区分pcie和hglink custom allreduce的使用

vllm/distributed/device_communicators/custom_all_reduce.py

@@ -264,10 +264,13 @@ class CustomAllreduce:
             return None
         if self._IS_CAPTURING:
             if torch.cuda.is_current_stream_capturing():
-                if not envs.VLLM_CUSTOM_CACHE:
-                    return self.all_reduce(input, registered=False)
-                else:
+                if envs.VLLM_CUSTOM_CACHE:
                     return self.all_reduce(input, registered=True)
+                else:
+                    if not self.fully_connected:
+                        return self.all_reduce(input, registered=False)
+                    else:
+                        return self.all_reduce(input, registered=True)
             else:
                 # If warm up, mimic the allocation pattern since custom
                 # allreduce is out-of-place.

vllm/envs.py

@@ -1080,7 +1080,7 @@ environment_variables: dict[str, Callable[[], Any]] = {
 
     # flag to control vllm to use optimized kernels
     "VLLM_CUSTOM_CACHE":
-    lambda: bool(int(os.environ.get("VLLM_CUSTOM_CACHE", "1"))),
+    lambda: bool(int(os.environ.get("VLLM_CUSTOM_CACHE", "0"))),
     
     # flag to control vllm to use optimized kernels
     "VLLM_CUSTOM_ALLREDUCE_SUPPORTED_WORLDSIZE_MAX":