release v0.6.0 (real)

Former-commit-id: 34e06bf408ccd21e674f896703f1c7b62e97e1ca

.dockerignore

@@ -0,0 +1,11 @@
+.vscode
+.git
+.github
+.venv
+cache
+data
+examples
+.dockerignore
+.gitattributes
+.gitignore
+Dockerfile

.github/CONTRIBUTING.md

@@ -0,0 +1,21 @@
+# Contributing to LLaMA Factory
+
+Everyone is welcome to contribute, and we value everybody's contribution. Code contributions are not the only way to help the community. Answering questions, helping others, and improving the documentation are also immensely valuable.
+
+It also helps us if you spread the word! Reference the library in blog posts about the awesome projects it made possible, shout out on Twitter every time it has helped you, or simply ⭐️ the repository to say thank you.
+
+However you choose to contribute, please be mindful and respect our [code of conduct](CODE_OF_CONDUCT.md).
+
+**This guide was heavily inspired by [transformers guide to contributing](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md).**
+
+## Ways to contribute
+
+There are several ways you can contribute to LLaMA Factory:
+
+* Fix outstanding issues with the existing code.
+* Submit issues related to bugs or desired new features.
+* Contribute to the examples or to the documentation.
+
+### Style guide
+
+LLaMA Factory follows the [Google Python Style Guide](https://google.github.io/styleguide/pyguide.html), check it for details.

.github/PULL_REQUEST_TEMPLATE.md

@@ -0,0 +1,7 @@
+# What does this PR do?
+
+Fixes # (issue)
+
+## Before submitting
+
+- [ ] Did you read the [contributor guideline](https://github.com/hiyouga/LLaMA-Factory/blob/main/.github/CONTRIBUTING.md)?

.github/SECURITY.md

@@ -0,0 +1,7 @@
+# Reporting Security Issues
+
+To report a security issue, please use the GitHub Security Advisory ["Report a Vulnerability"](https://github.com/electron/electron/security/advisories/new) tab.
+
+We will send a response indicating the next steps in handling your report. After the initial reply to your report, the security team will keep you informed of the progress towards a fix and full announcement, and may ask for additional information or guidance.
+
+Report security bugs in third-party modules to the person or team maintaining the module.

.github/workflows/tests.yml

@@ -0,0 +1,29 @@
+name: tests
+
+on:
+  push:
+    branches: [ "main" ]
+  pull_request:
+    branches: [ "main" ]
+
+jobs:
+  check_code_quality:
+
+    runs-on: ubuntu-latest
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.8"
+
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          python -m pip install ruff
+
+      - name: Check quality
+        run: |
+            make style && make quality

CITATION.cff

@@ -0,0 +1,37 @@
+cff-version: 1.2.0
+date-released: 2024-03
+message: "If you use this software, please cite it as below."
+authors:
+- family-names: "Zheng"
+  given-names: "Yaowei"
+- family-names: "Zhang"
+  given-names: "Richong"
+- family-names: "Zhang"
+  given-names: "Junhao"
+- family-names: "Ye"
+  given-names: "Yanhan"
+- family-names: "Luo"
+  given-names: "Zheyan"
+- family-names: "Ma"
+  given-names: "Yongqiang"
+title: "LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models"
+url: "https://arxiv.org/abs/2403.13372"
+preferred-citation:
+  type: article
+  authors:
+    - family-names: "Zheng"
+      given-names: "Yaowei"
+    - family-names: "Zhang"
+      given-names: "Richong"
+    - family-names: "Zhang"
+      given-names: "Junhao"
+    - family-names: "Ye"
+      given-names: "Yanhan"
+    - family-names: "Luo"
+      given-names: "Zheyan"
+    - family-names: "Ma"
+      given-names: "Yongqiang"
+  journal: "arXiv preprint arXiv:2403.13372"
+  title: "LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models"
+  url: "https://arxiv.org/abs/2403.13372"
+  year: 2024

Dockerfile

@@ -0,0 +1,14 @@
+FROM nvcr.io/nvidia/pytorch:24.01-py3
+
+WORKDIR /app
+
+COPY requirements.txt /app/
+RUN pip install -r requirements.txt
+
+COPY . /app/
+RUN pip install -e .[deepspeed,metrics,bitsandbytes,qwen]
+
+VOLUME [ "/root/.cache/huggingface/", "/app/data", "/app/output" ]
+EXPOSE 7860
+
+CMD [ "python", "src/train_web.py" ]

Makefile

@@ -1,11 +1,11 @@
 .PHONY: quality style
 
-check_dirs := src tests
+check_dirs := scripts src tests
 
 quality:
-	black --check $(check_dirs)
-	ruff $(check_dirs)
+	ruff check $(check_dirs)
+	ruff format --check $(check_dirs)
 
 style:
-	black $(check_dirs)
-	ruff $(check_dirs) --fix
+	ruff check $(check_dirs) --fix
+	ruff format $(check_dirs)

README.md

@@ -5,27 +5,30 @@
 [![GitHub last commit](https://img.shields.io/github/last-commit/hiyouga/LLaMA-Factory)](https://github.com/hiyouga/LLaMA-Factory/commits/main)
 [![PyPI](https://img.shields.io/pypi/v/llmtuner)](https://pypi.org/project/llmtuner/)
 [![Downloads](https://static.pepy.tech/badge/llmtuner)](https://pypi.org/project/llmtuner/)
+[![Citation](https://img.shields.io/badge/citation-26-green)](#projects-using-llama-factory)
 [![GitHub pull request](https://img.shields.io/badge/PRs-welcome-blue)](https://github.com/hiyouga/LLaMA-Factory/pulls)
 [![Discord](https://dcbadge.vercel.app/api/server/rKfvV9r9FK?compact=true&style=flat)](https://discord.gg/rKfvV9r9FK)
-[![Spaces](https://img.shields.io/badge/🤗-Open%20In%20Spaces-blue)](https://huggingface.co/spaces/hiyouga/LLaMA-Board)
-[![Studios](https://img.shields.io/badge/ModelScope-Open%20In%20Studios-blue)](https://modelscope.cn/studios/hiyouga/LLaMA-Board)
+[![Twitter](https://img.shields.io/twitter/follow/llamafactory_ai)](https://twitter.com/llamafactory_ai)
+[![Spaces](https://img.shields.io/badge/🤗-Open%20in%20Spaces-blue)](https://huggingface.co/spaces/hiyouga/LLaMA-Board)
+[![Studios](https://img.shields.io/badge/ModelScope-Open%20in%20Studios-blue)](https://modelscope.cn/studios/hiyouga/LLaMA-Board)
+[![Open in Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1eRTPn37ltBbYsISy9Aw2NuI2Aq5CQrD9?usp=sharing)
 
 👋 Join our [WeChat](assets/wechat.jpg).
 
 \[ English | [中文](README_zh.md) \]
 
-## LLaMA Board: A One-stop Web UI for Getting Started with LLaMA Factory
+**Fine-tuning a large language model can be easy as...**
 
-Preview LLaMA Board at **[🤗 Spaces](https://huggingface.co/spaces/hiyouga/LLaMA-Board)** or **[ModelScope](https://modelscope.cn/studios/hiyouga/LLaMA-Board)**.
+https://github.com/hiyouga/LLaMA-Factory/assets/16256802/9840a653-7e9c-41c8-ae89-7ace5698baf6
 
-Launch LLaMA Board via `CUDA_VISIBLE_DEVICES=0 python src/train_web.py`. (multiple GPUs are not supported yet in this mode)
+Choose your path:
 
-Here is an example of altering the self-cognition of an instruction-tuned language model within 10 minutes on a single GPU.
-
-https://github.com/hiyouga/LLaMA-Factory/assets/16256802/6ba60acc-e2e2-4bec-b846-2d88920d5ba1
+- **Colab**: https://colab.research.google.com/drive/1eRTPn37ltBbYsISy9Aw2NuI2Aq5CQrD9?usp=sharing
+- **Local machine**: Please refer to [usage](#getting-started)
 
 ## Table of Contents
 
+- [Features](#features)
 - [Benchmark](#benchmark)
 - [Changelog](#changelog)
 - [Supported Models](#supported-models)
@@ -38,6 +41,16 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/6ba60acc-e2e2-4bec-b846
 - [Citation](#citation)
 - [Acknowledgement](#acknowledgement)
 
+## Features
+
+- **Various models**: LLaMA, Mistral, Mixtral-MoE, Qwen, Yi, Gemma, Baichuan, ChatGLM, Phi, etc.
+- **Integrated methods**: (Continuous) pre-training, supervised fine-tuning, reward modeling, PPO and DPO.
+- **Scalable resources**: 32-bit full-tuning, 16-bit freeze-tuning, 16-bit LoRA and 2/4/8-bit QLoRA via AQLM/AWQ/GPTQ/LLM.int8.
+- **Advanced algorithms**: GaLore, DoRA, LongLoRA, LLaMA Pro, LoRA+, LoftQ and Agent tuning.
+- **Practical tricks**: FlashAttention-2, Unsloth, RoPE scaling, NEFTune and rsLoRA.
+- **Experiment monitors**: LlamaBoard, TensorBoard, Wandb, MLflow, etc.
+- **Faster inference**: OpenAI-style API, Gradio UI and CLI with vLLM worker.
+
 ## Benchmark
 
 Compared to ChatGLM's [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/ptuning), LLaMA-Factory's LoRA tuning offers up to **3.7 times faster** training speed with a better Rouge score on the advertising text generation task. By leveraging 4-bit quantization technique, LLaMA-Factory's QLoRA further improves the efficiency regarding the GPU memory.
@@ -55,14 +68,30 @@ Compared to ChatGLM's [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/
 
 ## Changelog
 
-[24/01/18] We supported **agent tuning** for most models, equipping model with tool using abilities by fine-tuning with `--dataset glaive_toolcall`.
+[24/03/21] Our paper "[LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models](https://arxiv.org/abs/2403.13372)" is available at arXiv!
 
-[23/12/23] We supported **[unsloth](https://github.com/unslothai/unsloth)**'s implementation to boost LoRA tuning for the LLaMA, Mistral and Yi models. Try `--use_unsloth` argument to activate unsloth patch. It achieves 1.7x speed in our benchmark, check [this page](https://github.com/hiyouga/LLaMA-Factory/wiki/Performance-comparison) for details.
+[24/03/20] We supported **FSDP+QLoRA** that fine-tunes a 70B model on 2x24GB GPUs. See `examples/fsdp_qlora` for usage.
 
-[23/12/12] We supported fine-tuning the latest MoE model **[Mixtral 8x7B](https://huggingface.co/mistralai/Mixtral-8x7B-v0.1)** in our framework. See hardware requirement [here](#hardware-requirement).
+[24/03/13] We supported **[LoRA+](https://arxiv.org/abs/2402.12354)**. Try `loraplus_lr_ratio=16.0` to enable LoRA+ algorithm.
+
+[24/03/07] We supported gradient low-rank projection (**[GaLore](https://arxiv.org/abs/2403.03507)**) algorithm. Try `--use_galore` to use the memory-efficient optimizer.
+
+[24/03/07] We integrated **[vLLM](https://github.com/vllm-project/vllm)** for faster and concurrent inference. Try `--infer_backend vllm` to enjoy **270%** inference speed. (LoRA is not yet supported, merge it first.)
 
 <details><summary>Full Changelog</summary>
 
+[24/02/28] We supported weight-decomposed LoRA (**[DoRA](https://arxiv.org/abs/2402.09353)**). Try `--use_dora` to activate DoRA training.
+
+[24/02/15] We supported **block expansion** proposed by [LLaMA Pro](https://github.com/TencentARC/LLaMA-Pro). See `examples/extras/llama_pro` for usage.
+
+[24/02/05] Qwen1.5 (Qwen2 beta version) series models are supported in LLaMA-Factory. Check this [blog post](https://qwenlm.github.io/blog/qwen1.5/) for details.
+
+[24/01/18] We supported **agent tuning** for most models, equipping model with tool using abilities by fine-tuning with `--dataset glaive_toolcall`.
+
+[23/12/23] We supported **[unsloth](https://github.com/unslothai/unsloth)**'s implementation to boost LoRA tuning for the LLaMA, Mistral and Yi models. Try `--use_unsloth` argument to activate unsloth patch. It achieves **170%** speed in our benchmark, check [this page](https://github.com/hiyouga/LLaMA-Factory/wiki/Performance-comparison) for details.
+
+[23/12/12] We supported fine-tuning the latest MoE model **[Mixtral 8x7B](https://huggingface.co/mistralai/Mixtral-8x7B-v0.1)** in our framework. See hardware requirement [here](#hardware-requirement).
+
 [23/12/01] We supported downloading pre-trained models and datasets from the **[ModelScope Hub](https://modelscope.cn/models)** for Chinese mainland users. See [this tutorial](#use-modelscope-hub-optional) for usage.
 
 [23/10/21] We supported **[NEFTune](https://arxiv.org/abs/2310.05914)** trick for fine-tuning. Try `--neftune_noise_alpha` argument to activate NEFTune, e.g., `--neftune_noise_alpha 5`.
@@ -103,15 +132,19 @@ Compared to ChatGLM's [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/
 | [ChatGLM3](https://huggingface.co/THUDM/chatglm3-6b)     | 6B                          | query_key_value   | chatglm3  |
 | [DeepSeek (MoE)](https://huggingface.co/deepseek-ai)     | 7B/16B/67B                  | q_proj,v_proj     | deepseek  |
 | [Falcon](https://huggingface.co/tiiuae)                  | 7B/40B/180B                 | query_key_value   | falcon    |
+| [Gemma](https://huggingface.co/google)                   | 2B/7B                       | q_proj,v_proj     | gemma     |
 | [InternLM2](https://huggingface.co/internlm)             | 7B/20B                      | wqkv              | intern2   |
 | [LLaMA](https://github.com/facebookresearch/llama)       | 7B/13B/33B/65B              | q_proj,v_proj     | -         |
 | [LLaMA-2](https://huggingface.co/meta-llama)             | 7B/13B/70B                  | q_proj,v_proj     | llama2    |
 | [Mistral](https://huggingface.co/mistralai)              | 7B                          | q_proj,v_proj     | mistral   |
 | [Mixtral](https://huggingface.co/mistralai)              | 8x7B                        | q_proj,v_proj     | mistral   |
+| [OLMo](https://huggingface.co/allenai)                   | 1B/7B                       | att_proj          | olmo      |
 | [Phi-1.5/2](https://huggingface.co/microsoft)            | 1.3B/2.7B                   | q_proj,v_proj     | -         |
 | [Qwen](https://huggingface.co/Qwen)                      | 1.8B/7B/14B/72B             | c_attn            | qwen      |
+| [Qwen1.5](https://huggingface.co/Qwen)                   | 0.5B/1.8B/4B/7B/14B/72B     | q_proj,v_proj     | qwen      |
+| [StarCoder2](https://huggingface.co/bigcode)             | 3B/7B/15B                   | q_proj,v_proj     | -         |
 | [XVERSE](https://huggingface.co/xverse)                  | 7B/13B/65B                  | q_proj,v_proj     | xverse    |
-| [Yi](https://huggingface.co/01-ai)                       | 6B/34B                      | q_proj,v_proj     | yi        |
+| [Yi](https://huggingface.co/01-ai)                       | 6B/9B/34B                   | q_proj,v_proj     | yi        |
 | [Yuan](https://huggingface.co/IEITYuan)                  | 2B/51B/102B                 | q_proj,v_proj     | yuan      |
 
 > [!NOTE]
@@ -121,9 +154,11 @@ Compared to ChatGLM's [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/
 
 Please refer to [constants.py](src/llmtuner/extras/constants.py) for a full list of models we supported.
 
+You also can add a custom chat template to [template.py](src/llmtuner/data/template.py).
+
 ## Supported Training Approaches
 
-| Approach               |   Full-parameter   | Partial-parameter  |       LoRA         |       QLoRA        |
+| Approach               |     Full-tuning    |    Freeze-tuning   |       LoRA         |       QLoRA        |
 | ---------------------- | ------------------ | ------------------ | ------------------ | ------------------ |
 | Pre-Training           | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
 | Supervised Fine-Tuning | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
@@ -154,8 +189,8 @@ Please refer to [constants.py](src/llmtuner/extras/constants.py) for a full list
 
 - [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)
-- [Self-cognition (zh)](data/self_cognition.json)
+- [Alpaca GPT4 (en&zh)](https://github.com/Instruction-Tuning-with-GPT-4/GPT-4-LLM)
+- [Self Cognition (zh)](data/self_cognition.json)
 - [Open Assistant (multilingual)](https://huggingface.co/datasets/OpenAssistant/oasst1)
 - [ShareGPT (zh)](https://huggingface.co/datasets/QingyiSi/Alpaca-CoT/tree/main/Chinese-instruction-collection)
 - [Guanaco Dataset (multilingual)](https://huggingface.co/datasets/JosephusCheung/GuanacoDataset)
@@ -171,8 +206,10 @@ Please refer to [constants.py](src/llmtuner/extras/constants.py) for a full list
 - [CodeAlpaca 20k (en)](https://huggingface.co/datasets/sahil2801/CodeAlpaca-20k)
 - [Alpaca CoT (multilingual)](https://huggingface.co/datasets/QingyiSi/Alpaca-CoT)
 - [OpenOrca (en)](https://huggingface.co/datasets/Open-Orca/OpenOrca)
+- [SlimOrca (en)](https://huggingface.co/datasets/Open-Orca/SlimOrca)
 - [MathInstruct (en)](https://huggingface.co/datasets/TIGER-Lab/MathInstruct)
 - [Firefly 1.1M (zh)](https://huggingface.co/datasets/YeungNLP/firefly-train-1.1M)
+- [Wiki QA (en)](https://huggingface.co/datasets/wiki_qa)
 - [Web QA (zh)](https://huggingface.co/datasets/suolyer/webqa)
 - [WebNovel (zh)](https://huggingface.co/datasets/zxbsmk/webnovel_cn)
 - [Nectar (en)](https://huggingface.co/datasets/berkeley-nest/Nectar)
@@ -185,6 +222,16 @@ Please refer to [constants.py](src/llmtuner/extras/constants.py) for a full list
 - [LMSYS Chat 1M (en)](https://huggingface.co/datasets/lmsys/lmsys-chat-1m)
 - [Evol Instruct V2 (en)](https://huggingface.co/datasets/WizardLM/WizardLM_evol_instruct_V2_196k)
 - [Glaive Function Calling V2 (en)](https://huggingface.co/datasets/glaiveai/glaive-function-calling-v2)
+- [Cosmopedia (en)](https://huggingface.co/datasets/HuggingFaceTB/cosmopedia)
+- [Open Assistant (de)](https://huggingface.co/datasets/mayflowergmbh/oasst_de)
+- [Dolly 15k (de)](https://huggingface.co/datasets/mayflowergmbh/dolly-15k_de)
+- [Alpaca GPT4 (de)](https://huggingface.co/datasets/mayflowergmbh/alpaca-gpt4_de)
+- [OpenSchnabeltier (de)](https://huggingface.co/datasets/mayflowergmbh/openschnabeltier_de)
+- [Evol Instruct (de)](https://huggingface.co/datasets/mayflowergmbh/evol-instruct_de)
+- [Dolphin (de)](https://huggingface.co/datasets/mayflowergmbh/dolphin_de)
+- [Booksum (de)](https://huggingface.co/datasets/mayflowergmbh/booksum_de)
+- [Airoboros (de)](https://huggingface.co/datasets/mayflowergmbh/airoboros-3.0_de)
+- [Ultrachat (de)](https://huggingface.co/datasets/mayflowergmbh/ultra-chat_de)
 
 </details>
 
@@ -193,7 +240,9 @@ Please refer to [constants.py](src/llmtuner/extras/constants.py) for a full list
 - [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)
+- [Orca DPO (en)](https://huggingface.co/datasets/Intel/orca_dpo_pairs)
 - [Nectar (en)](https://huggingface.co/datasets/berkeley-nest/Nectar)
+- [Orca DPO (de)](https://huggingface.co/datasets/mayflowergmbh/intel_orca_dpo_pairs_de)
 
 </details>
 
@@ -208,22 +257,37 @@ huggingface-cli login
 
 ## Requirement
 
-- Python 3.8+ and PyTorch 1.13.1+
-- 🤗Transformers, Datasets, Accelerate, PEFT and TRL
-- sentencepiece, protobuf and tiktoken
-- jieba, rouge-chinese and nltk (used at evaluation and predict)
-- gradio and matplotlib (used in web UI)
-- uvicorn, fastapi and sse-starlette (used in API)
+| Mandatory    | Minimum | Recommend |
+| ------------ | ------- | --------- |
+| python       | 3.8     | 3.10      |
+| torch        | 1.13.1  | 2.2.0     |
+| transformers | 4.37.2  | 4.39.1    |
+| datasets     | 2.14.3  | 2.17.1    |
+| accelerate   | 0.27.2  | 0.28.0    |
+| peft         | 0.9.0   | 0.10.0    |
+| trl          | 0.8.1   | 0.8.1     |
+
+| Optional     | Minimum | Recommend |
+| ------------ | ------- | --------- |
+| CUDA         | 11.6    | 12.2      |
+| deepspeed    | 0.10.0  | 0.14.0    |
+| bitsandbytes | 0.39.0  | 0.43.0    |
+| flash-attn   | 2.3.0   | 2.5.6     |
 
 ### Hardware Requirement
 
-| Method | Bits |   7B  |  13B  |  30B  |   65B  |   8x7B |
+\* *estimated*
+
+| Method | Bits |   7B  |  13B  |  30B  |   70B  |   8x7B |
 | ------ | ---- | ----- | ----- | ----- | ------ | ------ |
-| Full   |  16  | 160GB | 320GB | 600GB | 1200GB |  900GB |
-| Freeze |  16  |  20GB |  40GB | 120GB |  240GB |  200GB |
-| LoRA   |  16  |  16GB |  32GB |  80GB |  160GB |  120GB |
-| QLoRA  |   8  |  10GB |  16GB |  40GB |   80GB |   80GB |
-| QLoRA  |   4  |   6GB |  12GB |  24GB |   48GB |   32GB |
+| Full   | AMP  | 120GB | 240GB | 600GB | 1200GB |  900GB |
+| Full   |  16  |  60GB | 120GB | 300GB |  600GB |  400GB |
+| GaLore |  16  |  16GB |  32GB |  64GB |  160GB |  120GB |
+| Freeze |  16  |  20GB |  40GB |  80GB |  200GB |  160GB |
+| LoRA   |  16  |  16GB |  32GB |  64GB |  160GB |  120GB |
+| QLoRA  |   8  |  10GB |  20GB |  40GB |   80GB |   60GB |
+| QLoRA  |   4  |   6GB |  12GB |  24GB |   48GB |   30GB |
+| QLoRA  |   2  |   4GB |   8GB |  16GB |   24GB |   18GB |
 
 ## Getting Started
 
@@ -244,12 +308,14 @@ cd LLaMA-Factory
 pip install -r requirements.txt
 ```
 
-If you want to enable the quantized LoRA (QLoRA) on the Windows platform, you will be required to install a pre-built version of `bitsandbytes` library, which supports CUDA 11.1 to 12.1.
+If you want to enable the quantized LoRA (QLoRA) on the Windows platform, you will be required to install a pre-built version of `bitsandbytes` library, which supports CUDA 11.1 to 12.2, please select the appropriate [release version](https://github.com/jllllll/bitsandbytes-windows-webui/releases/tag/wheels) based on your CUDA version.
 
 ```bash
-pip install https://github.com/jllllll/bitsandbytes-windows-webui/releases/download/wheels/bitsandbytes-0.39.1-py3-none-win_amd64.whl
+pip install https://github.com/jllllll/bitsandbytes-windows-webui/releases/download/wheels/bitsandbytes-0.41.2.post2-py3-none-win_amd64.whl
 ```
 
+To enable FlashAttention-2 on the Windows platform, you need to install the precompiled `flash-attn` library, which supports CUDA 12.1 to 12.2. Please download the corresponding version from [flash-attention](https://github.com/bdashore3/flash-attention/releases) based on your requirements.
+
 ### Use ModelScope Hub (optional)
 
 If you have trouble with downloading models and datasets from Hugging Face, you can use LLaMA-Factory together with ModelScope in the following manner.
@@ -263,7 +329,7 @@ Then you can train the corresponding model by specifying a model ID of the Model
 ```bash
 CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --model_name_or_path modelscope/Llama-2-7b-ms \
-    ... # arguments (same as above)
+    ... # arguments (same as below)
 ```
 
 LLaMA Board also supports using the models and datasets on the ModelScope Hub.
@@ -277,6 +343,13 @@ CUDA_VISIBLE_DEVICES=0 USE_MODELSCOPE_HUB=1 python src/train_web.py
 > [!IMPORTANT]
 > If you want to train models on multiple GPUs, please refer to [Distributed Training](#distributed-training).
 
+
+#### LLaMA Board GUI
+
+```bash
+CUDA_VISIBLE_DEVICES=0 python src/train_web.py
+```
+
 #### Pre-Training
 
 ```bash
@@ -343,7 +416,7 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --lr_scheduler_type cosine \
     --logging_steps 10 \
     --save_steps 1000 \
-    --learning_rate 1e-6 \
+    --learning_rate 1e-5 \
     --num_train_epochs 1.0 \
     --plot_loss \
     --fp16
@@ -377,6 +450,9 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --fp16
 ```
 
+> [!TIP]
+> Use `--adapter_name_or_path path_to_sft_checkpoint,path_to_ppo_checkpoint` to infer the fine-tuned model.
+
 > [!WARNING]
 > Use `--per_device_train_batch_size=1` for LLaMA-2 models in fp16 PPO training.
 
@@ -405,19 +481,24 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --fp16
 ```
 
+> [!TIP]
+> Use `--adapter_name_or_path path_to_sft_checkpoint,path_to_dpo_checkpoint` to infer the fine-tuned model.
+
 ### Distributed Training
 
 #### Use Huggingface Accelerate
 
 ```bash
-accelerate config # configure the environment
-accelerate launch src/train_bash.py # arguments (same as above)
+accelerate launch --config_file config.yaml src/train_bash.py \
+    --ddp_timeout 180000000 \
+    ... # arguments (same as above)
 ```
 
-<details><summary>Example config for LoRA training</summary>
+<details><summary>Example config.yaml for LoRA training</summary>
 
 ```yaml
 compute_environment: LOCAL_MACHINE
+debug: false
 distributed_type: MULTI_GPU
 downcast_bf16: 'no'
 gpu_ids: all
@@ -436,15 +517,19 @@ use_cpu: false
 
 </details>
 
+> [!TIP]
+> We commend using Accelerate for LoRA tuning.
+
 #### Use DeepSpeed
 
 ```bash
-deepspeed --num_gpus 8 --master_port=9901 src/train_bash.py \
+deepspeed --num_gpus 8 src/train_bash.py \
     --deepspeed ds_config.json \
+    --ddp_timeout 180000000 \
     ... # arguments (same as above)
 ```
 
-<details><summary>Example config for full-parameter training with DeepSpeed ZeRO-2</summary>
+<details><summary>Example ds_config.json for full-parameter training with DeepSpeed ZeRO-2</summary>
 
 ```json
 {
@@ -456,29 +541,36 @@ deepspeed --num_gpus 8 --master_port=9901 src/train_bash.py \
   "fp16": {
     "enabled": "auto",
     "loss_scale": 0,
-    "initial_scale_power": 16,
     "loss_scale_window": 1000,
+    "initial_scale_power": 16,
     "hysteresis": 2,
     "min_loss_scale": 1
   },
+  "bf16": {
+    "enabled": "auto"
+  },
   "zero_optimization": {
     "stage": 2,
     "allgather_partitions": true,
     "allgather_bucket_size": 5e8,
+    "overlap_comm": true,
     "reduce_scatter": true,
     "reduce_bucket_size": 5e8,
-    "overlap_comm": false,
-    "contiguous_gradients": true
+    "contiguous_gradients": true,
+    "round_robin_gradients": true
   }
 }
 ```
 
 </details>
 
+> [!TIP]
+> Refer to [examples](examples) for more training scripts.
+
 ### Merge LoRA weights and export model
 
 ```bash
-python src/export_model.py \
+CUDA_VISIBLE_DEVICES=0 python src/export_model.py \
     --model_name_or_path path_to_llama_model \
     --adapter_name_or_path path_to_checkpoint \
     --template default \
@@ -492,12 +584,14 @@ python src/export_model.py \
 > Merging LoRA weights into a quantized model is not supported.
 
 > [!TIP]
-> Use `--export_quantization_bit 4` and `--export_quantization_dataset data/c4_demo.json` to quantize the model after merging the LoRA weights.
+> Use `--model_name_or_path path_to_export` solely to use the exported model.
+> 
+> Use `--export_quantization_bit 4` and `--export_quantization_dataset data/c4_demo.json` to quantize the model with AutoGPTQ after merging the LoRA weights.
 
-### API Demo
+### Inference with OpenAI-style API
 
 ```bash
-python src/api_demo.py \
+CUDA_VISIBLE_DEVICES=0 API_PORT=8000 python src/api_demo.py \
     --model_name_or_path path_to_llama_model \
     --adapter_name_or_path path_to_checkpoint \
     --template default \
@@ -507,20 +601,20 @@ python src/api_demo.py \
 > [!TIP]
 > Visit `http://localhost:8000/docs` for API documentation.
 
-### CLI Demo
+### Inference with command line
 
 ```bash
-python src/cli_demo.py \
+CUDA_VISIBLE_DEVICES=0 python src/cli_demo.py \
     --model_name_or_path path_to_llama_model \
     --adapter_name_or_path path_to_checkpoint \
     --template default \
     --finetuning_type lora
 ```
 
-### Web Demo
+### Inference with web browser
 
 ```bash
-python src/web_demo.py \
+CUDA_VISIBLE_DEVICES=0 python src/web_demo.py \
     --model_name_or_path path_to_llama_model \
     --adapter_name_or_path path_to_checkpoint \
     --template default \
@@ -554,7 +648,7 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --template default \
     --finetuning_type lora \
     --output_dir path_to_predict_result \
-    --per_device_eval_batch_size 8 \
+    --per_device_eval_batch_size 1 \
     --max_samples 100 \
     --predict_with_generate \
     --fp16
@@ -566,13 +660,60 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
 > [!TIP]
 > We recommend using `--per_device_eval_batch_size=1` and `--max_target_length 128` at 4/8-bit predict.
 
+### Dockerize Training
+
+#### Get ready
+
+Necessary dockerized environment is needed, such as Docker or Docker Compose.
+
+#### Docker support
+
+```bash
+docker build -f ./Dockerfile -t llama-factory:latest .
+
+docker run --gpus=all -v ./hf_cache:/root/.cache/huggingface/ -v ./data:/app/data -v ./output:/app/output -p 7860:7860 --shm-size 16G --name llama_factory -d llama-factory:latest
+```
+
+#### Docker Compose support
+
+```bash
+docker compose -f ./docker-compose.yml up -d
+```
+
+> [!TIP]
+> Details about volume:
+> * hf_cache: Utilize Huggingface cache on the host machine. Reassignable if a cache already exists in a different directory.
+> * data: Place datasets on this dir of the host machine so that they can be selected on LLaMA Board GUI.
+> * output: Set export dir to this location so that the merged result can be accessed directly on the host machine.
+
 ## Projects using LLaMA Factory
 
-- **[StarWhisper](https://github.com/Yu-Yang-Li/StarWhisper)**: A large language model for Astronomy, based on ChatGLM2-6B and Qwen-14B.
-- **[DISC-LawLLM](https://github.com/FudanDISC/DISC-LawLLM)**: A large language model specialized in Chinese legal domain, based on Baichuan-13B, is capable of retrieving and reasoning on legal knowledge.
-- **[Sunsimiao](https://github.com/thomas-yanxin/Sunsimiao)**: A large language model specialized in Chinese medical domain, based on Baichuan-7B and ChatGLM-6B.
-- **[CareGPT](https://github.com/WangRongsheng/CareGPT)**: A series of large language models for Chinese medical domain, based on LLaMA2-7B and Baichuan-13B.
-- **[MachineMindset](https://github.com/PKU-YuanGroup/Machine-Mindset/)**: A series of MBTI Personality large language models, capable of giving any LLM 16 different personality types based on different datasets and training methods.
+1. Wang et al. ESRL: Efficient Sampling-based Reinforcement Learning for Sequence Generation. 2023. [[arxiv]](https://arxiv.org/abs/2308.02223)
+1. Yu et al. Open, Closed, or Small Language Models for Text Classification? 2023. [[arxiv]](https://arxiv.org/abs/2308.10092)
+1. Wang et al. UbiPhysio: Support Daily Functioning, Fitness, and Rehabilitation with Action Understanding and Feedback in Natural Language. 2023. [[arxiv]](https://arxiv.org/abs/2308.10526)
+1. Luceri et al. Leveraging Large Language Models to Detect Influence Campaigns in Social Media. 2023. [[arxiv]](https://arxiv.org/abs/2311.07816)
+1. Zhang et al. Alleviating Hallucinations of Large Language Models through Induced Hallucinations. 2023. [[arxiv]](https://arxiv.org/abs/2312.15710)
+1. Wang et al. Know Your Needs Better: Towards Structured Understanding of Marketer Demands with Analogical Reasoning Augmented LLMs. 2024. [[arxiv]](https://arxiv.org/abs/2401.04319)
+1. Wang et al. CANDLE: Iterative Conceptualization and Instantiation Distillation from Large Language Models for Commonsense Reasoning. 2024. [[arxiv]](https://arxiv.org/abs/2401.07286)
+1. Choi et al. FACT-GPT: Fact-Checking Augmentation via Claim Matching with LLMs. 2024. [[arxiv]](https://arxiv.org/abs/2402.05904)
+1. Zhang et al. AutoMathText: Autonomous Data Selection with Language Models for Mathematical Texts. 2024. [[arxiv]](https://arxiv.org/abs/2402.07625)
+1. Lyu et al. KnowTuning: Knowledge-aware Fine-tuning for Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2402.11176)
+1. Yang et al. LaCo: Large Language Model Pruning via Layer Collaps. 2024. [[arxiv]](https://arxiv.org/abs/2402.11187)
+1. Bhardwaj et al. Language Models are Homer Simpson! Safety Re-Alignment of Fine-tuned Language Models through Task Arithmetic. 2024. [[arxiv]](https://arxiv.org/abs/2402.11746)
+1. Yang et al. Enhancing Empathetic Response Generation by Augmenting LLMs with Small-scale Empathetic Models. 2024. [[arxiv]](https://arxiv.org/abs/2402.11801)
+1. Yi et al. Generation Meets Verification: Accelerating Large Language Model Inference with Smart Parallel Auto-Correct Decoding. 2024. [[arxiv]](https://arxiv.org/abs/2402.11809)
+1. Cao et al. Head-wise Shareable Attention for Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2402.11819)
+1. Zhang et al. Enhancing Multilingual Capabilities of Large Language Models through Self-Distillation from Resource-Rich Languages. 2024. [[arxiv]](https://arxiv.org/abs/2402.12204)
+1. Kim et al. Efficient and Effective Vocabulary Expansion Towards Multilingual Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2402.14714)
+1. Yu et al. KIEval: A Knowledge-grounded Interactive Evaluation Framework for Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2402.15043)
+1. Huang et al. Key-Point-Driven Data Synthesis with its Enhancement on Mathematical Reasoning. 2024. [[arxiv]](https://arxiv.org/abs/2403.02333)
+1. Duan et al. Negating Negatives: Alignment without Human Positive Samples via Distributional Dispreference Optimization. 2024. [[arxiv]](https://arxiv.org/abs/2403.03419)
+1. Xie and Schwertfeger. Empowering Robotics with Large Language Models: osmAG Map Comprehension with LLMs. 2024. [[arxiv]](https://arxiv.org/abs/2403.08228)
+1. **[StarWhisper](https://github.com/Yu-Yang-Li/StarWhisper)**: A large language model for Astronomy, based on ChatGLM2-6B and Qwen-14B.
+1. **[DISC-LawLLM](https://github.com/FudanDISC/DISC-LawLLM)**: A large language model specialized in Chinese legal domain, based on Baichuan-13B, is capable of retrieving and reasoning on legal knowledge.
+1. **[Sunsimiao](https://github.com/thomas-yanxin/Sunsimiao)**: A large language model specialized in Chinese medical domain, based on Baichuan-7B and ChatGLM-6B.
+1. **[CareGPT](https://github.com/WangRongsheng/CareGPT)**: A series of large language models for Chinese medical domain, based on LLaMA2-7B and Baichuan-13B.
+1. **[MachineMindset](https://github.com/PKU-YuanGroup/Machine-Mindset/)**: A series of MBTI Personality large language models, capable of giving any LLM 16 different personality types based on different datasets and training methods.
 
 > [!TIP]
 > If you have a project that should be incorporated, please contact via email or create a pull request.
@@ -581,18 +722,19 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
 
 This repository is licensed under the [Apache-2.0 License](LICENSE).
 
-Please follow the model licenses to use the corresponding model weights: [Baichuan2](https://huggingface.co/baichuan-inc/Baichuan2-7B-Base/blob/main/Community%20License%20for%20Baichuan%202%20Model.pdf) / [BLOOM](https://huggingface.co/spaces/bigscience/license) / [ChatGLM3](https://github.com/THUDM/ChatGLM3/blob/main/MODEL_LICENSE) / [DeepSeek](https://github.com/deepseek-ai/DeepSeek-LLM/blob/main/LICENSE-MODEL) / [Falcon](https://huggingface.co/tiiuae/falcon-180B/blob/main/LICENSE.txt) / [InternLM2](https://github.com/InternLM/InternLM#license) / [LLaMA](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md) / [LLaMA-2](https://ai.meta.com/llama/license/) / [Mistral](LICENSE) / [Phi-1.5/2](https://huggingface.co/microsoft/phi-1_5/resolve/main/Research%20License.docx) / [Qwen](https://github.com/QwenLM/Qwen/blob/main/Tongyi%20Qianwen%20LICENSE%20AGREEMENT) / [XVERSE](https://github.com/xverse-ai/XVERSE-13B/blob/main/MODEL_LICENSE.pdf) / [Yi](https://huggingface.co/01-ai/Yi-6B/blob/main/LICENSE) / [Yuan](https://github.com/IEIT-Yuan/Yuan-2.0/blob/main/LICENSE-Yuan)
+Please follow the model licenses to use the corresponding model weights: [Baichuan2](https://huggingface.co/baichuan-inc/Baichuan2-7B-Base/blob/main/Community%20License%20for%20Baichuan%202%20Model.pdf) / [BLOOM](https://huggingface.co/spaces/bigscience/license) / [ChatGLM3](https://github.com/THUDM/ChatGLM3/blob/main/MODEL_LICENSE) / [DeepSeek](https://github.com/deepseek-ai/DeepSeek-LLM/blob/main/LICENSE-MODEL) / [Falcon](https://huggingface.co/tiiuae/falcon-180B/blob/main/LICENSE.txt) / [Gemma](https://ai.google.dev/gemma/terms) / [InternLM2](https://github.com/InternLM/InternLM#license) / [LLaMA](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md) / [LLaMA-2](https://ai.meta.com/llama/license/) / [Mistral](LICENSE) / [OLMo](LICENSE) / [Phi-1.5/2](https://huggingface.co/microsoft/phi-1_5/resolve/main/Research%20License.docx) / [Qwen](https://github.com/QwenLM/Qwen/blob/main/Tongyi%20Qianwen%20LICENSE%20AGREEMENT) / [StarCoder2](https://huggingface.co/spaces/bigcode/bigcode-model-license-agreement) / [XVERSE](https://github.com/xverse-ai/XVERSE-13B/blob/main/MODEL_LICENSE.pdf) / [Yi](https://huggingface.co/01-ai/Yi-6B/blob/main/LICENSE) / [Yuan](https://github.com/IEIT-Yuan/Yuan-2.0/blob/main/LICENSE-Yuan)
 
 ## Citation
 
 If this work is helpful, please kindly cite as:
 
 ```bibtex
-@Misc{llama-factory,
-  title = {LLaMA Factory},
-  author = {hiyouga},
-  howpublished = {\url{https://github.com/hiyouga/LLaMA-Factory}},
-  year = {2023}
+@article{zheng2024llamafactory,
+  title={LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models}, 
+  author={Yaowei Zheng and Richong Zhang and Junhao Zhang and Yanhan Ye and Zheyan Luo and Yongqiang Ma},
+  journal={arXiv preprint arXiv:2403.13372},
+  year={2024},
+  url={http://arxiv.org/abs/2403.13372}
 }
 ```
 

README_zh.md

@@ -5,27 +5,30 @@
 [![GitHub last commit](https://img.shields.io/github/last-commit/hiyouga/LLaMA-Factory)](https://github.com/hiyouga/LLaMA-Factory/commits/main)
 [![PyPI](https://img.shields.io/pypi/v/llmtuner)](https://pypi.org/project/llmtuner/)
 [![Downloads](https://static.pepy.tech/badge/llmtuner)](https://pypi.org/project/llmtuner/)
+[![Citation](https://img.shields.io/badge/citation-26-green)](#使用了-llama-factory-的项目)
 [![GitHub pull request](https://img.shields.io/badge/PRs-welcome-blue)](https://github.com/hiyouga/LLaMA-Factory/pulls)
 [![Discord](https://dcbadge.vercel.app/api/server/rKfvV9r9FK?compact=true&style=flat)](https://discord.gg/rKfvV9r9FK)
-[![Spaces](https://img.shields.io/badge/🤗-Open%20In%20Spaces-blue)](https://huggingface.co/spaces/hiyouga/LLaMA-Board)
-[![Studios](https://img.shields.io/badge/ModelScope-Open%20In%20Studios-blue)](https://modelscope.cn/studios/hiyouga/LLaMA-Board)
+[![Twitter](https://img.shields.io/twitter/follow/llamafactory_ai)](https://twitter.com/llamafactory_ai)
+[![Spaces](https://img.shields.io/badge/🤗-Open%20in%20Spaces-blue)](https://huggingface.co/spaces/hiyouga/LLaMA-Board)
+[![Studios](https://img.shields.io/badge/ModelScope-Open%20in%20Studios-blue)](https://modelscope.cn/studios/hiyouga/LLaMA-Board)
+[![Open in Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1eRTPn37ltBbYsISy9Aw2NuI2Aq5CQrD9?usp=sharing)
 
 👋 加入我们的[微信群](assets/wechat.jpg)。
 
 \[ [English](README.md) | 中文 \]
 
-## LLaMA Board: 通过一站式网页界面快速上手 LLaMA Factory
+**微调大模型可以像这样轻松…**
 
-通过 **[🤗 Spaces](https://huggingface.co/spaces/hiyouga/LLaMA-Board)** 或 **[ModelScope](https://modelscope.cn/studios/hiyouga/LLaMA-Board)** 预览 LLaMA Board。
+https://github.com/hiyouga/LLaMA-Factory/assets/16256802/ec36a9dd-37f4-4f72-81bd-d76c6d0a6594
 
-使用 `CUDA_VISIBLE_DEVICES=0 python src/train_web.py` 启动 LLaMA Board。（该模式目前仅支持单卡训练）
+选择你的打开方式：
 
-下面是使用单张 GPU 在 10 分钟内更改对话式大型语言模型自我认知的示例。
-
-https://github.com/hiyouga/LLaMA-Factory/assets/16256802/6ba60acc-e2e2-4bec-b846-2d88920d5ba1
+- **Colab**：https://colab.research.google.com/drive/1eRTPn37ltBbYsISy9Aw2NuI2Aq5CQrD9?usp=sharing
+- **本地机器**：请见[如何使用](#如何使用)
 
 ## 目录
 
+- [项目特色](#项目特色)
 - [性能指标](#性能指标)
 - [更新日志](#更新日志)
 - [模型](#模型)
@@ -38,6 +41,16 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/6ba60acc-e2e2-4bec-b846
 - [引用](#引用)
 - [致谢](#致谢)
 
+## 项目特色
+
+- **多种模型**：LLaMA、Mistral、Mixtral-MoE、Qwen、Yi、Gemma、Baichuan、ChatGLM、Phi 等等。
+- **集成方法**：（增量）预训练、指令监督微调、奖励模型训练、PPO 训练和 DPO 训练。
+- **多种精度**：32 比特全参数微调、16 比特冻结微调、16 比特 LoRA 微调和基于 AQLM/AWQ/GPTQ/LLM.int8 的 2/4/8 比特 QLoRA 微调。
+- **先进算法**：GaLore、DoRA、LongLoRA、LLaMA Pro、LoRA+、LoftQ 和 Agent 微调。
+- **实用技巧**：FlashAttention-2、Unsloth、RoPE scaling、NEFTune 和 rsLoRA。
+- **实验监控**：LlamaBoard、TensorBoard、Wandb、MLflow 等等。
+- **极速推理**：基于 vLLM 的 OpenAI 风格 API、浏览器界面和命令行接口。
+
 ## 性能指标
 
 与 ChatGLM 官方的 [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/ptuning) 微调相比，LLaMA-Factory 的 LoRA 微调提供了 **3.7 倍**的加速比，同时在广告文案生成任务上取得了更高的 Rouge 分数。结合 4 比特量化技术，LLaMA-Factory 的 QLoRA 微调进一步降低了 GPU 显存消耗。
@@ -55,14 +68,30 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/6ba60acc-e2e2-4bec-b846
 
 ## 更新日志
 
-[24/01/18] 我们针对绝大多数模型实现了 **Agent 微调**，微调时指定 `--dataset glaive_toolcall` 即可使模型获得工具调用能力。
+[24/03/21] 我们的论文 "[LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models](https://arxiv.org/abs/2403.13372)" 可在 arXiv 上查看！
 
-[23/12/23] 我们针对 LLaMA, Mistral 和 Yi 模型支持了 **[unsloth](https://github.com/unslothai/unsloth)** 的 LoRA 训练加速。请使用 `--use_unsloth` 参数启用 unsloth 优化。该方法可提供 1.7 倍的训练速度，详情请查阅[此页面](https://github.com/hiyouga/LLaMA-Factory/wiki/Performance-comparison)。
+[24/03/20] 我们支持了能在 2x24GB GPU 上微调 70B 模型的 **FSDP+QLoRA**。详细用法请参照 `examples/fsdp_qlora`。
 
-[23/12/12] 我们支持了微调最新的混合专家模型 **[Mixtral 8x7B](https://huggingface.co/mistralai/Mixtral-8x7B-v0.1)**。硬件需求请查阅[此处](#硬件依赖)。
+[24/03/13] 我们支持了 **[LoRA+](https://arxiv.org/abs/2402.12354)**。请使用 `loraplus_lr_ratio=16.0` 参数开启 LoRA+ 方法。
+
+[24/03/07] 我们支持了梯度低秩投影（**[GaLore](https://arxiv.org/abs/2403.03507)**）算法。请使用 `--use_galore` 参数切换显存高效的优化器。
+
+[24/03/07] 我们集成了 **[vLLM](https://github.com/vllm-project/vllm)** 以实现极速并发推理。请使用 `--infer_backend vllm` 来获得 **270%** 的推理速度。（尚不支持 LoRA，请先合并权重。）
 
