[CI] Exercise FA3 FP8 attention on SM90 (#49231)

Signed-off-by: Simon Mo <simon@inferact.ai>
This commit is contained in:
Simon Mo
2026-07-21 10:26:55 +08:00
committed by GitHub
parent e78a0c8e59
commit 72d16aee15
+6 -3
View File
@@ -93,9 +93,6 @@ def test_online_quantization(
use_rocm_aiter: bool,
monkeypatch,
) -> None:
if kv_cache_dtype == "fp8" and current_platform.is_device_capability_family(90):
pytest.skip("FA3 currently rejects FP8 KV cache output dtype on SM90")
if use_rocm_aiter:
monkeypatch.setenv("VLLM_ROCM_USE_AITER", "1")
@@ -105,9 +102,15 @@ def test_online_quantization(
if force_marlin:
monkeypatch.setenv("VLLM_TEST_FORCE_FP8_MARLIN", "1")
model_dtype = "auto"
if kv_cache_dtype == "fp8" and current_platform.is_device_capability_family(90):
# FA3 requires BF16 output when the query input is FP8.
model_dtype = "bfloat16"
with vllm_runner(
"facebook/opt-125m",
quantization="fp8",
dtype=model_dtype,
enforce_eager=True,
kv_cache_dtype=kv_cache_dtype,
) as llm: