Revert ipex.optimize_transformers

environments/ipex.yml

@@ -27,8 +27,6 @@ dependencies:
     - --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/
     - torch==2.1.0a0
     - intel-extension-for-pytorch==2.1.10+xpu
-    - oneccl-bind-pt==2.1.100+xpu; sys_platform == 'linux'
-    - deepspeed; sys_platform == 'linux'
     - openvino
     - onnxruntime-openvino
     - flask-cloudflared==0.0.10
@@ -57,8 +55,6 @@ dependencies:
     - einops
     - peft==0.3.0
     - scipy
-    - https://github.com/0cc4m/exllama/releases/download/0.0.7/exllama-0.0.7-cp38-cp38-linux_x86_64.whl; sys_platform == 'linux'
-    - https://github.com/0cc4m/exllama/releases/download/0.0.7/exllama-0.0.7-cp38-cp38-win_amd64.whl; sys_platform == 'win32'
     - windows-curses; sys_platform == 'win32'
     - pynvml
     - omegaconf

modeling/inference_models/hf_torch.py

@@ -423,14 +423,6 @@ 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 = model.to(memory_format=torch.channels_last)
-                        if hasattr(ipex, "optimize_transformers"):
-                            model = ipex.optimize_transformers(model.eval(), dtype=torch.float16, device="xpu", inplace=True)
-                        else:
-                            model = ipex.optimize(model.eval(), dtype=torch.float16, inplace=True)
             except Exception as e:
                 # ...but fall back to stock HF if lazyloader fails.
                 if utils.args.panic:
@@ -447,14 +439,6 @@ 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 = model.to(memory_format=torch.channels_last)
-                    if hasattr(ipex, "optimize_transformers"):
-                        model = ipex.optimize_transformers(model.eval(), dtype=torch.float16, device="xpu", inplace=True)
-                    else:
-                        model = ipex.optimize(model.eval(), dtype=torch.float16, 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: