support DPO training (2305.18290)

README.md

@@ -12,6 +12,8 @@
 
 ## Changelog
 
+[23/08/11] Now we support **[DPO training](https://arxiv.org/abs/2305.18290)** for instruction-tuned models. See [this example](#dpo-training) to train your models (experimental feature).
+
 [23/08/03] Now we support training the **Qwen-7B** model in this repo. Try `--model_name_or_path Qwen/Qwen-7B-Chat` and `--lora_target c_attn` arguments to train the Qwen-7B model. Remember to use `--template chatml` argument when you are using the Qwen-7B-Chat model.
 
 [23/07/31] Now we support dataset streaming. Try `--streaming` and `--max_steps 100` arguments to stream your dataset.
@@ -54,24 +56,18 @@
 | [Qwen](https://github.com/QwenLM/Qwen-7B)                | 7B                          | c_attn            | chatml   |
 | [XVERSE](https://github.com/xverse-ai/XVERSE-13B)        | 13B                         | q_proj,v_proj     | -        |
 
-> * **Default module** is used for the `--lora_target` argument. Please use `python src/train_bash.py -h` to see all available options.
+- **Default module** is used for the `--lora_target` argument. Please use `python src/train_bash.py -h` to see all available options.
-> * For the "base" models, the `--template` argument can be chosen from `default`, `alpaca`, `vicuna` etc.
+- For the "base" models, the `--template` argument can be chosen from `default`, `alpaca`, `vicuna` etc. But make sure to use the corresponding template for the "chat" models.
 
 ## Supported Training Approaches
 
-- [(Continually) pre-training](https://s3-us-west-2.amazonaws.com/openai-assets/research-covers/language-unsupervised/language_understanding_paper.pdf)
+| Approach               | Full-parameter | Partial-parameter | LoRA | QLoRA |
-  - Full-parameter tuning
+| ---------------------- | -------------- | ----------------- | ---- | ----- |
-  - Partial-parameter tuning
+| Pre-Training           | ✅            | ✅                | ✅   | ✅   |
-  - [LoRA](https://arxiv.org/abs/2106.09685)
+| Supervised Fine-Tuning | ✅            | ✅                | ✅   | ✅   |
-  - [QLoRA](https://arxiv.org/abs/2305.14314)
+| Reward Model Training  |                |                   | ✅   | ✅   |
-- [Supervised fine-tuning](https://arxiv.org/abs/2109.01652)
+| PPO Training           |                |                   | ✅   | ✅   |
-  - Full-parameter tuning
+| DPO Training           | ✅            |                    | ✅   | ✅   |
-  - Partial-parameter tuning
-  - [LoRA](https://arxiv.org/abs/2106.09685)
-  - [QLoRA](https://arxiv.org/abs/2305.14314)
-- [RLHF](https://arxiv.org/abs/2203.02155)
-  - [LoRA](https://arxiv.org/abs/2106.09685)
-  - [QLoRA](https://arxiv.org/abs/2305.14314)
 
 ## Provided Datasets
 
@@ -88,7 +84,6 @@
   - [Open Assistant (multilingual)](https://huggingface.co/datasets/OpenAssistant/oasst1)
   - [Self-cognition (zh)](data/self_cognition.json)
   - [ShareGPT (zh)](https://huggingface.co/datasets/QingyiSi/Alpaca-CoT/tree/main/Chinese-instruction-collection)
-  - [RefGPT (zh)](https://github.com/sufengniu/RefGPT)
   - [Guanaco Dataset (multilingual)](https://huggingface.co/datasets/JosephusCheung/GuanacoDataset)
   - [BELLE 2M (zh)](https://huggingface.co/datasets/BelleGroup/train_2M_CN)
   - [BELLE 1M (zh)](https://huggingface.co/datasets/BelleGroup/train_1M_CN)
@@ -103,7 +98,7 @@
   - [Web QA (zh)](https://huggingface.co/datasets/suolyer/webqa)
   - [UltraChat (en)](https://github.com/thunlp/UltraChat)
   - [WebNovel (zh)](https://huggingface.co/datasets/zxbsmk/webnovel_cn)
-- For reward modelling:
+- For reward modelling or DPO training:
   - [HH-RLHF (en)](https://huggingface.co/datasets/Anthropic/hh-rlhf)
   - [Open Assistant (multilingual)](https://huggingface.co/datasets/OpenAssistant/oasst1)
   - [GPT-4 Generated Data (en&zh)](https://github.com/Instruction-Tuning-with-GPT-4/GPT-4-LLM)
@@ -139,7 +134,6 @@ Note: please update `data/dataset_info.json` to use your custom dataset. About t
 ### Dependence Installation (optional)
 
 ```bash
-git lfs install
 git clone https://github.com/hiyouga/LLaMA-Efficient-Tuning.git
 conda create -n llama_etuning python=3.10
 conda activate llama_etuning
@@ -161,7 +155,7 @@ CUDA_VISIBLE_DEVICES=0 python src/train_web.py
 
 Currently the web UI only supports training on **a single GPU**.
 
-### (Continually) Pre-Training
+### Pre-Training
 
 ```bash
 CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
@@ -222,7 +216,7 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --resume_lora_training False \
     --checkpoint_dir path_to_sft_checkpoint \
     --output_dir path_to_rm_checkpoint \
-    --per_device_train_batch_size 4 \
+    --per_device_train_batch_size 2 \
     --gradient_accumulation_steps 4 \
     --lr_scheduler_type cosine \
     --logging_steps 10 \
@@ -233,7 +227,7 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --fp16
 ```
 
-### PPO Training (RLHF)
+### PPO Training
 
 ```bash
 CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
@@ -257,6 +251,30 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --plot_loss
 ```
 
+### DPO Training
+
+```bash
+CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
+    --stage dpo \
+    --model_name_or_path path_to_your_model \
+    --do_train \
+    --dataset comparison_gpt4_en \
+    --template default \
+    --finetuning_type lora \
+    --resume_lora_training False \
+    --checkpoint_dir path_to_sft_checkpoint \
+    --output_dir path_to_dpo_checkpoint \
+    --per_device_train_batch_size 2 \
+    --gradient_accumulation_steps 4 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --save_steps 1000 \
+    --learning_rate 1e-5 \
+    --num_train_epochs 1.0 \
+    --plot_loss \
+    --fp16
+```
+
 ### Distributed Training
 
 ```bash

README_zh.md

@@ -12,7 +12,9 @@
 
 ## 更新日志
 
-[23/08/03] 现在我们支持了 **Qwen-7B** 模型的训练。请尝试使用 `--model_name_or_path Qwen/Qwen-7B-Chat` 和 `--lora_target c_attn` 参数。请注意使用 Qwen-7B-Chat 模型需要添加 `--template chatml` 参数。
+[23/08/11] 现在我们支持了指令模型的 **[DPO 训练](https://arxiv.org/abs/2305.18290)**。详情请参阅[此示例](#dpo-training)（实验性功能）。
+
+[23/08/03] 现在我们支持了 **Qwen-7B** 模型的训练。请尝试使用 `--model_name_or_path Qwen/Qwen-7B-Chat` 和 `--lora_target c_attn` 参数。使用 Qwen-7B-Chat 模型请添加 `--template chatml` 参数。
 
 [23/07/31] 现在我们支持了训练数据流式加载。请尝试使用 `--streaming` 和 `--max_steps 100` 参数来流式加载数据集。
 
@@ -54,41 +56,34 @@
 | [Qwen](https://github.com/QwenLM/Qwen-7B)                | 7B                          | c_attn            | chatml   |
 | [XVERSE](https://github.com/xverse-ai/XVERSE-13B)        | 13B                         | q_proj,v_proj     | -        |
 
-> * **默认模块**是 `--lora_target` 参数的默认值。请使用 `python src/train_bash.py -h` 查看全部可选项。
+- **默认模块**是 `--lora_target` 参数的部分可选项。请使用 `python src/train_bash.py -h` 查看全部可选项。
-> * 对于所有“基座”模型，`--template` 参数可以是 `default`, `alpaca`, `vicuna` 等值。
+- 对于所有“基座”（Base）模型，`--template` 参数可以是 `default`, `alpaca`, `vicuna` 等任意值。但“对话”（Chat）模型请务必使用对应的模板。
 
-## 微调方法
+## 训练方法
 
-- [二次预训练](https://s3-us-west-2.amazonaws.com/openai-assets/research-covers/language-unsupervised/language_understanding_paper.pdf)
+| 方法        | 全参数训练 | 部分参数训练 | LoRA | QLoRA |
-  - 全参数微调
+| ---------- | ---------- | ----------- | ---- | ----- |
-  - 部分参数微调
+| 预训练      | ✅        | ✅         | ✅   | ✅   |
-  - [LoRA](https://arxiv.org/abs/2106.09685)
+| 指令监督微调 | ✅        | ✅         | ✅   | ✅   |
-  - [QLoRA](https://arxiv.org/abs/2305.14314)
+| 奖励模型训练 |           |             | ✅   | ✅   |
-- [指令监督微调](https://arxiv.org/abs/2109.01652)
+| PPO 训练    |           |             | ✅   | ✅   |
-  - 全参数微调
+| DPO 训练    | ✅        |             | ✅   | ✅   |
-  - 部分参数微调
-  - [LoRA](https://arxiv.org/abs/2106.09685)
-  - [QLoRA](https://arxiv.org/abs/2305.14314)
-- [人类反馈的强化学习（RLHF）](https://arxiv.org/abs/2203.02155)
-  - [LoRA](https://arxiv.org/abs/2106.09685)
-  - [QLoRA](https://arxiv.org/abs/2305.14314)
 
 ## 数据集
 
-- 用于二次预训练:
+- 用于预训练：
   - [Wiki Demo (en)](data/wiki_demo.txt)
   - [RefinedWeb (en)](https://huggingface.co/datasets/tiiuae/falcon-refinedweb)
   - [StarCoder (en)](https://huggingface.co/datasets/bigcode/starcoderdata)
   - [Wikipedia (en)](https://huggingface.co/datasets/olm/olm-wikipedia-20221220)
   - [Wikipedia (zh)](https://huggingface.co/datasets/pleisto/wikipedia-cn-20230720-filtered)
-- 用于指令监督微调:
+- 用于指令监督微调：
   - [Stanford Alpaca (en)](https://github.com/tatsu-lab/stanford_alpaca)
   - [Stanford Alpaca (zh)](https://github.com/ymcui/Chinese-LLaMA-Alpaca)
   - [GPT-4 Generated Data (en&zh)](https://github.com/Instruction-Tuning-with-GPT-4/GPT-4-LLM)
   - [Open Assistant (multilingual)](https://huggingface.co/datasets/OpenAssistant/oasst1)
   - [Self-cognition (zh)](data/self_cognition.json)
   - [ShareGPT (zh)](https://huggingface.co/datasets/QingyiSi/Alpaca-CoT/tree/main/Chinese-instruction-collection)
-  - [RefGPT (zh)](https://github.com/sufengniu/RefGPT)
   - [Guanaco Dataset (multilingual)](https://huggingface.co/datasets/JosephusCheung/GuanacoDataset)
   - [BELLE 2M (zh)](https://huggingface.co/datasets/BelleGroup/train_2M_CN)
   - [BELLE 1M (zh)](https://huggingface.co/datasets/BelleGroup/train_1M_CN)
@@ -103,7 +98,7 @@
   - [Web QA (zh)](https://huggingface.co/datasets/suolyer/webqa)
   - [UltraChat (en)](https://github.com/thunlp/UltraChat)
   - [WebNovel (zh)](https://huggingface.co/datasets/zxbsmk/webnovel_cn)
-- 用于奖励模型训练:
+- 用于奖励模型或 DPO 训练：
   - [HH-RLHF (en)](https://huggingface.co/datasets/Anthropic/hh-rlhf)
   - [Open Assistant (multilingual)](https://huggingface.co/datasets/OpenAssistant/oasst1)
   - [GPT-4 Generated Data (en&zh)](https://github.com/Instruction-Tuning-with-GPT-4/GPT-4-LLM)
@@ -139,7 +134,6 @@ huggingface-cli login
 ### 环境搭建（可跳过）
 
 ```bash
-git lfs install
 git clone https://github.com/hiyouga/LLaMA-Efficient-Tuning.git
 conda create -n llama_etuning python=3.10
 conda activate llama_etuning
@@ -161,7 +155,7 @@ CUDA_VISIBLE_DEVICES=0 python src/train_web.py
 
 目前网页 UI 仅支持**单卡训练**。
 
-### 二次预训练
+### 预训练
 
 ```bash
 CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
@@ -222,7 +216,7 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --resume_lora_training False \
     --checkpoint_dir path_to_sft_checkpoint \
     --output_dir path_to_rm_checkpoint \
-    --per_device_train_batch_size 4 \
+    --per_device_train_batch_size 2 \
     --gradient_accumulation_steps 4 \
     --lr_scheduler_type cosine \
     --logging_steps 10 \
@@ -233,7 +227,7 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --fp16
 ```
 
-### RLHF 训练
+### PPO 训练
 
 ```bash
 CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
@@ -257,6 +251,30 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --plot_loss
 ```
 
+### DPO 训练
+
+```bash
+CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
+    --stage dpo \
+    --model_name_or_path path_to_your_model \
+    --do_train \
+    --dataset comparison_gpt4_zh \
+    --template default \
+    --finetuning_type lora \
+    --resume_lora_training False \
+    --checkpoint_dir path_to_sft_checkpoint \
+    --output_dir path_to_dpo_checkpoint \
+    --per_device_train_batch_size 2 \
+    --gradient_accumulation_steps 4 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --save_steps 1000 \
+    --learning_rate 1e-5 \
+    --num_train_epochs 1.0 \
+    --plot_loss \
+    --fp16
+```
+
 ### 多 GPU 分布式训练
 
 ```bash

data/dataset_info.json

@@ -49,26 +49,6 @@
       "history": "history"
     }
   },
-  "refgpt_zh_p1": {
-    "file_name": "refgpt_zh_50k_p1.json",
-    "file_sha1": "b40f4f4d0ffacd16da7c275b056d5b6670021752",
-    "columns": {
-      "prompt": "instruction",
-      "query": "input",
-      "response": "output",
-      "history": "history"
-    }
-  },
-  "refgpt_zh_p2": {
-    "file_name": "refgpt_zh_50k_p2.json",
-    "file_sha1": "181f32b2c60264a29f81f59d3c76095793eae1b0",
-    "columns": {
-      "prompt": "instruction",
-      "query": "input",
-      "response": "output",
-      "history": "history"
-    }
-  },
   "lima": {
     "file_name": "lima.json",
     "file_sha1": "9db59f6b7007dc4b17529fc63379b9cd61640f37",

requirements.txt

@@ -3,7 +3,7 @@ transformers>=4.29.1
 datasets>=2.12.0
 accelerate>=0.21.0
 peft>=0.4.0
-trl>=0.4.7
+trl>=0.5.0
 scipy
 sentencepiece
 tiktoken
@@ -16,4 +16,3 @@ pydantic==1.10.11
 fastapi==0.95.1
 sse-starlette
 matplotlib
-huggingface_hub

src/api_demo.py

@@ -7,7 +7,7 @@ def main():
     chat_model = ChatModel()
     app = create_app(chat_model)
     uvicorn.run(app, host="0.0.0.0", port=8000, workers=1)
-    # Visit http://localhost:8000/docs for document.
+    print("Visit http://localhost:8000/docs for API document.")
 
