Skip to content

bench : cache the llama_context state at computed depth #16944

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

Merged
merged 3 commits into from
Nov 7, 2025
Merged

bench : cache the llama_context state at computed depth #16944

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
7d2daf8
bench : cache llama_context state at depth
ggerganov Nov 3, 2025
a99f7ce
cont : handle failures to restore the old state
ggerganov Nov 3, 2025
9c6bc80
cont : print information when the state is being reused
ggerganov Nov 5, 2025
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
47 changes: 40 additions & 7 deletions tools/llama-bench/llama-bench.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1919,6 +1919,12 @@ struct sql_printer : public printer {
}
};

struct ctx_state {
int depth = 0; // in tokens

std::vector<uint8_t> buf; // the llama_context state buffer
};

static bool test_prompt(llama_context * ctx, int n_prompt, int n_batch, int n_threads) {
llama_set_n_threads(ctx, n_threads, n_threads);

Expand Down Expand Up @@ -2051,6 +2057,10 @@ int main(int argc, char ** argv) {
llama_model * lmodel = nullptr;
const cmd_params_instance * prev_inst = nullptr;

// store the llama_context state at the previous depth that we performed a test
// ref: https://github.com/ggml-org/llama.cpp/pull/16944#issuecomment-3478151721
ctx_state cstate;

int params_idx = 0;
auto params_count = params_instances.size();
for (const auto & inst : params_instances) {
Expand Down Expand Up @@ -2134,14 +2144,37 @@ int main(int argc, char ** argv) {
llama_memory_clear(llama_get_memory(ctx), false);

if (t.n_depth > 0) {
if (params.progress) {
fprintf(stderr, "llama-bench: benchmark %d/%zu: depth run %d/%d\n", params_idx, params_count,
i + 1, params.reps);
bool is_cached = t.n_depth == cstate.depth;
Copy link
Member

@slaren slaren Nov 5, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Wouldn't this need also to check if the model is the same, using the same KV type, or other parameters that may make the cache incompatible with the current test?

Copy link
Member Author

@ggerganov ggerganov Nov 5, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

llama_state_seq_set_data should (in theory) return error (i.e. 0) when the state is incompatible with the current llama_context. I did try a few cases (different models, different KV cache types) and it seems to work as expected.

But it is a bit risky if somehow it's internal checks fail to detect an incompatibility, which can lead to invalid benches. So not sure - we can simply the logic to just reuse the state for the repetitions of the same test?


if (is_cached) {
// if previously we have computed at this depth, just restore the state
const size_t ret = llama_state_seq_set_data(ctx, cstate.buf.data(), cstate.buf.size(), 0);
if (ret == 0) {
// if the old state is incompatible with the current context - reprocess from scratch
is_cached = false;
}
}
bool res = test_prompt(ctx, t.n_depth, t.n_batch, t.n_threads);
if (!res) {
fprintf(stderr, "%s: error: failed to run depth\n", __func__);
exit(1);

if (!is_cached) {
if (params.progress) {
fprintf(stderr, "llama-bench: benchmark %d/%zu: depth run %d/%d\n", params_idx, params_count,
i + 1, params.reps);
}
bool res = test_prompt(ctx, t.n_depth, t.n_batch, t.n_threads);
if (!res) {
fprintf(stderr, "%s: error: failed to run depth\n", __func__);
exit(1);
}

// store the context state for reuse in later runs
cstate.depth = t.n_depth;
cstate.buf.resize(llama_state_seq_get_size(ctx, 0));
llama_state_seq_get_data(ctx, cstate.buf.data(), cstate.buf.size(), 0);
} else {
if (params.progress) {
fprintf(stderr, "llama-bench: benchmark %d/%zu: depth run %d/%d (cached)\n", params_idx, params_count,
i + 1, params.reps);
}
}
}

Expand Down
Loading