added finetuning example for gemma-2b on ARC. (#1328)

Signed-off-by: Ye, Xinyu <xinyu.ye@intel.com>

intel_extension_for_transformers/llm/finetuning/finetuning.py

@@ -61,6 +61,7 @@
     is_deepspeed_available,
 )
 from intel_extension_for_transformers.utils.device_utils import is_hpu_available
+from intel_extension_for_transformers.neural_chat.utils.common import get_device_type
 
 
 if is_bitsandbytes_available():
@@ -76,10 +77,7 @@ def __init__(self, finetuning_config: BaseFinetuningConfig):
             finetuning_config.finetune_args
         )
         if finetuning_config.finetune_args.device == "auto":
-            if torch.cuda.is_available():
-                finetuning_config.finetune_args.device = "cuda"
-            else:
-                finetuning_config.finetune_args.device = "cpu"
+            finetuning_config.finetune_args.device = get_device_type()
         if finetuning_config.finetune_args.device == "cpu":
             Arguments = type(finetuning_config.training_args)
             training_args = {

intel_extension_for_transformers/neural_chat/examples/finetuning/instruction/README.md

@@ -625,6 +625,35 @@ For finetuning on SPR, add `--bf16` argument will speedup the finetuning process
 You could also indicate `--peft` to switch peft method in P-tuning, Prefix tuning, Prompt tuning, LLama Adapter, LoRA,
 see https://github.com/huggingface/peft. Note for MPT, only LoRA is supported.
 
+## Fine-tuning on Intel Arc GPUs
+
+### 1. Single Card Fine-tuning
+
+Follow the installation guidance in [intel-extension-for-pytorch](https://github.com/intel/intel-extension-for-pytorch) to install intel-extension-for-pytorch for GPU.
+
+For `google/gemma-2b`, use the below command line for finetuning on the Alpaca dataset.
+
+```bash
+python finetune_clm.py \
+        --model_name_or_path "google/gemma-2b" \
+        --train_file "/path/to/alpaca_data.json" \
+        --dataset_concatenation \
+        --per_device_train_batch_size 2 \
+        --per_device_eval_batch_size 2 \
+        --gradient_accumulation_steps 4 \
+        --evaluation_strategy "no" \
+        --save_strategy "steps" \
+        --save_steps 2000 \
+        --save_total_limit 1 \
+        --learning_rate 1e-4  \
+        --do_train \
+        --num_train_epochs 3 \
+        --overwrite_output_dir \
+        --log_level info \
+        --output_dir ./gemma-2b_peft_finetuned_model \
+        --peft lora \
+        --gradient_checkpointing True
+```
 
 # Evaluation Metrics