Freeze IPEX version for INT8 SQ support (#2221)

Signed-off-by: Kaihui-intel <kaihui.tang@intel.com>

Author: Kaihui-intel

examples/3.x_api/pytorch/nlp/huggingface_models/language-modeling/quantization/smooth_quant/README.md

@@ -49,7 +49,7 @@ python run_clm_no_trainer.py \
 ```
 
 ### LLAMA2-7b/13b/70b
->Note: LLAMA requires IPEX requirements >= 2.1 to get better accuracy.
+>Note: LLAMA requires IPEX requirements >= 2.1 to get better accuracy. LLAMA requires transformers < 4.48.0.
 #### Quantization
 
 ```bash

examples/3.x_api/pytorch/nlp/huggingface_models/language-modeling/quantization/smooth_quant/requirements.txt

@@ -2,13 +2,13 @@ accelerate
 protobuf
 sentencepiece != 0.1.92
 datasets >= 1.1.3
-torch >= 1.10
-transformers < 4.48.0  # TODO: ILITV-3858
+torch == 2.7.0
+transformers
 pytest
 wandb
 einops
 neural-compressor
 lm_eval <= 0.4.7
 peft
 optimum-intel
-intel_extension_for_pytorch
+intel_extension_for_pytorch == 2.7.0

examples/3.x_api/pytorch/nlp/huggingface_models/language-modeling/quantization/smooth_quant/run_clm_no_trainer.py

@@ -30,6 +30,7 @@
     "--approach", type=str, default="static", help="Select from ['dynamic', 'static', 'weight-only']"
 )
 parser.add_argument("--optimized", action="store_true")
+parser.add_argument("--autotune", action="store_true", help="Use autotune to find the best alpha for SmoothQuant.")
 parser.add_argument("--ipex", action="store_true", help="Use intel extension for pytorch.")
 parser.add_argument("--load", action="store_true", help="Load quantized model.")
 parser.add_argument("--accuracy", action="store_true")
@@ -204,15 +205,25 @@ def eval_func(model):
 
     example_inputs = get_example_inputs(user_model, calib_dataloader)
 
-    from neural_compressor.torch.quantization import SmoothQuantConfig, autotune, TuningConfig
-    tune_config = TuningConfig(config_set=SmoothQuantConfig.get_config_set_for_tuning())
-    user_model = autotune(
-        user_model, 
-        tune_config=tune_config,
-        eval_fn=eval_func,
-        run_fn=run_fn,
-        example_inputs=example_inputs,
-    )
+    if args.autotune:
+        from neural_compressor.torch.quantization import SmoothQuantConfig, autotune, TuningConfig
+        tune_config = TuningConfig(config_set=SmoothQuantConfig.get_config_set_for_tuning())
+        user_model = autotune(
+            user_model, 
+            tune_config=tune_config,
+            eval_fn=eval_func,
+            run_fn=run_fn,
+            example_inputs=example_inputs,
+        )
+    else:        
+        from neural_compressor.torch.quantization import SmoothQuantConfig, prepare, convert
+        args.alpha = eval(args.alpha)
+        excluded_precisions = [] if args.int8_bf16_mixed else ["bf16"]
+        quant_config = SmoothQuantConfig(alpha=args.alpha, folding=False, excluded_precisions=excluded_precisions)
+
+        user_model = prepare(model=user_model, quant_config=quant_config, example_inputs=example_inputs)
+        run_fn(user_model)
+        user_model = convert(user_model)
     user_model.save(args.output_dir)
 
 

examples/3.x_api/pytorch/nlp/huggingface_models/language-modeling/quantization/smooth_quant/run_quant.sh

@@ -47,10 +47,10 @@ function run_tuning {
         extra_cmd=$extra_cmd" --ipex --sq --alpha 0.5"
     elif [ "${topology}" = "llama2_7b_ipex_sq" ]; then
         model_name_or_path="meta-llama/Llama-2-7b-hf"
-        extra_cmd=$extra_cmd" --ipex --sq --alpha 0.8"
+        extra_cmd=$extra_cmd" --ipex --sq --alpha 0.65"
     elif [ "${topology}" = "gpt_j_ipex_sq" ]; then
         model_name_or_path="EleutherAI/gpt-j-6b"
-        extra_cmd=$extra_cmd" --ipex --sq --alpha 1.0"
+        extra_cmd=$extra_cmd" --ipex --sq --alpha 0.5"
     fi
 
     python -u run_clm_no_trainer.py \