[CI] Exercise FA3 FP8 attention on SM90 (#49231)
Signed-off-by: Simon Mo <simon@inferact.ai>
Simon Mo committed
Jul 20, 2026 at 19:26 UTC
72d16aee1576283444352583f36ae65b8ff58455
1 file changed
+6
-3
tests/quantization/test_fp8.py
+6
-3
@@ -93,9 +93,6 @@ def test_online_quantization(
93
use_rocm_aiter: bool,
94
monkeypatch,
95
) -> None:
96
- if kv_cache_dtype == "fp8" and current_platform.is_device_capability_family(90):
97
- pytest.skip("FA3 currently rejects FP8 KV cache output dtype on SM90")
98
-
96
if use_rocm_aiter:
97
monkeypatch.setenv("VLLM_ROCM_USE_AITER", "1")
98
@@ -105,9 +102,15 @@ def test_online_quantization(
102
if force_marlin:
103
monkeypatch.setenv("VLLM_TEST_FORCE_FP8_MARLIN", "1")
104
105
+ model_dtype = "auto"
106
+ if kv_cache_dtype == "fp8" and current_platform.is_device_capability_family(90):
107
+ # FA3 requires BF16 output when the query input is FP8.
108
+ model_dtype = "bfloat16"
109
+
110
with vllm_runner(
111
"facebook/opt-125m",
112
quantization="fp8",
113
+ dtype=model_dtype,
114
enforce_eager=True,
115
kv_cache_dtype=kv_cache_dtype,
116
) as llm: