add ppo rl_training part (#634)

Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>

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

@@ -43,3 +43,25 @@ multi card finetunes
 ```
 python ../instruction/gaudi_spawn.py --world_size 8 --use_mpi reward_modeling.py --model_name_or_path meta-llama/Llama-2-7b-hf --output_dir <output>  --log_level info  --num_train_epochs 1 --use_habana --use_lazy_mode --hf_access_token xxxxxx --ddp_find_unused_parameters True
 ```
+
+## 5. Reinforcement Fine-tuning
+
+### Training on CUDA
+```
+accelerate launch --multi_gpu --num_machines 1  --num_processes 8 rl_training.py --log_with=wandb --model_name=meta-llama/Llama-2-7b-hf  --reward_model_name=output_se --adafactor=False --tokenizer_name=meta-llama/Llama-2-7b-hf  --save_freq=100 --output_max_length=128 --batch_size=8 --gradient_accumulation_steps=8 --batched_gen=True --ppo_epochs=4 --seed=0 --learning_rate=1.4e-5 --early_stopping=True --output_dir=llama-se-rl-finetune-128-8-8-1.4e-5_adam --hf_access_token xxxxxx
+```
+
+### Training on Habana
+
+Follow install guidance in [optimum-habana](https://github.com/huggingface/optimum-habana)
+
+single card finetune
+
+```
+python3 rl_training.py  --model_name=meta-llama/Llama-2-7b-hf --reward_model_name=<output_rm> --adafactor=False --tokenizer_name=meta-llama/Llama-2-7b-hf --save_freq=100 --output_max_length=128 --batch_size=8 --mini_batch_size=1 --gradient_accumulation_steps=8 --batched_gen=True --ppo_epochs=4 --seed=0 --learning_rate=1.4e-5 --early_stopping=True --output_dir=llama-se-rl-finetune-128-8-8-1.4e-5_adam --hf_access_token xxxxxx --use_habana
+```
+
+multi card finetunes
+```
+python3 ../instruction/gaudi_spawn.py --world_size 8 --use_mpi rl_training.py  --model_name=meta-llama/Llama-2-7b-hf --reward_model_name=<output_rm> --adafactor=False --tokenizer_name=meta-llama/Llama-2-7b-hf --save_freq=100 --output_max_length=128 --batch_size=8 --mini_batch_size=1 --gradient_accumulation_steps=8 --batched_gen=True --ppo_epochs=4 --seed=0 --learning_rate=1.4e-5 --early_stopping=True --output_dir=llama-se-rl-finetune-128-8-8-1.4e-5_adam --hf_access_token xxxxxx --use_habana
+```

intel_extension_for_transformers/neural_chat/examples/finetuning/ppo_pipeline/requirements.txt

@@ -5,3 +5,5 @@ datasets
 bitsandbytes
 evaluate
 scikit-learn
+intel-extension-for-transformers
+tyro

intel_extension_for_transformers/neural_chat/examples/finetuning/ppo_pipeline/reward_modeling.py

@@ -199,14 +199,14 @@ def preprocess_function(examples):
         "input_ids_k": [],
         "attention_mask_k": [],
     }
-    for question, response_j, response_k in zip(
-        examples["question"], examples["chatgpt"], examples["llama2-13b-chat"]
+    for system, question, response_j, response_k in zip(
+        examples["system"], examples["question"], examples["chatgpt"], examples["llama2-13b-chat"]
     ):
         tokenized_j = tokenizer(
-            "Question: " + question + "\n\nAnswer: " + response_j, truncation=True
+            system + question + response_j, truncation=True
         )
         tokenized_k = tokenizer(
-            "Question: " + question + "\n\nAnswer: " + response_k, truncation=True
+            system + question + response_k, truncation=True
         )
 
         new_examples["input_ids_j"].append(tokenized_j["input_ids"])