Fix Regression: Disable p2p for 4090 (#531)

Co-authored-by: Qubitium <417764+Qubitium@users.noreply.github.com>

python/sglang/srt/managers/controller/model_runner.py

@@ -241,7 +241,7 @@ class ModelRunner:
         logger.info(f"[gpu_id={self.gpu_id}] Set cuda device.")
         torch.cuda.set_device(self.gpu_id)
         logger.info(f"[gpu_id={self.gpu_id}] Init nccl begin.")
-        monkey_patch_vllm_p2p_access_check()
+        monkey_patch_vllm_p2p_access_check(self.gpu_id)
         init_distributed_environment(
             backend="nccl",
             world_size=self.tp_size,

python/sglang/srt/utils.py

@@ -453,13 +453,17 @@ def kill_parent_process():
     os.kill(parent_process.pid, 9)
 
 
-def monkey_patch_vllm_p2p_access_check():
+def monkey_patch_vllm_p2p_access_check(gpu_id: int):
     """
     Monkey patch the slow p2p access check in vllm.
     NOTE: We assume the p2p access is always allowed, which can be wrong for some setups.
     """
-    import vllm.distributed.device_communicators.custom_all_reduce_utils as tgt
 
+    # TODO: need a better check than just dev str name match
+    # compat: skip RTX 40 series as they do not have P2P feature and even checking for them may cause errors
+    device_name = torch.cuda.get_device_name(gpu_id)
+    if "RTX 40" not in device_name:
+        import vllm.distributed.device_communicators.custom_all_reduce_utils as tgt
         setattr(tgt, "gpu_p2p_access_check", lambda *arg, **kwargs: True)