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

[perf]消除sparse mla build时的拷贝调度空泡

See merge request dcutoolkit/deeplearing/vllm!510

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

@@ -456,6 +456,12 @@ class FlashMLASparseMetadataBuilder(AttentionMetadataBuilder[FlashMLASparseMetad
             dtype=torch.int32,
             device=device,
         )
+        self.req_id_per_token_buffer_cpu = torch.zeros((vllm_config.scheduler_config.max_num_batched_tokens,), 
+                                                       dtype=torch.int32, 
+                                                       device="cpu", 
+                                                       pin_memory=True)
+        self.req_id_per_token_buffer_np = self.req_id_per_token_buffer_cpu.numpy()
+
 
     def _build_fp8_mixed_decode_prefill(
         self,
@@ -651,9 +657,10 @@ class FlashMLASparseMetadataBuilder(AttentionMetadataBuilder[FlashMLASparseMetad
         )
         # Zero-fill for cudagraphs
         self.req_id_per_token_buffer.fill_(0)
+        self.req_id_per_token_buffer_np[: req_id_per_token.shape[0]] = req_id_per_token
         self.req_id_per_token_buffer[: req_id_per_token.shape[0]].copy_(
-            torch.from_numpy(req_id_per_token), non_blocking=True
-        )
+            self.req_id_per_token_buffer_cpu[: req_id_per_token.shape[0]], non_blocking=True)
+
         req_id_per_token = self.req_id_per_token_buffer[:num_tokens]
 
         fp8_extra_metadata: (