A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 5, 18:30 UTC
[WIP][CI][Bugfix] Fix `test_run_eagle_dp` (#38584)
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Matthew Bonanni committed
Mar 31, 2026 at 06:30 UTC
7d65463528a05260167d0f37997ecdc1ef0ae8b2
2 files changed
+6
-5
tests/v1/distributed/test_eagle_dp.py
+1
-3
@@ -69,9 +69,7 @@ async def test_run_eagle_dp(monkeypatch: pytest.MonkeyPatch, attn_backend: str):
69
)
70
71
prompt = "This is a test of data parallel with eagle"
72
- # This test might be flaky, see
73
- # https://github.com/vllm-project/vllm/issues/31913
74
- num_expected_tokens = 20
72
+ num_expected_tokens = 100
73
sampling_params = SamplingParams(
74
max_tokens=num_expected_tokens,
75
ignore_eos=True,
vllm/v1/attention/backends/flash_attn.py
+5
-2
@@ -389,8 +389,11 @@ class FlashAttentionMetadataBuilder(AttentionMetadataBuilder[FlashAttentionMetad
389
slot_mapping = common_attn_metadata.slot_mapping
390
causal = common_attn_metadata.causal
391
392
- # the overhead of the aot schedule is not worth it for spec-decode
393
- aot_schedule = self.aot_schedule and not fast_build
392
+ # Disable AOT schedule for spec-decode proposer (not worth the overhead)
393
+ # and for batch invariance (schedule varies with max_seqlen_q/k).
394
+ aot_schedule = (
395
+ self.aot_schedule and not fast_build and not envs.VLLM_BATCH_INVARIANT
396
+ )
397
398
if self.aot_sliding_window is None:
399
self.aot_sliding_window = (-1, -1)