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 8, 21:30 UTC
[Bugfix] fix DeepSeek R1 with CUTLASS MLA Broken on B200 (#33637)
Signed-off-by: chaunceyjiang <chaunceyjiang@gmail.com>
Chauncey committed
Feb 5, 2026 at 09:28 UTC
a7be77beef5f59d9d349818b4f2860483551b255
1 file changed
+1
-4
vllm/model_executor/layers/attention/mla_attention.py
+1
-4
@@ -293,7 +293,6 @@ class MLAAttention(nn.Module, AttentionLayerBase):
293
prefix: str = "",
294
use_sparse: bool = False,
295
indexer: object | None = None,
296
- q_pad_num_heads: int | None = None,
296
**extra_impl_args,
297
):
298
super().__init__()
@@ -308,7 +307,6 @@ class MLAAttention(nn.Module, AttentionLayerBase):
307
self.head_size = kv_lora_rank + qk_rope_head_dim
308
self.layer_name = prefix
309
self.indexer = indexer
311
- self.q_pad_num_heads = q_pad_num_heads
310
311
self.num_kv_heads = 1
312
self.qk_head_dim = self.qk_nope_head_dim + self.qk_rope_head_dim
@@ -375,10 +373,9 @@ class MLAAttention(nn.Module, AttentionLayerBase):
373
v_head_dim=self.v_head_dim,
374
kv_b_proj=kv_b_proj,
375
indexer=indexer,
378
- q_pad_num_heads=q_pad_num_heads,
376
**extra_impl_args,
377
)
381
-
378
+ self.q_pad_num_heads = getattr(self.impl, "q_pad_num_heads", None)
379
self.use_direct_call = not current_platform.opaque_attention_op()
380
381
compilation_config = get_current_vllm_config().compilation_config