 <details><summary>展开日志</summary>
 
+[24/02/28] 我们支持了 **[DoRA](https://arxiv.org/abs/2402.09353)** 微调。请使用 `--use_dora` 参数进行 DoRA 微调。
+
+[24/02/15] 我们支持了 [LLaMA Pro](https://github.com/TencentARC/LLaMA-Pro) 提出的**块扩展**方法。详细用法请参照 `examples/extras/llama_pro`。
+
+[24/02/05] Qwen1.5（Qwen2 测试版）系列模型已在 LLaMA-Factory 中实现微调支持。详情请查阅该[博客页面](https://qwenlm.github.io/zh/blog/qwen1.5/)。
+
+[24/01/18] 我们针对绝大多数模型实现了 **Agent 微调**，微调时指定 `--dataset glaive_toolcall` 即可使模型获得工具调用能力。
+
+[23/12/23] 我们针对 LLaMA, Mistral 和 Yi 模型支持了 **[unsloth](https://github.com/unslothai/unsloth)** 的 LoRA 训练加速。请使用 `--use_unsloth` 参数启用 unsloth 优化。该方法可提供 **170%** 的训练速度，详情请查阅[此页面](https://github.com/hiyouga/LLaMA-Factory/wiki/Performance-comparison)。
+
+[23/12/12] 我们支持了微调最新的混合专家模型 **[Mixtral 8x7B](https://huggingface.co/mistralai/Mixtral-8x7B-v0.1)**。硬件需求请查阅[此处](#硬件依赖)。
+
 [23/12/01] 我们支持了从 **[魔搭社区](https://modelscope.cn/models)** 下载预训练模型和数据集。详细用法请参照 [此教程](#使用魔搭社区可跳过)。
 
 [23/10/21] 我们支持了 **[NEFTune](https://arxiv.org/abs/2310.05914)** 训练技巧。请使用 `--neftune_noise_alpha` 参数启用 NEFTune，例如 `--neftune_noise_alpha 5`。
@@ -103,15 +132,19 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/6ba60acc-e2e2-4bec-b846
 | [ChatGLM3](https://huggingface.co/THUDM/chatglm3-6b)     | 6B                          | query_key_value   | chatglm3  |
 | [DeepSeek (MoE)](https://huggingface.co/deepseek-ai)     | 7B/16B/67B                  | q_proj,v_proj     | deepseek  |
 | [Falcon](https://huggingface.co/tiiuae)                  | 7B/40B/180B                 | query_key_value   | falcon    |
+| [Gemma](https://huggingface.co/google)                   | 2B/7B                       | q_proj,v_proj     | gemma     |
 | [InternLM2](https://huggingface.co/internlm)             | 7B/20B                      | wqkv              | intern2   |
 | [LLaMA](https://github.com/facebookresearch/llama)       | 7B/13B/33B/65B              | q_proj,v_proj     | -         |
 | [LLaMA-2](https://huggingface.co/meta-llama)             | 7B/13B/70B                  | q_proj,v_proj     | llama2    |
 | [Mistral](https://huggingface.co/mistralai)              | 7B                          | q_proj,v_proj     | mistral   |
 | [Mixtral](https://huggingface.co/mistralai)              | 8x7B                        | q_proj,v_proj     | mistral   |
+| [OLMo](https://huggingface.co/allenai)                   | 1B/7B                       | att_proj          | olmo      |
 | [Phi-1.5/2](https://huggingface.co/microsoft)            | 1.3B/2.7B                   | q_proj,v_proj     | -         |
 | [Qwen](https://huggingface.co/Qwen)                      | 1.8B/7B/14B/72B             | c_attn            | qwen      |
+| [Qwen1.5](https://huggingface.co/Qwen)                   | 0.5B/1.8B/4B/7B/14B/72B     | q_proj,v_proj     | qwen      |
+| [StarCoder2](https://huggingface.co/bigcode)             | 3B/7B/15B                   | q_proj,v_proj     | -         |
 | [XVERSE](https://huggingface.co/xverse)                  | 7B/13B/65B                  | q_proj,v_proj     | xverse    |
-| [Yi](https://huggingface.co/01-ai)                       | 6B/34B                      | q_proj,v_proj     | yi        |
+| [Yi](https://huggingface.co/01-ai)                       | 6B/9B/34B                   | q_proj,v_proj     | yi        |
 | [Yuan](https://huggingface.co/IEITYuan)                  | 2B/51B/102B                 | q_proj,v_proj     | yuan      |
 
 > [!NOTE]
@@ -121,6 +154,8 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/6ba60acc-e2e2-4bec-b846
 
 项目所支持模型的完整列表请参阅 [constants.py](src/llmtuner/extras/constants.py)。
 
+您也可以在 [template.py](src/llmtuner/data/template.py) 中添加自己的对话模板。
+
 ## 训练方法
 
 | 方法                   |     全参数训练      |    部分参数训练     |       LoRA         |       QLoRA        |
@@ -154,8 +189,8 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/6ba60acc-e2e2-4bec-b846
 
 - [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)
-- [Self-cognition (zh)](data/self_cognition.json)
+- [Alpaca GPT4 (en&zh)](https://github.com/Instruction-Tuning-with-GPT-4/GPT-4-LLM)
+- [Self Cognition (zh)](data/self_cognition.json)
 - [Open Assistant (multilingual)](https://huggingface.co/datasets/OpenAssistant/oasst1)
 - [ShareGPT (zh)](https://huggingface.co/datasets/QingyiSi/Alpaca-CoT/tree/main/Chinese-instruction-collection)
 - [Guanaco Dataset (multilingual)](https://huggingface.co/datasets/JosephusCheung/GuanacoDataset)
@@ -171,8 +206,10 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/6ba60acc-e2e2-4bec-b846
 - [CodeAlpaca 20k (en)](https://huggingface.co/datasets/sahil2801/CodeAlpaca-20k)
 - [Alpaca CoT (multilingual)](https://huggingface.co/datasets/QingyiSi/Alpaca-CoT)
 - [OpenOrca (en)](https://huggingface.co/datasets/Open-Orca/OpenOrca)
+- [SlimOrca (en)](https://huggingface.co/datasets/Open-Orca/SlimOrca)
 - [MathInstruct (en)](https://huggingface.co/datasets/TIGER-Lab/MathInstruct)
 - [Firefly 1.1M (zh)](https://huggingface.co/datasets/YeungNLP/firefly-train-1.1M)
+- [Wiki QA (en)](https://huggingface.co/datasets/wiki_qa)
 - [Web QA (zh)](https://huggingface.co/datasets/suolyer/webqa)
 - [WebNovel (zh)](https://huggingface.co/datasets/zxbsmk/webnovel_cn)
 - [Nectar (en)](https://huggingface.co/datasets/berkeley-nest/Nectar)
@@ -185,6 +222,16 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/6ba60acc-e2e2-4bec-b846
 - [LMSYS Chat 1M (en)](https://huggingface.co/datasets/lmsys/lmsys-chat-1m)
 - [Evol Instruct V2 (en)](https://huggingface.co/datasets/WizardLM/WizardLM_evol_instruct_V2_196k)
 - [Glaive Function Calling V2 (en)](https://huggingface.co/datasets/glaiveai/glaive-function-calling-v2)
+- [Cosmopedia (en)](https://huggingface.co/datasets/HuggingFaceTB/cosmopedia)
+- [Open Assistant (de)](https://huggingface.co/datasets/mayflowergmbh/oasst_de)
+- [Dolly 15k (de)](https://huggingface.co/datasets/mayflowergmbh/dolly-15k_de)
+- [Alpaca GPT4 (de)](https://huggingface.co/datasets/mayflowergmbh/alpaca-gpt4_de)
+- [OpenSchnabeltier (de)](https://huggingface.co/datasets/mayflowergmbh/openschnabeltier_de)
+- [Evol Instruct (de)](https://huggingface.co/datasets/mayflowergmbh/evol-instruct_de)
+- [Dolphin (de)](https://huggingface.co/datasets/mayflowergmbh/dolphin_de)
+- [Booksum (de)](https://huggingface.co/datasets/mayflowergmbh/booksum_de)
+- [Airoboros (de)](https://huggingface.co/datasets/mayflowergmbh/airoboros-3.0_de)
+- [Ultrachat (de)](https://huggingface.co/datasets/mayflowergmbh/ultra-chat_de)
 
 </details>
 
@@ -193,7 +240,9 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/6ba60acc-e2e2-4bec-b846
 - [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)
+- [Orca DPO (en)](https://huggingface.co/datasets/Intel/orca_dpo_pairs)
 - [Nectar (en)](https://huggingface.co/datasets/berkeley-nest/Nectar)
+- [Orca DPO (de)](https://huggingface.co/datasets/mayflowergmbh/intel_orca_dpo_pairs_de)
 
 </details>
 
@@ -208,22 +257,37 @@ huggingface-cli login
 
 ## 软硬件依赖
 
-- Python 3.8+ 和 PyTorch 1.13.1+
-- 🤗Transformers, Datasets, Accelerate, PEFT 和 TRL
-- sentencepiece, protobuf 和 tiktoken
-- jieba, rouge-chinese 和 nltk (用于评估及预测)
-- gradio 和 matplotlib (用于网页端交互)
-- uvicorn, fastapi 和 sse-starlette (用于 API)
+| 必需项       | 至少     | 推荐      |
+| ------------ | ------- | --------- |
+| python       | 3.8     | 3.10      |
+| torch        | 1.13.1  | 2.2.0     |
+| transformers | 4.37.2  | 4.39.1    |
+| datasets     | 2.14.3  | 2.17.1    |
+| accelerate   | 0.27.2  | 0.28.0    |
+| peft         | 0.9.0   | 0.10.0    |
+| trl          | 0.8.1   | 0.8.1     |
+
+| 可选项       | 至少     | 推荐      |
+| ------------ | ------- | --------- |
+| CUDA         | 11.6    | 12.2      |
+| deepspeed    | 0.10.0  | 0.14.0    |
+| bitsandbytes | 0.39.0  | 0.43.0    |
+| flash-attn   | 2.3.0   | 2.5.6     |
 
 ### 硬件依赖
 
-| 训练方法 | 精度 |   7B  |  13B  |  30B  |   65B  |   8x7B |
+\* *估算值*
+
+| 训练方法 | 精度 |   7B  |  13B  |  30B  |   70B  |   8x7B |
 | ------- | ---- | ----- | ----- | ----- | ------ | ------ |
-| 全参数   |  16  | 160GB | 320GB | 600GB | 1200GB |  900GB |
-| 部分参数 |  16  |  20GB |  40GB | 120GB |  240GB |  200GB |
-| LoRA    |  16  |  16GB |  32GB |  80GB |  160GB |  120GB |
-| QLoRA   |   8  |  10GB |  16GB |  40GB |   80GB |   80GB |
-| QLoRA   |   4  |   6GB |  12GB |  24GB |   48GB |   32GB |
+| 全参数   | AMP  | 120GB | 240GB | 600GB | 1200GB |  900GB |
+| 全参数   |  16  |  60GB | 120GB | 300GB |  600GB |  400GB |
+| GaLore  |  16  |  16GB |  32GB |  64GB |  160GB |  120GB |
+| 部分参数 |  16  |  20GB |  40GB |  80GB |  200GB |  160GB |
+| LoRA    |  16  |  16GB |  32GB |  64GB |  160GB |  120GB |
+| QLoRA   |   8  |  10GB |  20GB |  40GB |   80GB |   60GB |
+| QLoRA   |   4  |   6GB |  12GB |  24GB |   48GB |   30GB |
+| QLoRA   |   2  |   4GB |   8GB |  16GB |   24GB |   18GB |
 
 ## 如何使用
 
@@ -244,12 +308,14 @@ cd LLaMA-Factory
 pip install -r requirements.txt
 ```
 
-如果要在 Windows 平台上开启量化 LoRA（QLoRA），需要安装预编译的 `bitsandbytes` 库, 支持 CUDA 11.1 到 12.1.
+如果要在 Windows 平台上开启量化 LoRA（QLoRA），需要安装预编译的 `bitsandbytes` 库, 支持 CUDA 11.1 到 12.2, 请根据您的 CUDA 版本情况选择适合的[发布版本](https://github.com/jllllll/bitsandbytes-windows-webui/releases/tag/wheels)。
 
 ```bash
-pip install https://github.com/jllllll/bitsandbytes-windows-webui/releases/download/wheels/bitsandbytes-0.39.1-py3-none-win_amd64.whl
+pip install https://github.com/jllllll/bitsandbytes-windows-webui/releases/download/wheels/bitsandbytes-0.41.2.post2-py3-none-win_amd64.whl
 ```
 
+如果要在 Windows 平台上开启 FlashAttention-2，需要安装预编译的 `flash-attn` 库，支持 CUDA 12.1 到 12.2，请根据需求到 [flash-attention](https://github.com/bdashore3/flash-attention/releases) 下载对应版本安装。
+
 ### 使用魔搭社区（可跳过）
 
 如果您在 Hugging Face 模型和数据集的下载中遇到了问题，可以通过下述方法使用魔搭社区。
@@ -263,7 +329,7 @@ export USE_MODELSCOPE_HUB=1 # Windows 使用 `set USE_MODELSCOPE_HUB=1`
 ```bash
 CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --model_name_or_path modelscope/Llama-2-7b-ms \
-    ... # 参数同上
+    ... # 参数同下
 ```
 
 LLaMA Board 同样支持魔搭社区的模型和数据集下载。
@@ -277,6 +343,12 @@ CUDA_VISIBLE_DEVICES=0 USE_MODELSCOPE_HUB=1 python src/train_web.py
 > [!IMPORTANT]
 > 如果您使用多张 GPU 训练模型，请移步[多 GPU 分布式训练](#多-gpu-分布式训练)部分。
 
+#### LLaMA Board GUI
+
+```bash
+CUDA_VISIBLE_DEVICES=0 python src/train_web.py
+```
+
 #### 预训练
 
 ```bash
@@ -343,7 +415,7 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --lr_scheduler_type cosine \
     --logging_steps 10 \
     --save_steps 1000 \
-    --learning_rate 1e-6 \
+    --learning_rate 1e-5 \
     --num_train_epochs 1.0 \
     --plot_loss \
     --fp16
@@ -377,6 +449,9 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --fp16
 ```
 
+> [!TIP]
+> 使用 `--adapter_name_or_path path_to_sft_checkpoint,path_to_ppo_checkpoint` 来进行微调模型的推理。
+
 > [!WARNING]
 > 如果使用 fp16 精度进行 LLaMA-2 模型的 PPO 训练，请使用 `--per_device_train_batch_size=1`。
 
@@ -405,19 +480,24 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --fp16
 ```
 
+> [!TIP]
+> 使用 `--adapter_name_or_path path_to_sft_checkpoint,path_to_dpo_checkpoint` 来进行微调模型的推理。
+
 ### 多 GPU 分布式训练
 
 #### 使用 Huggingface Accelerate
 
 ```bash
-accelerate config # 首先配置分布式环境
-accelerate launch src/train_bash.py # 参数同上
+accelerate launch --config_file config.yaml src/train_bash.py \
+    --ddp_timeout 180000000 \
+    ... # 参数同上
 ```
 
-<details><summary>LoRA 训练的 Accelerate 配置示例</summary>
+<details><summary>使用 Accelerate 进行 LoRA 训练的 config.yaml 示例</summary>
 
 ```yaml
 compute_environment: LOCAL_MACHINE
+debug: false
 distributed_type: MULTI_GPU
 downcast_bf16: 'no'
 gpu_ids: all
@@ -436,15 +516,19 @@ use_cpu: false
 
 </details>
 
+> [!TIP]
+> 我们推荐使用 Accelerate 进行 LoRA 训练。
+
 #### 使用 DeepSpeed
 
 ```bash
-deepspeed --num_gpus 8 --master_port=9901 src/train_bash.py \
+deepspeed --num_gpus 8 src/train_bash.py \
     --deepspeed ds_config.json \
+    --ddp_timeout 180000000 \
     ... # 参数同上
 ```
 
-<details><summary>使用 DeepSpeed ZeRO-2 进行全参数训练的 DeepSpeed 配置示例</summary>
+<details><summary>使用 DeepSpeed ZeRO-2 进行全参数训练的 ds_config.json 示例</summary>
 
 ```json
 {
@@ -456,29 +540,36 @@ deepspeed --num_gpus 8 --master_port=9901 src/train_bash.py \
   "fp16": {
     "enabled": "auto",
     "loss_scale": 0,
-    "initial_scale_power": 16,
     "loss_scale_window": 1000,
+    "initial_scale_power": 16,
     "hysteresis": 2,
     "min_loss_scale": 1
   },
+  "bf16": {
+    "enabled": "auto"
+  },
   "zero_optimization": {
     "stage": 2,
     "allgather_partitions": true,
     "allgather_bucket_size": 5e8,
+    "overlap_comm": true,
     "reduce_scatter": true,
     "reduce_bucket_size": 5e8,
-    "overlap_comm": false,
-    "contiguous_gradients": true
+    "contiguous_gradients": true,
+    "round_robin_gradients": true
   }
 }
 ```
 
 </details>
 
+> [!TIP]
+> 更多训练脚本请查看 [examples](examples)。
+
 ### 合并 LoRA 权重并导出模型
 
 ```bash
-python src/export_model.py \
+CUDA_VISIBLE_DEVICES=0 python src/export_model.py \
     --model_name_or_path path_to_llama_model \
     --adapter_name_or_path path_to_checkpoint \
     --template default \
@@ -492,12 +583,14 @@ python src/export_model.py \
 > 尚不支持量化模型的 LoRA 权重合并及导出。
 
 > [!TIP]
-> 合并 LoRA 权重之后可再次使用 `--export_quantization_bit 4` 和 `--export_quantization_dataset data/c4_demo.json` 量化模型。
+> 仅使用 `--model_name_or_path path_to_export` 来加载导出后的模型。
+> 
+> 合并 LoRA 权重之后可再次使用 `--export_quantization_bit 4` 和 `--export_quantization_dataset data/c4_demo.json` 基于 AutoGPTQ 量化模型。
 
-### API 服务
+### 使用 OpenAI 风格 API 推理
 
 ```bash
-python src/api_demo.py \
+CUDA_VISIBLE_DEVICES=0 API_PORT=8000 python src/api_demo.py \
     --model_name_or_path path_to_llama_model \
     --adapter_name_or_path path_to_checkpoint \
     --template default \
@@ -507,20 +600,20 @@ python src/api_demo.py \
 > [!TIP]
 > 关于 API 文档请见 `http://localhost:8000/docs`。
 
-### 命令行测试
+### 使用命令行推理
 
 ```bash
-python src/cli_demo.py \
+CUDA_VISIBLE_DEVICES=0 python src/cli_demo.py \
     --model_name_or_path path_to_llama_model \
     --adapter_name_or_path path_to_checkpoint \
     --template default \
     --finetuning_type lora
 ```
 
-### 浏览器测试
+### 使用浏览器推理
 
 ```bash
-python src/web_demo.py \
+CUDA_VISIBLE_DEVICES=0 python src/web_demo.py \
     --model_name_or_path path_to_llama_model \
     --adapter_name_or_path path_to_checkpoint \
     --template default \
@@ -554,7 +647,7 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --template default \
     --finetuning_type lora \
     --output_dir path_to_predict_result \
-    --per_device_eval_batch_size 8 \
+    --per_device_eval_batch_size 1 \
     --max_samples 100 \
     --predict_with_generate \
     --fp16
@@ -568,11 +661,32 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
 
 ## 使用了 LLaMA Factory 的项目
 
-- **[StarWhisper](https://github.com/Yu-Yang-Li/StarWhisper)**: 天文大模型 StarWhisper，基于 ChatGLM2-6B 和 Qwen-14B 在天文数据上微调而得。
-- **[DISC-LawLLM](https://github.com/FudanDISC/DISC-LawLLM)**: 中文法律领域大模型 DISC-LawLLM，基于 Baichuan-13B 微调而得，具有法律推理和知识检索能力。
-- **[Sunsimiao](https://github.com/thomas-yanxin/Sunsimiao)**: 孙思邈中文医疗大模型 Sumsimiao，基于 Baichuan-7B 和 ChatGLM-6B 在中文医疗数据上微调而得。
-- **[CareGPT](https://github.com/WangRongsheng/CareGPT)**: 医疗大模型项目 CareGPT，基于 LLaMA2-7B 和 Baichuan-13B 在中文医疗数据上微调而得。
-- **[MachineMindset](https://github.com/PKU-YuanGroup/Machine-Mindset/)**：MBTI性格大模型项目，根据数据集与训练方式让任意 LLM 拥有 16 个不同的性格类型。
+1. Wang et al. ESRL: Efficient Sampling-based Reinforcement Learning for Sequence Generation. 2023. [[arxiv]](https://arxiv.org/abs/2308.02223)
+1. Yu et al. Open, Closed, or Small Language Models for Text Classification? 2023. [[arxiv]](https://arxiv.org/abs/2308.10092)
+1. Wang et al. UbiPhysio: Support Daily Functioning, Fitness, and Rehabilitation with Action Understanding and Feedback in Natural Language. 2023. [[arxiv]](https://arxiv.org/abs/2308.10526)
+1. Luceri et al. Leveraging Large Language Models to Detect Influence Campaigns in Social Media. 2023. [[arxiv]](https://arxiv.org/abs/2311.07816)
+1. Zhang et al. Alleviating Hallucinations of Large Language Models through Induced Hallucinations. 2023. [[arxiv]](https://arxiv.org/abs/2312.15710)
+1. Wang et al. Know Your Needs Better: Towards Structured Understanding of Marketer Demands with Analogical Reasoning Augmented LLMs. 2024. [[arxiv]](https://arxiv.org/abs/2401.04319)
+1. Wang et al. CANDLE: Iterative Conceptualization and Instantiation Distillation from Large Language Models for Commonsense Reasoning. 2024. [[arxiv]](https://arxiv.org/abs/2401.07286)
+1. Choi et al. FACT-GPT: Fact-Checking Augmentation via Claim Matching with LLMs. 2024. [[arxiv]](https://arxiv.org/abs/2402.05904)
+1. Zhang et al. AutoMathText: Autonomous Data Selection with Language Models for Mathematical Texts. 2024. [[arxiv]](https://arxiv.org/abs/2402.07625)
+1. Lyu et al. KnowTuning: Knowledge-aware Fine-tuning for Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2402.11176)
+1. Yang et al. LaCo: Large Language Model Pruning via Layer Collaps. 2024. [[arxiv]](https://arxiv.org/abs/2402.11187)
+1. Bhardwaj et al. Language Models are Homer Simpson! Safety Re-Alignment of Fine-tuned Language Models through Task Arithmetic. 2024. [[arxiv]](https://arxiv.org/abs/2402.11746)
+1. Yang et al. Enhancing Empathetic Response Generation by Augmenting LLMs with Small-scale Empathetic Models. 2024. [[arxiv]](https://arxiv.org/abs/2402.11801)
+1. Yi et al. Generation Meets Verification: Accelerating Large Language Model Inference with Smart Parallel Auto-Correct Decoding. 2024. [[arxiv]](https://arxiv.org/abs/2402.11809)
+1. Cao et al. Head-wise Shareable Attention for Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2402.11819)
+1. Zhang et al. Enhancing Multilingual Capabilities of Large Language Models through Self-Distillation from Resource-Rich Languages. 2024. [[arxiv]](https://arxiv.org/abs/2402.12204)
+1. Kim et al. Efficient and Effective Vocabulary Expansion Towards Multilingual Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2402.14714)
+1. Yu et al. KIEval: A Knowledge-grounded Interactive Evaluation Framework for Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2402.15043)
+1. Huang et al. Key-Point-Driven Data Synthesis with its Enhancement on Mathematical Reasoning. 2024. [[arxiv]](https://arxiv.org/abs/2403.02333)
+1. Duan et al. Negating Negatives: Alignment without Human Positive Samples via Distributional Dispreference Optimization. 2024. [[arxiv]](https://arxiv.org/abs/2403.03419)
+1. Xie and Schwertfeger. Empowering Robotics with Large Language Models: osmAG Map Comprehension with LLMs. 2024. [[arxiv]](https://arxiv.org/abs/2403.08228)
+1. **[StarWhisper](https://github.com/Yu-Yang-Li/StarWhisper)**: 天文大模型 StarWhisper，基于 ChatGLM2-6B 和 Qwen-14B 在天文数据上微调而得。
+1. **[DISC-LawLLM](https://github.com/FudanDISC/DISC-LawLLM)**: 中文法律领域大模型 DISC-LawLLM，基于 Baichuan-13B 微调而得，具有法律推理和知识检索能力。
+1. **[Sunsimiao](https://github.com/thomas-yanxin/Sunsimiao)**: 孙思邈中文医疗大模型 Sumsimiao，基于 Baichuan-7B 和 ChatGLM-6B 在中文医疗数据上微调而得。
+1. **[CareGPT](https://github.com/WangRongsheng/CareGPT)**: 医疗大模型项目 CareGPT，基于 LLaMA2-7B 和 Baichuan-13B 在中文医疗数据上微调而得。
+1. **[MachineMindset](https://github.com/PKU-YuanGroup/Machine-Mindset/)**：MBTI性格大模型项目，根据数据集与训练方式让任意 LLM 拥有 16 个不同的性格类型。
 
 > [!TIP]
 > 如果您有项目希望添加至上述列表，请通过邮件联系或者创建一个 PR。
@@ -581,18 +695,19 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
 
 本仓库的代码依照 [Apache-2.0](LICENSE) 协议开源。
 
-使用模型权重时，请遵循对应的模型协议：[Baichuan2](https://huggingface.co/baichuan-inc/Baichuan2-7B-Base/blob/main/Community%20License%20for%20Baichuan%202%20Model.pdf) / [BLOOM](https://huggingface.co/spaces/bigscience/license) / [ChatGLM3](https://github.com/THUDM/ChatGLM3/blob/main/MODEL_LICENSE) / [DeepSeek](https://github.com/deepseek-ai/DeepSeek-LLM/blob/main/LICENSE-MODEL) / [Falcon](https://huggingface.co/tiiuae/falcon-180B/blob/main/LICENSE.txt) / [InternLM2](https://github.com/InternLM/InternLM#license) / [LLaMA](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md) / [LLaMA-2](https://ai.meta.com/llama/license/) / [Mistral](LICENSE) / [Phi-1.5/2](https://huggingface.co/microsoft/phi-1_5/resolve/main/Research%20License.docx) / [Qwen](https://github.com/QwenLM/Qwen/blob/main/Tongyi%20Qianwen%20LICENSE%20AGREEMENT) / [XVERSE](https://github.com/xverse-ai/XVERSE-13B/blob/main/MODEL_LICENSE.pdf) / [Yi](https://huggingface.co/01-ai/Yi-6B/blob/main/LICENSE) / [Yuan](https://github.com/IEIT-Yuan/Yuan-2.0/blob/main/LICENSE-Yuan)
+使用模型权重时，请遵循对应的模型协议：[Baichuan2](https://huggingface.co/baichuan-inc/Baichuan2-7B-Base/blob/main/Community%20License%20for%20Baichuan%202%20Model.pdf) / [BLOOM](https://huggingface.co/spaces/bigscience/license) / [ChatGLM3](https://github.com/THUDM/ChatGLM3/blob/main/MODEL_LICENSE) / [DeepSeek](https://github.com/deepseek-ai/DeepSeek-LLM/blob/main/LICENSE-MODEL) / [Falcon](https://huggingface.co/tiiuae/falcon-180B/blob/main/LICENSE.txt) / [Gemma](https://ai.google.dev/gemma/terms) / [InternLM2](https://github.com/InternLM/InternLM#license) / [LLaMA](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md) / [LLaMA-2](https://ai.meta.com/llama/license/) / [Mistral](LICENSE) / [OLMo](LICENSE) / [Phi-1.5/2](https://huggingface.co/microsoft/phi-1_5/resolve/main/Research%20License.docx) / [Qwen](https://github.com/QwenLM/Qwen/blob/main/Tongyi%20Qianwen%20LICENSE%20AGREEMENT) / [StarCoder2](https://huggingface.co/spaces/bigcode/bigcode-model-license-agreement) / [XVERSE](https://github.com/xverse-ai/XVERSE-13B/blob/main/MODEL_LICENSE.pdf) / [Yi](https://huggingface.co/01-ai/Yi-6B/blob/main/LICENSE) / [Yuan](https://github.com/IEIT-Yuan/Yuan-2.0/blob/main/LICENSE-Yuan)
 
 ## 引用
 
 如果您觉得此项目有帮助，请考虑以下列格式引用
 
 ```bibtex
-@Misc{llama-factory,
-  title = {LLaMA Factory},
-  author = {hiyouga},
-  howpublished = {\url{https://github.com/hiyouga/LLaMA-Factory}},
-  year = {2023}
+@article{zheng2024llamafactory,
+  title={LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models}, 
+  author={Yaowei Zheng and Richong Zhang and Junhao Zhang and Yanhan Ye and Zheyan Luo and Yongqiang Ma},
+  journal={arXiv preprint arXiv:2403.13372},
+  year={2024},
+  url={http://arxiv.org/abs/2403.13372}
 }
 ```
 

data/README.md

@@ -11,15 +11,23 @@ If you are using a custom dataset, please provide your dataset definition in the
   "folder": "the name of the folder of the dataset repository on the Hugging Face hub. (optional, default: None)",
   "ranking": "whether the dataset is a preference dataset or not. (default: false)",
   "formatting": "the format of the dataset. (optional, default: alpaca, can be chosen from {alpaca, sharegpt})",
-  "columns": {
-    "prompt": "the column name in the dataset containing the prompts. (default: instruction, for alpaca)",
-    "query": "the column name in the dataset containing the queries. (default: input, for alpaca)",
-    "response": "the column name in the dataset containing the responses. (default: output, for alpaca)",
-    "history": "the column name in the dataset containing the histories. (default: None, for alpaca)",
-    "messages": "the column name in the dataset containing the messages. (default: conversations, for sharegpt)",
-    "role": "the key in the message represents the identity. (default: from, for sharegpt)",
-    "content": "the key in the message represents the content. (default: value, for sharegpt)",
-    "system": "the column name in the dataset containing the system prompts. (default: None, for both)"
+  "columns (optional)": {
+    "prompt": "the column name in the dataset containing the prompts. (default: instruction)",
+    "query": "the column name in the dataset containing the queries. (default: input)",
+    "response": "the column name in the dataset containing the responses. (default: output)",
+    "history": "the column name in the dataset containing the histories. (default: None)",
+    "messages": "the column name in the dataset containing the messages. (default: conversations)",
+    "system": "the column name in the dataset containing the system prompts. (default: None)",
+    "tools": "the column name in the dataset containing the tool description. (default: None)"
+  },
+  "tags (optional, used for the sharegpt format)": {
+    "role_tag": "the key in the message represents the identity. (default: from)",
+    "content_tag": "the key in the message represents the content. (default: value)",
+    "user_tag": "the value of the role_tag represents the user. (default: human)",
+    "assistant_tag": "the value of the role_tag represents the assistant. (default: gpt)",
+    "observation_tag": "the value of the role_tag represents the tool results. (default: observation)",
+    "function_tag": "the value of the role_tag represents the function call. (default: function_call)",
+    "system_tag": "the value of the role_tag represents the system prompt. (default: system, can override system column)"
   }
 }
 ```
@@ -57,9 +65,9 @@ Regarding the above dataset, the `columns` in `dataset_info.json` should be:
 }
 ```
 
-where the `prompt` and `response` columns should contain non-empty values, represent instruction and response respectively. The `query` column will be concatenated with the `prompt` column and used as input for the model.
+The `query` column will be concatenated with the `prompt` column and used as the user prompt, then the user prompt would be `prompt\nquery`. The `response` column represents the model response.
 
-The `system` column will be used as the system prompt in the template. The `history` column is a list consisting string tuples representing query-response pairs in history. Note that the responses **in each round will be used for training**.
+The `system` column will be used as the system prompt. The `history` column is a list consisting string tuples representing prompt-response pairs in the history. Note that the responses in the history **will also be used for training**.
 
 For the pre-training datasets, only the `prompt` column will be used for training.
 
@@ -91,7 +99,8 @@ The dataset in sharegpt format should follow the below format:
         "value": "model response"
       }
     ],
-    "system": "system prompt (optional)"
+    "system": "system prompt (optional)",
+    "tools": "tool description (optional)"
   }
 ]
 ```
@@ -102,13 +111,18 @@ Regarding the above dataset, the `columns` in `dataset_info.json` should be:
 "dataset_name": {
   "columns": {
     "messages": "conversations",
-    "role": "from",
-    "content": "value",
-    "system": "system"
+    "system": "system",
+    "tools": "tools"
+  },
+  "tags": {
+    "role_tag": "from",
+    "content_tag": "value",
+    "user_tag": "human",
+    "assistant_tag": "gpt"
   }
 }
 ```
 
-where the `messages` column should be a list whose length is even, and follow the `u/a/u/a/u/a` order.
+where the `messages` column should be a list following the `u/a/u/a/u/a` order.
 
 Pre-training datasets and preference datasets are incompatible with the sharegpt format yet.

data/README_zh.md

@@ -11,15 +11,23 @@
   "folder": "Hugging Face 仓库的文件夹名称（可选，默认：None）",
   "ranking": "是否为偏好数据集（可选，默认：False）",
   "formatting": "数据集格式（可选，默认：alpaca，可以为 alpaca 或 sharegpt）",
-  "columns": {
-    "prompt": "数据集代表提示词的表头名称（默认：instruction，用于 alpaca 格式）",
-    "query": "数据集代表请求的表头名称（默认：input，用于 alpaca 格式）",
-    "response": "数据集代表回答的表头名称（默认：output，用于 alpaca 格式）",
-    "history": "数据集代表历史对话的表头名称（默认：None，用于 alpaca 格式）",
-    "messages": "数据集代表消息列表的表头名称（默认：conversations，用于 sharegpt 格式）",
-    "role": "消息中代表发送者身份的键名（默认：from，用于 sharegpt 格式）",
-    "content": "消息中代表文本内容的键名（默认：value，用于 sharegpt 格式）",
-    "system": "数据集代表系统提示的表头名称（默认：None，用于两种格式）"
+  "columns（可选）": {
+    "prompt": "数据集代表提示词的表头名称（默认：instruction）",
+    "query": "数据集代表请求的表头名称（默认：input）",
+    "response": "数据集代表回答的表头名称（默认：output）",
+    "history": "数据集代表历史对话的表头名称（默认：None）",
+    "messages": "数据集代表消息列表的表头名称（默认：conversations）",
+    "system": "数据集代表系统提示的表头名称（默认：None）",
+    "tools": "数据集代表工具描述的表头名称（默认：None）"
+  },
+  "tags（可选，用于 sharegpt 格式）": {
+    "role_tag": "消息中代表发送者身份的键名（默认：from）",
+    "content_tag": "消息中代表文本内容的键名（默认：value）",
+    "user_tag": "消息中代表用户的 role_tag（默认：human）",
+    "assistant_tag": "消息中代表助手的 role_tag（默认：gpt）",
+    "observation_tag": "消息中代表工具返回结果的 role_tag（默认：observation）",
+    "function_tag": "消息中代表工具调用的 role_tag（默认：function_call）",
+    "system_tag": "消息中代表系统提示的 role_tag（默认：system，会覆盖 system 列）"
   }
 }
 ```
@@ -57,9 +65,9 @@
 }
 ```
 
-其中 `prompt` 和 `response` 列应当是非空的字符串，分别代表用户指令和模型回答。`query` 列的内容将会和 `prompt` 列拼接作为模型输入。
+其中 `query` 列对应的内容会与 `prompt` 列对应的内容拼接后作为用户指令，即用户指令为 `prompt\nquery`。`response` 列对应的内容为模型回答。
 
-`system` 为模板中的系统提示词。`history` 列是由多个字符串二元组构成的列表，分别代表历史消息中每轮的指令和回答。注意每轮的模型回答**均会被用于训练**。
+`system` 列对应的内容将被作为系统提示词。`history` 列是由多个字符串二元组构成的列表，分别代表历史消息中每轮的指令和回答。注意历史消息中的回答**也会被用于训练**。
 
 对于预训练数据集，仅 `prompt` 列中的内容会用于模型训练。
 
@@ -91,7 +99,8 @@
         "value": "模型回答"
       }
     ],
-    "system": "系统提示词（选填）"
+    "system": "系统提示词（选填）",
+    "tools": "工具描述（选填）"
   }
 ]
 ```
@@ -102,13 +111,18 @@
 "数据集名称": {
   "columns": {
     "messages": "conversations",
-    "role": "from",
-    "content": "value",
-    "system": "system"
+    "system": "system",
+    "tools": "tools"
+  },
+  "tags": {
+    "role_tag": "from",
+    "content_tag": "value",
+    "user_tag": "human",
+    "assistant_tag": "gpt"
   }
 }
 ```
 
-其中 `messages` 列必须为偶数长度的列表，且符合 `用户/模型/用户/模型/用户/模型` 的顺序。
+其中 `messages` 列应当是一个列表，且符合 `用户/模型/用户/模型/用户/模型` 的顺序。
 
 预训练数据集和偏好数据集尚不支持 sharegpt 格式。

data/alpaca_data_zh_51k.json.REMOVED.git-id

@@ -1 +1 @@
-fc9a6a3458caca2af8dafc6181773fe10c6d8657
+34c723573fbc2d7601f6d9c882ccf5aa4f9bcc4b

data/belle_multiturn/belle_multiturn.py

@@ -1,7 +1,10 @@
+import os
 import json
 import datasets
 
 
+_HF_ENDPOINT = os.getenv("HF_ENDPOINT", "https://huggingface.co")
+
 _DESCRIPTION = "BELLE multiturn chat dataset."
 
 _CITATION = """\
@@ -13,9 +16,9 @@ _CITATION = """\
 }
 """
 
-_HOMEPAGE = "https://huggingface.co/datasets/BelleGroup/multiturn_chat_0.8M"
+_HOMEPAGE = "{}/datasets/BelleGroup/multiturn_chat_0.8M".format(_HF_ENDPOINT)
 _LICENSE = "gpl-3.0"
-_URL = "https://huggingface.co/datasets/BelleGroup/multiturn_chat_0.8M/resolve/main/multiturn_chat_0.8M.json"
+_URL = "{}/datasets/BelleGroup/multiturn_chat_0.8M/resolve/main/multiturn_chat_0.8M.json".format(_HF_ENDPOINT)
 
 
 class BelleMultiturn(datasets.GeneratorBasedBuilder):

data/example_dataset/example_dataset.py

@@ -1,6 +1,6 @@
 import json
 import datasets
-from typing import Any, Dict, List
+from typing import Any, Dict, Generator, List, Tuple
 
 
 _DESCRIPTION = "An example of dataset."
@@ -40,7 +40,7 @@ class ExampleDataset(datasets.GeneratorBasedBuilder):
             )
         ]
 
-    def _generate_examples(self, filepath: str) -> Dict[int, Dict[str, Any]]:
+    def _generate_examples(self, filepath: str) -> Generator[Tuple[int, Dict[str, Any]], None, None]:
         example_dataset = json.load(open(filepath, "r", encoding="utf-8"))
         for key, example in enumerate(example_dataset):
             yield key, example

data/hh_rlhf_en/hh_rlhf_en.py

@@ -1,13 +1,14 @@
+import os
 import json
 import datasets
 from typing import List
 
-
+_HF_ENDPOINT = os.getenv("HF_ENDPOINT", "https://huggingface.co")
 _DESCRIPTION = "Human preference data about helpfulness and harmlessness."
 _CITATION = ""
-_HOMEPAGE = "https://huggingface.co/datasets/Anthropic/hh-rlhf"
+_HOMEPAGE = "{}/datasets/Anthropic/hh-rlhf".format(_HF_ENDPOINT)
 _LICENSE = "mit"
-_URL = "https://huggingface.co/datasets/Anthropic/hh-rlhf/resolve/main/"
+_URL = "{}/datasets/Anthropic/hh-rlhf/resolve/main/".format(_HF_ENDPOINT)
 _URLS = {
     "train": [
         _URL + "harmless-base/train.jsonl.gz",

data/orca_rlhf.json.REMOVED.git-id

@@ -0,0 +1 @@
+736bcedea2b24a1414765c6d69cbdafaea839f3c

data/ultra_chat/ultra_chat.py

@@ -1,7 +1,9 @@
+import os
 import json
 import datasets
 from typing import List
 
+_HF_ENDPOINT = os.getenv("HF_ENDPOINT", "https://huggingface.co")
 
 _DESCRIPTION = "UltraChat: Large-scale, Informative, and Diverse Multi-round Dialogue Data."
 
@@ -16,9 +18,9 @@ _CITATION = """\
 }
 """
 
-_HOMEPAGE = "https://huggingface.co/datasets/stingning/ultrachat"
+_HOMEPAGE = "{}/datasets/stingning/ultrachat".format(_HF_ENDPOINT)
 _LICENSE = "cc-by-nc-4.0"
-_BASE_DATA_URL = "https://huggingface.co/datasets/stingning/ultrachat/resolve/main/train_{idx}.jsonl"
+_BASE_DATA_URL = "{}/datasets/stingning/ultrachat/resolve/main/train_{{idx}}.jsonl".format(_HF_ENDPOINT)
 
 
 class UltraChat(datasets.GeneratorBasedBuilder):

docker-compose.yml

@@ -0,0 +1,23 @@
+version: '3.8'
+
+services:
+  llama-factory:
+    build:
+      dockerfile: Dockerfile
+      context: .
+    container_name: llama_factory
+    volumes:
+      - ./hf_cache:/root/.cache/huggingface/
+      - ./data:/app/data
+      - ./output:/app/output
+    ports:
+      - "7860:7860"
+    ipc: host
+    deploy:
+      resources:
+        reservations:
+          devices:
+          - driver: nvidia
+            count: "all"
+            capabilities: [gpu]
+    restart: unless-stopped

examples/accelerate/fsdp_config.yaml

@@ -0,0 +1,25 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+distributed_type: FSDP
+downcast_bf16: 'no'
+fsdp_config:
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_backward_prefetch: BACKWARD_PRE
+  fsdp_cpu_ram_efficient_loading: true
+  fsdp_forward_prefetch: false
+  fsdp_offload_params: true
+  fsdp_sharding_strategy: FULL_SHARD
+  fsdp_state_dict_type: SHARDED_STATE_DICT
+  fsdp_sync_module_states: true
+  fsdp_use_orig_params: false
+machine_rank: 0
+main_training_function: main
+mixed_precision: fp16
+num_machines: 1
+num_processes: 2
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false

examples/accelerate/master_config.yaml

@@ -0,0 +1,18 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+distributed_type: MULTI_GPU
+downcast_bf16: 'no'
+gpu_ids: all
+machine_rank: 0
+main_process_ip: 192.168.0.1
+main_process_port: 29555
+main_training_function: main
+mixed_precision: fp16
+num_machines: 2
+num_processes: 16
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false

examples/accelerate/single_config.yaml

@@ -0,0 +1,16 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+distributed_type: MULTI_GPU
+downcast_bf16: 'no'
+gpu_ids: all
+machine_rank: 0
+main_training_function: main
+mixed_precision: fp16
+num_machines: 1
+num_processes: 4
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false

examples/accelerate/slave_config.yaml

@@ -0,0 +1,18 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+distributed_type: MULTI_GPU
+downcast_bf16: 'no'
+gpu_ids: all
+machine_rank: 1
+main_process_ip: 192.168.0.1
+main_process_port: 29555
+main_training_function: main
+mixed_precision: fp16
+num_machines: 2
+num_processes: 16
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false

examples/extras/galore/adamw.sh

@@ -0,0 +1,31 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../../src/train_bash.py \
+    --stage sft \
+    --do_train \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --dataset alpaca_gpt4_en,glaive_toolcall \
+    --dataset_dir ../../../data \
+    --template default \
+    --finetuning_type full \
+    --output_dir ../../../saves/LLaMA2-7B/galore/sft \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --preprocessing_num_workers 16 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 1 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --warmup_steps 20 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --load_best_model_at_end \
+    --learning_rate 5e-5 \
+    --num_train_epochs 3.0 \
+    --max_samples 3000 \
+    --val_size 0.1 \
+    --plot_loss \
+    --fp16

examples/extras/galore/adamw_8bit_bf16.sh

@@ -0,0 +1,32 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../../src/train_bash.py \
+    --stage sft \
+    --do_train \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --dataset alpaca_gpt4_en,glaive_toolcall \
+    --dataset_dir ../../../data \
+    --template default \
+    --finetuning_type full \
+    --optim adamw_8bit \
+    --output_dir ../../../saves/LLaMA2-7B/galore/sft \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --preprocessing_num_workers 16 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 1 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --warmup_steps 20 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --load_best_model_at_end \
+    --learning_rate 5e-5 \
+    --num_train_epochs 3.0 \
+    --max_samples 3000 \
+    --val_size 0.1 \
+    --plot_loss \
+    --pure_bf16

examples/extras/galore/galore_adamw.sh

@@ -0,0 +1,35 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../../src/train_bash.py \
+    --stage sft \
+    --do_train \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --dataset alpaca_gpt4_en,glaive_toolcall \
+    --dataset_dir ../../../data \
+    --template default \
+    --finetuning_type full \
+    --use_galore \
+    --galore_layerwise \
+    --galore_target mlp,self_attn \
+    --galore_rank 128 \
+    --output_dir ../../../saves/LLaMA2-7B/galore/sft \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --preprocessing_num_workers 16 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 1 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --warmup_steps 20 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --load_best_model_at_end \
+    --learning_rate 5e-5 \
+    --num_train_epochs 3.0 \
+    --max_samples 3000 \
+    --val_size 0.1 \
+    --plot_loss \
+    --fp16

examples/extras/galore/galore_adamw_8bit_bf16.sh

@@ -0,0 +1,36 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../../src/train_bash.py \
+    --stage sft \
+    --do_train \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --dataset alpaca_gpt4_en,glaive_toolcall \
+    --dataset_dir ../../../data \
+    --template default \
+    --finetuning_type full \
+    --optim adamw_8bit \
+    --use_galore \
+    --galore_layerwise \
+    --galore_target mlp,self_attn \
+    --galore_rank 128 \
+    --output_dir ../../../saves/LLaMA2-7B/galore/sft \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --preprocessing_num_workers 16 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 1 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --warmup_steps 20 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --load_best_model_at_end \
+    --learning_rate 5e-5 \
+    --num_train_epochs 3.0 \
+    --max_samples 3000 \
+    --val_size 0.1 \
+    --plot_loss \
+    --pure_bf16

examples/extras/llama_pro/expand.sh

@@ -0,0 +1,6 @@
+#!/bin/bash
+
+python ../../../scripts/llama_pro.py \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --output_dir ../../../models/llama2-7b-pro \
+    --num_expand 8

examples/extras/llama_pro/sft.sh

@@ -0,0 +1,34 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../../src/train_bash.py \
+    --stage sft \
+    --do_train \
+    --model_name_or_path ../../../models/llama2-7b-pro \
+    --dataset alpaca_gpt4_en,glaive_toolcall \
+    --dataset_dir ../../../data \
+    --template default \
+    --finetuning_type freeze \
+    --name_module_trainable all \
+    --num_layer_trainable 8 \
+    --use_llama_pro \
+    --output_dir ../../../saves/LLaMA2-7B-Pro/lora/sft \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --preprocessing_num_workers 16 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 8 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --warmup_steps 20 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --load_best_model_at_end \
+    --learning_rate 5e-5 \
+    --num_train_epochs 3.0 \
+    --max_samples 3000 \
+    --val_size 0.1 \
+    --plot_loss \
+    --fp16

examples/extras/loraplus/sft.sh

@@ -0,0 +1,33 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../src/train_bash.py \
+    --stage sft \
+    --do_train \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --dataset alpaca_gpt4_en,glaive_toolcall \
+    --dataset_dir ../../data \
+    --template default \
+    --finetuning_type lora \
+    --lora_target q_proj,v_proj \
+    --output_dir ../../saves/LLaMA2-7B/loraplus/sft \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --preprocessing_num_workers 16 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 8 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --warmup_steps 20 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --load_best_model_at_end \
+    --learning_rate 5e-5 \
+    --num_train_epochs 3.0 \
+    --max_samples 3000 \
+    --val_size 0.1 \
+    --plot_loss \
+    --fp16 \
+    --loraplus_lr_ratio 16.0

examples/fsdp_qlora/README.md

@@ -0,0 +1,5 @@
+```bash
+pip install git+https://github.com/huggingface/transformers.git
+pip install "accelerate>=0.28.0"
+pip install "bitsandbytes>=0.43.0"
+```

