERT(ne03 == ne13); - GGML_ASSERT(ne2 == ne12); GGML_ASSERT(ne3 == ne13); + GGML_ASSERT(ne03 == ne13); // we don't support permuted src0 or src1 GGML_ASSERT(nb00 == sizeof(float)); @@ -9615,18 +9627,25 @@ static void ggml_compute_forward_out_prod_f32( // GGML_ASSERT(nb1 <= nb2); // GGML_ASSERT(nb2 <= nb3); - GGML_ASSERT(ne0 == ne00); - GGML_ASSERT(ne1 == ne10); - GGML_ASSERT(ne2 == ne02); - GGML_ASSERT(ne3 == ne03); - // nb01 >= nb00 - src0 is not transposed // compute by src0 rows // TODO: #if defined(GGML_USE_CUBLAS) ggml_cuda_out_prod - // TODO: #if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST) + // TODO: #if defined(GGML_USE_CLBLAST) + +#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) + bool use_blas = ggml_is_matrix(src0) && + ggml_is_matrix(src1) && + ggml_is_contiguous(src0) && + (ggml_is_contiguous(src1) || ggml_is_transposed(src1)); +#endif if (params->type == GGML_TASK_INIT) { +#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) // gemm beta will zero dst + if (use_blas) { + return; + } +#endif ggml_vec_set_f32(ne0*ne1*ne2*ne3, dst->data, 0); return; } @@ -9635,6 +9654,50 @@ static void ggml_compute_forward_out_prod_f32( return; } +#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) + if (use_blas) { + if (params->ith != 0) { // All threads other than the first do no work. + return; + } + // Arguments to ggml_compute_forward_out_prod (expressed as major,minor) + // src0: (k,n) + // src1: (k,m) + // dst: (m,n) + // + // Arguments to sgemm (see https://github.com/Reference-LAPACK/lapack/blob/master/BLAS/SRC/sgemm.f) + // Also expressed as (major,minor) + // a: (m,k): so src1 transposed + // b: (k,n): so src0 + // c: (m,n) + // + // However, if ggml_is_transposed(src1) is true, then + // src1->data already contains a transposed version, so sgemm mustn't + // transpose it further. + + int n = src0->ne[0]; + int k = src0->ne[1]; + int m = src1->ne[0]; + + int transposeA, lda; + + if (!ggml_is_transposed(src1)) { + transposeA = CblasTrans; + lda = m; + } else { + transposeA = CblasNoTrans; + lda = k; + } + + float * a = (float *) ((char *) src1->data); + float * b = (float *) ((char *) src0->data); + float * c = (float *) ((char *) dst->data); + + cblas_sgemm(CblasRowMajor, transposeA, CblasNoTrans, m, n, k, 1.0, a, lda, b, n, 0.0, c, n); + + return; + } +#endif + // dst[:,:,:,:] = 0 // for i2,i3: // for i1: @@ -10901,30 +10964,75 @@ static void ggml_compute_forward_clamp( // ggml_compute_forward_rope +static float rope_yarn_ramp(const float low, const float high, const int i0) { + const float y = (i0 / 2 - low) / MAX(0.001f, high - low); + return 1 - MIN(1, MAX(0, y)); +} + +// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn +// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng. +static void rope_yarn( + float theta_extrap, float freq_scale, float corr_dims[2], int64_t i0, float ext_factor, float mscale, + float * cos_theta, float * sin_theta +) { + // Get n-d rotational scaling corrected for extrapolation + float theta_interp = freq_scale * theta_extrap; + float theta = theta_interp; + if (ext_factor != 0.0f) { + float ramp_mix = rope_yarn_ramp(corr_dims[0], corr_dims[1], i0) * ext_factor; + theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix; + + // Get n-d magnitude scaling corrected for interpolation + mscale *= 1.0f + 0.1f * logf(1.0f / freq_scale); + } + *cos_theta = cosf(theta) * mscale; + *sin_theta = sinf(theta) * mscale; +} + +// Apparently solving `n_rot = 2pi * x * base^((2 * max_pos_emb) / n_dims)` for x, we get +// `corr_dim(n_rot) = n_dims * log(max_pos_emb / (n_rot * 2pi)) / (2 * log(base))` +static float ggml_rope_yarn_corr_dim(int n_dims, int n_orig_ctx, float n_rot, float base) { + return n_dims * logf(n_orig_ctx / (n_rot * 2 * (float)M_PI)) / (2 * logf(base)); +} + +void ggml_rope_yarn_corr_dims( + int n_dims, int n_orig_ctx, float freq_base, float beta_fast, float beta_slow, float dims[2] +) { + // start and end correction dims + dims[0] = MAX(0, floorf(ggml_rope_yarn_corr_dim(n_dims, n_orig_ctx, beta_fast, freq_base))); + dims[1] = MIN(n_dims - 1, ceilf(ggml_rope_yarn_corr_dim(n_dims, n_orig_ctx, beta_slow, freq_base))); +} + static void ggml_compute_forward_rope_f32( const struct ggml_compute_params * params, const struct ggml_tensor * src0, const struct ggml_tensor * src1, - struct ggml_tensor * dst) { + struct ggml_tensor * dst, + const bool forward) { if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { return; } - float freq_base; - float freq_scale; + float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow; // these two only relevant for xPos RoPE: float xpos_base; bool xpos_down; - //const int n_past = ((int32_t *) dst->op_params)[0]; - const int n_dims = ((int32_t *) dst->op_params)[1]; - const int mode = ((int32_t *) dst->op_params)[2]; - const int n_ctx = ((int32_t *) dst->op_params)[3]; - memcpy(&freq_base, (int32_t *) dst->op_params + 4, sizeof(float)); - memcpy(&freq_scale, (int32_t *) dst->op_params + 5, sizeof(float)); - memcpy(&xpos_base, (int32_t *) dst->op_params + 6, sizeof(float)); - memcpy(&xpos_down, (int32_t *) dst->op_params + 7, sizeof(bool)); + //const int n_past = ((int32_t *) dst->op_params)[0]; + const int n_dims = ((int32_t *) dst->op_params)[1]; + const int mode = ((int32_t *) dst->op_params)[2]; + const int n_ctx = ((int32_t *) dst->op_params)[3]; + const int n_orig_ctx = ((int32_t *) dst->op_params)[4]; + + memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float)); + memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float)); + memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float)); + memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float)); + memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float)); + memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float)); + memcpy(&xpos_base, (int32_t *) dst->op_params + 11, sizeof(float)); + memcpy(&xpos_down, (int32_t *) dst->op_params + 12, sizeof(bool)); GGML_TENSOR_UNARY_OP_LOCALS @@ -10952,10 +11060,18 @@ static void ggml_compute_forward_rope_f32( int ir = 0; const float theta_scale = powf(freq_base, -2.0f/n_dims); + const float inv_ndims = -1.f/n_dims; + float corr_dims[2]; + ggml_rope_yarn_corr_dims(n_dims, n_orig_ctx, freq_base, beta_fast, beta_slow, corr_dims); const bool is_neox = mode & 2; const bool is_glm = mode & 4; + // backward process uses inverse rotation by cos and sin. + // cos and sin build a rotation matrix, where the inverse is the transpose. + // this essentially just switches the sign of sin. + const float sin_sign = forward ? 1.0f : -1.0f; + const int32_t * pos = (const int32_t *) src1->data; for (int64_t i3 = 0; i3 < ne3; i3++) { @@ -10965,18 +11081,18 @@ static void ggml_compute_forward_rope_f32( if (ir++ < ir0) continue; if (ir > ir1) break; - float theta = freq_scale * (float)p; + float theta_base = (float)p; if (is_glm) { - theta = MIN(p, n_ctx - 2); + theta_base = MIN(p, n_ctx - 2); float block_theta = MAX(p - (n_ctx - 2), 0); for (int64_t i0 = 0; i0 < ne0 / 4; i0++) { - const float cos_theta = cosf(theta); - const float sin_theta = sinf(theta); + const float cos_theta = cosf(theta_base); + const float sin_theta = sinf(theta_base) * sin_sign; const float cos_block_theta = cosf(block_theta); - const float sin_block_theta = sinf(block_theta); + const float sin_block_theta = sinf(block_theta) * sin_sign; - theta *= theta_scale; + theta_base *= theta_scale; block_theta *= theta_scale; const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); @@ -10994,13 +11110,17 @@ static void ggml_compute_forward_rope_f32( } } else if (!is_neox) { for (int64_t i0 = 0; i0 < ne0; i0 += 2) { - const float cos_theta = cosf(theta); - const float sin_theta = sinf(theta); + float cos_theta, sin_theta; + rope_yarn( + theta_base, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta + ); + sin_theta *= sin_sign; + // zeta scaling for xPos only: float zeta = xpos_base != 0.0f ? powf((i0 + 0.4f * ne0) / (1.4f * ne0), p / xpos_base) : 1.0f; if (xpos_down) zeta = 1.0f / zeta; - theta *= theta_scale; + theta_base *= theta_scale; const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); float * dst_data = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); @@ -11014,12 +11134,20 @@ static void ggml_compute_forward_rope_f32( } else { // TODO: this might be wrong for ne0 != n_dims - need double check // ref: https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt_neox/modeling_gpt_neox.py#LL251C1-L294C28 + theta_base *= freq_scale; for (int64_t ib = 0; ib < ne0/n_dims; ++ib) { for (int64_t ic = 0; ic < n_dims; ic += 2) { - const float cos_theta = cosf(theta); - const float sin_theta = sinf(theta); + // simplified from `(ib * n_dims + ic) * inv_ndims` + float cur_rot = inv_ndims * ic - ib; - theta *= theta_scale; + float cos_theta, sin_theta; + rope_yarn( + theta_base, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, + &cos_theta, &sin_theta + ); + sin_theta *= sin_sign; + + theta_base *= theta_scale; const int64_t i0 = ib*n_dims + ic/2; @@ -11043,20 +11171,25 @@ static void ggml_compute_forward_rope_f16( const struct ggml_compute_params * params, const struct ggml_tensor * src0, const struct ggml_tensor * src1, - struct ggml_tensor * dst) { + struct ggml_tensor * dst, + const bool forward) { if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { return; } - float freq_base; - float freq_scale; + float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow; - //const int n_past = ((int32_t *) dst->op_params)[0]; - const int n_dims = ((int32_t *) dst->op_params)[1]; - const int mode = ((int32_t *) dst->op_params)[2]; - const int n_ctx = ((int32_t *) dst->op_params)[3]; - memcpy(&freq_base, (int32_t *) dst->op_params + 4, sizeof(float)); - memcpy(&freq_scale, (int32_t *) dst->op_params + 5, sizeof(float)); + //const int n_past = ((int32_t *) dst->op_params)[0]; + const int n_dims = ((int32_t *) dst->op_params)[1]; + const int mode = ((int32_t *) dst->op_params)[2]; + const int n_ctx = ((int32_t *) dst->op_params)[3]; + const int n_orig_ctx = ((int32_t *) dst->op_params)[4]; + memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float)); + memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float)); + memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float)); + memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float)); + memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float)); + memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float)); GGML_TENSOR_UNARY_OP_LOCALS @@ -11084,10 +11217,18 @@ static void ggml_compute_forward_rope_f16( int ir = 0; const float theta_scale = powf(freq_base, -2.0f/n_dims); + const float inv_ndims = -1.f/n_dims; + float corr_dims[2]; + ggml_rope_yarn_corr_dims(n_dims, n_orig_ctx, freq_base, beta_fast, beta_slow, corr_dims); const bool is_neox = mode & 2; const bool is_glm = mode & 4; + // backward process uses inverse rotation by cos and sin. + // cos and sin build a rotation matrix, where the inverse is the transpose. + // this essentially just switches the sign of sin. + const float sin_sign = forward ? 1.0f : -1.0f; + const int32_t * pos = (const int32_t *) src1->data; for (int64_t i3 = 0; i3 < ne3; i3++) { @@ -11097,18 +11238,18 @@ static void ggml_compute_forward_rope_f16( if (ir++ < ir0) continue; if (ir > ir1) break; - float theta = freq_scale * (float)p; + float theta_base = (float)p; if (is_glm) { - theta = MIN(p, n_ctx - 2); + theta_base = MIN(p, n_ctx - 2); float block_theta = MAX(p - (n_ctx - 2), 0); for (int64_t i0 = 0; i0 < ne0 / 4; i0++) { - const float cos_theta = cosf(theta); - const float sin_theta = sinf(theta); + const float cos_theta = cosf(theta_base); + const float sin_theta = sinf(theta_base) * sin_sign; const float cos_block_theta = cosf(block_theta); - const float sin_block_theta = sinf(block_theta); + const float sin_block_theta = sinf(block_theta) * sin_sign; - theta *= theta_scale; + theta_base *= theta_scale; block_theta *= theta_scale; const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); @@ -11126,10 +11267,13 @@ static void ggml_compute_forward_rope_f16( } } else if (!is_neox) { for (int64_t i0 = 0; i0 < ne0; i0 += 2) { - const float cos_theta = cosf(theta); - const float sin_theta = sinf(theta); + float cos_theta, sin_theta; + rope_yarn( + theta_base, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta + ); + sin_theta *= sin_sign; - theta *= theta_scale; + theta_base *= theta_scale; const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); @@ -11143,12 +11287,20 @@ static void ggml_compute_forward_rope_f16( } else { // TODO: this might be wrong for ne0 != n_dims - need double check // ref: https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt_neox/modeling_gpt_neox.py#LL251C1-L294C28 + theta_base *= freq_scale; for (int64_t ib = 0; ib < ne0/n_dims; ++ib) { for (int64_t ic = 0; ic < n_dims; ic += 2) { - const float cos_theta = cosf(theta); - const float sin_theta = sinf(theta); + // simplified from `(ib * n_dims + ic) * inv_ndims` + float cur_rot = inv_ndims * ic - ib; - theta *= theta_scale; + float cos_theta, sin_theta; + rope_yarn( + theta_base, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, + &cos_theta, &sin_theta + ); + sin_theta *= sin_sign; + + theta_base *= theta_scale; const int64_t i0 = ib*n_dims + ic/2; @@ -11176,11 +11328,11 @@ static void ggml_compute_forward_rope( switch (src0->type) { case GGML_TYPE_F16: { - ggml_compute_forward_rope_f16(params, src0, src1, dst); + ggml_compute_forward_rope_f16(params, src0, src1, dst, true); } break; case GGML_TYPE_F32: { - ggml_compute_forward_rope_f32(params, src0, src1, dst); + ggml_compute_forward_rope_f32(params, src0, src1, dst, true); } break; default: { @@ -11191,215 +11343,6 @@ static void ggml_compute_forward_rope( // ggml_compute_forward_rope_back -static void ggml_compute_forward_rope_back_f32( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - - if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { - return; - } - - // y = rope(x, src1) - // dx = rope_back(dy, src1) - // src0 is dy, src1 contains options - - float freq_base; - float freq_scale; - - // these two only relevant for xPos RoPE: - float xpos_base; - bool xpos_down; - - //const int n_past = ((int32_t *) dst->op_params)[0]; - const int n_dims = ((int32_t *) dst->op_params)[1]; - const int mode = ((int32_t *) dst->op_params)[2]; - const int n_ctx = ((int32_t *) dst->op_params)[3]; UNUSED(n_ctx); - memcpy(&freq_base, (int32_t *) dst->op_params + 4, sizeof(float)); - memcpy(&freq_scale, (int32_t *) dst->op_params + 5, sizeof(float)); - memcpy(&xpos_base, (int32_t *) dst->op_params + 6, sizeof(float)); - memcpy(&xpos_down, (int32_t *) dst->op_params + 7, sizeof(bool)); - - GGML_TENSOR_UNARY_OP_LOCALS - - //printf("ne0: %d, ne1: %d, ne2: %d, ne3: %d\n", ne0, ne1, ne2, ne3); - //printf("n_past = %d, ne2 = %d\n", n_past, ne2); - - assert(nb0 == sizeof(float)); - - const int ith = params->ith; - const int nth = params->nth; - - const int nr = ggml_nrows(dst); - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - // row index used to determine which thread to use - int ir = 0; - - const float theta_scale = powf(freq_base, -2.0f/n_dims); - - const bool is_neox = mode & 2; - - const int32_t * pos = (const int32_t *) src1->data; - - for (int64_t i3 = 0; i3 < ne3; i3++) { - for (int64_t i2 = 0; i2 < ne2; i2++) { - const int64_t p = pos[i2]; - for (int64_t i1 = 0; i1 < ne1; i1++) { - if (ir++ < ir0) continue; - if (ir > ir1) break; - - float theta = freq_scale * (float)p; - - if (!is_neox) { - for (int64_t i0 = 0; i0 < ne0; i0 += 2) { - const float cos_theta = cosf(theta); - const float sin_theta = sinf(theta); - // zeta scaling for xPos only: - float zeta = xpos_base != 0.0f ? powf((i0 + 0.4f * ne0) / (1.4f * ne0), p / xpos_base) : 1.0f; - if (xpos_down) zeta = 1.0f / zeta; - - theta *= theta_scale; - - const float * const dy = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); - float * dx = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); - - const float dy0 = dy[0]; - const float dy1 = dy[1]; - - dx[0] = dy0*cos_theta*zeta + dy1*sin_theta*zeta; - dx[1] = - dy0*sin_theta*zeta + dy1*cos_theta*zeta; - } - } else { - for (int64_t ib = 0; ib < ne0/n_dims; ++ib) { - for (int64_t ic = 0; ic < n_dims; ic += 2) { - const float cos_theta = cosf(theta); - const float sin_theta = sinf(theta); - - theta *= theta_scale; - - const int64_t i0 = ib*n_dims + ic/2; - - const float * const dy = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); - float * dx = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); - - const float dy0 = dy[0]; - const float dy1 = dy[n_dims/2]; - - dx[0] = dy0*cos_theta + dy1*sin_theta; - dx[n_dims/2] = - dy0*sin_theta + dy1*cos_theta; - } - } - } - } - } - } -} - -static void ggml_compute_forward_rope_back_f16( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - - if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { - return; - } - - // y = rope(x, src1) - // dx = rope_back(dy, src1) - // src0 is dy, src1 contains options - - //const int n_past = ((int32_t *) dst->op_params)[0]; - const int n_dims = ((int32_t *) dst->op_params)[1]; - const int mode = ((int32_t *) dst->op_params)[2]; - - GGML_TENSOR_UNARY_OP_LOCALS - - //printf("ne0: %d, ne1: %d, ne2: %d, ne3: %d\n", ne0, ne1, ne2, ne3); - //printf("n_past = %d, ne2 = %d\n", n_past, ne2); - - assert(nb0 == sizeof(ggml_fp16_t)); - - const int ith = params->ith; - const int nth = params->nth; - - const int nr = ggml_nrows(dst); - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - // row index used to determine which thread to use - int ir = 0; - - const float theta_scale = powf(10000.0, -2.0f/n_dims); - - const bool is_neox = mode & 2; - - const int32_t * pos = (const int32_t *) src1->data; - - for (int64_t i3 = 0; i3 < ne3; i3++) { - for (int64_t i2 = 0; i2 < ne2; i2++) { - const int64_t p = pos[i2]; - for (int64_t i1 = 0; i1 < ne1; i1++) { - if (ir++ < ir0) continue; - if (ir > ir1) break; - - float theta = (float)p; - - if (!is_neox) { - for (int64_t i0 = 0; i0 < ne0; i0 += 2) { - const float cos_theta = cosf(theta); - const float sin_theta = sinf(theta); - - theta *= theta_scale; - - const ggml_fp16_t * const dy = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); - ggml_fp16_t * dx = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); - - const float dy0 = GGML_FP16_TO_FP32(dy[0]); - const float dy1 = GGML_FP16_TO_FP32(dy[1]); - - dx[0] = GGML_FP32_TO_FP16( dy0*cos_theta + dy1*sin_theta); - dx[1] = GGML_FP32_TO_FP16(-dy0*sin_theta + dy1*cos_theta); - } - } else { - for (int64_t ib = 0; ib < ne0/n_dims; ++ib) { - for (int64_t ic = 0; ic < n_dims; ic += 2) { - const float cos_theta = cosf(theta); - const float sin_theta = sinf(theta); - - theta *= theta_scale; - - const int64_t i0 = ib*n_dims + ic/2; - - const ggml_fp16_t * const dy = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); - ggml_fp16_t * dx = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); - - const float dy0 = GGML_FP16_TO_FP32(dy[0]); - const float dy1 = GGML_FP16_TO_FP32(dy[n_dims/2]); - - dx[0] = GGML_FP32_TO_FP16( dy0*cos_theta + dy1*sin_theta); - dx[n_dims/2] = GGML_FP32_TO_FP16(-dy0*sin_theta + dy1*cos_theta); - } - } - } - } - } - } -} - static void ggml_compute_forward_rope_back( const struct ggml_compute_params * params, const struct ggml_tensor * src0, @@ -11408,421 +11351,11 @@ static void ggml_compute_forward_rope_back( switch (src0->type) { case GGML_TYPE_F16: { - ggml_compute_forward_rope_back_f16(params, src0, src1, dst); + ggml_compute_forward_rope_f16(params, src0, src1, dst, false); } break; case GGML_TYPE_F32: { - ggml_compute_forward_rope_back_f32(params, src0, src1, dst); - } break; - default: - { - GGML_ASSERT(false); - } break; - } -} - -// ggml_compute_forward_conv_1d - -static void ggml_compute_forward_conv_1d_f16_f32( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - GGML_ASSERT(src0->type == GGML_TYPE_F16); - GGML_ASSERT(src1->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - int64_t t0 = ggml_perf_time_us(); - UNUSED(t0); - - GGML_TENSOR_BINARY_OP_LOCALS - - const int ith = params->ith; - const int nth = params->nth; - - const int nk = ne00; - - // size of the convolution row - the kernel size unrolled across all input channels - const int ew0 = nk*ne01; - - const int32_t s0 = ((const int32_t*)(dst->op_params))[0]; - const int32_t p0 = ((const int32_t*)(dst->op_params))[1]; - const int32_t d0 = ((const int32_t*)(dst->op_params))[2]; - - GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); - GGML_ASSERT(nb10 == sizeof(float)); - - if (params->type == GGML_TASK_INIT) { - memset(params->wdata, 0, params->wsize); - - ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0; - - for (int64_t i11 = 0; i11 < ne11; i11++) { - const float * const src = (float *)((char *) src1->data + i11*nb11); - ggml_fp16_t * dst_data = wdata; - - for (int64_t i0 = 0; i0 < ne0; i0++) { - for (int64_t ik = 0; ik < nk; ik++) { - const int idx0 = i0*s0 + ik*d0 - p0; - - if(!(idx0 < 0 || idx0 >= ne10)) { - dst_data[i0*ew0 + i11*nk + ik] = GGML_FP32_TO_FP16(src[idx0]); - } - } - } - } - - return; - } - - if (params->type == GGML_TASK_FINALIZE) { - return; - } - - // total rows in dst - const int nr = ne2; - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0; - - for (int i2 = 0; i2 < ne2; i2++) { - for (int i1 = ir0; i1 < ir1; i1++) { - float * dst_data = (float *)((char *) dst->data + i2*nb2 + i1*nb1); - - for (int i0 = 0; i0 < ne0; i0++) { - ggml_vec_dot_f16(ew0, dst_data + i0, - (ggml_fp16_t *) ((char *) src0->data + i1*nb02), - (ggml_fp16_t *) wdata + i2*nb2 + i0*ew0); - } - } - } -} - -static void ggml_compute_forward_conv_1d_f32( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT(src1->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - int64_t t0 = ggml_perf_time_us(); - UNUSED(t0); - - GGML_TENSOR_BINARY_OP_LOCALS - - const int ith = params->ith; - const int nth = params->nth; - - const int nk = ne00; - - const int ew0 = nk*ne01; - - const int32_t s0 = ((const int32_t*)(dst->op_params))[0]; - const int32_t p0 = ((const int32_t*)(dst->op_params))[1]; - const int32_t d0 = ((const int32_t*)(dst->op_params))[2]; - - GGML_ASSERT(nb00 == sizeof(float)); - GGML_ASSERT(nb10 == sizeof(float)); - - if (params->type == GGML_TASK_INIT) { - memset(params->wdata, 0, params->wsize); - - float * const wdata = (float *) params->wdata + 0; - - for (int64_t i11 = 0; i11 < ne11; i11++) { - const float * const src = (float *)((char *) src1->data + i11*nb11); - float * dst_data = wdata; - - for (int64_t i0 = 0; i0 < ne0; i0++) { - for (int64_t ik = 0; ik < nk; ik++) { - const int idx0 = i0*s0 + ik*d0 - p0; - - if(!(idx0 < 0 || idx0 >= ne10)) { - dst_data[i0*ew0 + i11*nk + ik] = src[idx0]; - } - } - } - } - - return; - } - - if (params->type == GGML_TASK_FINALIZE) { - return; - } - - // total rows in dst - const int nr = ne02; - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - float * const wdata = (float *) params->wdata + 0; - - for (int i2 = 0; i2 < ne2; i2++) { - for (int i1 = ir0; i1 < ir1; i1++) { - float * dst_data = (float *)((char *) dst->data + i2*nb2 + i1*nb1); - - for (int i0 = 0; i0 < ne0; i0++) { - ggml_vec_dot_f32(ew0, dst_data + i0, - (float *) ((char *) src0->data + i1*nb02), - (float *) wdata + i2*nb2 + i0*ew0); - } - } - } -} - -// TODO: reuse ggml_mul_mat or implement ggml_im2col and remove stage_0 and stage_1 -static void gemm_f16_out_f32(int64_t m, int64_t n, int64_t k, - ggml_fp16_t * A, - ggml_fp16_t * B, - float * C, - const int ith, const int nth) { - // does not seem to make a difference - int64_t m0, m1, n0, n1; - // patches per thread - if (m > n) { - n0 = 0; - n1 = n; - - // total patches in dst - const int np = m; - - // patches per thread - const int dp = (np + nth - 1)/nth; - - // patch range for this thread - m0 = dp*ith; - m1 = MIN(m0 + dp, np); - } else { - m0 = 0; - m1 = m; - - // total patches in dst - const int np = n; - - // patches per thread - const int dp = (np + nth - 1)/nth; - - // patch range for this thread - n0 = dp*ith; - n1 = MIN(n0 + dp, np); - } - - // block-tiling attempt - int64_t blck_n = 16; - int64_t blck_m = 16; - - // int64_t CACHE_SIZE = 2 * 1024 * 1024; // 2MB - // int64_t blck_size = CACHE_SIZE / (sizeof(float) + 2 * sizeof(ggml_fp16_t) * K); - // if (blck_size > 0) { - // blck_0 = 4; - // blck_1 = blck_size / blck_0; - // if (blck_1 < 0) { - // blck_1 = 1; - // } - // // blck_0 = (int64_t)sqrt(blck_size); - // // blck_1 = blck_0; - // } - // // printf("%zd %zd %zd %zd\n", blck_size, K, blck_0, blck_1); - - for (int j = n0; j < n1; j+=blck_n) { - for (int i = m0; i < m1; i+=blck_m) { - // printf("i j k => %d %d %d\n", i, j, K); - for (int ii = i; ii < i + blck_m && ii < m1; ii++) { - for (int jj = j; jj < j + blck_n && jj < n1; jj++) { - ggml_vec_dot_f16(k, - C + ii*n + jj, - A + ii * k, - B + jj * k); - } - } - } - } -} - -// src0: kernel [OC, IC, K] -// src1: signal [N, IC, IL] -// dst: result [N, OL, IC*K] -static void ggml_compute_forward_conv_1d_stage_0_f32( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - GGML_ASSERT(src0->type == GGML_TYPE_F16); - GGML_ASSERT(src1->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F16); - - int64_t t0 = ggml_perf_time_us(); - UNUSED(t0); - - GGML_TENSOR_BINARY_OP_LOCALS; - - const int64_t N = ne12; - const int64_t IC = ne11; - const int64_t IL = ne10; - - const int64_t K = ne00; - - const int64_t OL = ne1; - - const int ith = params->ith; - const int nth = params->nth; - - const int32_t s0 = ((const int32_t*)(dst->op_params))[0]; - const int32_t p0 = ((const int32_t*)(dst->op_params))[1]; - const int32_t d0 = ((const int32_t*)(dst->op_params))[2]; - - GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); - GGML_ASSERT(nb10 == sizeof(float)); - - if (params->type == GGML_TASK_INIT) { - memset(dst->data, 0, ggml_nbytes(dst)); - return; - } - - if (params->type == GGML_TASK_FINALIZE) { - return; - } - - // im2col: [N, IC, IL] => [N, OL, IC*K] - { - ggml_fp16_t * const wdata = (ggml_fp16_t *) dst->data; - - for (int64_t in = 0; in < N; in++) { - for (int64_t iol = 0; iol < OL; iol++) { - for (int64_t iic = ith; iic < IC; iic+=nth) { - - // micro kernel - ggml_fp16_t * dst_data = wdata + (in*OL + iol)*(IC*K); // [IC, K] - const float * const src_data = (float *)((char *) src1->data + in*nb12 + iic*nb11); // [IL] - - for (int64_t ik = 0; ik < K; ik++) { - const int64_t iil = iol*s0 + ik*d0 - p0; - - if (!(iil < 0 || iil >= IL)) { - dst_data[iic*K + ik] = GGML_FP32_TO_FP16(src_data[iil]); - } - } - } - } - } - } -} - -// gemm: [N, OC, OL] = [OC, IC * K] x [N*OL, IC * K] -// src0: [OC, IC, K] -// src1: [N, OL, IC * K] -// result: [N, OC, OL] -static void ggml_compute_forward_conv_1d_stage_1_f16( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - GGML_ASSERT(src0->type == GGML_TYPE_F16); - GGML_ASSERT(src1->type == GGML_TYPE_F16); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - int64_t t0 = ggml_perf_time_us(); - UNUSED(t0); - - if (params->type == GGML_TASK_INIT) { - return; - } - - if (params->type == GGML_TASK_FINALIZE) { - return; - } - - GGML_TENSOR_BINARY_OP_LOCALS; - - GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); - GGML_ASSERT(nb10 == sizeof(ggml_fp16_t)); - GGML_ASSERT(nb0 == sizeof(float)); - - const int N = ne12; - const int OL = ne11; - - const int OC = ne02; - const int IC = ne01; - const int K = ne00; - - const int ith = params->ith; - const int nth = params->nth; - - int64_t m = OC; - int64_t n = OL; - int64_t k = IC * K; - - // [N, OC, OL] = [OC, IC * K] x [N*OL, IC * K] - for (int i = 0; i < N; i++) { - ggml_fp16_t * A = (ggml_fp16_t *)src0->data; // [m, k] - ggml_fp16_t * B = (ggml_fp16_t *)src1->data + i * m * k; // [n, k] - float * C = (float *)dst->data + i * m * n; // [m, n] - - gemm_f16_out_f32(m, n, k, A, B, C, ith, nth); - } -} - -static void ggml_compute_forward_conv_1d( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - switch(src0->type) { - case GGML_TYPE_F16: - { - ggml_compute_forward_conv_1d_f16_f32(params, src0, src1, dst); - } break; - case GGML_TYPE_F32: - { - ggml_compute_forward_conv_1d_f32(params, src0, src1, dst); - } break; - default: - { - GGML_ASSERT(false); - } break; - } -} - -static void ggml_compute_forward_conv_1d_stage_0( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - switch(src0->type) { - case GGML_TYPE_F16: - { - ggml_compute_forward_conv_1d_stage_0_f32(params, src0, src1, dst); - } break; - default: - { - GGML_ASSERT(false); - } break; - } -} - -static void ggml_compute_forward_conv_1d_stage_1( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - switch(src0->type) { - case GGML_TYPE_F16: - { - ggml_compute_forward_conv_1d_stage_1_f16(params, src0, src1, dst); + ggml_compute_forward_rope_f32(params, src0, src1, dst, false); } break; default: { @@ -12042,12 +11575,10 @@ static void ggml_compute_forward_conv_transpose_1d( } } -// ggml_compute_forward_conv_2d - // src0: kernel [OC, IC, KH, KW] // src1: image [N, IC, IH, IW] // dst: result [N, OH, OW, IC*KH*KW] -static void ggml_compute_forward_conv_2d_stage_0_f32( +static void ggml_compute_forward_im2col_f16( const struct ggml_compute_params * params, const struct ggml_tensor * src0, const struct ggml_tensor * src1, @@ -12061,34 +11592,35 @@ static void ggml_compute_forward_conv_2d_stage_0_f32( GGML_TENSOR_BINARY_OP_LOCALS; - const int64_t N = ne13; - const int64_t IC = ne12; - const int64_t IH = ne11; - const int64_t IW = ne10; - - // const int64_t OC = ne03; - // const int64_t IC = ne02; - const int64_t KH = ne01; - const int64_t KW = ne00; - - const int64_t OH = ne2; - const int64_t OW = ne1; + const int32_t s0 = ((const int32_t *)(dst->op_params))[0]; + const int32_t s1 = ((const int32_t *)(dst->op_params))[1]; + const int32_t p0 = ((const int32_t *)(dst->op_params))[2]; + const int32_t p1 = ((const int32_t *)(dst->op_params))[3]; + const int32_t d0 = ((const int32_t *)(dst->op_params))[4]; + const int32_t d1 = ((const int32_t *)(dst->op_params))[5]; + const bool is_2D = ((const int32_t *)(dst->op_params))[6] == 1; const int ith = params->ith; const int nth = params->nth; - const int32_t s0 = ((const int32_t*)(dst->op_params))[0]; - const int32_t s1 = ((const int32_t*)(dst->op_params))[1]; - const int32_t p0 = ((const int32_t*)(dst->op_params))[2]; - const int32_t p1 = ((const int32_t*)(dst->op_params))[3]; - const int32_t d0 = ((const int32_t*)(dst->op_params))[4]; - const int32_t d1 = ((const int32_t*)(dst->op_params))[5]; + const int64_t N = is_2D ? ne13 : ne12; + const int64_t IC = is_2D ? ne12 : ne11; + const int64_t IH = is_2D ? ne11 : 1; + const int64_t IW = ne10; + + const int64_t KH = is_2D ? ne01 : 1; + const int64_t KW = ne00; + + const int64_t OH = is_2D ? ne2 : 1; + const int64_t OW = ne1; + + int ofs0 = is_2D ? nb13 : nb12; + int ofs1 = is_2D ? nb12 : nb11; GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); GGML_ASSERT(nb10 == sizeof(float)); if (params->type == GGML_TASK_INIT) { - memset(dst->data, 0, ggml_nbytes(dst)); return; } @@ -12101,20 +11633,22 @@ static void ggml_compute_forward_conv_2d_stage_0_f32( ggml_fp16_t * const wdata = (ggml_fp16_t *) dst->data; for (int64_t in = 0; in < N; in++) { - for (int64_t ioh = 0; ioh < OH; ioh++) { + for (int64_t ioh = 0; ioh < OH; ioh++) { // 1 for (int64_t iow = 0; iow < OW; iow++) { - for (int64_t iic = ith; iic < IC; iic+=nth) { + for (int64_t iic = ith; iic < IC; iic += nth) { // micro kernel ggml_fp16_t * dst_data = wdata + (in*OH*OW + ioh*OW + iow)*(IC*KH*KW); // [IC, KH, KW] - const float * const src_data = (float *)((char *) src1->data + in*nb13 + iic*nb12); // [IH, IW] + const float * const src_data = (float *)((char *) src1->data + in*ofs0 + iic*ofs1); // [IH, IW] - for (int64_t ikh = 0; ikh < KH; ikh++) { + for (int64_t ikh = 0; ikh < KH; ikh++) { // 1 for (int64_t ikw = 0; ikw < KW; ikw++) { const int64_t iiw = iow*s0 + ikw*d0 - p0; const int64_t iih = ioh*s1 + ikh*d1 - p1; - if (!(iih < 0 || iih >= IH || iiw < 0 || iiw >= IW)) { + if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) { + dst_data[iic*(KH*KW) + ikh*KW + ikw] = 0; + } else { dst_data[iic*(KH*KW) + ikh*KW + ikw] = GGML_FP32_TO_FP16(src_data[iih*IW + iiw]); } } @@ -12126,180 +11660,7 @@ static void ggml_compute_forward_conv_2d_stage_0_f32( } } -// gemm: [N, OC, OH, OW] = [OC, IC * KH * KW] x [N*OH*OW, IC * KH * KW] -// src0: [OC, IC, KH, KW] -// src1: [N, OH, OW, IC * KH * KW] -// result: [N, OC, OH, OW] -static void ggml_compute_forward_conv_2d_stage_1_f16( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - GGML_ASSERT(src0->type == GGML_TYPE_F16); - GGML_ASSERT(src1->type == GGML_TYPE_F16); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - int64_t t0 = ggml_perf_time_us(); - UNUSED(t0); - - if (params->type == GGML_TASK_INIT) { - return; - } - - if (params->type == GGML_TASK_FINALIZE) { - return; - } - - GGML_TENSOR_BINARY_OP_LOCALS; - - GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); - GGML_ASSERT(nb10 == sizeof(ggml_fp16_t)); - GGML_ASSERT(nb0 == sizeof(float)); - - const int N = ne13; - const int OH = ne12; - const int OW = ne11; - - const int OC = ne03; - const int IC = ne02; - const int KH = ne01; - const int KW = ne00; - - const int ith = params->ith; - const int nth = params->nth; - - int64_t m = OC; - int64_t n = OH * OW; - int64_t k = IC * KH * KW; - - // [N, OC, OH, OW] = [OC, IC * KH * KW] x [N*OH*OW, IC * KH * KW] - for (int i = 0; i < N; i++) { - ggml_fp16_t * A = (ggml_fp16_t *)src0->data; // [m, k] - ggml_fp16_t * B = (ggml_fp16_t *)src1->data + i * m * k; // [n, k] - float * C = (float *)dst->data + i * m * n; // [m, n] - - gemm_f16_out_f32(m, n, k, A, B, C, ith, nth); - } -} - -static void ggml_compute_forward_conv_2d_f16_f32( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - GGML_ASSERT(src0->type == GGML_TYPE_F16); - GGML_ASSERT(src1->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - int64_t t0 = ggml_perf_time_us(); - UNUSED(t0); - - GGML_TENSOR_BINARY_OP_LOCALS - - // src1: image [N, IC, IH, IW] - // src0: kernel [OC, IC, KH, KW] - // dst: result [N, OC, OH, OW] - // ne12: IC - // ne0: OW - // ne1: OH - // nk0: KW - // nk1: KH - // ne13: N - - const int N = ne13; - const int IC = ne12; - const int IH = ne11; - const int IW = ne10; - - const int OC = ne03; - // const int IC = ne02; - const int KH = ne01; - const int KW = ne00; - - const int OH = ne1; - const int OW = ne0; - - const int ith = params->ith; - const int nth = params->nth; - - // const int nk0 = ne00; - // const int nk1 = ne01; - - // size of the convolution row - the kernel size unrolled across all channels - // const int ew0 = nk0*nk1*ne02; - // ew0: IC*KH*KW - - const int32_t s0 = ((const int32_t*)(dst->op_params))[0]; - const int32_t s1 = ((const int32_t*)(dst->op_params))[1]; - const int32_t p0 = ((const int32_t*)(dst->op_params))[2]; - const int32_t p1 = ((const int32_t*)(dst->op_params))[3]; - const int32_t d0 = ((const int32_t*)(dst->op_params))[4]; - const int32_t d1 = ((const int32_t*)(dst->op_params))[5]; - - GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); - GGML_ASSERT(nb10 == sizeof(float)); - - if (params->type == GGML_TASK_INIT) { - memset(params->wdata, 0, params->wsize); - - // prepare source data (src1) - // im2col: [N, IC, IH, IW] => [N*OH*OW, IC*KH*KW] - - { - ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0; - - for (int in = 0; in < N; in++) { - for (int iic = 0; iic < IC; iic++) { - for (int ioh = 0; ioh < OH; ioh++) { - for (int iow = 0; iow < OW; iow++) { - - // micro kernel - ggml_fp16_t * dst_data = wdata + (in*OH*OW + ioh*OW + iow)*(IC*KH*KW); // [IC, KH, KW] - const float * const src_data = (float *)((char *) src1->data + in*nb13 + iic*nb12); // [IH, IW] - - for (int ikh = 0; ikh < KH; ikh++) { - for (int ikw = 0; ikw < KW; ikw++) { - const int iiw = iow*s0 + ikw*d0 - p0; - const int iih = ioh*s1 + ikh*d1 - p1; - - if (!(iih < 0 || iih >= IH || iiw < 0 || iiw >= IW)) { - dst_data[iic*(KH*KW) + ikh*KW + ikw] = GGML_FP32_TO_FP16(src_data[iih*IW + iiw]); - } - } - } - } - } - } - } - } - - return; - } - - if (params->type == GGML_TASK_FINALIZE) { - return; - } - - ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0; - // wdata: [N*OH*OW, IC*KH*KW] - // dst: result [N, OC, OH, OW] - // src0: kernel [OC, IC, KH, KW] - - int64_t m = OC; - int64_t n = OH * OW; - int64_t k = IC * KH * KW; - - // [N, OC, OH, OW] = [OC, IC * KH * KW] x [N*OH*OW, IC * KH * KW] - for (int i = 0; i < N; i++) { - ggml_fp16_t * A = (ggml_fp16_t *)src0->data; // [m, k] - ggml_fp16_t * B = (ggml_fp16_t *)wdata + i * m * k; // [n, k] - float * C = (float *)dst->data + i * m * n; // [m * k] - - gemm_f16_out_f32(m, n, k, A, B, C, ith, nth); - } -} - -static void ggml_compute_forward_conv_2d( +static void ggml_compute_forward_im2col( const struct ggml_compute_params * params, const struct ggml_tensor * src0, const struct ggml_tensor * src1, @@ -12307,50 +11668,7 @@ static void ggml_compute_forward_conv_2d( switch (src0->type) { case GGML_TYPE_F16: { - ggml_compute_forward_conv_2d_f16_f32(params, src0, src1, dst); - } break; - case GGML_TYPE_F32: - { - //ggml_compute_forward_conv_2d_f32(params, src0, src1, dst); - GGML_ASSERT(false); - } break; - default: - { - GGML_ASSERT(false); - } break; - } -} - -static void ggml_compute_forward_conv_2d_stage_0( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - switch (src0->type) { - case GGML_TYPE_F16: - { - ggml_compute_forward_conv_2d_stage_0_f32(params, src0, src1, dst); - } break; - case GGML_TYPE_F32: - { - GGML_ASSERT(false); - } break; - default: - { - GGML_ASSERT(false); - } break; - } -} - -static void ggml_compute_forward_conv_2d_stage_1( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - switch (src0->type) { - case GGML_TYPE_F16: - { - ggml_compute_forward_conv_2d_stage_1_f16(params, src0, src1, dst); + ggml_compute_forward_im2col_f16(params, src0, src1, dst); } break; case GGML_TYPE_F32: { @@ -12535,14 +11853,11 @@ static void ggml_compute_forward_pool_1d( ggml_compute_forward_pool_1d_sk_p0(params, op, src0, k0, dst); } -// ggml_compute_forward_pool_2d_sk_p0 +// ggml_compute_forward_pool_2d -static void ggml_compute_forward_pool_2d_sk_p0( +static void ggml_compute_forward_pool_2d( const struct ggml_compute_params * params, - const enum ggml_op_pool op, const struct ggml_tensor * src, - const int k0, - const int k1, struct ggml_tensor * dst) { assert(src->type == GGML_TYPE_F32); assert(params->ith == 0); @@ -12551,6 +11866,14 @@ static void ggml_compute_forward_pool_2d_sk_p0( return; } + const int32_t * opts = (const int32_t *)dst->op_params; + enum ggml_op_pool op = opts[0]; + const int k0 = opts[1]; + const int k1 = opts[2]; + const int s0 = opts[3]; + const int s1 = opts[4]; + const int p0 = opts[5]; + const int p1 = opts[6]; const char * cdata = (const char*)src->data; const char * const data_end = cdata + ggml_nbytes(src); @@ -12561,6 +11884,8 @@ static void ggml_compute_forward_pool_2d_sk_p0( float * dplane = (float *)dst->data; const int ka = k0 * k1; + const int offset0 = -p0; + const int offset1 = -p1; while (cdata < data_end) { for (int oy = 0; oy < py; ++oy) { @@ -12573,13 +11898,15 @@ static void ggml_compute_forward_pool_2d_sk_p0( case GGML_OP_POOL_COUNT: GGML_ASSERT(false); break; } - const int ix = ox * k0; - const int iy = oy * k1; + const int ix = offset0 + ox * s0; + const int iy = offset1 + oy * s1; for (int ky = 0; ky < k1; ++ky) { + if (iy + ky < 0 || iy + ky >= src->ne[1]) continue; const float * const srow = (const float *)(cdata + src->nb[1] * (iy + ky)); for (int kx = 0; kx < k0; ++kx) { int j = ix + kx; + if (j < 0 || j >= src->ne[0]) continue; switch (op) { case GGML_OP_POOL_AVG: *out += srow[j]; break; case GGML_OP_POOL_MAX: if (srow[j] > *out) *out = srow[j]; break; @@ -12600,29 +11927,6 @@ static void ggml_compute_forward_pool_2d_sk_p0( } } -// ggml_compute_forward_pool_2d - -static void ggml_compute_forward_pool_2d( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - struct ggml_tensor * dst) { - - const int32_t * opts = (const int32_t *)dst->op_params; - enum ggml_op_pool op = opts[0]; - const int k0 = opts[1]; - const int k1 = opts[2]; - const int s0 = opts[3]; - const int s1 = opts[4]; - const int p0 = opts[5]; - const int p1 = opts[6]; - GGML_ASSERT(p0 == 0); - GGML_ASSERT(p1 == 0); // padding not supported - GGML_ASSERT(k0 == s0); - GGML_ASSERT(k1 == s1); // only s = k supported - - ggml_compute_forward_pool_2d_sk_p0(params, op, src0, k0, k1, dst); -} - // ggml_compute_forward_upscale static void ggml_compute_forward_upscale_f32( @@ -13824,6 +13128,10 @@ static void ggml_compute_forward_unary( { ggml_compute_forward_silu(params, src0, dst); } break; + case GGML_UNARY_OP_LEAKY: + { + ggml_compute_forward_leaky(params, src0, dst); + } break; default: { GGML_ASSERT(false); @@ -14577,33 +13885,13 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm { ggml_compute_forward_clamp(params, tensor->src[0], tensor); } break; - case GGML_OP_CONV_1D: - { - ggml_compute_forward_conv_1d(params, tensor->src[0], tensor->src[1], tensor); - } break; - case GGML_OP_CONV_1D_STAGE_0: - { - ggml_compute_forward_conv_1d_stage_0(params, tensor->src[0], tensor->src[1], tensor); - } break; - case GGML_OP_CONV_1D_STAGE_1: - { - ggml_compute_forward_conv_1d_stage_1(params, tensor->src[0], tensor->src[1], tensor); - } break; case GGML_OP_CONV_TRANSPOSE_1D: { ggml_compute_forward_conv_transpose_1d(params, tensor->src[0], tensor->src[1], tensor); } break; - case GGML_OP_CONV_2D: + case GGML_OP_IM2COL: { - ggml_compute_forward_conv_2d(params, tensor->src[0], tensor->src[1], tensor); - } break; - case GGML_OP_CONV_2D_STAGE_0: - { - ggml_compute_forward_conv_2d_stage_0(params, tensor->src[0], tensor->src[1], tensor); - } break; - case GGML_OP_CONV_2D_STAGE_1: - { - ggml_compute_forward_conv_2d_stage_1(params, tensor->src[0], tensor->src[1], tensor); + ggml_compute_forward_im2col(params, tensor->src[0], tensor->src[1], tensor); } break; case GGML_OP_CONV_TRANSPOSE_2D: { @@ -14732,62 +14020,109 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm //////////////////////////////////////////////////////////////////////////////// -static_assert(GGML_GRAPH_HASHTABLE_SIZE > GGML_MAX_NODES * 2, "GGML_GRAPH_HT_SIZE is too small"); +static size_t ggml_hash_size(size_t min_sz) { + // next primes after powers of two + static const size_t primes[] = { + 2, 3, 5, 11, 17, 37, 67, 131, 257, 521, 1031, + 2053, 4099, 8209, 16411, 32771, 65537, 131101, + 262147, 524309, 1048583, 2097169, 4194319, 8388617, + 16777259, 33554467, 67108879, 134217757, 268435459, + 536870923, 1073741827, 2147483659 + }; + static const size_t n_primes = sizeof(primes)/sizeof(primes[0]); -static size_t hash(void * p) { - return (size_t)p % GGML_GRAPH_HASHTABLE_SIZE; + // find the smallest prime that is larger or equal to min_sz + size_t l = 0; + size_t r = n_primes; + while (l < r) { + size_t m = (l + r)/2; + if (primes[m] < min_sz) { + l = m + 1; + } else { + r = m; + } + } + size_t sz = l < n_primes ? primes[l] : min_sz | 1; + return sz; } -static size_t hash_find(void * hash_table[], void * p) { - size_t h = hash(p); +static size_t ggml_hash(const void * p) { + return (size_t)p; +} + +size_t ggml_hash_find(const struct ggml_hash_set hash_set, struct ggml_tensor * key) { + size_t h = ggml_hash(key) % hash_set.size; // linear probing size_t i = h; - while (hash_table[i] != NULL && hash_table[i] != p) { - i = (i + 1) % GGML_GRAPH_HASHTABLE_SIZE; + while (hash_set.keys[i] != NULL && hash_set.keys[i] != key) { + i = (i + 1) % hash_set.size; if (i == h) { // visited all hash table entries -> not found - return GGML_GRAPH_HASHTABLE_SIZE; + return GGML_HASHTABLE_FULL; } } return i; } -static bool hash_insert(void * hash_table[], void * p) { - size_t i = hash_find(hash_table, p); +bool ggml_hash_contains(struct ggml_hash_set hash_set, struct ggml_tensor * key) { + size_t i = ggml_hash_find(hash_set, key); + return i != GGML_HASHTABLE_FULL && hash_set.keys[i] == key; +} - GGML_ASSERT(i < GGML_GRAPH_HASHTABLE_SIZE); // assert that not full +size_t ggml_hash_insert(struct ggml_hash_set hash_set, struct ggml_tensor * key) { + size_t i = ggml_hash_find(hash_set, key); - if (hash_table[i] == p) { - return true; + GGML_ASSERT(i != GGML_HASHTABLE_FULL); + + if (hash_set.keys[i] == key) { + return GGML_HASHTABLE_ALREADY_EXISTS; } // insert - GGML_ASSERT(hash_table[i] == NULL); - hash_table[i] = p; - return false; + GGML_ASSERT(hash_set.keys[i] == NULL); + hash_set.keys[i] = key; + return i; } -static bool hash_contains(void * hash_table[], void * p) { - size_t i = hash_find(hash_table, p); - return (i < GGML_GRAPH_HASHTABLE_SIZE) && (hash_table[i] == p); +size_t ggml_hash_find_or_insert(struct ggml_hash_set hash_set, struct ggml_tensor * key) { + size_t i = ggml_hash_find(hash_set, key); + + GGML_ASSERT(i != GGML_HASHTABLE_FULL); + + hash_set.keys[i] = key; + return i; } -struct hash_map { - void * keys[GGML_GRAPH_HASHTABLE_SIZE]; - void * vals[GGML_GRAPH_HASHTABLE_SIZE]; -}; - -static struct hash_map * new_hash_map(void) { - struct hash_map * result = malloc(sizeof(struct hash_map)); - for (int i=0; ikeys[i] = NULL; - result->vals[i] = NULL; - } +static struct ggml_hash_set ggml_hash_set_new(size_t size) { + size = ggml_hash_size(size); + struct ggml_hash_set result; + result.size = size; + result.keys = malloc(sizeof(struct ggml_tensor *) * size); + memset(result.keys, 0, sizeof(struct ggml_tensor *) * size); return result; } -static void free_hash_map(struct hash_map * map) { +static void ggml_hash_set_free(struct ggml_hash_set hash_set) { + free(hash_set.keys); +} + +struct hash_map { + struct ggml_hash_set set; + struct ggml_tensor ** vals; +}; + +static struct hash_map * ggml_new_hash_map(size_t size) { + struct hash_map * result = malloc(sizeof(struct hash_map)); + result->set = ggml_hash_set_new(size); + result->vals = malloc(sizeof(struct ggml_tensor *) * result->set.size); + memset(result->vals, 0, sizeof(struct ggml_tensor *) * result->set.size); + return result; +} + +static void ggml_hash_map_free(struct hash_map * map) { + ggml_hash_set_free(map->set); + free(map->vals); free(map); } @@ -14807,7 +14142,7 @@ static struct ggml_tensor * ggml_recompute_graph_node( return node; } - if (!hash_contains(graph->visited_hash_table, node)) { + if (!ggml_hash_contains(graph->visited_hash_table, node)) { return node; } @@ -14822,17 +14157,17 @@ static struct ggml_tensor * ggml_recompute_graph_node( return node; } - size_t i = hash_find(replacements->keys, node); - GGML_ASSERT(i < GGML_GRAPH_HASHTABLE_SIZE); // assert that not full - if (replacements->keys[i] == node) { - return (struct ggml_tensor *) replacements->vals[i]; + size_t i = ggml_hash_find(replacements->set, node); + GGML_ASSERT(i != GGML_HASHTABLE_FULL); // assert that not full + if (replacements->set.keys[i] == node) { + return replacements->vals[i]; } struct ggml_tensor * clone = ggml_new_tensor(ctx, node->type, node->n_dims, node->ne); // insert clone into replacements - GGML_ASSERT(replacements->keys[i] == NULL); // assert that we don't overwrite - replacements->keys[i] = node; + GGML_ASSERT(replacements->set.keys[i] == NULL); // assert that we don't overwrite + replacements->set.keys[i] = node; replacements->vals[i] = clone; clone->op = node->op; @@ -14869,26 +14204,26 @@ void ggml_build_backward_gradient_checkpointing( struct ggml_cgraph * gb_tmp, struct ggml_tensor * * checkpoints, int n_checkpoints) { - *gb_tmp = *gf; + ggml_graph_cpy(gf, gb_tmp); ggml_build_backward_expand(ctx, gf, gb_tmp, true); if (n_checkpoints <= 0) { - *gb = *gb_tmp; + ggml_graph_cpy(gb_tmp, gb); return; } - struct hash_map * replacements = new_hash_map(); + struct hash_map * replacements = ggml_new_hash_map(gf->n_nodes + gf->n_leafs + n_checkpoints); // insert checkpoints in replacements for (int i = 0; i < n_checkpoints; ++i) { - size_t k = hash_find(replacements->keys, checkpoints[i]); - GGML_ASSERT(k < GGML_GRAPH_HASHTABLE_SIZE); // assert that not full - GGML_ASSERT(replacements->keys[k] == NULL); // assert that we don't overwrite - replacements->keys[k] = checkpoints[i]; - replacements->vals[k] = checkpoints[i]; + size_t k = ggml_hash_find(replacements->set, checkpoints[i]); + GGML_ASSERT(k != GGML_HASHTABLE_FULL); // assert that not full + GGML_ASSERT(replacements->set.keys[k] == NULL); // assert that we don't overwrite + replacements->set.keys[k] = checkpoints[i]; + replacements->vals[k] = checkpoints[i]; } - *gb = *gf; + ggml_graph_cpy(gf, gb); // rewrite gb_tmp->nodes[gf->n_nodes:gb_tmp->n_nodes], // replacing references to gb_tmp->nodes[0:gf->n_nodes] ( == gf->nodes[0:gf->n_nodes]), // by recomputing them from checkpoints @@ -14905,21 +14240,21 @@ void ggml_build_backward_gradient_checkpointing( ggml_build_forward_expand(gb, node); } - free_hash_map(replacements); + ggml_hash_map_free(replacements); } // functions to change gradients considering the case that input a might be initial gradient with zero value -static struct ggml_tensor * ggml_add_or_set(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, void * zero_table[]) { - if (hash_contains(zero_table, a)) { +static struct ggml_tensor * ggml_add_or_set(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, struct ggml_hash_set zero_table) { + if (ggml_hash_contains(zero_table, a)) { return b; } else { return ggml_add_impl(ctx, a, b, false); } } -static struct ggml_tensor * ggml_acc_or_set(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, size_t nb1, size_t nb2, size_t nb3, size_t offset, void * zero_table[]) { - if (hash_contains(zero_table, a)) { +static struct ggml_tensor * ggml_acc_or_set(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, size_t nb1, size_t nb2, size_t nb3, size_t offset, struct ggml_hash_set zero_table) { + if (ggml_hash_contains(zero_table, a)) { struct ggml_tensor * a_zero = ggml_scale(ctx, a, ggml_new_f32(ctx, 0)); return ggml_acc_impl(ctx, a_zero, b, nb1, nb2, nb3, offset, false); } else { @@ -14927,23 +14262,23 @@ static struct ggml_tensor * ggml_acc_or_set(struct ggml_context * ctx, struct gg } } -static struct ggml_tensor * ggml_add1_or_set(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, void * zero_table[]) { - if (hash_contains(zero_table, a)) { +static struct ggml_tensor * ggml_add1_or_set(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, struct ggml_hash_set zero_table) { + if (ggml_hash_contains(zero_table, a)) { return ggml_repeat(ctx, b, a); } else { return ggml_add1_impl(ctx, a, b, false); } } -static struct ggml_tensor * ggml_sub_or_set(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, void * zero_table[]) { - if (hash_contains(zero_table, a)) { +static struct ggml_tensor * ggml_sub_or_set(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, struct ggml_hash_set zero_table) { + if (ggml_hash_contains(zero_table, a)) { return ggml_neg(ctx, b); } else { return ggml_sub_impl(ctx, a, b, false); } } -static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor * tensor, void * zero_table[]) { +static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor * tensor, struct ggml_hash_set zero_table) { struct ggml_tensor * src0 = tensor->src[0]; struct ggml_tensor * src1 = tensor->src[1]; @@ -15455,17 +14790,20 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor // necessary for llama if (src0->grad) { //const int n_past = ((int32_t *) tensor->op_params)[0]; - const int n_dims = ((int32_t *) tensor->op_params)[1]; - const int mode = ((int32_t *) tensor->op_params)[2]; - const int n_ctx = ((int32_t *) tensor->op_params)[3]; - float freq_base; - float freq_scale; - float xpos_base; - bool xpos_down; - memcpy(&freq_base, (int32_t *) tensor->op_params + 4, sizeof(float)); - memcpy(&freq_scale, (int32_t *) tensor->op_params + 5, sizeof(float)); - memcpy(&xpos_base, (int32_t *) tensor->op_params + 6, sizeof(float)); - memcpy(&xpos_down, (int32_t *) tensor->op_params + 7, sizeof(bool)); + const int n_dims = ((int32_t *) tensor->op_params)[1]; + const int mode = ((int32_t *) tensor->op_params)[2]; + const int n_ctx = ((int32_t *) tensor->op_params)[3]; + const int n_orig_ctx = ((int32_t *) tensor->op_params)[4]; + float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow, xpos_base, xpos_down; + + memcpy(&freq_base, (int32_t *) tensor->op_params + 5, sizeof(float)); + memcpy(&freq_scale, (int32_t *) tensor->op_params + 6, sizeof(float)); + memcpy(&ext_factor, (int32_t *) tensor->op_params + 7, sizeof(float)); + memcpy(&attn_factor, (int32_t *) tensor->op_params + 8, sizeof(float)); + memcpy(&beta_fast, (int32_t *) tensor->op_params + 9, sizeof(float)); + memcpy(&beta_slow, (int32_t *) tensor->op_params + 10, sizeof(float)); + memcpy(&xpos_base, (int32_t *) tensor->op_params + 11, sizeof(float)); + memcpy(&xpos_down, (int32_t *) tensor->op_params + 12, sizeof(bool)); src0->grad = ggml_add_or_set(ctx, src0->grad, @@ -15475,8 +14813,13 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor n_dims, mode, n_ctx, + n_orig_ctx, freq_base, freq_scale, + ext_factor, + attn_factor, + beta_fast, + beta_slow, xpos_base, xpos_down), zero_table); @@ -15486,17 +14829,20 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor { if (src0->grad) { //const int n_past = ((int32_t *) tensor->op_params)[0]; - const int n_dims = ((int32_t *) tensor->op_params)[1]; - const int mode = ((int32_t *) tensor->op_params)[2]; - const int n_ctx = ((int32_t *) tensor->op_params)[3]; - float freq_base; - float freq_scale; - float xpos_base; - bool xpos_down; - memcpy(&freq_base, (int32_t *) tensor->op_params + 4, sizeof(float)); - memcpy(&freq_scale, (int32_t *) tensor->op_params + 5, sizeof(float)); - memcpy(&xpos_base, (int32_t *) tensor->op_params + 6, sizeof(float)); - memcpy(&xpos_down, (int32_t *) tensor->op_params + 7, sizeof(bool)); + const int n_dims = ((int32_t *) tensor->op_params)[1]; + const int mode = ((int32_t *) tensor->op_params)[2]; + const int n_ctx = ((int32_t *) tensor->op_params)[3]; + const int n_orig_ctx = ((int32_t *) tensor->op_params)[4]; + float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow, xpos_base, xpos_down; + + memcpy(&freq_base, (int32_t *) tensor->op_params + 5, sizeof(float)); + memcpy(&freq_scale, (int32_t *) tensor->op_params + 6, sizeof(float)); + memcpy(&ext_factor, (int32_t *) tensor->op_params + 7, sizeof(float)); + memcpy(&attn_factor, (int32_t *) tensor->op_params + 8, sizeof(float)); + memcpy(&beta_fast, (int32_t *) tensor->op_params + 9, sizeof(float)); + memcpy(&beta_slow, (int32_t *) tensor->op_params + 10, sizeof(float)); + memcpy(&xpos_base, (int32_t *) tensor->op_params + 11, sizeof(float)); + memcpy(&xpos_down, (int32_t *) tensor->op_params + 12, sizeof(bool)); src0->grad = ggml_add_or_set(ctx, src0->grad, @@ -15506,8 +14852,13 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor n_dims, mode, n_ctx, + n_orig_ctx, freq_base, freq_scale, + ext_factor, + attn_factor, + beta_fast, + beta_slow, xpos_base, xpos_down, false), @@ -15522,31 +14873,11 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor { GGML_ASSERT(false); // TODO: not implemented } break; - case GGML_OP_CONV_1D: - { - GGML_ASSERT(false); // TODO: not implemented - } break; - case GGML_OP_CONV_1D_STAGE_0: - { - GGML_ASSERT(false); // TODO: not implemented - } break; - case GGML_OP_CONV_1D_STAGE_1: - { - GGML_ASSERT(false); // TODO: not implemented - } break; case GGML_OP_CONV_TRANSPOSE_1D: { GGML_ASSERT(false); // TODO: not implemented } break; - case GGML_OP_CONV_2D: - { - GGML_ASSERT(false); // TODO: not implemented - } break; - case GGML_OP_CONV_2D_STAGE_0: - { - GGML_ASSERT(false); // TODO: not implemented - } break; - case GGML_OP_CONV_2D_STAGE_1: + case GGML_OP_IM2COL: { GGML_ASSERT(false); // TODO: not implemented } break; @@ -15760,7 +15091,7 @@ static void ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor * } // check if already visited - if (hash_insert(cgraph->visited_hash_table, node)) { + if (ggml_hash_insert(cgraph->visited_hash_table, node) == GGML_HASHTABLE_ALREADY_EXISTS) { return; } @@ -15776,7 +15107,7 @@ static void ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor * if (node->op == GGML_OP_NONE && node->grad == NULL) { // reached a leaf node, not part of the gradient graph (e.g. a constant) - GGML_ASSERT(cgraph->n_leafs < GGML_MAX_NODES); + GGML_ASSERT(cgraph->n_leafs < cgraph->size); if (strlen(node->name) == 0) { ggml_format_name(node, "leaf_%d", cgraph->n_leafs); @@ -15785,22 +15116,24 @@ static void ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor * cgraph->leafs[cgraph->n_leafs] = node; cgraph->n_leafs++; } else { - GGML_ASSERT(cgraph->n_nodes < GGML_MAX_NODES); + GGML_ASSERT(cgraph->n_nodes < cgraph->size); if (strlen(node->name) == 0) { ggml_format_name(node, "node_%d", cgraph->n_nodes); } cgraph->nodes[cgraph->n_nodes] = node; - cgraph->grads[cgraph->n_nodes] = node->grad; + if (cgraph->grads) { + cgraph->grads[cgraph->n_nodes] = node->grad; + } cgraph->n_nodes++; } } static void ggml_build_forward_impl(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor, bool expand) { if (!expand) { - cgraph->n_nodes = 0; - cgraph->n_leafs = 0; + // TODO: this branch isn't accessible anymore, maybe move this to ggml_build_forward_expand + ggml_graph_clear(cgraph); } const int n0 = cgraph->n_nodes; @@ -15821,25 +15154,6 @@ void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * ggml_build_forward_impl(cgraph, tensor, true); } -struct ggml_cgraph ggml_build_forward(struct ggml_tensor * tensor) { - struct ggml_cgraph result = { - /*.n_nodes =*/ 0, - /*.n_leafs =*/ 0, - /*.nodes =*/ { NULL }, - /*.grads =*/ { NULL }, - /*.leafs =*/ { NULL }, - /*.hash_table =*/ { NULL }, - /*.order =*/ GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT, - /*.perf_runs =*/ 0, - /*.perf_cycles =*/ 0, - /*.perf_time_us =*/ 0, - }; - - ggml_build_forward_impl(&result, tensor, false); - - return result; -} - void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep) { GGML_ASSERT(gf->n_nodes > 0); @@ -15856,11 +15170,10 @@ void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * } // remember original gradients which start with zero values - void ** zero_table = malloc(sizeof(void *) * GGML_GRAPH_HASHTABLE_SIZE); - memset(zero_table, 0, sizeof(void*) * GGML_GRAPH_HASHTABLE_SIZE); + struct ggml_hash_set zero_table = ggml_hash_set_new(gf->size); for (int i = 0; i < gf->n_nodes; i++) { if (gf->grads[i]) { - hash_insert(zero_table, gf->grads[i]); + ggml_hash_insert(zero_table, gf->grads[i]); } } @@ -15883,26 +15196,54 @@ void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * } } - free(zero_table); + ggml_hash_set_free(zero_table); } -struct ggml_cgraph ggml_build_backward(struct ggml_context * ctx, struct ggml_cgraph * gf, bool keep) { - struct ggml_cgraph result = *gf; - ggml_build_backward_expand(ctx, gf, &result, keep); - return result; +static size_t ggml_graph_nbytes(size_t size, bool grads) { + size_t nbytes = sizeof(struct ggml_cgraph); + nbytes += size * sizeof(struct ggml_tensor *) * 2; // leafs + nodes + if (grads) { + nbytes += size * sizeof(struct ggml_tensor *); // grads + } + nbytes += ggml_hash_size(size * 2) * sizeof(struct ggml_tensor *); // hash set + return nbytes; } -struct ggml_cgraph * ggml_new_graph(struct ggml_context * ctx) { - struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_GRAPH, GGML_GRAPH_SIZE); +size_t ggml_graph_overhead_custom(size_t size, bool grads) { + return GGML_OBJECT_SIZE + GGML_PAD(ggml_graph_nbytes(size, grads), GGML_MEM_ALIGN); +} + +size_t ggml_graph_overhead(void) { + return ggml_graph_overhead_custom(GGML_DEFAULT_GRAPH_SIZE, false); +} + +struct ggml_cgraph * ggml_new_graph_custom(struct ggml_context * ctx, size_t size, bool grads) { + const size_t obj_size = ggml_graph_nbytes(size, grads); + struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_GRAPH, obj_size); struct ggml_cgraph * cgraph = (struct ggml_cgraph *) ((char *) ctx->mem_buffer + obj->offs); + struct ggml_tensor ** data_start = (struct ggml_tensor **) (cgraph + 1); + + size_t hash_size = ggml_hash_size(size * 2); + struct ggml_tensor ** nodes_ptr = data_start; + struct ggml_tensor ** leafs_ptr = nodes_ptr + size; + struct ggml_tensor ** hash_keys_ptr = leafs_ptr + size; + struct ggml_tensor ** grads_ptr = grads ? hash_keys_ptr + hash_size : NULL; + + // check that we allocated the correct amount of memory + assert(obj_size == (size_t) ( + (grads ? (char *)(grads_ptr + size) : (char *)(hash_keys_ptr + hash_size)) - (char *)cgraph)); + + memset(hash_keys_ptr, 0, hash_size * sizeof(struct ggml_tensor *)); + *cgraph = (struct ggml_cgraph) { + /*.size =*/ size, /*.n_nodes =*/ 0, /*.n_leafs =*/ 0, - /*.nodes =*/ { NULL }, - /*.grads =*/ { NULL }, - /*.leafs =*/ { NULL }, - /*.hash_table =*/ { NULL }, + /*.nodes =*/ nodes_ptr, + /*.grads =*/ grads_ptr, + /*.leafs =*/ leafs_ptr, + /*.hash_table =*/ { hash_size, hash_keys_ptr }, /*.order =*/ GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT, /*.perf_runs =*/ 0, /*.perf_cycles =*/ 0, @@ -15912,14 +15253,85 @@ struct ggml_cgraph * ggml_new_graph(struct ggml_context * ctx) { return cgraph; } -struct ggml_cgraph * ggml_build_forward_ctx(struct ggml_context * ctx, struct ggml_tensor * tensor) { - struct ggml_cgraph * cgraph = ggml_new_graph(ctx); - ggml_build_forward_impl(cgraph, tensor, false); +struct ggml_cgraph * ggml_new_graph(struct ggml_context * ctx) { + return ggml_new_graph_custom(ctx, GGML_DEFAULT_GRAPH_SIZE, false); +} + +struct ggml_cgraph * ggml_graph_view(struct ggml_context * ctx, struct ggml_cgraph * cgraph0, int i0, int i1) { + const size_t obj_size = sizeof(struct ggml_cgraph); + struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_GRAPH, obj_size); + struct ggml_cgraph * cgraph = (struct ggml_cgraph *) ((char *) ctx->mem_buffer + obj->offs); + + *cgraph = (struct ggml_cgraph) { + /*.size =*/ 0, + /*.n_nodes =*/ i1 - i0, + /*.n_leafs =*/ 0, + /*.nodes =*/ cgraph0->nodes + i0, + /*.grads =*/ cgraph0->grads ? cgraph0->grads + i0 : NULL, + /*.leafs =*/ NULL, + /*.hash_table =*/ { 0, NULL }, + /*.order =*/ cgraph0->order, + /*.perf_runs =*/ 0, + /*.perf_cycles =*/ 0, + /*.perf_time_us =*/ 0, + }; + return cgraph; } -size_t ggml_graph_overhead(void) { - return GGML_OBJECT_SIZE + GGML_PAD(GGML_GRAPH_SIZE, GGML_MEM_ALIGN); +void ggml_graph_cpy(struct ggml_cgraph * src, struct ggml_cgraph * dst) { + GGML_ASSERT(dst->size >= src->n_leafs); + GGML_ASSERT(dst->size >= src->n_nodes); + GGML_ASSERT(dst->visited_hash_table.size >= src->visited_hash_table.size); + + dst->n_leafs = src->n_leafs; + dst->n_nodes = src->n_nodes; + dst->order = src->order; + + for (int i = 0; i < src->n_leafs; ++i) { + dst->leafs[i] = src->leafs[i]; + } + + for (int i = 0; i < src->n_nodes; ++i) { + dst->nodes[i] = src->nodes[i]; + } + + if (src->grads) { + GGML_ASSERT(dst->grads != NULL); + for (int i = 0; i < src->n_nodes; ++i) { + dst->grads[i] = src->grads[i]; + } + } + + for (size_t i = 0; i < src->visited_hash_table.size; ++i) { + if (src->visited_hash_table.keys[i]) { + ggml_hash_insert(dst->visited_hash_table, src->visited_hash_table.keys[i]); + } + } +} + +struct ggml_cgraph * ggml_graph_dup(struct ggml_context * ctx, struct ggml_cgraph * cgraph) { + struct ggml_cgraph * result = ggml_new_graph_custom(ctx, cgraph->size, cgraph->grads != NULL); + ggml_graph_cpy(cgraph, result); + return result; +} + +void ggml_graph_reset(struct ggml_cgraph * cgraph) { + GGML_ASSERT(cgraph->grads != NULL); + + for (int i = 0; i < cgraph->n_nodes; i++) { + struct ggml_tensor * grad = cgraph->grads[i]; + + if (grad) { + ggml_set_zero(grad); + } + } +} + +void ggml_graph_clear(struct ggml_cgraph * cgraph) { + cgraph->n_leafs = 0; + cgraph->n_nodes = 0; + memset(cgraph->visited_hash_table.keys, 0, cgraph->visited_hash_table.size * sizeof(struct ggml_tensor *)); } // @@ -16072,13 +15484,233 @@ static void ggml_graph_compute_perf_stats_node(struct ggml_tensor * node, const node->perf_time_us += time_us_cur; } +static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) { + int n_tasks = 0; + + switch (node->op) { + case GGML_OP_CPY: + case GGML_OP_DUP: + case GGML_OP_ADD: + case GGML_OP_ADD1: + case GGML_OP_ACC: + { + n_tasks = n_threads; + } break; + case GGML_OP_SUB: + case GGML_OP_DIV: + case GGML_OP_SQR: + case GGML_OP_SQRT: + case GGML_OP_LOG: + case GGML_OP_SUM: + case GGML_OP_SUM_ROWS: + case GGML_OP_MEAN: + case GGML_OP_ARGMAX: + case GGML_OP_REPEAT: + case GGML_OP_REPEAT_BACK: + { + n_tasks = 1; + } break; + case GGML_OP_UNARY: + switch (ggml_get_unary_op(node)) { + case GGML_UNARY_OP_ABS: + case GGML_UNARY_OP_SGN: + case GGML_UNARY_OP_NEG: + case GGML_UNARY_OP_STEP: + case GGML_UNARY_OP_TANH: + case GGML_UNARY_OP_ELU: + case GGML_UNARY_OP_RELU: + case GGML_UNARY_OP_LEAKY: + { + n_tasks = 1; + } break; + + case GGML_UNARY_OP_GELU: + case GGML_UNARY_OP_GELU_QUICK: + case GGML_UNARY_OP_SILU: + { + n_tasks = n_threads; + } break; + } + break; + case GGML_OP_SILU_BACK: + case GGML_OP_MUL: + case GGML_OP_NORM: + case GGML_OP_RMS_NORM: + case GGML_OP_RMS_NORM_BACK: + case GGML_OP_GROUP_NORM: + case GGML_OP_CONCAT: + { + n_tasks = n_threads; + } break; + case GGML_OP_MUL_MAT: + { + n_tasks = n_threads; + + // TODO: use different scheduling for different matrix sizes + //const int nr0 = ggml_nrows(node->src[0]); + //const int nr1 = ggml_nrows(node->src[1]); + + //n_tasks = MIN(n_threads, MAX(1, nr0/128)); + //printf("nr0 = %8d, nr1 = %8d, nr0*nr1 = %8d, n_tasks%d\n", nr0, nr1, nr0*nr1, n_tasks); + +#if defined(GGML_USE_CUBLAS) + if (ggml_cuda_can_mul_mat(node->src[0], node->src[1], node)) { + n_tasks = 1; // TODO: this actually is doing nothing + // the threads are still spinning + } +#elif defined(GGML_USE_CLBLAST) + if (ggml_cl_can_mul_mat(node->src[0], node->src[1], node)) { + n_tasks = 1; // TODO: this actually is doing nothing + // the threads are still spinning + } +#endif +#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) + if (ggml_compute_forward_mul_mat_use_blas(node->src[0], node->src[1], node)) { + n_tasks = 1; // TODO: this actually is doing nothing + // the threads are still spinning + } +#endif + } break; + case GGML_OP_OUT_PROD: + { + n_tasks = n_threads; + } break; + case GGML_OP_SCALE: + case GGML_OP_SET: + case GGML_OP_CONT: + case GGML_OP_RESHAPE: + case GGML_OP_VIEW: + case GGML_OP_PERMUTE: + case GGML_OP_TRANSPOSE: + case GGML_OP_GET_ROWS: + case GGML_OP_GET_ROWS_BACK: + case GGML_OP_DIAG: + { + n_tasks = 1; + } break; + case GGML_OP_DIAG_MASK_ZERO: + case GGML_OP_DIAG_MASK_INF: + case GGML_OP_SOFT_MAX: + case GGML_OP_SOFT_MAX_BACK: + case GGML_OP_ROPE: + case GGML_OP_ROPE_BACK: + case GGML_OP_ADD_REL_POS: + { + n_tasks = n_threads; + } break; + case GGML_OP_ALIBI: + { + n_tasks = 1; //TODO + } break; + case GGML_OP_CLAMP: + { + n_tasks = 1; //TODO + } break; + case GGML_OP_CONV_TRANSPOSE_1D: + { + n_tasks = n_threads; + } break; + case GGML_OP_IM2COL: + { + n_tasks = n_threads; + } break; + case GGML_OP_CONV_TRANSPOSE_2D: + { + n_tasks = n_threads; + } break; + case GGML_OP_POOL_1D: + case GGML_OP_POOL_2D: + { + n_tasks = 1; + } break; + case GGML_OP_UPSCALE: + { + n_tasks = n_threads; + } break; + case GGML_OP_FLASH_ATTN: + { + n_tasks = n_threads; + } break; + case GGML_OP_FLASH_FF: + { + n_tasks = n_threads; + } break; + case GGML_OP_FLASH_ATTN_BACK: + { + n_tasks = n_threads; + } break; + case GGML_OP_WIN_PART: + case GGML_OP_WIN_UNPART: + case GGML_OP_GET_REL_POS: + case GGML_OP_MAP_UNARY: + case GGML_OP_MAP_BINARY: + case GGML_OP_MAP_CUSTOM1_F32: + case GGML_OP_MAP_CUSTOM2_F32: + case GGML_OP_MAP_CUSTOM3_F32: + { + n_tasks = 1; + } break; + case GGML_OP_MAP_CUSTOM1: + { + struct ggml_map_custom1_op_params * p = (struct ggml_map_custom1_op_params *) node->op_params; + if (p->n_tasks == GGML_N_TASKS_MAX) { + n_tasks = n_threads; + } else { + n_tasks = MIN(p->n_tasks, n_threads); + } + } break; + case GGML_OP_MAP_CUSTOM2: + { + struct ggml_map_custom2_op_params * p = (struct ggml_map_custom2_op_params *) node->op_params; + if (p->n_tasks == GGML_N_TASKS_MAX) { + n_tasks = n_threads; + } else { + n_tasks = MIN(p->n_tasks, n_threads); + } + } break; + case GGML_OP_MAP_CUSTOM3: + { + struct ggml_map_custom3_op_params * p = (struct ggml_map_custom3_op_params *) node->op_params; + if (p->n_tasks == GGML_N_TASKS_MAX) { + n_tasks = n_threads; + } else { + n_tasks = MIN(p->n_tasks, n_threads); + } + } break; + case GGML_OP_CROSS_ENTROPY_LOSS: + { + n_tasks = n_threads; + } break; + case GGML_OP_CROSS_ENTROPY_LOSS_BACK: + { + n_tasks = n_threads; + } break; + case GGML_OP_NONE: + { + n_tasks = 1; + } break; + case GGML_OP_COUNT: + { + GGML_ASSERT(false); + } break; + default: + { + printf("%s: op %s not implemented\n", __func__, ggml_op_name(node->op)); + GGML_ASSERT(false); + } break; + } + + assert(n_tasks > 0); + + return n_tasks; +} + static thread_ret_t ggml_graph_compute_thread(void * data) { struct ggml_compute_state * state = (struct ggml_compute_state *) data; const struct ggml_cgraph * cgraph = state->shared->cgraph; const struct ggml_cplan * cplan = state->shared->cplan; - const int * n_tasks_arr = cplan->n_tasks; const int n_threads = state->shared->n_threads; set_numa_thread_affinity(state->ith, n_threads); @@ -16103,9 +15735,9 @@ static thread_ret_t ggml_graph_compute_thread(void * data) { if (node_n != -1) { /* FINALIZE */ - struct ggml_tensor * node = state->shared->cgraph->nodes[node_n]; + struct ggml_tensor * node = cgraph->nodes[node_n]; if (GGML_OP_HAS_FINALIZE[node->op]) { - params.nth = n_tasks_arr[node_n]; + params.nth = ggml_get_n_tasks(node, n_threads); ggml_compute_forward(¶ms, node); } ggml_graph_compute_perf_stats_node(node, state->shared); @@ -16116,7 +15748,7 @@ static thread_ret_t ggml_graph_compute_thread(void * data) { GGML_PRINT_DEBUG_5("%s: %d/%d\n", __func__, node_n, cgraph->n_nodes); struct ggml_tensor * node = cgraph->nodes[node_n]; - const int n_tasks = n_tasks_arr[node_n]; + const int n_tasks = ggml_get_n_tasks(node, n_threads); state->shared->perf_node_start_cycles = ggml_perf_cycles(); state->shared->perf_node_start_time_us = ggml_perf_time_us(); @@ -16174,7 +15806,7 @@ static thread_ret_t ggml_graph_compute_thread(void * data) { /* COMPUTE */ struct ggml_tensor * node = cgraph->nodes[node_n]; - const int n_tasks = n_tasks_arr[node_n]; + const int n_tasks = ggml_get_n_tasks(node, n_threads); struct ggml_compute_params params = { /*.type =*/ GGML_TASK_COMPUTE, @@ -16208,121 +15840,46 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) { struct ggml_tensor * node = cgraph->nodes[i]; + size_t cur = 0; + switch (node->op) { case GGML_OP_CPY: case GGML_OP_DUP: { n_tasks = n_threads; - size_t cur = 0; if (ggml_is_quantized(node->type)) { cur = ggml_type_size(GGML_TYPE_F32) * node->ne[0] * n_tasks; } - - work_size = MAX(work_size, cur); } break; case GGML_OP_ADD: case GGML_OP_ADD1: { n_tasks = n_threads; - size_t cur = 0; - if (ggml_is_quantized(node->src[0]->type)) { cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks; } - - work_size = MAX(work_size, cur); } break; case GGML_OP_ACC: { n_tasks = n_threads; - size_t cur = 0; - if (ggml_is_quantized(node->src[0]->type)) { cur = ggml_type_size(GGML_TYPE_F32) * node->src[1]->ne[0] * n_tasks; } - - work_size = MAX(work_size, cur); } break; - case GGML_OP_SUB: - case GGML_OP_DIV: - case GGML_OP_SQR: - case GGML_OP_SQRT: - case GGML_OP_LOG: - case GGML_OP_SUM: - case GGML_OP_SUM_ROWS: - case GGML_OP_MEAN: - case GGML_OP_ARGMAX: - case GGML_OP_REPEAT: - case GGML_OP_REPEAT_BACK: - { - n_tasks = 1; - } break; - - case GGML_OP_UNARY: - { - switch (ggml_get_unary_op(node)) { - case GGML_UNARY_OP_ABS: - case GGML_UNARY_OP_SGN: - case GGML_UNARY_OP_NEG: - case GGML_UNARY_OP_STEP: - case GGML_UNARY_OP_TANH: - case GGML_UNARY_OP_ELU: - case GGML_UNARY_OP_RELU: - { - n_tasks = 1; - } break; - - case GGML_UNARY_OP_GELU: - case GGML_UNARY_OP_GELU_QUICK: - case GGML_UNARY_OP_SILU: - { - n_tasks = n_threads; - } break; - } - } break; - case GGML_OP_SILU_BACK: - case GGML_OP_MUL: - case GGML_OP_NORM: - case GGML_OP_RMS_NORM: - case GGML_OP_RMS_NORM_BACK: - case GGML_OP_GROUP_NORM: - { - n_tasks = n_threads; - } break; - case GGML_OP_CONCAT: case GGML_OP_MUL_MAT: { - n_tasks = n_threads; - - // TODO: use different scheduling for different matrix sizes - //const int nr0 = ggml_nrows(node->src[0]); - //const int nr1 = ggml_nrows(node->src[1]); - - //n_tasks = MIN(n_threads, MAX(1, nr0/128)); - //printf("nr0 = %8d, nr1 = %8d, nr0*nr1 = %8d, n_tasks%d\n", nr0, nr1, nr0*nr1, n_tasks); - - size_t cur = 0; const enum ggml_type vec_dot_type = type_traits[node->src[0]->type].vec_dot_type; -#if defined(GGML_USE_CUBLAS) - if (ggml_cuda_can_mul_mat(node->src[0], node->src[1], node)) { - n_tasks = 1; // TODO: this actually is doing nothing - // the threads are still spinning - } else -#elif defined(GGML_USE_CLBLAST) +#if defined(GGML_USE_CLBLAST) if (ggml_cl_can_mul_mat(node->src[0], node->src[1], node)) { - n_tasks = 1; // TODO: this actually is doing nothing - // the threads are still spinning cur = ggml_cl_mul_mat_get_wsize(node->src[0], node->src[1], node); } else #endif #if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) if (ggml_compute_forward_mul_mat_use_blas(node->src[0], node->src[1], node)) { - n_tasks = 1; // TODO: this actually is doing nothing - // the threads are still spinning if (node->src[0]->type != GGML_TYPE_F32) { // here we need memory just for single 2D matrix from src0 cur = ggml_type_size(GGML_TYPE_F32)*(node->src[0]->ne[0]*node->src[0]->ne[1]); @@ -16331,108 +15888,18 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) { #endif if (node->src[1]->type != vec_dot_type) { cur = ggml_type_size(vec_dot_type)*ggml_nelements(node->src[1])/ggml_blck_size(vec_dot_type); - } else { - cur = 0; } - - work_size = MAX(work_size, cur); } break; case GGML_OP_OUT_PROD: { n_tasks = n_threads; - size_t cur = 0; - if (ggml_is_quantized(node->src[0]->type)) { cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks; } - - work_size = MAX(work_size, cur); - } break; - case GGML_OP_SCALE: - { - n_tasks = 1; - } break; - case GGML_OP_SET: - case GGML_OP_CONT: - case GGML_OP_RESHAPE: - case GGML_OP_VIEW: - case GGML_OP_PERMUTE: - case GGML_OP_TRANSPOSE: - case GGML_OP_GET_ROWS: - case GGML_OP_GET_ROWS_BACK: - case GGML_OP_DIAG: - { - n_tasks = 1; - } break; - case GGML_OP_DIAG_MASK_ZERO: - case GGML_OP_DIAG_MASK_INF: - case GGML_OP_SOFT_MAX: - case GGML_OP_SOFT_MAX_BACK: - case GGML_OP_ROPE: - case GGML_OP_ROPE_BACK: - case GGML_OP_ADD_REL_POS: - { - n_tasks = n_threads; - } break; - case GGML_OP_ALIBI: - { - n_tasks = 1; //TODO - } break; - case GGML_OP_CLAMP: - { - n_tasks = 1; //TODO - } break; - case GGML_OP_CONV_1D: - { - n_tasks = n_threads; - - GGML_ASSERT(node->src[0]->ne[3] == 1); - GGML_ASSERT(node->src[1]->ne[2] == 1); - GGML_ASSERT(node->src[1]->ne[3] == 1); - - const int64_t ne00 = node->src[0]->ne[0]; - const int64_t ne01 = node->src[0]->ne[1]; - const int64_t ne02 = node->src[0]->ne[2]; - - const int64_t ne10 = node->src[1]->ne[0]; - const int64_t ne11 = node->src[1]->ne[1]; - - const int64_t ne0 = node->ne[0]; - const int64_t ne1 = node->ne[1]; - const int64_t nk = ne00; - const int64_t ew0 = nk * ne01; - - UNUSED(ne02); - UNUSED(ne10); - UNUSED(ne11); - - size_t cur = 0; - - if (node->src[0]->type == GGML_TYPE_F16 && - node->src[1]->type == GGML_TYPE_F32) { - cur = sizeof(ggml_fp16_t)*(ne0*ne1*ew0); - } else if (node->src[0]->type == GGML_TYPE_F32 && - node->src[1]->type == GGML_TYPE_F32) { - cur = sizeof(float)*(ne0*ne1*ew0); - } else { - GGML_ASSERT(false); - } - - work_size = MAX(work_size, cur); - } break; - case GGML_OP_CONV_1D_STAGE_0: - { - n_tasks = n_threads; - } break; - case GGML_OP_CONV_1D_STAGE_1: - { - n_tasks = n_threads; } break; case GGML_OP_CONV_TRANSPOSE_1D: { - n_tasks = n_threads; - GGML_ASSERT(node->src[0]->ne[3] == 1); GGML_ASSERT(node->src[1]->ne[2] == 1); GGML_ASSERT(node->src[1]->ne[3] == 1); @@ -16444,7 +15911,6 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) { const int64_t ne10 = node->src[1]->ne[0]; // L const int64_t ne11 = node->src[1]->ne[1]; // Cin - size_t cur = 0; if (node->src[0]->type == GGML_TYPE_F16 && node->src[1]->type == GGML_TYPE_F32) { cur += sizeof(ggml_fp16_t)*ne00*ne01*ne02; @@ -16456,59 +15922,13 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) { } else { GGML_ASSERT(false); } - - work_size = MAX(work_size, cur); } break; - case GGML_OP_CONV_2D: - { - n_tasks = n_threads; - - const int64_t ne00 = node->src[0]->ne[0]; // W - const int64_t ne01 = node->src[0]->ne[1]; // H - const int64_t ne02 = node->src[0]->ne[2]; // C - const int64_t ne03 = node->src[0]->ne[3]; // N - - const int64_t ne10 = node->src[1]->ne[0]; // W - const int64_t ne11 = node->src[1]->ne[1]; // H - const int64_t ne12 = node->src[1]->ne[2]; // C - - const int64_t ne0 = node->ne[0]; - const int64_t ne1 = node->ne[1]; - const int64_t ne2 = node->ne[2]; - const int64_t ne3 = node->ne[3]; - const int64_t nk = ne00*ne01; - const int64_t ew0 = nk * ne02; - - UNUSED(ne03); - UNUSED(ne2); - - size_t cur = 0; - - if (node->src[0]->type == GGML_TYPE_F16 && - node->src[1]->type == GGML_TYPE_F32) { - // im2col: [N*OH*OW, IC*KH*KW] - cur = sizeof(ggml_fp16_t)*(ne3*ne0*ne1*ew0); - } else if (node->src[0]->type == GGML_TYPE_F32 && - node->src[1]->type == GGML_TYPE_F32) { - cur = sizeof(float)* (ne10*ne11*ne12); - } else { - GGML_ASSERT(false); - } - - work_size = MAX(work_size, cur); - } break; - case GGML_OP_CONV_2D_STAGE_0: - { - n_tasks = n_threads; - } break; - case GGML_OP_CONV_2D_STAGE_1: + case GGML_OP_IM2COL: { n_tasks = n_threads; } break; case GGML_OP_CONV_TRANSPOSE_2D: { - n_tasks = n_threads; - const int64_t ne00 = node->src[0]->ne[0]; // W const int64_t ne01 = node->src[0]->ne[1]; // H const int64_t ne02 = node->src[0]->ne[2]; // Channels Out @@ -16518,141 +15938,66 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) { const int64_t ne11 = node->src[1]->ne[1]; // H const int64_t ne12 = node->src[1]->ne[2]; // Channels In - size_t cur = 0; cur += sizeof(ggml_fp16_t)*ne00*ne01*ne02*ne03; cur += sizeof(ggml_fp16_t)*ne10*ne11*ne12; - - work_size = MAX(work_size, cur); - } break; - case GGML_OP_POOL_1D: - case GGML_OP_POOL_2D: - { - n_tasks = 1; - } break; - case GGML_OP_UPSCALE: - { - n_tasks = n_threads; } break; case GGML_OP_FLASH_ATTN: { n_tasks = n_threads; - size_t cur = 0; - const int64_t ne11 = ggml_up(node->src[1]->ne[1], GGML_SOFT_MAX_UNROLL); if (node->src[1]->type == GGML_TYPE_F32) { cur = sizeof(float)*ne11*n_tasks; // TODO: this can become (n_tasks-1) cur += sizeof(float)*ne11*n_tasks; // this is overestimated by x2 - } - - if (node->src[1]->type == GGML_TYPE_F16) { + } else if (node->src[1]->type == GGML_TYPE_F16) { cur = sizeof(float)*ne11*n_tasks; // TODO: this can become (n_tasks-1) cur += sizeof(float)*ne11*n_tasks; // this is overestimated by x2 } - - work_size = MAX(work_size, cur); } break; case GGML_OP_FLASH_FF: { n_tasks = n_threads; - size_t cur = 0; - if (node->src[1]->type == GGML_TYPE_F32) { cur = sizeof(float)*node->src[1]->ne[1]*n_tasks; // TODO: this can become (n_tasks-1) cur += sizeof(float)*node->src[1]->ne[1]*n_tasks; // this is overestimated by x2 - } - - if (node->src[1]->type == GGML_TYPE_F16) { + } else if (node->src[1]->type == GGML_TYPE_F16) { cur = sizeof(float)*node->src[1]->ne[1]*n_tasks; // TODO: this can become (n_tasks-1) cur += sizeof(float)*node->src[1]->ne[1]*n_tasks; // this is overestimated by x2 } - - work_size = MAX(work_size, cur); } break; case GGML_OP_FLASH_ATTN_BACK: { n_tasks = n_threads; - size_t cur = 0; - const int64_t D = node->src[0]->ne[0]; const int64_t ne11 = ggml_up(node->src[1]->ne[1], GGML_SOFT_MAX_UNROLL); const int64_t mxDn = MAX(D, ne11) * 2; // *2 because of S and SM in ggml_compute_forward_flash_attn_back if (node->src[1]->type == GGML_TYPE_F32) { cur = sizeof(float)*mxDn*n_tasks; // TODO: this can become (n_tasks-1) cur += sizeof(float)*mxDn*n_tasks; // this is overestimated by x2 - } - - if (node->src[1]->type == GGML_TYPE_F16) { + } else if (node->src[1]->type == GGML_TYPE_F16) { cur = sizeof(float)*mxDn*n_tasks; // TODO: this can become (n_tasks-1) cur += sizeof(float)*mxDn*n_tasks; // this is overestimated by x2 } + } break; - work_size = MAX(work_size, cur); - } break; - case GGML_OP_WIN_PART: - case GGML_OP_WIN_UNPART: - case GGML_OP_GET_REL_POS: - case GGML_OP_MAP_UNARY: - case GGML_OP_MAP_BINARY: - case GGML_OP_MAP_CUSTOM1_F32: - case GGML_OP_MAP_CUSTOM2_F32: - case GGML_OP_MAP_CUSTOM3_F32: - { - n_tasks = 1; - } break; - case GGML_OP_MAP_CUSTOM1: - { - struct ggml_map_custom1_op_params * p = (struct ggml_map_custom1_op_params *) node->op_params; - if (p->n_tasks == GGML_N_TASKS_MAX) { - n_tasks = n_threads; - } else { - n_tasks = MIN(p->n_tasks, n_threads); - } - } break; - case GGML_OP_MAP_CUSTOM2: - { - struct ggml_map_custom2_op_params * p = (struct ggml_map_custom2_op_params *) node->op_params; - if (p->n_tasks == GGML_N_TASKS_MAX) { - n_tasks = n_threads; - } else { - n_tasks = MIN(p->n_tasks, n_threads); - } - } break; - case GGML_OP_MAP_CUSTOM3: - { - struct ggml_map_custom3_op_params * p = (struct ggml_map_custom3_op_params *) node->op_params; - if (p->n_tasks == GGML_N_TASKS_MAX) { - n_tasks = n_threads; - } else { - n_tasks = MIN(p->n_tasks, n_threads); - } - } break; case GGML_OP_CROSS_ENTROPY_LOSS: { n_tasks = n_threads; - size_t cur = ggml_type_size(node->type)*(n_tasks + node->src[0]->ne[0]*n_tasks); - - work_size = MAX(work_size, cur); - } break; - case GGML_OP_CROSS_ENTROPY_LOSS_BACK: - { - n_tasks = n_threads; - } break; - case GGML_OP_NONE: - { - n_tasks = 1; + cur = ggml_type_size(node->type)*(n_tasks + node->src[0]->ne[0]*n_tasks); } break; case GGML_OP_COUNT: { GGML_ASSERT(false); } break; + default: + break; } - cplan.n_tasks[i] = n_tasks; + work_size = MAX(work_size, cur); } if (work_size > 0) { @@ -16674,12 +16019,6 @@ int ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan) { if (cplan->work_size > 0) { GGML_ASSERT(cplan->work_data); } - - for (int i = 0; i < cgraph->n_nodes; ++i) { - if (cgraph->nodes[i]->op != GGML_OP_NONE) { - GGML_ASSERT(cplan->n_tasks[i] > 0); - } - } } const int n_threads = cplan->n_threads; @@ -16752,16 +16091,6 @@ int ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan) { return compute_status; } -void ggml_graph_reset(struct ggml_cgraph * cgraph) { - for (int i = 0; i < cgraph->n_nodes; i++) { - struct ggml_tensor * grad = cgraph->grads[i]; - - if (grad) { - ggml_set_zero(grad); - } - } -} - void ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads) { struct ggml_cplan cplan = ggml_graph_plan(cgraph, n_threads); @@ -16888,12 +16217,12 @@ void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname) { const uint32_t magic = GGML_FILE_MAGIC; const uint32_t version = GGML_FILE_VERSION; const uint32_t n_leafs = cgraph->n_leafs; - const uint32_t nodes = cgraph->n_nodes; + const uint32_t n_nodes = cgraph->n_nodes; fwrite(&magic, sizeof(uint32_t), 1, fout); fwrite(&version, sizeof(uint32_t), 1, fout); fwrite(&n_leafs, sizeof(uint32_t), 1, fout); - fwrite(&nodes, sizeof(uint32_t), 1, fout); + fwrite(&n_nodes, sizeof(uint32_t), 1, fout); fwrite(&size_eval, sizeof(uint64_t), 1, fout); } @@ -16981,7 +16310,7 @@ void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname) { if (idx == -1) { for (int k = 0; k < cgraph->n_nodes; ++k) { if (args[j] == cgraph->nodes[k]) { - idx = GGML_MAX_NODES + k; + idx = cgraph->n_leafs + k; break; } } @@ -17008,11 +16337,11 @@ void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname) { } } -struct ggml_cgraph ggml_graph_import(const char * fname, struct ggml_context ** ctx_data, struct ggml_context ** ctx_eval) { +struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context ** ctx_data, struct ggml_context ** ctx_eval) { assert(*ctx_data == NULL); assert(*ctx_eval == NULL); - struct ggml_cgraph result = { 0 }; + struct ggml_cgraph * result = NULL; struct ggml_tensor * data = NULL; @@ -17084,13 +16413,11 @@ struct ggml_cgraph ggml_graph_import(const char * fname, struct ggml_context ** const uint32_t n_leafs = *(const uint32_t *) ptr; ptr += sizeof(n_leafs); const uint32_t n_nodes = *(const uint32_t *) ptr; ptr += sizeof(n_nodes); const uint64_t size_eval = *(const uint64_t *) ptr; ptr += sizeof(size_eval); - - result.n_leafs = n_leafs; - result.n_nodes = n_nodes; + const int graph_size = MAX(n_leafs, n_nodes); // create the data context { - const size_t overhead = (n_leafs + n_nodes)*ggml_tensor_overhead(); + const size_t overhead = (n_leafs + n_nodes)*ggml_tensor_overhead() + ggml_graph_overhead_custom(graph_size, false); struct ggml_init_params params = { .mem_size = size_eval + overhead, @@ -17106,6 +16433,12 @@ struct ggml_cgraph ggml_graph_import(const char * fname, struct ggml_context ** } } + result = ggml_new_graph_custom(*ctx_eval, graph_size, false); + + result->n_leafs = n_leafs; + result->n_nodes = n_nodes; + + // leafs { uint32_t type; @@ -17144,7 +16477,7 @@ struct ggml_cgraph ggml_graph_import(const char * fname, struct ggml_context ** tensor->nb[j] = nb[j]; } - result.leafs[i] = tensor; + result->leafs[i] = tensor; ptr += ggml_nbytes(tensor); @@ -17196,10 +16529,10 @@ struct ggml_cgraph ggml_graph_import(const char * fname, struct ggml_context ** continue; } - if (arg_idx < GGML_MAX_NODES) { - args[j] = result.leafs[arg_idx]; + if (arg_idx < result->n_leafs) { + args[j] = result->leafs[arg_idx]; } else { - args[j] = result.nodes[arg_idx - GGML_MAX_NODES]; + args[j] = result->nodes[arg_idx - result->n_leafs]; } } @@ -17251,7 +16584,7 @@ struct ggml_cgraph ggml_graph_import(const char * fname, struct ggml_context ** tensor->src[j] = args[j]; } - result.nodes[i] = tensor; + result->nodes[i] = tensor; fprintf(stderr, "%s: loaded node %d: '%16s', %3d dims, %9zu bytes\n", __func__, i, tensor->name, n_dims, ggml_nbytes(tensor)); } @@ -18156,10 +17489,11 @@ struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type) { case GGML_OPT_ADAM: { result = (struct ggml_opt_params) { - .type = GGML_OPT_ADAM, - .n_threads = 1, - .past = 0, - .delta = 1e-5f, + .type = GGML_OPT_ADAM, + .graph_size = GGML_DEFAULT_GRAPH_SIZE, + .n_threads = 1, // FIXME: GGML_DEFAULT_N_THREADS ? + .past = 0, + .delta = 1e-5f, .max_no_improvement = 100, @@ -18186,10 +17520,11 @@ struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type) { case GGML_OPT_LBFGS: { result = (struct ggml_opt_params) { - .type = GGML_OPT_LBFGS, - .n_threads = 1, - .past = 0, - .delta = 1e-5f, + .type = GGML_OPT_LBFGS, + .graph_size = GGML_DEFAULT_GRAPH_SIZE, + .n_threads = 1, + .past = 0, + .delta = 1e-5f, .max_no_improvement = 0, @@ -18331,14 +17666,11 @@ enum ggml_opt_result ggml_opt_resume( struct ggml_tensor * f) { // build forward + backward compute graphs - struct ggml_tensor * gfbuf = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, sizeof(struct ggml_cgraph) / ggml_type_size(GGML_TYPE_I32)+ (sizeof(struct ggml_cgraph) % ggml_type_size(GGML_TYPE_I32) ? 1 : 0)); - struct ggml_tensor * gbbuf = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, sizeof(struct ggml_cgraph) / ggml_type_size(GGML_TYPE_I32)+ (sizeof(struct ggml_cgraph) % ggml_type_size(GGML_TYPE_I32) ? 1 : 0)); + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx, opt->params.graph_size, true); + ggml_build_forward_expand(gf, f); - struct ggml_cgraph * gf = (struct ggml_cgraph *) gfbuf->data; - struct ggml_cgraph * gb = (struct ggml_cgraph *) gbbuf->data; - - *gf = ggml_build_forward (f); - *gb = ggml_build_backward(ctx, gf, true); + struct ggml_cgraph * gb = ggml_graph_dup(ctx, gf); + ggml_build_backward_expand(ctx, gf, gb, true); return ggml_opt_resume_g(ctx, opt, f, gf, gb, NULL, NULL); } @@ -18702,8 +18034,7 @@ static bool gguf_fread_el(FILE * file, void * dst, size_t size, size_t * offset) return n == size; } -// NOTE: temporary handling of GGUFv1 >> remove after Oct 2023 -static bool gguf_fread_str_cur(FILE * file, struct gguf_str * p, size_t * offset) { +static bool gguf_fread_str(FILE * file, struct gguf_str * p, size_t * offset) { p->n = 0; p->data = NULL; @@ -18715,19 +18046,6 @@ static bool gguf_fread_str_cur(FILE * file, struct gguf_str * p, size_t * offset return ok; } -static bool gguf_fread_str_v1(FILE * file, struct gguf_str * p, size_t * offset) { - p->n = 0; - p->data = NULL; - - bool ok = true; - - uint32_t n = 0; - ok = ok && gguf_fread_el(file, &n, sizeof(n), offset); p->data = calloc(n + 1, 1); p->n = n; - ok = ok && gguf_fread_el(file, p->data, p->n, offset); - - return ok; -} - struct gguf_context * gguf_init_empty(void) { struct gguf_context * ctx = GGML_ALIGNED_MALLOC(sizeof(struct gguf_context)); @@ -18786,20 +18104,14 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p ctx->data = NULL; ok = ok && gguf_fread_el(file, &ctx->header.version, sizeof(ctx->header.version), &offset); + ok = ok && gguf_fread_el(file, &ctx->header.n_tensors, sizeof(ctx->header.n_tensors), &offset); + ok = ok && gguf_fread_el(file, &ctx->header.n_kv, sizeof(ctx->header.n_kv), &offset); if (ctx->header.version == 1) { - // NOTE: temporary handling of GGUFv1 >> remove after Oct 2023 - uint32_t n_tensors = 0; - uint32_t n_kv = 0; - - ok = ok && gguf_fread_el(file, &n_tensors, sizeof(n_tensors), &offset); - ok = ok && gguf_fread_el(file, &n_kv, sizeof(n_kv), &offset); - - ctx->header.n_tensors = n_tensors; - ctx->header.n_kv = n_kv; - } else { - ok = ok && gguf_fread_el(file, &ctx->header.n_tensors, sizeof(ctx->header.n_tensors), &offset); - ok = ok && gguf_fread_el(file, &ctx->header.n_kv, sizeof(ctx->header.n_kv), &offset); + fprintf(stderr, "%s: GGUFv1 is no longer supported. please use a more up-to-date version\n", __func__); + fclose(file); + gguf_free(ctx); + return NULL; } if (!ok) { @@ -18810,17 +18122,11 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p } } - // NOTE: temporary handling of GGUFv1 >> remove after Oct 2023 - bool (* gguf_fread_str)(FILE *, struct gguf_str *, size_t *) = gguf_fread_str_cur; - if (ctx->header.version == 1) { - gguf_fread_str = gguf_fread_str_v1; - } - // read the kv pairs { ctx->kv = malloc(ctx->header.n_kv * sizeof(struct gguf_kv)); - for (uint32_t i = 0; i < ctx->header.n_kv; ++i) { + for (uint64_t i = 0; i < ctx->header.n_kv; ++i) { struct gguf_kv * kv = &ctx->kv[i]; //fprintf(stderr, "%s: reading kv %d\n", __func__, i); @@ -18846,15 +18152,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p case GGUF_TYPE_ARRAY: { ok = ok && gguf_fread_el(file, &kv->value.arr.type, sizeof(kv->value.arr.type), &offset); - - if (ctx->header.version == 1) { - // NOTE: temporary handling of GGUFv1 >> remove after Oct 2023 - uint32_t n = 0; - ok = ok && gguf_fread_el(file, &n, sizeof(n), &offset); - kv->value.arr.n = n; - } else { - ok = ok && gguf_fread_el(file, &kv->value.arr.n, sizeof(kv->value.arr.n), &offset); - } + ok = ok && gguf_fread_el(file, &kv->value.arr.n, sizeof(kv->value.arr.n), &offset); switch (kv->value.arr.type) { case GGUF_TYPE_UINT8: @@ -18875,7 +18173,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p case GGUF_TYPE_STRING: { kv->value.arr.data = malloc(kv->value.arr.n * sizeof(struct gguf_str)); - for (uint32_t j = 0; j < kv->value.arr.n; ++j) { + for (uint64_t j = 0; j < kv->value.arr.n; ++j) { ok = ok && gguf_fread_str(file, &((struct gguf_str *) kv->value.arr.data)[j], &offset); } } break; @@ -18903,7 +18201,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p { ctx->infos = malloc(ctx->header.n_tensors * sizeof(struct gguf_tensor_info)); - for (uint32_t i = 0; i < ctx->header.n_tensors; ++i) { + for (uint64_t i = 0; i < ctx->header.n_tensors; ++i) { struct gguf_tensor_info * info = &ctx->infos[i]; for (int j = 0; j < GGML_MAX_DIMS; ++j) { @@ -18913,14 +18211,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p ok = ok && gguf_fread_str(file, &info->name, &offset); ok = ok && gguf_fread_el (file, &info->n_dims, sizeof(info->n_dims), &offset); for (uint32_t j = 0; j < info->n_dims; ++j) { - if (ctx->header.version == 1) { - // NOTE: temporary handling of GGUFv1 >> remove after Oct 2023 - uint32_t t = 0; - ok = ok && gguf_fread_el(file, &t, sizeof(t), &offset); - info->ne[j] = t; - } else { - ok = ok && gguf_fread_el(file, &info->ne[j], sizeof(info->ne[j]), &offset); - } + ok = ok && gguf_fread_el(file, &info->ne[j], sizeof(info->ne[j]), &offset); } ok = ok && gguf_fread_el (file, &info->type, sizeof(info->type), &offset); ok = ok && gguf_fread_el (file, &info->offset, sizeof(info->offset), &offset); @@ -18957,7 +18248,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p // compute the total size of the data section, taking into account the alignment { ctx->size = 0; - for (uint32_t i = 0; i < ctx->header.n_tensors; ++i) { + for (uint64_t i = 0; i < ctx->header.n_tensors; ++i) { struct gguf_tensor_info * info = &ctx->infos[i]; const int64_t ne = @@ -19026,7 +18317,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p ggml_set_no_alloc(ctx_data, true); // create the tensors - for (uint32_t i = 0; i < ctx->header.n_tensors; ++i) { + for (uint64_t i = 0; i < ctx->header.n_tensors; ++i) { const int64_t ne[GGML_MAX_DIMS] = { ctx->infos[i].ne[0], ctx->infos[i].ne[1], @@ -19161,24 +18452,29 @@ int gguf_find_key(const struct gguf_context * ctx, const char * key) { } const char * gguf_get_key(const struct gguf_context * ctx, int key_id) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); return ctx->kv[key_id].key.data; } enum gguf_type gguf_get_kv_type(const struct gguf_context * ctx, int key_id) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); return ctx->kv[key_id].type; } enum gguf_type gguf_get_arr_type(const struct gguf_context * ctx, int key_id) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_ARRAY); return ctx->kv[key_id].value.arr.type; } const void * gguf_get_arr_data(const struct gguf_context * ctx, int key_id) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_ARRAY); return ctx->kv[key_id].value.arr.data; } const char * gguf_get_arr_str(const struct gguf_context * ctx, int key_id, int i) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_ARRAY); struct gguf_kv * kv = &ctx->kv[key_id]; struct gguf_str * str = &((struct gguf_str *) kv->value.arr.data)[i]; @@ -19186,70 +18482,90 @@ const char * gguf_get_arr_str(const struct gguf_context * ctx, int key_id, int i } int gguf_get_arr_n(const struct gguf_context * ctx, int key_id) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_ARRAY); return ctx->kv[key_id].value.arr.n; } uint8_t gguf_get_val_u8(const struct gguf_context * ctx, int key_id) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_UINT8); return ctx->kv[key_id].value.uint8; } int8_t gguf_get_val_i8(const struct gguf_context * ctx, int key_id) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_INT8); return ctx->kv[key_id].value.int8; } uint16_t gguf_get_val_u16(const struct gguf_context * ctx, int key_id) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_UINT16); return ctx->kv[key_id].value.uint16; } int16_t gguf_get_val_i16(const struct gguf_context * ctx, int key_id) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_INT16); return ctx->kv[key_id].value.int16; } uint32_t gguf_get_val_u32(const struct gguf_context * ctx, int key_id) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_UINT32); return ctx->kv[key_id].value.uint32; } int32_t gguf_get_val_i32(const struct gguf_context * ctx, int key_id) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_INT32); return ctx->kv[key_id].value.int32; } float gguf_get_val_f32(const struct gguf_context * ctx, int key_id) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_FLOAT32); return ctx->kv[key_id].value.float32; } uint64_t gguf_get_val_u64(const struct gguf_context * ctx, int key_id) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_UINT64); return ctx->kv[key_id].value.uint64; } int64_t gguf_get_val_i64(const struct gguf_context * ctx, int key_id) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_INT64); return ctx->kv[key_id].value.int64; } double gguf_get_val_f64(const struct gguf_context * ctx, int key_id) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_FLOAT64); return ctx->kv[key_id].value.float64; } bool gguf_get_val_bool(const struct gguf_context * ctx, int key_id) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_BOOL); return ctx->kv[key_id].value.bool_; } const char * gguf_get_val_str(const struct gguf_context * ctx, int key_id) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_STRING); return ctx->kv[key_id].value.str.data; } +const void * gguf_get_val_data(const struct gguf_context * ctx, int key_id) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); + GGML_ASSERT(ctx->kv[key_id].type != GGUF_TYPE_ARRAY); + GGML_ASSERT(ctx->kv[key_id].type != GGUF_TYPE_STRING); + return &ctx->kv[key_id].value; +} + int gguf_get_n_tensors(const struct gguf_context * ctx) { return ctx->header.n_tensors; } diff --git a/native/jni/src/ggml/ggml.h b/native/jni/src/ggml/ggml.h index 9d16c5a72..33cc3cab1 100644 --- a/native/jni/src/ggml/ggml.h +++ b/native/jni/src/ggml/ggml.h @@ -58,7 +58,8 @@ // { // ... // -// struct ggml_cgraph gf = ggml_build_forward(f); +// struct ggml_cgraph * gf = ggml_new_graph(ctx); +// ggml_build_forward_expand(gf, f); // // // set the input variable and parameter values // ggml_set_f32(x, 2.0f); @@ -213,19 +214,18 @@ #define GGML_QNT_VERSION 2 // bump this on quantization format changes #define GGML_QNT_VERSION_FACTOR 1000 // do not change this -#define GGML_MAX_DIMS 4 -#define GGML_MAX_NODES 16384 -#define GGML_MAX_PARAMS 1024 -#define GGML_MAX_CONTEXTS 64 -#define GGML_MAX_SRC 6 -#define GGML_MAX_NAME 64 -#define GGML_MAX_OP_PARAMS 32 -#define GGML_DEFAULT_N_THREADS 4 - +#define GGML_MAX_DIMS 4 +#define GGML_MAX_PARAMS 1024 +#define GGML_MAX_CONTEXTS 64 +#define GGML_MAX_SRC 6 +#define GGML_MAX_NAME 64 +#define GGML_MAX_OP_PARAMS 64 +#define GGML_DEFAULT_N_THREADS 4 +#define GGML_DEFAULT_GRAPH_SIZE 2048 #if UINTPTR_MAX == 0xFFFFFFFF - #define GGML_MEM_ALIGN 4 +#define GGML_MEM_ALIGN 4 #else - #define GGML_MEM_ALIGN 16 +#define GGML_MEM_ALIGN 16 #endif #define GGML_EXIT_SUCCESS 0 @@ -245,7 +245,10 @@ do { \ if (!