update common.py

vllm/attention/backends/mla/common.py

@@ -1031,7 +1031,6 @@ class MLACommonImpl(MLAAttentionImpl[T], Generic[T]):
         # and the one from vllm_flash_attn. The former is used on RoCM and the
         # latter has an additional parameter to control FA2 vs FA3
         self.flash_attn_varlen_func = flash_attn_varlen_func
-        # self.vllm_flash_attn_version = None
         self.vllm_flash_attn_version = get_flash_attn_version()
         if self.vllm_flash_attn_version is not None:
             self.flash_attn_varlen_func = \

vllm/v1/attention/backends/mla/common.py

@@ -656,7 +656,7 @@ class MLACommonImpl(MLAAttentionImpl[M], Generic[M]):
             q=q,
             k=k,
             v=maybe_padded_v,
-            return_softmax_lse=return_softmax_lse,
+            # return_softmax_lse=return_softmax_lse,
             softmax_scale=softmax_scale,
             **kwargs,
         )