biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Bugfix][Kernel] Fix CUDA 11.8 being broken by FA3 build (#12375)

Browse Source 
Signed-off-by: Lucas Wilkinson <lwilkinson@neuralmagic.com>
This commit is contained in:
Lucas Wilkinson
2025-01-24 10:27:59 -05:00
committed by GitHub
parent 3bb8e2c9a2
commit ab5bbf5ae3
6 changed files with 42 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
tests/kernels/test_cascade_flash_attn.py Normal file → Executable file
View File

@@ -6,7 +6,9 @@ import torch
from vllm.platforms import current_platform
from vllm.v1.attention.backends.flash_attn import (cascade_attention,
merge_attn_states)
from vllm.vllm_flash_attn import flash_attn_varlen_func
from vllm.vllm_flash_attn import (fa_version_unsupported_reason,
flash_attn_varlen_func,
is_fa_version_supported)
NUM_HEADS = [(4, 4), (8, 2), (16, 2)]
HEAD_SIZES = [128, 192, 256]
@@ -91,10 +93,9 @@ def test_cascade(
fa_version: int,
) -> None:
torch.set_default_device("cuda")
if fa_version == 3 and (torch.cuda.get_device_capability() == (8, 6)
or torch.cuda.get_device_capability() == (8, 9)):
pytest.skip("Flash attention version 3 fails on 8.6 and 8.9 due to "
"insufficient shared memory for some shapes")
if not is_fa_version_supported(fa_version):
pytest.skip(f"Flash attention version {fa_version} not supported due "
f"to: \"{fa_version_unsupported_reason(fa_version)}\"")
current_platform.seed_everything(0)