Merge branch 'v0.15.1-dev_dsa_mqa' into 'v0.15.1-dev'

feat:修复dsa的mqa接口兼容glm5

See merge request dcutoolkit/deeplearing/vllm!478

vllm/model_executor/layers/sparse_attn_indexer.py

@@ -121,9 +121,10 @@ def sparse_attn_indexer(
                     chunk.cu_seqlen_ke,
                     q_fp8[chunk.token_start:chunk.token_end].shape[0],
                     k.shape[0],
-                    64,
-                    128,
-                    True,
+                    q_fp8.shape[1],
+                    q_fp8.shape[2],
+                    None,
+                    True
                     )
             num_rows = logits.shape[0]