merge dev主干代码

vllm/v1/worker/gpu_model_runner.py

@@ -829,8 +829,12 @@ class GPUModelRunner(
                 non_blocking=True,
             )
             return
+        # self.mrope_positions.gpu[:, :num_tokens].copy_(
+        #     self.mrope_positions.cpu[:, :num_tokens],
+        #     non_blocking=True,
+        # )
         self.mrope_positions.gpu[:, :num_tokens].copy_(
-            self.mrope_positions.cpu[:, :num_tokens],
+            self.mrope_positions.cpu[:, :num_tokens].contiguous().pin_memory(),
             non_blocking=True,
         )
 
@@ -6286,7 +6290,7 @@ class GPUModelRunner(
         return kv_caches
 
     def _update_hybrid_attention_mamba_layout(
-        self, kv_caches: dict[str, Any]
+        self, kv_caches: dict[str, torch.Tensor]
     ) -> None:
         """
         Update the layout of attention layers from (2, num_blocks, ...) to
@@ -6300,8 +6304,6 @@ class GPUModelRunner(
             kv_cache_spec = group.kv_cache_spec
             for layer_name in group.layer_names:
                 kv_cache = kv_caches[layer_name]
-                if not isinstance(kv_cache, torch.Tensor):
-                    continue
                 if isinstance(kv_cache_spec, AttentionSpec) and kv_cache.shape[0] == 2:
                     assert kv_cache.shape[1] != 2, (
                         "Fail to determine whether the layout is "