(x)) { \ fprintf(stderr, "GGML_ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \ - abort(); \ + fflush(stderr); \ + fflush(stdout); \ + ggml_print_backtrace(); \ + exit(1); \ } \ } while (0) @@ -286,1813 +289,1848 @@ extern "C" { #endif #if defined(__ARM_NEON) && defined(__CUDACC__) - typedef half ggml_fp16_t; +typedef half ggml_fp16_t; #elif defined(__ARM_NEON) - typedef __fp16 ggml_fp16_t; +typedef __fp16 ggml_fp16_t; #else - typedef uint16_t ggml_fp16_t; +typedef uint16_t ggml_fp16_t; #endif - // convert FP16 <-> FP32 - GGML_API float ggml_fp16_to_fp32(ggml_fp16_t x); - GGML_API ggml_fp16_t ggml_fp32_to_fp16(float x); - - GGML_API void ggml_fp16_to_fp32_row(const ggml_fp16_t * x, float * y, int n); - GGML_API void ggml_fp32_to_fp16_row(const float * x, ggml_fp16_t * y, int n); - - struct ggml_object; - struct ggml_context; - - enum ggml_type { - GGML_TYPE_F32 = 0, - GGML_TYPE_F16 = 1, - GGML_TYPE_Q4_0 = 2, - GGML_TYPE_Q4_1 = 3, - // GGML_TYPE_Q4_2 = 4, support has been removed - // GGML_TYPE_Q4_3 (5) support has been removed - GGML_TYPE_Q5_0 = 6, - GGML_TYPE_Q5_1 = 7, - GGML_TYPE_Q8_0 = 8, - GGML_TYPE_Q8_1 = 9, - // k-quantizations - GGML_TYPE_Q2_K = 10, - GGML_TYPE_Q3_K = 11, - GGML_TYPE_Q4_K = 12, - GGML_TYPE_Q5_K = 13, - GGML_TYPE_Q6_K = 14, - GGML_TYPE_Q8_K = 15, - GGML_TYPE_I8, - GGML_TYPE_I16, - GGML_TYPE_I32, - GGML_TYPE_COUNT, - }; - - enum ggml_backend_type { - GGML_BACKEND_CPU = 0, - GGML_BACKEND_GPU = 10, - GGML_BACKEND_GPU_SPLIT = 20, - }; - - // model file types - enum ggml_ftype { - GGML_FTYPE_UNKNOWN = -1, - GGML_FTYPE_ALL_F32 = 0, - GGML_FTYPE_MOSTLY_F16 = 1, // except 1d tensors - GGML_FTYPE_MOSTLY_Q4_0 = 2, // except 1d tensors - GGML_FTYPE_MOSTLY_Q4_1 = 3, // except 1d tensors - GGML_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16 - GGML_FTYPE_MOSTLY_Q8_0 = 7, // except 1d tensors - GGML_FTYPE_MOSTLY_Q5_0 = 8, // except 1d tensors - GGML_FTYPE_MOSTLY_Q5_1 = 9, // except 1d tensors - GGML_FTYPE_MOSTLY_Q2_K = 10, // except 1d tensors - GGML_FTYPE_MOSTLY_Q3_K = 11, // except 1d tensors - GGML_FTYPE_MOSTLY_Q4_K = 12, // except 1d tensors - GGML_FTYPE_MOSTLY_Q5_K = 13, // except 1d tensors - GGML_FTYPE_MOSTLY_Q6_K = 14, // except 1d tensors - }; - - // available tensor operations: - enum ggml_op { - GGML_OP_NONE = 0, - - GGML_OP_DUP, - GGML_OP_ADD, - GGML_OP_ADD1, - GGML_OP_ACC, - GGML_OP_SUB, - GGML_OP_MUL, - GGML_OP_DIV, - GGML_OP_SQR, - GGML_OP_SQRT, - GGML_OP_LOG, - GGML_OP_SUM, - GGML_OP_SUM_ROWS, - GGML_OP_MEAN, - GGML_OP_ARGMAX, - GGML_OP_REPEAT, - GGML_OP_REPEAT_BACK, - GGML_OP_CONCAT, - GGML_OP_SILU_BACK, - GGML_OP_NORM, // normalize - GGML_OP_RMS_NORM, - GGML_OP_RMS_NORM_BACK, - GGML_OP_GROUP_NORM, - - GGML_OP_MUL_MAT, - GGML_OP_OUT_PROD, - - GGML_OP_SCALE, - GGML_OP_SET, - GGML_OP_CPY, - GGML_OP_CONT, - GGML_OP_RESHAPE, - GGML_OP_VIEW, - GGML_OP_PERMUTE, - GGML_OP_TRANSPOSE, - GGML_OP_GET_ROWS, - GGML_OP_GET_ROWS_BACK, - GGML_OP_DIAG, - GGML_OP_DIAG_MASK_INF, - GGML_OP_DIAG_MASK_ZERO, - GGML_OP_SOFT_MAX, - GGML_OP_SOFT_MAX_BACK, - GGML_OP_ROPE, - GGML_OP_ROPE_BACK, - GGML_OP_ALIBI, - GGML_OP_CLAMP, - GGML_OP_CONV_1D, - GGML_OP_CONV_1D_STAGE_0, // internal - GGML_OP_CONV_1D_STAGE_1, // internal - GGML_OP_CONV_TRANSPOSE_1D, - GGML_OP_CONV_2D, - GGML_OP_CONV_2D_STAGE_0, // internal - GGML_OP_CONV_2D_STAGE_1, // internal - GGML_OP_CONV_TRANSPOSE_2D, - GGML_OP_POOL_1D, - GGML_OP_POOL_2D, - - GGML_OP_UPSCALE, // nearest interpolate - - GGML_OP_FLASH_ATTN, - GGML_OP_FLASH_FF, - GGML_OP_FLASH_ATTN_BACK, - GGML_OP_WIN_PART, - GGML_OP_WIN_UNPART, - GGML_OP_GET_REL_POS, - GGML_OP_ADD_REL_POS, - - GGML_OP_UNARY, - - GGML_OP_MAP_UNARY, - GGML_OP_MAP_BINARY, - - GGML_OP_MAP_CUSTOM1_F32, - GGML_OP_MAP_CUSTOM2_F32, - GGML_OP_MAP_CUSTOM3_F32, - - GGML_OP_MAP_CUSTOM1, - GGML_OP_MAP_CUSTOM2, - GGML_OP_MAP_CUSTOM3, - - GGML_OP_CROSS_ENTROPY_LOSS, - GGML_OP_CROSS_ENTROPY_LOSS_BACK, - - GGML_OP_COUNT, - }; - - enum ggml_unary_op { - GGML_UNARY_OP_ABS, - GGML_UNARY_OP_SGN, - GGML_UNARY_OP_NEG, - GGML_UNARY_OP_STEP, - GGML_UNARY_OP_TANH, - GGML_UNARY_OP_ELU, - GGML_UNARY_OP_RELU, - GGML_UNARY_OP_GELU, - GGML_UNARY_OP_GELU_QUICK, - GGML_UNARY_OP_SILU, - }; - - enum ggml_object_type { - GGML_OBJECT_TENSOR, - GGML_OBJECT_GRAPH, - GGML_OBJECT_WORK_BUFFER - }; - - enum ggml_log_level { - GGML_LOG_LEVEL_ERROR = 2, - GGML_LOG_LEVEL_WARN = 3, - GGML_LOG_LEVEL_INFO = 4 - }; - - // ggml object - struct ggml_object { - size_t offs; - size_t size; - - struct ggml_object * next; - - enum ggml_object_type type; - - char padding[4]; - }; - - static const size_t GGML_OBJECT_SIZE = sizeof(struct ggml_object); - - // n-dimensional tensor - struct ggml_tensor { - enum ggml_type type; - enum ggml_backend_type backend; - - struct ggml_backend_buffer * buffer; - - int n_dims; - int64_t ne[GGML_MAX_DIMS]; // number of elements - size_t nb[GGML_MAX_DIMS]; // stride in bytes: - // nb[0] = ggml_type_size(type) - // nb[1] = nb[0] * (ne[0] / ggml_blck_size(type)) + padding - // nb[i] = nb[i-1] * ne[i-1] - - // compute data - enum ggml_op op; - - // op params - allocated as int32_t for alignment - int32_t op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)]; - - bool is_param; - - struct ggml_tensor * grad; - struct ggml_tensor * src[GGML_MAX_SRC]; +// convert FP16 <-> FP32 +GGML_API float ggml_fp16_to_fp32(ggml_fp16_t x); +GGML_API ggml_fp16_t ggml_fp32_to_fp16(float x); + +GGML_API void ggml_fp16_to_fp32_row(const ggml_fp16_t * x, float * y, int n); +GGML_API void ggml_fp32_to_fp16_row(const float * x, ggml_fp16_t * y, int n); + +struct ggml_object; +struct ggml_context; + +enum ggml_type { + GGML_TYPE_F32 = 0, + GGML_TYPE_F16 = 1, + GGML_TYPE_Q4_0 = 2, + GGML_TYPE_Q4_1 = 3, + // GGML_TYPE_Q4_2 = 4, support has been removed + // GGML_TYPE_Q4_3 (5) support has been removed + GGML_TYPE_Q5_0 = 6, + GGML_TYPE_Q5_1 = 7, + GGML_TYPE_Q8_0 = 8, + GGML_TYPE_Q8_1 = 9, + // k-quantizations + GGML_TYPE_Q2_K = 10, + GGML_TYPE_Q3_K = 11, + GGML_TYPE_Q4_K = 12, + GGML_TYPE_Q5_K = 13, + GGML_TYPE_Q6_K = 14, + GGML_TYPE_Q8_K = 15, + GGML_TYPE_I8, + GGML_TYPE_I16, + GGML_TYPE_I32, + GGML_TYPE_COUNT, +}; + +enum ggml_backend_type { + GGML_BACKEND_CPU = 0, + GGML_BACKEND_GPU = 10, + GGML_BACKEND_GPU_SPLIT = 20, +}; + +// model file types +enum ggml_ftype { + GGML_FTYPE_UNKNOWN = -1, + GGML_FTYPE_ALL_F32 = 0, + GGML_FTYPE_MOSTLY_F16 = 1, // except 1d tensors + GGML_FTYPE_MOSTLY_Q4_0 = 2, // except 1d tensors + GGML_FTYPE_MOSTLY_Q4_1 = 3, // except 1d tensors + GGML_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16 + GGML_FTYPE_MOSTLY_Q8_0 = 7, // except 1d tensors + GGML_FTYPE_MOSTLY_Q5_0 = 8, // except 1d tensors + GGML_FTYPE_MOSTLY_Q5_1 = 9, // except 1d tensors + GGML_FTYPE_MOSTLY_Q2_K = 10, // except 1d tensors + GGML_FTYPE_MOSTLY_Q3_K = 11, // except 1d tensors + GGML_FTYPE_MOSTLY_Q4_K = 12, // except 1d tensors + GGML_FTYPE_MOSTLY_Q5_K = 13, // except 1d tensors + GGML_FTYPE_MOSTLY_Q6_K = 14, // except 1d tensors +}; + +// available tensor operations: +enum ggml_op { + GGML_OP_NONE = 0, + + GGML_OP_DUP, + GGML_OP_ADD, + GGML_OP_ADD1, + GGML_OP_ACC, + GGML_OP_SUB, + GGML_OP_MUL, + GGML_OP_DIV, + GGML_OP_SQR, + GGML_OP_SQRT, + GGML_OP_LOG, + GGML_OP_SUM, + GGML_OP_SUM_ROWS, + GGML_OP_MEAN, + GGML_OP_ARGMAX, + GGML_OP_REPEAT, + GGML_OP_REPEAT_BACK, + GGML_OP_CONCAT, + GGML_OP_SILU_BACK, + GGML_OP_NORM, // normalize + GGML_OP_RMS_NORM, + GGML_OP_RMS_NORM_BACK, + GGML_OP_GROUP_NORM, + + GGML_OP_MUL_MAT, + GGML_OP_OUT_PROD, + + GGML_OP_SCALE, + GGML_OP_SET, + GGML_OP_CPY, + GGML_OP_CONT, + GGML_OP_RESHAPE, + GGML_OP_VIEW, + GGML_OP_PERMUTE, + GGML_OP_TRANSPOSE, + GGML_OP_GET_ROWS, + GGML_OP_GET_ROWS_BACK, + GGML_OP_DIAG, + GGML_OP_DIAG_MASK_INF, + GGML_OP_DIAG_MASK_ZERO, + GGML_OP_SOFT_MAX, + GGML_OP_SOFT_MAX_BACK, + GGML_OP_ROPE, + GGML_OP_ROPE_BACK, + GGML_OP_ALIBI, + GGML_OP_CLAMP, + GGML_OP_CONV_TRANSPOSE_1D, + GGML_OP_IM2COL, + GGML_OP_CONV_TRANSPOSE_2D, + GGML_OP_POOL_1D, + GGML_OP_POOL_2D, + + GGML_OP_UPSCALE, // nearest interpolate + + GGML_OP_FLASH_ATTN, + GGML_OP_FLASH_FF, + GGML_OP_FLASH_ATTN_BACK, + GGML_OP_WIN_PART, + GGML_OP_WIN_UNPART, + GGML_OP_GET_REL_POS, + GGML_OP_ADD_REL_POS, + + GGML_OP_UNARY, + + GGML_OP_MAP_UNARY, + GGML_OP_MAP_BINARY, + + GGML_OP_MAP_CUSTOM1_F32, + GGML_OP_MAP_CUSTOM2_F32, + GGML_OP_MAP_CUSTOM3_F32, + + GGML_OP_MAP_CUSTOM1, + GGML_OP_MAP_CUSTOM2, + GGML_OP_MAP_CUSTOM3, + + GGML_OP_CROSS_ENTROPY_LOSS, + GGML_OP_CROSS_ENTROPY_LOSS_BACK, + + GGML_OP_COUNT, +}; + +enum ggml_unary_op { + GGML_UNARY_OP_ABS, + GGML_UNARY_OP_SGN, + GGML_UNARY_OP_NEG, + GGML_UNARY_OP_STEP, + GGML_UNARY_OP_TANH, + GGML_UNARY_OP_ELU, + GGML_UNARY_OP_RELU, + GGML_UNARY_OP_GELU, + GGML_UNARY_OP_GELU_QUICK, + GGML_UNARY_OP_SILU, + GGML_UNARY_OP_LEAKY +}; + +enum ggml_object_type { + GGML_OBJECT_TENSOR, + GGML_OBJECT_GRAPH, + GGML_OBJECT_WORK_BUFFER +}; + +enum ggml_log_level { + GGML_LOG_LEVEL_ERROR = 2, + GGML_LOG_LEVEL_WARN = 3, + GGML_LOG_LEVEL_INFO = 4 +}; + +// ggml object +struct ggml_object { + size_t offs; + size_t size; + + struct ggml_object * next; + + enum ggml_object_type type; + + char padding[4]; +}; + +static const size_t GGML_OBJECT_SIZE = sizeof(struct ggml_object); + +// n-dimensional tensor +struct ggml_tensor { + enum ggml_type type; + enum ggml_backend_type backend; + + struct ggml_backend_buffer * buffer; + + int n_dims; + int64_t ne[GGML_MAX_DIMS]; // number of elements + size_t nb[GGML_MAX_DIMS]; // stride in bytes: + // nb[0] = ggml_type_size(type) + // nb[1] = nb[0] * (ne[0] / ggml_blck_size(type)) + padding + // nb[i] = nb[i-1] * ne[i-1] + + // compute data + enum ggml_op op; + + // op params - allocated as int32_t for alignment + int32_t op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)]; + + bool is_param; + + struct ggml_tensor * grad; + struct ggml_tensor * src[GGML_MAX_SRC]; + + // performance + int perf_runs; + int64_t perf_cycles; + int64_t perf_time_us; - // performance - int perf_runs; - int64_t perf_cycles; - int64_t perf_time_us; + struct ggml_tensor * view_src; + size_t view_offs; - struct ggml_tensor * view_src; - size_t view_offs; + void * data; - void * data; + char name[GGML_MAX_NAME]; - char name[GGML_MAX_NAME]; + void * extra; // extra things e.g. for ggml-cuda.cu - void * extra; // extra things e.g. for ggml-cuda.cu + char padding[12]; +}; - char padding[12]; - }; +static const size_t GGML_TENSOR_SIZE = sizeof(struct ggml_tensor); - static const size_t GGML_TENSOR_SIZE = sizeof(struct ggml_tensor); +// the compute plan that needs to be prepared for ggml_graph_compute() +// since https://github.com/ggerganov/ggml/issues/287 +struct ggml_cplan { + size_t work_size; // size of work buffer, calculated by `ggml_graph_plan()` + uint8_t * work_data; // work buffer, to be allocated by caller before calling to `ggml_graph_compute()` - // the compute plan that needs to be prepared for ggml_graph_compute() - // since https://github.com/ggerganov/ggml/issues/287 - struct ggml_cplan { - size_t work_size; // size of work buffer, calculated by `ggml_graph_plan()` - uint8_t * work_data; // work buffer, to be allocated by caller before calling to `ggml_graph_compute()` + int n_threads; - int n_threads; + // abort ggml_graph_compute when true + bool (*abort_callback)(void * data); + void * abort_callback_data; +}; - // the `n_tasks` of nodes, 1:1 mapping to cgraph nodes - int n_tasks[GGML_MAX_NODES]; +enum ggml_cgraph_eval_order { + GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT = 0, + GGML_CGRAPH_EVAL_ORDER_RIGHT_TO_LEFT, + GGML_CGRAPH_EVAL_ORDER_COUNT +}; - // abort ggml_graph_compute when true - bool (*abort_callback)(void * data); - void * abort_callback_data; - }; +struct ggml_hash_set { + size_t size; + struct ggml_tensor ** keys; +}; - // next prime after GGML_MAX_NODES - // #define GGML_GRAPH_HASHTABLE_SIZE 4099 - // next prime after GGML_MAX_NODES * 2 (nodes + leafs) - // #define GGML_GRAPH_HASHTABLE_SIZE 8273 - // #define GGML_GRAPH_HASHTABLE_SIZE 16411 - #define GGML_GRAPH_HASHTABLE_SIZE 32771 +// computation graph +struct ggml_cgraph { + int size; + int n_nodes; + int n_leafs; - enum ggml_cgraph_eval_order { - GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT = 0, - GGML_CGRAPH_EVAL_ORDER_RIGHT_TO_LEFT, - GGML_CGRAPH_EVAL_ORDER_COUNT - }; + struct ggml_tensor ** nodes; + struct ggml_tensor ** grads; + struct ggml_tensor ** leafs; - // computation graph - struct ggml_cgraph { - int n_nodes; - int n_leafs; + struct ggml_hash_set visited_hash_table; - struct ggml_tensor * nodes[GGML_MAX_NODES]; - struct ggml_tensor * grads[GGML_MAX_NODES]; - struct ggml_tensor * leafs[GGML_MAX_NODES]; + enum ggml_cgraph_eval_order order; - void * visited_hash_table[GGML_GRAPH_HASHTABLE_SIZE]; + // performance + int perf_runs; + int64_t perf_cycles; + int64_t perf_time_us; +}; - enum ggml_cgraph_eval_order order; +// scratch buffer +struct ggml_scratch { + size_t offs; + size_t size; + void * data; +}; - // performance - int perf_runs; - int64_t perf_cycles; - int64_t perf_time_us; - }; +struct ggml_init_params { + // memory pool + size_t mem_size; // bytes + void * mem_buffer; // if NULL, memory will be allocated internally + bool no_alloc; // don't allocate memory for the tensor data +}; - static const size_t GGML_GRAPH_SIZE = sizeof(struct ggml_cgraph); - // scratch buffer - struct ggml_scratch { - size_t offs; - size_t size; - void * data; - }; +// compute types - struct ggml_init_params { - // memory pool - size_t mem_size; // bytes - void * mem_buffer; // if NULL, memory will be allocated internally - bool no_alloc; // don't allocate memory for the tensor data - }; +// NOTE: the INIT or FINALIZE pass is not scheduled unless explicitly enabled. +// This behavior was changed since https://github.com/ggerganov/llama.cpp/pull/1995. +enum ggml_task_type { + GGML_TASK_INIT = 0, + GGML_TASK_COMPUTE, + GGML_TASK_FINALIZE, +}; +struct ggml_compute_params { + enum ggml_task_type type; - // compute types + // ith = thread index, nth = number of threads + int ith, nth; - // NOTE: the INIT or FINALIZE pass is not scheduled unless explicitly enabled. - // This behavior was changed since https://github.com/ggerganov/llama.cpp/pull/1995. - enum ggml_task_type { - GGML_TASK_INIT = 0, - GGML_TASK_COMPUTE, - GGML_TASK_FINALIZE, - }; + // work buffer for all threads + size_t wsize; + void * wdata; +}; - struct ggml_compute_params { - enum ggml_task_type type; +// misc - // ith = thread index, nth = number of threads - int ith, nth; +GGML_API void ggml_time_init(void); // call this once at the beginning of the program +GGML_API int64_t ggml_time_ms(void); +GGML_API int64_t ggml_time_us(void); +GGML_API int64_t ggml_cycles(void); +GGML_API int64_t ggml_cycles_per_ms(void); - // work buffer for all threads - size_t wsize; - void * wdata; - }; +GGML_API void ggml_print_backtrace(void); - // misc +GGML_API void ggml_numa_init(void); // call once for better performance on NUMA systems +GGML_API bool ggml_is_numa(void); // true if init detected that system has >1 NUMA node - GGML_API void ggml_time_init(void); // call this once at the beginning of the program - GGML_API int64_t ggml_time_ms(void); - GGML_API int64_t ggml_time_us(void); - GGML_API int64_t ggml_cycles(void); - GGML_API int64_t ggml_cycles_per_ms(void); +GGML_API void ggml_print_object (const struct ggml_object * obj); +GGML_API void ggml_print_objects(const struct ggml_context * ctx); - GGML_API void ggml_numa_init(void); // call once for better performance on NUMA systems - GGML_API bool ggml_is_numa(void); // true if init detected that system has >1 NUMA node +GGML_API int64_t ggml_nelements (const struct ggml_tensor * tensor); +GGML_API int64_t ggml_nrows (const struct ggml_tensor * tensor); +GGML_API size_t ggml_nbytes (const struct ggml_tensor * tensor); +GGML_API size_t ggml_nbytes_pad (const struct ggml_tensor * tensor); // same as ggml_nbytes() but padded to GGML_MEM_ALIGN +GGML_API size_t ggml_nbytes_split(const struct ggml_tensor * tensor, int nrows_split); - GGML_API void ggml_print_object (const struct ggml_object * obj); - GGML_API void ggml_print_objects(const struct ggml_context * ctx); +GGML_API int ggml_blck_size (enum ggml_type type); +GGML_API size_t ggml_type_size (enum ggml_type type); // size in bytes for all elements in a block +GGML_API float ggml_type_sizef(enum ggml_type type); // ggml_type_size()/ggml_blck_size() as float - GGML_API int64_t ggml_nelements (const struct ggml_tensor * tensor); - GGML_API int64_t ggml_nrows (const struct ggml_tensor * tensor); - GGML_API size_t ggml_nbytes (const struct ggml_tensor * tensor); - GGML_API size_t ggml_nbytes_pad (const struct ggml_tensor * tensor); // same as ggml_nbytes() but padded to GGML_MEM_ALIGN - GGML_API size_t ggml_nbytes_split(const struct ggml_tensor * tensor, int nrows_split); +GGML_API const char * ggml_type_name(enum ggml_type type); +GGML_API const char * ggml_op_name (enum ggml_op op); +GGML_API const char * ggml_op_symbol(enum ggml_op op); - GGML_API int ggml_blck_size (enum ggml_type type); - GGML_API size_t ggml_type_size (enum ggml_type type); // size in bytes for all elements in a block - GGML_API float ggml_type_sizef(enum ggml_type type); // ggml_type_size()/ggml_blck_size() as float +GGML_API size_t ggml_element_size(const struct ggml_tensor * tensor); - GGML_API const char * ggml_type_name(enum ggml_type type); - GGML_API const char * ggml_op_name (enum ggml_op op); - GGML_API const char * ggml_op_symbol(enum ggml_op op); +GGML_API bool ggml_is_quantized(enum ggml_type type); - GGML_API size_t ggml_element_size(const struct ggml_tensor * tensor); +// TODO: temporary until model loading of ggml examples is refactored +GGML_API enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype); - GGML_API bool ggml_is_quantized(enum ggml_type type); +GGML_API bool ggml_is_transposed(const struct ggml_tensor * tensor); +GGML_API bool ggml_is_contiguous(const struct ggml_tensor * tensor); +GGML_API bool ggml_is_permuted (const struct ggml_tensor * tensor); - // TODO: temporary until model loading of ggml examples is refactored - GGML_API enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype); +GGML_API bool ggml_are_same_shape(const struct ggml_tensor * t0, const struct ggml_tensor * t1); - GGML_API bool ggml_is_transposed(const struct ggml_tensor * tensor); - GGML_API bool ggml_is_contiguous(const struct ggml_tensor * tensor); - GGML_API bool ggml_is_permuted (const struct ggml_tensor * tensor); +// use this to compute the memory overhead of a tensor +GGML_API size_t ggml_tensor_overhead(void); - GGML_API bool ggml_are_same_shape(const struct ggml_tensor * t0, const struct ggml_tensor * t1); +// main - // use this to compute the memory overhead of a tensor - GGML_API size_t ggml_tensor_overhead(void); +GGML_API struct ggml_context * ggml_init(struct ggml_init_params params); +GGML_API void ggml_free(struct ggml_context * ctx); - // main +GGML_API size_t ggml_used_mem(const struct ggml_context * ctx); - GGML_API struct ggml_context * ggml_init(struct ggml_init_params params); - GGML_API void ggml_free(struct ggml_context * ctx); +GGML_API size_t ggml_set_scratch (struct ggml_context * ctx, struct ggml_scratch scratch); +GGML_API bool ggml_get_no_alloc(struct ggml_context * ctx); +GGML_API void ggml_set_no_alloc(struct ggml_context * ctx, bool no_alloc); - GGML_API size_t ggml_used_mem(const struct ggml_context * ctx); +GGML_API void * ggml_get_mem_buffer (const struct ggml_context * ctx); +GGML_API size_t ggml_get_mem_size (const struct ggml_context * ctx); +GGML_API size_t ggml_get_max_tensor_size(const struct ggml_context * ctx); - GGML_API size_t ggml_set_scratch (struct ggml_context * ctx, struct ggml_scratch scratch); - GGML_API bool ggml_get_no_alloc(struct ggml_context * ctx); - GGML_API void ggml_set_no_alloc(struct ggml_context * ctx, bool no_alloc); +GGML_API struct ggml_tensor * ggml_new_tensor( + struct ggml_context * ctx, + enum ggml_type type, + int n_dims, + const int64_t *ne); - GGML_API void * ggml_get_mem_buffer (const struct ggml_context * ctx); - GGML_API size_t ggml_get_mem_size (const struct ggml_context * ctx); - GGML_API size_t ggml_get_max_tensor_size(const struct ggml_context * ctx); +GGML_API struct ggml_tensor * ggml_new_tensor_1d( + struct ggml_context * ctx, + enum ggml_type type, + int64_t ne0); - GGML_API struct ggml_tensor * ggml_new_tensor( - struct ggml_context * ctx, - enum ggml_type type, - int n_dims, - const int64_t *ne); +GGML_API struct ggml_tensor * ggml_new_tensor_2d( + struct ggml_context * ctx, + enum ggml_type type, + int64_t ne0, + int64_t ne1); - GGML_API struct ggml_tensor * ggml_new_tensor_1d( - struct ggml_context * ctx, - enum ggml_type type, - int64_t ne0); - - GGML_API struct ggml_tensor * ggml_new_tensor_2d( - struct ggml_context * ctx, - enum ggml_type type, - int64_t ne0, - int64_t ne1); - - GGML_API struct ggml_tensor * ggml_new_tensor_3d( - struct ggml_context * ctx, - enum ggml_type type, - int64_t ne0, - int64_t ne1, - int64_t ne2); - - GGML_API struct ggml_tensor * ggml_new_tensor_4d( - struct ggml_context * ctx, - enum ggml_type type, - int64_t ne0, - int64_t ne1, - int64_t ne2, - int64_t ne3); - - GGML_API struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value); - GGML_API struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value); - - GGML_API struct ggml_tensor * ggml_dup_tensor (struct ggml_context * ctx, const struct ggml_tensor * src); - GGML_API struct ggml_tensor * ggml_view_tensor(struct ggml_context * ctx, struct ggml_tensor * src); - - // Context tensor enumeration and lookup - GGML_API struct ggml_tensor * ggml_get_first_tensor(struct ggml_context * ctx); - GGML_API struct ggml_tensor * ggml_get_next_tensor (struct ggml_context * ctx, struct ggml_tensor * tensor); - GGML_API struct ggml_tensor * ggml_get_tensor (struct ggml_context * ctx, const char * name); - - GGML_API struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor); - GGML_API struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value); - GGML_API struct ggml_tensor * ggml_set_f32 (struct ggml_tensor * tensor, float value); - - // Converts a flat index into coordinates - GGML_API void ggml_unravel_index(const struct ggml_tensor * tensor, int64_t i, int64_t * i0, int64_t * i1, int64_t * i2, int64_t * i3); - - GGML_API int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i); - GGML_API void ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value); - - GGML_API int32_t ggml_get_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3); - GGML_API void ggml_set_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, int32_t value); - - GGML_API float ggml_get_f32_1d(const struct ggml_tensor * tensor, int i); - GGML_API void ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value); - - GGML_API float ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3); - GGML_API void ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, float value); - - GGML_API void * ggml_get_data (const struct ggml_tensor * tensor); - GGML_API float * ggml_get_data_f32(const struct ggml_tensor * tensor); - - GGML_API enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor); - - GGML_API const char * ggml_get_name (const struct ggml_tensor * tensor); - GGML_API struct ggml_tensor * ggml_set_name ( struct ggml_tensor * tensor, const char * name); - GGML_ATTRIBUTE_FORMAT(2, 3) - GGML_API struct ggml_tensor * ggml_format_name( struct ggml_tensor * tensor, const char * fmt, ...); - - // - // operations on tensors with backpropagation - // - - GGML_API struct ggml_tensor * ggml_dup( - struct ggml_context * ctx, - struct ggml_tensor * a); - - // in-place, returns view(a) - GGML_API struct ggml_tensor * ggml_dup_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a); - - GGML_API struct ggml_tensor * ggml_add( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); - - GGML_API struct ggml_tensor * ggml_add_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); - - GGML_API struct ggml_tensor * ggml_add_cast( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - enum ggml_type type); - - GGML_API struct ggml_tensor * ggml_add1( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); - - GGML_API struct ggml_tensor * ggml_add1_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); - - GGML_API struct ggml_tensor * ggml_acc( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - size_t nb1, - size_t nb2, - size_t nb3, - size_t offset); - - GGML_API struct ggml_tensor * ggml_acc_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - size_t nb1, - size_t nb2, - size_t nb3, - size_t offset); - - GGML_API struct ggml_tensor * ggml_sub( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); - - GGML_API struct ggml_tensor * ggml_sub_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); - - GGML_API struct ggml_tensor * ggml_mul( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); - - GGML_API struct ggml_tensor * ggml_mul_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); - - GGML_API struct ggml_tensor * ggml_div( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); - - GGML_API struct ggml_tensor * ggml_div_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); - - GGML_API struct ggml_tensor * ggml_sqr( - struct ggml_context * ctx, - struct ggml_tensor * a); - - GGML_API struct ggml_tensor * ggml_sqr_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a); - - GGML_API struct ggml_tensor * ggml_sqrt( - struct ggml_context * ctx, - struct ggml_tensor * a); - - GGML_API struct ggml_tensor * ggml_sqrt_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a); - - GGML_API struct ggml_tensor * ggml_log( - struct ggml_context * ctx, - struct ggml_tensor * a); - - GGML_API struct ggml_tensor * ggml_log_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a); - - // return scalar - GGML_API struct ggml_tensor * ggml_sum( - struct ggml_context * ctx, - struct ggml_tensor * a); - - // sums along rows, with input shape [a,b,c,d] return shape [1,b,c,d] - GGML_API struct ggml_tensor * ggml_sum_rows( - struct ggml_context * ctx, - struct ggml_tensor * a); - - // mean along rows - GGML_API struct ggml_tensor * ggml_mean( - struct ggml_context * ctx, - struct ggml_tensor * a); - - // argmax along rows - GGML_API struct ggml_tensor * ggml_argmax( - struct ggml_context * ctx, - struct ggml_tensor * a); - - // if a is the same shape as b, and a is not parameter, return a - // otherwise, return a new tensor: repeat(a) to fit in b - GGML_API struct ggml_tensor * ggml_repeat( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); - - // sums repetitions in a into shape of b - GGML_API struct ggml_tensor * ggml_repeat_back( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); - - // concat a and b on dim 2 - // used in stable-diffusion - GGML_API struct ggml_tensor * ggml_concat( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); - - GGML_API struct ggml_tensor * ggml_abs( - struct ggml_context * ctx, - struct ggml_tensor * a); - - GGML_API struct ggml_tensor * ggml_abs_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a); - - GGML_API struct ggml_tensor * ggml_sgn( - struct ggml_context * ctx, - struct ggml_tensor * a); - - GGML_API struct ggml_tensor * ggml_sgn_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a); - - GGML_API struct ggml_tensor * ggml_neg( - struct ggml_context * ctx, - struct ggml_tensor * a); - - GGML_API struct ggml_tensor * ggml_neg_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a); - - GGML_API struct ggml_tensor * ggml_step( - struct ggml_context * ctx, - struct ggml_tensor * a); - - GGML_API struct ggml_tensor * ggml_step_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a); - - GGML_API struct ggml_tensor * ggml_tanh( - struct ggml_context * ctx, - struct ggml_tensor * a); - - GGML_API struct ggml_tensor * ggml_tanh_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a); - - GGML_API struct ggml_tensor * ggml_elu( - struct ggml_context * ctx, - struct ggml_tensor * a); - - GGML_API struct ggml_tensor * ggml_elu_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a); - - GGML_API struct ggml_tensor * ggml_relu( - struct ggml_context * ctx, - struct ggml_tensor * a); - - GGML_API struct ggml_tensor * ggml_relu_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a); - - // TODO: double-check this computation is correct - GGML_API struct ggml_tensor * ggml_gelu( - struct ggml_context * ctx, - struct ggml_tensor * a); - - GGML_API struct ggml_tensor * ggml_gelu_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a); - - GGML_API struct ggml_tensor * ggml_gelu_quick( - struct ggml_context * ctx, - struct ggml_tensor * a); - - GGML_API struct ggml_tensor * ggml_gelu_quick_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a); - - GGML_API struct ggml_tensor * ggml_silu( - struct ggml_context * ctx, - struct ggml_tensor * a); - - GGML_API struct ggml_tensor * ggml_silu_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a); - - // a - x - // b - dy - GGML_API struct ggml_tensor * ggml_silu_back( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); - - // normalize along rows - GGML_API struct ggml_tensor * ggml_norm( - struct ggml_context * ctx, - struct ggml_tensor * a, - float eps); - - GGML_API struct ggml_tensor * ggml_norm_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a, - float eps); - - GGML_API struct ggml_tensor * ggml_rms_norm( - struct ggml_context * ctx, - struct ggml_tensor * a, - float eps); - - GGML_API struct ggml_tensor * ggml_rms_norm_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a, - float eps); - - // group normalize along ne0*ne1*n_groups - // used in stable-diffusion - // TODO: eps is hardcoded to 1e-6 for now - GGML_API struct ggml_tensor * ggml_group_norm( - struct ggml_context * ctx, - struct ggml_tensor * a, - int n_groups); - - GGML_API struct ggml_tensor * ggml_group_norm_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a, - int n_groups); - - // a - x - // b - dy - GGML_API struct ggml_tensor * ggml_rms_norm_back( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - float eps); - - // A: k columns, n rows => [ne03, ne02, n, k] - // B: k columns, m rows (i.e. we transpose it internally) => [ne03 * x, ne02 * y, m, k] - // result is n columns, m rows => [ne03 * x, ne02 * y, m, n] - GGML_API struct ggml_tensor * ggml_mul_mat( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); - - // A: m columns, n rows, - // B: p columns, n rows, - // result is m columns, p rows - GGML_API struct ggml_tensor * ggml_out_prod( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); - - // - // operations on tensors without backpropagation - // - - GGML_API struct ggml_tensor * ggml_scale( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); - - // in-place, returns view(a) - GGML_API struct ggml_tensor * ggml_scale_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); - - // b -> view(a,offset,nb1,nb2,3), return modified a - GGML_API struct ggml_tensor * ggml_set( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - size_t nb1, - size_t nb2, - size_t nb3, - size_t offset); - - // b -> view(a,offset,nb1,nb2,3), return view(a) - GGML_API struct ggml_tensor * ggml_set_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - size_t nb1, - size_t nb2, - size_t nb3, - size_t offset); - - GGML_API struct ggml_tensor * ggml_set_1d( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - size_t offset); - - GGML_API struct ggml_tensor * ggml_set_1d_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - size_t offset); - - // b -> view(a,offset,nb1,nb2,3), return modified a - GGML_API struct ggml_tensor * ggml_set_2d( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - size_t nb1, - size_t offset); - - // b -> view(a,offset,nb1,nb2,3), return view(a) - GGML_API struct ggml_tensor * ggml_set_2d_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - size_t nb1, - size_t offset); - - // a -> b, return view(b) - GGML_API struct ggml_tensor * ggml_cpy( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); - - // a -> b, in-place, return view(b) - GGML_API struct ggml_tensor * ggml_cpy_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); - - // make contiguous - GGML_API struct ggml_tensor * ggml_cont( - struct ggml_context * ctx, - struct ggml_tensor * a); - - // make contiguous, in-place - GGML_API struct ggml_tensor * ggml_cont_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a); - - // make contiguous, with new shape - GGML_API struct ggml_tensor * ggml_cont_1d( - struct ggml_context * ctx, - struct ggml_tensor * a, - int64_t ne0); - - GGML_API struct ggml_tensor * ggml_cont_2d( - struct ggml_context * ctx, - struct ggml_tensor * a, - int64_t ne0, - int64_t ne1); - - GGML_API struct ggml_tensor * ggml_cont_3d( - struct ggml_context * ctx, - struct ggml_tensor * a, - int64_t ne0, - int64_t ne1, - int64_t ne2); - - GGML_API struct ggml_tensor * ggml_cont_4d( - struct ggml_context * ctx, - struct ggml_tensor * a, - int64_t ne0, - int64_t ne1, - int64_t ne2, - int64_t ne3); - - // return view(a), b specifies the new shape - // TODO: when we start computing gradient, make a copy instead of view - GGML_API struct ggml_tensor * ggml_reshape( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); - - // return view(a) - // TODO: when we start computing gradient, make a copy instead of view - GGML_API struct ggml_tensor * ggml_reshape_1d( - struct ggml_context * ctx, - struct ggml_tensor * a, - int64_t ne0); - - GGML_API struct ggml_tensor * ggml_reshape_2d( - struct ggml_context * ctx, - struct ggml_tensor * a, - int64_t ne0, - int64_t ne1); - - // return view(a) - // TODO: when we start computing gradient, make a copy instead of view - GGML_API struct ggml_tensor * ggml_reshape_3d( - struct ggml_context * ctx, - struct ggml_tensor * a, - int64_t ne0, - int64_t ne1, - int64_t ne2); - - GGML_API struct ggml_tensor * ggml_reshape_4d( - struct ggml_context * ctx, - struct ggml_tensor * a, - int64_t ne0, - int64_t ne1, - int64_t ne2, - int64_t ne3); - - // offset in bytes - GGML_API struct ggml_tensor * ggml_view_1d( - struct ggml_context * ctx, - struct ggml_tensor * a, - int64_t ne0, - size_t offset); - - GGML_API struct ggml_tensor * ggml_view_2d( - struct ggml_context * ctx, - struct ggml_tensor * a, - int64_t ne0, - int64_t ne1, - size_t nb1, // row stride in bytes - size_t offset); - - GGML_API struct ggml_tensor * ggml_view_3d( - struct ggml_context * ctx, - struct ggml_tensor * a, - int64_t ne0, - int64_t ne1, - int64_t ne2, - size_t nb1, // row stride in bytes - size_t nb2, // slice stride in bytes - size_t offset); - - GGML_API struct ggml_tensor * ggml_view_4d( - struct ggml_context * ctx, - struct ggml_tensor * a, - int64_t ne0, - int64_t ne1, - int64_t ne2, - int64_t ne3, - size_t nb1, // row stride in bytes - size_t nb2, // slice stride in bytes - size_t nb3, - size_t offset); - - GGML_API struct ggml_tensor * ggml_permute( - struct ggml_context * ctx, - struct ggml_tensor * a, - int axis0, - int axis1, - int axis2, - int axis3); - - // alias for ggml_permute(ctx, a, 1, 0, 2, 3) - GGML_API struct ggml_tensor * ggml_transpose( - struct ggml_context * ctx, - struct ggml_tensor * a); - - GGML_API struct ggml_tensor * ggml_get_rows( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); - - GGML_API struct ggml_tensor * ggml_get_rows_back( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - struct ggml_tensor * c); - - GGML_API struct ggml_tensor * ggml_diag( +GGML_API struct ggml_tensor * ggml_new_tensor_3d( + struct ggml_context * ctx, + enum ggml_type type, + int64_t ne0, + int64_t ne1, + int64_t ne2); + +GGML_API struct ggml_tensor * ggml_new_tensor_4d( + struct ggml_context * ctx, + enum ggml_type type, + int64_t ne0, + int64_t ne1, + int64_t ne2, + int64_t ne3); + +GGML_API struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value); +GGML_API struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value); + +GGML_API struct ggml_tensor * ggml_dup_tensor (struct ggml_context * ctx, const struct ggml_tensor * src); +GGML_API struct ggml_tensor * ggml_view_tensor(struct ggml_context * ctx, struct ggml_tensor * src); + +// Context tensor enumeration and lookup +GGML_API struct ggml_tensor * ggml_get_first_tensor(struct ggml_context * ctx); +GGML_API struct ggml_tensor * ggml_get_next_tensor (struct ggml_context * ctx, struct ggml_tensor * tensor); +GGML_API struct ggml_tensor * ggml_get_tensor(struct ggml_context * ctx, const char * name); + +GGML_API struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor); +GGML_API struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value); +GGML_API struct ggml_tensor * ggml_set_f32 (struct ggml_tensor * tensor, float value); + +// Converts a flat index into coordinates +GGML_API void ggml_unravel_index(const struct ggml_tensor * tensor, int64_t i, int64_t * i0, int64_t * i1, int64_t * i2, int64_t * i3); + +GGML_API int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i); +GGML_API void ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value); + +GGML_API int32_t ggml_get_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3); +GGML_API void ggml_set_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, int32_t value); + +GGML_API float ggml_get_f32_1d(const struct ggml_tensor * tensor, int i); +GGML_API void ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value); + +GGML_API float ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3); +GGML_API void ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, float value); + +GGML_API void * ggml_get_data (const struct ggml_tensor * tensor); +GGML_API float * ggml_get_data_f32(const struct ggml_tensor * tensor); + +GGML_API enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor); + +GGML_API const char * ggml_get_name (const struct ggml_tensor * tensor); +GGML_API struct ggml_tensor * ggml_set_name ( struct ggml_tensor * tensor, const char * name); +GGML_ATTRIBUTE_FORMAT(2, 3) +GGML_API struct ggml_tensor * ggml_format_name( struct ggml_tensor * tensor, const char * fmt, ...); + +// +// operations on tensors with backpropagation +// + +GGML_API struct ggml_tensor * ggml_dup( + struct ggml_context * ctx, + struct ggml_tensor * a); + +// in-place, returns view(a) +GGML_API struct ggml_tensor * ggml_dup_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a); + +GGML_API struct ggml_tensor * ggml_add( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + +GGML_API struct ggml_tensor * ggml_add_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + +GGML_API struct ggml_tensor * ggml_add_cast( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + enum ggml_type type); + +GGML_API struct ggml_tensor * ggml_add1( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + +GGML_API struct ggml_tensor * ggml_add1_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + +GGML_API struct ggml_tensor * ggml_acc( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t nb2, + size_t nb3, + size_t offset); + +GGML_API struct ggml_tensor * ggml_acc_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t nb2, + size_t nb3, + size_t offset); + +GGML_API struct ggml_tensor * ggml_sub( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + +GGML_API struct ggml_tensor * ggml_sub_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + +GGML_API struct ggml_tensor * ggml_mul( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + +GGML_API struct ggml_tensor * ggml_mul_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + +GGML_API struct ggml_tensor * ggml_div( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + +GGML_API struct ggml_tensor * ggml_div_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + +GGML_API struct ggml_tensor * ggml_sqr( + struct ggml_context * ctx, + struct ggml_tensor * a); + +GGML_API struct ggml_tensor * ggml_sqr_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a); + +GGML_API struct ggml_tensor * ggml_sqrt( + struct ggml_context * ctx, + struct ggml_tensor * a); + +GGML_API struct ggml_tensor * ggml_sqrt_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a); + +GGML_API struct ggml_tensor * ggml_log( + struct ggml_context * ctx, + struct ggml_tensor * a); + +GGML_API struct ggml_tensor * ggml_log_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a); + +// return scalar +GGML_API struct ggml_tensor * ggml_sum( + struct ggml_context * ctx, + struct ggml_tensor * a); + +// sums along rows, with input shape [a,b,c,d] return shape [1,b,c,d] +GGML_API struct ggml_tensor * ggml_sum_rows( + struct ggml_context * ctx, + struct ggml_tensor * a); + +// mean along rows +GGML_API struct ggml_tensor * ggml_mean( + struct ggml_context * ctx, + struct ggml_tensor * a); + +// argmax along rows +GGML_API struct ggml_tensor * ggml_argmax( + struct ggml_context * ctx, + struct ggml_tensor * a); + +// if a is the same shape as b, and a is not parameter, return a +// otherwise, return a new tensor: repeat(a) to fit in b +GGML_API struct ggml_tensor * ggml_repeat( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + +// sums repetitions in a into shape of b +GGML_API struct ggml_tensor * ggml_repeat_back( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + +// concat a and b on dim 2 +// used in stable-diffusion +GGML_API struct ggml_tensor * ggml_concat( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + +GGML_API struct ggml_tensor * ggml_abs( + struct ggml_context * ctx, + struct ggml_tensor * a); + +GGML_API struct ggml_tensor * ggml_abs_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a); + +GGML_API struct ggml_tensor * ggml_sgn( + struct ggml_context * ctx, + struct ggml_tensor * a); + +GGML_API struct ggml_tensor * ggml_sgn_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a); + +GGML_API struct ggml_tensor * ggml_neg( + struct ggml_context * ctx, + struct ggml_tensor * a); + +GGML_API struct ggml_tensor * ggml_neg_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a); + +GGML_API struct ggml_tensor * ggml_step( + struct ggml_context * ctx, + struct ggml_tensor * a); + +GGML_API struct ggml_tensor * ggml_step_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a); + +GGML_API struct ggml_tensor * ggml_tanh( + struct ggml_context * ctx, + struct ggml_tensor * a); + +GGML_API struct ggml_tensor * ggml_tanh_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a); + +GGML_API struct ggml_tensor * ggml_elu( + struct ggml_context * ctx, + struct ggml_tensor * a); + +GGML_API struct ggml_tensor * ggml_elu_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a); + +GGML_API struct ggml_tensor * ggml_relu( + struct ggml_context * ctx, + struct ggml_tensor * a); + +GGML_API struct ggml_tensor * ggml_leaky( + struct ggml_context * ctx, + struct ggml_tensor * a); + +GGML_API struct ggml_tensor * ggml_relu_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a); + +// TODO: double-check this computation is correct +GGML_API struct ggml_tensor * ggml_gelu( + struct ggml_context * ctx, + struct ggml_tensor * a); + +GGML_API struct ggml_tensor * ggml_gelu_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a); + +GGML_API struct ggml_tensor * ggml_gelu_quick( + struct ggml_context * ctx, + struct ggml_tensor * a); + +GGML_API struct ggml_tensor * ggml_gelu_quick_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a); + +GGML_API struct ggml_tensor * ggml_silu( + struct ggml_context * ctx, + struct ggml_tensor * a); + +GGML_API struct ggml_tensor * ggml_silu_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a); + +// a - x +// b - dy +GGML_API struct ggml_tensor * ggml_silu_back( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + +// normalize along rows +GGML_API struct ggml_tensor * ggml_norm( + struct ggml_context * ctx, + struct ggml_tensor * a, + float eps); + +GGML_API struct ggml_tensor * ggml_norm_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + float eps); + +GGML_API struct ggml_tensor * ggml_rms_norm( + struct ggml_context * ctx, + struct ggml_tensor * a, + float eps); + +GGML_API struct ggml_tensor * ggml_rms_norm_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + float eps); + +// group normalize along ne0*ne1*n_groups +// used in stable-diffusion +// TODO: eps is hardcoded to 1e-6 for now +GGML_API struct ggml_tensor * ggml_group_norm( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_groups); + +GGML_API struct ggml_tensor * ggml_group_norm_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_groups); + +// a - x +// b - dy +GGML_API struct ggml_tensor * ggml_rms_norm_back( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + float eps); + +// A: k columns, n rows => [ne03, ne02, n, k] +// B: k columns, m rows (i.e. we transpose it internally) => [ne03 * x, ne02 * y, m, k] +// result is n columns, m rows => [ne03 * x, ne02 * y, m, n] +GGML_API struct ggml_tensor * ggml_mul_mat( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + +// A: m columns, n rows, +// B: p columns, n rows, +// result is m columns, p rows +GGML_API struct ggml_tensor * ggml_out_prod( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + +// +// operations on tensors without backpropagation +// + +GGML_API struct ggml_tensor * ggml_scale( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + +// in-place, returns view(a) +GGML_API struct ggml_tensor * ggml_scale_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + +// b -> view(a,offset,nb1,nb2,3), return modified a +GGML_API struct ggml_tensor * ggml_set( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t nb2, + size_t nb3, + size_t offset); + +// b -> view(a,offset,nb1,nb2,3), return view(a) +GGML_API struct ggml_tensor * ggml_set_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t nb2, + size_t nb3, + size_t offset); + +GGML_API struct ggml_tensor * ggml_set_1d( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t offset); + +GGML_API struct ggml_tensor * ggml_set_1d_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t offset); + +// b -> view(a,offset,nb1,nb2,3), return modified a +GGML_API struct ggml_tensor * ggml_set_2d( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t offset); + +// b -> view(a,offset,nb1,nb2,3), return view(a) +GGML_API struct ggml_tensor * ggml_set_2d_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t offset); + +// a -> b, return view(b) +GGML_API struct ggml_tensor * ggml_cpy( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + +// a -> b, in-place, return view(b) +GGML_API struct ggml_tensor * ggml_cpy_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + +// make contiguous +GGML_API struct ggml_tensor * ggml_cont( + struct ggml_context * ctx, + struct ggml_tensor * a); + +// make contiguous, in-place +GGML_API struct ggml_tensor * ggml_cont_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a); + +// make contiguous, with new shape +GGML_API struct ggml_tensor * ggml_cont_1d( + struct ggml_context * ctx, + struct ggml_tensor * a, + int64_t ne0); + +GGML_API struct ggml_tensor * ggml_cont_2d( + struct ggml_context * ctx, + struct ggml_tensor * a, + int64_t ne0, + int64_t ne1); + +GGML_API struct ggml_tensor * ggml_cont_3d( + struct ggml_context * ctx, + struct ggml_tensor * a, + int64_t ne0, + int64_t ne1, + int64_t ne2); + +GGML_API struct ggml_tensor * ggml_cont_4d( + struct ggml_context * ctx, + struct ggml_tensor * a, + int64_t ne0, + int64_t ne1, + int64_t ne2, + int64_t ne3); + +// return view(a), b specifies the new shape +// TODO: when we start computing gradient, make a copy instead of view +GGML_API struct ggml_tensor * ggml_reshape( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + +// return view(a) +// TODO: when we start computing gradient, make a copy instead of view +GGML_API struct ggml_tensor * ggml_reshape_1d( + struct ggml_context * ctx, + struct ggml_tensor * a, + int64_t ne0); + +GGML_API struct ggml_tensor * ggml_reshape_2d( + struct ggml_context * ctx, + struct ggml_tensor * a, + int64_t ne0, + int64_t ne1); + +// return view(a) +// TODO: when we start computing gradient, make a copy instead of view +GGML_API struct ggml_tensor * ggml_reshape_3d( + struct ggml_context * ctx, + struct ggml_tensor * a, + int64_t ne0, + int64_t ne1, + int64_t ne2); + +GGML_API struct ggml_tensor * ggml_reshape_4d( + struct ggml_context * ctx, + struct ggml_tensor * a, + int64_t ne0, + int64_t ne1, + int64_t ne2, + int64_t ne3); + +// offset in bytes +GGML_API struct ggml_tensor * ggml_view_1d( + struct ggml_context * ctx, + struct ggml_tensor * a, + int64_t ne0, + size_t offset); + +GGML_API struct ggml_tensor * ggml_view_2d( + struct ggml_context * ctx, + struct ggml_tensor * a, + int64_t ne0, + int64_t ne1, + size_t nb1, // row stride in bytes + size_t offset); + +GGML_API struct ggml_tensor * ggml_view_3d( + struct ggml_context * ctx, + struct ggml_tensor * a, + int64_t ne0, + int64_t ne1, + int64_t ne2, + size_t nb1, // row stride in bytes + size_t nb2, // slice stride in bytes + size_t offset); + +GGML_API struct ggml_tensor * ggml_view_4d( + struct ggml_context * ctx, + struct ggml_tensor * a, + int64_t ne0, + int64_t ne1, + int64_t ne2, + int64_t ne3, + size_t nb1, // row stride in bytes + size_t nb2, // slice stride in bytes + size_t nb3, + size_t offset); + +GGML_API struct ggml_tensor * ggml_permute( + struct ggml_context * ctx, + struct ggml_tensor * a, + int axis0, + int axis1, + int axis2, + int axis3); + +// alias for ggml_permute(ctx, a, 1, 0, 2, 3) +GGML_API struct ggml_tensor * ggml_transpose( + struct ggml_context * ctx, + struct ggml_tensor * a); + +GGML_API struct ggml_tensor * ggml_get_rows( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + +GGML_API struct ggml_tensor * ggml_get_rows_back( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + struct ggml_tensor * c); + +GGML_API struct ggml_tensor * ggml_diag( struct ggml_context * ctx, struct ggml_tensor * a); - // set elements above the diagonal to -INF - GGML_API struct ggml_tensor * ggml_diag_mask_inf( - struct ggml_context * ctx, - struct ggml_tensor * a, - int n_past); +// set elements above the diagonal to -INF +GGML_API struct ggml_tensor * ggml_diag_mask_inf( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past); - // in-place, returns view(a) - GGML_API struct ggml_tensor * ggml_diag_mask_inf_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a, - int n_past); +// in-place, returns view(a) +GGML_API struct ggml_tensor * ggml_diag_mask_inf_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past); - // set elements above the diagonal to 0 - GGML_API struct ggml_tensor * ggml_diag_mask_zero( - struct ggml_context * ctx, - struct ggml_tensor * a, - int n_past); +// set elements above the diagonal to 0 +GGML_API struct ggml_tensor * ggml_diag_mask_zero( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past); - // in-place, returns view(a) - GGML_API struct ggml_tensor * ggml_diag_mask_zero_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a, - int n_past); +// in-place, returns view(a) +GGML_API struct ggml_tensor * ggml_diag_mask_zero_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past); - GGML_API struct ggml_tensor * ggml_soft_max( - struct ggml_context * ctx, - struct ggml_tensor * a); +GGML_API struct ggml_tensor * ggml_soft_max( + struct ggml_context * ctx, + struct ggml_tensor * a); - // in-place, returns view(a) - GGML_API struct ggml_tensor * ggml_soft_max_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a); +// in-place, returns view(a) +GGML_API struct ggml_tensor * ggml_soft_max_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a); - GGML_API struct ggml_tensor * ggml_soft_max_back( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); +GGML_API struct ggml_tensor * ggml_soft_max_back( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); - // in-place, returns view(a) - GGML_API struct ggml_tensor * ggml_soft_max_back_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); +// in-place, returns view(a) +GGML_API struct ggml_tensor * ggml_soft_max_back_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); - // rotary position embedding - // if mode & 1 == 1, skip n_past elements (DEPRECATED) - // if mode & 2 == 1, GPT-NeoX style - // if mode & 4 == 1, ChatGLM style - // - // b is an int32 vector with size a->ne[2], it contains the positions - GGML_API struct ggml_tensor * ggml_rope( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - int n_dims, - int mode, - int n_ctx); +// rotary position embedding +// if mode & 1 == 1, skip n_past elements (DEPRECATED) +// if mode & 2 == 1, GPT-NeoX style +// if mode & 4 == 1, ChatGLM style +// +// b is an int32 vector with size a->ne[2], it contains the positions +GGML_API struct ggml_tensor * ggml_rope( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + int n_dims, + int mode, + int n_ctx); - // in-place, returns view(a) - GGML_API struct ggml_tensor * ggml_rope_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - int n_dims, - int mode, - int n_ctx); +// in-place, returns view(a) +GGML_API struct ggml_tensor * ggml_rope_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + int n_dims, + int mode, + int n_ctx); - // custom RoPE - GGML_API struct ggml_tensor * ggml_rope_custom( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - int n_dims, - int mode, - int n_ctx, - float freq_base, - float freq_scale); +// custom RoPE +GGML_API struct ggml_tensor * ggml_rope_custom( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + int n_dims, + int mode, + int n_ctx, + int n_orig_ctx, + float freq_base, + float freq_scale, + float ext_factor, + float attn_factor, + float beta_fast, + float beta_slow); - // in-place, returns view(a) - GGML_API struct ggml_tensor * ggml_rope_custom_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - int n_dims, - int mode, - int n_ctx, - float freq_base, - float freq_scale); +// in-place, returns view(a) +GGML_API struct ggml_tensor * ggml_rope_custom_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + int n_dims, + int mode, + int n_ctx, + int n_orig_ctx, + float freq_base, + float freq_scale, + float ext_factor, + float attn_factor, + float beta_fast, + float beta_slow); - // xPos RoPE, in-place, returns view(a) - GGML_API struct ggml_tensor * ggml_rope_xpos_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - int n_dims, - float base, - bool down); +// compute correction dims for YaRN RoPE scaling +void ggml_rope_yarn_corr_dims( + int n_dims, int n_orig_ctx, float freq_base, float beta_fast, float beta_slow, float dims[2]); - // rotary position embedding backward, i.e compute dx from dy - // a - dy - GGML_API struct ggml_tensor * ggml_rope_back( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - int n_dims, - int mode, - int n_ctx, - float freq_base, - float freq_scale, - float xpos_base, - bool xpos_down); +// xPos RoPE, in-place, returns view(a) +GGML_API struct ggml_tensor * ggml_rope_xpos_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + int n_dims, + float base, + bool down); - // alibi position embedding - // in-place, returns view(a) - GGML_API struct ggml_tensor * ggml_alibi( - struct ggml_context * ctx, - struct ggml_tensor * a, - int n_past, - int n_head, - float bias_max); +// rotary position embedding backward, i.e compute dx from dy +// a - dy +GGML_API struct ggml_tensor * ggml_rope_back( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + int n_dims, + int mode, + int n_ctx, + int n_orig_ctx, + float freq_base, + float freq_scale, + float ext_factor, + float attn_factor, + float beta_fast, + float beta_slow, + float xpos_base, + bool xpos_down); - // clamp - // in-place, returns view(a) - GGML_API struct ggml_tensor * ggml_clamp( - struct ggml_context * ctx, - struct ggml_tensor * a, - float min, - float max); +// alibi position embedding +// in-place, returns view(a) +GGML_API struct ggml_tensor * ggml_alibi( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past, + int n_head, + float bias_max); - GGML_API struct ggml_tensor * ggml_conv_1d( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - int s0, // stride - int p0, // padding - int d0); // dilation +// clamp +// in-place, returns view(a) +GGML_API struct ggml_tensor * ggml_clamp( + struct ggml_context * ctx, + struct ggml_tensor * a, + float min, + float max); - // conv_1d with padding = half - // alias for ggml_conv_1d(a, b, s, a->ne[0]/2, d) - GGML_API struct ggml_tensor* ggml_conv_1d_ph( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - int s, - int d); +GGML_API struct ggml_tensor * ggml_im2col( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + int s0, + int s1, + int p0, + int p1, + int d0, + int d1, + bool is_2D); - GGML_API struct ggml_tensor * ggml_conv_transpose_1d( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - int s0, - int p0, - int d0); +GGML_API struct ggml_tensor * ggml_conv_1d( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + int s0, // stride + int p0, // padding + int d0); // dilation - GGML_API struct ggml_tensor * ggml_conv_2d( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - int s0, - int s1, - int p0, - int p1, - int d0, - int d1); +// conv_1d with padding = half +// alias for ggml_conv_1d(a, b, s, a->ne[0]/2, d) +GGML_API struct ggml_tensor* ggml_conv_1d_ph( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + int s, + int d); + +GGML_API struct ggml_tensor * ggml_conv_transpose_1d( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + int s0, + int p0, + int d0); + +GGML_API struct ggml_tensor * ggml_conv_2d( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + int s0, + int s1, + int p0, + int p1, + int d0, + int d1); - // kernel size is a->ne[0] x a->ne[1] - // stride is equal to kernel size - // padding is zero - // example: - // a: 16 16 3 768 - // b: 1024 1024 3 1 - // res: 64 64 768 1 - // used in sam - GGML_API struct ggml_tensor * ggml_conv_2d_sk_p0( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); +// kernel size is a->ne[0] x a->ne[1] +// stride is equal to kernel size +// padding is zero +// example: +// a: 16 16 3 768 +// b: 1024 1024 3 1 +// res: 64 64 768 1 +// used in sam +GGML_API struct ggml_tensor * ggml_conv_2d_sk_p0( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); - // kernel size is a->ne[0] x a->ne[1] - // stride is 1 - // padding is half - // example: - // a: 3 3 256 256 - // b: 64 64 256 1 - // res: 64 64 256 1 - // used in sam - GGML_API struct ggml_tensor * ggml_conv_2d_s1_ph( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); +// kernel size is a->ne[0] x a->ne[1] +// stride is 1 +// padding is half +// example: +// a: 3 3 256 256 +// b: 64 64 256 1 +// res: 64 64 256 1 +// used in sam +GGML_API struct ggml_tensor * ggml_conv_2d_s1_ph( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); - GGML_API struct ggml_tensor * ggml_conv_transpose_2d_p0( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - int stride); +GGML_API struct ggml_tensor * ggml_conv_transpose_2d_p0( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + int stride); - enum ggml_op_pool { - GGML_OP_POOL_MAX, - GGML_OP_POOL_AVG, - GGML_OP_POOL_COUNT, - }; +enum ggml_op_pool { + GGML_OP_POOL_MAX, + GGML_OP_POOL_AVG, + GGML_OP_POOL_COUNT, +}; - GGML_API struct ggml_tensor * ggml_pool_1d( - struct ggml_context * ctx, - struct ggml_tensor * a, - enum ggml_op_pool op, - int k0, // kernel size - int s0, // stride - int p0); // padding +GGML_API struct ggml_tensor * ggml_pool_1d( + struct ggml_context * ctx, + struct ggml_tensor * a, + enum ggml_op_pool op, + int k0, // kernel size + int s0, // stride + int p0); // padding - GGML_API struct ggml_tensor * ggml_pool_2d( - struct ggml_context * ctx, - struct ggml_tensor * a, - enum ggml_op_pool op, - int k0, - int k1, - int s0, - int s1, - int p0, - int p1); +// the result will have 2*p0 padding for the first dimension +// and 2*p1 padding for the second dimension +GGML_API struct ggml_tensor * ggml_pool_2d( + struct ggml_context * ctx, + struct ggml_tensor * a, + enum ggml_op_pool op, + int k0, + int k1, + int s0, + int s1, + float p0, + float p1); - // nearest interpolate - // used in stable-diffusion - GGML_API struct ggml_tensor * ggml_upscale( - struct ggml_context * ctx, - struct ggml_tensor * a, - int scale_factor); +// nearest interpolate +// used in stable-diffusion +GGML_API struct ggml_tensor * ggml_upscale( + struct ggml_context * ctx, + struct ggml_tensor * a, + int scale_factor); - GGML_API struct ggml_tensor * ggml_flash_attn( - struct ggml_context * ctx, - struct ggml_tensor * q, - struct ggml_tensor * k, - struct ggml_tensor * v, - bool masked); +GGML_API struct ggml_tensor * ggml_flash_attn( + struct ggml_context * ctx, + struct ggml_tensor * q, + struct ggml_tensor * k, + struct ggml_tensor * v, + bool masked); - GGML_API struct ggml_tensor * ggml_flash_attn_back( - struct