[Attention] FlashAttention 4 SM100 FP8 kv cache support (#42569)

Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Matthew Bonanni
2026-07-20 10:53:27 -07:00
committed by GitHub
co-authored by mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent b23bd73f54
commit bd091079cb
8 changed files with 194 additions and 33 deletions
+2 -8
View File
@@ -9,8 +9,8 @@ Note: these tests will only pass on L4 GPU.
import pytest
from tests.quantization.utils import is_quant_method_supported
from vllm.v1.attention.backends.fa_utils import flash_attn_supports_kv_cache_dtype
from vllm.platforms import current_platform
from vllm.v1.attention.backends.fa_utils import get_flash_attn_version
from ..utils import check_logprobs_close
@@ -70,13 +70,7 @@ def test_models(
if kv_cache_dtype == "fp8_e5m2" and current_platform.is_cuda():
pytest.skip(f"{kv_cache_dtype} is not supported by FLASH_ATTN on CUDA.")
if not (
current_platform.is_xpu()
or (
get_flash_attn_version() == 3
and current_platform.is_device_capability_family(90)
)
):
if not flash_attn_supports_kv_cache_dtype(kv_cache_dtype):
pytest.skip(
f"{kv_cache_dtype} is not supported on this GPU type with {backend} attention."
)