bugfix in triton flash attn func in MLA backend: the function takes no keyword arguments.

Signed-off-by: vllmellm <vllm.ellm@embeddedllm.com>

Author: vllmellm

vllm/attention/backends/mla/common.py

@@ -1076,7 +1076,14 @@ def _flash_attn_varlen_diff_headdims(self, q, k, v, softmax_scale,
                 q,
                 k,
                 maybe_padded_v,
-                **kwargs,
+                None,  # output
+                kwargs["cu_seqlens_q"],
+                kwargs["cu_seqlens_k"],
+                kwargs["max_seqlen_q"],
+                kwargs["max_seqlen_k"],
+                kwargs["causal"],
+                softmax_scale,
+                None,  # bias
             )
         if is_vllm_fa:
             attn_out = self.flash_attn_varlen_func(