root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-24 10:24:35 +00:00
Code Issues Actions 5 Packages Projects Releases Wiki Activity

Slightly faster imatrix (#5050)

Browse Source 
* imatrix: speedup by avoiding unnecessary allocations and copies

* imatrix: add --no-ppl option to skip PPL calculations altogether

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow 2024-01-21 08:01:20 +02:00 committed by GitHub
parent 942c0107a7
commit 726c0fa9a2
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 49 additions and 25 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

74
examples/imatrix/imatrix.cpp
View File

@ -248,7 +248,7 @@ static void process_logits(
} }
} }
static bool compute_imatrix(llama_context * ctx, const gpt_params & params) { static bool compute_imatrix(llama_context * ctx, const gpt_params & params, bool compute_ppl) {
const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx)); const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
const int n_ctx = llama_n_ctx(ctx); const int n_ctx = llama_n_ctx(ctx);
@ -269,10 +269,12 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
} }
std::vector<float> logit_history; std::vector<float> logit_history;
logit_history.resize(tokens.size());
std::vector<float> prob_history; std::vector<float> prob_history;
prob_history.resize(tokens.size());
if (compute_ppl) {
logit_history.resize(tokens.size());
prob_history.resize(tokens.size());
}
const int n_chunk_max = tokens.size() / n_ctx; const int n_chunk_max = tokens.size() / n_ctx;
@ -288,12 +290,17 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
std::vector<std::thread> workers(std::thread::hardware_concurrency() - 1); std::vector<std::thread> workers(std::thread::hardware_concurrency() - 1);
const int num_batches = (n_ctx + n_batch - 1) / n_batch;
std::vector<float> logits;
if (compute_ppl && num_batches > 1) {
logits.reserve((size_t)n_ctx * n_vocab);
}
for (int i = 0; i < n_chunk; ++i) { for (int i = 0; i < n_chunk; ++i) {
const int start = i * n_ctx; const int start = i * n_ctx;
const int end = start + n_ctx; const int end = start + n_ctx;
const int num_batches = (n_ctx + n_batch - 1) / n_batch;
std::vector<float> logits; std::vector<float> logits;
const auto t_start = std::chrono::high_resolution_clock::now(); const auto t_start = std::chrono::high_resolution_clock::now();
@ -321,8 +328,10 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
// restore the original token in case it was set to BOS // restore the original token in case it was set to BOS
tokens[batch_start] = token_org; tokens[batch_start] = token_org;
const auto * batch_logits = llama_get_logits(ctx); if (compute_ppl && num_batches > 1) {
logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab); const auto * batch_logits = llama_get_logits(ctx);
logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);
}
} }
const auto t_end = std::chrono::high_resolution_clock::now(); const auto t_end = std::chrono::high_resolution_clock::now();
@ -338,25 +347,32 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0); fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0);
} }
const int first = n_ctx/2; if (compute_ppl) {
process_logits(n_vocab, logits.data() + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first, const int first = n_ctx/2;
workers, nll, nll2, logit_history.data() + start + first, prob_history.data() + start + first); const auto all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx);
count += n_ctx - first - 1; process_logits(n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first,
workers, nll, nll2, logit_history.data() + start + first, prob_history.data() + start + first);
count += n_ctx - first - 1;
printf("[%d]%.4lf,", i + 1, std::exp(nll / count)); printf("[%d]%.4lf,", i + 1, std::exp(nll / count));
fflush(stdout); fflush(stdout);
logits.clear();
}
} }
printf("\n"); printf("\n");
nll2 /= count; if (compute_ppl) {
nll /= count; nll2 /= count;
const double ppl = exp(nll); nll /= count;
nll2 -= nll * nll; const double ppl = exp(nll);
if (nll2 > 0) { nll2 -= nll * nll;
nll2 = sqrt(nll2/(count-1)); if (nll2 > 0) {
printf("Final estimate: PPL = %.4lf +/- %.5lf\n", ppl, nll2*ppl); nll2 = sqrt(nll2/(count-1));
} else { printf("Final estimate: PPL = %.4lf +/- %.5lf\n", ppl, nll2*ppl);
printf("Unexpected negative standard deviation of log(prob)\n"); } else {
printf("Unexpected negative standard deviation of log(prob)\n");
}
} }
return true; return true;
@ -365,6 +381,7 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
int main(int argc, char ** argv) { int main(int argc, char ** argv) {
StatParams sparams; StatParams sparams;
bool compute_ppl = true;
std::vector<char*> args; std::vector<char*> args;
args.push_back(argv[0]); args.push_back(argv[0]);
int iarg = 1; int iarg = 1;
@ -381,12 +398,19 @@ int main(int argc, char ** argv) {
} }
else if (arg == "--verbosity") { else if (arg == "--verbosity") {
sparams.verbosity = std::stoi(argv[++iarg]); sparams.verbosity = std::stoi(argv[++iarg]);
} else if (arg == "--no-ppl") {
compute_ppl = false;
} else { } else {
args.push_back(argv[iarg]); args.push_back(argv[iarg]);
} }
} }
if (iarg < argc) { if (iarg < argc) {
args.push_back(argv[iarg]); std::string arg{argv[iarg]};
if (arg == "--no-ppl") {
compute_ppl = false;
} else {
args.push_back(argv[iarg]);
}
} }
gpt_params params; gpt_params params;
@ -448,7 +472,7 @@ int main(int argc, char ** argv) {
fprintf(stderr, "%s\n", get_system_info(params).c_str()); fprintf(stderr, "%s\n", get_system_info(params).c_str());
} }
bool OK = compute_imatrix(ctx, params); bool OK = compute_imatrix(ctx, params, compute_ppl);
if (!OK) { if (!OK) {
return 1; return 1;
} }