examples/fsdp_qlora/fsdp.sh

@@ -0,0 +1,33 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0,1 accelerate launch \
+    --config_file ../accelerate/fsdp_config.yaml \
+    ../../src/train_bash.py \
+    --stage sft \
+    --do_train \
+    --model_name_or_path meta-llama/Llama-2-70b-hf \
+    --dataset alpaca_gpt4_en,glaive_toolcall \
+    --dataset_dir ../../data \
+    --template default \
+    --finetuning_type lora \
+    --lora_target q_proj,v_proj \
+    --output_dir ../../saves/LLaMA2-70B/lora/sft \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 8 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --load_best_model_at_end \
+    --learning_rate 5e-5 \
+    --num_train_epochs 3.0 \
+    --max_samples 3000 \
+    --val_size 0.1 \
+    --quantization_bit 4 \
+    --plot_loss \
+    --fp16

examples/full_multi_gpu/multi_node.sh

@@ -0,0 +1,38 @@
+#!/bin/bash
+
+python -m torch.distributed.run \
+    --nproc_per_node $NPROC_PER_NODE \
+    --nnodes $NNODES \
+    --node_rank $RANK \
+    --master_addr $MASTER_ADDR \
+    --master_port $MASTER_PORT \
+    ../../src/train_bash.py \
+    --deepspeed ../deepspeed/ds_z3_config.json \
+    --stage sft \
+    --do_train \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --dataset alpaca_gpt4_en,glaive_toolcall \
+    --dataset_dir ../../data \
+    --template default \
+    --finetuning_type full \
+    --output_dir ../../saves/LLaMA2-7B/full/sft \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --preprocessing_num_workers 16 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 2 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --warmup_steps 20 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --learning_rate 5e-5 \
+    --num_train_epochs 3.0 \
+    --max_samples 3000 \
+    --val_size 0.1 \
+    --ddp_timeout 1800000 \
+    --plot_loss \
+    --fp16

examples/full_multi_gpu/single_node.sh

@@ -0,0 +1,32 @@
+#!/bin/bash
+
+deepspeed --num_gpus 4 ../../src/train_bash.py \
+    --deepspeed ../deepspeed/ds_z3_config.json \
+    --stage sft \
+    --do_train \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --dataset alpaca_gpt4_en,glaive_toolcall \
+    --dataset_dir ../../data \
+    --template default \
+    --finetuning_type full \
+    --output_dir ../../saves/LLaMA2-7B/full/sft \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --preprocessing_num_workers 16 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 2 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --warmup_steps 20 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --learning_rate 5e-5 \
+    --num_train_epochs 3.0 \
+    --max_samples 3000 \
+    --val_size 0.1 \
+    --ddp_timeout 1800000 \
+    --plot_loss \
+    --fp16

examples/lora_multi_gpu/multi_node.sh

@@ -0,0 +1,35 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0,1,2,3 accelerate launch \
+    --config_file ../accelerate/master_config.yaml \
+    ../../src/train_bash.py \
+    --stage sft \
+    --do_train \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --dataset alpaca_gpt4_en,glaive_toolcall \
+    --dataset_dir ../../data \
+    --template default \
+    --finetuning_type lora \
+    --lora_target q_proj,v_proj \
+    --output_dir ../../saves/LLaMA2-7B/lora/sft \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --preprocessing_num_workers 16 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 2 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --warmup_steps 20 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --load_best_model_at_end \
+    --learning_rate 5e-5 \
+    --num_train_epochs 3.0 \
+    --max_samples 3000 \
+    --val_size 0.1 \
+    --ddp_timeout 1800000 \
+    --plot_loss \
+    --fp16

examples/lora_multi_gpu/single_node.sh

@@ -0,0 +1,35 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 accelerate launch \
+    --config_file ../accelerate/single_config.yaml \
+    ../../src/train_bash.py \
+    --stage sft \
+    --do_train \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --dataset alpaca_gpt4_en,glaive_toolcall \
+    --dataset_dir ../../data \
+    --template default \
+    --finetuning_type lora \
+    --lora_target q_proj,v_proj \
+    --output_dir ../../saves/LLaMA2-7B/lora/sft \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --preprocessing_num_workers 16 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 2 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --warmup_steps 20 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --load_best_model_at_end \
+    --learning_rate 5e-5 \
+    --num_train_epochs 3.0 \
+    --max_samples 3000 \
+    --val_size 0.1 \
+    --ddp_timeout 1800000 \
+    --plot_loss \
+    --fp16

examples/lora_single_gpu/README.md

@@ -0,0 +1,8 @@
+Usage:
+
+- `pretrain.sh`: do pre-train (optional)
+- `sft.sh`: do supervised fine-tune
+- `reward.sh`: do reward modeling (must after sft.sh)
+- `ppo.sh`: do PPO training (must after sft.sh and reward.sh)
+- `dpo.sh`: do DPO training (must after sft.sh)
+- `predict.sh`: do predict (must after sft.sh and dpo.sh)

examples/lora_single_gpu/dpo.sh

@@ -0,0 +1,35 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../src/train_bash.py \
+    --stage dpo \
+    --do_train \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --adapter_name_or_path ../../saves/LLaMA2-7B/lora/sft \
+    --create_new_adapter \
+    --dataset comparison_gpt4_en \
+    --dataset_dir ../../data \
+    --template default \
+    --finetuning_type lora \
+    --lora_target q_proj,v_proj \
+    --output_dir ../../saves/LLaMA2-7B/lora/dpo \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --preprocessing_num_workers 16 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 8 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --warmup_steps 20 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --load_best_model_at_end \
+    --learning_rate 1e-5 \
+    --num_train_epochs 1.0 \
+    --max_samples 1000 \
+    --val_size 0.1 \
+    --dpo_ftx 1.0 \
+    --plot_loss \
+    --fp16

examples/lora_single_gpu/ppo.sh

@@ -0,0 +1,32 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../src/train_bash.py \
+    --stage ppo \
+    --do_train \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --adapter_name_or_path ../../saves/LLaMA2-7B/lora/sft \
+    --create_new_adapter \
+    --dataset alpaca_gpt4_en \
+    --dataset_dir ../../data \
+    --template default \
+    --finetuning_type lora \
+    --lora_target q_proj,v_proj \
+    --reward_model ../../saves/LLaMA2-7B/lora/reward \
+    --output_dir ../../saves/LLaMA2-7B/lora/ppo \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 512 \
+    --preprocessing_num_workers 16 \
+    --per_device_train_batch_size 1 \
+    --gradient_accumulation_steps 8 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --save_steps 100 \
+    --learning_rate 1e-5 \
+    --num_train_epochs 1.0 \
+    --max_samples 1000 \
+    --top_k 0 \
+    --top_p 0.9 \
+    --max_new_tokens 256 \
+    --plot_loss \
+    --fp16

examples/lora_single_gpu/predict.sh

@@ -0,0 +1,19 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../src/train_bash.py \
+    --stage sft \
+    --do_predict \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --adapter_name_or_path ../../saves/LLaMA2-7B/lora/sft,../../saves/LLaMA2-7B/lora/dpo \
+    --dataset alpaca_gpt4_en,glaive_toolcall \
+    --dataset_dir ../../data \
+    --template default \
+    --finetuning_type lora \
+    --output_dir ../../saves/LLaMA2-7B/lora/predict \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --preprocessing_num_workers 16 \
+    --per_device_eval_batch_size 1 \
+    --max_samples 20 \
+    --predict_with_generate

examples/lora_single_gpu/pretrain.sh

@@ -0,0 +1,31 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../src/train_bash.py \
+    --stage pt \
+    --do_train \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --dataset c4_demo \
+    --dataset_dir ../../data \
+    --finetuning_type lora \
+    --lora_target q_proj,v_proj \
+    --output_dir ../../saves/LLaMA2-7B/lora/pretrain \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --preprocessing_num_workers 16 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 8 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --warmup_steps 20 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --load_best_model_at_end \
+    --learning_rate 5e-5 \
+    --num_train_epochs 3.0 \
+    --max_samples 10000 \
+    --val_size 0.1 \
+    --plot_loss \
+    --fp16

examples/lora_single_gpu/reward.sh

@@ -0,0 +1,33 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../src/train_bash.py \
+    --stage rm \
+    --do_train \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --adapter_name_or_path ../../saves/LLaMA2-7B/lora/sft \
+    --create_new_adapter \
+    --dataset comparison_gpt4_en \
+    --dataset_dir ../../data \
+    --template default \
+    --finetuning_type lora \
+    --lora_target q_proj,v_proj \
+    --output_dir ../../saves/LLaMA2-7B/lora/reward \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --preprocessing_num_workers 16 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 8 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --warmup_steps 20 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --learning_rate 1e-5 \
+    --num_train_epochs 1.0 \
+    --max_samples 5000 \
+    --val_size 0.1 \
+    --plot_loss \
+    --fp16

examples/lora_single_gpu/sft.sh

@@ -0,0 +1,32 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../src/train_bash.py \
+    --stage sft \
+    --do_train \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --dataset alpaca_gpt4_en,glaive_toolcall \
+    --dataset_dir ../../data \
+    --template default \
+    --finetuning_type lora \
+    --lora_target q_proj,v_proj \
+    --output_dir ../../saves/LLaMA2-7B/lora/sft \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --preprocessing_num_workers 16 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 8 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --warmup_steps 20 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --load_best_model_at_end \
+    --learning_rate 5e-5 \
+    --num_train_epochs 3.0 \
+    --max_samples 3000 \
+    --val_size 0.1 \
+    --plot_loss \
+    --fp16

examples/merge_lora/README.md

@@ -0,0 +1,4 @@
+Usage:
+
+- `merge.sh`: merge the lora weights
+- `quantize.sh`: quantize the model with AutoGPTQ (must after merge.sh, optional)

examples/merge_lora/merge.sh

@@ -0,0 +1,10 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../src/export_model.py \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --adapter_name_or_path ../../saves/LLaMA2-7B/lora/sft \
+    --template default \
+    --finetuning_type lora \
+    --export_dir ../../models/llama2-7b-sft \
+    --export_size 2 \
+    --export_legacy_format False

examples/merge_lora/quantize.sh

@@ -0,0 +1,10 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../src/export_model.py \
+    --model_name_or_path ../../models/llama2-7b-sft \
+    --template default \
+    --export_dir ../../models/llama2-7b-sft-int4 \
+    --export_quantization_bit 4 \
+    --export_quantization_dataset ../../data/c4_demo.json \
+    --export_size 2 \
+    --export_legacy_format False

examples/qlora_single_gpu/aqlm.sh

@@ -0,0 +1,30 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../src/train_bash.py \
+    --stage sft \
+    --do_train \
+    --model_name_or_path BlackSamorez/Llama-2-7b-AQLM-2Bit-1x16-hf \
+    --dataset alpaca_gpt4_en,glaive_toolcall \
+    --dataset_dir ../../data \
+    --template default \
+    --finetuning_type lora \
+    --lora_target q_proj,v_proj \
+    --output_dir ../../saves/LLaMA2-7B/lora/sft \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 8 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --load_best_model_at_end \
+    --learning_rate 5e-5 \
+    --num_train_epochs 3.0 \
+    --max_samples 3000 \
+    --val_size 0.1 \
+    --plot_loss \
+    --fp16

examples/qlora_single_gpu/awq.sh

@@ -0,0 +1,30 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../src/train_bash.py \
+    --stage sft \
+    --do_train \
+    --model_name_or_path TheBloke/Llama-2-7B-AWQ \
+    --dataset alpaca_gpt4_en,glaive_toolcall \
+    --dataset_dir ../../data \
+    --template default \
+    --finetuning_type lora \
+    --lora_target q_proj,v_proj \
+    --output_dir ../../saves/LLaMA2-7B/lora/sft \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 8 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --load_best_model_at_end \
+    --learning_rate 5e-5 \
+    --num_train_epochs 3.0 \
+    --max_samples 3000 \
+    --val_size 0.1 \
+    --plot_loss \
+    --fp16

examples/qlora_single_gpu/bitsandbytes.sh

@@ -0,0 +1,31 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../src/train_bash.py \
+    --stage sft \
+    --do_train \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --dataset alpaca_gpt4_en,glaive_toolcall \
+    --dataset_dir ../../data \
+    --template default \
+    --finetuning_type lora \
+    --lora_target q_proj,v_proj \
+    --output_dir ../../saves/LLaMA2-7B/lora/sft \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 8 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --load_best_model_at_end \
+    --learning_rate 5e-5 \
+    --num_train_epochs 3.0 \
+    --max_samples 3000 \
+    --val_size 0.1 \
+    --quantization_bit 4 \
+    --plot_loss \
+    --fp16

examples/qlora_single_gpu/gptq.sh

@@ -0,0 +1,30 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../src/train_bash.py \
+    --stage sft \
+    --do_train \
+    --model_name_or_path TheBloke/Llama-2-7B-GPTQ \
+    --dataset alpaca_gpt4_en,glaive_toolcall \
+    --dataset_dir ../../data \
+    --template default \
+    --finetuning_type lora \
+    --lora_target q_proj,v_proj \
+    --output_dir ../../saves/LLaMA2-7B/lora/sft \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 8 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --load_best_model_at_end \
+    --learning_rate 5e-5 \
+    --num_train_epochs 3.0 \
+    --max_samples 3000 \
+    --val_size 0.1 \
+    --plot_loss \
+    --fp16

pyproject.toml

@@ -2,23 +2,19 @@
 requires = ["setuptools>=61.0"]
 build-backend = "setuptools.build_meta"
 
-[tool.black]
-line-length = 119
-target-version = ["py38"]
-
 [tool.ruff]
+target-version = "py38"
+line-length = 119
+indent-width = 4
+
+[tool.ruff.lint]
 ignore = ["C408", "C901", "E501", "E731", "E741", "W605"]
 select = ["C", "E", "F", "I", "W"]
-line-length = 119
 
-[tool.ruff.isort]
+[tool.ruff.lint.isort]
 lines-after-imports = 2
 known-first-party = ["llmtuner"]
-
-[isort]
-default_section = "FIRSTPARTY"
-known_first_party = "llmtuner"
-known_third_party = [
+known-third-party = [
     "accelerate",
     "datasets",
     "gradio",
@@ -28,10 +24,10 @@ known_third_party = [
     "transformers",
     "trl"
 ]
-line_length = 119
-lines_after_imports = 2
-multi_line_output = 3
-include_trailing_comma = true
-force_grid_wrap = 0
-use_parentheses = true
-ensure_newline_before_comments = true
+
+[tool.ruff.format]
+quote-style = "double"
+indent-style = "space"
+docstring-code-format = true
+skip-magic-trailing-comma = false
+line-ending = "auto"

requirements.txt

@@ -1,19 +1,18 @@
 torch>=1.13.1
-transformers>=4.36.2
+transformers>=4.37.2
 datasets>=2.14.3
-accelerate>=0.21.0
-peft>=0.7.0
-trl>=0.7.6
+accelerate>=0.27.2
+peft>=0.9.0
+trl>=0.8.1
 gradio>=3.38.0,<4.0.0
 scipy
 einops
 sentencepiece
 protobuf
-jieba
-rouge-chinese
-nltk
 uvicorn
 pydantic
 fastapi
 sse-starlette
 matplotlib
+fire
+galore-torch

scripts/cal_lr.py

@@ -10,7 +10,7 @@ import fire
 import torch
 from torch.utils.data import DataLoader
 from tqdm import tqdm
-from transformers import DataCollatorForSeq2Seq
+from transformers import DataCollatorForLanguageModeling, DataCollatorForSeq2Seq
 
 from llmtuner.data import get_dataset
 from llmtuner.extras.constants import IGNORE_INDEX
@@ -24,26 +24,35 @@ BASE_BS = 4_000_000  # from llama paper
 
 def calculate_lr(
     model_name_or_path: str,
-    dataset: str,
-    cutoff_len: int,  # i.e. maximum input length during training
     batch_size: int,  # total batch size, namely (batch size * gradient accumulation * world size)
-    is_mistral: bool,  # mistral model uses a smaller learning rate,
+    stage: Optional[str] = "sft",
+    dataset: Optional[str] = "alpaca_en",
     dataset_dir: Optional[str] = "data",
+    template: Optional[str] = "default",
+    cutoff_len: Optional[int] = 1024,  # i.e. maximum input length during training
+    is_mistral: Optional[bool] = False,  # mistral model uses a smaller learning rate,
 ):
     model_args, data_args, training_args, finetuning_args, _ = get_train_args(
         dict(
-            stage="sft",
+            stage=stage,
             model_name_or_path=model_name_or_path,
             dataset=dataset,
             dataset_dir=dataset_dir,
-            template="default",
+            template=template,
             cutoff_len=cutoff_len,
             output_dir="dummy_dir",
+            overwrite_cache=True,
         )
     )
     _, tokenizer = load_model_and_tokenizer(model_args, finetuning_args, is_trainable=False, add_valuehead=False)
-    trainset = get_dataset(tokenizer, model_args, data_args, training_args, stage="sft")
-    data_collator = DataCollatorForSeq2Seq(tokenizer=tokenizer, label_pad_token_id=IGNORE_INDEX)
+    trainset = get_dataset(tokenizer, model_args, data_args, training_args, stage=stage)
+    if stage == "pt":
+        data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)
+    elif stage == "sft":
+        data_collator = DataCollatorForSeq2Seq(tokenizer=tokenizer, label_pad_token_id=IGNORE_INDEX)
+    else:
+        raise NotImplementedError
+
     dataloader = DataLoader(
         dataset=trainset, batch_size=batch_size, shuffle=True, collate_fn=data_collator, pin_memory=True
     )

scripts/length_cdf.py

@@ -0,0 +1,52 @@
+# coding=utf-8
+# Calculates the distribution of the input lengths in the dataset.
+# Usage: python length_cdf.py --model_name_or_path path_to_model --dataset alpaca_en --template default
+
+from collections import defaultdict
+from typing import Optional
+
+import fire
+from tqdm import tqdm
+
+from llmtuner.data import get_dataset
+from llmtuner.hparams import get_train_args
+from llmtuner.model import load_model_and_tokenizer
+
+
+def length_cdf(
+    model_name_or_path: str,
+    dataset: Optional[str] = "alpaca_en",
+    dataset_dir: Optional[str] = "data",
+    template: Optional[str] = "default",
+    interval: Optional[int] = 1000,
+):
+    model_args, data_args, training_args, finetuning_args, _ = get_train_args(
+        dict(
+            stage="sft",
+            model_name_or_path=model_name_or_path,
+            dataset=dataset,
+            dataset_dir=dataset_dir,
+            template=template,
+            cutoff_len=1_000_000,
+            output_dir="dummy_dir",
+            overwrite_cache=True,
+        )
+    )
+    _, tokenizer = load_model_and_tokenizer(model_args, finetuning_args, is_trainable=False, add_valuehead=False)
+    trainset = get_dataset(tokenizer, model_args, data_args, training_args, stage="sft")
+    total_num = len(trainset)
+    length_dict = defaultdict(int)
+    for sample in tqdm(trainset["input_ids"]):
+        length_dict[len(sample) // interval * interval] += 1
+
+    length_tuples = list(length_dict.items())
+    length_tuples.sort()
+    count_accu, prob_accu = 0, 0
+    for length, count in length_tuples:
+        count_accu += count
+        prob_accu += count / total_num * 100
+        print("{:d} ({:.2f}%) samples have length < {}.".format(count_accu, prob_accu, length + interval))
+
+
+if __name__ == "__main__":
+    fire.Fire(length_cdf)

scripts/llama_pro.py

@@ -0,0 +1,115 @@
+# coding=utf-8
+# Performs block expansion for LLaMA, Mistral or Qwen1.5 models.
+# Usage: python llama_pro.py --model_name_or_path meta-llama/Llama-2-7b-hf --output_dir llama2_pro --num_expand 8
+# Inspired by: https://github.com/TencentARC/LLaMA-Pro/blob/main/scripts/block_expansion.py
+
+import json
+import os
+from collections import OrderedDict
+from typing import TYPE_CHECKING, Optional
+
+import fire
+import torch
+from safetensors.torch import save_file
+from tqdm import tqdm
+from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer
+from transformers.modeling_utils import (
+    SAFE_WEIGHTS_INDEX_NAME,
+    SAFE_WEIGHTS_NAME,
+    WEIGHTS_INDEX_NAME,
+    WEIGHTS_NAME,
+    shard_checkpoint,
+)
+
+
+if TYPE_CHECKING:
+    from transformers import PretrainedConfig, PreTrainedModel
+
+
+def change_name(name: str, old_index: int, new_index: int) -> str:
+    return name.replace(".{:d}.".format(old_index), ".{:d}.".format(new_index))
+
+
+def block_expansion(
+    model_name_or_path: str,
+    output_dir: str,
+    num_expand: int,
+    shard_size: Optional[str] = "2GB",
+    save_safetensors: Optional[bool] = False,
+):
+    config: "PretrainedConfig" = AutoConfig.from_pretrained(model_name_or_path)
+    num_layers = getattr(config, "num_hidden_layers")
+    setattr(config, "num_hidden_layers", num_layers + num_expand)
+    config.save_pretrained(output_dir)
+
+    tokenizer = AutoTokenizer.from_pretrained(model_name_or_path)
+    tokenizer.save_pretrained(output_dir)
+
+    config: "PretrainedConfig" = AutoConfig.from_pretrained(model_name_or_path)  # load the original one
+    if save_safetensors:
+        setattr(config, "tie_word_embeddings", False)  # safetensors does not allow shared weights
+
+    model: "PreTrainedModel" = AutoModelForCausalLM.from_pretrained(
+        model_name_or_path,
+        config=config,
+        torch_dtype="auto",
+        trust_remote_code=True,
+        low_cpu_mem_usage=True,
+    )
+    state_dict = model.state_dict()
+
+    if num_layers % num_expand != 0:
+        raise ValueError("`num_layers` {} should be divisible by `num_expand` {}.".format(num_layers, num_expand))
+
+    split = num_layers // num_expand
+    layer_cnt = 0
+    output_state_dict = OrderedDict()
+    for i in range(num_layers):
+        for key, value in state_dict.items():
+            if ".{:d}.".format(i) in key:
+                output_state_dict[change_name(key, i, layer_cnt)] = value
+
+        print("Add layer {} copied from layer {}".format(layer_cnt, i))
+        layer_cnt += 1
+        if (i + 1) % split == 0:
+            for key, value in state_dict.items():
+                if ".{:d}.".format(i) in key:
+                    if "down_proj" in key or "o_proj" in key:
+                        output_state_dict[change_name(key, i, layer_cnt)] = torch.zeros_like(value)
+                    else:
+                        output_state_dict[change_name(key, i, layer_cnt)] = torch.clone(value)
+
+            print("Add layer {} expanded from layer {}".format(layer_cnt, i))
+            layer_cnt += 1
+
+    for key, value in state_dict.items():
+        if key not in output_state_dict:
+            output_state_dict[key] = value
+
+    weights_name = SAFE_WEIGHTS_NAME if save_safetensors else WEIGHTS_NAME
+    shards, index = shard_checkpoint(output_state_dict, max_shard_size=shard_size, weights_name=weights_name)
+
+    for shard_file, shard in tqdm(shards.items(), desc="Save weights"):
+        if save_safetensors:
+            save_file(shard, os.path.join(output_dir, shard_file), metadata={"format": "pt"})
+        else:
+            torch.save(shard, os.path.join(output_dir, shard_file))
+
+    if index is None:
+        print("Model weights saved in {}".format(os.path.join(output_dir, weights_name)))
+    else:
+        index_name = SAFE_WEIGHTS_INDEX_NAME if save_safetensors else WEIGHTS_INDEX_NAME
+        with open(os.path.join(output_dir, index_name), "w", encoding="utf-8") as f:
+            json.dump(index, f, indent=2, sort_keys=True)
+        print("Model weights saved in {}".format(output_dir))
+
+    print("Fine-tune this model with:")
+    print("  --model_name_or_path {} \\".format(output_dir))
+    print("  --finetuning_type freeze \\")
+    print("  --name_module_trainable all \\")
+    print("  --num_layer_trainable {} \\".format(num_expand))
+    print("  --use_llama_pro")
+
+
+if __name__ == "__main__":
+    fire.Fire(block_expansion)

scripts/llamafy_baichuan2.py

@@ -1,6 +1,6 @@
 # coding=utf-8
 # Converts the Baichuan2-7B model in the same format as LLaMA2-7B.
-# Usage: python llamafy_baichuan2.py --input_dir input --output_dir output --shard_size 10GB
+# Usage: python llamafy_baichuan2.py --input_dir input --output_dir output
 # Inspired by: https://huggingface.co/fireballoon/baichuan-llama-7b/blob/main/convert_baichuan_to_llama.py
 # Converted model: https://huggingface.co/hiyouga/Baichuan2-7B-Base-LLaMAfied
 
@@ -76,7 +76,9 @@ def save_config(input_dir: str, output_dir: str):
     print("Model config saved in {}".format(os.path.join(output_dir, CONFIG_NAME)))
 
 
-def llamafy_baichuan2(input_dir: str, output_dir: str, shard_size: str, save_safetensors: Optional[bool] = False):
+def llamafy_baichuan2(
+    input_dir: str, output_dir: str, shard_size: Optional[str] = "2GB", save_safetensors: Optional[bool] = False
+):
     try:
         os.makedirs(output_dir, exist_ok=False)
     except Exception as e:

scripts/llamafy_internlm2.py

@@ -1,6 +1,6 @@
 # coding=utf-8
 # Converts the InternLM2 model in the same format as LLaMA2.
-# Usage: python llamafy_internlm2.py --input_dir input --output_dir output --shard_size 10GB
+# Usage: python llamafy_internlm2.py --input_dir input --output_dir output
 # Warning: We have found that the converted model cannot infer correctly. It will be fixed later.
 
 import json
@@ -98,7 +98,9 @@ def save_config(input_dir: str, output_dir: str):
     print("Model config saved in {}".format(os.path.join(output_dir, CONFIG_NAME)))
 
 
-def llamafy_internlm2(input_dir: str, output_dir: str, shard_size: str, save_safetensors: Optional[bool] = False):
+def llamafy_internlm2(
+    input_dir: str, output_dir: str, shard_size: Optional[str] = "2GB", save_safetensors: Optional[bool] = False
+):
     try:
         os.makedirs(output_dir, exist_ok=False)
     except Exception as e:

scripts/llamafy_qwen.py

@@ -1,6 +1,6 @@
 # coding=utf-8
 # Converts the Qwen models in the same format as LLaMA2.
-# Usage: python llamafy_qwen.py --input_dir input --output_dir output --shard_size 10GB
+# Usage: python llamafy_qwen.py --input_dir input --output_dir output
 # Converted model: https://huggingface.co/hiyouga/Qwen-14B-Chat-LLaMAfied
 
 import json
@@ -128,7 +128,9 @@ def save_config(input_dir: str, output_dir: str, torch_dtype: str):
     print("Model config saved in {}".format(os.path.join(output_dir, CONFIG_NAME)))
 
 
-def llamafy_qwen(input_dir: str, output_dir: str, shard_size: str, save_safetensors: Optional[bool] = False):
+def llamafy_qwen(
+    input_dir: str, output_dir: str, shard_size: Optional[str] = "2GB", save_safetensors: Optional[bool] = False
+):
     try:
         os.makedirs(output_dir, exist_ok=False)
     except Exception as e:

scripts/loftq_init.py

@@ -26,7 +26,7 @@ class Shell(nn.Module):
 
 
 def unwrap_model(model: nn.Module, pattern=".base_layer") -> None:
-    for name in set([k.split(pattern)[0] for k, _ in model.named_modules() if pattern in k]):  # noqa: C403
+    for name in {k.split(pattern)[0] for k, _ in model.named_modules() if pattern in k}:
         parent_name = ".".join(name.split(".")[:-1])
         child_name = name.split(".")[-1]
         parent_module = model.get_submodule(parent_name)

setup.py

@@ -1,13 +1,14 @@
 import os
 import re
-from setuptools import setup, find_packages
+
+from setuptools import find_packages, setup
 
 
 def get_version():
     with open(os.path.join("src", "llmtuner", "__init__.py"), "r", encoding="utf-8") as f:
         file_content = f.read()
         pattern = r"{0}\W*=\W*\"([^\"]+)\"".format("__version__")
-        version, = re.findall(pattern, file_content)
+        (version,) = re.findall(pattern, file_content)
         return version
 
 
@@ -18,8 +19,21 @@ def get_requires():
         return lines
 
 
-def main():
+extra_require = {
+    "deepspeed": ["deepspeed"],
+    "metrics": ["nltk", "jieba", "rouge-chinese"],
+    "unsloth": ["torch==2.2.0", "unsloth[cu121-ampere-torch220]"],
+    "vllm": ["vllm>=0.3.3"],
+    "bitsandbytes": ["bitsandbytes>=0.39.0"],
+    "gptq": ["optimum>=1.16.0", "auto-gptq>=0.5.0"],
+    "awq": ["autoawq"],
+    "aqlm": ["aqlm[gpu]>=1.1.0"],
+    "qwen": ["tiktoken", "transformers_stream_generator"],
+    "quality": ["ruff"],
+}
 
+
+def main():
     setup(
         name="llmtuner",
         version=get_version(),
@@ -35,8 +49,9 @@ def main():
         packages=find_packages("src"),
         python_requires=">=3.8.0",
         install_requires=get_requires(),
+        extras_require=extra_require,
         classifiers=[
-            "Development Status :: 3 - Alpha",
+            "Development Status :: 4 - Beta",
             "Intended Audience :: Developers",
             "Intended Audience :: Education",
             "Intended Audience :: Science/Research",
@@ -46,8 +61,9 @@ def main():
             "Programming Language :: Python :: 3.8",
             "Programming Language :: Python :: 3.9",
             "Programming Language :: Python :: 3.10",
+            "Programming Language :: Python :: 3.11",
             "Topic :: Scientific/Engineering :: Artificial Intelligence",
-        ]
+        ],
     )
 
 

src/llmtuner/__init__.py

@@ -7,5 +7,5 @@ from .train import export_model, run_exp
 from .webui import create_ui, create_web_demo
 
 
-__version__ = "0.5.0"
+__version__ = "0.6.0"
 __all__ = ["create_app", "ChatModel", "Evaluator", "export_model", "run_exp", "create_ui", "create_web_demo"]

src/llmtuner/api/app.py

@@ -1,4 +1,3 @@
-import asyncio
 import json
 import os
 from contextlib import asynccontextmanager
@@ -73,7 +72,13 @@ def create_app(chat_model: "ChatModel") -> "FastAPI":
         allow_headers=["*"],
     )
 
-    semaphore = asyncio.Semaphore(int(os.environ.get("MAX_CONCURRENT", 1)))
+    role_mapping = {
+        Role.USER: DataRole.USER.value,
+        Role.ASSISTANT: DataRole.ASSISTANT.value,
+        Role.SYSTEM: DataRole.SYSTEM.value,
+        Role.FUNCTION: DataRole.FUNCTION.value,
+        Role.TOOL: DataRole.OBSERVATION.value,
+    }
 
     @app.get("/v1/models", response_model=ModelList)
     async def list_models():
@@ -82,49 +87,47 @@ def create_app(chat_model: "ChatModel") -> "FastAPI":
 
     @app.post("/v1/chat/completions", response_model=ChatCompletionResponse, status_code=status.HTTP_200_OK)
     async def create_chat_completion(request: ChatCompletionRequest):
-        if not chat_model.can_generate:
+        if not chat_model.engine.can_generate:
             raise HTTPException(status_code=status.HTTP_405_METHOD_NOT_ALLOWED, detail="Not allowed")
 
-        if len(request.messages) == 0 or request.messages[-1].role != Role.USER:
+        if len(request.messages) == 0:
             raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="Invalid length")
 
-        messages = [dictify(message) for message in request.messages]
-        if len(messages) and messages[0]["role"] == Role.SYSTEM:
-            system = messages.pop(0)["content"]
+        if request.messages[0].role == Role.SYSTEM:
+            system = request.messages.pop(0).content
         else:
-            system = None
+            system = ""
 
-        if len(messages) % 2 == 0:
+        if len(request.messages) % 2 == 0:
             raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="Only supports u/a/u/a/u...")
 
-        for i in range(len(messages)):
-            if i % 2 == 0 and messages[i]["role"] not in [Role.USER, Role.TOOL]:
+        input_messages = []
+        for i, message in enumerate(request.messages):
+            if i % 2 == 0 and message.role not in [Role.USER, Role.TOOL]:
                 raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="Invalid role")
-            elif i % 2 == 1 and messages[i]["role"] not in [Role.ASSISTANT, Role.FUNCTION]:
+            elif i % 2 == 1 and message.role not in [Role.ASSISTANT, Role.FUNCTION]:
                 raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="Invalid role")
-            elif messages[i]["role"] == Role.TOOL:
-                messages[i]["role"] = DataRole.OBSERVATION
+
+            input_messages.append({"role": role_mapping[message.role], "content": message.content})
 
         tool_list = request.tools
-        if len(tool_list):
+        if isinstance(tool_list, list) and len(tool_list):
             try:
-                tools = json.dumps([tool_list[0]["function"]], ensure_ascii=False)
+                tools = json.dumps([tool["function"] for tool in tool_list], ensure_ascii=False)
             except Exception:
                 raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="Invalid tools")
         else:
             tools = ""
 
-        async with semaphore:
-            loop = asyncio.get_running_loop()
-            return await loop.run_in_executor(None, chat_completion, messages, system, tools, request)
-
-    def chat_completion(messages: Sequence[Dict[str, str]], system: str, tools: str, request: ChatCompletionRequest):
         if request.stream:
-            generate = stream_chat_completion(messages, system, tools, request)
+            if tools:
+                raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="Cannot stream function calls.")
+
+            generate = stream_chat_completion(input_messages, system, tools, request)
             return EventSourceResponse(generate, media_type="text/event-stream")
 
-        responses = chat_model.chat(
-            messages,
+        responses = await chat_model.achat(
+            input_messages,
             system,
             tools,
             do_sample=request.do_sample,
@@ -138,7 +141,7 @@ def create_app(chat_model: "ChatModel") -> "FastAPI":
         choices = []
         for i, response in enumerate(responses):
             if tools:
-                result = chat_model.template.format_tools.extract(response.response_text)
+                result = chat_model.engine.template.format_tools.extract(response.response_text)
             else:
                 result = response.response_text
 
@@ -167,7 +170,7 @@ def create_app(chat_model: "ChatModel") -> "FastAPI":
 
         return ChatCompletionResponse(model=request.model, choices=choices, usage=usage)
 
-    def stream_chat_completion(
+    async def stream_chat_completion(
         messages: Sequence[Dict[str, str]], system: str, tools: str, request: ChatCompletionRequest
     ):
         choice_data = ChatCompletionResponseStreamChoice(
@@ -176,7 +179,7 @@ def create_app(chat_model: "ChatModel") -> "FastAPI":
         chunk = ChatCompletionStreamResponse(model=request.model, choices=[choice_data])
         yield jsonify(chunk)
 
-        for new_text in chat_model.stream_chat(
+        async for new_token in chat_model.astream_chat(
             messages,
             system,
             tools,
@@ -185,11 +188,11 @@ def create_app(chat_model: "ChatModel") -> "FastAPI":
             top_p=request.top_p,
             max_new_tokens=request.max_tokens,
         ):
-            if len(new_text) == 0:
+            if len(new_token) == 0:
                 continue
 
             choice_data = ChatCompletionResponseStreamChoice(
-                index=0, delta=ChatCompletionMessage(content=new_text), finish_reason=None
+                index=0, delta=ChatCompletionMessage(content=new_token), finish_reason=None
             )
             chunk = ChatCompletionStreamResponse(model=request.model, choices=[choice_data])
             yield jsonify(chunk)
@@ -203,18 +206,13 @@ def create_app(chat_model: "ChatModel") -> "FastAPI":
 
     @app.post("/v1/score/evaluation", response_model=ScoreEvaluationResponse, status_code=status.HTTP_200_OK)
     async def create_score_evaluation(request: ScoreEvaluationRequest):
-        if chat_model.can_generate:
+        if chat_model.engine.can_generate:
             raise HTTPException(status_code=status.HTTP_405_METHOD_NOT_ALLOWED, detail="Not allowed")
 
         if len(request.messages) == 0:
             raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="Invalid request")
 
-        async with semaphore:
-            loop = asyncio.get_running_loop()
-            return await loop.run_in_executor(None, get_score, request)
-
-    def get_score(request: ScoreEvaluationRequest):
-        scores = chat_model.get_scores(request.messages, max_length=request.max_length)
+        scores = await chat_model.aget_scores(request.messages, max_length=request.max_length)
         return ScoreEvaluationResponse(model=request.model, scores=scores)
 
     return app

src/llmtuner/api/protocol.py

@@ -59,7 +59,7 @@ class ChatCompletionMessage(BaseModel):
 class ChatCompletionRequest(BaseModel):
     model: str
     messages: List[ChatMessage]
-    tools: Optional[list] = []
+    tools: list = []
     do_sample: bool = True
     temperature: Optional[float] = None
     top_p: Optional[float] = None

src/llmtuner/chat/__init__.py

@@ -1,4 +1,5 @@
+from .base_engine import BaseEngine
 from .chat_model import ChatModel
 
 
-__all__ = ["ChatModel"]
+__all__ = ["BaseEngine", "ChatModel"]

src/llmtuner/chat/base_engine.py

@@ -0,0 +1,69 @@
+from abc import ABC, abstractmethod
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Any, AsyncGenerator, Dict, List, Literal, Optional, Sequence, Union
+
+
+if TYPE_CHECKING:
+    from transformers import PreTrainedModel, PreTrainedTokenizer
+
+    from ..data import Template
+    from ..extras.packages import is_vllm_available
+    from ..hparams import DataArguments, FinetuningArguments, GeneratingArguments, ModelArguments
+
+    if is_vllm_available():
+        from vllm import AsyncLLMEngine
+
+
+@dataclass
+class Response:
+    response_text: str
+    response_length: int
+    prompt_length: int
+    finish_reason: Literal["stop", "length"]
+
+
+class BaseEngine(ABC):
+    model: Union["PreTrainedModel", "AsyncLLMEngine"]
+    tokenizer: "PreTrainedTokenizer"
+    can_generate: bool
+    template: "Template"
+    generating_args: Dict[str, Any]
+
+    @abstractmethod
+    def __init__(
+        self,
+        model_args: "ModelArguments",
+        data_args: "DataArguments",
+        finetuning_args: "FinetuningArguments",
+        generating_args: "GeneratingArguments",
+    ) -> None: ...
+
+    @abstractmethod
+    async def start(
+        self,
+    ) -> None: ...
+
+    @abstractmethod
+    async def chat(
+        self,
+        messages: Sequence[Dict[str, str]],
+        system: Optional[str] = None,
+        tools: Optional[str] = None,
+        **input_kwargs,
+    ) -> List["Response"]: ...
+
+    @abstractmethod
+    async def stream_chat(
+        self,
+        messages: Sequence[Dict[str, str]],
+        system: Optional[str] = None,
+        tools: Optional[str] = None,
+        **input_kwargs,
+    ) -> AsyncGenerator[str, None]: ...
+
+    @abstractmethod
+    async def get_scores(
+        self,
+        batch_input: List[str],
+        **input_kwargs,
+    ) -> List[float]: ...

src/llmtuner/chat/chat_model.py

@@ -1,121 +1,55 @@
-from dataclasses import dataclass
+import asyncio
 from threading import Thread
-from typing import Any, Dict, Generator, List, Literal, Optional, Sequence, Tuple
+from typing import TYPE_CHECKING, Any, AsyncGenerator, Dict, Generator, List, Optional, Sequence
 
-import torch
-from transformers import GenerationConfig, TextIteratorStreamer
-
-from ..data import get_template_and_fix_tokenizer
-from ..extras.misc import get_logits_processor
 from ..hparams import get_infer_args
-from ..model import dispatch_model, load_model_and_tokenizer
+from .hf_engine import HuggingfaceEngine
+from .vllm_engine import VllmEngine
 
 
-@dataclass
-class Response:
-    response_text: str
-    response_length: int
-    prompt_length: int
-    finish_reason: Literal["stop", "length"]
+if TYPE_CHECKING:
+    from .base_engine import BaseEngine, Response
+
+
+def _start_background_loop(loop: asyncio.AbstractEventLoop) -> None:
+    asyncio.set_event_loop(loop)
+    loop.run_forever()
 
 
 class ChatModel:
     def __init__(self, args: Optional[Dict[str, Any]] = None) -> None:
-        model_args, data_args, finetuning_args, self.generating_args = get_infer_args(args)
-        self.can_generate = finetuning_args.stage == "sft"
-        self.model, self.tokenizer = load_model_and_tokenizer(
-            model_args, finetuning_args, is_trainable=False, add_valuehead=(not self.can_generate)
-        )
-        self.tokenizer.padding_side = "left" if self.can_generate else "right"
-        self.model = dispatch_model(self.model)
-        self.template = get_template_and_fix_tokenizer(data_args.template, self.tokenizer)
+        model_args, data_args, finetuning_args, generating_args = get_infer_args(args)
+        if model_args.infer_backend == "huggingface":
+            self.engine: "BaseEngine" = HuggingfaceEngine(model_args, data_args, finetuning_args, generating_args)
+        elif model_args.infer_backend == "vllm":
+            self.engine: "BaseEngine" = VllmEngine(model_args, data_args, finetuning_args, generating_args)
+        else:
+            raise NotImplementedError("Unknown backend: {}".format(model_args.infer_backend))
 
-    def _process_args(
-        self,
-        messages: Sequence[Dict[str, str]],
-        system: Optional[str] = None,
-        tools: Optional[str] = None,
-        **input_kwargs,
-    ) -> Tuple[Dict[str, Any], int]:
-        paired_messages = messages + [{"role": "assistant", "content": ""}]
-        prompt, _ = self.template.encode_oneturn(
-            tokenizer=self.tokenizer, messages=paired_messages, system=system, tools=tools
-        )
-        prompt_length = len(prompt)
-        input_ids = torch.tensor([prompt], device=self.model.device)
+        self._loop = asyncio.new_event_loop()
+        self._thread = Thread(target=_start_background_loop, args=(self._loop,), daemon=True)
+        self._thread.start()
+        asyncio.run_coroutine_threadsafe(self.engine.start(), self._loop)
 
-        do_sample = input_kwargs.pop("do_sample", None)
-        temperature = input_kwargs.pop("temperature", None)
-        top_p = input_kwargs.pop("top_p", None)
-        top_k = input_kwargs.pop("top_k", None)
-        num_return_sequences = input_kwargs.pop("num_return_sequences", None)
-        repetition_penalty = input_kwargs.pop("repetition_penalty", None)
-        max_length = input_kwargs.pop("max_length", None)
-        max_new_tokens = input_kwargs.pop("max_new_tokens", None)
-
-        generating_args = self.generating_args.to_dict()
-        generating_args.update(
-            dict(
-                do_sample=do_sample if do_sample is not None else generating_args["do_sample"],
-                temperature=temperature or generating_args["temperature"],
-                top_p=top_p or generating_args["top_p"],
-                top_k=top_k or generating_args["top_k"],
-                num_return_sequences=num_return_sequences or 1,
-                repetition_penalty=repetition_penalty or generating_args["repetition_penalty"],
-                eos_token_id=[self.tokenizer.eos_token_id] + self.tokenizer.additional_special_tokens_ids,
-                pad_token_id=self.tokenizer.pad_token_id,
-            )
-        )
-
-        if isinstance(num_return_sequences, int) and num_return_sequences > 1:
-            generating_args["do_sample"] = True
-
-        if max_length:
-            generating_args.pop("max_new_tokens", None)
-            generating_args["max_length"] = max_length
-
-        if max_new_tokens:
-            generating_args.pop("max_length", None)
-            generating_args["max_new_tokens"] = max_new_tokens
-
-        gen_kwargs = dict(
-            inputs=input_ids,
-            generation_config=GenerationConfig(**generating_args),
-            logits_processor=get_logits_processor(),
-        )
-
-        return gen_kwargs, prompt_length
-
-    @torch.inference_mode()
     def chat(
         self,
         messages: Sequence[Dict[str, str]],
         system: Optional[str] = None,
         tools: Optional[str] = None,
         **input_kwargs,
-    ) -> List[Response]:
-        gen_kwargs, prompt_length = self._process_args(messages, system, tools, **input_kwargs)
-        generate_output = self.model.generate(**gen_kwargs)
-        response_ids = generate_output[:, prompt_length:]
-        response = self.tokenizer.batch_decode(
-            response_ids, skip_special_tokens=True, clean_up_tokenization_spaces=True
-        )
-        results = []
-        for i in range(len(response)):
-            eos_index = (response_ids[i] == self.tokenizer.eos_token_id).nonzero()
-            response_length = (eos_index[0].item() + 1) if len(eos_index) else len(response_ids[i])
-            results.append(
-                Response(
-                    response_text=response[i],
-                    response_length=response_length,
-                    prompt_length=prompt_length,
-                    finish_reason="stop" if len(eos_index) else "length",
-                )
-            )
+    ) -> List["Response"]:
+        task = asyncio.run_coroutine_threadsafe(self.achat(messages, system, tools, **input_kwargs), self._loop)
+        return task.result()
 
-        return results
+    async def achat(
+        self,
+        messages: Sequence[Dict[str, str]],
+        system: Optional[str] = None,
+        tools: Optional[str] = None,
+        **input_kwargs,
+    ) -> List["Response"]:
+        return await self.engine.chat(messages, system, tools, **input_kwargs)
 
-    @torch.inference_mode()
     def stream_chat(
         self,
         messages: Sequence[Dict[str, str]],
@@ -123,39 +57,35 @@ class ChatModel:
         tools: Optional[str] = None,
         **input_kwargs,
     ) -> Generator[str, None, None]:
-        gen_kwargs, _ = self._process_args(messages, system, tools, **input_kwargs)
-        streamer = TextIteratorStreamer(self.tokenizer, timeout=60.0, skip_prompt=True, skip_special_tokens=True)
-        gen_kwargs["streamer"] = streamer
+        generator = self.astream_chat(messages, system, tools, **input_kwargs)
+        while True:
+            try:
+                task = asyncio.run_coroutine_threadsafe(generator.__anext__(), self._loop)
+                yield task.result()
+            except StopAsyncIteration:
+                break
 
-        thread = Thread(target=self.model.generate, kwargs=gen_kwargs)
-        thread.start()
+    async def astream_chat(
+        self,
+        messages: Sequence[Dict[str, str]],
+        system: Optional[str] = None,
+        tools: Optional[str] = None,
+        **input_kwargs,
+    ) -> AsyncGenerator[str, None]:
+        async for new_token in self.engine.stream_chat(messages, system, tools, **input_kwargs):
+            yield new_token
 
-        yield from streamer
+    def get_scores(
+        self,
+        batch_input: List[str],
+        **input_kwargs,
+    ) -> List[float]:
+        task = asyncio.run_coroutine_threadsafe(self.aget_scores(batch_input, **input_kwargs), self._loop)
+        return task.result()
 
-    @torch.inference_mode()
-    def get_scores(self, batch_input: List[str], **input_kwargs) -> List[float]:
-        max_length = input_kwargs.pop("max_length", None)
-        device = getattr(self.model.pretrained_model, "device", "cuda")
-
-        inputs = self.tokenizer(
-            batch_input,
-            padding=True,
-            truncation=True,
-            max_length=max_length or getattr(self.model.config, "max_position_embeddings", 1024),
-            return_tensors="pt",
-            add_special_tokens=True,
-        ).to(device)
-
-        input_ids: torch.Tensor = inputs["input_ids"]
-        _, _, values = self.model(**inputs, output_hidden_states=True, return_dict=True)
-
-        if getattr(self.model.config, "model_type", None) == "chatglm":
-            values = torch.transpose(values, 0, 1)
-
-        scores = []
-        for i in range(input_ids.size(0)):
-            end_indexes = (input_ids[i] != self.tokenizer.pad_token_id).nonzero()
-            end_index = end_indexes[-1].item() if len(end_indexes) else 0
-            scores.append(values[i, end_index].nan_to_num().item())
-
-        return scores
+    async def aget_scores(
+        self,
+        batch_input: List[str],
+        **input_kwargs,
+    ) -> List[float]:
+        return await self.engine.get_scores(batch_input, **input_kwargs)

src/llmtuner/chat/hf_engine.py

@@ -0,0 +1,263 @@
+import asyncio
+import concurrent.futures
+import os
+from threading import Thread
+from typing import TYPE_CHECKING, Any, AsyncGenerator, Callable, Dict, List, Optional, Sequence, Tuple
+
+import torch
+from transformers import GenerationConfig, TextIteratorStreamer
+
+from ..data import get_template_and_fix_tokenizer
+from ..extras.misc import get_logits_processor
+from ..model import load_model_and_tokenizer
+from .base_engine import BaseEngine, Response
+
+
+if TYPE_CHECKING:
+    from transformers import PreTrainedModel, PreTrainedTokenizer
+    from trl import PreTrainedModelWrapper
+
+    from ..data import Template
+    from ..hparams import DataArguments, FinetuningArguments, GeneratingArguments, ModelArguments
+
+
+class HuggingfaceEngine(BaseEngine):
+    def __init__(
+        self,
+        model_args: "ModelArguments",
+        data_args: "DataArguments",
+        finetuning_args: "FinetuningArguments",
+        generating_args: "GeneratingArguments",
+    ) -> None:
+        self.can_generate = finetuning_args.stage == "sft"
+        self.model, self.tokenizer = load_model_and_tokenizer(
+            model_args, finetuning_args, is_trainable=False, add_valuehead=(not self.can_generate)
+        )
+        self.tokenizer.padding_side = "left" if self.can_generate else "right"
+        self.template = get_template_and_fix_tokenizer(self.tokenizer, data_args.template)
+        self.generating_args = generating_args.to_dict()
+
+    @staticmethod
+    def _process_args(
+        model: "PreTrainedModel",
+        tokenizer: "PreTrainedTokenizer",
+        template: "Template",
+        generating_args: Dict[str, Any],
+        messages: Sequence[Dict[str, str]],
+        system: Optional[str] = None,
+        tools: Optional[str] = None,
+        input_kwargs: Optional[Dict[str, Any]] = {},
+    ) -> Tuple[Dict[str, Any], int]:
+        paired_messages = messages + [{"role": "assistant", "content": ""}]
+        prompt_ids, _ = template.encode_oneturn(
+            tokenizer=tokenizer, messages=paired_messages, system=system, tools=tools
+        )
+        prompt_length = len(prompt_ids)
+        inputs = torch.tensor([prompt_ids], device=model.device)
+
+        do_sample = input_kwargs.pop("do_sample", None)
+        temperature = input_kwargs.pop("temperature", None)
+        top_p = input_kwargs.pop("top_p", None)
+        top_k = input_kwargs.pop("top_k", None)
+        num_return_sequences = input_kwargs.pop("num_return_sequences", None)
+        repetition_penalty = input_kwargs.pop("repetition_penalty", None)
+        max_length = input_kwargs.pop("max_length", None)
+        max_new_tokens = input_kwargs.pop("max_new_tokens", None)
+
+        generating_args.update(
+            dict(
+                do_sample=do_sample if do_sample is not None else generating_args["do_sample"],
+                temperature=temperature or generating_args["temperature"],
+                top_p=top_p or generating_args["top_p"],
+                top_k=top_k or generating_args["top_k"],
+                num_return_sequences=num_return_sequences or 1,
+                repetition_penalty=repetition_penalty or generating_args["repetition_penalty"],
+                eos_token_id=[tokenizer.eos_token_id] + tokenizer.additional_special_tokens_ids,
+                pad_token_id=tokenizer.pad_token_id,
+            )
+        )
+
+        if isinstance(num_return_sequences, int) and num_return_sequences > 1:
+            generating_args["do_sample"] = True
+
+        if max_length:
+            generating_args.pop("max_new_tokens", None)
+            generating_args["max_length"] = max_length
+
+        if max_new_tokens:
+            generating_args.pop("max_length", None)
+            generating_args["max_new_tokens"] = max_new_tokens
+
+        gen_kwargs = dict(
+            inputs=inputs,
+            generation_config=GenerationConfig(**generating_args),
+            logits_processor=get_logits_processor(),
+        )
+
+        return gen_kwargs, prompt_length
+
+    @staticmethod
+    @torch.inference_mode()
+    def _chat(
+        model: "PreTrainedModel",
+        tokenizer: "PreTrainedTokenizer",
+        template: "Template",
+        generating_args: Dict[str, Any],
+        messages: Sequence[Dict[str, str]],
+        system: Optional[str] = None,
+        tools: Optional[str] = None,
+        input_kwargs: Optional[Dict[str, Any]] = {},
+    ) -> List["Response"]:
+        gen_kwargs, prompt_length = HuggingfaceEngine._process_args(
+            model, tokenizer, template, generating_args, messages, system, tools, input_kwargs
+        )
+        generate_output = model.generate(**gen_kwargs)
+        response_ids = generate_output[:, prompt_length:]
+        response = tokenizer.batch_decode(response_ids, skip_special_tokens=True, clean_up_tokenization_spaces=True)
+        results = []
+        for i in range(len(response)):
+            eos_index = (response_ids[i] == tokenizer.eos_token_id).nonzero()
+            response_length = (eos_index[0].item() + 1) if len(eos_index) else len(response_ids[i])
+            results.append(
+                Response(
+                    response_text=response[i],
+                    response_length=response_length,
+                    prompt_length=prompt_length,
+                    finish_reason="stop" if len(eos_index) else "length",
+                )
+            )
+
+        return results
+
+    @staticmethod
+    @torch.inference_mode()
+    def _stream_chat(
+        model: "PreTrainedModel",
+        tokenizer: "PreTrainedTokenizer",
+        template: "Template",
+        generating_args: Dict[str, Any],
+        messages: Sequence[Dict[str, str]],
+        system: Optional[str] = None,
+        tools: Optional[str] = None,
+        input_kwargs: Optional[Dict[str, Any]] = {},
+    ) -> Callable[[], str]:
+        gen_kwargs, _ = HuggingfaceEngine._process_args(
+            model, tokenizer, template, generating_args, messages, system, tools, input_kwargs
+        )
+        streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
+        gen_kwargs["streamer"] = streamer
+        thread = Thread(target=model.generate, kwargs=gen_kwargs, daemon=True)
+        thread.start()
+
+        def stream():
+            try:
+                return streamer.__next__()
+            except StopIteration:
+                raise StopAsyncIteration()
+
+        return stream
+
+    @staticmethod
+    @torch.inference_mode()
+    def _get_scores(
+        model: "PreTrainedModelWrapper",
+        tokenizer: "PreTrainedTokenizer",
+        batch_input: List[str],
+        input_kwargs: Optional[Dict[str, Any]] = {},
+    ) -> List[float]:
+        max_length = input_kwargs.pop("max_length", None)
+        device = getattr(model.pretrained_model, "device", "cuda")
+        inputs = tokenizer(
+            batch_input,
+            padding=True,
+            truncation=True,
+            max_length=max_length or getattr(model.config, "max_position_embeddings", 1024),
+            return_tensors="pt",
+            add_special_tokens=True,
+        ).to(device)
+
+        input_ids: torch.Tensor = inputs["input_ids"]
+        _, _, values = model(**inputs, output_hidden_states=True, return_dict=True)
+
+        if getattr(model.config, "model_type", None) == "chatglm":
+            values = torch.transpose(values, 0, 1)
+
+        scores = []
+        for i in range(input_ids.size(0)):
+            end_indexes = (input_ids[i] != tokenizer.pad_token_id).nonzero()
+            end_index = end_indexes[-1].item() if len(end_indexes) else 0
+            scores.append(values[i, end_index].nan_to_num().item())
+
+        return scores
+
+    async def start(self) -> None:
+        self._semaphore = asyncio.Semaphore(int(os.environ.get("MAX_CONCURRENT", 1)))
+
+    async def chat(
+        self,
+        messages: Sequence[Dict[str, str]],
+        system: Optional[str] = None,
+        tools: Optional[str] = None,
+        **input_kwargs,
+    ) -> List["Response"]:
+        if not self.can_generate:
+            raise ValueError("The current model does not support `chat`.")
+
+        loop = asyncio.get_running_loop()
+        input_args = (
+            self.model,
+            self.tokenizer,
+            self.template,
+            self.generating_args,
+            messages,
+            system,
+            tools,
+            input_kwargs,
+        )
+        async with self._semaphore:
+            with concurrent.futures.ThreadPoolExecutor() as pool:
+                return await loop.run_in_executor(pool, self._chat, *input_args)
+
+    async def stream_chat(
+        self,
+        messages: Sequence[Dict[str, str]],
+        system: Optional[str] = None,
+        tools: Optional[str] = None,
+        **input_kwargs,
+    ) -> AsyncGenerator[str, None]:
+        if not self.can_generate:
+            raise ValueError("The current model does not support `stream_chat`.")
+
+        loop = asyncio.get_running_loop()
+        input_args = (
+            self.model,
+            self.tokenizer,
+            self.template,
+            self.generating_args,
+            messages,
+            system,
+            tools,
+            input_kwargs,
+        )
+        async with self._semaphore:
+            with concurrent.futures.ThreadPoolExecutor() as pool:
+                stream = self._stream_chat(*input_args)
+                while True:
+                    try:
+                        yield await loop.run_in_executor(pool, stream)
+                    except StopAsyncIteration:
+                        break
+
+    async def get_scores(
+        self,
+        batch_input: List[str],
+        **input_kwargs,
+    ) -> List[float]:
+        if self.can_generate:
+            raise ValueError("Cannot get scores using an auto-regressive model.")
+
+        loop = asyncio.get_running_loop()
+        input_args = (self.model, self.tokenizer, batch_input, input_kwargs)
+        async with self._semaphore:
+            with concurrent.futures.ThreadPoolExecutor() as pool:
+                return await loop.run_in_executor(pool, self._get_scores, *input_args)

src/llmtuner/chat/vllm_engine.py

@@ -0,0 +1,149 @@
+import uuid
+from typing import TYPE_CHECKING, AsyncGenerator, AsyncIterator, Dict, List, Optional, Sequence
+
+from transformers.utils.versions import require_version
+
+from ..data import get_template_and_fix_tokenizer
+from ..extras.misc import get_device_count
+from ..extras.packages import is_vllm_available
+from ..model import load_tokenizer
+from .base_engine import BaseEngine, Response
+
+
+if is_vllm_available():
+    from vllm import AsyncEngineArgs, AsyncLLMEngine, RequestOutput, SamplingParams
+
+if TYPE_CHECKING:
+    from ..hparams import DataArguments, FinetuningArguments, GeneratingArguments, ModelArguments
+
+
+class VllmEngine(BaseEngine):
+    def __init__(
+        self,
+        model_args: "ModelArguments",
+        data_args: "DataArguments",
+        finetuning_args: "FinetuningArguments",
+        generating_args: "GeneratingArguments",
+    ) -> None:
+        require_version("vllm>=0.3.3", "To fix: pip install vllm>=0.3.3")
+        self.can_generate = finetuning_args.stage == "sft"
+        engine_args = AsyncEngineArgs(
+            model=model_args.model_name_or_path,
+            trust_remote_code=True,
+            max_model_len=model_args.vllm_maxlen,
+            tensor_parallel_size=get_device_count() or 1,
+            gpu_memory_utilization=model_args.vllm_gpu_util,
+            disable_log_stats=True,
+            disable_log_requests=True,
+            enforce_eager=model_args.vllm_enforce_eager,
+        )
+        self.model = AsyncLLMEngine.from_engine_args(engine_args)
+        self.tokenizer = load_tokenizer(model_args)
+        self.tokenizer.padding_side = "left"
+        self.template = get_template_and_fix_tokenizer(self.tokenizer, data_args.template)
+        self.generating_args = generating_args.to_dict()
+
+    async def _generate(
+        self,
+        messages: Sequence[Dict[str, str]],
+        system: Optional[str] = None,
+        tools: Optional[str] = None,
+        **input_kwargs,
+    ) -> AsyncIterator["RequestOutput"]:
+        request_id = "chatcmpl-{}".format(uuid.uuid4().hex)
+        paired_messages = messages + [{"role": "assistant", "content": ""}]
+        prompt_ids, _ = self.template.encode_oneturn(
+            tokenizer=self.tokenizer, messages=paired_messages, system=system, tools=tools
+        )
+        prompt_length = len(prompt_ids)
+
+        temperature = input_kwargs.pop("temperature", None)
+        top_p = input_kwargs.pop("top_p", None)
+        top_k = input_kwargs.pop("top_k", None)
+        num_return_sequences = input_kwargs.pop("num_return_sequences", None)
+        repetition_penalty = input_kwargs.pop("repetition_penalty", None)
+        max_length = input_kwargs.pop("max_length", None)
+        max_new_tokens = input_kwargs.pop("max_new_tokens", None)
+
+        generating_args = self.generating_args.copy()
+        generating_args.update(
+            dict(
+                temperature=temperature or generating_args["temperature"],
+                top_p=top_p or generating_args["top_p"],
+                top_k=top_k or generating_args["top_k"],
+                num_return_sequences=num_return_sequences or 1,
+                repetition_penalty=repetition_penalty or generating_args["repetition_penalty"],
+            )
+        )
+
+        if max_length:
+            generating_args["max_new_tokens"] = max_length - prompt_length
+
+        if max_new_tokens:
+            generating_args["max_new_tokens"] = max_new_tokens
+
+        sampling_params = SamplingParams(
+            n=generating_args["num_return_sequences"],
+            repetition_penalty=generating_args["repetition_penalty"],
+            temperature=generating_args["temperature"],
+            top_p=generating_args["top_p"],
+            top_k=generating_args["top_k"],
+            use_beam_search=generating_args["num_beams"] > 1,
+            length_penalty=generating_args["length_penalty"],
+            stop_token_ids=[self.tokenizer.eos_token_id] + self.tokenizer.additional_special_tokens_ids,
+            max_tokens=generating_args["max_new_tokens"],
+            skip_special_tokens=True,
+        )
+        result_generator = self.model.generate(
+            prompt=None, sampling_params=sampling_params, request_id=request_id, prompt_token_ids=prompt_ids
+        )
+        return result_generator
+
+    async def start(self) -> None:
+        pass
+
+    async def chat(
+        self,
+        messages: Sequence[Dict[str, str]],
+        system: Optional[str] = None,
+        tools: Optional[str] = None,
+        **input_kwargs,
+    ) -> List["Response"]:
+        final_output = None
+        generator = await self._generate(messages, system, tools, **input_kwargs)
+        async for request_output in generator:
+            final_output = request_output
+
+        results = []
+        for output in final_output.outputs:
+            results.append(
+                Response(
+                    response_text=output.text,
+                    response_length=len(output.token_ids),
+                    prompt_length=len(final_output.prompt_token_ids),
+                    finish_reason=output.finish_reason,
+                )
+            )
+
+        return results
+
+    async def stream_chat(
+        self,
+        messages: Sequence[Dict[str, str]],
+        system: Optional[str] = None,
+        tools: Optional[str] = None,
+        **input_kwargs,
+    ) -> AsyncGenerator[str, None]:
+        generated_text = ""
+        generator = await self._generate(messages, system, tools, **input_kwargs)
+        async for result in generator:
+            delta_text = result.outputs[0].text[len(generated_text) :]
+            generated_text = result.outputs[0].text
+            yield delta_text
+
+    async def get_scores(
+        self,
+        batch_input: List[str],
+        **input_kwargs,
+    ) -> List[float]:
+        raise NotImplementedError("vLLM engine does not support get_scores.")

src/llmtuner/data/__init__.py

@@ -1,6 +1,6 @@
 from .loader import get_dataset
-from .template import get_template_and_fix_tokenizer, templates
+from .template import Template, get_template_and_fix_tokenizer, templates
 from .utils import Role, split_dataset
 
 
-__all__ = ["get_dataset", "get_template_and_fix_tokenizer", "templates", "Role", "split_dataset"]
+__all__ = ["get_dataset", "Template", "get_template_and_fix_tokenizer", "templates", "Role", "split_dataset"]

src/llmtuner/data/aligner.py

@@ -1,6 +1,8 @@
 from functools import partial
 from typing import TYPE_CHECKING, Any, Dict, List, Union
 
+from datasets import Features
+
 from .utils import Role
 
 
@@ -15,23 +17,26 @@ def convert_alpaca(examples: Dict[str, List[Any]], dataset_attr: "DatasetAttr")
     outputs = {"prompt": [], "response": [], "system": [], "tools": []}
     for i in range(len(examples[dataset_attr.prompt])):
         prompt = []
-        if dataset_attr.history:
+        if dataset_attr.history and isinstance(examples[dataset_attr.history][i], list):
             for old_prompt, old_response in examples[dataset_attr.history][i]:
-                prompt.append({"role": Role.USER, "content": old_prompt})
-                prompt.append({"role": Role.ASSISTANT, "content": old_response})
+                prompt.append({"role": Role.USER.value, "content": old_prompt})
+                prompt.append({"role": Role.ASSISTANT.value, "content": old_response})
+
+        content = []
+        if dataset_attr.prompt and examples[dataset_attr.prompt][i]:
+            content.append(examples[dataset_attr.prompt][i])
 
-        instruction = examples[dataset_attr.prompt][i]
         if dataset_attr.query and examples[dataset_attr.query][i]:
-            instruction += "\n" + examples[dataset_attr.query][i]
-        prompt.append({"role": Role.USER, "content": instruction})
+            content.append(examples[dataset_attr.query][i])
 
-        if dataset_attr.response:
-            if isinstance(examples[dataset_attr.response][i], list):
-                response = [
-                    {"role": Role.ASSISTANT, "content": content} for content in examples[dataset_attr.response][i]
-                ]
-            else:
-                response = [{"role": Role.ASSISTANT, "content": examples[dataset_attr.response][i]}]
+        prompt.append({"role": Role.USER.value, "content": "\n".join(content)})
+
+        if dataset_attr.response and isinstance(examples[dataset_attr.response][i], list):
+            response = [
+                {"role": Role.ASSISTANT.value, "content": content} for content in examples[dataset_attr.response][i]
+            ]
+        elif dataset_attr.response and isinstance(examples[dataset_attr.response][i], str):
+            response = [{"role": Role.ASSISTANT.value, "content": examples[dataset_attr.response][i]}]
         else:
             response = []
 
@@ -46,36 +51,38 @@ def convert_alpaca(examples: Dict[str, List[Any]], dataset_attr: "DatasetAttr")
 def convert_sharegpt(examples: Dict[str, List[Any]], dataset_attr: "DatasetAttr") -> Dict[str, List[Any]]:
     outputs = {"prompt": [], "response": [], "system": [], "tools": []}
     tag_mapping = {
-        dataset_attr.user_tag: Role.USER,
-        dataset_attr.assistant_tag: Role.ASSISTANT,
-        dataset_attr.observation_tag: Role.OBSERVATION,
-        dataset_attr.function_tag: Role.FUNCTION,
+        dataset_attr.user_tag: Role.USER.value,
+        dataset_attr.assistant_tag: Role.ASSISTANT.value,
+        dataset_attr.observation_tag: Role.OBSERVATION.value,
+        dataset_attr.function_tag: Role.FUNCTION.value,
+        dataset_attr.system_tag: Role.SYSTEM.value,
     }
+    odd_tags = (dataset_attr.user_tag, dataset_attr.observation_tag)
+    even_tags = (dataset_attr.assistant_tag, dataset_attr.function_tag)
+    accept_tags = (odd_tags, even_tags)
     for i, messages in enumerate(examples[dataset_attr.messages]):
+        if dataset_attr.system_tag and messages[0][dataset_attr.role_tag] == dataset_attr.system_tag:
+            system = messages[0][dataset_attr.content_tag]
+            messages = messages[1:]
+        else:
+            system = examples[dataset_attr.system][i] if dataset_attr.system else ""
+
         messages = messages[: len(messages) // 2 * 2]  # should be multiples of 2
         if len(messages) == 0:
             continue
 
-        prompt = []
-        response = []
+        aligned_messages = []
         for turn_idx, message in enumerate(messages):
-            if turn_idx % 2 == 0:
-                accept_tags = [dataset_attr.user_tag, dataset_attr.observation_tag]
-            else:
-                accept_tags = [dataset_attr.assistant_tag, dataset_attr.function_tag]
-
-            if message[dataset_attr.role_tag] not in accept_tags:
+            if message[dataset_attr.role_tag] not in accept_tags[turn_idx % 2]:
                 raise ValueError("Invalid role tag in {}.".format(messages))
 
-            prompt.append(
+            aligned_messages.append(
                 {"role": tag_mapping[message[dataset_attr.role_tag]], "content": message[dataset_attr.content_tag]}
             )
 
-        last_message = prompt.pop(-1)
-        response.append(last_message)
-        outputs["prompt"].append(prompt)
-        outputs["response"].append(response)
-        outputs["system"].append(examples[dataset_attr.system][i] if dataset_attr.system else "")
+        outputs["prompt"].append(aligned_messages[:-1])
+        outputs["response"].append(aligned_messages[-1:])
+        outputs["system"].append(system)
         outputs["tools"].append(examples[dataset_attr.tools][i] if dataset_attr.tools else "")
 
     return outputs
@@ -86,8 +93,8 @@ def align_dataset(
 ) -> Union["Dataset", "IterableDataset"]:
     r"""
     Aligned dataset:
-        prompt: [{"role": "user", "content": "..."}]
-        response: [{"role": "assistant", "content": "..."}]
+        prompt: [{"role": "user", "content": "..."}] * (2T - 1)
+        response: [{"role": "assistant", "content": "..."}] * N (N > 1 for ranking dataset)
         system: "..."
         tools: "..."
     """
@@ -97,6 +104,18 @@ def align_dataset(
         convert_func = partial(convert_sharegpt, dataset_attr=dataset_attr)
 
     column_names = list(next(iter(dataset)).keys())
+    features = Features.from_dict(
+        {
+            "prompt": [
+                {"role": {"dtype": "string", "_type": "Value"}, "content": {"dtype": "string", "_type": "Value"}}
+            ],
+            "response": [
+                {"role": {"dtype": "string", "_type": "Value"}, "content": {"dtype": "string", "_type": "Value"}}
+            ],
+            "system": {"dtype": "string", "_type": "Value"},
+            "tools": {"dtype": "string", "_type": "Value"},
+        }
+    )
     kwargs = {}
     if not data_args.streaming:
         kwargs = dict(
@@ -105,4 +124,10 @@ def align_dataset(
             desc="Converting format of dataset",
         )
 
-    return dataset.map(convert_func, batched=True, remove_columns=column_names, **kwargs)
+    return dataset.map(
+        convert_func,
+        batched=True,
+        remove_columns=column_names,
+        features=features,
+        **kwargs,
+    )

src/llmtuner/data/formatter.py

@@ -2,7 +2,7 @@ import json
 import re
 from abc import ABC, abstractmethod
 from dataclasses import dataclass, field
-from typing import Any, Dict, List, Literal, Sequence, Set, Tuple, Union
+from typing import Any, Dict, List, Literal, Optional, Sequence, Set, Tuple, Union
 
 
 SLOTS = Sequence[Union[str, Set[str], Dict[str, str]]]
@@ -15,11 +15,11 @@ JSON_FORMAT_PROMPT = (
 
 TOOL_SYSTEM_PROMPT = (
     "You have access to the following tools:\n{tool_text}"
-    "Use the following format to answer the question:\n"
+    "Use the following format if using a tool:\n"
+    "```\n"
+    "Action: tool name (one of [{tool_names}]).\n"
+    "Action Input: the input to the tool{format_prompt}.\n"
     "```\n"
-    "Action: the action to take, should be one of [{tool_names}] if using a tool.\n"
-    "Action Input: the input to the action{format_prompt}.\n"
-    "```"
 )
 
