Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Slightly faster imatrix #5050

Merged
merged 2 commits into from
Jan 21, 2024
Merged

Slightly faster imatrix #5050

Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
cdeac23
imatrix: speedup by avoiding unnecessary allocations and copies
Kawrakow Jan 20, 2024
3aa5656
imatrix: add --no-ppl option to skip PPL calculations altogether
Kawrakow Jan 20, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Next Next commit
imatrix: speedup by avoiding unnecessary allocations and copies
  • Loading branch information
@Kawrakow
Kawrakow committed Jan 20, 2024
commit cdeac23ef5dd97150da8406269af4fbd48fd9e97
20 changes: 15 additions & 5 deletions examples/imatrix/imatrix.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -288,12 +288,17 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {

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 (num_batches > 1) {
logits.reserve((size_t)n_ctx * n_vocab);
}

for (int i = 0; i < n_chunk; ++i) {
const int start = i * n_ctx;
const int end = start + n_ctx;

const int num_batches = (n_ctx + n_batch - 1) / n_batch;

std::vector<float> logits;

const auto t_start = std::chrono::high_resolution_clock::now();
Expand Down Expand Up @@ -321,8 +326,10 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
// restore the original token in case it was set to BOS
tokens[batch_start] = token_org;

const auto * batch_logits = llama_get_logits(ctx);
logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);
if (num_batches > 1) {
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();
Expand All @@ -339,12 +346,15 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
}

const int first = n_ctx/2;
process_logits(n_vocab, logits.data() + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first,
const auto all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx);
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));
fflush(stdout);

logits.clear();
}
printf("\n");

Expand Down