 
 if __name__ == "__main__":

src/llmtuner/chat/stream_chat.py

@@ -18,7 +18,6 @@ class ChatModel:
         self.model = self.model.eval() # change to eval mode
         self.template = get_template_and_fix_tokenizer(data_args.template, self.tokenizer)
         self.source_prefix = data_args.source_prefix
-        self.stop_ids = self.tokenizer.convert_tokens_to_ids(self.template.stop_words)
         self.model.generate = MethodType(PreTrainedModel.generate, self.model) # disable custom method (for Qwen)
 
     def process_args(
@@ -53,7 +52,7 @@ class ChatModel:
             top_k=top_k or gen_kwargs["top_k"],
             repetition_penalty=repetition_penalty or gen_kwargs["repetition_penalty"],
             logits_processor=get_logits_processor(),
-            stopping_criteria=get_stopping_criteria(self.stop_ids)
+            stopping_criteria=get_stopping_criteria(self.tokenizer.additional_special_tokens_ids)
         ))
 
         if max_length:

src/llmtuner/dsets/preprocess.py

@@ -46,7 +46,6 @@ def preprocess_dataset(
             k: [t[i: i + block_size] for i in range(0, total_length, block_size)]
             for k, t in concatenated_examples.items()
         }
-        result["labels"] = result["input_ids"].copy()
         return result
 
     def preprocess_supervised_dataset(examples: Dict[str, List[Any]]) -> Dict[str, Any]:
@@ -95,24 +94,22 @@ def preprocess_dataset(
         return model_inputs
 
     def preprocess_pairwise_dataset(examples):
-        # build input pairs with format `<bos> X Y1 <eos>` and `<bos> X Y2 <eos>`
+        # build input pairs with format `<bos> X`, `Y1 <eos>` and `Y2 <eos>`
-        model_inputs = {"accept_ids": [], "reject_ids": []}
+        model_inputs = {"prompt_ids": [], "chosen_ids": [], "rejected_ids": []}
         for query, response, history, prefix in construct_example(examples):
-            source_ids, accept_ids = template.encode_oneturn(tokenizer, query, response[0], history, prefix)
+            prompt_ids, chosen_ids = template.encode_oneturn(tokenizer, query, response[0], history, prefix)
-            source_ids, reject_ids = template.encode_oneturn(tokenizer, query, response[1], history, prefix)
+            _, rejected_ids = template.encode_oneturn(tokenizer, query, response[1], history, prefix)
 
-            if len(source_ids) > data_args.max_source_length:
+            if len(prompt_ids) > data_args.max_source_length:
-                source_ids = source_ids[:data_args.max_source_length]
+                prompt_ids = prompt_ids[:data_args.max_source_length]
-            if len(accept_ids) > data_args.max_target_length:
+            if len(chosen_ids) > data_args.max_target_length:
-                accept_ids = accept_ids[:data_args.max_target_length]
+                chosen_ids = chosen_ids[:data_args.max_target_length]
-            if len(reject_ids) > data_args.max_target_length:
+            if len(rejected_ids) > data_args.max_target_length:
-                reject_ids = reject_ids[:data_args.max_target_length]
+                rejected_ids = rejected_ids[:data_args.max_target_length]
 
-            accept_ids = source_ids + accept_ids
+            model_inputs["prompt_ids"].append(prompt_ids)
-            reject_ids = source_ids + reject_ids
+            model_inputs["chosen_ids"].append(chosen_ids)
-
+            model_inputs["rejected_ids"].append(rejected_ids)
-            model_inputs["accept_ids"].append(accept_ids)
-            model_inputs["reject_ids"].append(reject_ids)
         return model_inputs
 
     def print_supervised_dataset_example(example):
@@ -124,10 +121,12 @@ def preprocess_dataset(
         ], skip_special_tokens=False)))
 
     def print_pairwise_dataset_example(example):
-        print("accept_ids:\n{}".format(example["accept_ids"]))
+        print("prompt_ids:\n{}".format(example["prompt_ids"]))
-        print("accepts:\n{}".format(tokenizer.decode(example["accept_ids"], skip_special_tokens=False)))
+        print("prompt:\n{}".format(tokenizer.decode(example["prompt_ids"], skip_special_tokens=False)))
-        print("reject_ids:\n{}".format(example["reject_ids"]))
+        print("chosen_ids:\n{}".format(example["chosen_ids"]))
-        print("rejects:\n{}".format(tokenizer.decode(example["reject_ids"], skip_special_tokens=False)))
+        print("chosen:\n{}".format(tokenizer.decode(example["chosen_ids"], skip_special_tokens=False)))
+        print("rejected_ids:\n{}".format(example["rejected_ids"]))
+        print("rejected:\n{}".format(tokenizer.decode(example["rejected_ids"], skip_special_tokens=False)))
 
     def print_unsupervised_dataset_example(example):
         print("input_ids:\n{}".format(example["input_ids"]))

src/llmtuner/extras/callbacks.py

@@ -7,10 +7,16 @@ from datetime import timedelta
 from transformers import TrainerCallback
 from transformers.trainer_utils import has_length
 
+from llmtuner.extras.constants import LOG_FILE_NAME
+from llmtuner.extras.logging import get_logger
+
 if TYPE_CHECKING:
     from transformers import TrainingArguments, TrainerState, TrainerControl
 
 
+logger = get_logger(__name__)
+
+
 class LogCallback(TrainerCallback):
 
     def __init__(self, runner=None):
@@ -38,6 +44,9 @@ class LogCallback(TrainerCallback):
             self.in_training = True
             self.start_time = time.time()
             self.max_steps = state.max_steps
+            if os.path.exists(os.path.join(args.output_dir, LOG_FILE_NAME)):
+                logger.warning("Previous log file in this folder will be deleted.")
+                os.remove(os.path.join(args.output_dir, LOG_FILE_NAME))
 
     def on_train_end(self, args: "TrainingArguments", state: "TrainerState", control: "TrainerControl", **kwargs):
         r"""

src/llmtuner/extras/constants.py

@@ -1,10 +1,12 @@
 IGNORE_INDEX = -100
 
+LOG_FILE_NAME = "trainer_log.jsonl"
+
 VALUE_HEAD_FILE_NAME = "value_head.bin"
 
 FINETUNING_ARGS_NAME = "finetuning_args.json"
 
-LAYERNORM_NAMES = ["norm", "ln_f", "ln_attn", "ln_mlp"] # for LLaMA, BLOOM and Falcon settings
+LAYERNORM_NAMES = ["norm", "ln_f", "ln_attn", "ln_mlp"]
 
 METHODS = ["full", "freeze", "lora"]
 

src/llmtuner/extras/misc.py

@@ -1,7 +1,11 @@
 import torch
 from typing import TYPE_CHECKING, List, Optional, Tuple
-
+from transformers import (
-from transformers import LogitsProcessor, LogitsProcessorList, StoppingCriteria, StoppingCriteriaList
+    LogitsProcessor,
+    LogitsProcessorList,
+    StoppingCriteria,
+    StoppingCriteriaList
+)
 
 from llmtuner.extras.constants import LAYERNORM_NAMES
 

src/llmtuner/extras/template.py

@@ -61,7 +61,7 @@ class Template:
         prefix: Optional[str] = None
     ) -> Tuple[List[Union[str, Dict[str, str]]], List[Tuple[str, str]]]:
         r"""
-        Aligns inputs to a special format.
+        Aligns inputs to the standard format.
         """
         prefix = [prefix] if prefix else self.prefix # use prefix if provided
         history = history if (history and self.use_history) else []
@@ -92,28 +92,32 @@ class Template:
     ) -> List[Tuple[List[int], List[int]]]:
         r"""
         Encodes formatted inputs to pairs of token ids.
+        Turn 0: bos + prefix + sep + query    resp + eos
+        Turn t: sep + bos + query             resp + eos
         """
         bos_ids, eos_ids = self._get_special_ids(tokenizer)
         sep_ids = self._convert_inputs_to_ids(tokenizer, context=self.sep)
         encoded_pairs = []
         for turn_idx, (query, resp) in enumerate(history):
-            if turn_idx != 0:
+            if turn_idx == 0:
-                prefix_ids = sep_ids
+                if prefix: # has prefix
-            elif prefix:
+                    prefix_ids = bos_ids + self._convert_inputs_to_ids(tokenizer, context=prefix) + sep_ids
-                prefix_ids = self._convert_inputs_to_ids(tokenizer, context=prefix) + eos_ids + sep_ids
+                else:
+                    prefix_ids = bos_ids
             else:
-                prefix_ids = []
+                prefix_ids = sep_ids + bos_ids
 
