Changed ipex from ipex-xpu -> ipex-cpu (#90)

Author: Egor-Krivov

dl_bench/llm.py

@@ -26,7 +26,7 @@ def get_llm(name, dtype):
 
     kwargs = {}
     if name.startswith("llama2") and "HF_TOKEN" in os.environ:
-        kwargs = {"HF_TOKEN": os.environ.get("HF_TOKEN")}
+        kwargs = {"token": os.environ.get("HF_TOKEN")}
 
     model_name, M, T = name2params[name]
 
@@ -75,14 +75,15 @@ def inference(self, backend):
         # self.flops_per_sample = get_macs(self.model, self.in_shape, backend) * 2
         self.model = backend.prepare_eval_transformer(self.model)
 
-        self.model.eval()
         enabled = backend.dtype != torch.float32
 
         n_items = 0
         outputs = []
         fw_times = []
 
-        self.model.eval()
+
+        # Ipex gives error with eval, other backends have no effect
+        # self.model.eval()
         for i in range(self.n_iter):
             print(f"Epoch {i+1}/{self.n_iter}")
             cast = torch.autocast(enabled=enabled, device_type=backend.device_name)

dl_bench/utils.py

@@ -132,7 +132,7 @@ def prepare_eval_transformer(self, model):
         model = model.to(memory_format=torch.channels_last)
 
         model.to(self.device)
-        with torch.inference_mode():
+        with torch.no_grad():
             model.eval()
             return self._compile_transformer_model(
                 self.compile_mode, model, dtype=self.dtype
@@ -160,7 +160,9 @@ def _compile_transformer_model(compile_mode, model, dtype=torch.bfloat16):
             import intel_extension_for_pytorch as ipex
 
             params = {} if dtype != torch.bfloat16 else {"dtype": torch.bfloat16}
-            compiled_model = ipex.optimize_transformers(model, **params)
+            #compiled_model = ipex.llm.optimize(model, **params, inplace=True, deployment_mode=True)
+            compiled_model = ipex.llm.optimize(model, **params)
+            # compiled_model = ipex.optimize_transformers(model, **params)
             print("Compiled with ipex")
         elif compile_mode == "ipex_onednn_graph":
             raise NotImplementedError()

tests/conda-envs/ipex-xpu.yaml

@@ -0,0 +1,14 @@
+name: ipex
+channels:
+  - intel
+  - conda-forge
+dependencies:
+  - intel-aikit-pytorch
+  - pytorch>=2.0.1=*_xpu_*
+  - intel-extension-for-pytorch
+  - datasets
+  - accelerate
+  - sentencepiece
+# The following packages are required to run benchmarks
+  - sqlalchemy>=2.0.0
+  - pytest

tests/conda-envs/ipex.yaml

@@ -3,12 +3,20 @@ channels:
   - intel
   - conda-forge
 dependencies:
-  - intel-aikit-pytorch
-  - pytorch>=2.0.1=*_xpu_*
-  - intel-extension-for-pytorch
+  - python=3.11
   - datasets
   - accelerate
   - sentencepiece
 # The following packages are required to run benchmarks
   - sqlalchemy>=2.0.0
   - pytest
+  - pip
+  - pip:
+    - --extra-index-url https://download.pytorch.org/whl/cpu
+    - torch
+    - torchvision
+    - torchaudio
+    - transformers==4.35.2
+    - intel-extension-for-pytorch
+    - --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/cpu/us/
+    - oneccl_bind_pt