ggml_context * ctx, - struct ggml_tensor * q, - struct ggml_tensor * k, - struct ggml_tensor * v, - struct ggml_tensor * d, - bool masked); +GGML_API struct ggml_tensor * ggml_flash_attn_back( + struct ggml_context * ctx, + struct ggml_tensor * q, + struct ggml_tensor * k, + struct ggml_tensor * v, + struct ggml_tensor * d, + bool masked); - GGML_API struct ggml_tensor * ggml_flash_ff( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b0, - struct ggml_tensor * b1, - struct ggml_tensor * c0, - struct ggml_tensor * c1); +GGML_API struct ggml_tensor * ggml_flash_ff( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b0, + struct ggml_tensor * b1, + struct ggml_tensor * c0, + struct ggml_tensor * c1); - // partition into non-overlapping windows with padding if needed - // example: - // a: 768 64 64 1 - // w: 14 - // res: 768 14 14 25 - // used in sam - GGML_API struct ggml_tensor * ggml_win_part( - struct ggml_context * ctx, - struct ggml_tensor * a, - int w); +// partition into non-overlapping windows with padding if needed +// example: +// a: 768 64 64 1 +// w: 14 +// res: 768 14 14 25 +// used in sam +GGML_API struct ggml_tensor * ggml_win_part( + struct ggml_context * ctx, + struct ggml_tensor * a, + int w); - // reverse of ggml_win_part - // used in sam - GGML_API struct ggml_tensor * ggml_win_unpart( - struct ggml_context * ctx, - struct ggml_tensor * a, - int w0, - int h0, - int w); +// reverse of ggml_win_part +// used in sam +GGML_API struct ggml_tensor * ggml_win_unpart( + struct ggml_context * ctx, + struct ggml_tensor * a, + int w0, + int h0, + int w); - GGML_API struct ggml_tensor * ggml_unary( - struct ggml_context * ctx, - struct ggml_tensor * a, - enum ggml_unary_op op); +GGML_API struct ggml_tensor * ggml_unary( + struct ggml_context * ctx, + struct ggml_tensor * a, + enum ggml_unary_op op); - GGML_API struct ggml_tensor * ggml_unary_inplace( +GGML_API struct ggml_tensor * ggml_unary_inplace( struct ggml_context * ctx, struct ggml_tensor * a, enum ggml_unary_op op); - // used in sam - GGML_API struct ggml_tensor * ggml_get_rel_pos( - struct ggml_context * ctx, - struct ggml_tensor * a, - int qh, - int kh); +// used in sam +GGML_API struct ggml_tensor * ggml_get_rel_pos( + struct ggml_context * ctx, + struct ggml_tensor * a, + int qh, + int kh); - // used in sam +// used in sam - GGML_API struct ggml_tensor * ggml_add_rel_pos( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * pw, - struct ggml_tensor * ph); +GGML_API struct ggml_tensor * ggml_add_rel_pos( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * pw, + struct ggml_tensor * ph); - GGML_API struct ggml_tensor * ggml_add_rel_pos_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * pw, - struct ggml_tensor * ph); +GGML_API struct ggml_tensor * ggml_add_rel_pos_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * pw, + struct ggml_tensor * ph); - // custom operators +// custom operators - typedef void (*ggml_unary_op_f32_t) (const int, float *, const float *); - typedef void (*ggml_binary_op_f32_t)(const int, float *, const float *, const float *); +typedef void (*ggml_unary_op_f32_t) (const int, float *, const float *); +typedef void (*ggml_binary_op_f32_t)(const int, float *, const float *, const float *); - typedef void (*ggml_custom1_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *); - typedef void (*ggml_custom2_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *); - typedef void (*ggml_custom3_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *); +typedef void (*ggml_custom1_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *); +typedef void (*ggml_custom2_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *); +typedef void (*ggml_custom3_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *); - GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_unary_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - ggml_unary_op_f32_t fun), - "use ggml_map_custom1 instead"); +GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_unary_f32( + struct ggml_context * ctx, + struct ggml_tensor * a, + ggml_unary_op_f32_t fun), + "use ggml_map_custom1 instead"); - GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_unary_inplace_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - ggml_unary_op_f32_t fun), - "use ggml_map_custom1_inplace instead"); +GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_unary_inplace_f32( + struct ggml_context * ctx, + struct ggml_tensor * a, + ggml_unary_op_f32_t fun), + "use ggml_map_custom1_inplace instead"); - GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_binary_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - ggml_binary_op_f32_t fun), - "use ggml_map_custom2 instead"); +GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_binary_f32( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + ggml_binary_op_f32_t fun), + "use ggml_map_custom2 instead"); - GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_binary_inplace_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - ggml_binary_op_f32_t fun), - "use ggml_map_custom2_inplace instead"); +GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_binary_inplace_f32( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + ggml_binary_op_f32_t fun), + "use ggml_map_custom2_inplace instead"); - GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom1_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - ggml_custom1_op_f32_t fun), - "use ggml_map_custom1 instead"); +GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom1_f32( + struct ggml_context * ctx, + struct ggml_tensor * a, + ggml_custom1_op_f32_t fun), + "use ggml_map_custom1 instead"); - GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom1_inplace_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - ggml_custom1_op_f32_t fun), - "use ggml_map_custom1_inplace instead"); +GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom1_inplace_f32( + struct ggml_context * ctx, + struct ggml_tensor * a, + ggml_custom1_op_f32_t fun), + "use ggml_map_custom1_inplace instead"); - GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom2_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - ggml_custom2_op_f32_t fun), - "use ggml_map_custom2 instead"); +GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom2_f32( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + ggml_custom2_op_f32_t fun), + "use ggml_map_custom2 instead"); - GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom2_inplace_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - ggml_custom2_op_f32_t fun), - "use ggml_map_custom2_inplace instead"); +GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom2_inplace_f32( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + ggml_custom2_op_f32_t fun), + "use ggml_map_custom2_inplace instead"); - GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom3_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - struct ggml_tensor * c, - ggml_custom3_op_f32_t fun), - "use ggml_map_custom3 instead"); +GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom3_f32( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + struct ggml_tensor * c, + ggml_custom3_op_f32_t fun), + "use ggml_map_custom3 instead"); - GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom3_inplace_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - struct ggml_tensor * c, - ggml_custom3_op_f32_t fun), - "use ggml_map_custom3_inplace instead"); +GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom3_inplace_f32( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + struct ggml_tensor * c, + ggml_custom3_op_f32_t fun), + "use ggml_map_custom3_inplace instead"); - // custom operators v2 +// custom operators v2 - typedef void (*ggml_custom1_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, int ith, int nth, void * userdata); - typedef void (*ggml_custom2_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, const struct ggml_tensor * b, int ith, int nth, void * userdata); - typedef void (*ggml_custom3_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, const struct ggml_tensor * b, const struct ggml_tensor * c, int ith, int nth, void * userdata); +typedef void (*ggml_custom1_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, int ith, int nth, void * userdata); +typedef void (*ggml_custom2_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, const struct ggml_tensor * b, int ith, int nth, void * userdata); +typedef void (*ggml_custom3_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, const struct ggml_tensor * b, const struct ggml_tensor * c, int ith, int nth, void * userdata); - #define GGML_N_TASKS_MAX -1 +#define GGML_N_TASKS_MAX -1 - GGML_API struct ggml_tensor * ggml_map_custom1( - struct ggml_context * ctx, - struct ggml_tensor * a, - ggml_custom1_op_t fun, - int n_tasks, - void * userdata); +GGML_API struct ggml_tensor * ggml_map_custom1( + struct ggml_context * ctx, + struct ggml_tensor * a, + ggml_custom1_op_t fun, + int n_tasks, + void * userdata); - GGML_API struct ggml_tensor * ggml_map_custom1_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a, - ggml_custom1_op_t fun, - int n_tasks, - void * userdata); +GGML_API struct ggml_tensor * ggml_map_custom1_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + ggml_custom1_op_t fun, + int n_tasks, + void * userdata); - GGML_API struct ggml_tensor * ggml_map_custom2( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - ggml_custom2_op_t fun, - int n_tasks, - void * userdata); +GGML_API struct ggml_tensor * ggml_map_custom2( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + ggml_custom2_op_t fun, + int n_tasks, + void * userdata); - GGML_API struct ggml_tensor * ggml_map_custom2_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - ggml_custom2_op_t fun, - int n_tasks, - void * userdata); +GGML_API struct ggml_tensor * ggml_map_custom2_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + ggml_custom2_op_t fun, + int n_tasks, + void * userdata); - GGML_API struct ggml_tensor * ggml_map_custom3( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - struct ggml_tensor * c, - ggml_custom3_op_t fun, - int n_tasks, - void * userdata); +GGML_API struct ggml_tensor * ggml_map_custom3( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + struct ggml_tensor * c, + ggml_custom3_op_t fun, + int n_tasks, + void * userdata); - GGML_API struct ggml_tensor * ggml_map_custom3_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - struct ggml_tensor * c, - ggml_custom3_op_t fun, - int n_tasks, - void * userdata); +GGML_API struct ggml_tensor * ggml_map_custom3_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + struct ggml_tensor * c, + ggml_custom3_op_t fun, + int n_tasks, + void * userdata); - // loss function +// loss function - GGML_API struct ggml_tensor * ggml_cross_entropy_loss( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); +GGML_API struct ggml_tensor * ggml_cross_entropy_loss( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); - GGML_API struct ggml_tensor * ggml_cross_entropy_loss_back( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - struct ggml_tensor * c); +GGML_API struct ggml_tensor * ggml_cross_entropy_loss_back( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + struct ggml_tensor * c); +// +// automatic differentiation +// + +GGML_API void ggml_set_param( + struct ggml_context * ctx, + struct ggml_tensor * tensor); + + +GGML_API void ggml_build_forward_expand (struct ggml_cgraph * cgraph, struct ggml_tensor * tensor); +GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); + +// graph allocation in a context +GGML_API struct ggml_cgraph * ggml_new_graph (struct ggml_context * ctx); // size = GGML_DEFAULT_GRAPH_SIZE, grads = false +GGML_API struct ggml_cgraph * ggml_new_graph_custom (struct ggml_context * ctx, size_t size, bool grads); +GGML_API struct ggml_cgraph * ggml_graph_dup (struct ggml_context * ctx, struct ggml_cgraph * cgraph); +GGML_API struct ggml_cgraph * ggml_graph_view (struct ggml_context * ctx, struct ggml_cgraph * cgraph, int i0, int i1); +GGML_API void ggml_graph_cpy (struct ggml_cgraph * src, struct ggml_cgraph * dst); +GGML_API void ggml_graph_reset (struct ggml_cgraph * cgraph); // zero grads +GGML_API void ggml_graph_clear (struct ggml_cgraph * cgraph); + +GGML_API size_t ggml_graph_overhead(void); +GGML_API size_t ggml_graph_overhead_custom(size_t size, bool grads); + +// ggml_graph_plan() has to be called before ggml_graph_compute() +// when plan.work_size > 0, caller must allocate memory for plan.work_data +GGML_API struct ggml_cplan ggml_graph_plan (struct ggml_cgraph * cgraph, int n_threads /*= GGML_DEFAULT_N_THREADS*/); +GGML_API int ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan); + +// same as ggml_graph_compute() but the work data is allocated as a part of the context +// note: the drawback of this API is that you must have ensured that the context has enough memory for the work data +GGML_API void ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads); + +GGML_API struct ggml_tensor * ggml_graph_get_tensor(struct ggml_cgraph * cgraph, const char * name); + +GGML_API void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname); +GGML_API struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context ** ctx_data, struct ggml_context ** ctx_eval); + +// print info and performance information for the graph +GGML_API void ggml_graph_print(const struct ggml_cgraph * cgraph); + +// dump the graph into a file using the dot format +GGML_API void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph * gf, const char * filename); + +// build gradient checkpointing backward graph gb for gf using provided checkpoints +// gb_tmp will contain original backward graph with rewritten backward process nodes, +// but without the second forward pass nodes. +GGML_API void ggml_build_backward_gradient_checkpointing( + struct ggml_context * ctx, + struct ggml_cgraph * gf, + struct ggml_cgraph * gb, + struct ggml_cgraph * gb_tmp, + struct ggml_tensor * * checkpoints, + int n_checkpoints); +// +// optimization +// + +// optimization methods +enum ggml_opt_type { + GGML_OPT_ADAM, + GGML_OPT_LBFGS, +}; + +// linesearch methods +enum ggml_linesearch { + GGML_LINESEARCH_DEFAULT = 1, + + GGML_LINESEARCH_BACKTRACKING_ARMIJO = 0, + GGML_LINESEARCH_BACKTRACKING_WOLFE = 1, + GGML_LINESEARCH_BACKTRACKING_STRONG_WOLFE = 2, +}; + +// optimization return values +enum ggml_opt_result { + GGML_OPT_OK = 0, + GGML_OPT_DID_NOT_CONVERGE, + GGML_OPT_NO_CONTEXT, + GGML_OPT_INVALID_WOLFE, + GGML_OPT_FAIL, + GGML_OPT_CANCEL, + + GGML_LINESEARCH_FAIL = -128, + GGML_LINESEARCH_MINIMUM_STEP, + GGML_LINESEARCH_MAXIMUM_STEP, + GGML_LINESEARCH_MAXIMUM_ITERATIONS, + GGML_LINESEARCH_INVALID_PARAMETERS, +}; + +typedef void (*ggml_opt_callback)(void * data, int accum_step, float * sched, bool * cancel); +typedef void (*ggml_log_callback)(enum ggml_log_level level, const char * text, void * user_data); + +// optimization parameters +// +// see ggml.c (ggml_opt_default_params) for default values +// +struct ggml_opt_params { + enum ggml_opt_type type; + + size_t graph_size; + + int n_threads; + + // delta-based convergence test // - // automatic differentiation + // if past == 0 - disabled + // if past > 0: + // stop if |f(x) - f(x_past)| < delta * max(1, |f(x)|) // + int past; + float delta; - GGML_API void ggml_set_param( - struct ggml_context * ctx, - struct ggml_tensor * tensor); - - - GGML_API void ggml_build_forward_expand (struct ggml_cgraph * cgraph, struct ggml_tensor * tensor); - GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); - - GGML_API struct ggml_cgraph ggml_build_forward (struct ggml_tensor * tensor); - GGML_API struct ggml_cgraph ggml_build_backward(struct ggml_context * ctx, struct ggml_cgraph * gf, bool keep); - - // graph allocation in a context - GGML_API struct ggml_cgraph * ggml_new_graph (struct ggml_context * ctx); - GGML_API struct ggml_cgraph * ggml_build_forward_ctx(struct ggml_context * ctx, struct ggml_tensor * tensor); - GGML_API size_t ggml_graph_overhead(void); - - // ggml_graph_plan() has to be called before ggml_graph_compute() - // when plan.work_size > 0, caller must allocate memory for plan.work_data - GGML_API struct ggml_cplan ggml_graph_plan (struct ggml_cgraph * cgraph, int n_threads /*= GGML_DEFAULT_N_THREADS*/); - GGML_API int ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan); - GGML_API void ggml_graph_reset (struct ggml_cgraph * cgraph); - - // same as ggml_graph_compute() but the work data is allocated as a part of the context - // note: the drawback of this API is that you must have ensured that the context has enough memory for the work data - GGML_API void ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads); - - GGML_API struct ggml_tensor * ggml_graph_get_tensor(struct ggml_cgraph * cgraph, const char * name); - - GGML_API void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname); - GGML_API struct ggml_cgraph ggml_graph_import(const char * fname, struct ggml_context ** ctx_data, struct ggml_context ** ctx_eval); - - // print info and performance information for the graph - GGML_API void ggml_graph_print(const struct ggml_cgraph * cgraph); - - // dump the graph into a file using the dot format - GGML_API void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph * gf, const char * filename); - - // build gradient checkpointing backward graph gb for gf using provided checkpoints - // gb_tmp will contain original backward graph with rewritten backward process nodes, - // but without the second forward pass nodes. - GGML_API void ggml_build_backward_gradient_checkpointing( - struct ggml_context * ctx, - struct ggml_cgraph * gf, - struct ggml_cgraph * gb, - struct ggml_cgraph * gb_tmp, - struct ggml_tensor * * checkpoints, - int n_checkpoints); + // maximum number of iterations without improvement // - // optimization + // if 0 - disabled + // if > 0: + // assume convergence if no cost improvement in this number of iterations // + int max_no_improvement; - // optimization methods - enum ggml_opt_type { - GGML_OPT_ADAM, - GGML_OPT_LBFGS, - }; + bool print_forward_graph; + bool print_backward_graph; - // linesearch methods - enum ggml_linesearch { - GGML_LINESEARCH_DEFAULT = 1, + int n_gradient_accumulation; - GGML_LINESEARCH_BACKTRACKING_ARMIJO = 0, - GGML_LINESEARCH_BACKTRACKING_WOLFE = 1, - GGML_LINESEARCH_BACKTRACKING_STRONG_WOLFE = 2, - }; + // ADAM parameters + struct { + int n_iter; - // optimization return values - enum ggml_opt_result { - GGML_OPT_OK = 0, - GGML_OPT_DID_NOT_CONVERGE, - GGML_OPT_NO_CONTEXT, - GGML_OPT_INVALID_WOLFE, - GGML_OPT_FAIL, - GGML_OPT_CANCEL, + float sched; // schedule multiplier (fixed, decay or warmup) + float decay; // weight decay for AdamW, use 0.0f to disable + int decay_min_ndim; // minimum number of tensor dimension to apply weight decay + float alpha; // learning rate + float beta1; + float beta2; + float eps; // epsilon for numerical stability + float eps_f; // epsilon for convergence test + float eps_g; // epsilon for convergence test + float gclip; // gradient clipping + } adam; - GGML_LINESEARCH_FAIL = -128, - GGML_LINESEARCH_MINIMUM_STEP, - GGML_LINESEARCH_MAXIMUM_STEP, - GGML_LINESEARCH_MAXIMUM_ITERATIONS, - GGML_LINESEARCH_INVALID_PARAMETERS, - }; + // LBFGS parameters + struct { + int m; // number of corrections to approximate the inv. Hessian + int n_iter; + int max_linesearch; - typedef void (*ggml_opt_callback)(void * data, int accum_step, float * sched, bool * cancel); - typedef void (*ggml_log_callback)(enum ggml_log_level level, const char * text, void * user_data); + float eps; // convergence tolerance + float ftol; // line search tolerance + float wolfe; + float min_step; + float max_step; - // optimization parameters - // - // see ggml.c (ggml_opt_default_params) for default values - // - struct ggml_opt_params { - enum ggml_opt_type type; + enum ggml_linesearch linesearch; + } lbfgs; +}; - int n_threads; +struct ggml_opt_context { + struct ggml_context * ctx; + struct ggml_opt_params params; - // delta-based convergence test - // - // if past == 0 - disabled - // if past > 0: - // stop if |f(x) - f(x_past)| < delta * max(1, |f(x)|) - // - int past; - float delta; + int iter; + int64_t nx; // number of parameter elements - // maximum number of iterations without improvement - // - // if 0 - disabled - // if > 0: - // assume convergence if no cost improvement in this number of iterations - // - int max_no_improvement; + bool just_initialized; - bool print_forward_graph; - bool print_backward_graph; + float loss_before; + float loss_after; - int n_gradient_accumulation; + struct { + struct ggml_tensor * g; // current gradient + struct ggml_tensor * m; // first moment + struct ggml_tensor * v; // second moment + struct ggml_tensor * pf; // past function values + float fx_best; + float fx_prev; + int n_no_improvement; + } adam; - // ADAM parameters - struct { - int n_iter; + struct { + struct ggml_tensor * x; // current parameters + struct ggml_tensor * xp; // previous parameters + struct ggml_tensor * g; // current gradient + struct ggml_tensor * gp; // previous gradient + struct ggml_tensor * d; // search direction + struct ggml_tensor * pf; // past function values + struct ggml_tensor * lmal; // the L-BFGS memory alpha + struct ggml_tensor * lmys; // the L-BFGS memory ys + struct ggml_tensor * lms; // the L-BFGS memory s + struct ggml_tensor * lmy; // the L-BFGS memory y + float fx_best; + float step; + int j; + int k; + int end; + int n_no_improvement; + } lbfgs; +}; - float sched; // schedule multiplier (fixed, decay or warmup) - float decay; // weight decay for AdamW, use 0.0f to disable - int decay_min_ndim; // minimum number of tensor dimension to apply weight decay - float alpha; // learning rate - float beta1; - float beta2; - float eps; // epsilon for numerical stability - float eps_f; // epsilon for convergence test - float eps_g; // epsilon for convergence test - float gclip; // gradient clipping - } adam; +GGML_API struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type); - // LBFGS parameters - struct { - int m; // number of corrections to approximate the inv. Hessian - int n_iter; - int max_linesearch; +// optimize the function defined by the tensor f +GGML_API enum ggml_opt_result ggml_opt( + struct ggml_context * ctx, + struct ggml_opt_params params, + struct ggml_tensor * f); - float eps; // convergence tolerance - float ftol; // line search tolerance - float wolfe; - float min_step; - float max_step; +// initialize optimizer context +GGML_API void ggml_opt_init( + struct ggml_context * ctx, + struct ggml_opt_context * opt, + struct ggml_opt_params params, + int64_t nx); - enum ggml_linesearch linesearch; - } lbfgs; - }; +// continue optimizing the function defined by the tensor f +GGML_API enum ggml_opt_result ggml_opt_resume( + struct ggml_context * ctx, + struct ggml_opt_context * opt, + struct ggml_tensor * f); - struct ggml_opt_context { - struct ggml_context * ctx; - struct ggml_opt_params params; +// continue optimizing the function defined by the tensor f +GGML_API enum ggml_opt_result ggml_opt_resume_g( + struct ggml_context * ctx, + struct ggml_opt_context * opt, + struct ggml_tensor * f, + struct ggml_cgraph * gf, + struct ggml_cgraph * gb, + ggml_opt_callback callback, + void * callback_data); - int iter; - int64_t nx; // number of parameter elements +// +// quantization +// - bool just_initialized; +// TODO: these would probably get removed in favor of the more general ggml_quantize_chunk +GGML_API size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * hist); +GGML_API size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * hist); +GGML_API size_t ggml_quantize_q5_0(const float * src, void * dst, int n, int k, int64_t * hist); +GGML_API size_t ggml_quantize_q5_1(const float * src, void * dst, int n, int k, int64_t * hist); +GGML_API size_t ggml_quantize_q8_0(const float * src, void * dst, int n, int k, int64_t * hist); - float loss_before; - float loss_after; +GGML_API size_t ggml_quantize_q2_K(const float * src, void * dst, int n, int k, int64_t * hist); +GGML_API size_t ggml_quantize_q3_K(const float * src, void * dst, int n, int k, int64_t * hist); +GGML_API size_t ggml_quantize_q4_K(const float * src, void * dst, int n, int k, int64_t * hist); +GGML_API size_t ggml_quantize_q5_K(const float * src, void * dst, int n, int k, int64_t * hist); +GGML_API size_t ggml_quantize_q6_K(const float * src, void * dst, int n, int k, int64_t * hist); - struct { - struct ggml_tensor * g; // current gradient - struct ggml_tensor * m; // first moment - struct ggml_tensor * v; // second moment - struct ggml_tensor * pf; // past function values - float fx_best; - float fx_prev; - int n_no_improvement; - } adam; +GGML_API size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, int start, int n, int64_t * hist); - struct { - struct ggml_tensor * x; // current parameters - struct ggml_tensor * xp; // previous parameters - struct ggml_tensor * g; // current gradient - struct ggml_tensor * gp; // previous gradient - struct ggml_tensor * d; // search direction - struct ggml_tensor * pf; // past function values - struct ggml_tensor * lmal; // the L-BFGS memory alpha - struct ggml_tensor * lmys; // the L-BFGS memory ys - struct ggml_tensor * lms; // the L-BFGS memory s - struct ggml_tensor * lmy; // the L-BFGS memory y - float fx_best; - float step; - int j; - int k; - int end; - int n_no_improvement; - } lbfgs; - }; +// +// gguf +// - GGML_API struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type); +enum gguf_type { + GGUF_TYPE_UINT8 = 0, + GGUF_TYPE_INT8 = 1, + GGUF_TYPE_UINT16 = 2, + GGUF_TYPE_INT16 = 3, + GGUF_TYPE_UINT32 = 4, + GGUF_TYPE_INT32 = 5, + GGUF_TYPE_FLOAT32 = 6, + GGUF_TYPE_BOOL = 7, + GGUF_TYPE_STRING = 8, + GGUF_TYPE_ARRAY = 9, + GGUF_TYPE_UINT64 = 10, + GGUF_TYPE_INT64 = 11, + GGUF_TYPE_FLOAT64 = 12, + GGUF_TYPE_COUNT, // marks the end of the enum +}; - // optimize the function defined by the tensor f - GGML_API enum ggml_opt_result ggml_opt( - struct ggml_context * ctx, - struct ggml_opt_params params, - struct ggml_tensor * f); +struct gguf_context; - // initialize optimizer context - GGML_API void ggml_opt_init( - struct ggml_context * ctx, - struct ggml_opt_context * opt, - struct ggml_opt_params params, - int64_t nx); +struct gguf_init_params { + bool no_alloc; - // continue optimizing the function defined by the tensor f - GGML_API enum ggml_opt_result ggml_opt_resume( - struct ggml_context * ctx, - struct ggml_opt_context * opt, - struct ggml_tensor * f); + // if not NULL, create a ggml_context and allocate the tensor data in it + struct ggml_context ** ctx; +}; - // continue optimizing the function defined by the tensor f - GGML_API enum ggml_opt_result ggml_opt_resume_g( - struct ggml_context * ctx, - struct ggml_opt_context * opt, - struct ggml_tensor * f, - struct ggml_cgraph * gf, - struct ggml_cgraph * gb, - ggml_opt_callback callback, - void * callback_data); +GGML_API struct gguf_context * gguf_init_empty(void); +GGML_API struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_params params); +//GGML_API struct gguf_context * gguf_init_from_buffer(..); - // - // quantization - // +GGML_API void gguf_free(struct gguf_context * ctx); - // TODO: these would probably get removed in favor of the more general ggml_quantize_chunk - GGML_API size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * hist); - GGML_API size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * hist); - GGML_API size_t ggml_quantize_q5_0(const float * src, void * dst, int n, int k, int64_t * hist); - GGML_API size_t ggml_quantize_q5_1(const float * src, void * dst, int n, int k, int64_t * hist); - GGML_API size_t ggml_quantize_q8_0(const float * src, void * dst, int n, int k, int64_t * hist); +GGML_API const char * gguf_type_name(enum gguf_type type); - GGML_API size_t ggml_quantize_q2_K(const float * src, void * dst, int n, int k, int64_t * hist); - GGML_API size_t ggml_quantize_q3_K(const float * src, void * dst, int n, int k, int64_t * hist); - GGML_API size_t ggml_quantize_q4_K(const float * src, void * dst, int n, int k, int64_t * hist); - GGML_API size_t ggml_quantize_q5_K(const float * src, void * dst, int n, int k, int64_t * hist); - GGML_API size_t ggml_quantize_q6_K(const float * src, void * dst, int n, int k, int64_t * hist); +GGML_API int gguf_get_version (const struct gguf_context * ctx); +GGML_API size_t gguf_get_alignment (const struct gguf_context * ctx); +GGML_API size_t gguf_get_data_offset(const struct gguf_context * ctx); +GGML_API void * gguf_get_data (const struct gguf_context * ctx); - GGML_API size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, int start, int n, int64_t * hist); +GGML_API int gguf_get_n_kv(const struct gguf_context * ctx); +GGML_API int gguf_find_key(const struct gguf_context * ctx, const char * key); +GGML_API const char * gguf_get_key (const struct gguf_context * ctx, int key_id); - // - // gguf - // +GGML_API enum gguf_type gguf_get_kv_type (const struct gguf_context * ctx, int key_id); +GGML_API enum gguf_type gguf_get_arr_type(const struct gguf_context * ctx, int key_id); - enum gguf_type { - GGUF_TYPE_UINT8 = 0, - GGUF_TYPE_INT8 = 1, - GGUF_TYPE_UINT16 = 2, - GGUF_TYPE_INT16 = 3, - GGUF_TYPE_UINT32 = 4, - GGUF_TYPE_INT32 = 5, - GGUF_TYPE_FLOAT32 = 6, - GGUF_TYPE_BOOL = 7, - GGUF_TYPE_STRING = 8, - GGUF_TYPE_ARRAY = 9, - GGUF_TYPE_UINT64 = 10, - GGUF_TYPE_INT64 = 11, - GGUF_TYPE_FLOAT64 = 12, - GGUF_TYPE_COUNT, // marks the end of the enum - }; +// will abort if the wrong type is used for the key +GGML_API uint8_t gguf_get_val_u8 (const struct gguf_context * ctx, int key_id); +GGML_API int8_t gguf_get_val_i8 (const struct gguf_context * ctx, int key_id); +GGML_API uint16_t gguf_get_val_u16 (const struct gguf_context * ctx, int key_id); +GGML_API int16_t gguf_get_val_i16 (const struct gguf_context * ctx, int key_id); +GGML_API uint32_t gguf_get_val_u32 (const struct gguf_context * ctx, int key_id); +GGML_API int32_t gguf_get_val_i32 (const struct gguf_context * ctx, int key_id); +GGML_API float gguf_get_val_f32 (const struct gguf_context * ctx, int key_id); +GGML_API uint64_t gguf_get_val_u64 (const struct gguf_context * ctx, int key_id); +GGML_API int64_t gguf_get_val_i64 (const struct gguf_context * ctx, int key_id); +GGML_API double gguf_get_val_f64 (const struct gguf_context * ctx, int key_id); +GGML_API bool gguf_get_val_bool(const struct gguf_context * ctx, int key_id); +GGML_API const char * gguf_get_val_str (const struct gguf_context * ctx, int key_id); +GGML_API const void * gguf_get_val_data(const struct gguf_context * ctx, int key_id); +GGML_API int gguf_get_arr_n (const struct gguf_context * ctx, int key_id); +GGML_API const void * gguf_get_arr_data(const struct gguf_context * ctx, int key_id); +GGML_API const char * gguf_get_arr_str (const struct gguf_context * ctx, int key_id, int i); - struct gguf_context; +GGML_API int gguf_get_n_tensors (const struct gguf_context * ctx); +GGML_API int gguf_find_tensor (const struct gguf_context * ctx, const char * name); +GGML_API size_t gguf_get_tensor_offset(const struct gguf_context * ctx, int i); +GGML_API char * gguf_get_tensor_name (const struct gguf_context * ctx, int i); - struct gguf_init_params { - bool no_alloc; +// overrides existing values or adds a new one +GGML_API void gguf_set_val_u8 (struct gguf_context * ctx, const char * key, uint8_t val); +GGML_API void gguf_set_val_i8 (struct gguf_context * ctx, const char * key, int8_t val); +GGML_API void gguf_set_val_u16 (struct gguf_context * ctx, const char * key, uint16_t val); +GGML_API void gguf_set_val_i16 (struct gguf_context * ctx, const char * key, int16_t val); +GGML_API void gguf_set_val_u32 (struct gguf_context * ctx, const char * key, uint32_t val); +GGML_API void gguf_set_val_i32 (struct gguf_context * ctx, const char * key, int32_t val); +GGML_API void gguf_set_val_f32 (struct gguf_context * ctx, const char * key, float val); +GGML_API void gguf_set_val_u64 (struct gguf_context * ctx, const char * key, uint64_t val); +GGML_API void gguf_set_val_i64 (struct gguf_context * ctx, const char * key, int64_t val); +GGML_API void gguf_set_val_f64 (struct gguf_context * ctx, const char * key, double val); +GGML_API void gguf_set_val_bool(struct gguf_context * ctx, const char * key, bool val); +GGML_API void gguf_set_val_str (struct gguf_context * ctx, const char * key, const char * val); +GGML_API void gguf_set_arr_data(struct gguf_context * ctx, const char * key, enum gguf_type type, const void * data, int n); +GGML_API void gguf_set_arr_str (struct gguf_context * ctx, const char * key, const char ** data, int n); - // if not NULL, create a ggml_context and allocate the tensor data in it - struct ggml_context ** ctx; - }; +// set or add KV pairs from another context +GGML_API void gguf_set_kv(struct gguf_context * ctx, struct gguf_context * src); - GGML_API struct gguf_context * gguf_init_empty(void); - GGML_API struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_params params); - //GGML_API struct gguf_context * gguf_init_from_buffer(..); +// manage tensor info +GGML_API void gguf_add_tensor(struct gguf_context * ctx, const struct ggml_tensor * tensor); +GGML_API void gguf_set_tensor_type(struct gguf_context * ctx, const char * name, enum ggml_type type); +GGML_API void gguf_set_tensor_data(struct gguf_context * ctx, const char * name, const void * data, size_t size); - GGML_API void gguf_free(struct gguf_context * ctx); +// writing gguf files can be done in 2 ways: +// +// - write the entire gguf_context to a binary file in a single pass: +// +// gguf_write_to_file(ctx, fname); +// +// - first prepare a file with a placeholder for the meta data, write the tensor data, then write the meta data: +// +// FILE * f = fopen(fname, "wb"); +// fseek(f, gguf_get_meta_size(ctx), SEEK_SET); +// fwrite(f, ...); +// void * data = gguf_meta_get_meta_data(ctx); +// fseek(f, 0, SEEK_SET); +// fwrite(f, data, gguf_get_meta_size(ctx)); +// free(data); +// fclose(f); +// - GGML_API const char * gguf_type_name(enum gguf_type type); +// write the entire context to a binary file +GGML_API void gguf_write_to_file(const struct gguf_context * ctx, const char * fname, bool only_meta); - GGML_API int gguf_get_version (const struct gguf_context * ctx); - GGML_API size_t gguf_get_alignment (const struct gguf_context * ctx); - GGML_API size_t gguf_get_data_offset(const struct gguf_context * ctx); - GGML_API void * gguf_get_data (const struct gguf_context * ctx); +// get the size in bytes of the meta data (header, kv pairs, tensor info) including padding +GGML_API size_t gguf_get_meta_size(const struct gguf_context * ctx); +GGML_API void gguf_get_meta_data(const struct gguf_context * ctx, void * data); - GGML_API int gguf_get_n_kv(const struct gguf_context * ctx); - GGML_API int gguf_find_key(const struct gguf_context * ctx, const char * key); - GGML_API const char * gguf_get_key (const struct gguf_context * ctx, int key_id); +// +// system info +// - GGML_API enum gguf_type gguf_get_kv_type (const struct gguf_context * ctx, int key_id); - GGML_API enum gguf_type gguf_get_arr_type(const struct gguf_context * ctx, int key_id); +GGML_API int ggml_cpu_has_avx (void); +GGML_API int ggml_cpu_has_avx2 (void); +GGML_API int ggml_cpu_has_avx512 (void); +GGML_API int ggml_cpu_has_avx512_vbmi(void); +GGML_API int ggml_cpu_has_avx512_vnni(void); +GGML_API int ggml_cpu_has_fma (void); +GGML_API int ggml_cpu_has_neon (void); +GGML_API int ggml_cpu_has_arm_fma (void); +GGML_API int ggml_cpu_has_metal (void); +GGML_API int ggml_cpu_has_f16c (void); +GGML_API int ggml_cpu_has_fp16_va (void); +GGML_API int ggml_cpu_has_wasm_simd (void); +GGML_API int ggml_cpu_has_blas (void); +GGML_API int ggml_cpu_has_cublas (void); +GGML_API int ggml_cpu_has_clblast (void); +GGML_API int ggml_cpu_has_gpublas (void); +GGML_API int ggml_cpu_has_sse3 (void); +GGML_API int ggml_cpu_has_ssse3 (void); +GGML_API int ggml_cpu_has_vsx (void); - // will abort if the wrong type is used for the key - GGML_API uint8_t gguf_get_val_u8 (const struct gguf_context * ctx, int key_id); - GGML_API int8_t gguf_get_val_i8 (const struct gguf_context * ctx, int key_id); - GGML_API uint16_t gguf_get_val_u16 (const struct gguf_context * ctx, int key_id); - GGML_API int16_t gguf_get_val_i16 (const struct gguf_context * ctx, int key_id); - GGML_API uint32_t gguf_get_val_u32 (const struct gguf_context * ctx, int key_id); - GGML_API int32_t gguf_get_val_i32 (const struct gguf_context * ctx, int key_id); - GGML_API float gguf_get_val_f32 (const struct gguf_context * ctx, int key_id); - GGML_API uint64_t gguf_get_val_u64 (const struct gguf_context * ctx, int key_id); - GGML_API int64_t gguf_get_val_i64 (const struct gguf_context * ctx, int key_id); - GGML_API double gguf_get_val_f64 (const struct gguf_context * ctx, int key_id); - GGML_API bool gguf_get_val_bool(const struct gguf_context * ctx, int key_id); - GGML_API const char * gguf_get_val_str (const struct gguf_context * ctx, int key_id); - GGML_API int gguf_get_arr_n (const struct gguf_context * ctx, int key_id); - GGML_API const void * gguf_get_arr_data(const struct gguf_context * ctx, int key_id); - GGML_API const char * gguf_get_arr_str (const struct gguf_context * ctx, int key_id, int i); - - GGML_API int gguf_get_n_tensors (const struct gguf_context * ctx); - GGML_API int gguf_find_tensor (const struct gguf_context * ctx, const char * name); - GGML_API size_t gguf_get_tensor_offset(const struct gguf_context * ctx, int i); - GGML_API char * gguf_get_tensor_name (const struct gguf_context * ctx, int i); - - // overrides existing values or adds a new one - GGML_API void gguf_set_val_u8 (struct gguf_context * ctx, const char * key, uint8_t val); - GGML_API void gguf_set_val_i8 (struct gguf_context * ctx, const char * key, int8_t val); - GGML_API void gguf_set_val_u16 (struct gguf_context * ctx, const char * key, uint16_t val); - GGML_API void gguf_set_val_i16 (struct gguf_context * ctx, const char * key, int16_t val); - GGML_API void gguf_set_val_u32 (struct gguf_context * ctx, const char * key, uint32_t val); - GGML_API void gguf_set_val_i32 (struct gguf_context * ctx, const char * key, int32_t val); - GGML_API void gguf_set_val_f32 (struct gguf_context * ctx, const char * key, float val); - GGML_API void gguf_set_val_u64 (struct gguf_context * ctx, const char * key, uint64_t val); - GGML_API void gguf_set_val_i64 (struct gguf_context * ctx, const char * key, int64_t val); - GGML_API void gguf_set_val_f64 (struct gguf_context * ctx, const char * key, double val); - GGML_API void gguf_set_val_bool(struct gguf_context * ctx, const char * key, bool val); - GGML_API void gguf_set_val_str (struct gguf_context * ctx, const char * key, const char * val); - GGML_API void gguf_set_arr_data(struct gguf_context * ctx, const char * key, enum gguf_type type, const void * data, int n); - GGML_API void gguf_set_arr_str (struct gguf_context * ctx, const char * key, const char ** data, int n); - - // set or add KV pairs from another context - GGML_API void gguf_set_kv(struct gguf_context * ctx, struct gguf_context * src); - - // manage tensor info - GGML_API void gguf_add_tensor(struct gguf_context * ctx, const struct ggml_tensor * tensor); - GGML_API void gguf_set_tensor_type(struct gguf_context * ctx, const char * name, enum ggml_type type); - GGML_API void gguf_set_tensor_data(struct gguf_context * ctx, const char * name, const void * data, size_t size); - - // writing gguf files can be done in 2 ways: - // - // - write the entire gguf_context to a binary file in a single pass: - // - // gguf_write_to_file(ctx, fname); - // - // - first prepare a file with a placeholder for the meta data, write the tensor data, then write the meta data: - // - // FILE * f = fopen(fname, "wb"); - // fseek(f, gguf_get_meta_size(ctx), SEEK_SET); - // fwrite(f, ...); - // void * data = gguf_meta_get_meta_data(ctx); - // fseek(f, 0, SEEK_SET); - // fwrite(f, data, gguf_get_meta_size(ctx)); - // free(data); - // fclose(f); - // - - // write the entire context to a binary file - GGML_API void gguf_write_to_file(const struct gguf_context * ctx, const char * fname, bool only_meta); - - // get the size in bytes of the meta data (header, kv pairs, tensor info) including padding - GGML_API size_t gguf_get_meta_size(const struct gguf_context * ctx); - GGML_API void gguf_get_meta_data(const struct gguf_context * ctx, void * data); - - // - // system info - // - - GGML_API int ggml_cpu_has_avx (void); - GGML_API int ggml_cpu_has_avx2 (void); - GGML_API int ggml_cpu_has_avx512 (void); - GGML_API int ggml_cpu_has_avx512_vbmi(void); - GGML_API int ggml_cpu_has_avx512_vnni(void); - GGML_API int ggml_cpu_has_fma (void); - GGML_API int ggml_cpu_has_neon (void); - GGML_API int ggml_cpu_has_arm_fma (void); - GGML_API int ggml_cpu_has_metal (void); - GGML_API int ggml_cpu_has_f16c (void); - GGML_API int ggml_cpu_has_fp16_va (void); - GGML_API int ggml_cpu_has_wasm_simd (void); - GGML_API int ggml_cpu_has_blas (void); - GGML_API int ggml_cpu_has_cublas (void); - GGML_API int ggml_cpu_has_clblast (void); - GGML_API int ggml_cpu_has_gpublas (void); - GGML_API int ggml_cpu_has_sse3 (void); - GGML_API int ggml_cpu_has_ssse3 (void); - GGML_API int ggml_cpu_has_vsx (void); - - // - // Internal types and functions exposed for tests and benchmarks - // +// +// Internal types and functions exposed for tests and benchmarks +// #ifdef __cplusplus // restrict not standard in C++ @@ -2100,23 +2138,23 @@ extern "C" { #else #define GGML_RESTRICT restrict #endif - typedef void (*ggml_to_float_t) (const void * GGML_RESTRICT x, float * GGML_RESTRICT y, int k); - typedef void (*ggml_from_float_t)(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k); - typedef void (*ggml_vec_dot_t) (const int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT x, const void * GGML_RESTRICT y); +typedef void (*ggml_to_float_t) (const void * GGML_RESTRICT x, float * GGML_RESTRICT y, int k); +typedef void (*ggml_from_float_t)(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k); +typedef void (*ggml_vec_dot_t) (const int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT x, const void * GGML_RESTRICT y); - typedef struct { - const char * type_name; - int blck_size; - size_t type_size; - bool is_quantized; - ggml_to_float_t to_float; - ggml_from_float_t from_float; - ggml_from_float_t from_float_reference; - ggml_vec_dot_t vec_dot; - enum ggml_type vec_dot_type; - } ggml_type_traits_t; +typedef struct { + const char * type_name; + int blck_size; + size_t type_size; + bool is_quantized; + ggml_to_float_t to_float; + ggml_from_float_t from_float; + ggml_from_float_t from_float_reference; + ggml_vec_dot_t vec_dot; + enum ggml_type vec_dot_type; +} ggml_type_traits_t; - GGML_API ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type); +GGML_API ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type); #ifdef __cplusplus } diff --git a/native/jni/src/ggml/llama.cpp b/native/jni/src/ggml/llama.cpp index 1f038da1f..9b9ed2446 100644 --- a/native/jni/src/ggml/llama.cpp +++ b/native/jni/src/ggml/llama.cpp @@ -28,19 +28,19 @@ #endif #ifdef __has_include - #if __has_include() - #include - #if defined(_POSIX_MAPPED_FILES) - #include - #endif - #if defined(_POSIX_MEMLOCK_RANGE) - #include - #endif - #endif +#if __has_include() +#include +#if defined(_POSIX_MAPPED_FILES) +#include +#endif +#if defined(_POSIX_MEMLOCK_RANGE) +#include +#endif +#endif #endif #if defined(_WIN32) - #define WIN32_LEAN_AND_MEAN +#define WIN32_LEAN_AND_MEAN #ifndef NOMINMAX #define NOMINMAX #endif @@ -54,6 +54,7 @@ #include #include #include +#include #include #include #include @@ -90,6 +91,8 @@ #define LLAMA_ATTRIBUTE_FORMAT(...) #endif +#define LLAMA_MAX_NODES 8192 + // // logging // @@ -189,21 +192,23 @@ enum llm_arch { LLM_ARCH_PERSIMMON, LLM_ARCH_REFACT, LLM_ARCH_BLOOM, + LLM_ARCH_STABLELM, LLM_ARCH_UNKNOWN, }; static std::map LLM_ARCH_NAMES = { - { LLM_ARCH_LLAMA, "llama" }, - { LLM_ARCH_FALCON, "falcon" }, - { LLM_ARCH_GPT2, "gpt2" }, - { LLM_ARCH_GPTJ, "gptj" }, - { LLM_ARCH_GPTNEOX, "gptneox" }, - { LLM_ARCH_MPT, "mpt" }, - { LLM_ARCH_BAICHUAN, "baichuan" }, - { LLM_ARCH_STARCODER, "starcoder" }, - { LLM_ARCH_PERSIMMON, "persimmon" }, - { LLM_ARCH_REFACT, "refact" }, - { LLM_ARCH_BLOOM, "bloom" }, + { LLM_ARCH_LLAMA, "llama" }, + { LLM_ARCH_FALCON, "falcon" }, + { LLM_ARCH_GPT2, "gpt2" }, + { LLM_ARCH_GPTJ, "gptj" }, + { LLM_ARCH_GPTNEOX, "gptneox" }, + { LLM_ARCH_MPT, "mpt" }, + { LLM_ARCH_BAICHUAN, "baichuan" }, + { LLM_ARCH_STARCODER, "starcoder" }, + { LLM_ARCH_PERSIMMON, "persimmon" }, + { LLM_ARCH_REFACT, "refact" }, + { LLM_ARCH_BLOOM, "bloom" }, + { LLM_ARCH_STABLELM, "stablelm" }, }; enum llm_kv { @@ -235,6 +240,10 @@ enum llm_kv { LLM_KV_ROPE_DIMENSION_COUNT, LLM_KV_ROPE_FREQ_BASE, LLM_KV_ROPE_SCALE_LINEAR, + LLM_KV_ROPE_SCALING_TYPE, + LLM_KV_ROPE_SCALING_FACTOR, + LLM_KV_ROPE_SCALING_ORIG_CTX_LEN, + LLM_KV_ROPE_SCALING_FINETUNED, LLM_KV_TOKENIZER_MODEL, LLM_KV_TOKENIZER_LIST, @@ -246,52 +255,60 @@ enum llm_kv { LLM_KV_TOKENIZER_UNK_ID, LLM_KV_TOKENIZER_SEP_ID, LLM_KV_TOKENIZER_PAD_ID, + LLM_KV_TOKENIZER_ADD_BOS, + LLM_KV_TOKENIZER_ADD_EOS, LLM_KV_TOKENIZER_HF_JSON, LLM_KV_TOKENIZER_RWKV, }; static std::map LLM_KV_NAMES = { - { LLM_KV_GENERAL_ARCHITECTURE, "general.architecture" }, - { LLM_KV_GENERAL_QUANTIZATION_VERSION, "general.quantization_version" }, - { LLM_KV_GENERAL_ALIGNMENT, "general.alignment" }, - { LLM_KV_GENERAL_NAME, "general.name" }, - { LLM_KV_GENERAL_AUTHOR, "general.author" }, - { LLM_KV_GENERAL_URL, "general.url" }, - { LLM_KV_GENERAL_DESCRIPTION, "general.description" }, - { LLM_KV_GENERAL_LICENSE, "general.license" }, - { LLM_KV_GENERAL_SOURCE_URL, "general.source.url" }, - { LLM_KV_GENERAL_SOURCE_HF_REPO, "general.source.huggingface.repository" }, + { LLM_KV_GENERAL_ARCHITECTURE, "general.architecture" }, + { LLM_KV_GENERAL_QUANTIZATION_VERSION, "general.quantization_version" }, + { LLM_KV_GENERAL_ALIGNMENT, "general.alignment" }, + { LLM_KV_GENERAL_NAME, "general.name" }, + { LLM_KV_GENERAL_AUTHOR, "general.author" }, + { LLM_KV_GENERAL_URL, "general.url" }, + { LLM_KV_GENERAL_DESCRIPTION, "general.description" }, + { LLM_KV_GENERAL_LICENSE, "general.license" }, + { LLM_KV_GENERAL_SOURCE_URL, "general.source.url" }, + { LLM_KV_GENERAL_SOURCE_HF_REPO, "general.source.huggingface.repository" }, - { LLM_KV_CONTEXT_LENGTH, "%s.context_length" }, - { LLM_KV_EMBEDDING_LENGTH, "%s.embedding_length" }, - { LLM_KV_BLOCK_COUNT, "%s.block_count" }, - { LLM_KV_FEED_FORWARD_LENGTH, "%s.feed_forward_length" }, - { LLM_KV_USE_PARALLEL_RESIDUAL, "%s.use_parallel_residual" }, - { LLM_KV_TENSOR_DATA_LAYOUT, "%s.tensor_data_layout" }, + { LLM_KV_CONTEXT_LENGTH, "%s.context_length" }, + { LLM_KV_EMBEDDING_LENGTH, "%s.embedding_length" }, + { LLM_KV_BLOCK_COUNT, "%s.block_count" }, + { LLM_KV_FEED_FORWARD_LENGTH, "%s.feed_forward_length" }, + { LLM_KV_USE_PARALLEL_RESIDUAL, "%s.use_parallel_residual" }, + { LLM_KV_TENSOR_DATA_LAYOUT, "%s.tensor_data_layout" }, - { LLM_KV_ATTENTION_HEAD_COUNT, "%s.attention.head_count" }, - { LLM_KV_ATTENTION_HEAD_COUNT_KV, "%s.attention.head_count_kv" }, - { LLM_KV_ATTENTION_MAX_ALIBI_BIAS, "%s.attention.max_alibi_bias" }, - { LLM_KV_ATTENTION_CLAMP_KQV, "%s.attention.clamp_kqv" }, - { LLM_KV_ATTENTION_LAYERNORM_EPS, "%s.attention.layer_norm_epsilon" }, - { LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, "%s.attention.layer_norm_rms_epsilon" }, + { LLM_KV_ATTENTION_HEAD_COUNT, "%s.attention.head_count" }, + { LLM_KV_ATTENTION_HEAD_COUNT_KV, "%s.attention.head_count_kv" }, + { LLM_KV_ATTENTION_MAX_ALIBI_BIAS, "%s.attention.max_alibi_bias" }, + { LLM_KV_ATTENTION_CLAMP_KQV, "%s.attention.clamp_kqv" }, + { LLM_KV_ATTENTION_LAYERNORM_EPS, "%s.attention.layer_norm_epsilon" }, + { LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, "%s.attention.layer_norm_rms_epsilon" }, - { LLM_KV_ROPE_DIMENSION_COUNT, "%s.rope.dimension_count" }, - { LLM_KV_ROPE_FREQ_BASE, "%s.rope.freq_base" }, - { LLM_KV_ROPE_SCALE_LINEAR, "%s.rope.scale_linear" }, + { LLM_KV_ROPE_DIMENSION_COUNT, "%s.rope.dimension_count" }, + { LLM_KV_ROPE_FREQ_BASE, "%s.rope.freq_base" }, + { LLM_KV_ROPE_SCALE_LINEAR, "%s.rope.scale_linear" }, + { LLM_KV_ROPE_SCALING_TYPE, "%s.rope.scaling.type" }, + { LLM_KV_ROPE_SCALING_FACTOR, "%s.rope.scaling.factor" }, + { LLM_KV_ROPE_SCALING_ORIG_CTX_LEN, "%s.rope.scaling.original_context_length" }, + { LLM_KV_ROPE_SCALING_FINETUNED, "%s.rope.scaling.finetuned" }, - { LLM_KV_TOKENIZER_MODEL, "tokenizer.ggml.model" }, - { LLM_KV_TOKENIZER_LIST, "tokenizer.ggml.tokens" }, - { LLM_KV_TOKENIZER_TOKEN_TYPE, "tokenizer.ggml.token_type" }, - { LLM_KV_TOKENIZER_SCORES, "tokenizer.ggml.scores" }, - { LLM_KV_TOKENIZER_MERGES, "tokenizer.ggml.merges" }, - { LLM_KV_TOKENIZER_BOS_ID, "tokenizer.ggml.bos_token_id" }, - { LLM_KV_TOKENIZER_EOS_ID, "tokenizer.ggml.eos_token_id" }, - { LLM_KV_TOKENIZER_UNK_ID, "tokenizer.ggml.unknown_token_id" }, - { LLM_KV_TOKENIZER_SEP_ID, "tokenizer.ggml.seperator_token_id" }, - { LLM_KV_TOKENIZER_PAD_ID, "tokenizer.ggml.padding_token_id" }, - { LLM_KV_TOKENIZER_HF_JSON, "tokenizer.huggingface.json" }, - { LLM_KV_TOKENIZER_RWKV, "tokenizer.rwkv.world" }, + { LLM_KV_TOKENIZER_MODEL, "tokenizer.ggml.model" }, + { LLM_KV_TOKENIZER_LIST, "tokenizer.ggml.tokens" }, + { LLM_KV_TOKENIZER_TOKEN_TYPE, "tokenizer.ggml.token_type" }, + { LLM_KV_TOKENIZER_SCORES, "tokenizer.ggml.scores" }, + { LLM_KV_TOKENIZER_MERGES, "tokenizer.ggml.merges" }, + { LLM_KV_TOKENIZER_BOS_ID, "tokenizer.ggml.bos_token_id" }, + { LLM_KV_TOKENIZER_EOS_ID, "tokenizer.ggml.eos_token_id" }, + { LLM_KV_TOKENIZER_UNK_ID, "tokenizer.ggml.unknown_token_id" }, + { LLM_KV_TOKENIZER_SEP_ID, "tokenizer.ggml.seperator_token_id" }, + { LLM_KV_TOKENIZER_PAD_ID, "tokenizer.ggml.padding_token_id" }, + { LLM_KV_TOKENIZER_ADD_BOS, "tokenizer.ggml.add_bos_token" }, + { LLM_KV_TOKENIZER_ADD_EOS, "tokenizer.ggml.add_eos_token" }, + { LLM_KV_TOKENIZER_HF_JSON, "tokenizer.huggingface.json" }, + { LLM_KV_TOKENIZER_RWKV, "tokenizer.rwkv.world" }, }; struct LLM_KV { @@ -328,168 +345,187 @@ enum llm_tensor { }; static std::map> LLM_TENSOR_NAMES = { - { - LLM_ARCH_LLAMA, { - { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, - { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, - { LLM_TENSOR_OUTPUT, "output" }, - { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, - { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, - { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, - { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, - { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, - { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, - { LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" }, - { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, - { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, - { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, - { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + LLM_ARCH_LLAMA, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, }, - }, - { - LLM_ARCH_BAICHUAN, { - { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, - { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, - { LLM_TENSOR_OUTPUT, "output" }, - { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, - { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, - { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, - { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, - { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, - { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, - { LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" }, - { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, - { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, - { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, - { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + LLM_ARCH_BAICHUAN, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, }, - }, - { - LLM_ARCH_FALCON, { - { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, - { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, - { LLM_TENSOR_OUTPUT, "output" }, - { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, - { LLM_TENSOR_ATTN_NORM_2, "blk.%d.attn_norm_2" }, - { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, - { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, - { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, - { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + LLM_ARCH_FALCON, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_NORM_2, "blk.