[Misc] Clean up cruft from previous FlashMLA sparse implementation (#26125)

Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
This commit is contained in:
Lucas Wilkinson
2025-10-08 10:09:34 +08:00
committed by GitHub
parent 31a4b3e6c4
commit f80e7866c0
9 changed files with 80 additions and 141 deletions
@@ -165,10 +165,10 @@ def test_env(
pytest.skip("FlashMLA only supports block_size 64")
else:
from vllm.v1.attention.backends.mla.flashmla import (
is_flashmla_supported,
is_flashmla_dense_supported,
)
is_supported, _ = is_flashmla_supported()
is_supported, _ = is_flashmla_dense_supported()
if not is_supported:
pytest.skip("FlashMLA not supported on this platform")
else:
+6 -4
View File
@@ -10,7 +10,7 @@ import torch
from vllm.attention.ops.flashmla import (
flash_mla_with_kvcache,
get_mla_metadata,
is_flashmla_supported,
is_flashmla_dense_supported,
)
from vllm.triton_utils import triton
@@ -27,13 +27,15 @@ def cal_diff(
FLASH_MLA_UNSUPPORTED_REASON = (
is_flashmla_supported()[1]
if not is_flashmla_supported()[0]
is_flashmla_dense_supported()[1]
if not is_flashmla_dense_supported()[0]
else "FlashMLA is supported"
)
@pytest.mark.skipif(not is_flashmla_supported()[0], reason=FLASH_MLA_UNSUPPORTED_REASON)
@pytest.mark.skipif(
not is_flashmla_dense_supported()[0], reason=FLASH_MLA_UNSUPPORTED_REASON
)
@pytest.mark.parametrize("b", [128])
@pytest.mark.parametrize("s_q", [1, 2])
@pytest.mark.parametrize("mean_sk", [4096, 8192, 16384])
@@ -4,19 +4,12 @@ import pytest
import torch
def _cuda_sm90_available() -> bool:
if not torch.cuda.is_available():
return False
major, _ = torch.cuda.get_device_capability()
return major == 9
def test_sparse_flashmla_metadata_smoke():
import vllm.attention.ops.flashmla as fm
ok, reason = fm.is_flashmla_supported()
if not ok or not _cuda_sm90_available():
pytest.skip(reason or "SM90 not available")
ok, reason = fm.is_flashmla_sparse_supported()
if not ok:
pytest.skip(reason)
device = torch.device("cuda")
batch_size = 1
@@ -43,9 +36,9 @@ def test_sparse_flashmla_metadata_smoke():
def test_sparse_flashmla_decode_smoke():
import vllm.attention.ops.flashmla as fm
ok, reason = fm.is_flashmla_supported()
if not ok or not _cuda_sm90_available():
pytest.skip(reason or "SM90 not available")
ok, reason = fm.is_flashmla_sparse_supported()
if not ok:
pytest.skip(reason)
device = torch.device("cuda")
batch_size = 1
@@ -106,9 +99,9 @@ def test_sparse_flashmla_decode_smoke():
def test_sparse_flashmla_prefill_smoke():
import vllm.attention.ops.flashmla as fm
ok, reason = fm.is_flashmla_supported()
if not ok or not _cuda_sm90_available():
pytest.skip(reason or "SM90 not available")
ok, reason = fm.is_flashmla_sparse_supported()
if not ok:
pytest.skip(reason)
device = torch.device("cuda")
s_q = 1