Revert bundle-in MKL and DPCPP

environments/ipex.yml

@@ -27,11 +27,7 @@ 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
-    - 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

modeling/inference_models/hf_torch.py

@@ -426,7 +426,11 @@ class HFTorchInferenceModel(HFInferenceModel):
 
                     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)
+                        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:
@@ -445,7 +449,11 @@ class HFTorchInferenceModel(HFInferenceModel):
 
                 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)
+                    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

play-ipex.sh

@@ -4,7 +4,17 @@ if [ ! -f "runtime/envs/koboldai-ipex/bin/python" ]; then
 ./install_requirements.sh ipex
 fi
 
-export LD_LIBRARY_PATH=$(realpath "runtime/envs/koboldai-ipex")/lib/:$LD_LIBRARY_PATH
+#Set OneAPI environmet if it's not set by the user
+if [ ! -x "$(command -v sycl-ls)" ]
+then
+    echo "Setting OneAPI environment"
+    if [[ -z "$ONEAPI_ROOT" ]]
+    then
+        ONEAPI_ROOT=/opt/intel/oneapi
+    fi
+    source $ONEAPI_ROOT/setvars.sh
+fi
+
 export LD_PRELOAD=/usr/lib/libstdc++.so
 export NEOReadDebugKeys=1
 export ClDeviceGlobalMemSizeAvailablePercent=100