-            query_ids = self._convert_inputs_to_ids(tokenizer, context=self.prompt, query=query)
+            query_ids = self._convert_inputs_to_ids(tokenizer, context=self.prompt, query=query, idx=str(turn_idx))
             resp_ids = self._convert_inputs_to_ids(tokenizer, context=[resp])
-            encoded_pairs.append((bos_ids + prefix_ids + query_ids, resp_ids + eos_ids))
+            encoded_pairs.append((prefix_ids + query_ids, resp_ids + eos_ids))
         return encoded_pairs
 
     def _convert_inputs_to_ids(
         self,
         tokenizer: "PreTrainedTokenizer",
         context: List[Union[str, Dict[str, str]]],
-        query: Optional[str] = ""
+        query: Optional[str] = "",
+        idx: Optional[str] = ""
     ) -> List[int]:
         r"""
         Converts context to token ids.
@@ -127,6 +131,7 @@ class Template:
         for elem in context:
             if isinstance(elem, str):
                 elem = elem.replace("{{query}}", query, 1)
+                elem = elem.replace("{{idx}}", idx, 1)
                 token_ids = token_ids + tokenizer.encode(elem, **kwargs)
             elif isinstance(elem, dict):
                 token_ids = token_ids + [tokenizer.convert_tokens_to_ids(elem.get("token"))]
@@ -146,10 +151,12 @@ class Llama2Template(Template):
     ) -> List[Tuple[List[int], List[int]]]:
         r"""
         Encodes formatted inputs to pairs of token ids.
+        Turn 0: bos + prefix + query    resp + eos
+        Turn t: bos + query             resp + eos
         """
         bos_ids, eos_ids = self._get_special_ids(tokenizer)
         encoded_pairs = []
-        assert isinstance(prefix[0], str), "LLaMA-2 template only accepts list containing a single str."
+        assert isinstance(prefix[0], str), "LLaMA-2 template only accepts list containing a single string."
         for turn_idx, (query, resp) in enumerate(history):
             if turn_idx == 0: # llama2 template has not sep_ids
                 query = prefix[0] + query
@@ -187,11 +194,12 @@ def get_template_and_fix_tokenizer(
     template = templates.get(name, None)
     assert template is not None, "Template {} does not exist.".format(name)
 
-    if tokenizer.eos_token_id is None: # inplace method
+    if len(template.stop_words): # inplace method
-        if len(template.stop_words):
+        tokenizer.eos_token = template.stop_words[0]
-            tokenizer.eos_token = template.stop_words[0]
+        logger.info("Replace eos token: {}".format(tokenizer.eos_token))
-        else:
+
-            tokenizer.eos_token = "<|endoftext|>"
+    if tokenizer.eos_token_id is None:
+        tokenizer.eos_token = "<|endoftext|>"
         logger.info("Add eos token: {}".format(tokenizer.eos_token))
 
     if tokenizer.pad_token_id is None:
@@ -422,12 +430,13 @@ register_template(
     name="baichuan",
     prefix=[],
     prompt=[
-        {"token": "<reserved_102>"},
+        {"token": "<reserved_102>"}, # user token (a little difference in position)
-        "{{query}}",
+        "{{query}}"
-        {"token": "<reserved_103>"}
     ],
     sep=[],
-    stop_words=[],
+    stop_words=[
+        "<reserved_103>" # assistant token
+    ],
     use_history=True
 )
 
@@ -440,7 +449,8 @@ register_template(
     name="starchat",
     prefix=[
         {"token": "<|system|>"},
-        "\n"
+        "\n",
+        {"token": "<|end|>"}
     ],
     prompt=[
         {"token": "<|user|>"},
@@ -466,7 +476,8 @@ register_template(
     name="chatml",
     prefix=[
         {"token": "<|im_start|>"},
-        "system\nYou are a helpful assistant."
+        "system\nYou are a helpful assistant.",
+        {"token": "<|im_end|>"}
     ],
     prompt=[
         {"token": "<|im_start|>"},
@@ -484,3 +495,23 @@ register_template(
     ],
     use_history=True
 )
+
+
+r"""
+Supports: https://huggingface.co/THUDM/chatglm2-6b
+"""
+register_template(
+    name="chatglm2",
+    prefix=[
+        {"token": "[gMASK]"},
+        {"token": "sop"}
+    ],
+    prompt=[
+        "[Round {{idx}}]\n\n问：{{query}}\n\n答："
+    ],
+    sep=[
+        "\n\n"
+    ],
+    stop_words=[],
+    use_history=True
+)

src/llmtuner/hparams/data_args.py

@@ -24,7 +24,7 @@ class DatasetAttr:
 
 @dataclass
 class DataArguments:
-    """
+    r"""
     Arguments pertaining to what data we are going to input our model for training and evaluation.
     """
     template: str = field(

src/llmtuner/hparams/finetuning_args.py

@@ -5,7 +5,7 @@ from dataclasses import asdict, dataclass, field
 
 @dataclass
 class FinetuningArguments:
-    """
+    r"""
     Arguments pertaining to which techniques we are going to fine-tuning with.
     """
     finetuning_type: Optional[Literal["none", "freeze", "lora", "full"]] = field(
@@ -14,7 +14,7 @@ class FinetuningArguments:
     )
     num_hidden_layers: Optional[int] = field(
         default=32,
-        metadata={"help": "Number of decoder blocks in the model. \
+        metadata={"help": "Number of decoder blocks in the model for partial-parameter (freeze) fine-tuning. \
                   LLaMA choices: [\"32\", \"40\", \"60\", \"80\"], \
                   LLaMA-2 choices: [\"32\", \"40\", \"80\"], \
                   BLOOM choices: [\"24\", \"30\", \"70\"], \
@@ -25,16 +25,16 @@ class FinetuningArguments:
     )
     num_layer_trainable: Optional[int] = field(
         default=3,
-        metadata={"help": "Number of trainable layers for Freeze fine-tuning."}
+        metadata={"help": "Number of trainable layers for partial-parameter (freeze) fine-tuning."}
     )
     name_module_trainable: Optional[Literal["mlp", "self_attn", "self_attention"]] = field(
         default="mlp",
-        metadata={"help": "Name of trainable modules for Freeze fine-tuning. \
+        metadata={"help": "Name of trainable modules for partial-parameter (freeze) fine-tuning. \
-                  LLaMA & LLaMA-2 choices: [\"mlp\", \"self_attn\"], \
+                  LLaMA choices: [\"mlp\", \"self_attn\"], \
                   BLOOM & Falcon choices: [\"mlp\", \"self_attention\"], \
                   Baichuan choices: [\"mlp\", \"self_attn\"], \
                   Qwen choices: [\"mlp\", \"attn\"], \
-                  InternLM, XVERSE choices: the same as LLaMA."}
+                  LLaMA-2, InternLM, XVERSE choices: the same as LLaMA."}
     )
     lora_rank: Optional[int] = field(
         default=8,
@@ -51,11 +51,15 @@ class FinetuningArguments:
     lora_target: Optional[str] = field(
         default="q_proj,v_proj",
         metadata={"help": "Name(s) of target modules to apply LoRA. Use commas to separate multiple modules. \
-                  LLaMA & LLaMA-2 choices: [\"q_proj\", \"k_proj\", \"v_proj\", \"o_proj\", \"gate_proj\", \"up_proj\", \"down_proj\"], \
+                  LLaMA choices: [\"q_proj\", \"k_proj\", \"v_proj\", \"o_proj\", \"gate_proj\", \"up_proj\", \"down_proj\"], \
                   BLOOM & Falcon choices: [\"query_key_value\", \"self_attention.dense\", \"mlp.dense\"], \
                   Baichuan choices: [\"W_pack\", \"o_proj\", \"gate_proj\", \"up_proj\", \"down_proj\"], \
                   Qwen choices: [\"c_attn\", \"attn.c_proj\", \"w1\", \"w2\", \"mlp.c_proj\"], \
-                  InternLM, XVERSE choices: the same as LLaMA."}
+                  LLaMA-2, InternLM, XVERSE choices: the same as LLaMA."}
+    )
+    dpo_beta: Optional[float] = field(
+        default=0.1,
+        metadata={"help": "The beta parameter for the DPO loss."}
     )
 
     def __post_init__(self):
@@ -72,14 +76,14 @@ class FinetuningArguments:
         assert self.finetuning_type in ["none", "freeze", "lora", "full"], "Invalid fine-tuning method."
 
     def save_to_json(self, json_path: str):
-        """Saves the content of this instance in JSON format inside `json_path`."""
+        r"""Saves the content of this instance in JSON format inside `json_path`."""
         json_string = json.dumps(asdict(self), indent=2, sort_keys=True) + "\n"
         with open(json_path, "w", encoding="utf-8") as f:
             f.write(json_string)
 
     @classmethod
     def load_from_json(cls, json_path: str):
-        """Creates an instance from the content of `json_path`."""
+        r"""Creates an instance from the content of `json_path`."""
         with open(json_path, "r", encoding="utf-8") as f:
             text = f.read()
         return cls(**json.loads(text))

src/llmtuner/hparams/general_args.py

@@ -4,10 +4,10 @@ from dataclasses import dataclass, field
 
 @dataclass
 class GeneralArguments:
-    """
+    r"""
     Arguments pertaining to which stage we are going to perform.
     """
-    stage: Optional[Literal["pt", "sft", "rm", "ppo"]] = field(
+    stage: Optional[Literal["pt", "sft", "rm", "ppo", "dpo"]] = field(
         default="sft",
         metadata={"help": "Which stage will be performed in training."}
     )

src/llmtuner/hparams/generating_args.py

@@ -4,7 +4,7 @@ from dataclasses import asdict, dataclass, field
 
 @dataclass
 class GeneratingArguments:
-    """
+    r"""
     Arguments pertaining to specify the decoding parameters.
     """
     do_sample: Optional[bool] = field(

src/llmtuner/hparams/model_args.py

@@ -1,12 +1,11 @@
 import torch
 from typing import Literal, Optional
 from dataclasses import dataclass, field
-from huggingface_hub.hf_api import HfFolder
 
 
 @dataclass
 class ModelArguments:
-    """
+    r"""
     Arguments pertaining to which model/config/tokenizer we are going to fine-tune.
     """
     model_name_or_path: str = field(
@@ -64,12 +63,11 @@ class ModelArguments:
         default=False,
         metadata={"help": "Whether to plot the training loss after fine-tuning or not."}
     )
-    hf_hub_token  : Optional[str] = field(
+    hf_auth_token: Optional[str] = field(
         default=None,
-        metadata={"help": "Path to the directory containing the checkpoints of the reward model."}
+        metadata={"help": "Auth token to log in with Hugging Face Hub."}
     )
 
-
     def __post_init__(self):
         if self.checkpoint_dir is not None: # support merging multiple lora weights
             self.checkpoint_dir = [cd.strip() for cd in self.checkpoint_dir.split(",")]
@@ -77,5 +75,6 @@ class ModelArguments:
         if self.quantization_bit is not None:
             assert self.quantization_bit in [4, 8], "We only accept 4-bit or 8-bit quantization."
 
-        if self.use_auth_token == True and self.hf_hub_token != None:
+        if self.use_auth_token == True and self.hf_auth_token is not None:
-            HfFolder.save_token(self.hf_hub_token)
+            from huggingface_hub.hf_api import HfFolder # lazy load
+            HfFolder.save_token(self.hf_auth_token)

src/llmtuner/tuner/core/adapter.py

@@ -39,7 +39,7 @@ def init_adapter(
     if finetuning_args.finetuning_type == "none" and is_trainable:
         raise ValueError("You cannot use finetuning_type=none while training.")
 
-    if finetuning_args.finetuning_type == "full":
+    if finetuning_args.finetuning_type == "full" and is_trainable:
         logger.info("Fine-tuning method: Full")
         model = model.float()
 

src/llmtuner/tuner/core/loader.py

@@ -34,7 +34,7 @@ check_min_version("4.29.1")
 require_version("datasets>=2.12.0", "To fix: pip install datasets>=2.12.0")
 require_version("accelerate>=0.21.0", "To fix: pip install accelerate>=0.21.0")
 require_version("peft>=0.4.0", "To fix: pip install peft>=0.4.0")
-require_version("trl>=0.4.7", "To fix: pip install trl>=0.4.7")
+require_version("trl>=0.5.0", "To fix: pip install trl>=0.5.0")
 
 
 def load_model_and_tokenizer(
@@ -52,9 +52,6 @@ def load_model_and_tokenizer(
         logger.warning("Checkpoint is not found at evaluation, load the original model.")
         finetuning_args = FinetuningArguments(finetuning_type="none")
 
-    assert stage in ["pt", "sft"] or finetuning_args.finetuning_type == "lora", \
-        "RM and PPO training can only be performed with the LoRA method."
-
     config_kwargs = {
         "trust_remote_code": True,
         "cache_dir": model_args.cache_dir,
@@ -132,8 +129,6 @@ def load_model_and_tokenizer(
                 })
 
         if stage == "ppo": # load reward model
-            assert is_trainable, "PPO stage cannot be performed at evaluation."
-            assert model_args.reward_model is not None, "Reward model is necessary for PPO training."
             logger.info("Load reward model from {}".format(model_args.reward_model))
             model.pretrained_model.load_adapter(model_args.reward_model, "reward", is_trainable=False)
             assert load_valuehead_params(model, model_args.reward_model), "Reward model is not correctly loaded."

src/llmtuner/tuner/core/parser.py

@@ -19,7 +19,7 @@ from llmtuner.hparams import (
 logger = get_logger(__name__)
 
 
-def _parse_args(parser: HfArgumentParser, args: Optional[Dict[str, Any]] = None):
+def _parse_args(parser: HfArgumentParser, args: Optional[Dict[str, Any]] = None) -> Tuple[Any]:
     if args is not None:
         return parser.parse_dict(args)
     elif len(sys.argv) == 2 and sys.argv[1].endswith(".yaml"):
@@ -32,26 +32,53 @@ def _parse_args(parser: HfArgumentParser, args: Optional[Dict[str, Any]] = None)
 
 def parse_train_args(
     args: Optional[Dict[str, Any]] = None
-) -> Tuple[ModelArguments, DataArguments, Seq2SeqTrainingArguments, FinetuningArguments, GeneralArguments]:
+) -> Tuple[
+    ModelArguments,
+    DataArguments,
+    Seq2SeqTrainingArguments,
+    FinetuningArguments,
+    GeneratingArguments,
+    GeneralArguments
+]:
     parser = HfArgumentParser((
-        ModelArguments, DataArguments, Seq2SeqTrainingArguments, FinetuningArguments, GeneralArguments
+        ModelArguments,
+        DataArguments,
+        Seq2SeqTrainingArguments,
+        FinetuningArguments,
+        GeneratingArguments,
+        GeneralArguments
     ))
     return _parse_args(parser, args)
 
 
 def parse_infer_args(
     args: Optional[Dict[str, Any]] = None
-) -> Tuple[ModelArguments, DataArguments, FinetuningArguments, GeneratingArguments]:
+) -> Tuple[
+    ModelArguments,
+    DataArguments,
+    FinetuningArguments,
+    GeneratingArguments
+]:
     parser = HfArgumentParser((
-        ModelArguments, DataArguments, FinetuningArguments, GeneratingArguments
+        ModelArguments,
+        DataArguments,
+        FinetuningArguments,
+        GeneratingArguments
     ))
     return _parse_args(parser, args)
 
 
 def get_train_args(
     args: Optional[Dict[str, Any]] = None
-) -> Tuple[ModelArguments, DataArguments, Seq2SeqTrainingArguments, FinetuningArguments, GeneralArguments]:
+) -> Tuple[
-    model_args, data_args, training_args, finetuning_args, general_args = parse_train_args(args)
+    ModelArguments,
+    DataArguments,
+    Seq2SeqTrainingArguments,
+    FinetuningArguments,
+    GeneratingArguments,
+    GeneralArguments
+]:
+    model_args, data_args, training_args, finetuning_args, generating_args, general_args = parse_train_args(args)
 
     # Setup logging
     if training_args.should_log:
@@ -68,7 +95,7 @@ def get_train_args(
     data_args.init_for_training()
 
     if general_args.stage != "sft" and training_args.predict_with_generate:
-        raise ValueError("`predict_with_generate` cannot be set as True at PT, RM and PPO stages.")
+        raise ValueError("`predict_with_generate` cannot be set as True except SFT.")
 
     if training_args.do_train and training_args.predict_with_generate:
         raise ValueError("`predict_with_generate` cannot be set as True while training.")
@@ -76,6 +103,15 @@ def get_train_args(
     if general_args.stage == "sft" and training_args.do_predict and not training_args.predict_with_generate:
         raise ValueError("Please enable `predict_with_generate` to save model predictions.")
 
+    if general_args.stage in ["rm", "ppo"] and finetuning_args.finetuning_type != "lora":
+        raise ValueError("RM and PPO training can only be performed with the LoRA method.")
+
+    if general_args.stage in ["ppo", "dpo"] and not training_args.do_train:
+        raise ValueError("PPO and DPO stage can only be performed at training.")
+
+    if general_args.stage == "ppo" and model_args.reward_model is None:
+        raise ValueError("Reward model is necessary for PPO training.")
+
     if training_args.max_steps == -1 and data_args.streaming:
         raise ValueError("Please specify `max_steps` in streaming mode.")
 
@@ -133,12 +169,17 @@ def get_train_args(
     # Set seed before initializing model.
     transformers.set_seed(training_args.seed)
 
-    return model_args, data_args, training_args, finetuning_args, general_args
+    return model_args, data_args, training_args, finetuning_args, generating_args, general_args
 
 
 def get_infer_args(
     args: Optional[Dict[str, Any]] = None
-) -> Tuple[ModelArguments, DataArguments, FinetuningArguments, GeneratingArguments]:
+) -> Tuple[
+    ModelArguments,
+    DataArguments,
+    FinetuningArguments,
+    GeneratingArguments
+]:
     model_args, data_args, finetuning_args, generating_args = parse_infer_args(args)
 
     if model_args.quantization_bit is not None and finetuning_args.finetuning_type != "lora":

src/llmtuner/tuner/core/trainer.py

@@ -13,26 +13,25 @@ from llmtuner.extras.logging import get_logger
 from llmtuner.extras.save_and_load import get_state_dict, load_trainable_params
 
 if TYPE_CHECKING:
+    from transformers import PreTrainedTokenizer, Seq2SeqTrainingArguments, TrainerState
     from llmtuner.hparams import FinetuningArguments
 
 
 logger = get_logger(__name__)
 
 
-class PeftTrainer(Seq2SeqTrainer):
+class PeftModelMixin:
     r"""
-    Inherits Seq2SeqTrainer to support parameter-efficient checkpoints.
+    Patches the save and load methods in Hugging Face Trainer for PeftModel and ModelWithValueHead.
     """
 
-    def __init__(self, finetuning_args: "FinetuningArguments", **kwargs):
+    def __init__(self) -> None: # for type checking
-        super().__init__(**kwargs)
+        self.model: PreTrainedModel = None
-        self.finetuning_args = finetuning_args
+        self.tokenizer: "PreTrainedTokenizer" = None
-        self._remove_log()
+        self.args: "Seq2SeqTrainingArguments" = None
-
+        self.finetuning_args: "FinetuningArguments" = None
-    def _remove_log(self):
+        self.state: "TrainerState" = None
-        if self.is_world_process_zero() and os.path.exists(os.path.join(self.args.output_dir, "trainer_log.jsonl")):
+        raise AssertionError("Mixin should not be initialized.")
-            logger.warning("Previous log file in this folder will be deleted.")
-            os.remove(os.path.join(self.args.output_dir, "trainer_log.jsonl"))
 
     def _save(self, output_dir: Optional[str] = None, state_dict: Optional[Dict[str, torch.Tensor]] = None) -> None:
         r"""
@@ -96,3 +95,13 @@ class PeftTrainer(Seq2SeqTrainer):
             model.load_adapter(self.state.best_model_checkpoint, model.active_adapter)
         else: # freeze/full-tuning
             load_trainable_params(model, self.state.best_model_checkpoint)
+
+
+class PeftTrainer(PeftModelMixin, Seq2SeqTrainer):
+    r"""
+    Inherits Seq2SeqTrainer to support parameter-efficient checkpoints.
+    """
+
+    def __init__(self, finetuning_args: "FinetuningArguments", **kwargs):
+        Seq2SeqTrainer.__init__(self, **kwargs)
+        self.finetuning_args = finetuning_args

src/llmtuner/tuner/dpo/__init__.py

@@ -0,0 +1 @@
+from llmtuner.tuner.dpo.workflow import run_dpo

src/llmtuner/tuner/dpo/collator.py

@@ -0,0 +1,51 @@
+import torch
+from dataclasses import dataclass
+from typing import Any, Dict, List, Sequence, Tuple
+from transformers import DataCollatorForSeq2Seq
+
+
+@dataclass
+class DPODataCollatorWithPadding(DataCollatorForSeq2Seq):
+    r"""
+    Data collator for pairwise data.
+    """
+
+    def _pad_labels(self, batch: torch.Tensor, positions: List[Tuple[int, int]]) -> torch.Tensor:
+        padded_labels = []
+        for feature, (prompt_len, answer_len) in zip(batch, positions):
+            if self.tokenizer.padding_side == "left":
+                start, end = feature.size(0) - answer_len, feature.size(0)
+            else:
+                start, end = prompt_len, answer_len
+            padded_tensor = self.label_pad_token_id * torch.ones_like(feature)
+            padded_tensor[start:end] = feature[start:end]
+            padded_labels.append(padded_tensor)
+        return torch.stack(padded_labels, dim=0).contiguous() # in contiguous memory
+
+    def __call__(self, features: Sequence[Dict[str, Any]]) -> Dict[str, torch.Tensor]:
+        r"""
+        Pads batched data to the longest sequence in the batch.
+
+        We generate 2 * n examples where the first n examples represent chosen examples and
+        the last n examples represent rejected examples.
+        """
+        concatenated_features = []
+        label_positions = []
+        for key in ("chosen_ids", "rejected_ids"):
+            for feature in features:
+                prompt_len, answer_len = len(feature["prompt_ids"]), len(feature[key])
+                concatenated_features.append({
+                    "input_ids": feature["prompt_ids"] + feature[key],
+                    "attention_mask": [1] * (prompt_len + answer_len)
+                })
+                label_positions.append((prompt_len, answer_len))
+
+        batch = self.tokenizer.pad(
+            concatenated_features,
+            padding=self.padding,
+            max_length=self.max_length,
+            pad_to_multiple_of=self.pad_to_multiple_of,
+            return_tensors=self.return_tensors,
+        )
+        batch["labels"] = self._pad_labels(batch["input_ids"], label_positions)
+        return batch

src/llmtuner/tuner/dpo/trainer.py

@@ -0,0 +1,75 @@
+import torch
+from collections import defaultdict
+from peft import PeftModel
+from typing import TYPE_CHECKING, Dict, Optional, Tuple, Union
+from transformers import Trainer
+from trl import DPOTrainer
+
+from llmtuner.extras.constants import IGNORE_INDEX
+from llmtuner.tuner.core.trainer import PeftModelMixin
+
+if TYPE_CHECKING:
+    from transformers import PreTrainedModel
+    from llmtuner.hparams import FinetuningArguments, GeneratingArguments
+
+
+class DPOPeftTrainer(PeftModelMixin, DPOTrainer):
+
+    def __init__(
+        self,
+        finetuning_args: "FinetuningArguments",
+        generating_args: "GeneratingArguments",
+        ref_model: Optional[Union["PreTrainedModel", torch.nn.Module]] = None,
+        **kwargs
+    ):
+        self.finetuning_args = finetuning_args
+        self.generating_args = generating_args
+        self.ref_model = ref_model
+        self.use_dpo_data_collator = True # hack to avoid warning
+        self.label_pad_token_id = IGNORE_INDEX
+        self.padding_value = 0
+        self.beta = finetuning_args.dpo_beta
+        self._stored_metrics = defaultdict(lambda: defaultdict(list))
+
+        Trainer.__init__(self, **kwargs)
+        if ref_model is not None:
+            if hasattr(self, "accelerator"):
+                self.ref_model = self.accelerator.prepare_model(self.ref_model, evaluation_mode=True)
+            else:
+                raise AttributeError("Please update `transformers`.")
+
+    def concatenated_forward(
+        self,
+        model: Optional[torch.nn.Module] = None,
+        batch: Optional[Dict[str, torch.Tensor]] = None
+    ) -> Tuple[torch.FloatTensor, torch.FloatTensor, torch.FloatTensor, torch.FloatTensor]:
+        unwrapped_model: "PreTrainedModel" = self.accelerator.unwrap_model(self.model)
+        if not torch.is_grad_enabled():
+            unwrapped_model.gradient_checkpointing_disable()
+
+        if model is None and isinstance(unwrapped_model, PeftModel): # peft model has no ref_model
+            with unwrapped_model.disable_adapter():
+                all_logits: torch.Tensor = self.model(
+                    batch["input_ids"],
+                    attention_mask=batch["attention_mask"],
+                    return_dict=True
+                ).logits.to(torch.float32)
+        else:
+            all_logits: torch.Tensor = model(
+                batch["input_ids"],
+                attention_mask=batch["attention_mask"],
+                return_dict=True
+            ).logits.to(torch.float32)
+
+        if not torch.is_grad_enabled():
+            unwrapped_model.gradient_checkpointing_enable()
+
+        all_logps = self._get_batch_logps(
+            all_logits,
+            batch["labels"],
+            average_log_prob=False
+        )
+        batch_size = batch["input_ids"].size(0) // 2
+        chosen_logps, rejected_logps = all_logps.split(batch_size, dim=0)
+        chosen_logits, rejected_logits = all_logits.split(batch_size, dim=0)
+        return chosen_logps, rejected_logps, chosen_logits, rejected_logits

