Add `--enable-p2p-check` option (#599)

README.md

@@ -362,7 +362,7 @@ python -m sglang.launch_server --model-path meta-llama/Llama-2-7b-chat-hf --port
 ```
 
 ### Additional Arguments
-- Add `--tp 2` to enable tensor parallelism.
+- Add `--tp 2` to enable tensor parallelism. If it indicates `peer access is not supported between these two devices`, add `--enable-p2p-check` option.
 ```
 python -m sglang.launch_server --model-path meta-llama/Llama-2-7b-chat-hf --port 30000 --tp 2
 ```

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

@@ -259,7 +259,10 @@ 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.")
+
+        if not server_args.enable_p2p_check:
             monkey_patch_vllm_p2p_access_check(self.gpu_id)
+
         if server_args.nccl_init_addr:
             nccl_init_method = f"tcp://{server_args.nccl_init_addr}"
         else:

python/sglang/srt/server_args.py

@@ -55,6 +55,7 @@ class ServerArgs:
     disable_regex_jump_forward: bool = False
     disable_disk_cache: bool = False
     attention_reduce_in_fp32: bool = False
+    enable_p2p_check: bool = False
 
     # Distributed args
     nccl_init_addr: Optional[str] = None
@@ -304,6 +305,11 @@ class ServerArgs:
             help="Cast the intermidiate attention results to fp32 to avoid possible crashes related to fp16."
             "This only affects Triton attention kernels",
         )
+        parser.add_argument(
+            "--enable-p2p-check",
+            action="store_true",
+            help="Enable P2P check for GPU access, otherwise the p2p access is allowed by default.",
+        )
 
     @classmethod
     def from_cli_args(cls, args: argparse.Namespace):

python/sglang/srt/utils.py

@@ -458,12 +458,7 @@ def monkey_patch_vllm_p2p_access_check(gpu_id: int):
     NOTE: We assume the p2p access is always allowed, which can be wrong for some setups.
     """
 
-    # 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)