Bug/test eagle dp v0 (#38938)

Signed-off-by: Monishver Chandrasekaran <monishverchandrasekaran@gmail.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Co-authored-by: Nicolò Lucchesi <nlucches@redhat.com>
This commit is contained in:
Monishver
2026-04-13 20:50:08 +00:00
committed by GitHub
co-authored by mergify[bot] <37929162+mergify[bot]@users.noreply.github.com> Nicolò Lucchesi
parent 3693f922ff
commit 8213e8f880
3 changed files with 13 additions and 8 deletions
+6 -8
View File
@@ -20,16 +20,14 @@ if current_platform.is_rocm():
else:
ATTN_BACKENDS = ["FLASH_ATTN"]
# On SM<90 (e.g., L4), batch invariance does not support CUDA graphs.
# See https://github.com/vllm-project/vllm/pull/30018 and
# tests/v1/determinism/utils.py for the documented limitation.
IS_DEVICE_CAPABILITY_BELOW_90 = not current_platform.has_device_capability(90)
@pytest.mark.asyncio
@pytest.mark.parametrize("attn_backend", ATTN_BACKENDS)
@pytest.mark.xfail(
not current_platform.is_rocm(),
reason="EAGLE + DP > 1 produces wrong outputs when async spec decode "
"correction is active. Root cause under investigation. "
"See: https://github.com/vllm-project/vllm/issues/31913",
strict=False,
)
@pytest.mark.xfail(
current_platform.is_rocm(),
reason="Test may fail on ROCm until batch invariance is enabled. "
@@ -57,7 +55,7 @@ async def test_run_eagle_dp(monkeypatch: pytest.MonkeyPatch, attn_backend: str):
engine_args = AsyncEngineArgs(
model=target_model,
tokenizer_mode="auto",
enforce_eager=False,
enforce_eager=IS_DEVICE_CAPABILITY_BELOW_90,
tensor_parallel_size=int(os.getenv("TP_SIZE", 1)),
data_parallel_size=DP_SIZE,
data_parallel_backend="mp", # ray takes more time