src/llmtuner/tuner/dpo/workflow.py

@@ -0,0 +1,59 @@
+# Inspired by: https://github.com/huggingface/trl/blob/main/examples/research_projects/stack_llama_2/scripts/dpo_llama2.py
+
+from copy import deepcopy
+from peft import PeftModel
+from typing import TYPE_CHECKING, Optional, List
+
+from llmtuner.dsets import get_dataset, preprocess_dataset, split_dataset
+from llmtuner.extras.constants import IGNORE_INDEX
+from llmtuner.extras.ploting import plot_loss
+from llmtuner.tuner.core import load_model_and_tokenizer
+from llmtuner.tuner.dpo.collator import DPODataCollatorWithPadding
+from llmtuner.tuner.dpo.trainer import DPOPeftTrainer
+
+if TYPE_CHECKING:
+    from transformers import Seq2SeqTrainingArguments, TrainerCallback
+    from llmtuner.hparams import ModelArguments, DataArguments, FinetuningArguments, GeneratingArguments
+
+
+def run_dpo(
+    model_args: "ModelArguments",
+    data_args: "DataArguments",
+    training_args: "Seq2SeqTrainingArguments",
+    finetuning_args: "FinetuningArguments",
+    generating_args: "GeneratingArguments",
+    callbacks: Optional[List["TrainerCallback"]] = None
+):
+    dataset = get_dataset(model_args, data_args)
+    model, tokenizer = load_model_and_tokenizer(model_args, finetuning_args, training_args.do_train, stage="sft")
+    dataset = preprocess_dataset(dataset, tokenizer, data_args, training_args, stage="rm")
+    data_collator = DPODataCollatorWithPadding(
+        tokenizer=tokenizer,
+        label_pad_token_id=IGNORE_INDEX if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id
+    )
+
+    training_args.remove_unused_columns = False # important for pairwise dataset
+    ref_model = deepcopy(model) if not isinstance(model, PeftModel) else None
+
+    # Initialize our Trainer
+    trainer = DPOPeftTrainer(
+        finetuning_args=finetuning_args,
+        generating_args=generating_args,
+        ref_model=ref_model,
+        model=model,
+        args=training_args,
+        tokenizer=tokenizer,
+        data_collator=data_collator,
+        callbacks=callbacks,
+        **split_dataset(dataset, data_args, training_args)
+    )
+
+    # Training
+    if training_args.do_train:
+        train_result = trainer.train()
+        trainer.log_metrics("train", train_result.metrics)
+        trainer.save_metrics("train", train_result.metrics)
+        trainer.save_state()
+        trainer.save_model()
+        if trainer.is_world_process_zero() and model_args.plot_loss:
+            plot_loss(training_args.output_dir, keys=["loss", "eval_loss"])

src/llmtuner/tuner/ppo/trainer.py

@@ -10,7 +10,7 @@ from trl import PPOTrainer
 from trl.core import LengthSampler
 
 from llmtuner.extras.logging import get_logger
-from llmtuner.extras.misc import AverageMeter, count_parameters, get_logits_processor
+from llmtuner.extras.misc import AverageMeter, count_parameters, get_logits_processor, get_stopping_criteria
 from llmtuner.tuner.core.trainer import PeftTrainer
 from llmtuner.tuner.ppo.utils import cast_layernorm_dtype, replace_model
 
@@ -18,7 +18,7 @@ if TYPE_CHECKING:
     from transformers import Seq2SeqTrainingArguments
     from trl import AutoModelForCausalLMWithValueHead
     from llmtuner.extras.callbacks import LogCallback
-    from llmtuner.hparams import FinetuningArguments
+    from llmtuner.hparams import FinetuningArguments, GeneratingArguments
 
 
 logger = get_logger(__name__)
@@ -33,16 +33,17 @@ class PPOPeftTrainer(PPOTrainer, PeftTrainer):
         self,
         training_args: "Seq2SeqTrainingArguments",
         finetuning_args: "FinetuningArguments",
+        generating_args: "GeneratingArguments",
         callbacks: List["LogCallback"],
         **kwargs
     ):
         PPOTrainer.__init__(self, **kwargs)
         self.args = training_args
         self.finetuning_args = finetuning_args
+        self.generating_args = generating_args
         self.log_callback = callbacks[0]
         self.state = TrainerState()
         self.control = TrainerControl()
-        self._remove_log()
 
     def ppo_train(self, max_target_length: int) -> None:
         r"""
@@ -72,14 +73,10 @@ class PPOPeftTrainer(PPOTrainer, PeftTrainer):
             logger.info(f"  Number of trainable parameters = {count_parameters(self.model)[0]}")
 
         # Keyword arguments for `model.generate`
-        gen_kwargs = {
+        gen_kwargs = self.generating_args.to_dict()
-            "top_k": 0.0,
+        gen_kwargs["logits_processor"] = get_logits_processor()
-            "top_p": 1.0,
+        gen_kwargs["stopping_criteria"] = get_stopping_criteria(self.tokenizer.additional_special_tokens_ids)
-            "do_sample": True,
+
-            "pad_token_id": self.tokenizer.pad_token_id,
-            "eos_token_id": self.tokenizer.eos_token_id,
-            "logits_processor": get_logits_processor()
-        }
         length_sampler = LengthSampler(max_target_length // 2, max_target_length)
         unwrapped_model: "AutoModelForCausalLMWithValueHead" = self.accelerator.unwrap_model(self.model)
 

src/llmtuner/tuner/ppo/workflow.py

@@ -1,11 +1,9 @@
-# Inspired by:
+# Inspired by: https://github.com/lvwerra/trl/blob/main/examples/research_projects/stack_llama/scripts/rl_training.py
-# https://github.com/lvwerra/trl/blob/main/examples/research_projects/stack_llama/scripts/rl_training.py
 
 import math
-from typing import TYPE_CHECKING
 from trl import PPOConfig
 from torch.optim import AdamW
-from typing import Optional, List
+from typing import TYPE_CHECKING, Optional, List
 from transformers import DataCollatorForSeq2Seq
 from transformers.optimization import get_scheduler
 
@@ -16,7 +14,7 @@ from llmtuner.tuner.ppo.trainer import PPOPeftTrainer
 
 if TYPE_CHECKING:
     from transformers import Seq2SeqTrainingArguments, TrainerCallback
-    from llmtuner.hparams import ModelArguments, DataArguments, FinetuningArguments
+    from llmtuner.hparams import ModelArguments, DataArguments, FinetuningArguments, GeneratingArguments
 
 
 def run_ppo(
@@ -24,6 +22,7 @@ def run_ppo(
     data_args: "DataArguments",
     training_args: "Seq2SeqTrainingArguments",
     finetuning_args: "FinetuningArguments",
+    generating_args: "GeneratingArguments",
     callbacks: Optional[List["TrainerCallback"]] = None
 ):
     dataset = get_dataset(model_args, data_args)
@@ -42,8 +41,9 @@ def run_ppo(
     )
 
     optimizer = AdamW(filter(lambda p: p.requires_grad, model.parameters()), lr=training_args.learning_rate)
-    total_train_batch_size = \
+    total_train_batch_size = (
         training_args.per_device_train_batch_size * training_args.gradient_accumulation_steps * training_args.world_size
+    )
     num_training_steps = training_args.num_train_epochs * math.ceil(len(dataset) / total_train_batch_size)
     lr_scheduler = get_scheduler(
         training_args.lr_scheduler_type,
@@ -56,6 +56,7 @@ def run_ppo(
     ppo_trainer = PPOPeftTrainer(
         training_args=training_args,
         finetuning_args=finetuning_args,
+        generating_args=generating_args,
         callbacks=callbacks,
         config=ppo_config,
         model=model,
@@ -67,8 +68,10 @@ def run_ppo(
         lr_scheduler=lr_scheduler
     )
 
