Roll back to use vllm custom allreduce (#3006)

python/sglang/srt/distributed/__init__.py

@@ -1,3 +1,3 @@
-from .communication_op import *
-from .parallel_state import *
-from .utils import *
+from sglang.srt.distributed.communication_op import *
+from sglang.srt.distributed.parallel_state import *
+from sglang.srt.distributed.utils import *

python/sglang/srt/distributed/communication_op.py

@@ -4,7 +4,7 @@ from typing import Any, Dict, Optional, Union
 import torch
 import torch.distributed
 
-from .parallel_state import get_tp_group
+from sglang.srt.distributed.parallel_state import get_tp_group
 
 
 def tensor_model_parallel_all_reduce(input_: torch.Tensor) -> torch.Tensor:

python/sglang/srt/distributed/device_communicators/custom_all_reduce_utils.py

@@ -7,7 +7,6 @@ import pickle
 import subprocess
 import sys
 import tempfile
-from functools import lru_cache
 from itertools import product
 from typing import Dict, List, Optional, Sequence
 

python/sglang/srt/distributed/device_communicators/pynccl_wrapper.py

@@ -57,7 +57,7 @@ def find_nccl_library() -> str:
             so_file = "librccl.so.1"
         else:
             raise ValueError("NCCL only supports CUDA and ROCm backends.")
-        logger.info("Found nccl from library %s", so_file)
+        logger.debug("Found nccl from library %s", so_file)
     return so_file
 
 

python/sglang/srt/distributed/device_communicators/shm_broadcast.py

@@ -313,7 +313,7 @@ class MessageQueue:
             remote_subscribe_port=remote_subscribe_port,
         )
 
-        logger.info("vLLM message queue communication handle: %s", self.handle)
+        logger.debug("Message queue communication handle: %s", self.handle)
 
     def export_handle(self) -> Handle:
         return self.handle