[Bugfix] Bound GDN speculative decode query widths (#55504)

Signed-off-by: DimensionSTP <65501090+DimensionSTP@users.noreply.github.com>
Co-authored-by: OpenAI Codex <noreply@openai.com>
This commit is contained in:
DimensionSTP
2026-10-10 11:00:16 +08:00
committed by GitHub
co-authored by OpenAI Codex
parent 02aa10539f
commit 10cc2f6ae2
2 changed files with 19 additions and 1 deletions
@@ -314,6 +314,17 @@ def test_padded_rows_keep_the_sentinel(monkeypatch):
assert metadata.num_decode_draft_tokens_cpu.tolist() == [0, 0, -1, -1]
def test_rows_wider_than_the_speculative_state_window_keep_the_sentinel(
monkeypatch: pytest.MonkeyPatch,
) -> None:
state = _mamba_hybrid_state(num_speculative_tokens=7)
input_batch = _input_batch([12, 8, 1], None, [False, False, False])
metadata = _prepare_attn_metadata(state, monkeypatch, input_batch)
assert metadata.num_decode_draft_tokens_cpu.tolist() == [-1, 0, 0]
def test_chunked_prefill_tail_of_two_or_three_tokens_keeps_the_sentinel(
monkeypatch,
):