A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[Bug] Fix `is_flashmla_supported` Check Error (#24774)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Sep 15, 2025 at 22:10 UTC
b42566f44039ed8dc5f6c124203d5d71740c6fe0
2 files changed
+4
-26
vllm/attention/backends/flashmla.py
+2
-13
@@ -17,7 +17,6 @@ from vllm.attention.backends.mla.common import (MLACommonBackend,
17
from vllm.attention.ops.flashmla import (flash_mla_with_kvcache,
18
get_mla_metadata,
19
is_flashmla_supported)
20
-from vllm.platforms.cuda import CudaPlatform
20
21
22
class FlashMLABackend(MLACommonBackend):
@@ -179,18 +178,8 @@ class FlashMLAImpl(MLACommonImpl[FlashMLAMetadata]):
178
logits_soft_cap, attn_type,
179
kv_sharing_target_layer_name, **mla_args)
180
182
- assert is_flashmla_supported(), \
183
- "FlashMLA is not supported on this device"
184
-
185
- # disallow FlashMLA on NVIDIA Blackwell (SM 10.0+) GPUs
186
- # context:
187
- # https://github.com/deepseek-ai/FlashMLA/issues/83
188
- # https://github.com/vllm-project/vllm/issues/24513
189
- if CudaPlatform.has_device_capability(100):
190
- raise NotImplementedError(
191
- "FlashMLA is temporarily disabled on Blackwell (SM 10.0). "
192
- "Please use CUTLASS_MLA or TRITON_MLA instead. "
193
- "Example: `export VLLM_ATTENTION_BACKEND=CUTLASS_MLA`")
181
+ is_supported, reason = is_flashmla_supported()
182
+ assert is_supported, reason
183
184
unsupported_features = [alibi_slopes, sliding_window, logits_soft_cap]
185
if any(unsupported_features):
vllm/v1/attention/backends/mla/flashmla.py
+2
-13
@@ -12,7 +12,6 @@ from vllm.attention.ops.flashmla import (flash_mla_with_kvcache,
12
is_flashmla_supported)
13
from vllm.config import VllmConfig
14
from vllm.logger import init_logger
15
-from vllm.platforms.cuda import CudaPlatform
15
from vllm.v1.attention.backends.mla.common import (MLACommonBackend,
16
MLACommonDecodeMetadata,
17
MLACommonImpl,
@@ -156,18 +155,8 @@ class FlashMLAImpl(MLACommonImpl[FlashMLAMetadata]):
155
logits_soft_cap, attn_type,
156
kv_sharing_target_layer_name, **mla_args)
157
159
- assert is_flashmla_supported(), \
160
- "FlashMLA is not supported on this device"
161
-
162
- # disallow FlashMLA on NVIDIA Blackwell (SM 10.0+) GPUs
163
- # context:
164
- # https://github.com/deepseek-ai/FlashMLA/issues/83
165
- # https://github.com/vllm-project/vllm/issues/24513
166
- if CudaPlatform.has_device_capability(100):
167
- raise NotImplementedError(
168
- "FlashMLA is temporarily disabled on Blackwell (SM 10.0). "
169
- "Please use CUTLASS_MLA or TRITON_MLA instead. "
170
- "Example: `export VLLM_ATTENTION_BACKEND=CUTLASS_MLA`")
158
+ is_supported, reason = is_flashmla_supported()
159
+ assert is_supported, reason
160
161
unsupported_features = [alibi_slopes, sliding_window, logits_soft_cap]
162
if any(unsupported_features):