Skip to content

spec : reset ngram-cache state between requests #27866

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

Open
Beatrice0377 wants to merge 2 commits into
base: master
Choose a base branch
from
+40 −2
Open

spec : reset ngram-cache state between requests #27866

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
459ed0a
spec : fix ngram-cache state leaking across requests
Beatrice0377 Aug 28, 2026
8492c2e
tests : move ngram-cache regression test into test-arg-parser
Beatrice0377 Aug 28, 2026
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
8 changes: 6 additions & 2 deletions common/speculative.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2142,8 +2142,12 @@ struct common_speculative_impl_ngram_cache : public common_speculative_impl {
}
}

void begin(llama_seq_id /*seq_id*/, const llama_tokens & /*prompt*/) override {
// noop
void begin(llama_seq_id seq_id, const llama_tokens & /*prompt*/) override {
auto & sinfo = sinfos[seq_id];

// per-request state - do not carry it over to the next request on this sequence
sinfo.ngram_cache_context.clear();
sinfo.cache_size = 0;
}

void draft_one(
Expand Down
34 changes: 34 additions & 0 deletions tests/test-arg-parser.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -101,6 +101,40 @@ static void test(void) {
assert(draft.n_outputs_max_per_seq == 1);
}

// ngram-cache keeps per-sequence state across requests; begin() must reset it
{
common_params_speculative spec;
spec.types = { COMMON_SPECULATIVE_TYPE_NGRAM_CACHE };

common_speculative_ptr drafter(common_speculative_init(spec, 1));
assert(drafter != nullptr);

// request 1: ingest (104, 105, 106) -> 107 into the context cache
llama_tokens prompt1 = { 101, 102, 103, 104, 105, 106 };
common_speculative_begin(drafter.get(), 0, prompt1);

llama_tokens result1;
auto & dp = common_speculative_get_draft_params(drafter.get(), 0);
dp.drafting = true;
dp.prompt = &prompt1;
dp.id_last = 107;
dp.result = &result1;
common_speculative_draft(drafter.get());

// request 2: same seq_id, ending n-gram only exists in request 1's cache, so stale state drafts 107
llama_tokens prompt2 = { 100, 104, 105 };
common_speculative_begin(drafter.get(), 0, prompt2);

llama_tokens result2;
dp.drafting = true;
dp.prompt = &prompt2;
dp.id_last = 106;
dp.result = &result2;
common_speculative_draft(drafter.get());

assert(result2.empty());
}

printf("test-arg-parser: make sure there is no duplicated arguments in any examples\n\n");
for (int ex = 0; ex < LLAMA_EXAMPLE_COUNT; ex++) {
try {
Expand Down