From a7be77beef5f59d9d349818b4f2860483551b255 Mon Sep 17 00:00:00 2001 From: Chauncey Date: Thu, 5 Feb 2026 09:28:36 +0800 Subject: [PATCH] [Bugfix] fix DeepSeek R1 with CUTLASS MLA Broken on B200 (#33637) Signed-off-by: chaunceyjiang --- vllm/model_executor/layers/attention/mla_attention.py | 5 +---- 1 file changed, 1 insertion(+), 4 deletions(-) diff --git a/vllm/model_executor/layers/attention/mla_attention.py b/vllm/model_executor/layers/attention/mla_attention.py index 1b719330e2..febad38216 100644 --- a/vllm/model_executor/layers/attention/mla_attention.py +++ b/vllm/model_executor/layers/attention/mla_attention.py @@ -293,7 +293,6 @@ class MLAAttention(nn.Module, AttentionLayerBase): prefix: str = "", use_sparse: bool = False, indexer: object | None = None, - q_pad_num_heads: int | None = None, **extra_impl_args, ): super().__init__() @@ -308,7 +307,6 @@ class MLAAttention(nn.Module, AttentionLayerBase): self.head_size = kv_lora_rank + qk_rope_head_dim self.layer_name = prefix self.indexer = indexer - self.q_pad_num_heads = q_pad_num_heads self.num_kv_heads = 1 self.qk_head_dim = self.qk_nope_head_dim + self.qk_rope_head_dim @@ -375,10 +373,9 @@ class MLAAttention(nn.Module, AttentionLayerBase): v_head_dim=self.v_head_dim, kv_b_proj=kv_b_proj, indexer=indexer, - q_pad_num_heads=q_pad_num_heads, **extra_impl_args, ) - + self.q_pad_num_heads = getattr(self.impl, "q_pad_num_heads", None) self.use_direct_call = not current_platform.opaque_attention_op() compilation_config = get_current_vllm_config().compilation_config