%d.attn_norm_2" }, + { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, }, - }, - { - LLM_ARCH_GPT2, { - { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + LLM_ARCH_GPT2, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + }, }, - }, - { - LLM_ARCH_GPTJ, { - { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + LLM_ARCH_GPTJ, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + }, }, - }, - { - LLM_ARCH_GPTNEOX, { - { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, - { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, - { LLM_TENSOR_OUTPUT, "output" }, - { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, - { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, - { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, - { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, - { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, - { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + LLM_ARCH_GPTNEOX, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, }, - }, - { - LLM_ARCH_PERSIMMON, { - { LLM_TENSOR_TOKEN_EMBD, "token_embd"}, - { LLM_TENSOR_OUTPUT_NORM, "output_norm"}, - { LLM_TENSOR_OUTPUT, "output"}, - { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm"}, - { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv"}, - { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output"}, - { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm"}, - { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm"}, - { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm"}, - { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down"}, - { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up"}, - { LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd"}, + LLM_ARCH_PERSIMMON, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd"}, + { LLM_TENSOR_OUTPUT_NORM, "output_norm"}, + { LLM_TENSOR_OUTPUT, "output"}, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm"}, + { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv"}, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output"}, + { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm"}, + { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm"}, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm"}, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down"}, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up"}, + { LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd"}, + }, }, - }, - { - LLM_ARCH_MPT, { - { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, - { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, - { LLM_TENSOR_OUTPUT, "output" }, - { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, - { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, - { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, - { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, - { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, - { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + LLM_ARCH_MPT, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, }, - }, - { - LLM_ARCH_STARCODER, { - { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, - { LLM_TENSOR_POS_EMBD, "position_embd" }, - { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, - { LLM_TENSOR_OUTPUT, "output" }, - { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, - { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, - { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, - { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, - { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, - { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + LLM_ARCH_STARCODER, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_POS_EMBD, "position_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + }, }, - }, - { - LLM_ARCH_REFACT, { - { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, - { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, - { LLM_TENSOR_OUTPUT, "output" }, - { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, - { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, - { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, - { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, - { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, - { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, - { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, - { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, - { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + LLM_ARCH_REFACT, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, }, - }, - { - LLM_ARCH_BLOOM, { - { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, - { LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" }, - { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, - { LLM_TENSOR_OUTPUT, "output" }, - { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, - { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, - { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, - { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, - { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, - { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + LLM_ARCH_BLOOM, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + }, }, - }, - { - LLM_ARCH_UNKNOWN, { - { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + LLM_ARCH_STABLELM, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, + + { + LLM_ARCH_UNKNOWN, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + }, }, - }, }; static llm_arch llm_arch_from_string(const std::string & name) { @@ -552,6 +588,76 @@ do { \ } \ } while (0) +static std::map LLAMA_ROPE_SCALING_TYPES = { + { LLAMA_ROPE_SCALING_NONE, "none" }, + { LLAMA_ROPE_SCALING_LINEAR, "linear" }, + { LLAMA_ROPE_SCALING_YARN, "yarn" }, +}; + +static int8_t llama_rope_scaling_type_from_string(const std::string & name) { + for (const auto & kv : LLAMA_ROPE_SCALING_TYPES) { + if (kv.second == name) { + return kv.first; + } + } + + return LLAMA_ROPE_SCALING_UNSPECIFIED; +} + +static std::string gguf_data_to_str(enum gguf_type type, const void * data, int i) { + switch (type) { + case GGUF_TYPE_UINT8: return std::to_string(((const uint8_t *)data)[i]); + case GGUF_TYPE_INT8: return std::to_string(((const int8_t *)data)[i]); + case GGUF_TYPE_UINT16: return std::to_string(((const uint16_t *)data)[i]); + case GGUF_TYPE_INT16: return std::to_string(((const int16_t *)data)[i]); + case GGUF_TYPE_UINT32: return std::to_string(((const uint32_t *)data)[i]); + case GGUF_TYPE_INT32: return std::to_string(((const int32_t *)data)[i]); + case GGUF_TYPE_UINT64: return std::to_string(((const uint64_t *)data)[i]); + case GGUF_TYPE_INT64: return std::to_string(((const int64_t *)data)[i]); + case GGUF_TYPE_FLOAT32: return std::to_string(((const float *)data)[i]); + case GGUF_TYPE_FLOAT64: return std::to_string(((const double *)data)[i]); + case GGUF_TYPE_BOOL: return ((const bool *)data)[i] ? "true" : "false"; + default: return format("unknown type %d", type); + } +} + +static std::string gguf_kv_to_str(struct gguf_context * ctx_gguf, int i) { + const enum gguf_type type = gguf_get_kv_type(ctx_gguf, i); + + switch (type) { + case GGUF_TYPE_STRING: + return gguf_get_val_str(ctx_gguf, i); + case GGUF_TYPE_ARRAY: + { + const enum gguf_type arr_type = gguf_get_arr_type(ctx_gguf, i); + int arr_n = gguf_get_arr_n(ctx_gguf, i); + const void * data = gguf_get_arr_data(ctx_gguf, i); + std::stringstream ss; + ss << "["; + for (int j = 0; j < arr_n; j++) { + if (arr_type == GGUF_TYPE_STRING) { + std::string val = gguf_get_arr_str(ctx_gguf, i, j); + // escape quotes + replace_all(val, "\\", "\\\\"); + replace_all(val, "\"", "\\\""); + ss << '"' << val << '"'; + } else if (arr_type == GGUF_TYPE_ARRAY) { + ss << "???"; + } else { + ss << gguf_data_to_str(arr_type, data, j); + } + if (j < arr_n - 1) { + ss << ", "; + } + } + ss << "]"; + return ss.str(); + } + default: + return gguf_data_to_str(type, gguf_get_val_data(ctx_gguf, i), 0); + } +} + // // ggml helpers // @@ -571,19 +677,37 @@ static void ggml_graph_compute_helper(std::vector & buf, ggml_cgraph * // llama helpers // +inline void * llama_host_malloc(size_t n) { #ifdef GGML_USE_CUBLAS -# define llama_host_malloc(n) ggml_cuda_host_malloc(n) -# define llama_host_free(data) ggml_cuda_host_free(data) + if (ggml_cublas_loaded()) { + return ggml_cuda_host_malloc(n); + } else { + return malloc(n); + } #elif GGML_USE_METAL -# define llama_host_malloc(n) ggml_metal_host_malloc(n) -# define llama_host_free(data) ggml_metal_host_free(data) + return ggml_metal_host_malloc(n); #elif GGML_USE_CPU_HBM -# define llama_host_malloc(n) hbw_malloc(n) -# define llama_host_free(data) if (data != NULL) hbw_free(data) + return hbw_malloc(n); #else -# define llama_host_malloc(n) malloc(n) -# define llama_host_free(data) free(data) + return malloc(n); #endif +} + +inline void llama_host_free(void * ptr) { +#ifdef GGML_USE_CUBLAS + if (ggml_cublas_loaded()) { + return ggml_cuda_host_free(ptr); + } else { + return free(ptr); + } +#elif GGML_USE_METAL + return ggml_metal_host_free(ptr); +#elif GGML_USE_CPU_HBM + return hbw_free(ptr); +#else + return free(ptr); +#endif +} #if defined(_WIN32) static std::string llama_format_win_err(DWORD err) { @@ -864,14 +988,14 @@ struct llama_mlock { return (size_t) sysconf(_SC_PAGESIZE); } - #ifdef __APPLE__ - #define MLOCK_SUGGESTION \ +#ifdef __APPLE__ + #define MLOCK_SUGGESTION \ "Try increasing the sysctl values 'vm.user_wire_limit' and 'vm.global_user_wire_limit' and/or " \ "decreasing 'vm.global_no_user_wire_amount'. Also try increasing RLIMIT_MLOCK (ulimit -l).\n" - #else - #define MLOCK_SUGGESTION \ +#else +#define MLOCK_SUGGESTION \ "Try increasing RLIMIT_MLOCK ('ulimit -l' as root).\n" - #endif +#endif bool raw_lock(const void * addr, size_t size) const { if (!mlock(addr, size)) { @@ -895,7 +1019,7 @@ struct llama_mlock { return false; } - #undef MLOCK_SUGGESTION +#undef MLOCK_SUGGESTION static void raw_unlock(void * addr, size_t size) { if (munlock(addr, size)) { @@ -1017,9 +1141,9 @@ enum e_model { MODEL_70B, }; -static const size_t kB = 1024; -static const size_t MB = 1024*kB; -static const size_t GB = 1024*MB; +static const size_t kiB = 1024; +static const size_t MiB = 1024*kiB; +static const size_t GiB = 1024*MiB; struct llama_hparams { bool vocab_only; @@ -1035,8 +1159,11 @@ struct llama_hparams { float f_norm_eps; float f_norm_rms_eps; - float rope_freq_base_train; - float rope_freq_scale_train; + float rope_freq_base_train; + float rope_freq_scale_train; + uint32_t n_yarn_orig_ctx; + int8_t rope_scaling_type_train : 3; + bool rope_finetuned : 1; float f_clamp_kqv; float f_max_alibi_bias; @@ -1051,6 +1178,8 @@ struct llama_hparams { if (this->n_layer != other.n_layer) return true; if (this->n_rot != other.n_rot) return true; if (this->n_ff != other.n_ff) return true; + if (this->rope_finetuned != other.rope_finetuned) return true; + if (this->n_yarn_orig_ctx != other.n_yarn_orig_ctx) return true; const float EPSILON = 1e-9; @@ -1081,8 +1210,16 @@ struct llama_cparams { uint32_t n_threads; // number of threads to use for generation uint32_t n_threads_batch; // number of threads to use for batch processing - float rope_freq_base; - float rope_freq_scale; + float rope_freq_base; + float rope_freq_scale; + + uint32_t n_yarn_orig_ctx; + // These hyperparameters are not exposed in GGUF, because all + // existing YaRN models use the same values for them. + float yarn_ext_factor; + float yarn_attn_factor; + float yarn_beta_fast; + float yarn_beta_slow; bool mul_mat_q; }; @@ -1162,9 +1299,11 @@ struct llama_kv_cache { } #ifdef GGML_USE_CUBLAS - ggml_cuda_free_data(k); - ggml_cuda_free_data(v); -#endif // GGML_USE_CUBLAS + if (ggml_cublas_loaded()) { + ggml_cuda_free_data(k); + ggml_cuda_free_data(v); + } +#endif } }; @@ -1195,6 +1334,9 @@ struct llama_vocab { id special_sep_id = -1; id special_pad_id = -1; + int special_add_bos = -1; // -1 unknown, 1 add, 0 don't add. + int special_add_eos = -1; // -1 unknown, 1 add, 0 don't add. + id linefeed_id = 13; id special_prefix_id = 32007; id special_middle_id = 32009; @@ -1239,6 +1381,9 @@ struct llama_model { int n_gpu_layers; + // gguf metadata + std::unordered_map gguf_kv; + // context struct ggml_context * ctx = NULL; @@ -1264,11 +1409,15 @@ struct llama_model { } #ifdef GGML_USE_CUBLAS - for (size_t i = 0; i < tensors_by_name.size(); ++i) { - ggml_cuda_free_data(tensors_by_name[i].second); + if (ggml_cublas_loaded()) { + for (size_t i = 0; i < tensors_by_name.size(); ++i) { + ggml_cuda_free_data(tensors_by_name[i].second); + } + ggml_cuda_free_scratch(); } - ggml_cuda_free_scratch(); -#elif defined(GGML_USE_CLBLAST) +#endif + +#if defined(GGML_USE_CLBLAST) for (size_t i = 0; i < tensors_by_name.size(); ++i) { ggml_cl_free_data(tensors_by_name[i].second); } @@ -1341,10 +1490,10 @@ struct llama_context { static bool llama_kv_cache_init( const struct llama_hparams & hparams, - struct llama_kv_cache & cache, - ggml_type wtype, - uint32_t n_ctx, - int n_gpu_layers) { + struct llama_kv_cache & cache, + ggml_type wtype, + uint32_t n_ctx, + int n_gpu_layers) { const uint32_t n_embd = hparams.n_embd_gqa(); const uint32_t n_layer = hparams.n_layer; @@ -1380,23 +1529,26 @@ static bool llama_kv_cache_init( ggml_set_name(cache.v, "cache_v"); (void) n_gpu_layers; -#ifdef GGML_USE_CUBLAS - size_t vram_kv_cache = 0; - if (n_gpu_layers > (int)n_layer + 1) { - ggml_cuda_assign_buffers_no_scratch(cache.v); - LLAMA_LOG_INFO("%s: offloading v cache to GPU\n", __func__); - vram_kv_cache += ggml_nbytes(cache.v); +#ifdef GGML_USE_CUBLAS + if (ggml_cublas_loaded()) { + size_t vram_kv_cache = 0; + + if (n_gpu_layers > (int)n_layer + 1) { + ggml_cuda_assign_buffers_no_scratch(cache.v); + LLAMA_LOG_INFO("%s: offloading v cache to GPU\n", __func__); + vram_kv_cache += ggml_nbytes(cache.v); + } + if (n_gpu_layers > (int)n_layer + 2) { + ggml_cuda_assign_buffers_no_scratch(cache.k); + LLAMA_LOG_INFO("%s: offloading k cache to GPU\n", __func__); + vram_kv_cache += ggml_nbytes(cache.k); + } + if (vram_kv_cache > 0) { + LLAMA_LOG_INFO("%s: VRAM kv self = %.2f MiB\n", __func__, vram_kv_cache / 1024.0 / 1024.0); + } } - if (n_gpu_layers > (int)n_layer + 2) { - ggml_cuda_assign_buffers_no_scratch(cache.k); - LLAMA_LOG_INFO("%s: offloading k cache to GPU\n", __func__); - vram_kv_cache += ggml_nbytes(cache.k); - } - if (vram_kv_cache > 0) { - LLAMA_LOG_INFO("%s: VRAM kv self = %.2f MB\n", __func__, vram_kv_cache / 1024.0 / 1024.0); - } -#endif // GGML_USE_CUBLAS +#endif return true; } @@ -1406,7 +1558,7 @@ static bool llama_kv_cache_init( // Note: On success, it's important that cache.head points // to the first cell of the slot. static bool llama_kv_cache_find_slot( - struct llama_kv_cache & cache, + struct llama_kv_cache & cache, const struct llama_batch & batch) { const uint32_t n_ctx = cache.size; const uint32_t n_tokens = batch.n_tokens; @@ -1477,9 +1629,9 @@ static void llama_kv_cache_clear(struct llama_kv_cache & cache) { static void llama_kv_cache_seq_rm( struct llama_kv_cache & cache, - llama_seq_id seq_id, - llama_pos p0, - llama_pos p1) { + llama_seq_id seq_id, + llama_pos p0, + llama_pos p1) { uint32_t new_head = cache.size; if (p0 < 0) p0 = 0; @@ -1507,10 +1659,10 @@ static void llama_kv_cache_seq_rm( static void llama_kv_cache_seq_cp( struct llama_kv_cache & cache, - llama_seq_id seq_id_src, - llama_seq_id seq_id_dst, - llama_pos p0, - llama_pos p1) { + llama_seq_id seq_id_src, + llama_seq_id seq_id_dst, + llama_pos p0, + llama_pos p1) { if (p0 < 0) p0 = 0; if (p1 < 0) p1 = std::numeric_limits::max(); @@ -1543,10 +1695,10 @@ static void llama_kv_cache_seq_keep(struct llama_kv_cache & cache, llama_seq_id static void llama_kv_cache_seq_shift( struct llama_kv_cache & cache, - llama_seq_id seq_id, - llama_pos p0, - llama_pos p1, - llama_pos delta) { + llama_seq_id seq_id, + llama_pos p0, + llama_pos p1, + llama_pos delta) { uint32_t new_head = cache.size; if (p0 < 0) p0 = 0; @@ -1630,8 +1782,8 @@ struct llama_model_loader { llama_model_loader(const std::string & fname, bool use_mmap) : file(fname.c_str(), "rb") { struct gguf_init_params params = { - /*.no_alloc = */ true, - /*.ctx = */ &ctx_meta, + /*.no_alloc = */ true, + /*.ctx = */ &ctx_meta, }; ctx_gguf = gguf_init_from_file(fname.c_str(), params); @@ -1652,7 +1804,7 @@ struct llama_model_loader { } LLAMA_LOG_INFO("%s: loaded meta data with %d key-value pairs and %d tensors from %s (version %s)\n", - __func__, n_kv, n_tensors, fname.c_str(), llama_file_version_name(fver)); + __func__, n_kv, n_tensors, fname.c_str(), llama_file_version_name(fver)); // determine file type based on the number of tensors for each quantization and print meta data // TODO: make optional @@ -1690,10 +1842,10 @@ struct llama_model_loader { case GGML_TYPE_Q5_K: ftype = LLAMA_FTYPE_MOSTLY_Q5_K_M; break; case GGML_TYPE_Q6_K: ftype = LLAMA_FTYPE_MOSTLY_Q6_K; break; default: - { - LLAMA_LOG_WARN("%s: unknown type %s\n", __func__, ggml_type_name(type_max)); - ftype = LLAMA_FTYPE_ALL_F32; - } break; + { + LLAMA_LOG_WARN("%s: unknown type %s\n", __func__, ggml_type_name(type_max)); + ftype = LLAMA_FTYPE_ALL_F32; + } break; } // this is a way to mark that we have "guessed" the file type @@ -1707,10 +1859,21 @@ struct llama_model_loader { } for (int i = 0; i < n_kv; i++) { - const char * name = gguf_get_key(ctx_gguf, i); - const enum gguf_type type = gguf_get_kv_type(ctx_gguf, i); + const char * name = gguf_get_key(ctx_gguf, i); + const enum gguf_type type = gguf_get_kv_type(ctx_gguf, i); + const std::string type_name = + type == GGUF_TYPE_ARRAY + ? format("%s[%s,%d]", gguf_type_name(type), gguf_type_name(gguf_get_arr_type(ctx_gguf, i)), gguf_get_arr_n(ctx_gguf, i)) + : gguf_type_name(type); - LLAMA_LOG_INFO("%s: - kv %3d: %42s %-8s\n", __func__, i, name, gguf_type_name(type)); + std::string value = gguf_kv_to_str(ctx_gguf, i); + const size_t MAX_VALUE_LEN = 40; + if (value.size() > MAX_VALUE_LEN) { + value = format("%s...", value.substr(0, MAX_VALUE_LEN - 3).c_str()); + } + replace_all(value, "\n", "\\n"); + + LLAMA_LOG_INFO("%s: - kv %3d: %42s %-16s = %s\n", __func__, i, name, type_name.c_str(), value.c_str()); } // print type counts @@ -1799,6 +1962,12 @@ struct llama_model_loader { throw std::runtime_error(format("%s: tensor '%s' not found", __func__, name.c_str())); } + if (backend == GGML_BACKEND_GPU_SPLIT) { + if (ne.size() == 1) { + throw std::runtime_error(format("%s: 1-dimensional tensor '%s' cannot be split on the GPU", __func__, name.c_str())); + } + } + { bool is_ok = true; for (size_t i = 0; i < ne.size(); ++i) { @@ -1810,9 +1979,9 @@ struct llama_model_loader { if (!is_ok) { throw std::runtime_error( format("%s: tensor '%s' has wrong shape; expected %s, got %s", - __func__, name.c_str(), - llama_format_tensor_shape(ne).c_str(), - llama_format_tensor_shape(cur).c_str())); + __func__, name.c_str(), + llama_format_tensor_shape(ne).c_str(), + llama_format_tensor_shape(cur).c_str())); } } @@ -1878,11 +2047,11 @@ struct llama_model_loader { // allocate temp buffer if not using mmap if (!use_mmap && cur->data == NULL) { GGML_ASSERT(cur->backend != GGML_BACKEND_CPU); - #ifdef GGML_USE_CPU_HBM +#ifdef GGML_USE_CPU_HBM cur->data = (uint8_t*)hbw_malloc(ggml_nbytes(cur)); - #else +#else cur->data = (uint8_t*)malloc(ggml_nbytes(cur)); - #endif +#endif } load_data_for(cur); @@ -1895,7 +2064,7 @@ struct llama_model_loader { } break; #ifdef GGML_USE_CUBLAS - case GGML_BACKEND_GPU: + case GGML_BACKEND_GPU: case GGML_BACKEND_GPU_SPLIT: // old code: //ggml_cuda_transform_tensor(lt.data, lt.ggml_tensor); @@ -1907,7 +2076,7 @@ struct llama_model_loader { } break; #elif defined(GGML_USE_CLBLAST) - case GGML_BACKEND_GPU: + case GGML_BACKEND_GPU: ggml_cl_transform_tensor(cur->data, cur); if (!use_mmap) { free(cur->data); @@ -1946,12 +2115,12 @@ static std::string llama_model_ftype_name(llama_ftype ftype) { case LLAMA_FTYPE_MOSTLY_Q4_0: return "mostly Q4_0"; case LLAMA_FTYPE_MOSTLY_Q4_1: return "mostly Q4_1"; case LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16: - return "mostly Q4_1, some F16"; + return "mostly Q4_1, some F16"; case LLAMA_FTYPE_MOSTLY_Q5_0: return "mostly Q5_0"; case LLAMA_FTYPE_MOSTLY_Q5_1: return "mostly Q5_1"; case LLAMA_FTYPE_MOSTLY_Q8_0: return "mostly Q8_0"; - // K-quants + // K-quants case LLAMA_FTYPE_MOSTLY_Q2_K: return "mostly Q2_K"; case LLAMA_FTYPE_MOSTLY_Q3_K_S: return "mostly Q3_K - Small"; case LLAMA_FTYPE_MOSTLY_Q3_K_M: return "mostly Q3_K - Medium"; @@ -1999,6 +2168,17 @@ static void llm_load_hparams( auto & hparams = model.hparams; + // get metadata as string + for (int i = 0; i < gguf_get_n_kv(ctx); i++) { + enum gguf_type type = gguf_get_kv_type(ctx, i); + if (type == GGUF_TYPE_ARRAY) { + continue; + } + const char * name = gguf_get_key(ctx, i); + const std::string value = gguf_kv_to_str(ctx, i); + model.gguf_kv.emplace(name, value); + } + // get general kv GGUF_GET_KEY(ctx, model.name, gguf_get_val_str, GGUF_TYPE_STRING, false, kv(LLM_KV_GENERAL_NAME)); @@ -2014,14 +2194,30 @@ static void llm_load_hparams( hparams.n_head_kv = hparams.n_head; GGUF_GET_KEY(ctx, hparams.n_head_kv, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_ATTENTION_HEAD_COUNT_KV)); + hparams.rope_finetuned = false; + GGUF_GET_KEY(ctx, hparams.rope_finetuned, gguf_get_val_bool, GGUF_TYPE_BOOL, false, + kv(LLM_KV_ROPE_SCALING_FINETUNED)); + + hparams.n_yarn_orig_ctx = hparams.n_ctx_train; + GGUF_GET_KEY(ctx, hparams.n_yarn_orig_ctx, gguf_get_val_u32, GGUF_TYPE_UINT32, false, + kv(LLM_KV_ROPE_SCALING_ORIG_CTX_LEN)); + // rope_freq_base (optional) hparams.rope_freq_base_train = 10000.0f; GGUF_GET_KEY(ctx, hparams.rope_freq_base_train, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_FREQ_BASE)); + std::string rope_scaling("linear"); + GGUF_GET_KEY(ctx, rope_scaling, gguf_get_val_str, GGUF_TYPE_STRING, false, kv(LLM_KV_ROPE_SCALING_TYPE)); + hparams.rope_scaling_type_train = llama_rope_scaling_type_from_string(rope_scaling); + GGML_ASSERT(hparams.rope_scaling_type_train != LLAMA_ROPE_SCALING_UNSPECIFIED); + // rope_freq_scale (inverse of the kv) is optional - float ropescale = 1.0f; - GGUF_GET_KEY(ctx, ropescale, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_SCALE_LINEAR)); - hparams.rope_freq_scale_train = 1.0f/ropescale; + float ropescale = 0.0f; + GGUF_GET_KEY(ctx, ropescale, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_SCALING_FACTOR)); + if (ropescale == 0.0f) { // try the old key name + GGUF_GET_KEY(ctx, ropescale, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_SCALE_LINEAR)); + } + hparams.rope_freq_scale_train = ropescale == 0.0f ? 1.0f : 1.0f/ropescale; // sanity check for n_rot (optional) { @@ -2041,92 +2237,102 @@ static void llm_load_hparams( // arch-specific KVs switch (model.arch) { case LLM_ARCH_LLAMA: - { - GGUF_GET_KEY(ctx, hparams.f_norm_rms_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS)); + { + GGUF_GET_KEY(ctx, hparams.f_norm_rms_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS)); - switch (hparams.n_layer) { - case 26: model.type = e_model::MODEL_3B; break; - case 32: model.type = e_model::MODEL_7B; break; - case 40: model.type = e_model::MODEL_13B; break; - case 48: model.type = e_model::MODEL_34B; break; - case 60: model.type = e_model::MODEL_30B; break; - case 80: model.type = hparams.n_head == hparams.n_head_kv ? e_model::MODEL_65B : e_model::MODEL_70B; break; - default: model.type = e_model::MODEL_UNKNOWN; - } - } break; + switch (hparams.n_layer) { + case 26: model.type = e_model::MODEL_3B; break; + case 32: model.type = e_model::MODEL_7B; break; + case 40: model.type = e_model::MODEL_13B; break; + case 48: model.type = e_model::MODEL_34B; break; + case 60: model.type = e_model::MODEL_30B; break; + case 80: model.type = hparams.n_head == hparams.n_head_kv ? e_model::MODEL_65B : e_model::MODEL_70B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; case LLM_ARCH_FALCON: - { - GGUF_GET_KEY(ctx, hparams.f_norm_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_EPS)); + { + GGUF_GET_KEY(ctx, hparams.f_norm_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_EPS)); - switch (hparams.n_layer) { - case 32: model.type = e_model::MODEL_7B; break; - case 60: model.type = e_model::MODEL_40B; break; - default: model.type = e_model::MODEL_UNKNOWN; - } - } break; + switch (hparams.n_layer) { + case 32: model.type = e_model::MODEL_7B; break; + case 60: model.type = e_model::MODEL_40B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; case LLM_ARCH_BAICHUAN: - { - GGUF_GET_KEY(ctx, hparams.f_norm_rms_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS)); - switch (hparams.n_layer) { - case 32: model.type = e_model::MODEL_7B; break; - case 40: model.type = e_model::MODEL_13B; break; - default: model.type = e_model::MODEL_UNKNOWN; - } - } break; + { + GGUF_GET_KEY(ctx, hparams.f_norm_rms_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS)); + switch (hparams.n_layer) { + case 32: model.type = e_model::MODEL_7B; break; + case 40: model.type = e_model::MODEL_13B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; case LLM_ARCH_STARCODER: - { - GGUF_GET_KEY(ctx, hparams.f_norm_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_EPS)); - switch (hparams.n_layer) { - case 24: model.type = e_model::MODEL_1B; break; - case 36: model.type = e_model::MODEL_3B; break; - case 42: model.type = e_model::MODEL_7B; break; - case 40: model.type = e_model::MODEL_15B; break; - default: model.type = e_model::MODEL_UNKNOWN; - } - } break; + { + GGUF_GET_KEY(ctx, hparams.f_norm_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_EPS)); + switch (hparams.n_layer) { + case 24: model.type = e_model::MODEL_1B; break; + case 36: model.type = e_model::MODEL_3B; break; + case 42: model.type = e_model::MODEL_7B; break; + case 40: model.type = e_model::MODEL_15B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; case LLM_ARCH_PERSIMMON: - { - GGUF_GET_KEY(ctx, hparams.f_norm_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_EPS)); - switch (hparams.n_layer) { - case 36: model.type = e_model::MODEL_8B; break; - default: model.type = e_model::MODEL_UNKNOWN; - } - } break; + { + GGUF_GET_KEY(ctx, hparams.f_norm_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_EPS)); + switch (hparams.n_layer) { + case 36: model.type = e_model::MODEL_8B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; case LLM_ARCH_REFACT: - { - GGUF_GET_KEY(ctx, hparams.f_norm_rms_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS)); - switch (hparams.n_layer) { - case 32: model.type = e_model::MODEL_1B; break; - default: model.type = e_model::MODEL_UNKNOWN; - } - } break; + { + GGUF_GET_KEY(ctx, hparams.f_norm_rms_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS)); + switch (hparams.n_layer) { + case 32: model.type = e_model::MODEL_1B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; case LLM_ARCH_BLOOM: - { - GGUF_GET_KEY(ctx, hparams.f_norm_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_EPS)); + { + GGUF_GET_KEY(ctx, hparams.f_norm_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_EPS)); - switch (hparams.n_layer) { - case 24: model.type = e_model::MODEL_1B; break; - case 30: - switch (hparams.n_embd) { - case 2560: model.type = e_model::MODEL_3B; break; - case 4096: model.type = e_model::MODEL_7B; break; - } break; - } - } break; + switch (hparams.n_layer) { + case 24: model.type = e_model::MODEL_1B; break; + case 30: + switch (hparams.n_embd) { + case 2560: model.type = e_model::MODEL_3B; break; + case 4096: model.type = e_model::MODEL_7B; break; + } break; + } + } break; case LLM_ARCH_MPT: - { - hparams.f_clamp_kqv = 0.0f; + { + hparams.f_clamp_kqv = 0.0f; - GGUF_GET_KEY(ctx, hparams.f_norm_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_EPS)); - GGUF_GET_KEY(ctx, hparams.f_clamp_kqv, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ATTENTION_CLAMP_KQV)); - GGUF_GET_KEY(ctx, hparams.f_max_alibi_bias, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_MAX_ALIBI_BIAS)); + GGUF_GET_KEY(ctx, hparams.f_norm_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_EPS)); + GGUF_GET_KEY(ctx, hparams.f_clamp_kqv, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ATTENTION_CLAMP_KQV)); + GGUF_GET_KEY(ctx, hparams.f_max_alibi_bias, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_MAX_ALIBI_BIAS)); + + switch (hparams.n_layer) { + case 32: model.type = e_model::MODEL_7B; break; + case 48: model.type = e_model::MODEL_30B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_STABLELM: + { + GGUF_GET_KEY(ctx, hparams.f_norm_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_EPS)); + + switch (hparams.n_layer) { + case 32: model.type = e_model::MODEL_3B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; - switch (hparams.n_layer) { - case 32: model.type = e_model::MODEL_7B; break; - case 48: model.type = e_model::MODEL_30B; break; - default: model.type = e_model::MODEL_UNKNOWN; - } - } break; default: (void)0; } @@ -2249,11 +2455,11 @@ static void llm_load_vocab( // special tokens { const std::vector> special_token_types = { - { LLM_KV_TOKENIZER_BOS_ID, vocab.special_bos_id }, - { LLM_KV_TOKENIZER_EOS_ID, vocab.special_eos_id }, - { LLM_KV_TOKENIZER_UNK_ID, vocab.special_unk_id }, - { LLM_KV_TOKENIZER_SEP_ID, vocab.special_sep_id }, - { LLM_KV_TOKENIZER_PAD_ID, vocab.special_pad_id }, + { LLM_KV_TOKENIZER_BOS_ID, vocab.special_bos_id }, + { LLM_KV_TOKENIZER_EOS_ID, vocab.special_eos_id }, + { LLM_KV_TOKENIZER_UNK_ID, vocab.special_unk_id }, + { LLM_KV_TOKENIZER_SEP_ID, vocab.special_sep_id }, + { LLM_KV_TOKENIZER_PAD_ID, vocab.special_pad_id }, }; for (const auto & it : special_token_types) { const std::string & key = kv(std::get<0>(it)); @@ -2265,9 +2471,26 @@ static void llm_load_vocab( // validating that. if (size_t(id + 1) > vocab.id_to_token.size()) { LLAMA_LOG_WARN("%s: bad special token: '%s' = %d, using default id %d\n", - __func__, key.c_str(), id, old_id); + __func__, key.c_str(), id, old_id); id = old_id; } + + } + + // Handle add_bos_token and add_eos_token + std::string key = kv(LLM_KV_TOKENIZER_ADD_BOS); + int kid = gguf_find_key(ctx, key.c_str()); + enum gguf_type ktype = kid < 0 ? GGUF_TYPE_COUNT : gguf_get_kv_type(ctx, kid); + vocab.special_add_bos = ktype == GGUF_TYPE_BOOL ? gguf_get_val_bool(ctx, kid) : -1; + if (ktype != GGUF_TYPE_BOOL && ktype != GGUF_TYPE_COUNT) { + LLAMA_LOG_WARN("%s: bad field type %d for '%s' - ignoring\n", __func__, ktype, key.c_str()); + } + key = kv(LLM_KV_TOKENIZER_ADD_EOS); + kid = gguf_find_key(ctx, key.c_str()); + ktype = kid < 0 ? GGUF_TYPE_COUNT : gguf_get_kv_type(ctx, kid); + vocab.special_add_eos = ktype == GGUF_TYPE_BOOL ? gguf_get_val_bool(ctx, kid) : -1; + if (ktype != GGUF_TYPE_BOOL && ktype != GGUF_TYPE_COUNT) { + LLAMA_LOG_WARN("%s: bad field type %d for '%s' - ignoring\n", __func__, ktype, key.c_str()); } } @@ -2354,14 +2577,14 @@ static void llm_load_vocab( if (special_tokens_definition_mismatch || special_tokens_count_from_verification != special_tokens_count_by_type) { LLAMA_LOG_WARN("%s: mismatch in special tokens definition ( %u/%zu vs %u/%zu ).\n", - __func__, - special_tokens_count_from_verification, vocab.id_to_token.size(), - special_tokens_count_by_type, vocab.id_to_token.size() + __func__, + special_tokens_count_from_verification, vocab.id_to_token.size(), + special_tokens_count_by_type, vocab.id_to_token.size() ); } else { LLAMA_LOG_INFO("%s: special tokens definition check successful ( %u/%zu ).\n", - __func__, - special_tokens_count_from_verification, vocab.id_to_token.size() + __func__, + special_tokens_count_from_verification, vocab.id_to_token.size() ); } } @@ -2371,6 +2594,8 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) { const auto & hparams = model.hparams; const auto & vocab = model.vocab; + const auto rope_scaling_type = LLAMA_ROPE_SCALING_TYPES.at(hparams.rope_scaling_type_train); + // hparams LLAMA_LOG_INFO("%s: format = %s\n", __func__, llama_file_version_name(ml.fver)); LLAMA_LOG_INFO("%s: arch = %s\n", __func__, LLM_ARCH_NAMES.at(model.arch).c_str()); @@ -2389,13 +2614,16 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) { LLAMA_LOG_INFO("%s: f_clamp_kqv = %.1e\n", __func__, hparams.f_clamp_kqv); LLAMA_LOG_INFO("%s: f_max_alibi_bias = %.1e\n", __func__, hparams.f_max_alibi_bias); LLAMA_LOG_INFO("%s: n_ff = %u\n", __func__, hparams.n_ff); + LLAMA_LOG_INFO("%s: rope scaling = %s\n", __func__, rope_scaling_type.c_str()); LLAMA_LOG_INFO("%s: freq_base_train = %.1f\n", __func__, hparams.rope_freq_base_train); LLAMA_LOG_INFO("%s: freq_scale_train = %g\n", __func__, hparams.rope_freq_scale_train); + LLAMA_LOG_INFO("%s: n_yarn_orig_ctx = %u\n", __func__, hparams.n_yarn_orig_ctx); + LLAMA_LOG_INFO("%s: rope_finetuned = %s\n", __func__, hparams.rope_finetuned ? "yes" : "unknown"); LLAMA_LOG_INFO("%s: model type = %s\n", __func__, llama_model_type_name(model.type)); LLAMA_LOG_INFO("%s: model ftype = %s\n", __func__, llama_model_ftype_name(model.ftype).c_str()); LLAMA_LOG_INFO("%s: model params = %.2f B\n", __func__, ml.n_elements*1e-9); - if (ml.n_bytes < GB) { - LLAMA_LOG_INFO("%s: model size = %.2f MiB (%.2f BPW) \n", __func__, ml.n_bytes/1024.0/1024.0, ml.n_bytes*8.0/ml.n_elements); + if (ml.n_bytes < GiB) { + LLAMA_LOG_INFO("%s: model size = %.2f MiB (%.2f BPW) \n", __func__, ml.n_bytes/1024.0/1024.0, ml.n_bytes*8.0/ml.n_elements); } else { LLAMA_LOG_INFO("%s: model size = %.2f GiB (%.2f BPW) \n", __func__, ml.n_bytes/1024.0/1024.0/1024.0, ml.n_bytes*8.0/ml.n_elements); } @@ -2433,7 +2661,7 @@ static void llm_load_tensors( ml.calc_sizes(ctx_size, mmapped_size); - LLAMA_LOG_INFO("%s: ggml ctx size = %7.2f MB\n", __func__, ctx_size/1024.0/1024.0); + LLAMA_LOG_INFO("%s: ggml ctx size = %7.2f MiB\n", __func__, ctx_size/1024.0/1024.0); // create the ggml context { @@ -2444,9 +2672,9 @@ static void llm_load_tensors( } struct ggml_init_params params = { - /*.mem_size =*/ model.buf.size, - /*.mem_buffer =*/ model.buf.data, - /*.no_alloc =*/ ml.use_mmap, + /*.mem_size =*/ model.buf.size, + /*.mem_buffer =*/ model.buf.data, + /*.no_alloc =*/ ml.use_mmap, }; model.ctx = ggml_init(params); @@ -2456,18 +2684,22 @@ static void llm_load_tensors( } (void) main_gpu; + + enum ggml_backend_type llama_backend_offload = GGML_BACKEND_CPU; + enum ggml_backend_type llama_backend_offload_split = GGML_BACKEND_CPU; + #ifdef GGML_USE_CUBLAS - LLAMA_LOG_INFO("%s: using " GGML_CUDA_NAME " for GPU acceleration\n", __func__); - ggml_cuda_set_main_device(main_gpu); -#define LLAMA_BACKEND_OFFLOAD GGML_BACKEND_GPU -#define LLAMA_BACKEND_OFFLOAD_SPLIT GGML_BACKEND_GPU_SPLIT + if (ggml_cublas_loaded()) { + LLAMA_LOG_INFO("%s: using " GGML_CUDA_NAME " for GPU acceleration\n", __func__); + ggml_cuda_set_main_device(main_gpu); + + llama_backend_offload = GGML_BACKEND_GPU; + llama_backend_offload_split = GGML_BACKEND_GPU_SPLIT; + } #elif defined(GGML_USE_CLBLAST) LLAMA_LOG_INFO("%s: using OpenCL for GPU acceleration\n", __func__); -#define LLAMA_BACKEND_OFFLOAD GGML_BACKEND_GPU -#define LLAMA_BACKEND_OFFLOAD_SPLIT GGML_BACKEND_GPU -#else -#define LLAMA_BACKEND_OFFLOAD GGML_BACKEND_CPU -#define LLAMA_BACKEND_OFFLOAD_SPLIT GGML_BACKEND_CPU + llama_backend_offload = GGML_BACKEND_GPU; + llama_backend_offload_split = GGML_BACKEND_GPU; #endif // prepare memory for the weights @@ -2482,506 +2714,588 @@ static void llm_load_tensors( switch (model.arch) { case LLM_ARCH_LLAMA: case LLM_ARCH_REFACT: + { + model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); + + // output { - model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); + ggml_backend_type backend_norm; + ggml_backend_type backend_output; - // output - { - ggml_backend_type backend_norm; - ggml_backend_type backend_output; - - if (n_gpu_layers > int(n_layer)) { - // norm is not performance relevant on its own but keeping it in VRAM reduces data copying - // on Windows however this is detrimental unless everything is on the GPU + if (n_gpu_layers > int(n_layer)) { + // norm is not performance relevant on its own but keeping it in VRAM reduces data copying + // on Windows however this is detrimental unless everything is on the GPU #ifndef _WIN32 - backend_norm = LLAMA_BACKEND_OFFLOAD; + backend_norm = llama_backend_offload; #else - backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; + backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : llama_backend_offload; #endif // _WIN32 - backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT; - } else { - backend_norm = GGML_BACKEND_CPU; - backend_output = GGML_BACKEND_CPU; - } - - model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm); - model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output); - - if (backend_norm == GGML_BACKEND_GPU) { - vram_weights += ggml_nbytes(model.output_norm); - } - if (backend_output == GGML_BACKEND_GPU_SPLIT) { - vram_weights += ggml_nbytes(model.output); - } + backend_output = llama_backend_offload_split; + } else { + backend_norm = GGML_BACKEND_CPU; + backend_output = GGML_BACKEND_CPU; } - const uint32_t n_ff = hparams.n_ff; + model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm); + model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output); - const int i_gpu_start = n_layer - n_gpu_layers; + if (backend_norm == GGML_BACKEND_GPU) { + vram_weights += ggml_nbytes(model.output_norm); + } + if (backend_output == GGML_BACKEND_GPU_SPLIT) { + vram_weights += ggml_nbytes(model.output); + } + } - model.layers.resize(n_layer); + const uint32_t n_ff = hparams.n_ff; - for (uint32_t i = 0; i < n_layer; ++i) { - const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT - const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT + const int i_gpu_start = n_layer - n_gpu_layers; - auto & layer = model.layers[i]; + model.layers.resize(n_layer); - layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend); + for (uint32_t i = 0; i < n_layer; ++i) { + const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload; // NOLINT + const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload_split; // NOLINT - layer.wq = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}, backend_split); - layer.wk = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}, backend_split); - layer.wv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}, backend_split); - layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split); + auto & layer = model.layers[i]; - layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend); + layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend); - layer.ffn_gate = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, backend_split); - layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); - layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + layer.wq = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}, backend_split); + layer.wk = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}, backend_split); + layer.wv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}, backend_split); + layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split); - if (backend == GGML_BACKEND_GPU) { - vram_weights += + layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend); + + layer.ffn_gate = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, backend_split); + layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); + layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + + if (backend == GGML_BACKEND_GPU) { + vram_weights += ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.wq) + ggml_nbytes(layer.wk) + ggml_nbytes(layer.wv) + ggml_nbytes(layer.wo) + ggml_nbytes(layer.ffn_norm) + ggml_nbytes(layer.ffn_gate) + ggml_nbytes(layer.ffn_down) + ggml_nbytes(layer.ffn_up); - } } - } break; + } + } break; case LLM_ARCH_BAICHUAN: + { + model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); { - model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); - { - ggml_backend_type backend_norm; - ggml_backend_type backend_output; + ggml_backend_type backend_norm; + ggml_backend_type backend_output; - if (n_gpu_layers > int(n_layer)) { - // norm is not performance relevant on its own but keeping it in VRAM reduces data copying - // on Windows however this is detrimental unless everything is on the GPU + if (n_gpu_layers > int(n_layer)) { + // norm is not performance relevant on its own but keeping it in VRAM reduces data copying + // on Windows however this is detrimental unless everything is on the GPU #ifndef _WIN32 - backend_norm = LLAMA_BACKEND_OFFLOAD; + backend_norm = llama_backend_offload; #else - backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; + backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : llama_backend_offload; #endif // _WIN32 - backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT; - } else { - backend_norm = GGML_BACKEND_CPU; - backend_output = GGML_BACKEND_CPU; - } - - model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm); - model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output); - - if (backend_norm == GGML_BACKEND_GPU) { - vram_weights += ggml_nbytes(model.output_norm); - } - if (backend_output == GGML_BACKEND_GPU_SPLIT) { - vram_weights += ggml_nbytes(model.output); - } + backend_output = llama_backend_offload_split; + } else { + backend_norm = GGML_BACKEND_CPU; + backend_output = GGML_BACKEND_CPU; } - const uint32_t n_ff = hparams.n_ff; + model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm); + model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output); - const int i_gpu_start = n_layer - n_gpu_layers; + if (backend_norm == GGML_BACKEND_GPU) { + vram_weights += ggml_nbytes(model.output_norm); + } + if (backend_output == GGML_BACKEND_GPU_SPLIT) { + vram_weights += ggml_nbytes(model.output); + } + } - model.layers.resize(n_layer); + const uint32_t n_ff = hparams.n_ff; - for (uint32_t i = 0; i < n_layer; ++i) { - const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT - const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT + const int i_gpu_start = n_layer - n_gpu_layers; - auto & layer = model.layers[i]; + model.layers.resize(n_layer); - layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend); + for (uint32_t i = 0; i < n_layer; ++i) { + const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload; // NOLINT + const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload_split; // NOLINT - layer.wq = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}, backend_split); - layer.wk = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}, backend_split); - layer.wv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}, backend_split); - layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split); + auto & layer = model.layers[i]; - layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend); + layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend); - layer.ffn_gate = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, backend_split); - layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); - layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + layer.wq = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}, backend_split); + layer.wk = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}, backend_split); + layer.wv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}, backend_split); + layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split); - if (backend == GGML_BACKEND_GPU) { - vram_weights += + layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend); + + layer.ffn_gate = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, backend_split); + layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); + layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + + if (backend == GGML_BACKEND_GPU) { + vram_weights += ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.wq) + ggml_nbytes(layer.wk) + ggml_nbytes(layer.wv) + ggml_nbytes(layer.wo) + ggml_nbytes(layer.ffn_norm) + ggml_nbytes(layer.ffn_gate) + ggml_nbytes(layer.ffn_down) + ggml_nbytes(layer.ffn_up); - } } - } break; + } + } break; case LLM_ARCH_FALCON: + { + // TODO: CPU-only for now + + model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); + + // output { - // TODO: CPU-only for now + ggml_backend_type backend_norm; + ggml_backend_type backend_output; - model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); - - // output - { - ggml_backend_type backend_norm; - ggml_backend_type backend_output; - - if (n_gpu_layers > int(n_layer)) { - // norm is not performance relevant on its own but keeping it in VRAM reduces data copying - // on Windows however this is detrimental unless everything is on the GPU + if (n_gpu_layers > int(n_layer)) { + // norm is not performance relevant on its own but keeping it in VRAM reduces data copying + // on Windows however this is detrimental unless everything is on the GPU #ifndef _WIN32 - backend_norm = LLAMA_BACKEND_OFFLOAD; + backend_norm = llama_backend_offload; #else - backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; + backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : llama_backend_offload; #endif // _WIN32 - backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT; - } else { - backend_norm = GGML_BACKEND_CPU; - backend_output = GGML_BACKEND_CPU; - } + backend_output = llama_backend_offload_split; + } else { + backend_norm = GGML_BACKEND_CPU; + backend_output = GGML_BACKEND_CPU; + } - model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm); - model.output_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, backend_norm); - model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output); + model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm); + model.output_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, backend_norm); + model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output); - if (backend_norm == GGML_BACKEND_GPU) { - vram_weights += ggml_nbytes(model.output_norm); - vram_weights += ggml_nbytes(model.output_norm_b); - } - if (backend_output == GGML_BACKEND_GPU_SPLIT) { - vram_weights += ggml_nbytes(model.output); + if (backend_norm == GGML_BACKEND_GPU) { + vram_weights += ggml_nbytes(model.output_norm); + vram_weights += ggml_nbytes(model.output_norm_b); + } + if (backend_output == GGML_BACKEND_GPU_SPLIT) { + vram_weights += ggml_nbytes(model.output); + } + } + + const uint32_t n_ff = hparams.n_ff; + + const int i_gpu_start = n_layer - n_gpu_layers; + + model.layers.resize(n_layer); + + for (uint32_t i = 0; i < n_layer; ++i) { + const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload; // NOLINT + const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload_split; // NOLINT + + auto & layer = model.layers[i]; + + layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend); + layer.attn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, backend); + + if (gguf_find_tensor(ml.ctx_gguf, tn(LLM_TENSOR_ATTN_NORM_2, "weight", i).c_str()) >= 0) { + layer.attn_norm_2 = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM_2, "weight", i), {n_embd}, backend); + layer.attn_norm_2_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM_2, "bias", i), {n_embd}, backend); + + if (backend == GGML_BACKEND_GPU) { + vram_weights += ggml_nbytes(layer.attn_norm_2); + vram_weights += ggml_nbytes(layer.attn_norm_2_b); } } - const uint32_t n_ff = hparams.n_ff; + layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, backend_split); + layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split); - const int i_gpu_start = n_layer - n_gpu_layers; + layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); + layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); - model.layers.resize(n_layer); - - for (uint32_t i = 0; i < n_layer; ++i) { - const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT - const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT - - auto & layer = model.layers[i]; - - layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend); - layer.attn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, backend); - - if (gguf_find_tensor(ml.ctx_gguf, tn(LLM_TENSOR_ATTN_NORM_2, "weight", i).c_str()) >= 0) { - layer.attn_norm_2 = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM_2, "weight", i), {n_embd}, backend); - layer.attn_norm_2_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM_2, "bias", i), {n_embd}, backend); - - if (backend == GGML_BACKEND_GPU) { - vram_weights += ggml_nbytes(layer.attn_norm_2); - vram_weights += ggml_nbytes(layer.attn_norm_2_b); - } - } - - layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, backend_split); - layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split); - - layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); - layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); - - if (backend == GGML_BACKEND_GPU) { - vram_weights += + if (backend == GGML_BACKEND_GPU) { + vram_weights += ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.attn_norm_b) + ggml_nbytes(layer.wqkv) + ggml_nbytes(layer.wo) + ggml_nbytes(layer.ffn_down) + ggml_nbytes(layer.ffn_up); - } } - } break; + } + } break; case LLM_ARCH_STARCODER: + { + model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); + model.pos_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_POS_EMBD, "weight"), {n_embd, hparams.n_ctx_train}, GGML_BACKEND_CPU); + + // output { - model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); - model.pos_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_POS_EMBD, "weight"), {n_embd, hparams.n_ctx_train}, GGML_BACKEND_CPU); + ggml_backend_type backend_norm; + ggml_backend_type backend_output; - // output - { - ggml_backend_type backend_norm; - ggml_backend_type backend_output; - - if (n_gpu_layers > int(n_layer)) { - // norm is not performance relevant on its own but keeping it in VRAM reduces data copying - // on Windows however this is detrimental unless everything is on the GPU + if (n_gpu_layers > int(n_layer)) { + // norm is not performance relevant on its own but keeping it in VRAM reduces data copying + // on Windows however this is detrimental unless everything is on the GPU #ifndef _WIN32 - backend_norm = LLAMA_BACKEND_OFFLOAD; + backend_norm = llama_backend_offload; #else - backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; + backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : llama_backend_offload; #endif // _WIN32 - backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT; - } else { - backend_norm = GGML_BACKEND_CPU; - backend_output = GGML_BACKEND_CPU; - } - - model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm); - model.output_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, backend_norm); - model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output); - - if (backend_norm == GGML_BACKEND_GPU) { - vram_weights += ggml_nbytes(model.output_norm); - vram_weights += ggml_nbytes(model.output_norm_b); - } - if (backend_output == GGML_BACKEND_GPU_SPLIT) { - vram_weights += ggml_nbytes(model.output); - } + backend_output = llama_backend_offload_split; + } else { + backend_norm = GGML_BACKEND_CPU; + backend_output = GGML_BACKEND_CPU; } - const uint32_t n_ff = hparams.n_ff; + model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm); + model.output_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, backend_norm); + model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output); - const int i_gpu_start = n_layer - n_gpu_layers; + if (backend_norm == GGML_BACKEND_GPU) { + vram_weights += ggml_nbytes(model.output_norm); + vram_weights += ggml_nbytes(model.output_norm_b); + } + if (backend_output == GGML_BACKEND_GPU_SPLIT) { + vram_weights += ggml_nbytes(model.output); + } + } - model.layers.resize(n_layer); + const uint32_t n_ff = hparams.n_ff; - for (uint32_t i = 0; i < n_layer; ++i) { - const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT - const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT + const int i_gpu_start = n_layer - n_gpu_layers; - auto & layer = model.layers[i]; + model.layers.resize(n_layer); - layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend); - layer.attn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, backend); + for (uint32_t i = 0; i < n_layer; ++i) { + const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload; // NOLINT + const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload_split; // NOLINT - layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, backend_split); - layer.bqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, backend); + auto & layer = model.layers[i]; - layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split); - layer.bo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, backend); + layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend); + layer.attn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, backend); - layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend); - layer.ffn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}, backend); + layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, backend_split); + layer.bqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, backend); - layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, backend_split); - layer.ffn_down_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, backend); + layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split); + layer.bo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, backend); - layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); - layer.ffn_up_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, backend); + layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend); + layer.ffn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}, backend); - if (backend == GGML_BACKEND_GPU) { - vram_weights += + layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, backend_split); + layer.ffn_down_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, backend); + + layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + layer.ffn_up_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, backend); + + if (backend == GGML_BACKEND_GPU) { + vram_weights += ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.attn_norm_b) + ggml_nbytes(layer.wqkv) + ggml_nbytes(layer.bqkv) + ggml_nbytes(layer.wo) + ggml_nbytes(layer.bo) + ggml_nbytes(layer.ffn_norm) + ggml_nbytes(layer.ffn_norm_b) + ggml_nbytes(layer.ffn_down) + ggml_nbytes(layer.ffn_down_b) + ggml_nbytes(layer.ffn_up) + ggml_nbytes(layer.ffn_up_b); - } } - } break; + } + } break; case LLM_ARCH_PERSIMMON: + { + model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); + { - model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); + ggml_backend_type backend_norm; + ggml_backend_type backend_output; - { - ggml_backend_type backend_norm; - ggml_backend_type backend_output; - - if (n_gpu_layers > int(n_layer)) { - // norm is not performance relevant on its own but keeping it in VRAM reduces data copying - // on Windows however this is detrimental unless everything is on the GPU + if (n_gpu_layers > int(n_layer)) { +#ifdef GGML_USE_CUBLAS + if (n_gpu_layers > int(n_layer + 1)) { + LLAMA_LOG_ERROR("%s: CUDA backend missing Persimmon CUDA ops, can offload at most %ld layers. See: https://github.com/ggerganov/llama.cpp/issues/4038\n", + __func__, n_layer + 1); + throw std::runtime_error("Persimmon CUDA offload failed"); + } +#endif + // norm is not performance relevant on its own but keeping it in VRAM reduces data copying + // on Windows however this is detrimental unless everything is on the GPU #ifndef _WIN32 - backend_norm = LLAMA_BACKEND_OFFLOAD; + backend_norm = llama_backend_offload; #else - backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; + backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : llama_backend_offload; #endif // _WIN32 - backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT; - } else { - backend_norm = GGML_BACKEND_CPU; - backend_output = GGML_BACKEND_CPU; - } - - model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm); - model.output_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, backend_norm); - model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output); - - if (backend_norm == GGML_BACKEND_GPU) { - vram_weights += ggml_nbytes(model.output_norm); - vram_weights += ggml_nbytes(model.output_norm_b); - } - if (backend_output == GGML_BACKEND_GPU_SPLIT) { - vram_weights += ggml_nbytes(model.output); - } + backend_output = llama_backend_offload_split; + } else { + backend_norm = GGML_BACKEND_CPU; + backend_output = GGML_BACKEND_CPU; } - const uint32_t n_ff = hparams.n_ff; - const int i_gpu_start = n_layer - n_gpu_layers; - model.layers.resize(n_layer); - for (uint32_t i = 0; i < n_layer; ++i) { - const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; - const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; - auto & layer = model.layers[i]; - layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend); - layer.attn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, backend); - layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, backend_split); - layer.bqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, backend_split); - layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split); - layer.bo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, backend_split); - layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, backend_split); - layer.ffn_down_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, backend_split); - layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); - layer.ffn_up_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, backend_split); - layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend); - layer.ffn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}, backend); - layer.attn_q_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {64}, backend); - layer.attn_q_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_Q_NORM, "bias", i), {64}, backend); - layer.attn_k_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {64}, backend); - layer.attn_k_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_K_NORM, "bias", i), {64}, backend); + model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm); + model.output_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, backend_norm); + model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output); + + if (backend_norm == GGML_BACKEND_GPU) { + vram_weights += ggml_nbytes(model.output_norm); + vram_weights += ggml_nbytes(model.output_norm_b); } - } break; + if (backend_output == GGML_BACKEND_GPU_SPLIT) { + vram_weights += ggml_nbytes(model.output); + } + } + + const uint32_t n_ff = hparams.n_ff; + const int i_gpu_start = n_layer - n_gpu_layers; + model.layers.resize(n_layer); + for (uint32_t i = 0; i < n_layer; ++i) { + const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload; + const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload_split; + auto & layer = model.layers[i]; + layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend); + layer.attn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, backend); + layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, backend_split); + layer.bqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, backend); + layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split); + layer.bo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, backend); + layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, backend_split); + layer.ffn_down_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, backend); + layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + layer.ffn_up_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, backend); + layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend); + layer.ffn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}, backend); + layer.attn_q_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {64}, backend); + layer.attn_q_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_Q_NORM, "bias", i), {64}, backend); + layer.attn_k_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {64}, backend); + layer.attn_k_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_K_NORM, "bias", i), {64}, backend); + } + } break; case LLM_ARCH_BLOOM: + { + // TODO: CPU-only for now + + model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); + model.tok_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}, GGML_BACKEND_CPU); + model.tok_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias"), {n_embd}, GGML_BACKEND_CPU); + + // output { - // TODO: CPU-only for now + ggml_backend_type backend_norm; + ggml_backend_type backend_output; - model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); - model.tok_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}, GGML_BACKEND_CPU); - model.tok_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias"), {n_embd}, GGML_BACKEND_CPU); - - // output - { - ggml_backend_type backend_norm; - ggml_backend_type backend_output; - - if (n_gpu_layers > int(n_layer)) { - // norm is not performance relevant on its own but keeping it in VRAM reduces data copying - // on Windows however this is detrimental unless everything is on the GPU + if (n_gpu_layers > int(n_layer)) { + // norm is not performance relevant on its own but keeping it in VRAM reduces data copying + // on Windows however this is detrimental unless everything is on the GPU #ifndef _WIN32 - backend_norm = LLAMA_BACKEND_OFFLOAD; + backend_norm = llama_backend_offload; #else - backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; + backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : llama_backend_offload; #endif // _WIN32 - backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT; - } else { - backend_norm = GGML_BACKEND_CPU; - backend_output = GGML_BACKEND_CPU; - } - - model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm); - model.output_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, backend_norm); - model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output); - - if (backend_norm == GGML_BACKEND_GPU) { - vram_weights += ggml_nbytes(model.output_norm); - vram_weights += ggml_nbytes(model.output_norm_b); - } - if (backend_output == GGML_BACKEND_GPU_SPLIT) { - vram_weights += ggml_nbytes(model.output); - } + backend_output = llama_backend_offload_split; + } else { + backend_norm = GGML_BACKEND_CPU; + backend_output = GGML_BACKEND_CPU; } - const uint32_t n_ff = hparams.n_ff; + model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm); + model.output_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, backend_norm); + model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output); - const int i_gpu_start = n_layer - n_gpu_layers; + if (backend_norm == GGML_BACKEND_GPU) { + vram_weights += ggml_nbytes(model.output_norm); + vram_weights += ggml_nbytes(model.output_norm_b); + } + if (backend_output == GGML_BACKEND_GPU_SPLIT) { + vram_weights += ggml_nbytes(model.output); + } + } - model.layers.resize(n_layer); + const uint32_t n_ff = hparams.n_ff; - for (uint32_t i = 0; i < n_layer; ++i) { - const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT - const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT + const int i_gpu_start = n_layer - n_gpu_layers; - auto & layer = model.layers[i]; + model.layers.resize(n_layer); - layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend); - layer.attn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, backend); + for (uint32_t i = 0; i < n_layer; ++i) { + const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload; // NOLINT + const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload_split; // NOLINT - layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, backend_split); - layer.bqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, backend_split); + auto & layer = model.layers[i]; - layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split); - layer.bo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, backend_split); + layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend); + layer.attn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, backend); - layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend); - layer.ffn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}, backend); + layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, backend_split); + layer.bqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, backend); - layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, backend_split); - layer.ffn_down_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, backend_split); + layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split); + layer.bo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, backend); - layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); - layer.ffn_up_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, backend_split); + layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend); + layer.ffn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}, backend); - if (backend == GGML_BACKEND_GPU) { - vram_weights += + layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, backend_split); + layer.ffn_down_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, backend); + + layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + layer.ffn_up_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, backend); + + if (backend == GGML_BACKEND_GPU) { + vram_weights += ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.attn_norm_b) + ggml_nbytes(layer.wqkv) + ggml_nbytes(layer.bqkv) + ggml_nbytes(layer.wo) + ggml_nbytes(layer.bo) + ggml_nbytes(layer.ffn_norm) + ggml_nbytes(layer.ffn_norm_b) + ggml_nbytes(layer.ffn_up) + ggml_nbytes(layer.ffn_up_b) + ggml_nbytes(layer.ffn_down) + ggml_nbytes(layer.ffn_down_b); - } } - } break; + } + } break; case LLM_ARCH_MPT: + { + model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); + + // output { - model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); + ggml_backend_type backend_norm; + ggml_backend_type backend_output; - // output - { - ggml_backend_type backend_norm; - ggml_backend_type backend_output; - - if (n_gpu_layers > int(n_layer)) { - // norm is not performance relevant on its own but keeping it in VRAM reduces data copying - // on Windows however this is detrimental unless everything is on the GPU + if (n_gpu_layers > int(n_layer)) { + // norm is not performance relevant on its own but keeping it in VRAM reduces data copying + // on Windows however this is detrimental unless everything is on the GPU #ifndef _WIN32 - backend_norm = LLAMA_BACKEND_OFFLOAD; + backend_norm = llama_backend_offload; #else - backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; + backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : llama_backend_offload; #endif // _WIN32 - backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT; - } else { - backend_norm = GGML_BACKEND_CPU; - backend_output = GGML_BACKEND_CPU; - } - - model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm); - model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output); - - if (backend_norm == GGML_BACKEND_GPU) { - vram_weights += ggml_nbytes(model.output_norm); - } - if (backend_output == GGML_BACKEND_GPU_SPLIT) { - vram_weights += ggml_nbytes(model.output); - } + backend_output = llama_backend_offload_split; + } else { + backend_norm = GGML_BACKEND_CPU; + backend_output = GGML_BACKEND_CPU; } - const uint32_t n_ff = hparams.n_ff; + model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm); + model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output); - const int i_gpu_start = n_layer - n_gpu_layers; + if (backend_norm == GGML_BACKEND_GPU) { + vram_weights += ggml_nbytes(model.output_norm); + } + if (backend_output == GGML_BACKEND_GPU_SPLIT) { + vram_weights += ggml_nbytes(model.output); + } + } - model.layers.resize(n_layer); + const uint32_t n_ff = hparams.n_ff; - for (uint32_t i = 0; i < n_layer; ++i) { - const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT - const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT + const int i_gpu_start = n_layer - n_gpu_layers; - auto & layer = model.layers[i]; + model.layers.resize(n_layer); - layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend); - layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, backend_split); - layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split); + for (uint32_t i = 0; i < n_layer; ++i) { + const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload; // NOLINT + const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload_split; // NOLINT - layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend); + auto & layer = model.layers[i]; - layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); - layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend); + layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, backend_split); + layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split); - if (backend == GGML_BACKEND_GPU) { - vram_weights += + layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend); + + layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); + layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + + if (backend == GGML_BACKEND_GPU) { + vram_weights += ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.wqkv) + ggml_nbytes(layer.wo) + ggml_nbytes(layer.ffn_norm) + ggml_nbytes(layer.ffn_down) + ggml_nbytes(layer.ffn_up); - } } - } break; + } + } break; + case LLM_ARCH_STABLELM: + { + model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); + + // output + { + ggml_backend_type backend_norm; + ggml_backend_type backend_output; + + if (n_gpu_layers > int(n_layer)) { + // norm is not performance relevant on its own but keeping it in VRAM reduces data copying + // on Windows however this is detrimental unless everything is on the GPU +#ifndef _WIN32 + backend_norm = llama_backend_offload; +#else + backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : llama_backend_offload; +#endif // _WIN32 + + backend_output = llama_backend_offload_split; + } else { + backend_norm = GGML_BACKEND_CPU; + backend_output = GGML_BACKEND_CPU; + } + + model.output_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, backend_norm); + model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm); + model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output); + + if (backend_norm == GGML_BACKEND_GPU) { + vram_weights += ggml_nbytes(model.output_norm); + } + if (backend_output == GGML_BACKEND_GPU_SPLIT) { + vram_weights += ggml_nbytes(model.output); + } + } + + const uint32_t n_ff = hparams.n_ff; + + const int i_gpu_start = n_layer - n_gpu_layers; + + model.layers.resize(n_layer); + + for (uint32_t i = 0; i < n_layer; ++i) { + /* + llama_model_loader: - tensor 4: blk.0.attn_output.weight f16 [ 2560, 2560, 1, 1 ] + */ + const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload; // NOLINT + const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload_split; // NOLINT + + auto & layer = model.layers[i]; + + layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend); + layer.attn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, backend); + + layer.wq = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}, backend_split); + layer.wk = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}, backend_split); + layer.wv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}, backend_split); + layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split); + + layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend); + layer.ffn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}, backend); + + layer.ffn_gate = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, backend_split); + layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); + layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + + if (backend == GGML_BACKEND_GPU) { + vram_weights += + ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.wq) + ggml_nbytes(layer.wk) + + ggml_nbytes(layer.wv) + ggml_nbytes(layer.wo) + ggml_nbytes(layer.ffn_norm) + + ggml_nbytes(layer.ffn_gate) + ggml_nbytes(layer.ffn_down) + ggml_nbytes(layer.ffn_up); + } + } + } break; + default: throw std::runtime_error("unknown architecture"); } @@ -2993,10 +3307,10 @@ static void llm_load_tensors( { // this is the total memory required to run the inference size_t mem_required = - ctx_size + - mmapped_size - vram_weights; // weights in VRAM not in memory + ctx_size + + mmapped_size - vram_weights; // weights in VRAM not in memory - LLAMA_LOG_INFO("%s: mem required = %7.2f MB\n", __func__, mem_required / 1024.0 / 1024.0); + LLAMA_LOG_INFO("%s: mem required = %7.2f MiB\n", __func__, mem_required / 1024.0 / 1024.0); #if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) const int n_gpu = std::min(n_gpu_layers, int(hparams.n_layer)); @@ -3015,7 +3329,7 @@ static void llm_load_tensors( #endif // GGML_USE_CUBLAS LLAMA_LOG_INFO("%s: offloaded %d/%d layers to GPU\n", __func__, std::min(n_gpu_layers, max_offloadable_layers), max_backend_supported_layers); - LLAMA_LOG_INFO("%s: VRAM used: %.2f MB\n", __func__, vram_weights / 1024.0 / 1024.0); + LLAMA_LOG_INFO("%s: VRAM used: %.2f MiB\n", __func__, vram_weights / 1024.0 / 1024.0); #else (void) n_gpu_layers; #endif // defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) @@ -3047,21 +3361,11 @@ static void llm_load_tensors( model.t_load_us = ggml_time_us() - model.t_start_us; } -static bool llama_model_load( - const std::string & fname, - llama_model & model, - int n_gpu_layers, - int main_gpu, - const float * tensor_split, - bool use_mmap, - bool use_mlock, - bool vocab_only, - llama_progress_callback progress_callback, - void *progress_callback_user_data) { +static bool llama_model_load(const std::string & fname, llama_model & model, const llama_model_params & params) { try { - llama_model_loader ml(fname, use_mmap); + llama_model_loader ml(fname, params.use_mmap); - model.hparams.vocab_only = vocab_only; + model.hparams.vocab_only = params.vocab_only; llm_load_arch (ml, model); llm_load_hparams(ml, model); @@ -3073,15 +3377,15 @@ static bool llama_model_load( throw std::runtime_error("vocab size mismatch"); } - if (vocab_only) { + if (params.vocab_only) { LLAMA_LOG_INFO("%s: vocab only - skipping tensors\n", __func__); return true; } llm_load_tensors( - ml, model, n_gpu_layers, - main_gpu, tensor_split, - use_mlock, progress_callback, progress_callback_user_data); + ml, model, params.n_gpu_layers, params.main_gpu, params.tensor_split, params.use_mlock, + params.progress_callback, params.progress_callback_user_data + ); } catch (const std::exception & err) { LLAMA_LOG_ERROR("error loading model: %s\n", err.what()); return false; @@ -3122,9 +3426,9 @@ enum llm_norm_type { static struct ggml_tensor * llm_build_inp_embd( struct ggml_context * ctx, const llama_hparams & hparams, - const llama_batch & batch, - struct ggml_tensor * tok_embd, - const llm_build_cb & cb) { + const llama_batch & batch, + struct ggml_tensor * tok_embd, + const llm_build_cb & cb) { const int64_t n_embd = hparams.n_embd; struct ggml_tensor * inpL; @@ -3148,20 +3452,26 @@ static struct ggml_tensor * llm_build_inp_embd( // Persimmon: n_rot = n_embd_head/2 // Other: n_rot = n_embd_head static void llm_build_k_shift( - struct ggml_context * ctx, - const llama_hparams & hparams, - const llama_kv_cache & kv, - struct ggml_cgraph * graph, - llm_rope_type type, - int64_t n_ctx, - int64_t n_rot, - float freq_base, - float freq_scale, - const llm_build_cb & cb) { + struct ggml_context * ctx, + const llama_hparams & hparams, + const llama_cparams & cparams, + const llama_kv_cache & kv, + struct ggml_cgraph * graph, + llm_rope_type type, + int64_t n_ctx, + int64_t n_rot, + float freq_base, + float freq_scale, + const llm_build_cb & cb) { const int64_t n_layer = hparams.n_layer; const int64_t n_head_kv = hparams.n_head_kv; const int64_t n_embd_gqa = hparams.n_embd_gqa(); const int64_t n_embd_head = hparams.n_embd_head(); + const int32_t n_orig_ctx = cparams.n_yarn_orig_ctx; + const float ext_factor = cparams.yarn_ext_factor; + const float attn_factor = cparams.yarn_attn_factor; + const float beta_fast = cparams.yarn_beta_fast; + const float beta_slow = cparams.yarn_beta_slow; GGML_ASSERT(n_embd_head % n_rot == 0); @@ -3178,14 +3488,15 @@ static void llm_build_k_shift( for (int il = 0; il < n_layer; ++il) { struct ggml_tensor * tmp = - // we rotate only the first n_rot dimensions - ggml_rope_custom_inplace(ctx, - ggml_view_3d(ctx, kv.k, - n_rot, n_head_kv, n_ctx, - ggml_element_size(kv.k)*n_embd_head, - ggml_element_size(kv.k)*n_embd_gqa, - ggml_element_size(kv.k)*n_embd_gqa*n_ctx*il), - K_shift, n_rot, rope_type, 0, freq_base, freq_scale); + // we rotate only the first n_rot dimensions + ggml_rope_custom_inplace(ctx, + ggml_view_3d(ctx, kv.k, + n_rot, n_head_kv, n_ctx, + ggml_element_size(kv.k)*n_embd_head, + ggml_element_size(kv.k)*n_embd_gqa, + ggml_element_size(kv.k)*n_embd_gqa*n_ctx*il), + K_shift, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); cb(tmp, "K_shifted", il); ggml_build_forward_expand(graph, tmp); } @@ -3194,15 +3505,15 @@ static void llm_build_k_shift( static void llm_build_kv_store( struct ggml_context * ctx, const llama_hparams & hparams, - const llama_kv_cache & kv, - struct ggml_cgraph * graph, - struct ggml_tensor * k_cur, - struct ggml_tensor * v_cur, - int64_t n_ctx, - int32_t n_tokens, - int32_t kv_head, - const llm_build_cb & cb, - int64_t il) { + const llama_kv_cache & kv, + struct ggml_cgraph * graph, + struct ggml_tensor * k_cur, + struct ggml_tensor * v_cur, + int64_t n_ctx, + int32_t n_tokens, + int32_t kv_head, + const llm_build_cb & cb, + int64_t il) { const int64_t n_embd_gqa = hparams.n_embd_gqa(); // compute the transposed [n_tokens, n_embd] V matrix @@ -3211,12 +3522,12 @@ static void llm_build_kv_store( cb(v_cur_t, "v_cur_t", il); struct ggml_tensor * k_cache_view = ggml_view_1d(ctx, kv.k, n_tokens*n_embd_gqa, - (ggml_element_size(kv.k)*n_embd_gqa)*(il*n_ctx + kv_head)); + (ggml_element_size(kv.k)*n_embd_gqa)*(il*n_ctx + kv_head)); cb(k_cache_view, "k_cache_view", il); struct ggml_tensor * v_cache_view = ggml_view_2d(ctx, kv.v, n_tokens, n_embd_gqa, - ( n_ctx)*ggml_element_size(kv.v), - (il*n_ctx)*ggml_element_size(kv.v)*n_embd_gqa + kv_head*ggml_element_size(kv.v)); + ( n_ctx)*ggml_element_size(kv.v), + (il*n_ctx)*ggml_element_size(kv.v)*n_embd_gqa + kv_head*ggml_element_size(kv.v)); cb(v_cache_view, "v_cache_view", il); // important: storing RoPE-ed version of K in the KV cache! @@ -3226,13 +3537,13 @@ static void llm_build_kv_store( static struct ggml_tensor * llm_build_norm( struct ggml_context * ctx, - struct ggml_tensor * cur, + struct ggml_tensor * cur, const llama_hparams & hparams, - struct ggml_tensor * mw, - struct ggml_tensor * mb, - llm_norm_type type, - const llm_build_cb & cb, - int il) { + struct ggml_tensor * mw, + struct ggml_tensor * mb, + llm_norm_type type, + const llm_build_cb & cb, + int il) { switch (type) { case LLM_NORM: cur = ggml_norm (ctx, cur, hparams.f_norm_eps); break; case LLM_NORM_RMS: cur = ggml_rms_norm(ctx, cur, hparams.f_norm_rms_eps); break; @@ -3258,17 +3569,17 @@ static struct ggml_tensor * llm_build_norm( static struct ggml_tensor * llm_build_ffn( struct ggml_context * ctx, - struct ggml_tensor * cur, - struct ggml_tensor * up, - struct ggml_tensor * up_b, - struct ggml_tensor * gate, - struct ggml_tensor * gate_b, - struct ggml_tensor * down, - struct ggml_tensor * down_b, - llm_ffn_op_type type_op, - llm_ffn_gate_type type_gate, - const llm_build_cb & cb, - int il) { + struct ggml_tensor * cur, + struct ggml_tensor * up, + struct ggml_tensor * up_b, + struct ggml_tensor * gate, + struct ggml_tensor * gate_b, + struct ggml_tensor * down, + struct ggml_tensor * down_b, + llm_ffn_op_type type_op, + llm_ffn_gate_type type_gate, + const llm_build_cb & cb, + int il) { struct ggml_tensor * tmp = ggml_mul_mat(ctx, up, cur); cb(tmp, "ffn_up", il); @@ -3280,15 +3591,15 @@ static struct ggml_tensor * llm_build_ffn( if (gate) { switch (type_gate) { case LLM_FFN_SEQ: - { - cur = ggml_mul_mat(ctx, gate, tmp); - cb(cur, "ffn_gate", il); - } break; + { + cur = ggml_mul_mat(ctx, gate, tmp); + cb(cur, "ffn_gate", il); + } break; case LLM_FFN_PAR: - { - cur = ggml_mul_mat(ctx, gate, cur); - cb(cur, "ffn_gate", il); - } break; + { + cur = ggml_mul_mat(ctx, gate, cur); + cb(cur, "ffn_gate", il); + } break; } if (gate_b) { @@ -3301,28 +3612,28 @@ static struct ggml_tensor * llm_build_ffn( switch (type_op) { case LLM_FFN_SILU: - { - cur = ggml_silu(ctx, cur); - cb(cur, "ffn_silu", il); - } break; + { + cur = ggml_silu(ctx, cur); + cb(cur, "ffn_silu", il); + } break; case LLM_FFN_GELU: - { - cur = ggml_gelu(ctx, cur); - cb(cur, "ffn_gelu", il); - } break; + { + cur = ggml_gelu(ctx, cur); + cb(cur, "ffn_gelu", il); + } break; case LLM_FFN_RELU: - { - cur = ggml_relu(ctx, cur); - cb(cur, "ffn_relu", il); - } break; + { + cur = ggml_relu(ctx, cur); + cb(cur, "ffn_relu", il); + } break; case LLM_FFN_RELU_SQR: - { - cur = ggml_relu(ctx, cur); - cb(cur, "ffn_relu", il); + { + cur = ggml_relu(ctx, cur); + cb(cur, "ffn_relu", il); - cur = ggml_sqr(ctx, cur); - cb(cur, "ffn_sqr(relu)", il); - } break; + cur = ggml_sqr(ctx, cur); + cb(cur, "ffn_sqr(relu)", il); + } break; } if (type_gate == LLM_FFN_PAR) { @@ -3346,18 +3657,18 @@ static struct ggml_tensor * llm_build_ffn( static struct ggml_tensor * llm_build_kqv( struct ggml_context * ctx, const llama_hparams & hparams, - const llama_kv_cache & kv, - struct ggml_tensor * wo, - struct ggml_tensor * wo_b, - struct ggml_tensor * q_cur, - struct ggml_tensor * kq_scale, - struct ggml_tensor * kq_mask, - int64_t n_ctx, - int32_t n_tokens, - int32_t n_kv, - float max_alibi_bias, - const llm_build_cb & cb, - int il) { + const llama_kv_cache & kv, + struct ggml_tensor * wo, + struct ggml_tensor * wo_b, + struct ggml_tensor * q_cur, + struct ggml_tensor * kq_scale, + struct ggml_tensor * kq_mask, + int64_t n_ctx, + int32_t n_tokens, + int32_t n_kv, + float max_alibi_bias, + const llm_build_cb & cb, + int il) { const int64_t n_embd = hparams.n_embd; const int64_t n_head = hparams.n_head; const int64_t n_head_kv = hparams.n_head_kv; @@ -3368,11 +3679,11 @@ static struct ggml_tensor * llm_build_kqv( cb(q, "q", il); struct ggml_tensor * k = - ggml_view_3d(ctx, kv.k, - n_embd_head, n_kv, n_head_kv, - ggml_element_size(kv.k)*n_embd_gqa, - ggml_element_size(kv.k)*n_embd_head, - ggml_element_size(kv.k)*n_embd_gqa*n_ctx*il); + ggml_view_3d(ctx, kv.k, + n_embd_head, n_kv, n_head_kv, + ggml_element_size(kv.k)*n_embd_gqa, + ggml_element_size(kv.k)*n_embd_head, + ggml_element_size(kv.k)*n_embd_gqa*n_ctx*il); cb(k, "k", il); struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q); @@ -3397,11 +3708,11 @@ static struct ggml_tensor * llm_build_kqv( // split cached v into n_head heads struct ggml_tensor * v = - ggml_view_3d(ctx, kv.v, - n_kv, n_embd_head, n_head_kv, - ggml_element_size(kv.v)*n_ctx, - ggml_element_size(kv.v)*n_ctx*n_embd_head, - ggml_element_size(kv.v)*n_ctx*n_embd_gqa*il); + ggml_view_3d(ctx, kv.v, + n_kv, n_embd_head, n_head_kv, + ggml_element_size(kv.v)*n_ctx, + ggml_element_size(kv.v)*n_ctx*n_embd_head, + ggml_element_size(kv.v)*n_ctx*n_embd_gqa*il); cb(v, "v", il); struct ggml_tensor * kqv = ggml_mul_mat(ctx, v, kq); @@ -3442,12 +3753,17 @@ struct llm_build_context { const float freq_base; const float freq_scale; + const float ext_factor; + const float attn_factor; + const float beta_fast; + const float beta_slow; const float norm_eps; const float norm_rms_eps; const int32_t n_tokens; const int32_t n_kv; // size of KV cache to consider (n_kv <= n_ctx) const int32_t kv_head; // index of where we store new KV data in the cache + const int32_t n_orig_ctx; const bool do_rope_shift; @@ -3459,42 +3775,47 @@ struct llm_build_context { // TODO: consider making the entire interface noexcept llm_build_context( - llama_context & lctx, - const llama_batch & batch, - const llm_build_cb & cb, - bool worst_case) : - model (lctx.model), - hparams (model.hparams), - cparams (lctx.cparams), - batch (batch), - kv_self (lctx.kv_self), - n_embd (hparams.n_embd), - n_layer (hparams.n_layer), - n_ctx (cparams.n_ctx), - n_head (hparams.n_head), - n_head_kv (hparams.n_head_kv), - n_embd_head (hparams.n_embd_head()), - n_embd_gqa (hparams.n_embd_gqa()), - freq_base (cparams.rope_freq_base), - freq_scale (cparams.rope_freq_scale), - norm_eps (hparams.f_norm_eps), - norm_rms_eps (hparams.f_norm_rms_eps), - n_tokens (batch.n_tokens), - n_kv (worst_case ? n_ctx : kv_self.n), - kv_head (worst_case ? n_ctx - n_tokens : kv_self.head), - do_rope_shift (worst_case || kv_self.has_shift), - cb (cb), - buf_compute (lctx.buf_compute) { - GGML_ASSERT(!!kv_self.ctx); + llama_context & lctx, + const llama_batch & batch, + const llm_build_cb & cb, + bool worst_case) : + model (lctx.model), + hparams (model.hparams), + cparams (lctx.cparams), + batch (batch), + kv_self (lctx.kv_self), + n_embd (hparams.n_embd), + n_layer (hparams.n_layer), + n_ctx (cparams.n_ctx), + n_head (hparams.n_head), + n_head_kv (hparams.n_head_kv), + n_embd_head (hparams.n_embd_head()), + n_embd_gqa (hparams.n_embd_gqa()), + freq_base (cparams.rope_freq_base), + freq_scale (cparams.rope_freq_scale), + ext_factor (cparams.yarn_ext_factor), + attn_factor (cparams.yarn_attn_factor), + beta_fast (cparams.yarn_beta_fast), + beta_slow (cparams.yarn_beta_slow), + norm_eps (hparams.f_norm_eps), + norm_rms_eps (hparams.f_norm_rms_eps), + n_tokens (batch.n_tokens), + n_kv (worst_case ? n_ctx : kv_self.n), + kv_head (worst_case ? n_ctx - n_tokens : kv_self.head), + n_orig_ctx (cparams.n_yarn_orig_ctx), + do_rope_shift (worst_case || kv_self.has_shift), + cb (cb), + buf_compute (lctx.buf_compute) { + GGML_ASSERT(!!kv_self.ctx); - // all initializations should be done in init() - } + // all initializations should be done in init() + } void init() { struct ggml_init_params params = { - /*.mem_size =*/ buf_compute.size, - /*.mem_buffer =*/ buf_compute.data, - /*.no_alloc =*/ true, + /*.mem_size =*/ buf_compute.size, + /*.mem_buffer =*/ buf_compute.data, + /*.no_alloc =*/ true, }; ctx0 = ggml_init(params); @@ -3508,7 +3829,7 @@ struct llm_build_context { } struct ggml_cgraph * build_llama() { - struct ggml_cgraph * gf = ggml_new_graph(ctx0); + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); GGML_ASSERT(n_embd_head == hparams.n_rot); @@ -3532,7 +3853,7 @@ struct llm_build_context { // shift the entire K-cache if needed if (do_rope_shift) { - llm_build_k_shift(ctx0, hparams, kv_self, gf, LLM_ROPE, n_ctx, n_embd_head, freq_base, freq_scale, cb); + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, n_embd_head, freq_base, freq_scale, cb); } for (int il = 0; il < n_layer; ++il) { @@ -3540,8 +3861,8 @@ struct llm_build_context { // norm cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, cb, il); + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, cb, il); cb(cur, "attn_norm", il); // self-attention @@ -3556,17 +3877,25 @@ struct llm_build_context { struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur); cb(Vcur, "Vcur", il); - Qcur = ggml_rope_custom(ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, n_embd_head, 0, 0, freq_base, freq_scale); + Qcur = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, + n_embd_head, 0, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); cb(Qcur, "Qcur", il); - Kcur = ggml_rope_custom(ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, n_embd_head, 0, 0, freq_base, freq_scale); + Kcur = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, + n_embd_head, 0, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); cb(Kcur, "Kcur", il); llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); cur = llm_build_kqv(ctx0, hparams, kv_self, - model.layers[il].wo, NULL, - Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il); + model.layers[il].wo, NULL, + Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il); cb(cur, "kqv_out", il); } @@ -3576,15 +3905,15 @@ struct llm_build_context { // feed-forward network { cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, cb, il); + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, cb, il); cb(cur, "ffn_norm", il); cur = llm_build_ffn(ctx0, cur, - model.layers[il].ffn_up, NULL, - model.layers[il].ffn_gate, NULL, - model.layers[il].ffn_down, NULL, - LLM_FFN_SILU, LLM_FFN_PAR, cb, il); + model.layers[il].ffn_up, NULL, + model.layers[il].ffn_gate, NULL, + model.layers[il].ffn_down, NULL, + LLM_FFN_SILU, LLM_FFN_PAR, cb, il); cb(cur, "ffn_out", il); } @@ -3598,8 +3927,8 @@ struct llm_build_context { cur = inpL; cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, NULL, - LLM_NORM_RMS, cb, -1); + model.output_norm, NULL, + LLM_NORM_RMS, cb, -1); cb(cur, "result_norm", -1); // lm_head @@ -3612,7 +3941,7 @@ struct llm_build_context { } struct ggml_cgraph * build_baichuan() { - struct ggml_cgraph * gf = ggml_new_graph(ctx0); + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); struct ggml_tensor * cur; struct ggml_tensor * inpL; @@ -3634,15 +3963,15 @@ struct llm_build_context { // shift the entire K-cache if needed if (do_rope_shift) { - llm_build_k_shift(ctx0, hparams, kv_self, gf, LLM_ROPE, n_ctx, n_embd_head, freq_base, freq_scale, cb); + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, n_embd_head, freq_base, freq_scale, cb); } for (int il = 0; il < n_layer; ++il) { struct ggml_tensor * inpSA = inpL; cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, cb, il); + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, cb, il); cb(cur, "attn_norm", il); // self-attention @@ -3658,8 +3987,16 @@ struct llm_build_context { switch (model.type) { case MODEL_7B: - Qcur = ggml_rope_custom(ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, n_embd_head, 0, 0, freq_base, freq_scale); - Kcur = ggml_rope_custom(ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, n_embd_head, 0, 0, freq_base, freq_scale); + Qcur = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, + n_embd_head, 0, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + Kcur = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, + n_embd_head, 0, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); break; case MODEL_13B: Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd/n_head, n_head, n_tokens); @@ -3677,8 +4014,8 @@ struct llm_build_context { const float max_alibi_bias = model.type == MODEL_13B ? 8.0f : -1.0f; cur = llm_build_kqv(ctx0, hparams, kv_self, - model.layers[il].wo, NULL, - Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, max_alibi_bias, cb, il); + model.layers[il].wo, NULL, + Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, max_alibi_bias, cb, il); cb(cur, "kqv_out", il); } @@ -3688,15 +4025,15 @@ struct llm_build_context { // feed-forward network { cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, cb, il); + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, cb, il); cb(cur, "ffn_norm", il); cur = llm_build_ffn(ctx0, cur, - model.layers[il].ffn_up, NULL, - model.layers[il].ffn_gate, NULL, - model.layers[il].ffn_down, NULL, - LLM_FFN_SILU, LLM_FFN_PAR, cb, il); + model.layers[il].ffn_up, NULL, + model.layers[il].ffn_gate, NULL, + model.layers[il].ffn_down, NULL, + LLM_FFN_SILU, LLM_FFN_PAR, cb, il); cb(cur, "ffn_out", il); } @@ -3710,8 +4047,8 @@ struct llm_build_context { cur = inpL; cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, NULL, - LLM_NORM_RMS, cb, -1); + model.output_norm, NULL, + LLM_NORM_RMS, cb, -1); cb(cur, "result_norm", -1); // lm_head @@ -3724,7 +4061,7 @@ struct llm_build_context { } struct ggml_cgraph * build_falcon() { - struct ggml_cgraph * gf = ggml_new_graph(ctx0); + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); struct ggml_tensor * cur; struct ggml_tensor * inpL; @@ -3746,16 +4083,16 @@ struct llm_build_context { // shift the entire K-cache if needed if (do_rope_shift) { - llm_build_k_shift(ctx0, hparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, n_embd_head, freq_base, freq_scale, cb); + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, n_embd_head, freq_base, freq_scale, cb); } for (int il = 0; il < n_layer; ++il) { struct ggml_tensor * attn_norm; attn_norm = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, - model.layers[il].attn_norm_b, - LLM_NORM, cb, il); + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, cb, il); cb(attn_norm, "attn_norm", il); // self-attention @@ -3763,9 +4100,9 @@ struct llm_build_context { if (model.layers[il].attn_norm_2) { // Falcon-40B cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm_2, - model.layers[il].attn_norm_2_b, - LLM_NORM, cb, il); + model.layers[il].attn_norm_2, + model.layers[il].attn_norm_2_b, + LLM_NORM, cb, il); cb(cur, "attn_norm_2", il); } else { cur = attn_norm; @@ -3786,17 +4123,23 @@ struct llm_build_context { Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); // using mode = 2 for neox mode - Qcur = ggml_rope_custom(ctx0, Qcur, inp_pos, n_embd_head, 2, 0, freq_base, freq_scale); + Qcur = ggml_rope_custom( + ctx0, Qcur, inp_pos, n_embd_head, 2, 0, n_orig_ctx, + freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow + ); cb(Qcur, "Qcur", il); - Kcur = ggml_rope_custom(ctx0, Kcur, inp_pos, n_embd_head, 2, 0, freq_base, freq_scale); + Kcur = ggml_rope_custom( + ctx0, Kcur, inp_pos, n_embd_head, 2, 0, n_orig_ctx, + freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow + ); cb(Kcur, "Kcur", il); llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); cur = llm_build_kqv(ctx0, hparams, kv_self, - model.layers[il].wo, NULL, - Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il); + model.layers[il].wo, NULL, + Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il); cb(cur, "kqv_out", il); } @@ -3805,10 +4148,10 @@ struct llm_build_context { // feed forward { cur = llm_build_ffn(ctx0, attn_norm, // !! use the attn norm, not the result - model.layers[il].ffn_up, NULL, - NULL, NULL, - model.layers[il].ffn_down, NULL, - LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); + model.layers[il].ffn_up, NULL, + NULL, NULL, + model.layers[il].ffn_down, NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); cb(cur, "ffn_out", il); } @@ -3826,9 +4169,9 @@ struct llm_build_context { // norm cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, - model.output_norm_b, - LLM_NORM, cb, -1); + model.output_norm, + model.output_norm_b, + LLM_NORM, cb, -1); cb(cur, "result_norm", -1); cur = ggml_mul_mat(ctx0, model.output, cur); @@ -3840,7 +4183,7 @@ struct llm_build_context { } struct ggml_cgraph * build_starcoder() { - struct ggml_cgraph * gf = ggml_new_graph(ctx0); + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); struct ggml_tensor * cur; struct ggml_tensor * pos; @@ -3869,9 +4212,9 @@ struct llm_build_context { for (int il = 0; il < n_layer; ++il) { cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, - model.layers[il].attn_norm_b, - LLM_NORM, cb, il); + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, cb, il); cb(cur, "attn_norm", il); // self-attention @@ -3895,8 +4238,8 @@ struct llm_build_context { llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); cur = llm_build_kqv(ctx0, hparams, kv_self, - model.layers[il].wo, model.layers[il].bo, - Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il); + model.layers[il].wo, model.layers[il].bo, + Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il); cb(cur, "kqv_out", il); } @@ -3907,16 +4250,16 @@ struct llm_build_context { // FF { cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, - model.layers[il].ffn_norm_b, - LLM_NORM, cb, il); + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, cb, il); cb(cur, "ffn_norm", il); cur = llm_build_ffn(ctx0, cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, - NULL, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, - LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); + model.layers[il].ffn_up, model.layers[il].ffn_up_b, + NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, + LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); cb(cur, "ffn_out", il); } @@ -3925,9 +4268,9 @@ struct llm_build_context { } cur = llm_build_norm(ctx0, inpL, hparams, - model.output_norm, - model.output_norm_b, - LLM_NORM, cb, -1); + model.output_norm, + model.output_norm_b, + LLM_NORM, cb, -1); cb(cur, "result_norm", -1); cur = ggml_mul_mat(ctx0, model.output, cur); @@ -3939,7 +4282,7 @@ struct llm_build_context { } struct ggml_cgraph * build_persimmon() { - struct ggml_cgraph * gf = ggml_new_graph(ctx0); + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); const int64_t n_rot = n_embd_head / 2; @@ -3960,16 +4303,16 @@ struct llm_build_context { cb(KQ_mask, "KQ_mask", -1); if (do_rope_shift) { - llm_build_k_shift(ctx0, hparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, n_embd_head, freq_base, freq_scale, cb); + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, n_embd_head, freq_base, freq_scale, cb); } for (int il = 0; il < n_layer; ++il) { struct ggml_tensor * residual = inpL; cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, - model.layers[il].attn_norm_b, - LLM_NORM, cb, il); + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, cb, il); cb(cur, "attn_norm", il); // self attention @@ -3994,7 +4337,7 @@ struct llm_build_context { ggml_element_size(tmpqkv_perm) * n_embd_head, ggml_element_size(tmpqkv_perm) * n_embd_head * n_head, 0 - ); + ); cb(tmpq, "tmpq", il); struct ggml_tensor * tmpk = ggml_view_3d( @@ -4002,20 +4345,20 @@ struct llm_build_context { ggml_element_size(tmpqkv_perm) * n_embd_head, ggml_element_size(tmpqkv_perm) * n_embd_head * n_head, ggml_element_size(tmpqkv_perm) * n_embd_head * n_head * n_tokens - ); + ); cb(tmpk, "tmpk", il); // Q/K Layernorm tmpq = llm_build_norm(ctx0, tmpq, hparams, - model.layers[il].attn_q_norm, - model.layers[il].attn_q_norm_b, - LLM_NORM, cb, il); + model.layers[il].attn_q_norm, + model.layers[il].attn_q_norm_b, + LLM_NORM, cb, il); cb(tmpq, "tmpq", il); tmpk = llm_build_norm(ctx0, tmpk, hparams, - model.layers[il].attn_k_norm, - model.layers[il].attn_k_norm_b, - LLM_NORM, cb, il); + model.layers[il].attn_k_norm, + model.layers[il].attn_k_norm_b, + LLM_NORM, cb, il); cb(tmpk, "tmpk", il); // RoPE the first n_rot of q/k, pass the other half, and concat. @@ -4024,7 +4367,7 @@ struct llm_build_context { ggml_element_size(tmpq) * n_embd_head, ggml_element_size(tmpq) * n_embd_head * n_head, 0 - ); + ); cb(qrot, "qrot", il); struct ggml_tensor * krot = ggml_view_3d( @@ -4032,7 +4375,7 @@ struct llm_build_context { ggml_element_size(tmpk) * n_embd_head, ggml_element_size(tmpk) * n_embd_head * n_head, 0 - ); + ); cb(krot, "krot", il); // get the second half of tmpq, e.g tmpq[n_rot:, :, :] @@ -4041,7 +4384,7 @@ struct llm_build_context { ggml_element_size(tmpq) * n_embd_head, ggml_element_size(tmpq) * n_embd_head * n_head, ggml_element_size(tmpq) * n_rot - ); + ); cb(qpass, "qpass", il); struct ggml_tensor * kpass = ggml_view_3d( @@ -4049,17 +4392,19 @@ struct llm_build_context { ggml_element_size(tmpk) * n_embd_head, ggml_element_size(tmpk) * n_embd_head * n_head, ggml_element_size(tmpk) * n_rot - ); + ); cb(kpass, "kpass", il); struct ggml_tensor * qrotated = ggml_rope_custom( - ctx0, qrot, inp_pos, n_rot, 2, 0, freq_base, freq_scale - ); + ctx0, qrot, inp_pos, n_rot, 2, 0, n_orig_ctx, + freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow + ); cb(qrotated, "qrotated", il); struct ggml_tensor * krotated = ggml_rope_custom( - ctx0, krot, inp_pos, n_rot, 2, 0, freq_base, freq_scale - ); + ctx0, krot, inp_pos, n_rot, 2, 0, n_orig_ctx, + freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow + ); cb(krotated, "krotated", il); // ggml currently only supports concatenation on dim=2 @@ -4082,7 +4427,7 @@ struct llm_build_context { struct ggml_tensor * Kcur = ggml_concat(ctx0, krotated, kpass); cb(Kcur, "Kcur", il); - struct ggml_tensor * Q = ggml_cont(ctx0, ggml_permute(ctx0, Qcur, 1, 2, 0, 3)); + struct ggml_tensor * Q = ggml_cont(ctx0, ggml_permute(ctx0, Qcur, 2, 1, 0, 3)); cb(Q, "Q", il); Kcur = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 2, 1, 0, 3)); @@ -4093,15 +4438,15 @@ struct llm_build_context { ggml_element_size(tmpqkv_perm) * n_embd_head, ggml_element_size(tmpqkv_perm) * n_embd_head * n_head, ggml_element_size(tmpqkv_perm) * n_embd_head * n_head * n_tokens * 2 - ); + ); cb(Vcur, "Vcur", il); llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); // TODO: not tested, could be broken cur = llm_build_kqv(ctx0, hparams, kv_self, - model.layers[il].wo, model.layers[il].bo, - Q, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il); + model.layers[il].wo, model.layers[il].bo, + Q, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il); cb(cur, "kqv_out", il); } @@ -4111,16 +4456,16 @@ struct llm_build_context { // feed-forward network { cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, - model.layers[il].ffn_norm_b, - LLM_NORM, cb, il); + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, cb, il); cb(cur, "ffn_norm", il); cur = llm_build_ffn(ctx0, cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, - NULL, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, - LLM_FFN_RELU_SQR, LLM_FFN_SEQ, cb, il); + model.layers[il].ffn_up, model.layers[il].ffn_up_b, + NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, + LLM_FFN_RELU_SQR, LLM_FFN_SEQ, cb, il); cb(cur, "ffn_out", il); } @@ -4133,9 +4478,9 @@ struct llm_build_context { cur = inpL; cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, - model.output_norm_b, - LLM_NORM, cb, -1); + model.output_norm, + model.output_norm_b, + LLM_NORM, cb, -1); cb(cur, "result_norm", -1); cur = ggml_mul_mat(ctx0, model.output, cur); @@ -4147,7 +4492,7 @@ struct llm_build_context { } struct ggml_cgraph * build_refact() { - struct ggml_cgraph * gf = ggml_new_graph(ctx0); + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); struct ggml_tensor * cur; struct ggml_tensor * inpL; @@ -4167,8 +4512,8 @@ struct llm_build_context { struct ggml_tensor * inpSA = inpL; cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, cb, il); + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, cb, il); cb(cur, "attn_norm", il); // self-attention @@ -4191,8 +4536,8 @@ struct llm_build_context { llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); cur = llm_build_kqv(ctx0, hparams, kv_self, - model.layers[il].wo, NULL, - Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, 8.0f, cb, il); + model.layers[il].wo, NULL, + Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, 8.0f, cb, il); cb(cur, "kqv_out", il); } @@ -4202,15 +4547,15 @@ struct llm_build_context { // feed-forward network { cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, cb, il); + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, cb, il); cb(cur, "ffn_norm", il); cur = llm_build_ffn(ctx0, cur, - model.layers[il].ffn_up, NULL, - model.layers[il].ffn_gate, NULL, - model.layers[il].ffn_down, NULL, - LLM_FFN_SILU, LLM_FFN_PAR, cb, il); + model.layers[il].ffn_up, NULL, + model.layers[il].ffn_gate, NULL, + model.layers[il].ffn_down, NULL, + LLM_FFN_SILU, LLM_FFN_PAR, cb, il); cb(cur, "ffn_out", il); } @@ -4224,8 +4569,8 @@ struct llm_build_context { cur = inpL; cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, NULL, - LLM_NORM_RMS, cb, -1); + model.output_norm, NULL, + LLM_NORM_RMS, cb, -1); cb(cur, "result_norm", -1); // lm_head @@ -4238,7 +4583,7 @@ struct llm_build_context { } struct ggml_cgraph * build_bloom() { - struct ggml_cgraph * gf = ggml_new_graph(ctx0); + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); struct ggml_tensor * cur; struct ggml_tensor * inpL; @@ -4255,16 +4600,16 @@ struct llm_build_context { cb(KQ_mask, "KQ_mask", -1); inpL = llm_build_norm(ctx0, inpL, hparams, - model.tok_norm, - model.tok_norm_b, - LLM_NORM, cb, -1); + model.tok_norm, + model.tok_norm_b, + LLM_NORM, cb, -1); cb(inpL, "inp_norm", -1); for (int il = 0; il < n_layer; ++il) { cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, - model.layers[il].attn_norm_b, - LLM_NORM, cb, il); + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, cb, il); cb(cur, "attn_norm", il); // self-attention @@ -4288,8 +4633,8 @@ struct llm_build_context { llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); cur = llm_build_kqv(ctx0, hparams, kv_self, - model.layers[il].wo, model.layers[il].bo, - Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, 8.0f, cb, il); + model.layers[il].wo, model.layers[il].bo, + Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, 8.0f, cb, il); cb(cur, "kqv_out", il); } @@ -4300,16 +4645,16 @@ struct llm_build_context { // FF { cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, - model.layers[il].ffn_norm_b, - LLM_NORM, cb, il); + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, cb, il); cb(cur, "ffn_norm", il); cur = llm_build_ffn(ctx0, cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, - NULL, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, - LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); + model.layers[il].ffn_up, model.layers[il].ffn_up_b, + NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, + LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); cb(cur, "ffn_out", il); } @@ -4318,9 +4663,9 @@ struct llm_build_context { } cur = llm_build_norm(ctx0, inpL, hparams, - model.output_norm, - model.output_norm_b, - LLM_NORM, cb, -1); + model.output_norm, + model.output_norm_b, + LLM_NORM, cb, -1); cb(cur, "result_norm", -1); cur = ggml_mul_mat(ctx0, model.output, cur); @@ -4332,7 +4677,7 @@ struct llm_build_context { } struct ggml_cgraph * build_mpt() { - struct ggml_cgraph * gf = ggml_new_graph(ctx0); + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); struct ggml_tensor * cur; struct ggml_tensor * inpL; @@ -4352,9 +4697,9 @@ struct llm_build_context { struct ggml_tensor * attn_norm; attn_norm = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, - NULL, - LLM_NORM, cb, il); + model.layers[il].attn_norm, + NULL, + LLM_NORM, cb, il); cb(attn_norm, "attn_norm", il); // self-attention @@ -4382,8 +4727,8 @@ struct llm_build_context { llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); cur = llm_build_kqv(ctx0, hparams, kv_self, - model.layers[il].wo, NULL, - Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, hparams.f_max_alibi_bias, cb, il); + model.layers[il].wo, NULL, + Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, hparams.f_max_alibi_bias, cb, il); cb(cur, "kqv_out", il); } @@ -4394,16 +4739,16 @@ struct llm_build_context { // feed forward { cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, - NULL, - LLM_NORM, cb, il); + model.layers[il].ffn_norm, + NULL, + LLM_NORM, cb, il); cb(cur, "ffn_norm", il); cur = llm_build_ffn(ctx0, cur, - model.layers[il].ffn_up, NULL, - NULL, NULL, - model.layers[il].ffn_down, NULL, - LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); + model.layers[il].ffn_up, NULL, + NULL, NULL, + model.layers[il].ffn_down, NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); cb(cur, "ffn_out", il); } @@ -4417,9 +4762,9 @@ struct llm_build_context { cur = inpL; cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, - NULL, - LLM_NORM, cb, -1); + model.output_norm, + NULL, + LLM_NORM, cb, -1); cb(cur, "result_norm", -1); cur = ggml_mul_mat(ctx0, model.output, cur); @@ -4429,6 +4774,119 @@ struct llm_build_context { return gf; } + + struct ggml_cgraph * build_stablelm() { + struct ggml_cgraph * gf = ggml_new_graph(ctx0); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + cb(inpL, "inp_embd", -1); + + // inp_pos - contains the positions + struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + cb(inp_pos, "inp_pos", -1); + + // KQ_scale + struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); + cb(KQ_scale, "KQ_scale", -1); + + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + cb(KQ_mask, "KQ_mask", -1); + + // shift the entire K-cache if needed + if (do_rope_shift) { + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, hparams.n_rot, freq_base, freq_scale, cb); + } + + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * inpSA = inpL; + + // norm + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, cb, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, + hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, + hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); + + cur = llm_build_kqv(ctx0, hparams, kv_self, + model.layers[il].wo, NULL, + Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il); + cb(cur, "kqv_out", il); + } + + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, cb, il); + cb(cur, "ffn_norm", il); + + cur = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up, NULL, + model.layers[il].ffn_gate, NULL, + model.layers[il].ffn_down, NULL, + LLM_FFN_SILU, LLM_FFN_PAR, cb, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = llm_build_norm(ctx0, cur, hparams, + model.output_norm, + model.output_norm_b, + LLM_NORM, cb, -1); + cb(cur, "result_norm", -1); + + // lm_head + cur = ggml_mul_mat(ctx0, model.output, cur); + cb(cur, "result_output", -1); + + ggml_build_forward_expand(gf, cur); + + return gf; + } }; // @@ -4522,87 +4980,87 @@ struct llm_offload_trie { // TODO: will be removed with backend v2 static const std::unordered_map k_offload_map = { - //{ "inp_tokens", OFFLOAD_FUNC_NR }, // TODO: missing K-quants get_rows kernel - //{ "inp_embd", OFFLOAD_FUNC_NR }, // TODO: missing K-quants get_rows kernel - { "pos_embd", OFFLOAD_FUNC_NR }, + //{ "inp_tokens", OFFLOAD_FUNC_NR }, // TODO: missing K-quants get_rows kernel + //{ "inp_embd", OFFLOAD_FUNC_NR }, // TODO: missing K-quants get_rows kernel + { "pos_embd", OFFLOAD_FUNC_NR }, - { "inp_pos", OFFLOAD_FUNC_KQ }, // this is often used for KQ ops (e.g. rope) - { "KQ_scale", OFFLOAD_FUNC_KQ }, - { "KQ_mask", OFFLOAD_FUNC_KQ }, - { "K_shift", OFFLOAD_FUNC_KQ }, - { "K_shifted", OFFLOAD_FUNC_KQ }, + { "inp_pos", OFFLOAD_FUNC_KQ }, // this is often used for KQ ops (e.g. rope) + { "KQ_scale", OFFLOAD_FUNC_KQ }, + { "KQ_mask", OFFLOAD_FUNC_KQ }, + { "K_shift", OFFLOAD_FUNC_KQ }, + { "K_shifted", OFFLOAD_FUNC_KQ }, - { "inp_norm", OFFLOAD_FUNC_NR }, - { "inp_norm_w", OFFLOAD_FUNC_NR }, - { "inp_norm_wb", OFFLOAD_FUNC_NR }, + { "inp_norm", OFFLOAD_FUNC_NR }, + { "inp_norm_w", OFFLOAD_FUNC_NR }, + { "inp_norm_wb", OFFLOAD_FUNC_NR }, - { "norm", OFFLOAD_FUNC }, - { "norm_w", OFFLOAD_FUNC }, - { "norm_wb", OFFLOAD_FUNC }, + { "norm", OFFLOAD_FUNC }, + { "norm_w", OFFLOAD_FUNC }, + { "norm_wb", OFFLOAD_FUNC }, - { "attn_norm", OFFLOAD_FUNC }, - { "attn_norm_2", OFFLOAD_FUNC }, + { "attn_norm", OFFLOAD_FUNC }, + { "attn_norm_2", OFFLOAD_FUNC }, - { "wqkv", OFFLOAD_FUNC_KQ }, - { "bqkv", OFFLOAD_FUNC_KQ }, - { "wqkv_clamped", OFFLOAD_FUNC_KQ }, + { "wqkv", OFFLOAD_FUNC_KQ }, + { "bqkv", OFFLOAD_FUNC_KQ }, + { "wqkv_clamped", OFFLOAD_FUNC_KQ }, - { "tmpk", OFFLOAD_FUNC_KQ }, - { "tmpq", OFFLOAD_FUNC_KQ }, - { "tmpv", OFFLOAD_FUNC_V }, - { "Kcur", OFFLOAD_FUNC_KQ }, - { "Qcur", OFFLOAD_FUNC_KQ }, - { "Vcur", OFFLOAD_FUNC_V }, + { "tmpk", OFFLOAD_FUNC_KQ }, + { "tmpq", OFFLOAD_FUNC_KQ }, + { "tmpv", OFFLOAD_FUNC_V }, + { "Kcur", OFFLOAD_FUNC_KQ }, + { "Qcur", OFFLOAD_FUNC_KQ }, + { "Vcur", OFFLOAD_FUNC_V }, - { "krot", OFFLOAD_FUNC_KQ }, - { "qrot", OFFLOAD_FUNC_KQ }, - { "kpass", OFFLOAD_FUNC_KQ }, - { "qpass", OFFLOAD_FUNC_KQ }, - { "krotated", OFFLOAD_FUNC_KQ }, - { "qrotated", OFFLOAD_FUNC_KQ }, + { "krot", OFFLOAD_FUNC_KQ }, + { "qrot", OFFLOAD_FUNC_KQ }, + { "kpass", OFFLOAD_FUNC_KQ }, + { "qpass", OFFLOAD_FUNC_KQ }, + { "krotated", OFFLOAD_FUNC_KQ }, + { "qrotated", OFFLOAD_FUNC_KQ }, - { "q", OFFLOAD_FUNC_KQ }, - { "k", OFFLOAD_FUNC_KQ }, - { "kq", OFFLOAD_FUNC_KQ }, - { "kq_scaled", OFFLOAD_FUNC_KQ }, - { "kq_scaled_alibi", OFFLOAD_FUNC_KQ }, - { "kq_masked", OFFLOAD_FUNC_KQ }, - { "kq_soft_max", OFFLOAD_FUNC_V }, - { "v", OFFLOAD_FUNC_V }, - { "kqv", OFFLOAD_FUNC_V }, - { "kqv_merged", OFFLOAD_FUNC_V }, - { "kqv_merged_cont", OFFLOAD_FUNC_V }, - { "kqv_wo", OFFLOAD_FUNC_V }, - { "kqv_out", OFFLOAD_FUNC_V }, + { "q", OFFLOAD_FUNC_KQ }, + { "k", OFFLOAD_FUNC_KQ }, + { "kq", OFFLOAD_FUNC_KQ }, + { "kq_scaled", OFFLOAD_FUNC_KQ }, + { "kq_scaled_alibi", OFFLOAD_FUNC_KQ }, + { "kq_masked", OFFLOAD_FUNC_KQ }, + { "kq_soft_max", OFFLOAD_FUNC_V }, + { "v", OFFLOAD_FUNC_V }, + { "kqv", OFFLOAD_FUNC_V }, + { "kqv_merged", OFFLOAD_FUNC_V }, + { "kqv_merged_cont", OFFLOAD_FUNC_V }, + { "kqv_wo", OFFLOAD_FUNC_V }, + { "kqv_out", OFFLOAD_FUNC_V }, - { "ffn_inp", OFFLOAD_FUNC }, - { "ffn_norm", OFFLOAD_FUNC }, + { "ffn_inp", OFFLOAD_FUNC }, + { "ffn_norm", OFFLOAD_FUNC }, - { "ffn_up", OFFLOAD_FUNC }, - { "ffn_up_b", OFFLOAD_FUNC }, - { "ffn_gate", OFFLOAD_FUNC }, - { "ffn_gate_b", OFFLOAD_FUNC }, - { "ffn_gate_par", OFFLOAD_FUNC }, - { "ffn_down", OFFLOAD_FUNC }, - { "ffn_down_b", OFFLOAD_FUNC }, - { "ffn_out", OFFLOAD_FUNC }, + { "ffn_up", OFFLOAD_FUNC }, + { "ffn_up_b", OFFLOAD_FUNC }, + { "ffn_gate", OFFLOAD_FUNC }, + { "ffn_gate_b", OFFLOAD_FUNC }, + { "ffn_gate_par", OFFLOAD_FUNC }, + { "ffn_down", OFFLOAD_FUNC }, + { "ffn_down_b", OFFLOAD_FUNC }, + { "ffn_out", OFFLOAD_FUNC }, - { "ffn_silu", OFFLOAD_FUNC }, - { "ffn_gelu", OFFLOAD_FUNC }, - { "ffn_relu", OFFLOAD_FUNC }, - { "ffn_sqr(relu)", OFFLOAD_FUNC }, + { "ffn_silu", OFFLOAD_FUNC }, + { "ffn_gelu", OFFLOAD_FUNC }, + { "ffn_relu", OFFLOAD_FUNC }, + { "ffn_sqr(relu)", OFFLOAD_FUNC }, - { "l_out", OFFLOAD_FUNC }, + { "l_out", OFFLOAD_FUNC }, - { "result_norm", OFFLOAD_FUNC_EMB }, - { "result_output", OFFLOAD_FUNC_OUT }, + { "result_norm", OFFLOAD_FUNC_EMB }, + { "result_output", OFFLOAD_FUNC_OUT }, }; static llm_offload_trie k_offload_func_trie(k_offload_map); static struct ggml_cgraph * llama_build_graph( - llama_context & lctx, - const llama_batch & batch) { + llama_context & lctx, + const llama_batch & batch) { const auto & model = lctx.model; // check if we should build the worst-case graph (for memory measurement) @@ -4762,20 +5220,20 @@ static struct ggml_cgraph * llama_build_graph( const bool offload_emb = lctx.embedding.empty(); static const std::unordered_map> k_offload_func_name = { - { OFFLOAD_FUNC_NOP, "CPU" }, - { OFFLOAD_FUNC_OUT, "CPU" }, + { OFFLOAD_FUNC_NOP, "CPU" }, + { OFFLOAD_FUNC_OUT, "CPU" }, #ifdef GGML_USE_CUBLAS - { OFFLOAD_FUNC, "GPU (CUDA)" }, + { OFFLOAD_FUNC, "GPU (CUDA)" }, { OFFLOAD_FUNC_KQ, "GPU (CUDA) KQ" }, { OFFLOAD_FUNC_V, "GPU (CUDA) V" }, { OFFLOAD_FUNC_NR, "GPU (CUDA) NR" }, { OFFLOAD_FUNC_EMB, "GPU (CUDA) EMB" }, #else - { OFFLOAD_FUNC, "CPU" }, - { OFFLOAD_FUNC_KQ, "CPU" }, - { OFFLOAD_FUNC_V, "CPU" }, - { OFFLOAD_FUNC_NR, "CPU" }, - { OFFLOAD_FUNC_EMB, "CPU" }, + { OFFLOAD_FUNC, "CPU" }, + { OFFLOAD_FUNC_KQ, "CPU" }, + { OFFLOAD_FUNC_V, "CPU" }, + { OFFLOAD_FUNC_NR, "CPU" }, + { OFFLOAD_FUNC_EMB, "CPU" }, #endif // GGML_USE_CUBLAS }; @@ -4867,37 +5325,41 @@ static struct ggml_cgraph * llama_build_graph( switch (model.arch) { case LLM_ARCH_LLAMA: - { - result = llm.build_llama(); - } break; + { + result = llm.build_llama(); + } break; case LLM_ARCH_BAICHUAN: - { - result = llm.build_baichuan(); - } break; + { + result = llm.build_baichuan(); + } break; case LLM_ARCH_FALCON: - { - result = llm.build_falcon(); - } break; + { + result = llm.build_falcon(); + } break; case LLM_ARCH_STARCODER: - { - result = llm.build_starcoder(); - } break; + { + result = llm.build_starcoder(); + } break; case LLM_ARCH_PERSIMMON: - { - result = llm.build_persimmon(); - } break; + { + result = llm.build_persimmon(); + } break; case LLM_ARCH_REFACT: - { - result = llm.build_refact(); - } break; + { + result = llm.build_refact(); + } break; case LLM_ARCH_BLOOM: - { - result = llm.build_bloom(); - } break; + { + result = llm.build_bloom(); + } break; case LLM_ARCH_MPT: - { - result = llm.build_mpt(); - } break; + { + result = llm.build_mpt(); + } break; + case LLM_ARCH_STABLELM: + { + result = llm.build_stablelm(); + } break; default: GGML_ASSERT(false); } @@ -4938,8 +5400,8 @@ static struct ggml_cgraph * llama_build_graph( // return negative int on error // static int llama_decode_internal( - llama_context & lctx, - llama_batch batch) { + llama_context & lctx, + llama_batch batch) { const uint32_t n_tokens = batch.n_tokens; if (n_tokens == 0) { @@ -5068,11 +5530,13 @@ static int llama_decode_internal( // If all tensors can be run on the GPU then using more than 1 thread is detrimental. const bool full_offload_supported = - model.arch == LLM_ARCH_LLAMA || - model.arch == LLM_ARCH_BAICHUAN || - model.arch == LLM_ARCH_FALCON || - model.arch == LLM_ARCH_REFACT || - model.arch == LLM_ARCH_MPT; + model.arch == LLM_ARCH_LLAMA || + model.arch == LLM_ARCH_BAICHUAN || + model.arch == LLM_ARCH_FALCON || + model.arch == LLM_ARCH_REFACT || + model.arch == LLM_ARCH_MPT || + model.arch == LLM_ARCH_STARCODER || + model.arch == LLM_ARCH_STABLELM; const bool fully_offloaded = model.n_gpu_layers >= (int) hparams.n_layer + 3; if (ggml_cpu_has_cublas() && full_offload_supported && fully_offloaded) { @@ -5210,31 +5674,31 @@ static uint8_t llama_token_to_byte(const llama_vocab& vocab, llama_token id) { GGML_ASSERT(llama_is_byte_token(vocab, id)); const auto& token_data = vocab.id_to_token.at(id); switch (llama_vocab_get_type(vocab)) { - case LLAMA_VOCAB_TYPE_SPM: { - auto buf = token_data.text.substr(3, 2); - return strtol(buf.c_str(), NULL, 16); - } - case LLAMA_VOCAB_TYPE_BPE: { - GGML_ASSERT(false); - return unicode_to_bytes_bpe(token_data.text); - } - default: - GGML_ASSERT(false); + case LLAMA_VOCAB_TYPE_SPM: { + auto buf = token_data.text.substr(3, 2); + return strtol(buf.c_str(), NULL, 16); + } + case LLAMA_VOCAB_TYPE_BPE: { + GGML_ASSERT(false); + return unicode_to_bytes_bpe(token_data.text); + } + default: + GGML_ASSERT(false); } } static llama_token llama_byte_to_token(const llama_vocab & vocab, uint8_t ch) { static const char * hex = "0123456789ABCDEF"; switch (llama_vocab_get_type(vocab)) { - case LLAMA_VOCAB_TYPE_SPM: { - const char buf[7] = { '<', '0', 'x', hex[ch >> 4], hex[ch & 15], '>', 0 }; - return vocab.token_to_id.at(buf); - } - case LLAMA_VOCAB_TYPE_BPE: { - return vocab.token_to_id.at(bytes_to_unicode_bpe(ch)); - } - default: - GGML_ASSERT(false); + case LLAMA_VOCAB_TYPE_SPM: { + const char buf[7] = { '<', '0', 'x', hex[ch >> 4], hex[ch & 15], '>', 0 }; + return vocab.token_to_id.at(buf); + } + case LLAMA_VOCAB_TYPE_BPE: { + return vocab.token_to_id.at(bytes_to_unicode_bpe(ch)); + } + default: + GGML_ASSERT(false); } } @@ -5602,10 +6066,10 @@ private: if (!split_condition && bytes_remain >= 3) { // 're|'ve|'ll if (utf_char == "\'" && ( - (utf_char_next == "r" && utf_char_next_next == "e") || - (utf_char_next == "v" && utf_char_next_next == "e") || - (utf_char_next == "l" && utf_char_next_next == "l")) - ) { + (utf_char_next == "r" && utf_char_next_next == "e") || + (utf_char_next == "v" && utf_char_next_next == "e") || + (utf_char_next == "l" && utf_char_next_next == "l")) + ) { split_condition = true; } if (split_condition) { @@ -5631,9 +6095,9 @@ private: collecting = true; } else if ( - ((codepoint_type(utf_char) != CODEPOINT_TYPE_LETTER && codepoint_type(utf_char) != CODEPOINT_TYPE_DIGIT) && (codepoint_type(utf_char) != CODEPOINT_TYPE_WHITESPACE)) || - (!token.size() && utf_char == " " && codepoint_type(utf_char_next) != CODEPOINT_TYPE_LETTER && codepoint_type(utf_char_next) != CODEPOINT_TYPE_DIGIT && codepoint_type(utf_char_next) != CODEPOINT_TYPE_WHITESPACE) - ) { + ((codepoint_type(utf_char) != CODEPOINT_TYPE_LETTER && codepoint_type(utf_char) != CODEPOINT_TYPE_DIGIT) && (codepoint_type(utf_char) != CODEPOINT_TYPE_WHITESPACE)) || + (!token.size() && utf_char == " " && codepoint_type(utf_char_next) != CODEPOINT_TYPE_LETTER && codepoint_type(utf_char_next) != CODEPOINT_TYPE_DIGIT && codepoint_type(utf_char_next) != CODEPOINT_TYPE_WHITESPACE) + ) { collecting_special = true; collecting = true; } @@ -5707,23 +6171,23 @@ typedef enum FRAGMENT_BUFFER_VARIANT_TYPE{ struct fragment_buffer_variant{ fragment_buffer_variant(llama_vocab::id _token) - : - type(FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN), - token(_token), - raw_text(_dummy), - offset(0), - length(0){} + : + type(FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN), + token(_token), + raw_text(_dummy), + offset(0), + length(0){} fragment_buffer_variant(const std::string & _raw_text, int64_t _offset, int64_t _length) - : - type(FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT), - token((llama_vocab::id)-1), - raw_text(_raw_text), - offset(_offset), - length(_length){ - GGML_ASSERT( _offset >= 0 ); - GGML_ASSERT( _length >= 1 ); - GGML_ASSERT( offset + length <= raw_text.length() ); - } + : + type(FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT), + token((llama_vocab::id)-1), + raw_text(_raw_text), + offset(_offset), + length(_length){ + GGML_ASSERT( _offset >= 0 ); + GGML_ASSERT( _length >= 1 ); + GGML_ASSERT( offset + length <= raw_text.length() ); + } const FRAGMENT_BUFFER_VARIANT_TYPE type; const llama_vocab::id token; @@ -5853,52 +6317,55 @@ static std::vector llama_tokenize_internal(const llama_vocab & switch (vocab.type) { case LLAMA_VOCAB_TYPE_SPM: + { + for (const auto & fragment: fragment_buffer) { - for (const auto & fragment: fragment_buffer) + if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) { - if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) - { - // without adding this leading whitespace, we do not get the same results as the original tokenizer + // without adding this leading whitespace, we do not get the same results as the original tokenizer - // TODO: It's likely possible to get rid of this string copy entirely - // by modifying llm_tokenizer_x to operate with string offsets like pre-tokenizer - // and passing 'add space prefix' as bool argument - // - auto raw_text = (special ? "" : " ") + fragment.raw_text.substr(fragment.offset, fragment.length); + // TODO: It's likely possible to get rid of this string copy entirely + // by modifying llm_tokenizer_x to operate with string offsets like pre-tokenizer + // and passing 'add space prefix' as bool argument + // + auto raw_text = fragment.raw_text.substr(fragment.offset, fragment.length); + if (&fragment == &fragment_buffer.front()) { + raw_text = " " + raw_text; // prefix with space if the first token is not special + } #ifdef PRETOKENIZERDEBUG - fprintf(stderr,"TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str()); + fprintf(stderr,"TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str()); #endif - llm_tokenizer_spm tokenizer(vocab); - llama_escape_whitespace(raw_text); - tokenizer.tokenize(raw_text, output); - } - else // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN) - { - output.push_back(fragment.token); - } + llm_tokenizer_spm tokenizer(vocab); + llama_escape_whitespace(raw_text); + tokenizer.tokenize(raw_text, output); } - } break; + else // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN) + { + output.push_back(fragment.token); + } + } + } break; case LLAMA_VOCAB_TYPE_BPE: + { + for (const auto & fragment: fragment_buffer) { - for (const auto & fragment: fragment_buffer) + if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) { - if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) - { - auto raw_text = fragment.raw_text.substr(fragment.offset, fragment.length); + auto raw_text = fragment.raw_text.substr(fragment.offset, fragment.length); #ifdef PRETOKENIZERDEBUG - fprintf(stderr,"TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str()); + fprintf(stderr,"TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str()); #endif - llm_tokenizer_bpe tokenizer(vocab); - tokenizer.tokenize(raw_text, output); - } - else // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN) - { - output.push_back(fragment.token); - } + llm_tokenizer_bpe tokenizer(vocab); + tokenizer.tokenize(raw_text, output); } - } break; + else // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN) + { + output.push_back(fragment.token); + } + } + } break; } return output; @@ -5959,7 +6426,7 @@ static std::pair, llama_partial_utf8> decode_utf8( while (*pos != 0) { uint8_t first_byte = static_cast(*pos); uint8_t highbits = first_byte >> 4; - n_remain = lookup[highbits] - 1; + n_remain = lookup[highbits] - 1; if (n_remain < 0) { // invalid sequence, abort @@ -5969,7 +6436,7 @@ static std::pair, llama_partial_utf8> decode_utf8( } uint8_t mask = (1 << (7 - n_remain)) - 1; - value = first_byte & mask; + value = first_byte & mask; ++pos; while (*pos != 0 && n_remain > 0) { value = (value << 6) + (static_cast(*pos) & 0x3F); @@ -6186,7 +6653,7 @@ static std::vector llama_grammar_reject_candidates_for_ // reached end of full codepoints in token, reject iff it ended in a partial sequence // that cannot satisfy this position in grammar if (tok.partial_utf8.n_remain != 0 && - !llama_grammar_match_partial_char(stack_pos, tok.partial_utf8)) { + !llama_grammar_match_partial_char(stack_pos, tok.partial_utf8)) { rejects.push_back(tok); } } else if (llama_grammar_match_char(stack_pos, *tok.code_points).first) { @@ -6237,9 +6704,9 @@ static std::vector llama_grammar_reject_candidates( // struct llama_grammar * llama_grammar_init( - const llama_grammar_element ** rules, - size_t n_rules, - size_t start_rule_index) { + const llama_grammar_element ** rules, + size_t n_rules, + size_t start_rule_index) { const llama_grammar_element * pos; // copy rule definitions into vectors @@ -6289,7 +6756,7 @@ struct llama_grammar * llama_grammar_copy(const struct llama_grammar * grammar) for (size_t ir0 = 0; ir0 < grammar->rules.size(); ir0++) { for (size_t ir1 = 0; ir1 < grammar->rules[ir0].size(); ir1++) { if (grammar->stacks[is][ie] == &grammar->rules[ir0][ir1]) { - result->stacks[is][ie] = &result->rules[ir0][ir1]; + result->stacks[is][ie] = &result->rules[ir0][ir1]; } } } @@ -6561,13 +7028,13 @@ void llama_sample_temperature(struct llama_context * ctx, llama_token_data_array } void llama_sample_repetition_penalties( - struct llama_context * ctx, - llama_token_data_array * candidates, - const llama_token * last_tokens, - size_t penalty_last_n, - float penalty_repeat, - float penalty_freq, - float penalty_present) { + struct llama_context * ctx, + llama_token_data_array * candidates, + const llama_token * last_tokens, + size_t penalty_last_n, + float penalty_repeat, + float penalty_freq, + float penalty_present) { if (penalty_last_n == 0 || (penalty_repeat == 1.0f && penalty_freq == 0.0f && penalty_present == 0.0f)) { return; } @@ -6662,10 +7129,10 @@ static void llama_log_softmax(float * array, size_t size) { } void llama_sample_classifier_free_guidance( - struct llama_context * ctx, + struct llama_context * ctx, llama_token_data_array * candidates, - struct llama_context * guidance_ctx, - float scale) { + struct llama_context * guidance_ctx, + float scale) { int64_t t_start_sample_us = ggml_time_us(); GGML_ASSERT(ctx); @@ -6885,11 +7352,11 @@ struct llama_logit_info { float operator()(float sum, float l) const { return sum + std::exp(l - max_l); } }; llama_logit_info(llama_context * ctx) - : logits(llama_get_logits(ctx)) - , n_vocab(llama_n_vocab(llama_get_model(ctx))) - , max_l(*std::max_element(logits, logits + n_vocab)) - , normalizer(1.0f / std::accumulate(logits, logits + n_vocab, 0.0f, sum_exp{max_l})) - { } + : logits(llama_get_logits(ctx)) + , n_vocab(llama_n_vocab(llama_get_model(ctx))) + , max_l(*std::max_element(logits, logits + n_vocab)) + , normalizer(1.0f / std::accumulate(logits, logits + n_vocab, 0.0f, sum_exp{max_l})) + { } llama_token_data get_token_data(const llama_token token_id) const { constexpr auto p = std::numeric_limits::quiet_NaN(); // never used return {token_id, logits[token_id], p}; @@ -6934,11 +7401,11 @@ struct llama_beam_search_data { std::vector beam_views; llama_beam_search_data(llama_context * ctx, size_t n_beams, int n_past, int n_predict) - : ctx(ctx) - , n_beams(n_beams) - , n_past(n_past) - , n_predict(n_predict) - , beam_views(n_beams) { + : ctx(ctx) + , n_beams(n_beams) + , n_past(n_past) + , n_predict(n_predict) + , beam_views(n_beams) { beams.reserve(n_beams); next_beams.reserve(n_beams); } @@ -7044,7 +7511,7 @@ struct llama_beam_search_data { beams.push_back({{}, 1.0f, false}); // Start with one empty beam w/ probability = 1.0 and !eob. const auto not_eob = [](const llama_beam & beam) { return !beam.eob; }; for (int i = 0 ; i < n_predict && std::any_of(beams.begin(),beams.end(),not_eob) && - !beams[top_beam_index()].eob ; ++i) { + !beams[top_beam_index()].eob ; ++i) { callback(callback_data, get_beams_state(false)); // Sets common_prefix_length update_beams_from_beam_views(); // Update values (p,eob) that callback may have changed. if (common_prefix_length) { @@ -7124,14 +7591,14 @@ struct quantize_state_internal { int n_fallback = 0; quantize_state_internal(const llama_model & model, const llama_model_quantize_params * params) - : model(model) - , params(params) - {} + : model(model) + , params(params) + {} }; static void llama_convert_tensor_internal( - struct ggml_tensor * tensor, std::vector> & output, std::vector & workers, - const size_t nelements, const int nthread + struct ggml_tensor * tensor, std::vector> & output, std::vector & workers, + const size_t nelements, const int nthread ) { if (output.size() < nelements) { output.resize(nelements); @@ -7188,8 +7655,8 @@ static void llama_convert_tensor_internal( } static ggml_type get_k_quant_type( - quantize_state_internal & qs, - ggml_type new_type, const ggml_tensor * tensor, llama_ftype ftype + quantize_state_internal & qs, + ggml_type new_type, const ggml_tensor * tensor, llama_ftype ftype ) { const std::string name = ggml_get_name(tensor); // TODO: avoid hardcoded tensor names - use the TN_* constants @@ -7215,10 +7682,10 @@ static ggml_type get_k_quant_type( } else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K; else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) && - use_more_bits(qs.i_attention_wv, qs.n_attention_wv)) new_type = GGML_TYPE_Q6_K; + use_more_bits(qs.i_attention_wv, qs.n_attention_wv)) new_type = GGML_TYPE_Q6_K; else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && qs.i_attention_wv < 4) new_type = GGML_TYPE_Q5_K; else if (QK_K == 64 && (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S) && - (qs.i_attention_wv < qs.n_attention_wv/8 || qs.i_attention_wv >= 7*qs.n_attention_wv/8)) new_type = GGML_TYPE_Q6_K; + (qs.i_attention_wv < qs.n_attention_wv/8 || qs.i_attention_wv >= 7*qs.n_attention_wv/8)) new_type = GGML_TYPE_Q6_K; if (qs.model.type == MODEL_70B) { // In the 70B model we have 8 heads sharing the same attn_v weights. As a result, the attn_v.weight tensor is // 8x smaller compared to attn_q.weight. Hence, we can get a nice boost in quantization accuracy with @@ -7230,8 +7697,8 @@ static ggml_type get_k_quant_type( if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K; else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) { new_type = qs.i_feed_forward_w2 < 2 ? GGML_TYPE_Q5_K - : arch != LLM_ARCH_FALCON || use_more_bits(qs.i_feed_forward_w2, qs.n_feed_forward_w2) ? GGML_TYPE_Q4_K - : GGML_TYPE_Q3_K; + : arch != LLM_ARCH_FALCON || use_more_bits(qs.i_feed_forward_w2, qs.n_feed_forward_w2) ? GGML_TYPE_Q4_K + : GGML_TYPE_Q3_K; } else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) { new_type = arch == LLM_ARCH_FALCON ? GGML_TYPE_Q4_K : GGML_TYPE_Q5_K; @@ -7312,7 +7779,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s case LLAMA_FTYPE_MOSTLY_F16: quantized_type = GGML_TYPE_F16; break; case LLAMA_FTYPE_ALL_F32: quantized_type = GGML_TYPE_F32; break; - // K-quants + // K-quants case LLAMA_FTYPE_MOSTLY_Q2_K: quantized_type = GGML_TYPE_Q2_K; break; case LLAMA_FTYPE_MOSTLY_Q3_K_S: case LLAMA_FTYPE_MOSTLY_Q3_K_M: @@ -7378,7 +7845,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s } if (qs.n_attention_wv != qs.n_feed_forward_w2 || (uint32_t)qs.n_attention_wv != model.hparams.n_layer) { LLAMA_LOG_WARN("%s ============ Strange model: n_attention_wv = %d, n_feed_forward_w2 = %d, hparams.n_layer = %d\n", - __func__, qs.n_attention_wv, qs.n_feed_forward_w2, model.hparams.n_layer); + __func__, qs.n_attention_wv, qs.n_feed_forward_w2, model.hparams.n_layer); } size_t total_size_org = 0; @@ -7425,10 +7892,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s ml.load_data_for(tensor); LLAMA_LOG_INFO("[%4d/%4d] %36s - [%s], type = %6s, ", - ++idx, ml.n_tensors, - ggml_get_name(tensor), - llama_format_tensor_shape(tensor).c_str(), - ggml_type_name(tensor->type)); + ++idx, ml.n_tensors, + ggml_get_name(tensor), + llama_format_tensor_shape(tensor).c_str(), + ggml_type_name(tensor->type)); // This used to be a regex, but has an extreme cost to compile times. bool quantize = name.rfind("weight") == name.size() - 6; // ends with 'weight'? @@ -7516,7 +7983,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s workers.clear(); } - LLAMA_LOG_INFO("size = %8.2f MB -> %8.2f MB | hist: ", ggml_nbytes(tensor)/1024.0/1024.0, new_size/1024.0/1024.0); + LLAMA_LOG_INFO("size = %8.2f MiB -> %8.2f MiB | hist: ", ggml_nbytes(tensor)/1024.0/1024.0, new_size/1024.0/1024.0); int64_t tot_count = 0; for (size_t i = 0; i < hist_cur.size(); i++) { hist_all[i] += hist_cur[i]; @@ -7575,12 +8042,12 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s if (qs.n_fallback > 0) { LLAMA_LOG_WARN("%s: WARNING: %d of %d tensor(s) incompatible with k-quants and required fallback quantization\n", - __func__, qs.n_fallback, qs.n_k_quantized + qs.n_fallback); + __func__, qs.n_fallback, qs.n_k_quantized + qs.n_fallback); } } static int llama_apply_lora_from_file_internal( - const struct llama_model & model, const char * path_lora, float scale, const char * path_base_model, int n_threads + const struct llama_model & model, const char * path_lora, float scale, const char * path_base_model, int n_threads ) { LLAMA_LOG_INFO("%s: applying lora adapter from '%s' - please wait ...\n", __func__, path_lora); @@ -7710,11 +8177,11 @@ static int llama_apply_lora_from_file_internal( case 0: wtype = GGML_TYPE_F32; break; case 1: wtype = GGML_TYPE_F16; break; default: - { - LLAMA_LOG_ERROR("%s: invalid tensor data type '%d'\n", + { + LLAMA_LOG_ERROR("%s: invalid tensor data type '%d'\n", __func__, ftype); - return false; - } + return false; + } } ggml_tensor * lora_tensor; if (n_dims == 2) { @@ -7791,7 +8258,7 @@ static int llama_apply_lora_from_file_internal( if (base_t->ne[0] != loraA->ne[1] || base_t->ne[1] != loraB->ne[1]) { LLAMA_LOG_ERROR("%s: incompatible tensor dimensions (%" PRId64 " and %" PRId64 ");" - " are you sure that this adapter is for this model?\n", __func__, base_t->ne[0], loraA->ne[1]); + " are you sure that this adapter is for this model?\n", __func__, base_t->ne[0], loraA->ne[1]); return 1; } @@ -7859,14 +8326,14 @@ static int llama_apply_lora_from_file_internal( // struct llama_model_params llama_model_default_params() { struct llama_model_params result = { - /*.n_gpu_layers =*/ 0, - /*.main_gpu =*/ 0, - /*.tensor_split =*/ nullptr, - /*.progress_callback =*/ nullptr, - /*.progress_callback_user_data =*/ nullptr, - /*.vocab_only =*/ false, - /*.use_mmap =*/ true, - /*.use_mlock =*/ false, + /*.n_gpu_layers =*/ 0, + /*.main_gpu =*/ 0, + /*.tensor_split =*/ nullptr, + /*.progress_callback =*/ nullptr, + /*.progress_callback_user_data =*/ nullptr, + /*.vocab_only =*/ false, + /*.use_mmap =*/ true, + /*.use_mlock =*/ false, }; #ifdef GGML_USE_METAL @@ -7878,17 +8345,23 @@ struct llama_model_params llama_model_default_params() { struct llama_context_params llama_context_default_params() { struct llama_context_params result = { - /*.seed =*/ LLAMA_DEFAULT_SEED, - /*.n_ctx =*/ 512, - /*.n_batch =*/ 512, - /*.n_threads =*/ GGML_DEFAULT_N_THREADS, // TODO: better default - /*.n_threads_batch =*/ GGML_DEFAULT_N_THREADS, - /*.rope_freq_base =*/ 0.0f, - /*.rope_freq_scale =*/ 0.0f, - /*.mul_mat_q =*/ true, - /*.f16_kv =*/ true, - /*.logits_all =*/ false, - /*.embedding =*/ false, + /*.seed =*/ LLAMA_DEFAULT_SEED, + /*.n_ctx =*/ 512, + /*.n_batch =*/ 512, + /*.n_threads =*/ GGML_DEFAULT_N_THREADS, // TODO: better default + /*.n_threads_batch =*/ GGML_DEFAULT_N_THREADS, + /*.rope_scaling_type =*/ LLAMA_ROPE_SCALING_UNSPECIFIED, + /*.rope_freq_base =*/ 0.0f, + /*.rope_freq_scale =*/ 0.0f, + /*.yarn_ext_factor =*/ -1.0f, + /*.yarn_attn_factor =*/ 1.0f, + /*.yarn_beta_fast =*/ 32.0f, + /*.yarn_beta_slow =*/ 1.0f, + /*.yarn_orig_ctx =*/ 0, + /*.mul_mat_q =*/ true, + /*.f16_kv =*/ true, + /*.logits_all =*/ false, + /*.embedding =*/ false, }; return result; @@ -7896,12 +8369,12 @@ struct llama_context_params llama_context_default_params() { struct llama_model_quantize_params llama_model_quantize_default_params() { struct llama_model_quantize_params result = { - /*.nthread =*/ 0, - /*.ftype =*/ LLAMA_FTYPE_MOSTLY_Q5_1, - /*.allow_requantize =*/ false, - /*.quantize_output_tensor =*/ true, - /*.only_copy =*/ false, - /*.pure =*/ false, + /*.nthread =*/ 0, + /*.ftype =*/ LLAMA_FTYPE_MOSTLY_Q5_1, + /*.allow_requantize =*/ false, + /*.quantize_output_tensor =*/ true, + /*.only_copy =*/ false, + /*.pure =*/ false, }; return result; @@ -7949,8 +8422,8 @@ int64_t llama_time_us(void) { } struct llama_model * llama_load_model_from_file( - const char * path_model, - struct llama_model_params params) { + const char * path_model, + struct llama_model_params params) { ggml_time_init(); llama_model * model = new llama_model; @@ -7971,10 +8444,7 @@ struct llama_model * llama_load_model_from_file( }; } - if (!llama_model_load(path_model, *model, params.n_gpu_layers, - params.main_gpu, params.tensor_split, - params.use_mmap, params.use_mlock, params.vocab_only, - params.progress_callback, params.progress_callback_user_data)) { + if (!llama_model_load(path_model, *model, params)) { LLAMA_LOG_ERROR("%s: failed to load model\n", __func__); delete model; return nullptr; @@ -7988,7 +8458,7 @@ void llama_free_model(struct llama_model * model) { } struct llama_context * llama_new_context_with_model( - struct llama_model * model, + struct llama_model * model, struct llama_context_params params) { if (!model) { @@ -8000,13 +8470,35 @@ struct llama_context * llama_new_context_with_model( const auto & hparams = model->hparams; auto & cparams = ctx->cparams; - cparams.n_batch = params.n_batch; - cparams.n_ctx = params.n_ctx == 0 ? hparams.n_ctx_train : params.n_ctx; - cparams.rope_freq_base = params.rope_freq_base == 0 ? hparams.rope_freq_base_train : params.rope_freq_base; - cparams.rope_freq_scale = params.rope_freq_scale == 0 ? hparams.rope_freq_scale_train : params.rope_freq_scale; - cparams.n_threads = params.n_threads; - cparams.n_threads_batch = params.n_threads_batch; - cparams.mul_mat_q = params.mul_mat_q; + cparams.n_batch = params.n_batch; + cparams.n_threads = params.n_threads; + cparams.n_threads_batch = params.n_threads_batch; + cparams.yarn_ext_factor = params.yarn_ext_factor; + cparams.yarn_attn_factor = params.yarn_attn_factor; + cparams.yarn_beta_fast = params.yarn_beta_fast; + cparams.yarn_beta_slow = params.yarn_beta_slow; + cparams.mul_mat_q = params.mul_mat_q; + + cparams.n_ctx = params.n_ctx == 0 ? hparams.n_ctx_train : params.n_ctx; + cparams.rope_freq_base = params.rope_freq_base == 0.0f ? hparams.rope_freq_base_train : params.rope_freq_base; + cparams.rope_freq_scale = params.rope_freq_scale == 0.0f ? hparams.rope_freq_scale_train : params.rope_freq_scale; + + cparams.n_yarn_orig_ctx = params.yarn_orig_ctx != 0 ? params.yarn_orig_ctx : + hparams.n_yarn_orig_ctx != 0 ? hparams.n_yarn_orig_ctx : + hparams.n_ctx_train; + + auto rope_scaling_type = params.rope_scaling_type; + if (rope_scaling_type == LLAMA_ROPE_SCALING_UNSPECIFIED) { + rope_scaling_type = hparams.rope_scaling_type_train; + } + + if (rope_scaling_type == LLAMA_ROPE_SCALING_NONE) { + cparams.rope_freq_scale = 1.0f; // never scale if scaling type is none + } + + if (cparams.yarn_ext_factor < 0.0f) { // negative indicates 'not set' + cparams.yarn_ext_factor = rope_scaling_type == LLAMA_ROPE_SCALING_YARN ? 1.0f : 0.0f; + } if (params.seed == LLAMA_DEFAULT_SEED) { params.seed = time(NULL); @@ -8031,7 +8523,7 @@ struct llama_context * llama_new_context_with_model( { const size_t memory_size = ggml_nbytes(ctx->kv_self.k) + ggml_nbytes(ctx->kv_self.v); - LLAMA_LOG_INFO("%s: kv self size = %7.2f MB\n", __func__, memory_size / 1024.0 / 1024.0); + LLAMA_LOG_INFO("%s: kv self size = %7.2f MiB\n", __func__, memory_size / 1024.0 / 1024.0); } // resized during inference @@ -8048,7 +8540,7 @@ struct llama_context * llama_new_context_with_model( { static const size_t tensor_alignment = 32; // the compute buffer is used to store the tensor and graph structs, while the allocator buffer is used for the tensor data - ctx->buf_compute.resize(ggml_tensor_overhead()*GGML_MAX_NODES + ggml_graph_overhead()); + ctx->buf_compute.resize(ggml_tensor_overhead()*LLAMA_MAX_NODES + ggml_graph_overhead()); // create measure allocator ctx->alloc = ggml_allocr_new_measure(tensor_alignment); @@ -8076,7 +8568,7 @@ struct llama_context * llama_new_context_with_model( // measure memory requirements for the graph size_t alloc_size = ggml_allocr_alloc_graph(ctx->alloc, gf) + tensor_alignment; - LLAMA_LOG_INFO("%s: compute buffer total size = %.2f MB\n", __func__, (ctx->buf_compute.size + alloc_size) / 1024.0 / 1024.0); + LLAMA_LOG_INFO("%s: compute buffer total size = %.2f MiB\n", __func__, (ctx->buf_compute.size + alloc_size) / 1024.0 / 1024.0); // recreate allocator with exact memory requirements ggml_allocr_free(ctx->alloc); @@ -8090,7 +8582,7 @@ struct llama_context * llama_new_context_with_model( #endif #ifdef GGML_USE_CUBLAS ggml_cuda_set_scratch_size(alloc_size); - LLAMA_LOG_INFO("%s: VRAM scratch buffer: %.2f MB\n", __func__, alloc_size / 1024.0 / 1024.0); + LLAMA_LOG_INFO("%s: VRAM scratch buffer: %.2f MiB\n", __func__, alloc_size / 1024.0 / 1024.0); // calculate total VRAM usage auto add_tensor = [](const ggml_tensor * t, size_t & size) { @@ -8110,10 +8602,10 @@ struct llama_context * llama_new_context_with_model( size_t ctx_vram_size = alloc_size + kv_vram_size; size_t total_vram_size = model_vram_size + ctx_vram_size; - LLAMA_LOG_INFO("%s: total VRAM used: %.2f MB (model: %.2f MB, context: %.2f MB)\n", __func__, + LLAMA_LOG_INFO("%s: total VRAM used: %.2f MiB (model: %.2f MiB, context: %.2f MiB)\n", __func__, total_vram_size / 1024.0 / 1024.0, model_vram_size / 1024.0 / 1024.0, - ctx_vram_size / 1024.0 / 1024.0); + ctx_vram_size / 1024.0 / 1024.0); #endif } @@ -8134,7 +8626,7 @@ struct llama_context * llama_new_context_with_model( const size_t max_size = ggml_get_max_tensor_size(ctx->model.ctx); - LLAMA_LOG_INFO("%s: max tensor size = %8.2f MB\n", __func__, max_size/1024.0/1024.0); + LLAMA_LOG_INFO("%s: max tensor size = %8.2f MiB\n", __func__, max_size/1024.0/1024.0); #define LLAMA_METAL_CHECK_BUF(result) \ if (!(result)) { \ @@ -8200,11 +8692,50 @@ float llama_rope_freq_scale_train(const struct llama_model * model) { return model->hparams.rope_freq_scale_train; } +int llama_model_meta_val_str(const struct llama_model * model, const char * key, char * buf, size_t buf_size) { + const auto & it = model->gguf_kv.find(key); + if (it == model->gguf_kv.end()) { + if (buf_size > 0) { + buf[0] = '\0'; + } + return -1; + } + return snprintf(buf, buf_size, "%s", it->second.c_str()); +} + +int llama_model_meta_count(const struct llama_model * model) { + return (int)model->gguf_kv.size(); +} + +int llama_model_meta_key_by_index(const struct llama_model * model, int i, char * buf, size_t buf_size) { + if (i < 0 || i >= (int)model->gguf_kv.size()) { + if (buf_size > 0) { + buf[0] = '\0'; + } + return -1; + } + auto it = model->gguf_kv.begin(); + std::advance(it, i); + return snprintf(buf, buf_size, "%s", it->first.c_str()); +} + +int llama_model_meta_val_str_by_index(const struct llama_model * model, int i, char * buf, size_t buf_size) { + if (i < 0 || i >= (int)model->gguf_kv.size()) { + if (buf_size > 0) { + buf[0] = '\0'; + } + return -1; + } + auto it = model->gguf_kv.begin(); + std::advance(it, i); + return snprintf(buf, buf_size, "%s", it->second.c_str()); +} + int llama_model_desc(const struct llama_model * model, char * buf, size_t buf_size) { return snprintf(buf, buf_size, "%s %s %s", - llama_model_arch_name(model->arch).c_str(), - llama_model_type_name(model->type), - llama_model_ftype_name(model->ftype).c_str()); + llama_model_arch_name(model->arch).c_str(), + llama_model_type_name(model->type), + llama_model_ftype_name(model->ftype).c_str()); } uint64_t llama_model_size(const struct llama_model * model) { @@ -8301,16 +8832,16 @@ size_t llama_get_state_size(const struct llama_context * ctx) { const size_t s_kv = ctx->kv_self.buf.size; const size_t s_total = ( - + s_rng_size - + s_rng - + s_logits_capacity - + s_logits_size - + s_logits - + s_embedding_size - + s_embedding - + s_kv_size - + s_kv_ntok - + s_kv + + s_rng_size + + s_rng + + s_logits_capacity + + s_logits_size + + s_logits + + s_embedding_size + + s_embedding + + s_kv_size + + s_kv_ntok + + s_kv ); return s_total; @@ -8437,8 +8968,8 @@ static void llama_copy_state_data_internal(struct llama_context * ctx, llama_dat if (kv_buf_size) { const size_t elt_size = ggml_element_size(kv_self.k); - ggml_context * cpy_ctx = ggml_init({ 4096, NULL, /* no_alloc */ true }); - ggml_cgraph gf{}; + ggml_context * cpy_ctx = ggml_init({ 6*ggml_tensor_overhead() + ggml_graph_overhead(), NULL, /* no_alloc */ true }); + ggml_cgraph * gf = ggml_new_graph(cpy_ctx); ggml_tensor * kout3d = ggml_new_tensor_3d(cpy_ctx, kv_self.k->type, n_embd, kv_head, n_layer); std::vector kout3d_data(ggml_nbytes(kout3d), 0); @@ -8449,16 +8980,16 @@ static void llama_copy_state_data_internal(struct llama_context * ctx, llama_dat vout3d->data = vout3d_data.data(); ggml_tensor * k3d = ggml_view_3d(cpy_ctx, kv_self.k, - n_embd, kv_head, n_layer, - elt_size*n_embd, elt_size*n_embd*n_ctx, 0); + n_embd, kv_head, n_layer, + elt_size*n_embd, elt_size*n_embd*n_ctx, 0); ggml_tensor * v3d = ggml_view_3d(cpy_ctx, kv_self.v, - kv_head, n_embd, n_layer, - elt_size*n_ctx, elt_size*n_ctx*n_embd, 0); + kv_head, n_embd, n_layer, + elt_size*n_ctx, elt_size*n_ctx*n_embd, 0); - ggml_build_forward_expand(&gf, ggml_cpy(cpy_ctx, k3d, kout3d)); - ggml_build_forward_expand(&gf, ggml_cpy(cpy_ctx, v3d, vout3d)); - ggml_graph_compute_helper(ctx->work_buffer, &gf, /*n_threads*/ 1); + ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, k3d, kout3d)); + ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, v3d, vout3d)); + ggml_graph_compute_helper(ctx->work_buffer, gf, /*n_threads*/ 1); ggml_free(cpy_ctx); @@ -8565,8 +9096,8 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) { const size_t elt_size = ggml_element_size(kv_self.k); - ggml_context * cpy_ctx = ggml_init({ 4096, NULL, /* no_alloc */ true }); - ggml_cgraph gf{}; + ggml_context * cpy_ctx = ggml_init({ 6*ggml_tensor_overhead() + ggml_graph_overhead(), NULL, /* no_alloc */ true }); + ggml_cgraph * gf = ggml_new_graph(cpy_ctx); ggml_tensor * kin3d = ggml_new_tensor_3d(cpy_ctx, kv_self.k->type, n_embd, kv_head, n_layer); kin3d->data = (void *) inp; @@ -8577,16 +9108,16 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) { inp += ggml_nbytes(vin3d); ggml_tensor * k3d = ggml_view_3d(cpy_ctx, kv_self.k, - n_embd, kv_head, n_layer, - elt_size*n_embd, elt_size*n_embd*n_ctx, 0); + n_embd, kv_head, n_layer, + elt_size*n_embd, elt_size*n_embd*n_ctx, 0); ggml_tensor * v3d = ggml_view_3d(cpy_ctx, kv_self.v, - kv_head, n_embd, n_layer, - elt_size*n_ctx, elt_size*n_ctx*n_embd, 0); + kv_head, n_embd, n_layer, + elt_size*n_ctx, elt_size*n_ctx*n_embd, 0); - ggml_build_forward_expand(&gf, ggml_cpy(cpy_ctx, kin3d, k3d)); - ggml_build_forward_expand(&gf, ggml_cpy(cpy_ctx, vin3d, v3d)); - ggml_graph_compute_helper(ctx->work_buffer, &gf, /*n_threads*/ 1); + ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, kin3d, k3d)); + ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, vin3d, v3d)); + ggml_graph_compute_helper(ctx->work_buffer, gf, /*n_threads*/ 1); ggml_free(cpy_ctx); } @@ -8706,9 +9237,9 @@ bool llama_save_session_file(struct llama_context * ctx, const char * path_sessi int llama_eval( struct llama_context * ctx, - llama_token * tokens, - int32_t n_tokens, - int n_past) { + llama_token * tokens, + int32_t n_tokens, + int n_past) { llama_kv_cache_seq_rm(ctx->kv_self, -1, n_past, -1); const int ret = llama_decode_internal(*ctx, llama_batch_get_one(tokens, n_tokens, n_past, 0)); @@ -8720,10 +9251,10 @@ int llama_eval( } int llama_eval_embd( - struct llama_context * ctx, - float * embd, - int32_t n_tokens, - int n_past) { + struct llama_context * ctx, + float * embd, + int32_t n_tokens, + int n_past) { llama_kv_cache_seq_rm(ctx->kv_self, -1, n_past, -1); llama_batch batch = { n_tokens, nullptr, embd, nullptr, nullptr, nullptr, nullptr, n_past, 1, 0, }; @@ -8742,21 +9273,21 @@ void llama_set_n_threads(struct llama_context * ctx, uint32_t n_threads, uint32_ } struct llama_batch llama_batch_get_one( - llama_token * tokens, - int32_t n_tokens, - llama_pos pos_0, - llama_seq_id seq_id) { + llama_token * tokens, + int32_t n_tokens, + llama_pos pos_0, + llama_seq_id seq_id) { return { - /*n_tokens =*/ n_tokens, - /*tokens =*/ tokens, - /*embd =*/ nullptr, - /*pos =*/ nullptr, - /*n_seq_id =*/ nullptr, - /*seq_id =*/ nullptr, - /*logits =*/ nullptr, - /*all_pos_0 =*/ pos_0, - /*all_pos_1 =*/ 1, - /*all_seq_id =*/ seq_id, + /*n_tokens =*/ n_tokens, + /*tokens =*/ tokens, + /*embd =*/ nullptr, + /*pos =*/ nullptr, + /*n_seq_id =*/ nullptr, + /*seq_id =*/ nullptr, + /*logits =*/ nullptr, + /*all_pos_0 =*/ pos_0, + /*all_pos_1 =*/ 1, + /*all_seq_id =*/ seq_id, }; } @@ -8796,7 +9327,7 @@ void llama_batch_free(struct llama_batch batch) { int llama_decode( struct llama_context * ctx, - struct llama_batch batch) { + struct llama_batch batch) { const int ret = llama_decode_internal(*ctx, batch); if (ret < 0) { LLAMA_LOG_ERROR("%s: failed to decode, ret = %d\n", __func__, ret); @@ -8841,6 +9372,14 @@ llama_token llama_token_nl(const struct llama_model * model) { return model->vocab.linefeed_id; } +int llama_add_bos_token(const struct llama_model * model) { + return model->vocab.special_add_bos; +} + +int llama_add_eos_token(const struct llama_model * model) { + return model->vocab.special_add_eos; +} + llama_token llama_token_prefix(const struct llama_model * model) { return model->vocab.special_prefix_id; } @@ -8858,13 +9397,13 @@ llama_token llama_token_eot(const struct llama_model * model) { } int llama_tokenize( - const struct llama_model * model, - const char * text, - int text_len, - llama_token * tokens, - int n_max_tokens, - bool add_bos, - bool special) { + const struct llama_model * model, + const char * text, + int text_len, + llama_token * tokens, + int n_max_tokens, + bool add_bos, + bool special) { auto res = llama_tokenize_internal(model->vocab, std::string(text, text_len), add_bos, special); if (n_max_tokens < (int) res.size()) { @@ -8893,56 +9432,56 @@ static std::string llama_decode_text(const std::string & text) { int llama_token_to_piece(const struct llama_model * model, llama_token token, char * buf, int length) { if (0 <= token && token < llama_n_vocab(model)) { switch (llama_vocab_get_type(model->vocab)) { - case LLAMA_VOCAB_TYPE_SPM: { - if (llama_is_normal_token(model->vocab, token)) { - std::string result = model->vocab.id_to_token[token].text; - llama_unescape_whitespace(result); - if (length < (int) result.length()) { - return -result.length(); + case LLAMA_VOCAB_TYPE_SPM: { + if (llama_is_normal_token(model->vocab, token)) { + std::string result = model->vocab.id_to_token[token].text; + llama_unescape_whitespace(result); + if (length < (int) result.length()) { + return -result.length(); + } + memcpy(buf, result.c_str(), result.length()); + return result.length(); + } else if (llama_is_unknown_token(model->vocab, token)) { // NOLINT + if (length < 3) { + return -3; + } + memcpy(buf, "\xe2\x96\x85", 3); + return 3; + } else if (llama_is_control_token(model->vocab, token)) { + ; + } else if (llama_is_byte_token(model->vocab, token)) { + if (length < 1) { + return -1; + } + buf[0] = llama_token_to_byte(model->vocab, token); + return 1; + } else { + // TODO: for now we accept all unsupported token types, + // suppressing them like CONTROL tokens. + // GGML_ASSERT(false); } - memcpy(buf, result.c_str(), result.length()); - return result.length(); - } else if (llama_is_unknown_token(model->vocab, token)) { // NOLINT - if (length < 3) { - return -3; - } - memcpy(buf, "\xe2\x96\x85", 3); - return 3; - } else if (llama_is_control_token(model->vocab, token)) { - ; - } else if (llama_is_byte_token(model->vocab, token)) { - if (length < 1) { - return -1; - } - buf[0] = llama_token_to_byte(model->vocab, token); - return 1; - } else { - // TODO: for now we accept all unsupported token types, - // suppressing them like CONTROL tokens. - // GGML_ASSERT(false); + break; } - break; - } - case LLAMA_VOCAB_TYPE_BPE: { - if (llama_is_normal_token(model->vocab, token)) { - std::string result = model->vocab.id_to_token[token].text; - result = llama_decode_text(result); - if (length < (int) result.length()) { - return -result.length(); + case LLAMA_VOCAB_TYPE_BPE: { + if (llama_is_normal_token(model->vocab, token)) { + std::string result = model->vocab.id_to_token[token].text; + result = llama_decode_text(result); + if (length < (int) result.length()) { + return -result.length(); + } + memcpy(buf, result.c_str(), result.length()); + return result.length(); + } else if (llama_is_control_token(model->vocab, token)) { + ; + } else { + // TODO: for now we accept all unsupported token types, + // suppressing them like CONTROL tokens. + // GGML_ASSERT(false); } - memcpy(buf, result.c_str(), result.length()); - return result.length(); - } else if (llama_is_control_token(model->vocab, token)) { - ; - } else { - // TODO: for now we accept all unsupported token types, - // suppressing them like CONTROL tokens. - // GGML_ASSERT(false); + break; } - break; - } - default: - GGML_ASSERT(false); + default: + GGML_ASSERT(false); } } return 0; @@ -8950,16 +9489,16 @@ int llama_token_to_piece(const struct llama_model * model, llama_token token, ch struct llama_timings llama_get_timings(struct llama_context * ctx) { struct llama_timings result = { - /*.t_start_ms =*/ 1e-3 * ctx->t_start_us, - /*.t_end_ms =*/ 1.00 * ggml_time_ms(), - /*.t_load_ms =*/ 1e-3 * ctx->t_load_us, - /*.t_sample_ms =*/ 1e-3 * ctx->t_sample_us, - /*.t_p_eval_ms =*/ 1e-3 * ctx->t_p_eval_us, - /*.t_eval_ms =*/ 1e-3 * ctx->t_eval_us, + /*.t_start_ms =*/ 1e-3 * ctx->t_start_us, + /*.t_end_ms =*/ 1.00 * ggml_time_ms(), + /*.t_load_ms =*/ 1e-3 * ctx->t_load_us, + /*.t_sample_ms =*/ 1e-3 * ctx->t_sample_us, + /*.t_p_eval_ms =*/ 1e-3 * ctx->t_p_eval_us, + /*.t_eval_ms =*/ 1e-3 * ctx->t_eval_us, - /*.n_sample =*/ std::max(1, ctx->n_sample), - /*.n_p_eval =*/ std::max(1, ctx->n_p_eval), - /*.n_eval =*/ std::max(1, ctx->n_eval), + /*.n_sample =*/ std::max(1, ctx->n_sample), + /*.n_p_eval =*/ std::max(1, ctx->n_p_eval), + /*.n_eval =*/ std::max(1, ctx->n_eval), }; return result; @@ -8971,11 +9510,11 @@ void llama_print_timings(struct llama_context * ctx) { LLAMA_LOG_INFO("\n"); LLAMA_LOG_INFO("%s: load time = %10.2f ms\n", __func__, timings.t_load_ms); LLAMA_LOG_INFO("%s: sample time = %10.2f ms / %5d runs (%8.2f ms per token, %8.2f tokens per second)\n", - __func__, timings.t_sample_ms, timings.n_sample, timings.t_sample_ms / timings.n_sample, 1e3 / timings.t_sample_ms * timings.n_sample); + __func__, timings.t_sample_ms, timings.n_sample, timings.t_sample_ms / timings.n_sample, 1e3 / timings.t_sample_ms * timings.n_sample); LLAMA_LOG_INFO("%s: prompt eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n", - __func__, timings.t_p_eval_ms, timings.n_p_eval, timings.t_p_eval_ms / timings.n_p_eval, 1e3 / timings.t_p_eval_ms * timings.n_p_eval); + __func__, timings.t_p_eval_ms, timings.n_p_eval, timings.t_p_eval_ms / timings.n_p_eval, 1e3 / timings.t_p_eval_ms * timings.n_p_eval); LLAMA_LOG_INFO("%s: eval time = %10.2f ms / %5d runs (%8.2f ms per token, %8.2f tokens per second)\n", - __func__, timings.t_eval_ms, timings.n_eval, timings.t_eval_ms / timings.n_eval, 1e3 / timings.t_eval_ms * timings.n_eval); + __func__, timings.t_eval_ms, timings.n_eval, timings.t_eval_ms / timings.n_eval, 1e3 / timings.t_eval_ms * timings.n_eval); LLAMA_LOG_INFO("%s: total time = %10.2f ms\n", __func__, (timings.t_end_ms - timings.t_start_ms)); } @@ -9039,7 +9578,7 @@ void llama_dump_timing_info_yaml(FILE * stream, const llama_context * ctx) { // For internal test use const std::vector> & llama_internal_get_tensor_map( - struct llama_context * ctx + struct llama_context * ctx ) { return ctx->model.tensors_by_name; } @@ -9081,8 +9620,6 @@ static void llama_log_callback_default(ggml_log_level level, const char * text, } - - static int save_llama_model_gguf(struct gguf_context * fctx, const char * fn_vocab_model, struct llama_model * model) { const char * arch = "llama"; const auto kv = LLM_KV(LLM_ARCH_LLAMA); diff --git a/native/jni/src/ggml/llama.h b/native/jni/src/ggml/llama.h index 60ba3d69b..0ac2bed19 100644 --- a/native/jni/src/ggml/llama.h +++ b/native/jni/src/ggml/llama.h @@ -53,690 +53,727 @@ extern "C" { #endif +// +// C interface +// +// TODO: show sample usage +// + +struct llama_model; +struct llama_context; + +typedef int32_t llama_pos; +typedef int32_t llama_token; +typedef int32_t llama_seq_id; + +enum llama_vocab_type { + LLAMA_VOCAB_TYPE_SPM = 0, // SentencePiece + LLAMA_VOCAB_TYPE_BPE = 1, // Byte Pair Encoding +}; + +enum llama_token_type { + LLAMA_TOKEN_TYPE_UNDEFINED = 0, + LLAMA_TOKEN_TYPE_NORMAL = 1, + LLAMA_TOKEN_TYPE_UNKNOWN = 2, + LLAMA_TOKEN_TYPE_CONTROL = 3, + LLAMA_TOKEN_TYPE_USER_DEFINED = 4, + LLAMA_TOKEN_TYPE_UNUSED = 5, + LLAMA_TOKEN_TYPE_BYTE = 6, +}; + +// model file types +enum llama_ftype { + LLAMA_FTYPE_ALL_F32 = 0, + LLAMA_FTYPE_MOSTLY_F16 = 1, // except 1d tensors + LLAMA_FTYPE_MOSTLY_Q4_0 = 2, // except 1d tensors + LLAMA_FTYPE_MOSTLY_Q4_1 = 3, // except 1d tensors + LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16 + // LLAMA_FTYPE_MOSTLY_Q4_2 = 5, // support has been removed + // LLAMA_FTYPE_MOSTLY_Q4_3 = 6, // support has been removed + LLAMA_FTYPE_MOSTLY_Q8_0 = 7, // except 1d tensors + LLAMA_FTYPE_MOSTLY_Q5_0 = 8, // except 1d tensors + LLAMA_FTYPE_MOSTLY_Q5_1 = 9, // except 1d tensors + LLAMA_FTYPE_MOSTLY_Q2_K = 10, // except 1d tensors + LLAMA_FTYPE_MOSTLY_Q3_K_S = 11, // except 1d tensors + LLAMA_FTYPE_MOSTLY_Q3_K_M = 12, // except 1d tensors + LLAMA_FTYPE_MOSTLY_Q3_K_L = 13, // except 1d tensors + LLAMA_FTYPE_MOSTLY_Q4_K_S = 14, // except 1d tensors + LLAMA_FTYPE_MOSTLY_Q4_K_M = 15, // except 1d tensors + LLAMA_FTYPE_MOSTLY_Q5_K_S = 16, // except 1d tensors + LLAMA_FTYPE_MOSTLY_Q5_K_M = 17, // except 1d tensors + LLAMA_FTYPE_MOSTLY_Q6_K = 18, // except 1d tensors + + LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file +}; + +enum llama_rope_scaling_type { + LLAMA_ROPE_SCALING_UNSPECIFIED = -1, + LLAMA_ROPE_SCALING_NONE = 0, + LLAMA_ROPE_SCALING_LINEAR = 1, + LLAMA_ROPE_SCALING_YARN = 2, + LLAMA_ROPE_SCALING_MAX_VALUE = LLAMA_ROPE_SCALING_YARN, +}; + +typedef struct llama_token_data { + llama_token id; // token id + float logit; // log-odds of the token + float p; // probability of the token +} llama_token_data; + +typedef struct llama_token_data_array { + llama_token_data * data; + size_t size; + bool sorted; +} llama_token_data_array; + +typedef void (*llama_progress_callback)(float progress, void *ctx); + +// Input data for llama_decode +// A llama_batch object can contain input about one or many sequences +// The provided arrays (i.e. token, embd, pos, etc.) must have size of n_tokens +// +// - token : the token ids of the input (used when embd is NULL) +// - embd : token embeddings (i.e. float vector of size n_embd) (used when token is NULL) +// - pos : the positions of the respective token in the sequence +// - seq_id : the sequence to which the respective token belongs +// - logits : if zero, the logits for the respective token will not be output +// +typedef struct llama_batch { + int32_t n_tokens; + + llama_token * token; + float * embd; + llama_pos * pos; + int32_t * n_seq_id; + llama_seq_id ** seq_id; + int8_t * logits; + + // NOTE: helpers for smooth API transition - can be deprecated in the future + // for future-proof code, use the above fields instead and ignore everything below // - // C interface - // - // TODO: show sample usage + // pos[i] = all_pos_0 + i*all_pos_1 // + llama_pos all_pos_0; // used if pos == NULL + llama_pos all_pos_1; // used if pos == NULL + llama_seq_id all_seq_id; // used if seq_id == NULL +} llama_batch; - struct llama_model; - struct llama_context; +struct llama_model_params { + int32_t n_gpu_layers; // number of layers to store in VRAM + int32_t main_gpu; // the GPU that is used for scratch and small tensors + const float * tensor_split; // how to split layers across multiple GPUs (size: LLAMA_MAX_DEVICES) - typedef int32_t llama_pos; - typedef int32_t llama_token; - typedef int32_t llama_seq_id; + // called with a progress value between 0 and 1, pass NULL to disable + llama_progress_callback progress_callback; + // context pointer passed to the progress callback + void * progress_callback_user_data; - enum llama_vocab_type { - LLAMA_VOCAB_TYPE_SPM = 0, // SentencePiece - LLAMA_VOCAB_TYPE_BPE = 1, // Byte Pair Encoding - }; + // Keep the booleans together to avoid misalignment during copy-by-value. + bool vocab_only; // only load the vocabulary, no weights + bool use_mmap; // use mmap if possible + bool use_mlock; // force system to keep model in RAM +}; - enum llama_token_type { - LLAMA_TOKEN_TYPE_UNDEFINED = 0, - LLAMA_TOKEN_TYPE_NORMAL = 1, - LLAMA_TOKEN_TYPE_UNKNOWN = 2, - LLAMA_TOKEN_TYPE_CONTROL = 3, - LLAMA_TOKEN_TYPE_USER_DEFINED = 4, - LLAMA_TOKEN_TYPE_UNUSED = 5, - LLAMA_TOKEN_TYPE_BYTE = 6, - }; +struct llama_context_params { + uint32_t seed; // RNG seed, -1 for random + uint32_t n_ctx; // text context, 0 = from model + uint32_t n_batch; // prompt processing maximum batch size + uint32_t n_threads; // number of threads to use for generation + uint32_t n_threads_batch; // number of threads to use for batch processing + int8_t rope_scaling_type; // RoPE scaling type, from `enum llama_rope_scaling_type` - // model file types - enum llama_ftype { - LLAMA_FTYPE_ALL_F32 = 0, - LLAMA_FTYPE_MOSTLY_F16 = 1, // except 1d tensors - LLAMA_FTYPE_MOSTLY_Q4_0 = 2, // except 1d tensors - LLAMA_FTYPE_MOSTLY_Q4_1 = 3, // except 1d tensors - LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16 - // LLAMA_FTYPE_MOSTLY_Q4_2 = 5, // support has been removed - // LLAMA_FTYPE_MOSTLY_Q4_3 = 6, // support has been removed - LLAMA_FTYPE_MOSTLY_Q8_0 = 7, // except 1d tensors - LLAMA_FTYPE_MOSTLY_Q5_0 = 8, // except 1d tensors - LLAMA_FTYPE_MOSTLY_Q5_1 = 9, // except 1d tensors - LLAMA_FTYPE_MOSTLY_Q2_K = 10, // except 1d tensors - LLAMA_FTYPE_MOSTLY_Q3_K_S = 11, // except 1d tensors - LLAMA_FTYPE_MOSTLY_Q3_K_M = 12, // except 1d tensors - LLAMA_FTYPE_MOSTLY_Q3_K_L = 13, // except 1d tensors - LLAMA_FTYPE_MOSTLY_Q4_K_S = 14, // except 1d tensors - LLAMA_FTYPE_MOSTLY_Q4_K_M = 15, // except 1d tensors - LLAMA_FTYPE_MOSTLY_Q5_K_S = 16, // except 1d tensors - LLAMA_FTYPE_MOSTLY_Q5_K_M = 17, // except 1d tensors - LLAMA_FTYPE_MOSTLY_Q6_K = 18, // except 1d tensors + // ref: https://github.com/ggerganov/llama.cpp/pull/2054 + float rope_freq_base; // RoPE base frequency, 0 = from model + float rope_freq_scale; // RoPE frequency scaling factor, 0 = from model + float yarn_ext_factor; // YaRN extrapolation mix factor, NaN = from model + float yarn_attn_factor; // YaRN magnitude scaling factor + float yarn_beta_fast; // YaRN low correction dim + float yarn_beta_slow; // YaRN high correction dim + uint32_t yarn_orig_ctx; // YaRN original context size - LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file - }; + // Keep the booleans together to avoid misalignment during copy-by-value. + bool mul_mat_q; // if true, use experimental mul_mat_q kernels (DEPRECATED - always true) + bool f16_kv; // use fp16 for KV cache, fp32 otherwise + bool logits_all; // the llama_eval() call computes all logits, not just the last one + bool embedding; // embedding mode only +}; - typedef struct llama_token_data { - llama_token id; // token id - float logit; // log-odds of the token - float p; // probability of the token - } llama_token_data; +// model quantization parameters +typedef struct llama_model_quantize_params { + int nthread; // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency() + enum llama_ftype ftype; // quantize to this llama_ftype + bool allow_requantize; // allow quantizing non-f32/f16 tensors + bool quantize_output_tensor; // quantize output.weight + bool only_copy; // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored + bool pure; // disable k-quant mixtures and quantize all tensors to the same type +} llama_model_quantize_params; - typedef struct llama_token_data_array { - llama_token_data * data; - size_t size; - bool sorted; - } llama_token_data_array; +// grammar types +struct llama_grammar; - typedef void (*llama_progress_callback)(float progress, void *ctx); +// grammar element type +enum llama_gretype { + // end of rule definition + LLAMA_GRETYPE_END = 0, - // Input data for llama_decode - // A llama_batch object can contain input about one or many sequences - // The provided arrays (i.e. token, embd, pos, etc.) must have size of n_tokens - // - // - token : the token ids of the input (used when embd is NULL) - // - embd : token embeddings (i.e. float vector of size n_embd) (used when token is NULL) - // - pos : the positions of the respective token in the sequence - // - seq_id : the sequence to which the respective token belongs - // - logits : if zero, the logits for the respective token will not be output - // - typedef struct llama_batch { - int32_t n_tokens; + // start of alternate definition for rule + LLAMA_GRETYPE_ALT = 1, - llama_token * token; - float * embd; - llama_pos * pos; - int32_t * n_seq_id; - llama_seq_id ** seq_id; - int8_t * logits; + // non-terminal element: reference to rule + LLAMA_GRETYPE_RULE_REF = 2, - // NOTE: helpers for smooth API transition - can be deprecated in the future - // for future-proof code, use the above fields instead and ignore everything below - // - // pos[i] = all_pos_0 + i*all_pos_1 - // - llama_pos all_pos_0; // used if pos == NULL - llama_pos all_pos_1; // used if pos == NULL - llama_seq_id all_seq_id; // used if seq_id == NULL - } llama_batch; + // terminal element: character (code point) + LLAMA_GRETYPE_CHAR = 3, - struct llama_model_params { - int32_t n_gpu_layers; // number of layers to store in VRAM - int32_t main_gpu; // the GPU that is used for scratch and small tensors - const float * tensor_split; // how to split layers across multiple GPUs (size: LLAMA_MAX_DEVICES) + // inverse char(s) ([^a], [^a-b] [^abc]) + LLAMA_GRETYPE_CHAR_NOT = 4, - // called with a progress value between 0 and 1, pass NULL to disable - llama_progress_callback progress_callback; - // context pointer passed to the progress callback - void * progress_callback_user_data; + // modifies a preceding LLAMA_GRETYPE_CHAR or LLAMA_GRETYPE_CHAR_ALT to + // be an inclusive range ([a-z]) + LLAMA_GRETYPE_CHAR_RNG_UPPER = 5, - // Keep the booleans together to avoid misalignment during copy-by-value. - bool vocab_only; // only load the vocabulary, no weights - bool use_mmap; // use mmap if possible - bool use_mlock; // force system to keep model in RAM - }; + // modifies a preceding LLAMA_GRETYPE_CHAR or + // LLAMA_GRETYPE_CHAR_RNG_UPPER to add an alternate char to match ([ab], [a-zA]) + LLAMA_GRETYPE_CHAR_ALT = 6, +}; - struct llama_context_params { - uint32_t seed; // RNG seed, -1 for random - uint32_t n_ctx; // text context, 0 = from model - uint32_t n_batch; // prompt processing maximum batch size - uint32_t n_threads; // number of threads to use for generation - uint32_t n_threads_batch; // number of threads to use for batch processing +typedef struct llama_grammar_element { + enum llama_gretype type; + uint32_t value; // Unicode code point or rule ID +} llama_grammar_element; - // ref: https://github.com/ggerganov/llama.cpp/pull/2054 - float rope_freq_base; // RoPE base frequency, 0 = from model - float rope_freq_scale; // RoPE frequency scaling factor, 0 = from model +// performance timing information +struct llama_timings { + double t_start_ms; + double t_end_ms; + double t_load_ms; + double t_sample_ms; + double t_p_eval_ms; + double t_eval_ms; - // Keep the booleans together to avoid misalignment during copy-by-value. - bool mul_mat_q; // if true, use experimental mul_mat_q kernels (DEPRECATED - always true) - bool f16_kv; // use fp16 for KV cache, fp32 otherwise - bool logits_all; // the llama_eval() call computes all logits, not just the last one - bool embedding; // embedding mode only - }; + int32_t n_sample; + int32_t n_p_eval; + int32_t n_eval; +}; - // model quantization parameters - typedef struct llama_model_quantize_params { - int nthread; // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency() - enum llama_ftype ftype; // quantize to this llama_ftype - bool allow_requantize; // allow quantizing non-f32/f16 tensors - bool quantize_output_tensor; // quantize output.weight - bool only_copy; // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored - bool pure; // disable k-quant mixtures and quantize all tensors to the same type - } llama_model_quantize_params; +// Helpers for getting default parameters +LLAMA_API struct llama_model_params llama_model_default_params(void); +LLAMA_API struct llama_context_params llama_context_default_params(void); +LLAMA_API struct llama_model_quantize_params llama_model_quantize_default_params(void); - // grammar types - struct llama_grammar; +// Initialize the llama + ggml backend +// If numa is true, use NUMA optimizations +// Call once at the start of the program +LLAMA_API void llama_backend_init(bool numa); - // grammar element type - enum llama_gretype { - // end of rule definition - LLAMA_GRETYPE_END = 0, +// Call once at the end of the program - currently only used for MPI +LLAMA_API void llama_backend_free(void); - // start of alternate definition for rule - LLAMA_GRETYPE_ALT = 1, +LLAMA_API struct llama_model * llama_load_model_from_file( + const char * path_model, + struct llama_model_params params); - // non-terminal element: reference to rule - LLAMA_GRETYPE_RULE_REF = 2, +LLAMA_API void llama_free_model(struct llama_model * model); - // terminal element: character (code point) - LLAMA_GRETYPE_CHAR = 3, +LLAMA_API struct llama_context * llama_new_context_with_model( + struct llama_model * model, + struct llama_context_params params); - // inverse char(s) ([^a], [^a-b] [^abc]) - LLAMA_GRETYPE_CHAR_NOT = 4, +// Frees all allocated memory +LLAMA_API void llama_free(struct llama_context * ctx); - // modifies a preceding LLAMA_GRETYPE_CHAR or LLAMA_GRETYPE_CHAR_ALT to - // be an inclusive range ([a-z]) - LLAMA_GRETYPE_CHAR_RNG_UPPER = 5, +LLAMA_API int64_t llama_time_us(void); - // modifies a preceding LLAMA_GRETYPE_CHAR or - // LLAMA_GRETYPE_CHAR_RNG_UPPER to add an alternate char to match ([ab], [a-zA]) - LLAMA_GRETYPE_CHAR_ALT = 6, - }; +LLAMA_API int llama_max_devices (void); +LLAMA_API bool llama_mmap_supported (void); +LLAMA_API bool llama_mlock_supported(void); - typedef struct llama_grammar_element { - enum llama_gretype type; - uint32_t value; // Unicode code point or rule ID - } llama_grammar_element; +LLAMA_API const struct llama_model * llama_get_model(const struct llama_context * ctx); - // performance timing information - struct llama_timings { - double t_start_ms; - double t_end_ms; - double t_load_ms; - double t_sample_ms; - double t_p_eval_ms; - double t_eval_ms; +LLAMA_API int llama_n_ctx (const struct llama_context * ctx); - int32_t n_sample; - int32_t n_p_eval; - int32_t n_eval; - }; +LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_model * model); - // Helpers for getting default parameters - LLAMA_API struct llama_model_params llama_model_default_params(void); - LLAMA_API struct llama_context_params llama_context_default_params(void); - LLAMA_API struct llama_model_quantize_params llama_model_quantize_default_params(void); +LLAMA_API int llama_n_vocab (const struct llama_model * model); +LLAMA_API int llama_n_ctx_train(const struct llama_model * model); +LLAMA_API int llama_n_embd (const struct llama_model * model); - // Initialize the llama + ggml backend - // If numa is true, use NUMA optimizations - // Call once at the start of the program - LLAMA_API void llama_backend_init(bool numa); +// Get the model's RoPE frequency scaling factor +LLAMA_API float llama_rope_freq_scale_train(const struct llama_model * model); - // Call once at the end of the program - currently only used for MPI - LLAMA_API void llama_backend_free(void); +// Functions to access the model's GGUF metadata scalar values +// - The functions return the length of the string on success, or -1 on failure +// - The output string is always null-terminated and cleared on failure +// - GGUF array values are not supported by these functions - LLAMA_API struct llama_model * llama_load_model_from_file( - const char * path_model, - struct llama_model_params params); +// Get metadata value as a string by key name +LLAMA_API int llama_model_meta_val_str(const struct llama_model * model, const char * key, char * buf, size_t buf_size); - LLAMA_API void llama_free_model(struct llama_model * model); +// Get the number of metadata key/value pairs +LLAMA_API int llama_model_meta_count(const struct llama_model * model); - LLAMA_API struct llama_context * llama_new_context_with_model( - struct llama_model * model, - struct llama_context_params params); +// Get metadata key name by index +LLAMA_API int llama_model_meta_key_by_index(const struct llama_model * model, int i, char * buf, size_t buf_size); - // Frees all allocated memory - LLAMA_API void llama_free(struct llama_context * ctx); +// Get metadata value as a string by index +LLAMA_API int llama_model_meta_val_str_by_index(const struct llama_model * model, int i, char * buf, size_t buf_size); - LLAMA_API int64_t llama_time_us(void); +// Get a string describing the model type +LLAMA_API int llama_model_desc(const struct llama_model * model, char * buf, size_t buf_size); - LLAMA_API int llama_max_devices (void); - LLAMA_API bool llama_mmap_supported (void); - LLAMA_API bool llama_mlock_supported(void); +// Returns the total size of all the tensors in the model in bytes +LLAMA_API uint64_t llama_model_size(const struct llama_model * model); - LLAMA_API const struct llama_model * llama_get_model(const struct llama_context * ctx); +// Returns the total number of parameters in the model +LLAMA_API uint64_t llama_model_n_params(const struct llama_model * model); - LLAMA_API int llama_n_ctx (const struct llama_context * ctx); +// Get a llama model tensor +LLAMA_API struct ggml_tensor * llama_get_model_tensor(struct llama_model * model, const char * name); - LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_model * model); +// Returns 0 on success +LLAMA_API int llama_model_quantize( + const char * fname_inp, + const char * fname_out, + const llama_model_quantize_params * params); - LLAMA_API int llama_n_vocab (const struct llama_model * model); - LLAMA_API int llama_n_ctx_train(const struct llama_model * model); - LLAMA_API int llama_n_embd (const struct llama_model * model); - - // Get the model's RoPE frequency scaling factor - LLAMA_API float llama_rope_freq_scale_train(const struct llama_model * model); - - // Get a string describing the model type - LLAMA_API int llama_model_desc(const struct llama_model * model, char * buf, size_t buf_size); - - // Returns the total size of all the tensors in the model in bytes - LLAMA_API uint64_t llama_model_size(const struct llama_model * model); - - // Returns the total number of parameters in the model - LLAMA_API uint64_t llama_model_n_params(const struct llama_model * model); - - // Get a llama model tensor - LLAMA_API struct ggml_tensor * llama_get_model_tensor(struct llama_model * model, const char * name); - - // Returns 0 on success - LLAMA_API int llama_model_quantize( - const char * fname_inp, - const char * fname_out, - const llama_model_quantize_params * params); - - // Apply a LoRA adapter to a loaded model - // path_base_model is the path to a higher quality model to use as a base for - // the layers modified by the adapter. Can be NULL to use the current loaded model. - // The model needs to be reloaded before applying a new adapter, otherwise the adapter - // will be applied on top of the previous one - // Returns 0 on success - LLAMA_API DEPRECATED(int llama_apply_lora_from_file( - struct llama_context * ctx, - const char * path_lora, - float scale, - const char * path_base_model, +// Apply a LoRA adapter to a loaded model +// path_base_model is the path to a higher quality model to use as a base for +// the layers modified by the adapter. Can be NULL to use the current loaded model. +// The model needs to be reloaded before applying a new adapter, otherwise the adapter +// will be applied on top of the previous one +// Returns 0 on success +LLAMA_API DEPRECATED(int llama_apply_lora_from_file( + struct llama_context * ctx, + const char * path_lora, + float scale, + const char * path_base_model, int n_threads), - "use llama_model_apply_lora_from_file instead"); + "use llama_model_apply_lora_from_file instead"); - LLAMA_API int llama_model_apply_lora_from_file( - const struct llama_model * model, - const char * path_lora, - float scale, - const char * path_base_model, - int n_threads); - - // - // KV cache - // - - // Returns the number of tokens in the KV cache - LLAMA_API DEPRECATED(int llama_get_kv_cache_token_count(const struct llama_context * ctx), - "avoid using this, it will be removed in the future, instead - count the tokens in user code"); - - // Clear the KV cache - LLAMA_API void llama_kv_cache_clear( - struct llama_context * ctx); - - // Removes all tokens that belong to the specified sequence and have positions in [p0, p1) - // seq_id < 0 : match any sequence - // p0 < 0 : [0, p1] - // p1 < 0 : [p0, inf) - LLAMA_API void llama_kv_cache_seq_rm( - struct llama_context * ctx, - llama_seq_id seq_id, - llama_pos p0, - llama_pos p1); - - // Copy all tokens that belong to the specified sequence to another sequence - // Note that this does not allocate extra KV cache memory - it simply assigns the tokens to the new sequence - // p0 < 0 : [0, p1] - // p1 < 0 : [p0, inf) - LLAMA_API void llama_kv_cache_seq_cp( - struct llama_context * ctx, - llama_seq_id seq_id_src, - llama_seq_id seq_id_dst, - llama_pos p0, - llama_pos p1); - - // Removes all tokens that do not belong to the specified sequence - LLAMA_API void llama_kv_cache_seq_keep( - struct llama_context * ctx, - llama_seq_id seq_id); - - // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1) - // If the KV cache is RoPEd, the KV data is updated accordingly - // p0 < 0 : [0, p1] - // p1 < 0 : [p0, inf) - LLAMA_API void llama_kv_cache_seq_shift( - struct llama_context * ctx, - llama_seq_id seq_id, - llama_pos p0, - llama_pos p1, - llama_pos delta); - - // - // State / sessions - // - - // Returns the maximum size in bytes of the state (rng, logits, embedding - // and kv_cache) - will often be smaller after compacting tokens - LLAMA_API size_t llama_get_state_size(const struct llama_context * ctx); - - // Copies the state to the specified destination address. - // Destination needs to have allocated enough memory. - // Returns the number of bytes copied - LLAMA_API size_t llama_copy_state_data( - struct llama_context * ctx, - uint8_t * dst); - - // Set the state reading from the specified address - // Returns the number of bytes read - LLAMA_API size_t llama_set_state_data( - struct llama_context * ctx, - uint8_t * src); - - // Save/load session file - LLAMA_API bool llama_load_session_file( - struct llama_context * ctx, - const char * path_session, - llama_token * tokens_out, - size_t n_token_capacity, - size_t * n_token_count_out); - - LLAMA_API bool llama_save_session_file( - struct llama_context * ctx, - const char * path_session, - const llama_token * tokens, - size_t n_token_count); - - // - // Decoding - // - - // Run the llama inference to obtain the logits and probabilities for the next token(s). - // tokens + n_tokens is the provided batch of new tokens to process - // n_past is the number of tokens to use from previous eval calls - // Returns 0 on success - // DEPRECATED: use llama_decode() instead - LLAMA_API DEPRECATED(int llama_eval( - struct llama_context * ctx, - llama_token * tokens, - int32_t n_tokens, - int n_past), - "use llama_decode() instead"); - - // Same as llama_eval, but use float matrix input directly. - // DEPRECATED: use llama_decode() instead - LLAMA_API DEPRECATED(int llama_eval_embd( - struct llama_context * ctx, - float * embd, - int32_t n_tokens, - int n_past), - "use llama_decode() instead"); - - // Return batch for single sequence of tokens starting at pos_0 - // - // NOTE: this is a helper function to facilitate transition to the new batch API - avoid using it - // - LLAMA_API struct llama_batch llama_batch_get_one( - llama_token * tokens, - int32_t n_tokens, - llama_pos pos_0, - llama_seq_id seq_id); - - // Allocates a batch of tokens on the heap that can hold a maximum of n_tokens - // Each token can be assigned up to n_seq_max sequence ids - // The batch has to be freed with llama_batch_free() - // If embd != 0, llama_batch.embd will be allocated with size of n_tokens * embd * sizeof(float) - // Otherwise, llama_batch.token will be allocated to store n_tokens llama_token - // The rest of the llama_batch members are allocated with size n_tokens - // All members are left uninitialized - LLAMA_API struct llama_batch llama_batch_init( - int32_t n_tokens, - int32_t embd, - int32_t n_seq_max); - - // Frees a batch of tokens allocated with llama_batch_init() - LLAMA_API void llama_batch_free(struct llama_batch batch); - - // Positive return values does not mean a fatal error, but rather a warning. - // 0 - success - // 1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context) - // < 0 - error - LLAMA_API int llama_decode( - struct llama_context * ctx, - struct llama_batch batch); - - // Set the number of threads used for decoding - // n_threads is the number of threads used for generation (single token) - // n_threads_batch is the number of threads used for prompt and batch processing (multiple tokens) - LLAMA_API void llama_set_n_threads(struct llama_context * ctx, uint32_t n_threads, uint32_t n_threads_batch); - - // Token logits obtained from the last call to llama_eval() - // The logits for the last token are stored in the last row - // Logits for which llama_batch.logits[i] == 0 are undefined - // Rows: n_tokens provided with llama_batch - // Cols: n_vocab - LLAMA_API float * llama_get_logits(struct llama_context * ctx); - - // Logits for the ith token. Equivalent to: - // llama_get_logits(ctx) + i*n_vocab - LLAMA_API float * llama_get_logits_ith(struct llama_context * ctx, int32_t i); - - // Get the embeddings for the input - // shape: [n_embd] (1-dimensional) - LLAMA_API float * llama_get_embeddings(struct llama_context * ctx); - - // - // Vocab - // - - LLAMA_API const char * llama_token_get_text(const struct llama_model * model, llama_token token); - - LLAMA_API float llama_token_get_score(const struct llama_model * model, llama_token token); - - LLAMA_API enum llama_token_type llama_token_get_type(const struct llama_model * model, llama_token token); - - // Special tokens - LLAMA_API llama_token llama_token_bos(const struct llama_model * model); // beginning-of-sentence - LLAMA_API llama_token llama_token_eos(const struct llama_model * model); // end-of-sentence - LLAMA_API llama_token llama_token_nl (const struct llama_model * model); // next-line - - // codellama infill tokens - LLAMA_API llama_token llama_token_prefix(const struct llama_model * model); // Beginning of infill prefix - LLAMA_API llama_token llama_token_middle(const struct llama_model * model); // Beginning of infill middle - LLAMA_API llama_token llama_token_suffix(const struct llama_model * model); // Beginning of infill suffix - LLAMA_API llama_token llama_token_eot (const struct llama_model * model); // End of infill middle - - // - // Tokenization - // - - /// @details Convert the provided text into tokens. - /// @param tokens The tokens pointer must be large enough to hold the resulting tokens. - /// @return Returns the number of tokens on success, no more than n_max_tokens - /// @return Returns a negative number on failure - the number of tokens that would have been returned - /// @param special Allow tokenizing special and/or control tokens which otherwise are not exposed and treated as plaintext. - /// Does not insert a leading space. - LLAMA_API int llama_tokenize( +LLAMA_API int llama_model_apply_lora_from_file( const struct llama_model * model, - const char * text, - int text_len, - llama_token * tokens, - int n_max_tokens, - bool add_bos, - bool special); + const char * path_lora, + float scale, + const char * path_base_model, + int n_threads); - // Token Id -> Piece. - // Uses the vocabulary in the provided context. - // Does not write null terminator to the buffer. - // User code is responsible to remove the leading whitespace of the first non-BOS token when decoding multiple tokens. - LLAMA_API int llama_token_to_piece( - const struct llama_model * model, - llama_token token, - char * buf, - int length); +// +// KV cache +// - // - // Grammar - // +// Returns the number of tokens in the KV cache +LLAMA_API DEPRECATED(int llama_get_kv_cache_token_count(const struct llama_context * ctx), + "avoid using this, it will be removed in the future, instead - count the tokens in user code"); - LLAMA_API struct llama_grammar * llama_grammar_init( - const llama_grammar_element ** rules, - size_t n_rules, - size_t start_rule_index); +// Clear the KV cache +LLAMA_API void llama_kv_cache_clear( + struct llama_context * ctx); - LLAMA_API void llama_grammar_free(struct llama_grammar * grammar); +// Removes all tokens that belong to the specified sequence and have positions in [p0, p1) +// seq_id < 0 : match any sequence +// p0 < 0 : [0, p1] +// p1 < 0 : [p0, inf) +LLAMA_API void llama_kv_cache_seq_rm( + struct llama_context * ctx, + llama_seq_id seq_id, + llama_pos p0, + llama_pos p1); - LLAMA_API struct llama_grammar * llama_grammar_copy(const struct llama_grammar * grammar); +// Copy all tokens that belong to the specified sequence to another sequence +// Note that this does not allocate extra KV cache memory - it simply assigns the tokens to the new sequence +// p0 < 0 : [0, p1] +// p1 < 0 : [p0, inf) +LLAMA_API void llama_kv_cache_seq_cp( + struct llama_context * ctx, + llama_seq_id seq_id_src, + llama_seq_id seq_id_dst, + llama_pos p0, + llama_pos p1); - // - // Sampling functions - // +// Removes all tokens that do not belong to the specified sequence +LLAMA_API void llama_kv_cache_seq_keep( + struct llama_context * ctx, + llama_seq_id seq_id); - // Sets the current rng seed. - LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, uint32_t seed); +// Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1) +// If the KV cache is RoPEd, the KV data is updated accordingly +// p0 < 0 : [0, p1] +// p1 < 0 : [p0, inf) +LLAMA_API void llama_kv_cache_seq_shift( + struct llama_context * ctx, + llama_seq_id seq_id, + llama_pos p0, + llama_pos p1, + llama_pos delta); - /// @details Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix. - /// @details Frequency and presence penalties described in OpenAI API https://platform.openai.com/docs/api-reference/parameter-details. - LLAMA_API void llama_sample_repetition_penalties( - struct llama_context * ctx, - llama_token_data_array * candidates, - const llama_token * last_tokens, - size_t penalty_last_n, - float penalty_repeat, - float penalty_freq, - float penalty_present); +// +// State / sessions +// - /// @details Apply classifier-free guidance to the logits as described in academic paper "Stay on topic with Classifier-Free Guidance" https://arxiv.org/abs/2306.17806 - /// @param candidates A vector of `llama_token_data` containing the candidate tokens, the logits must be directly extracted from the original generation context without being sorted. - /// @params guidance_ctx A separate context from the same model. Other than a negative prompt at the beginning, it should have all generated and user input tokens copied from the main context. - /// @params scale Guidance strength. 1.0f means no guidance. Higher values mean stronger guidance. - LLAMA_API void llama_sample_classifier_free_guidance( - struct llama_context * ctx, - llama_token_data_array * candidates, - struct llama_context * guidance_ctx, - float scale); +// Returns the maximum size in bytes of the state (rng, logits, embedding +// and kv_cache) - will often be smaller after compacting tokens +LLAMA_API size_t llama_get_state_size(const struct llama_context * ctx); - /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits. - LLAMA_API void llama_sample_softmax( - struct llama_context * ctx, - llama_token_data_array * candidates); +// Copies the state to the specified destination address. +// Destination needs to have allocated enough memory. +// Returns the number of bytes copied +LLAMA_API size_t llama_copy_state_data( + struct llama_context * ctx, + uint8_t * dst); - /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751 - LLAMA_API void llama_sample_top_k( - struct llama_context * ctx, - llama_token_data_array * candidates, - int k, - size_t min_keep); +// Set the state reading from the specified address +// Returns the number of bytes read +LLAMA_API size_t llama_set_state_data( + struct llama_context * ctx, + uint8_t * src); - /// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751 - LLAMA_API void llama_sample_top_p( - struct llama_context * ctx, - llama_token_data_array * candidates, - float p, - size_t min_keep); +// Save/load session file +LLAMA_API bool llama_load_session_file( + struct llama_context * ctx, + const char * path_session, + llama_token * tokens_out, + size_t n_token_capacity, + size_t * n_token_count_out); - /// @details Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841 - LLAMA_API void llama_sample_min_p( - struct llama_context * ctx, - llama_token_data_array * candidates, - float p, - size_t min_keep); +LLAMA_API bool llama_save_session_file( + struct llama_context * ctx, + const char * path_session, + const llama_token * tokens, + size_t n_token_count); - /// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/. - LLAMA_API void llama_sample_tail_free( - struct llama_context * ctx, - llama_token_data_array * candidates, - float z, - size_t min_keep); +// +// Decoding +// - /// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666. - LLAMA_API void llama_sample_typical( - struct llama_context * ctx, - llama_token_data_array * candidates, - float p, - size_t min_keep); +// Run the llama inference to obtain the logits and probabilities for the next token(s). +// tokens + n_tokens is the provided batch of new tokens to process +// n_past is the number of tokens to use from previous eval calls +// Returns 0 on success +// DEPRECATED: use llama_decode() instead +LLAMA_API DEPRECATED(int llama_eval( + struct llama_context * ctx, + llama_token * tokens, + int32_t n_tokens, + int n_past), + "use llama_decode() instead"); - LLAMA_API void llama_sample_temp( - struct llama_context * ctx, - llama_token_data_array * candidates, - float temp); +// Same as llama_eval, but use float matrix input directly. +// DEPRECATED: use llama_decode() instead +LLAMA_API DEPRECATED(int llama_eval_embd( + struct llama_context * ctx, + float * embd, + int32_t n_tokens, + int n_past), + "use llama_decode() instead"); - LLAMA_API DEPRECATED(void llama_sample_temperature( - struct llama_context * ctx, - llama_token_data_array * candidates, - float temp), - "use llama_sample_temp instead"); +// Return batch for single sequence of tokens starting at pos_0 +// +// NOTE: this is a helper function to facilitate transition to the new batch API - avoid using it +// +LLAMA_API struct llama_batch llama_batch_get_one( + llama_token * tokens, + int32_t n_tokens, + llama_pos pos_0, + llama_seq_id seq_id); - /// @details Apply constraints from grammar - LLAMA_API void llama_sample_grammar( - struct llama_context * ctx, - llama_token_data_array * candidates, - const struct llama_grammar * grammar); +// Allocates a batch of tokens on the heap that can hold a maximum of n_tokens +// Each token can be assigned up to n_seq_max sequence ids +// The batch has to be freed with llama_batch_free() +// If embd != 0, llama_batch.embd will be allocated with size of n_tokens * embd * sizeof(float) +// Otherwise, llama_batch.token will be allocated to store n_tokens llama_token +// The rest of the llama_batch members are allocated with size n_tokens +// All members are left uninitialized +LLAMA_API struct llama_batch llama_batch_init( + int32_t n_tokens, + int32_t embd, + int32_t n_seq_max); - /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words. - /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text. - /// @param tau The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text. - /// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates. - /// @param m The number of tokens considered in the estimation of `s_hat`. This is an arbitrary value that is used to calculate `s_hat`, which in turn helps to calculate the value of `k`. In the paper, they use `m = 100`, but you can experiment with different values to see how it affects the performance of the algorithm. - /// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal. - LLAMA_API llama_token llama_sample_token_mirostat( - struct llama_context * ctx, - llama_token_data_array * candidates, - float tau, - float eta, - int m, - float * mu); +// Frees a batch of tokens allocated with llama_batch_init() +LLAMA_API void llama_batch_free(struct llama_batch batch); - /// @details Mirostat 2.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words. - /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text. - /// @param tau The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text. - /// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates. - /// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal. - LLAMA_API llama_token llama_sample_token_mirostat_v2( - struct llama_context * ctx, - llama_token_data_array * candidates, - float tau, - float eta, - float * mu); +// Positive return values does not mean a fatal error, but rather a warning. +// 0 - success +// 1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context) +// < 0 - error +LLAMA_API int llama_decode( + struct llama_context * ctx, + struct llama_batch batch); - /// @details Selects the token with the highest probability. - /// Does not compute the token probabilities. Use llama_sample_softmax() instead. - LLAMA_API llama_token llama_sample_token_greedy( - struct llama_context * ctx, - llama_token_data_array * candidates); +// Set the number of threads used for decoding +// n_threads is the number of threads used for generation (single token) +// n_threads_batch is the number of threads used for prompt and batch processing (multiple tokens) +LLAMA_API void llama_set_n_threads(struct llama_context * ctx, uint32_t n_threads, uint32_t n_threads_batch); - /// @details Randomly selects a token from the candidates based on their probabilities. - LLAMA_API llama_token llama_sample_token( - struct llama_context * ctx, - llama_token_data_array * candidates); +// Token logits obtained from the last call to llama_eval() +// The logits for the last token are stored in the last row +// Logits for which llama_batch.logits[i] == 0 are undefined +// Rows: n_tokens provided with llama_batch +// Cols: n_vocab +LLAMA_API float * llama_get_logits(struct llama_context * ctx); - /// @details Accepts the sampled token into the grammar - LLAMA_API void llama_grammar_accept_token( - struct llama_context * ctx, - struct llama_grammar * grammar, - llama_token token); +// Logits for the ith token. Equivalent to: +// llama_get_logits(ctx) + i*n_vocab +LLAMA_API float * llama_get_logits_ith(struct llama_context * ctx, int32_t i); - // - // Beam search - // +// Get the embeddings for the input +// shape: [n_embd] (1-dimensional) +LLAMA_API float * llama_get_embeddings(struct llama_context * ctx); - struct llama_beam_view { - const llama_token * tokens; +// +// Vocab +// - size_t n_tokens; - float p; // Cumulative beam probability (renormalized relative to all beams) - bool eob; // Callback should set this to true when a beam is at end-of-beam. - }; +LLAMA_API const char * llama_token_get_text(const struct llama_model * model, llama_token token); - // Passed to beam_search_callback function. - // Whenever 0 < common_prefix_length, this number of tokens should be copied from any of the beams - // (e.g. beams[0]) as they will be removed (shifted) from all beams in all subsequent callbacks. - // These pointers are valid only during the synchronous callback, so should not be saved. - struct llama_beams_state { - struct llama_beam_view * beam_views; +LLAMA_API float llama_token_get_score(const struct llama_model * model, llama_token token); - size_t n_beams; // Number of elements in beam_views[]. - size_t common_prefix_length; // Current max length of prefix tokens shared by all beams. - bool last_call; // True iff this is the last callback invocation. - }; +LLAMA_API enum llama_token_type llama_token_get_type(const struct llama_model * model, llama_token token); - // Type of pointer to the beam_search_callback function. - // void* callback_data is any custom data passed to llama_beam_search, that is subsequently - // passed back to beam_search_callback. This avoids having to use global variables in the callback. - typedef void (*llama_beam_search_callback_fn_t)(void * callback_data, struct llama_beams_state); +// Special tokens +LLAMA_API llama_token llama_token_bos(const struct llama_model * model); // beginning-of-sentence +LLAMA_API llama_token llama_token_eos(const struct llama_model * model); // end-of-sentence +LLAMA_API llama_token llama_token_nl (const struct llama_model * model); // next-line - /// @details Deterministically returns entire sentence constructed by a beam search. - /// @param ctx Pointer to the llama_context. - /// @param callback Invoked for each iteration of the beam_search loop, passing in beams_state. - /// @param callback_data A pointer that is simply passed back to callback. - /// @param n_beams Number of beams to use. - /// @param n_past Number of tokens already evaluated. - /// @param n_predict Maximum number of tokens to predict. EOS may occur earlier. - LLAMA_API void llama_beam_search( - struct llama_context * ctx, +// Returns -1 if unknown, 1 for true or 0 for false. +LLAMA_API int llama_add_bos_token(const struct llama_model * model); + +// Returns -1 if unknown, 1 for true or 0 for false. +LLAMA_API int llama_add_eos_token(const struct llama_model * model); + +// codellama infill tokens +LLAMA_API llama_token llama_token_prefix(const struct llama_model * model); // Beginning of infill prefix +LLAMA_API llama_token llama_token_middle(const struct llama_model * model); // Beginning of infill middle +LLAMA_API llama_token llama_token_suffix(const struct llama_model * model); // Beginning of infill suffix +LLAMA_API llama_token llama_token_eot (const struct llama_model * model); // End of infill middle + +// +// Tokenization +// + +/// @details Convert the provided text into tokens. +/// @param tokens The tokens pointer must be large enough to hold the resulting tokens. +/// @return Returns the number of tokens on success, no more than n_max_tokens +/// @return Returns a negative number on failure - the number of tokens that would have been returned +/// @param special Allow tokenizing special and/or control tokens which otherwise are not exposed and treated as plaintext. +/// Does not insert a leading space. +LLAMA_API int llama_tokenize( + const struct llama_model * model, + const char * text, + int text_len, + llama_token * tokens, + int n_max_tokens, + bool add_bos, + bool special); + +// Token Id -> Piece. +// Uses the vocabulary in the provided context. +// Does not write null terminator to the buffer. +// User code is responsible to remove the leading whitespace of the first non-BOS token when decoding multiple tokens. +LLAMA_API int llama_token_to_piece( + const struct llama_model * model, + llama_token token, + char * buf, + int length); + +// +// Grammar +// + +LLAMA_API struct llama_grammar * llama_grammar_init( + const llama_grammar_element ** rules, + size_t n_rules, + size_t start_rule_index); + +LLAMA_API void llama_grammar_free(struct llama_grammar * grammar); + +LLAMA_API struct llama_grammar * llama_grammar_copy(const struct llama_grammar * grammar); + +// +// Sampling functions +// + +// Sets the current rng seed. +LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, uint32_t seed); + +/// @details Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix. +/// @details Frequency and presence penalties described in OpenAI API https://platform.openai.com/docs/api-reference/parameter-details. +LLAMA_API void llama_sample_repetition_penalties( + struct llama_context * ctx, + llama_token_data_array * candidates, + const llama_token * last_tokens, + size_t penalty_last_n, + float penalty_repeat, + float penalty_freq, + float penalty_present); + +/// @details Apply classifier-free guidance to the logits as described in academic paper "Stay on topic with Classifier-Free Guidance" https://arxiv.org/abs/2306.17806 +/// @param candidates A vector of `llama_token_data` containing the candidate tokens, the logits must be directly extracted from the original generation context without being sorted. +/// @params guidance_ctx A separate context from the same model. Other than a negative prompt at the beginning, it should have all generated and user input tokens copied from the main context. +/// @params scale Guidance strength. 1.0f means no guidance. Higher values mean stronger guidance. +LLAMA_API void llama_sample_classifier_free_guidance( + struct llama_context * ctx, + llama_token_data_array * candidates, + struct llama_context * guidance_ctx, + float scale); + +/// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits. +LLAMA_API void llama_sample_softmax( + struct llama_context * ctx, + llama_token_data_array * candidates); + +/// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751 +LLAMA_API void llama_sample_top_k( + struct llama_context * ctx, + llama_token_data_array * candidates, + int k, + size_t min_keep); + +/// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751 +LLAMA_API void llama_sample_top_p( + struct llama_context * ctx, + llama_token_data_array * candidates, + float p, + size_t min_keep); + +/// @details Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841 +LLAMA_API void llama_sample_min_p( + struct llama_context * ctx, + llama_token_data_array * candidates, + float p, + size_t min_keep); + +/// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/. +LLAMA_API void llama_sample_tail_free( + struct llama_context * ctx, + llama_token_data_array * candidates, + float z, + size_t min_keep); + +/// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666. +LLAMA_API void llama_sample_typical( + struct llama_context * ctx, + llama_token_data_array * candidates, + float p, + size_t min_keep); + +LLAMA_API void llama_sample_temp( + struct llama_context * ctx, + llama_token_data_array * candidates, + float temp); + +LLAMA_API DEPRECATED(void llama_sample_temperature( + struct llama_context * ctx, + llama_token_data_array * candidates, + float temp), + "use llama_sample_temp instead"); + +/// @details Apply constraints from grammar +LLAMA_API void llama_sample_grammar( + struct llama_context * ctx, + llama_token_data_array * candidates, + const struct llama_grammar * grammar); + +/// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words. +/// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text. +/// @param tau The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text. +/// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates. +/// @param m The number of tokens considered in the estimation of `s_hat`. This is an arbitrary value that is used to calculate `s_hat`, which in turn helps to calculate the value of `k`. In the paper, they use `m = 100`, but you can experiment with different values to see how it affects the performance of the algorithm. +/// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal. +LLAMA_API llama_token llama_sample_token_mirostat( + struct llama_context * ctx, + llama_token_data_array * candidates, + float tau, + float eta, + int m, + float * mu); + +/// @details Mirostat 2.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words. +/// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text. +/// @param tau The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text. +/// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates. +/// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal. +LLAMA_API llama_token llama_sample_token_mirostat_v2( + struct llama_context * ctx, + llama_token_data_array * candidates, + float tau, + float eta, + float * mu); + +/// @details Selects the token with the highest probability. +/// Does not compute the token probabilities. Use llama_sample_softmax() instead. +LLAMA_API llama_token llama_sample_token_greedy( + struct llama_context * ctx, + llama_token_data_array * candidates); + +/// @details Randomly selects a token from the candidates based on their probabilities. +LLAMA_API llama_token llama_sample_token( + struct llama_context * ctx, + llama_token_data_array * candidates); + +/// @details Accepts the sampled token into the grammar +LLAMA_API void llama_grammar_accept_token( + struct llama_context * ctx, + struct llama_grammar * grammar, + llama_token token); + +// +// Beam search +// + +struct llama_beam_view { + const llama_token * tokens; + + size_t n_tokens; + float p; // Cumulative beam probability (renormalized relative to all beams) + bool eob; // Callback should set this to true when a beam is at end-of-beam. +}; + +// Passed to beam_search_callback function. +// Whenever 0 < common_prefix_length, this number of tokens should be copied from any of the beams +// (e.g. beams[0]) as they will be removed (shifted) from all beams in all subsequent callbacks. +// These pointers are valid only during the synchronous callback, so should not be saved. +struct llama_beams_state { + struct llama_beam_view * beam_views; + + size_t n_beams; // Number of elements in beam_views[]. + size_t common_prefix_length; // Current max length of prefix tokens shared by all beams. + bool last_call; // True iff this is the last callback invocation. +}; + +// Type of pointer to the beam_search_callback function. +// void* callback_data is any custom data passed to llama_beam_search, that is subsequently +// passed back to beam_search_callback. This avoids having to use global variables in the callback. +typedef void (*llama_beam_search_callback_fn_t)(void * callback_data, struct llama_beams_state); + +/// @details Deterministically returns entire sentence constructed by a beam search. +/// @param ctx Pointer to the llama_context. +/// @param callback Invoked for each iteration of the beam_search loop, passing in beams_state. +/// @param callback_data A pointer that is simply passed back to callback. +/// @param n_beams Number of beams to use. +/// @param n_past Number of tokens already evaluated. +/// @param n_predict Maximum number of tokens to predict. EOS may occur earlier. +LLAMA_API void llama_beam_search( + struct llama_context * ctx, llama_beam_search_callback_fn_t callback, - void * callback_data, - size_t n_beams, - int n_past, - int n_predict); + void * callback_data, + size_t n_beams, + int n_past, + int n_predict); - // Performance information - LLAMA_API struct llama_timings llama_get_timings(struct llama_context * ctx); +// Performance information +LLAMA_API struct llama_timings llama_get_timings(struct llama_context * ctx); - LLAMA_API void llama_print_timings(struct llama_context * ctx); - LLAMA_API void llama_reset_timings(struct llama_context * ctx); +LLAMA_API void llama_print_timings(struct llama_context * ctx); +LLAMA_API void llama_reset_timings(struct llama_context * ctx); - // Print system information - LLAMA_API const char * llama_print_system_info(void); +// Print system information +LLAMA_API const char * llama_print_system_info(void); - // Set callback for all future logging events. - // If this is not called, or NULL is supplied, everything is output on stderr. - LLAMA_API void llama_log_set(ggml_log_callback log_callback, void * user_data); +// Set callback for all future logging events. +// If this is not called, or NULL is supplied, everything is output on stderr. +LLAMA_API void llama_log_set(ggml_log_callback log_callback, void * user_data); - LLAMA_API void llama_dump_timing_info_yaml(FILE * stream, const struct llama_context * ctx); +LLAMA_API void llama_dump_timing_info_yaml(FILE * stream, const struct llama_context * ctx); #ifdef __cplusplus } @@ -751,7 +788,7 @@ extern "C" { struct ggml_tensor; const std::vector> & llama_internal_get_tensor_map( - struct llama_context * ctx + struct llama_context * ctx ); #endif // LLAMA_API_INTERNAL diff --git a/native/jni/src/ggml/train.cpp b/native/jni/src/ggml/train.cpp index b8b7286e5..c90ed84f8 100644 --- a/native/jni/src/ggml/train.cpp +++ b/native/jni/src/ggml/train.cpp @@ -33,6 +33,7 @@ struct train_state * init_train_state() { state->opt = new struct ggml_opt_context; state->opt->ctx = NULL; state->opt->params = ggml_opt_default_params(GGML_OPT_ADAM); + state->opt->params.graph_size = LLAMA_TRAIN_MAX_NODES; state->opt->loss_after = 0.0f; return state; @@ -1136,6 +1137,7 @@ void print_common_train_usage(int /*argc*/, char ** /*argv*/, const struct train fprintf(stderr, " --adam-beta2 N AdamW beta2 in interval [0,1). How much to smooth the second moment of gradients. (default %f)\n", params->adam_beta2); fprintf(stderr, " --adam-gclip N AdamW gradient clipping. Disabled when zero. (default %f)\n", params->adam_gclip); fprintf(stderr, " --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. (default %f)\n", params->adam_eps_f); + fprintf(stderr, " -ngl N, --n-gpu-layers N Number of model layers to offload to GPU (default %d)", params->n_gpu_layers); fprintf(stderr, "\n"); } @@ -1355,6 +1357,17 @@ bool consume_common_train_arg( return true; } params->adam_gclip = std::stof(argv[i]); + } else if (arg == "-ngl" || arg == "--n-gpu-layers") { + if (++i >= argc) { + *invalid_param = true; + return true; + } +#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD + params->n_gpu_layers = std::stoi(argv[i]); +#else + fprintf(stderr, "warning: not compiled with GPU offload support, --n-gpu-layers option will be ignored\n"); + fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n"); +#endif } else if (arg == "-h" || arg == "--help") { params->print_usage = true; return true; diff --git a/native/jni/src/ggml/train.h b/native/jni/src/ggml/train.h index 800440306..010075ae8 100644 --- a/native/jni/src/ggml/train.h +++ b/native/jni/src/ggml/train.h @@ -9,6 +9,8 @@ #include "ggml.h" #include "llama.h" +#define LLAMA_TRAIN_MAX_NODES 16384 + typedef std::string mt19937_state; struct train_state {