xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-09 22:25:32 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bug] Fix is_flashmla_supported Check Error (#24774)

Browse Source 
Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
Wentao Ye 2025-09-15 22:10:55 -04:00 committed by GitHub
parent d96e11167d
commit b42566f440
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 4 additions and 26 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
vllm/attention/backends/flashmla.py
View File

@ -17,7 +17,6 @@ from vllm.attention.backends.mla.common import (MLACommonBackend,
from vllm.attention.ops.flashmla import (flash_mla_with_kvcache,
get_mla_metadata,
is_flashmla_supported)
from vllm.platforms.cuda import CudaPlatform
class FlashMLABackend(MLACommonBackend):
@ -179,18 +178,8 @@ class FlashMLAImpl(MLACommonImpl[FlashMLAMetadata]):
logits_soft_cap, attn_type,
kv_sharing_target_layer_name, **mla_args)
assert is_flashmla_supported(), \
"FlashMLA is not supported on this device"
# disallow FlashMLA on NVIDIA Blackwell (SM 10.0+) GPUs
# context:
# https://github.com/deepseek-ai/FlashMLA/issues/83
# https://github.com/vllm-project/vllm/issues/24513
if CudaPlatform.has_device_capability(100):
raise NotImplementedError(
"FlashMLA is temporarily disabled on Blackwell (SM 10.0). "
"Please use CUTLASS_MLA or TRITON_MLA instead. "
"Example: `export VLLM_ATTENTION_BACKEND=CUTLASS_MLA`")
is_supported, reason = is_flashmla_supported()
assert is_supported, reason
unsupported_features = [alibi_slopes, sliding_window, logits_soft_cap]
if any(unsupported_features):

15
vllm/v1/attention/backends/mla/flashmla.py
View File

@ -12,7 +12,6 @@ from vllm.attention.ops.flashmla import (flash_mla_with_kvcache,
is_flashmla_supported)
from vllm.config import VllmConfig
from vllm.logger import init_logger
from vllm.platforms.cuda import CudaPlatform
from vllm.v1.attention.backends.mla.common import (MLACommonBackend,
MLACommonDecodeMetadata,
MLACommonImpl,
@ -156,18 +155,8 @@ class FlashMLAImpl(MLACommonImpl[FlashMLAMetadata]):
logits_soft_cap, attn_type,
kv_sharing_target_layer_name, **mla_args)
assert is_flashmla_supported(), \
"FlashMLA is not supported on this device"
# disallow FlashMLA on NVIDIA Blackwell (SM 10.0+) GPUs
# context:
# https://github.com/deepseek-ai/FlashMLA/issues/83
# https://github.com/vllm-project/vllm/issues/24513
if CudaPlatform.has_device_capability(100):
raise NotImplementedError(
"FlashMLA is temporarily disabled on Blackwell (SM 10.0). "
"Please use CUTLASS_MLA or TRITON_MLA instead. "
"Example: `export VLLM_ATTENTION_BACKEND=CUTLASS_MLA`")
is_supported, reason = is_flashmla_supported()
assert is_supported, reason
unsupported_features = [alibi_slopes, sliding_window, logits_soft_cap]
if any(unsupported_features):