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
[Feature] Disallow FlashMLA on Blackwell (#24521)
Signed-off-by: yewentao256 <zhyanwentao@126.com> Signed-off-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com> Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Wentao Ye committed
Sep 9, 2025 at 14:59 UTC
15de5ff9ea30950212b724818c716084528a2f7b
2 files changed
+22
vllm/attention/backends/flashmla.py
+11
@@ -17,6 +17,7 @@ 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
21
22
23
class FlashMLABackend(MLACommonBackend):
@@ -181,6 +182,16 @@ class FlashMLAImpl(MLACommonImpl[FlashMLAMetadata]):
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`")
194
+
195
unsupported_features = [alibi_slopes, sliding_window, logits_soft_cap]
196
if any(unsupported_features):
197
raise NotImplementedError(
vllm/v1/attention/backends/mla/flashmla.py
+11
@@ -12,6 +12,7 @@ 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
16
from vllm.v1.attention.backends.mla.common import (MLACommonBackend,
17
MLACommonDecodeMetadata,
18
MLACommonImpl,
@@ -158,6 +159,16 @@ class FlashMLAImpl(MLACommonImpl[FlashMLAMetadata]):
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`")
171
+
172
unsupported_features = [alibi_slopes, sliding_window, logits_soft_cap]
173
if any(unsupported_features):
174
raise NotImplementedError(