Fix IPEX auto importer (#11192)

* Fix ipex auto importer with Python builtins.
* Raise errors if the user imports ipex manually before importing ipex_llm. Do nothing if they import ipex after importing ipex_llm.
* Remove import ipex in examples.

python/llm/example/CPU/Applications/streaming-llm/streaming_llm/utils.py

@@ -51,7 +51,6 @@ import json
 # code change to import from IPEX-LLM API instead of using transformers API
 from ipex_llm.transformers import AutoModelForCausalLM
 from transformers import LlamaTokenizer
-import intel_extension_for_pytorch as ipex
 
 
 def load(model_name_or_path):

python/llm/example/CPU/Deepspeed-AutoTP/deepspeed_autotp.py

@@ -47,7 +47,6 @@ from transformers import AutoModelForCausalLM, LlamaTokenizer, AutoTokenizer
 import deepspeed
 from ipex_llm import optimize_model
 import torch
-import intel_extension_for_pytorch as ipex
 import time
 import argparse
 

python/llm/example/GPU/HF-Transformers-AutoModels/Model/codeshell/server.py

@@ -274,9 +274,6 @@ if __name__ == "__main__":
     multi_turn = args.multi_turn
     max_context = args.max_context
 
-    if device == 'xpu':
-        import intel_extension_for_pytorch as ipex
-
     model = model.to(device)
 
     model.generation_config = GenerationConfig.from_pretrained(

python/llm/example/GPU/HF-Transformers-AutoModels/Model/internlm2/generate.py

@@ -20,7 +20,6 @@ import argparse
 
 from transformers import AutoTokenizer
 from ipex_llm import optimize_model
-import intel_extension_for_pytorch as ipex
 
 # you could tune the prompt based on your own model,
 # here the prompt tuning refers to https://huggingface.co/internlm/internlm-chat-7b/blob/main/modeling_internlm.py#L1053

python/llm/example/GPU/HF-Transformers-AutoModels/Model/phixtral/generate.py

@@ -20,7 +20,6 @@ import argparse
 import numpy as np
 
 from transformers import AutoTokenizer, GenerationConfig
-import intel_extension_for_pytorch as ipex
 
 
 # you could tune the prompt based on your own model,

python/llm/example/GPU/HF-Transformers-AutoModels/Model/qwen1.5/generate.py

@@ -20,7 +20,6 @@ import argparse
 
 from transformers import AutoTokenizer
 from ipex_llm import optimize_model
-import intel_extension_for_pytorch as ipex
 import numpy as np
 
 

python/llm/example/GPU/HF-Transformers-AutoModels/Model/solar/generate.py

@@ -15,7 +15,6 @@
 #
 
 import torch
-import intel_extension_for_pytorch as ipex
 import time
 import argparse
 

python/llm/example/GPU/HF-Transformers-AutoModels/Model/yuan2/generate.py

@@ -16,7 +16,6 @@
 
 import torch, transformers
 import sys, os, time
-import intel_extension_for_pytorch as ipex
 import argparse
 from transformers import LlamaTokenizer
 from ipex_llm.transformers import AutoModelForCausalLM

python/llm/example/GPU/Pipeline-Parallel-FastAPI/pipeline_models.py

@@ -1,7 +1,6 @@
 from torch import nn
 import torch
 import torch.distributed as dist
-import intel_extension_for_pytorch as ipex
 
 from typing import List, Optional, Tuple, Union, Iterator
 import time

python/llm/example/GPU/Pipeline-Parallel-FastAPI/pipeline_serving.py

@@ -2,7 +2,6 @@ from pipeline_models import ModelRunner
 import torch.nn.parallel
 import torch.distributed as dist
 import os
-import intel_extension_for_pytorch as ipex
 
 import oneccl_bindings_for_pytorch
 

python/llm/example/GPU/Pipeline-Parallel-Inference/generate.py

@@ -16,7 +16,6 @@
 #
 
 import torch
-import intel_extension_for_pytorch as ipex
 import time
 import argparse
 

python/llm/example/GPU/PyTorch-Models/Model/internlm2/generate.py

@@ -20,7 +20,6 @@ import argparse
 
 from transformers import AutoTokenizer
 from ipex_llm import optimize_model
-import intel_extension_for_pytorch as ipex
 
