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

Signed-off-by: Louis Ulmer <[email protected]>

Author: Akashcodes732

vllm/distributed/parallel_state.py

@@ -329,9 +329,17 @@ def all_reduce(self, input_: torch.Tensor) -> torch.Tensor:
             return input_
 
         if input_.is_cpu:
-            import intel_extension_for_pytorch as ipex
-            ipex.distributed.all_reduce(input_, group=self.device_group)
-            return input_
+            try:
+                import intel_extension_for_pytorch as ipex
+                ipex.distributed.all_reduce(input_, group=self.device_group)
+                return input_
+            except ImportError:
+                """
+                Intel IPEX not found. Falling back to PyTorch native 
+                all_reduce for CPU
+                """
+                torch.distributed.all_reduce(input_, group=self.device_group)
+                return input_
 
         if self.tpu_communicator is not None and \
             not self.tpu_communicator.disabled: