IPEX add ipex.optimize_transformers and DeepSpeed support

environments/ipex.yml

@@ -29,6 +29,10 @@ dependencies:
     - intel-extension-for-pytorch==2.1.10+xpu
     - mkl==2024.0.0
     - mkl-dpcpp==2024.0.0
+    - oneccl-bind-pt==2.1.100+xpu; sys_platform == 'linux'
+    - impi-devel==2021.11.0; sys_platform == 'linux'
+    - oneccl-devel==2021.11.1; sys_platform == 'linux'
+    - deepspeed; sys_platform == 'linux'
     - openvino
     - onnxruntime-openvino
     - flask-cloudflared==0.0.10

modeling/inference_models/hf_torch.py

@@ -423,6 +423,10 @@ class HFTorchInferenceModel(HFInferenceModel):
                         torch_dtype=self._get_target_dtype(),
                         **tf_kwargs,
                     )
+
+                    if hasattr(torch, "xpu") and torch.xpu.is_available and os.environ.get('DISABLE_IPEX_OPTIMIZE', None) is None:
+                        import intel_extension_for_pytorch as ipex
+                        model = ipex.optimize_transformers(model.eval(), dtype=torch.float16, device="xpu", inplace=True)
             except Exception as e:
                 # ...but fall back to stock HF if lazyloader fails.
                 if utils.args.panic:
@@ -439,6 +443,10 @@ class HFTorchInferenceModel(HFInferenceModel):
                     **tf_kwargs,
                 )
 
+                if hasattr(torch, "xpu") and torch.xpu.is_available and os.environ.get('DISABLE_IPEX_OPTIMIZE', None) is None:
+                    import intel_extension_for_pytorch as ipex
+                    model = ipex.optimize_transformers(model.eval(), dtype=torch.float16, device="xpu", inplace=True)
+
             if not self.lazy_load and not self.breakmodel:
                 # We need to move the model to the desired device
                 if (not self.usegpu) or torch.cuda.device_count() <= 0: