diff --git a/tests/v1/e2e/general/test_mamba_prefix_cache.py b/tests/v1/e2e/general/test_mamba_prefix_cache.py index 6c208e19fb..cf8aee141b 100644 --- a/tests/v1/e2e/general/test_mamba_prefix_cache.py +++ b/tests/v1/e2e/general/test_mamba_prefix_cache.py @@ -129,6 +129,10 @@ def get_fake_propose_draft_token_ids_fn(): first_token_id_index = ( num_computed_tokens + 1 ) # bonus token isn't considered as computed + # num_accepted_tokens_cpu is filled by a non-blocking D2H copy in the + # same step; wait for it or a lagging GPU yields stale counts. + if self.num_accepted_tokens_event is not None: + self.num_accepted_tokens_event.synchronize() first_token_id_index += self.input_batch.num_accepted_tokens_cpu[0].item() proposed_draft_token_ids = [ prompt_token_ids[