@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 6ef09b08f8
Fork 0

A high-throughput and memory-efficient inference and serving engine for LLMs

Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git

Synced Nov 17, 01:45 UTC

[Core][Distributed] fix pynccl del error (#4508)

youkaichao committed May 1, 2024 at 15:23 UTC 6ef09b08f88b675f84b7140238286e5d4c5304c8
1 file changed +4 -8
vllm/distributed/device_communicators/pynccl.py
+4 -8
@@ -200,6 +200,10 @@ _c_ncclAllReduce.argtypes = [
200 ncclDataType_t, ctypes.c_void_p, ctypes.c_void_p
201 ]
202
203 +# be cautious! this is a collective call, it will block until all
204 +# processes in the communicator have called this function.
205 +# because Python object destruction can happen in random order,
206 +# it is better not to call it at all.
207 # equivalent to c declaration:
208 # ncclResult_t ncclCommDestroy(ncclComm_t comm);
209 _c_ncclCommDestroy = nccl.ncclCommDestroy
@@ -278,11 +282,3 @@ class NCCLCommunicator:
282 ncclDataTypeEnum.from_torch(tensor.dtype),
283 ncclRedOpTypeEnum.from_torch(op), self.comm,
284 ctypes.c_void_p(stream.cuda_stream)))
281 -
282 - def __del__(self):
283 - # `dist` module might have been already destroyed
284 - if hasattr(dist, 'destroy_process_group'):
285 - dist.destroy_process_group()
286 - # function might have been already destroyed
287 - if _c_ncclCommDestroy is not None:
288 - _c_ncclCommDestroy(self.comm)