 # you could tune the prompt based on your own model,
 # here the prompt tuning refers to https://huggingface.co/internlm/internlm-chat-7b/blob/main/modeling_internlm.py#L1053

python/llm/example/GPU/PyTorch-Models/Model/mamba/generate.py

@@ -17,7 +17,6 @@
 import argparse
 import time
 import torch
-import intel_extension_for_pytorch as ipex
 from ipex_llm import optimize_model
 from transformers import AutoTokenizer
 

python/llm/example/GPU/PyTorch-Models/Model/phixtral/generate.py

@@ -20,7 +20,6 @@ import argparse
 import numpy as np
 
 from transformers import AutoTokenizer, GenerationConfig
-import intel_extension_for_pytorch as ipex
 from ipex_llm import optimize_model
 
 

python/llm/example/GPU/PyTorch-Models/Model/qwen1.5/generate.py

@@ -20,7 +20,6 @@ import argparse
 
 from transformers import AutoTokenizer
 from ipex_llm import optimize_model
-import intel_extension_for_pytorch as ipex
 import numpy as np
 
 

python/llm/example/GPU/PyTorch-Models/Model/solar/generate.py

@@ -15,7 +15,6 @@
 #
 
 import torch
-import intel_extension_for_pytorch as ipex
 import time
 import argparse
 

python/llm/example/GPU/PyTorch-Models/Model/yuan2/generate.py

@@ -16,7 +16,6 @@
 
 import torch, transformers
 import sys, os, time
-import intel_extension_for_pytorch as ipex
 import argparse
 from transformers import LlamaTokenizer, AutoModelForCausalLM
 from ipex_llm import optimize_model

python/llm/example/GPU/Speculative-Decoding/EAGLE/evaluation/gen_ea_answer_llama2chat.py

@@ -47,7 +47,6 @@ from eagle.model.ea_model import EaModel
 from eagle.model.utils import *
 from eagle.model.kv_cache import initialize_past_key_values
 from eagle.model.choices import *
-import intel_extension_for_pytorch as ipex
 from ipex_llm import optimize_model
 
 def ea_forward(input_ids, model, tokenizer, tree_choices, logits_processor=None, max_steps=512):

python/llm/src/ipex_llm/utils/ipex_importer.py

@@ -16,6 +16,9 @@
 
 from importlib.metadata import distribution, PackageNotFoundError
 import logging
+import builtins
+import sys
+from ipex_llm.utils.common import log4Error
 
 
 class IPEXImporter:
@@ -51,15 +54,36 @@ class IPEXImporter:
 
     def import_ipex(self):
         """
-        Try to import Intel Extension for PyTorch as ipex
+        Try to import Intel Extension for PyTorch as ipex for XPU
 
-        Raises ImportError if failed
+        Raises ImportError and invalidInputError if failed
         """
         if self.is_xpu_version_installed():
-            import intel_extension_for_pytorch as ipex
+            # Check if user import ipex manually
+            if 'ipex' in sys.modules or 'intel_extension_for_pytorch' in sys.modules:
+                logging.error("ipex_llm will automatically import intel_extension_for_pytorch.")
+                log4Error.invalidInputError(False,
+                                            "Please import ipex_llm before importing \
+                                                intel_extension_for_pytorch!")
+            self.directly_import_ipex()
             self.ipex_version = ipex.__version__
             logging.info("intel_extension_for_pytorch auto imported")
 
+    def directly_import_ipex(self):
+        """
+        Try to import Intel Extension for PyTorch as ipex
+
+        Raises ImportError and invalidInputError if failed
+        """
+        # import ipex
+        import intel_extension_for_pytorch as ipex
+        if ipex is not None:
+            # Expose ipex to Python builtins
+            builtins.ipex = ipex
+        else:
+            log4Error.invalidInputError(False,
+                                        "Can not import intel_extension_for_pytorch.")
+
     def get_ipex_version(self):
         """
         Get ipex version
@@ -69,11 +93,8 @@ class IPEXImporter:
         if self.ipex_version is not None:
             return self.ipex_version
         # try to import Intel Extension for PyTorch and get version
-        try:
+        self.directly_import_ipex()
-            import intel_extension_for_pytorch as ipex
+        self.ipex_version = ipex.__version__
-            self.ipex_version = ipex.__version__
-        except ImportError:
-            self.ipex_version = None
         return self.ipex_version