[BugFix] Preserve spec decoding uniform decode when scheduling (#29759)

Signed-off-by: Nick Hill <nhill@redhat.com>
This commit is contained in:
Nick Hill
2025-12-01 17:15:52 -08:00
committed by GitHub
parent 342c4f1472
commit 44822d7ff2
3 changed files with 25 additions and 17 deletions

View File

@@ -191,8 +191,8 @@ def test_suffix_decoding_acceptance(
# Expect the acceptance rate to improve.
assert first_accept_rate < last_accept_rate
# Heuristic: expect at least 85% acceptance rate at the end.
assert last_accept_rate > 0.85
# Heuristic: expect at least 82.5% acceptance rate at the end.
assert last_accept_rate > 0.825
del spec_llm
torch.cuda.empty_cache()