Merge branch 'v0.9.2-dev-wm' into 'v0.9.2-dev'

[fix]避免mla中cudagraph的适配影响非并行解码的逻辑

See merge request dcutoolkit/deeplearing/vllm!165

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

@@ -690,7 +690,8 @@ class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]):
 
     def can_run_in_cudagraph(
             self, common_attn_metadata: CommonAttentionMetadata) -> bool:
-        #return common_attn_metadata.max_query_len == 1
+        if not self.use_spec_decode:
+            return common_attn_metadata.max_query_len == 1
         return self._num_prefills == 0