-    ppo_trainer.ppo_train(max_target_length=data_args.max_target_length)
+    # Training
-    ppo_trainer.save_model()
+    if training_args.do_train:
-    ppo_trainer.save_state() # must be after save_model
+        ppo_trainer.ppo_train(max_target_length=data_args.max_target_length)
-    if ppo_trainer.is_world_process_zero() and model_args.plot_loss:
+        ppo_trainer.save_model()
-        plot_loss(training_args.output_dir, keys=["loss", "reward"])
+        ppo_trainer.save_state() # must be called after save_model to have a folder
+        if ppo_trainer.is_world_process_zero() and model_args.plot_loss:
+            plot_loss(training_args.output_dir, keys=["loss", "reward"])

src/llmtuner/tuner/pt/workflow.py

@@ -2,10 +2,9 @@
 
 import math
 from typing import TYPE_CHECKING, Optional, List
-from transformers import DataCollatorForSeq2Seq
+from transformers import DataCollatorForLanguageModeling
 
 from llmtuner.dsets import get_dataset, preprocess_dataset, split_dataset
-from llmtuner.extras.constants import IGNORE_INDEX
 from llmtuner.extras.ploting import plot_loss
 from llmtuner.tuner.core import load_model_and_tokenizer
 from llmtuner.tuner.core.trainer import PeftTrainer
@@ -25,10 +24,7 @@ def run_pt(
     dataset = get_dataset(model_args, data_args)
     model, tokenizer = load_model_and_tokenizer(model_args, finetuning_args, training_args.do_train, stage="pt")
     dataset = preprocess_dataset(dataset, tokenizer, data_args, training_args, stage="pt")
-    data_collator = DataCollatorForSeq2Seq(
+    data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)
-        tokenizer=tokenizer,
-        label_pad_token_id=IGNORE_INDEX if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id
-    )
 
     # Initialize our Trainer
     trainer = PeftTrainer(

src/llmtuner/tuner/rm/collator.py

@@ -1,8 +1,10 @@
 import torch
+from dataclasses import dataclass
 from typing import Any, Dict, Sequence
 from transformers import DataCollatorWithPadding
 
 
+@dataclass
 class PairwiseDataCollatorWithPadding(DataCollatorWithPadding):
     r"""
     Data collator for pairwise data.
@@ -16,7 +18,10 @@ class PairwiseDataCollatorWithPadding(DataCollatorWithPadding):
         the last n examples represent rejected examples.
         """
         features = [
-            {"input_ids": feature[key], "attention_mask": [1] * len(feature[key])}
+            {
-            for key in ("accept_ids", "reject_ids") for feature in features
+                "input_ids": feature["prompt_ids"] + feature[key],
+                "attention_mask": [1] * (len(feature["prompt_ids"]) + len(feature[key]))
+            }
+            for key in ("chosen_ids", "rejected_ids") for feature in features
         ]
         return super().__call__(features)

src/llmtuner/tuner/sft/trainer.py

@@ -79,7 +79,7 @@ class Seq2SeqPeftTrainer(PeftTrainer):
 
         padded_tensor = pad_token_id * torch.ones_like(tgt_tensor)
         padded_tensor[:, -src_tensor.shape[-1]:] = src_tensor # adopt left-padding
-        return padded_tensor.contiguous()
+        return padded_tensor.contiguous() # in contiguous memory
 
     def save_predictions(
         self,

src/llmtuner/tuner/sft/workflow.py

@@ -5,7 +5,7 @@ from transformers import DataCollatorForSeq2Seq
 
 from llmtuner.dsets import get_dataset, preprocess_dataset, split_dataset
 from llmtuner.extras.constants import IGNORE_INDEX
-from llmtuner.extras.misc import get_logits_processor
+from llmtuner.extras.misc import get_logits_processor, get_stopping_criteria
 from llmtuner.extras.ploting import plot_loss
 from llmtuner.tuner.core import load_model_and_tokenizer
 from llmtuner.tuner.sft.metric import ComputeMetrics
@@ -13,7 +13,7 @@ from llmtuner.tuner.sft.trainer import Seq2SeqPeftTrainer
 
 if TYPE_CHECKING:
     from transformers import Seq2SeqTrainingArguments, TrainerCallback
-    from llmtuner.hparams import ModelArguments, DataArguments, FinetuningArguments
+    from llmtuner.hparams import ModelArguments, DataArguments, FinetuningArguments, GeneratingArguments
 
 
 def run_sft(
@@ -21,6 +21,7 @@ def run_sft(
     data_args: "DataArguments",
     training_args: "Seq2SeqTrainingArguments",
     finetuning_args: "FinetuningArguments",
+    generating_args: "GeneratingArguments",
     callbacks: Optional[List["TrainerCallback"]] = None
 ):
     dataset = get_dataset(model_args, data_args)
@@ -50,13 +51,9 @@ def run_sft(
     )
 
     # Keyword arguments for `model.generate`
-    gen_kwargs = {
+    gen_kwargs = generating_args.to_dict()
-        "do_sample": True,
+    gen_kwargs["logits_processor"] = get_logits_processor()
-        "top_p": 0.7,
+    gen_kwargs["stopping_criteria"] = get_stopping_criteria(tokenizer.additional_special_tokens_ids)
-        "max_new_tokens": data_args.max_target_length + 1,
-        "temperature": 0.95,
-        "logits_processor": get_logits_processor()
-    }
 
     # Training
     if training_args.do_train:

src/llmtuner/tuner/tune.py

@@ -1,35 +1,47 @@
 from typing import TYPE_CHECKING, Any, Dict, List, Optional
 
 from llmtuner.extras.callbacks import LogCallback
+from llmtuner.extras.logging import get_logger
 from llmtuner.tuner.core import get_train_args, load_model_and_tokenizer
 from llmtuner.tuner.pt import run_pt
 from llmtuner.tuner.sft import run_sft
 from llmtuner.tuner.rm import run_rm
 from llmtuner.tuner.ppo import run_ppo
+from llmtuner.tuner.dpo import run_dpo
 
 if TYPE_CHECKING:
     from transformers import TrainerCallback
 
 
+logger = get_logger(__name__)
+
+
 def run_exp(args: Optional[Dict[str, Any]] = None, callbacks: Optional[List["TrainerCallback"]] = None):
-    model_args, data_args, training_args, finetuning_args, general_args = get_train_args(args)
+    model_args, data_args, training_args, finetuning_args, generating_args, general_args = get_train_args(args)
-    callbacks = [LogCallback()] if callbacks is None else callbacks
+    callbacks = [LogCallback()] if callbacks is None else callbacks + [LogCallback()]
 
     if general_args.stage == "pt":
         run_pt(model_args, data_args, training_args, finetuning_args, callbacks)
     elif general_args.stage == "sft":
-        run_sft(model_args, data_args, training_args, finetuning_args, callbacks)
+        run_sft(model_args, data_args, training_args, finetuning_args, generating_args, callbacks)
     elif general_args.stage == "rm":
         run_rm(model_args, data_args, training_args, finetuning_args, callbacks)
     elif general_args.stage == "ppo":
-        run_ppo(model_args, data_args, training_args, finetuning_args, callbacks)
+        run_ppo(model_args, data_args, training_args, finetuning_args, generating_args, callbacks)
+    elif general_args.stage == "dpo":
+        run_dpo(model_args, data_args, training_args, finetuning_args, callbacks)
+    else:
+        raise ValueError("Unknown task.")
 
 
 def export_model(args: Optional[Dict[str, Any]] = None, max_shard_size: Optional[str] = "10GB"):
     model_args, _, training_args, finetuning_args, _ = get_train_args(args)
     model, tokenizer = load_model_and_tokenizer(model_args, finetuning_args)
     model.save_pretrained(training_args.output_dir, max_shard_size=max_shard_size)
-    tokenizer.save_pretrained(training_args.output_dir)
+    try:
+        tokenizer.save_pretrained(training_args.output_dir)
+    except:
+        logger.warning("Cannot save tokenizer, please copy the files manually.")
 
 
 if __name__ == "__main__":

src/llmtuner/webui/runner.py

@@ -4,7 +4,7 @@ import threading
 import time
 import transformers
 from transformers.trainer import TRAINING_ARGS_NAME
-from typing import Generator, List, Optional, Tuple
+from typing import Generator, List, Tuple
 
 from llmtuner.extras.callbacks import LogCallback
 from llmtuner.extras.constants import DEFAULT_MODULE