@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 022bcc701a
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 16, 00:45 UTC

[Bugfix] Fix 'ModuleNotFoundError: No module named 'intel_extension_for_pytorch'' for --tensor-parallel-size more than 1 (#12546)

Akash kaothalkar committed Feb 5, 2025 at 12:41 UTC 022bcc701a948f96e68af678eee686837f393d07
1 file changed +11 -3
vllm/distributed/parallel_state.py
+11 -3
@@ -329,9 +329,17 @@ class GroupCoordinator:
329 return input_
330
331 if input_.is_cpu:
332 - import intel_extension_for_pytorch as ipex
333 - ipex.distributed.all_reduce(input_, group=self.device_group)
334 - return input_
332 + try:
333 + import intel_extension_for_pytorch as ipex
334 + ipex.distributed.all_reduce(input_, group=self.device_group)
335 + return input_
336 + except ImportError:
337 + """
338 + Intel IPEX not found. Falling back to PyTorch native
339 + all_reduce for CPU
340 + """
341 + torch.distributed.all_reduce(input_, group=self.device_group)
342 + return input_
343
344 if self.tpu_communicator is not None and \
345 not self.tpu_communicator.disabled: