Reduce logging

This commit is contained in:
Aleksandras Kostarevas 2024-04-11 00:43:56 -05:00
parent 725c679eef
commit 9308bcbfb0
2 changed files with 22 additions and 16 deletions

View File

@ -13,11 +13,16 @@
#include "jni_utils.h" #include "jni_utils.h"
#define EPS 0.0001 #define EPS 0.0001
#if false
#define TIME_START(name) const int64_t start_##name = ggml_time_us(); #define TIME_START(name) const int64_t start_##name = ggml_time_us();
#define TIME_END(name) const int64_t end_##name = ggml_time_us(); \ #define TIME_END(name) const int64_t end_##name = ggml_time_us(); \
const int64_t time_taken_##name = (end_##name - start_##name) / 1000L; \ const int64_t time_taken_##name = (end_##name - start_##name) / 1000L; \
AKLOGI("%s: Time taken by %s: %d ms\n", __func__, #name, (int)time_taken_##name); AKLOGI("%s: Time taken by %s: %d ms\n", __func__, #name, (int)time_taken_##name);
#else
#define TIME_START(name)
#define TIME_END(name)
#endif
#define RETURNVAL_AUTOCORRECT "autocorrect" #define RETURNVAL_AUTOCORRECT "autocorrect"
#define RETURNVAL_UNCERTAIN "uncertain" #define RETURNVAL_UNCERTAIN "uncertain"
@ -349,7 +354,7 @@ struct LanguageModelState {
return {}; return {};
} }
} else { } else {
AKLOGI("No need to recompute prompt, proceeding to mixes"); //AKLOGI("No need to recompute prompt, proceeding to mixes");
} }
transformer_context_apply(model->transformerContext, prompt_ff); transformer_context_apply(model->transformerContext, prompt_ff);
@ -362,7 +367,7 @@ struct LanguageModelState {
std::vector<float> embeds; std::vector<float> embeds;
bool useEncoder = !llamaAdapter->encoder_weight.empty(); bool useEncoder = !llamaAdapter->encoder_weight.empty();
AKLOGI("DecodePromptAndMixes: useEncoder=%d", useEncoder); //AKLOGI("DecodePromptAndMixes: useEncoder=%d", useEncoder);
for(auto &mix : mixes) { for(auto &mix : mixes) {
@ -474,13 +479,13 @@ struct LanguageModelState {
ASSERT(head == prompt_ff.first.size() - 1); ASSERT(head == prompt_ff.first.size() - 1);
} }
AKLOGI("-- Decode"); //AKLOGI("-- Decode");
AKLOGI("First we processed the prompt (%d):", prompt_ff.first.size()); //AKLOGI("First we processed the prompt (%d):", prompt_ff.first.size());
for(auto t : prompt) { //for(auto t : prompt) {
AKLOGI(" - [%s]", model->getToken(t)); // AKLOGI(" - [%s]", model->getToken(t));
} //}
AKLOGI("Then %d embeds (cached %d)", embeds.size(), n_past); //AKLOGI("Then %d embeds (cached %d)", embeds.size(), n_past);
AKLOGI("The final size is %d and head is %d", size, head); //AKLOGI("The final size is %d and head is %d", size, head);
TIME_START(FinishRm) TIME_START(FinishRm)
@ -509,7 +514,7 @@ struct LanguageModelState {
auto priorTxt = model->decode(prior); auto priorTxt = model->decode(prior);
auto nextTxt = model->decode({next}); auto nextTxt = model->decode({next});
auto bannedTxt = model->decode(banned_sequence.sequence); auto bannedTxt = model->decode(banned_sequence.sequence);
AKLOGI("Tokens [%s] + [%s] matches banned wildcard [%s]", priorTxt.c_str(), nextTxt.c_str(), bannedTxt.c_str()); //AKLOGI("Tokens [%s] + [%s] matches banned wildcard [%s]", priorTxt.c_str(), nextTxt.c_str(), bannedTxt.c_str());
return true; return true;
} }
}else if((banned_sequence.sequence.size() == prior.size() + 1) && (banned_sequence.hash == new_hash)) { }else if((banned_sequence.sequence.size() == prior.size() + 1) && (banned_sequence.hash == new_hash)) {
@ -526,7 +531,7 @@ struct LanguageModelState {
auto priorTxt = model->decode(prior); auto priorTxt = model->decode(prior);
auto nextTxt = model->decode({next}); auto nextTxt = model->decode({next});
auto bannedTxt = model->decode(banned_sequence.sequence); auto bannedTxt = model->decode(banned_sequence.sequence);
AKLOGI("Tokens [%s] + [%s] matches banned [%s]", priorTxt.c_str(), nextTxt.c_str(), bannedTxt.c_str()); //AKLOGI("Tokens [%s] + [%s] matches banned [%s]", priorTxt.c_str(), nextTxt.c_str(), bannedTxt.c_str());
return true; return true;
} }
} }
@ -547,7 +552,7 @@ struct LanguageModelState {
bool allow_correction_token = decodeResult.logits_head == 0; bool allow_correction_token = decodeResult.logits_head == 0;
float *logits = llama_get_logits_ith(ctx, decodeResult.logits_head); float *logits = llama_get_logits_ith(ctx, decodeResult.logits_head);
AKLOGI("Value of [the ] before transform: %f", logits[561]); //AKLOGI("Value of [the ] before transform: %f", logits[561]);
bool is_bugged = logits[561] == 0.0f; bool is_bugged = logits[561] == 0.0f;
@ -565,7 +570,7 @@ struct LanguageModelState {
return { }; return { };
} }
AKLOGI("Value of [the ] after transform: %f", logits[561]); //AKLOGI("Value of [the ] after transform: %f", logits[561]);
std::vector<std::pair<float, int>> index_value; std::vector<std::pair<float, int>> index_value;
index_value.clear(); index_value.clear();
@ -993,7 +998,7 @@ namespace latinime {
}else if(c >= 'A' && c <= 'Z') { }else if(c >= 'A' && c <= 'Z') {
results.mixes[j].token = (state->specialTokens.LETTERS_TO_IDS[c - 'A']); results.mixes[j].token = (state->specialTokens.LETTERS_TO_IDS[c - 'A']);
} else { } else {
AKLOGI("ignoring character in partial word [%c]", c); //AKLOGI("ignoring character in partial word [%c]", c);
results.mixes[j].weight = 0.0f; results.mixes[j].weight = 0.0f;
} }
} }

View File

@ -175,7 +175,8 @@ static inline void showStackTrace() {
#endif // defined(FLAG_DO_PROFILE) || defined(FLAG_DBG) #endif // defined(FLAG_DO_PROFILE) || defined(FLAG_DBG)
#ifdef FLAG_DBG #ifdef FLAG_DBG
#define DEBUG_DICT true //#define DEBUG_DICT true
#define DEBUG_DICT false
#define DEBUG_DICT_FULL false #define DEBUG_DICT_FULL false
#define DEBUG_EDIT_DISTANCE false #define DEBUG_EDIT_DISTANCE false
#define DEBUG_NODE DEBUG_DICT_FULL #define DEBUG_NODE DEBUG_DICT_FULL