 
@@ -31,12 +31,16 @@ def default_tool_formatter(tools: List[Dict[str, Any]]) -> str:
         for name, param in tool["parameters"]["properties"].items():
             required = ", required" if name in tool["parameters"].get("required", []) else ""
             enum = ", should be one of [{}]".format(", ".join(param["enum"])) if param.get("enum", None) else ""
-            param_text += "  - {name} ({type}{required}): {desc}{enum}\n".format(
+            items = (
+                ", where each item should be {}".format(param["items"].get("type", "")) if param.get("items") else ""
+            )
+            param_text += "  - {name} ({type}{required}): {desc}{enum}{items}\n".format(
                 name=name,
                 type=param.get("type", ""),
                 required=required,
                 desc=param.get("description", ""),
                 enum=enum,
+                items=items,
             )
 
         tool_text += "> Tool Name: {name}\nTool Description: {desc}\nTool Args:\n{args}\n".format(
@@ -68,11 +72,10 @@ def default_tool_extractor(content: str) -> Union[str, Tuple[str, str]]:
 @dataclass
 class Formatter(ABC):
     slots: SLOTS = field(default_factory=list)
-    tool_format: Literal["default"] = "default"
+    tool_format: Optional[Literal["default"]] = None
 
     @abstractmethod
-    def apply(self, **kwargs) -> SLOTS:
-        ...
+    def apply(self, **kwargs) -> SLOTS: ...
 
     def extract(self, content: str) -> Union[str, Tuple[str, str]]:
         raise NotImplementedError
@@ -80,29 +83,61 @@ class Formatter(ABC):
 
 @dataclass
 class EmptyFormatter(Formatter):
+    def __post_init__(self):
+        has_placeholder = False
+        for slot in filter(lambda s: isinstance(s, str), self.slots):
+            if re.search(r"\{\{[a-zA-Z_][a-zA-Z0-9_]*\}\}", slot):
+                has_placeholder = True
+
+        if has_placeholder:
+            raise ValueError("Empty formatter should not contain any placeholder.")
+
     def apply(self, **kwargs) -> SLOTS:
         return self.slots
 
 
 @dataclass
 class StringFormatter(Formatter):
+    def __post_init__(self):
+        has_placeholder = False
+        for slot in filter(lambda s: isinstance(s, str), self.slots):
+            if re.search(r"\{\{[a-zA-Z_][a-zA-Z0-9_]*\}\}", slot):
+                has_placeholder = True
+
+        if not has_placeholder:
+            raise ValueError("A placeholder is required in the string formatter.")
+
     def apply(self, **kwargs) -> SLOTS:
         elements = []
         for slot in self.slots:
             if isinstance(slot, str):
                 for name, value in kwargs.items():
+                    if not isinstance(value, str):
+                        raise RuntimeError("Expected a string, got {}".format(value))
+
                     slot = slot.replace("{{" + name + "}}", value, 1)
                 elements.append(slot)
             elif isinstance(slot, (dict, set)):
                 elements.append(slot)
             else:
-                raise ValueError("Input must be string, set[str] or dict[str, str], got {}".format(type(slot)))
+                raise RuntimeError("Input must be string, set[str] or dict[str, str], got {}".format(type(slot)))
 
         return elements
 
 
 @dataclass
 class FunctionFormatter(Formatter):
+    def __post_init__(self):
+        has_name, has_args = False, False
+        for slot in filter(lambda s: isinstance(s, str), self.slots):
+            if "{{name}}" in slot:
+                has_name = True
+            if "{{arguments}}" in slot:
+                has_args = True
+
+        if not has_name or not has_args:
+            raise ValueError("Name and arguments placeholders are required in the function formatter.")
+
     def apply(self, **kwargs) -> SLOTS:
         content = kwargs.pop("content")
         try:
@@ -120,13 +155,17 @@ class FunctionFormatter(Formatter):
             elif isinstance(slot, (dict, set)):
                 elements.append(slot)
             else:
-                raise ValueError("Input must be string, set[str] or dict[str, str], got {}".format(type(slot)))
+                raise RuntimeError("Input must be string, set[str] or dict[str, str], got {}".format(type(slot)))
 
         return elements
 
 
 @dataclass
 class ToolFormatter(Formatter):
+    def __post_init__(self):
+        if self.tool_format is None:
+            raise ValueError("Tool format was not found.")
+
     def apply(self, **kwargs) -> SLOTS:
         content = kwargs.pop("content")
         try:

src/llmtuner/data/loader.py

@@ -1,8 +1,8 @@
 import inspect
 import os
-from typing import TYPE_CHECKING, List, Literal, Union
+from typing import TYPE_CHECKING, Literal, Union
 
-from datasets import concatenate_datasets, interleave_datasets, load_dataset, load_from_disk
+from datasets import load_dataset, load_from_disk
 
 from ..extras.constants import FILEEXT2TYPE
 from ..extras.logging import get_logger
@@ -10,7 +10,7 @@ from .aligner import align_dataset
 from .parser import get_dataset_list
 from .preprocess import get_preprocess_and_print_func
 from .template import get_template_and_fix_tokenizer
-from .utils import checksum
+from .utils import checksum, merge_dataset
 
 
 if TYPE_CHECKING:
@@ -29,7 +29,8 @@ def load_single_dataset(
     dataset_attr: "DatasetAttr",
     model_args: "ModelArguments",
     data_args: "DataArguments",
-):
+) -> Union["Dataset", "IterableDataset"]:
+    logger.info("Loading dataset {}...".format(dataset_attr))
     data_path, data_name, data_dir, data_files = None, None, None, None
     if dataset_attr.load_from in ["hf_hub", "ms_hub"]:
         data_path = dataset_attr.dataset_name
@@ -43,7 +44,7 @@ def load_single_dataset(
 
     elif dataset_attr.load_from == "file":
         data_files = []
-        local_path: str = os.path.join(data_args.dataset_dir, dataset_attr.dataset_name)
+        local_path = os.path.join(data_args.dataset_dir, dataset_attr.dataset_name)
         if os.path.isdir(local_path):  # is directory
             for file_name in os.listdir(local_path):
                 data_files.append(os.path.join(local_path, file_name))
@@ -60,7 +61,7 @@ def load_single_dataset(
         if data_path is None:
             raise ValueError("File extension must be txt, csv, json or jsonl.")
 
-        checksum(data_files, dataset_attr.dataset_sha1)
+        checksum(data_files, dataset_attr.file_sha1)
     else:
         raise NotImplementedError
 
@@ -110,30 +111,6 @@ def load_single_dataset(
     return align_dataset(dataset, dataset_attr, data_args)
 
 
-def merge_dataset(
-    all_datasets: List[Union["Dataset", "IterableDataset"]],
-    data_args: "DataArguments",
-    training_args: "Seq2SeqTrainingArguments",
-) -> Union["Dataset", "IterableDataset"]:
-    if len(all_datasets) == 1:
-        return all_datasets[0]
-    elif data_args.mix_strategy == "concat":
-        if data_args.streaming:
-            logger.warning("The samples between different datasets will not be mixed in streaming mode.")
-        return concatenate_datasets(all_datasets)
-    elif data_args.mix_strategy.startswith("interleave"):
-        if not data_args.streaming:
-            logger.warning("We recommend using `mix_strategy=concat` in non-streaming mode.")
-        return interleave_datasets(
-            datasets=all_datasets,
-            probabilities=data_args.interleave_probs,
-            seed=training_args.seed,
-            stopping_strategy="first_exhausted" if data_args.mix_strategy.endswith("under") else "all_exhausted",
-        )
-    else:
-        raise ValueError("Unknown mixing strategy.")
-
-
 def get_dataset(
     tokenizer: "PreTrainedTokenizer",
     model_args: "ModelArguments",
@@ -142,7 +119,7 @@ def get_dataset(
     stage: Literal["pt", "sft", "rm", "ppo"],
     # split: Optional[str] = "train", # TODO: add split
 ) -> Union["Dataset", "IterableDataset"]:
-    template = get_template_and_fix_tokenizer(data_args.template, tokenizer)
+    template = get_template_and_fix_tokenizer(tokenizer, data_args.template)
     if data_args.train_on_prompt and template.efficient_eos:
         raise ValueError("Current template does not support `train_on_prompt`.")
 
@@ -156,11 +133,11 @@ def get_dataset(
             return dataset
 
         if data_args.streaming:
-            raise ValueError("Turn off dataset streaming to save cache files.")
+            raise ValueError("Turn off `streaming` when saving dataset to disk.")
 
     with training_args.main_process_first(desc="load dataset"):
         all_datasets = []
-        for dataset_attr in get_dataset_list(data_args):  # TODO: add split
+        for dataset_attr in get_dataset_list(data_args):
             all_datasets.append(load_single_dataset(dataset_attr, model_args, data_args))
         dataset = merge_dataset(all_datasets, data_args, training_args)
 
@@ -188,6 +165,6 @@ def get_dataset(
             try:
                 print_function(next(iter(dataset)))
             except StopIteration:
-                raise RuntimeError("Empty dataset!")
+                raise RuntimeError("Cannot find valid samples, check `data/README.md` for the data format.")
 
         return dataset

src/llmtuner/data/parser.py

@@ -1,7 +1,7 @@
 import json
 import os
 from dataclasses import dataclass
-from typing import TYPE_CHECKING, List, Literal, Optional
+from typing import TYPE_CHECKING, Any, Dict, List, Literal, Optional
 
 from ..extras.constants import DATA_CONFIG
 from ..extras.misc import use_modelscope
@@ -13,34 +13,44 @@ if TYPE_CHECKING:
 
 @dataclass
 class DatasetAttr:
+    r"""
+    Dataset attributes.
+    """
+
+    """ basic configs """
     load_from: Literal["hf_hub", "ms_hub", "script", "file"]
-    dataset_name: Optional[str] = None
-    dataset_sha1: Optional[str] = None
+    dataset_name: str
+    """ extra configs """
+    file_sha1: Optional[str] = None
     subset: Optional[str] = None
     folder: Optional[str] = None
-    ranking: Optional[bool] = False
-    formatting: Optional[Literal["alpaca", "sharegpt"]] = "alpaca"
-
+    ranking: bool = False
+    formatting: Literal["alpaca", "sharegpt"] = "alpaca"
+    """ columns """
     system: Optional[str] = None
-
+    """ columns for the alpaca format """
     prompt: Optional[str] = "instruction"
     query: Optional[str] = "input"
     response: Optional[str] = "output"
     history: Optional[str] = None
-
+    """ columns for the sharegpt format """
     messages: Optional[str] = "conversations"
     tools: Optional[str] = None
-
+    """ tags for the sharegpt format """
     role_tag: Optional[str] = "from"
     content_tag: Optional[str] = "value"
     user_tag: Optional[str] = "human"
     assistant_tag: Optional[str] = "gpt"
     observation_tag: Optional[str] = "observation"
     function_tag: Optional[str] = "function_call"
+    system_tag: Optional[str] = "system"
 
     def __repr__(self) -> str:
         return self.dataset_name
 
+    def set_attr(self, key: str, obj: Dict[str, Any], default: Optional[Any] = None) -> None:
+        setattr(self, key, obj.get(key, default))
+
 
 def get_dataset_list(data_args: "DataArguments") -> List["DatasetAttr"]:
     dataset_names = [ds.strip() for ds in data_args.dataset.split(",")] if data_args.dataset is not None else []
@@ -73,30 +83,36 @@ def get_dataset_list(data_args: "DataArguments") -> List["DatasetAttr"]:
         elif "script_url" in dataset_info[name]:
             dataset_attr = DatasetAttr("script", dataset_name=dataset_info[name]["script_url"])
         else:
-            dataset_attr = DatasetAttr(
-                "file",
-                dataset_name=dataset_info[name]["file_name"],
-                dataset_sha1=dataset_info[name].get("file_sha1", None),
-            )
+            dataset_attr = DatasetAttr("file", dataset_name=dataset_info[name]["file_name"])
 
-        dataset_attr.subset = dataset_info[name].get("subset", None)
-        dataset_attr.folder = dataset_info[name].get("folder", None)
-        dataset_attr.ranking = dataset_info[name].get("ranking", False)
-        dataset_attr.formatting = dataset_info[name].get("formatting", "alpaca")
+        dataset_attr.set_attr("file_sha1", dataset_info[name])
+        dataset_attr.set_attr("subset", dataset_info[name])
+        dataset_attr.set_attr("folder", dataset_info[name])
+        dataset_attr.set_attr("ranking", dataset_info[name], default=False)
+        dataset_attr.set_attr("formatting", dataset_info[name], default="alpaca")
 
         if "columns" in dataset_info[name]:
+            column_names = ["system"]
             if dataset_attr.formatting == "alpaca":
-                column_names = ["prompt", "query", "response", "history"]
+                column_names.extend(["prompt", "query", "response", "history"])
             else:
-                column_names = ["messages", "tools"]
+                column_names.extend(["messages", "tools"])
 
-            column_names += ["system"]
             for column_name in column_names:
-                setattr(dataset_attr, column_name, dataset_info[name]["columns"].get(column_name, None))
+                dataset_attr.set_attr(column_name, dataset_info[name]["columns"])
 
         if dataset_attr.formatting == "sharegpt" and "tags" in dataset_info[name]:
-            for tag in ["role_tag", "content_tag", "user_tag", "assistant_tag", "observation_tag", "function_tag"]:
-                setattr(dataset_attr, tag, dataset_info[name]["tags"].get(tag, None))
+            tag_names = (
+                "role_tag",
+                "content_tag",
+                "user_tag",
+                "assistant_tag",
+                "observation_tag",
+                "function_tag",
+                "system_tag",
+            )
+            for tag in tag_names:
+                dataset_attr.set_attr(tag, dataset_info[name]["tags"])
 
         dataset_list.append(dataset_attr)
 

src/llmtuner/data/preprocess.py

@@ -4,6 +4,7 @@ from typing import TYPE_CHECKING, Any, Callable, Dict, List, Literal, Tuple
 
 from ..extras.constants import IGNORE_INDEX
 from ..extras.logging import get_logger
+from .utils import Role
 
 
 if TYPE_CHECKING:
@@ -20,13 +21,12 @@ logger = get_logger(__name__)
 def preprocess_pretrain_dataset(
     examples: Dict[str, List[Any]], tokenizer: "PreTrainedTokenizer", data_args: "DataArguments"
 ) -> Dict[str, List[List[int]]]:
-    # build grouped texts with format `X1 X2 X3 ...`
-    text_examples = [examples["prompt"][i][0]["content"] for i in range(len(examples["prompt"]))]
-    tokenized_examples = tokenizer(text_examples, add_special_tokens=False)
-    for i in range(len(tokenized_examples["input_ids"])):
-        tokenized_examples["input_ids"][i] += [tokenizer.eos_token_id]
-        tokenized_examples["attention_mask"][i] += [1]
+    # build grouped texts with format `X1 X2 X3 ...` if packing is enabled
+    text_examples = [messages[0]["content"] + tokenizer.eos_token for messages in examples["prompt"]]
+    if not data_args.packing:
+        return tokenizer(text_examples, add_special_tokens=False, max_length=data_args.cutoff_len)
 
+    tokenized_examples = tokenizer(text_examples, add_special_tokens=False)
     concatenated_examples = {k: list(chain(*tokenized_examples[k])) for k in tokenized_examples.keys()}
     total_length = len(concatenated_examples[list(concatenated_examples.keys())[0]])
     block_size = data_args.cutoff_len
@@ -37,6 +37,10 @@ def preprocess_pretrain_dataset(
         k: [t[i : i + block_size] for i in range(0, total_length, block_size)]
         for k, t in concatenated_examples.items()
     }
+    if data_args.template == "gemma":
+        for i in range(len(result["input_ids"])):
+            result["input_ids"][i][0] = tokenizer.bos_token_id
+
     return result
 
 
@@ -51,14 +55,19 @@ def preprocess_supervised_dataset(
     model_inputs = {"input_ids": [], "attention_mask": [], "labels": []}
 
     for i in range(len(examples["prompt"])):
-        if len(examples["prompt"][i]) == 0 or len(examples["response"][i]) != 1:
+        if len(examples["prompt"][i]) % 2 != 1 or len(examples["response"][i]) != 1:
             continue
 
         messages = examples["prompt"][i] + examples["response"][i]
         input_ids, labels = [], []
         for turn_idx, (source_ids, target_ids) in enumerate(
             template.encode_multiturn(
-                tokenizer, messages, examples["system"][i], examples["tools"][i], data_args.cutoff_len
+                tokenizer,
+                messages,
+                examples["system"][i],
+                examples["tools"][i],
+                data_args.cutoff_len,
+                data_args.reserved_label_len,
             )
         ):
             if data_args.train_on_prompt:
@@ -93,16 +102,16 @@ def preprocess_packed_supervised_dataset(
     model_inputs = {"input_ids": [], "attention_mask": [], "labels": []}
     input_ids, labels = [], []
     for i in range(len(examples["prompt"])):
-        if len(examples["prompt"][i]) == 0 or len(examples["response"][i]) != 1:
+        if len(examples["prompt"][i]) % 2 != 1 or len(examples["response"][i]) != 1:
             continue
 
         messages = examples["prompt"][i] + examples["response"][i]
-        for turn_idx, (source_ids, target_ids) in enumerate(
-            template.encode_multiturn(tokenizer, messages, examples["system"][i], examples["tools"][i])
+        for source_ids, target_ids in template.encode_multiturn(
+            tokenizer, messages, examples["system"][i], examples["tools"][i]
         ):
             if data_args.train_on_prompt:
                 source_mask = source_ids
-            elif turn_idx != 0 and template.efficient_eos:
+            elif len(input_ids) != 0 and template.efficient_eos:
                 source_mask = [tokenizer.eos_token_id] + [IGNORE_INDEX] * (len(source_ids) - 1)
             else:
                 source_mask = [IGNORE_INDEX] * len(source_ids)
@@ -120,9 +129,10 @@ def preprocess_packed_supervised_dataset(
     total_length = (total_length // block_size) * block_size
     # split by chunks of cutoff_len
     for i in range(0, total_length, block_size):
-        model_inputs["input_ids"].append(input_ids[i : i + block_size])
-        model_inputs["attention_mask"].append([1] * block_size)
-        model_inputs["labels"].append(labels[i : i + block_size])
+        if not all(label == IGNORE_INDEX for label in labels[i : i + block_size]):
+            model_inputs["input_ids"].append(input_ids[i : i + block_size])
+            model_inputs["attention_mask"].append([1] * block_size)
+            model_inputs["labels"].append(labels[i : i + block_size])
 
     return model_inputs
 
@@ -137,12 +147,21 @@ def preprocess_unsupervised_dataset(
     model_inputs = {"input_ids": [], "attention_mask": [], "labels": []}
 
     for i in range(len(examples["prompt"])):
-        if len(examples["prompt"][i]) == 0 or len(examples["response"][i]) != 1:
+        if len(examples["prompt"][i]) % 2 != 1:
             continue
 
-        messages = examples["prompt"][i] + examples["response"][i]
+        if len(examples["response"][i]) == 1:
+            messages = examples["prompt"][i] + examples["response"][i]
+        else:
+            messages = examples["prompt"][i] + [{"role": Role.ASSISTANT.value, "content": ""}]
+
         input_ids, labels = template.encode_oneturn(
-            tokenizer, messages, examples["system"][i], examples["tools"][i], data_args.cutoff_len
+            tokenizer,
+            messages,
+            examples["system"][i],
+            examples["tools"][i],
+            data_args.cutoff_len,
+            data_args.reserved_label_len,
         )
 
         if template.efficient_eos:
@@ -164,17 +183,26 @@ def preprocess_pairwise_dataset(
     # build input pairs with format `<bos> X`, `Y1 <eos>` and `Y2 <eos>`
     model_inputs = {"prompt_ids": [], "chosen_ids": [], "rejected_ids": []}
     for i in range(len(examples["prompt"])):
-        if len(examples["prompt"][i]) == 0 or len(examples["response"][i]) < 2:
+        if len(examples["prompt"][i]) % 2 != 1 or len(examples["response"][i]) < 2:
             continue
 
         chosen_messages = examples["prompt"][i] + [examples["response"][i][0]]
         rejected_messages = examples["prompt"][i] + [examples["response"][i][1]]
-
         prompt_ids, chosen_ids = template.encode_oneturn(
-            tokenizer, chosen_messages, examples["system"][i], examples["tools"][i], data_args.cutoff_len
+            tokenizer,
+            chosen_messages,
+            examples["system"][i],
+            examples["tools"][i],
+            data_args.cutoff_len,
+            data_args.reserved_label_len,
         )
         _, rejected_ids = template.encode_oneturn(
-            tokenizer, rejected_messages, examples["system"][i], examples["tools"][i], data_args.cutoff_len
+            tokenizer,
+            rejected_messages,
+            examples["system"][i],
+            examples["tools"][i],
+            data_args.cutoff_len,
+            data_args.reserved_label_len,
         )
 
         if template.efficient_eos:
@@ -224,7 +252,7 @@ def get_preprocess_and_print_func(
         preprocess_func = partial(preprocess_pretrain_dataset, tokenizer=tokenizer, data_args=data_args)
         print_function = partial(print_unsupervised_dataset_example, tokenizer=tokenizer)
     elif stage == "sft" and not training_args.predict_with_generate:
-        if data_args.sft_packing:
+        if data_args.packing:
             preprocess_func = partial(
                 preprocess_packed_supervised_dataset, tokenizer=tokenizer, template=template, data_args=data_args
             )

src/llmtuner/data/template.py

@@ -9,7 +9,7 @@ from .utils import Role, infer_max_len
 if TYPE_CHECKING:
     from transformers import PreTrainedTokenizer
 
-    from .formatter import Formatter
+    from .formatter import SLOTS, Formatter
 
 
 logger = get_logger(__name__)
@@ -36,8 +36,8 @@ class Template:
         messages: List[Dict[str, str]],
         system: Optional[str] = None,
         tools: Optional[str] = None,
-        cutoff_len: Optional[int] = 1_000_000,
-        reserved_label_len: Optional[int] = 16,
+        cutoff_len: int = 1_000_000,
+        reserved_label_len: int = 1,
     ) -> Tuple[List[int], List[int]]:
         r"""
         Returns a single pair of token ids representing prompt and response respectively.
@@ -56,8 +56,8 @@ class Template:
         messages: List[Dict[str, str]],
         system: Optional[str] = None,
         tools: Optional[str] = None,
-        cutoff_len: Optional[int] = 1_000_000,
-        reserved_label_len: Optional[int] = 16,
+        cutoff_len: int = 1_000_000,
+        reserved_label_len: int = 1,
     ) -> Sequence[Tuple[List[int], List[int]]]:
         r"""
         Returns multiple pairs of token ids representing prompts and responses respectively.
@@ -88,16 +88,16 @@ class Template:
             elif i > 0 and i % 2 == 0:
                 elements += self.format_separator.apply()
 
-            if message["role"] == Role.USER:
+            if message["role"] == Role.USER.value:
                 elements += self.format_user.apply(content=message["content"], idx=str(i // 2))
-            elif message["role"] == Role.ASSISTANT:
+            elif message["role"] == Role.ASSISTANT.value:
                 elements += self.format_assistant.apply(content=message["content"])
-            elif message["role"] == Role.OBSERVATION:
+            elif message["role"] == Role.OBSERVATION.value:
                 elements += self.format_observation.apply(content=message["content"])
-            elif message["role"] == Role.FUNCTION:
+            elif message["role"] == Role.FUNCTION.value:
                 elements += self.format_function.apply(content=message["content"])
             else:
-                raise NotImplementedError
+                raise NotImplementedError("Unexpected role: {}".format(message["role"]))
 
             encoded_messages.append(self._convert_elements_to_ids(tokenizer, elements))
 
@@ -117,9 +117,9 @@ class Template:
             elif isinstance(elem, dict):
                 token_ids += [tokenizer.convert_tokens_to_ids(elem.get("token"))]
             elif isinstance(elem, set):
-                if "bos_token" in elem and tokenizer.bos_token_id:
+                if "bos_token" in elem and tokenizer.bos_token_id is not None:
                     token_ids += [tokenizer.bos_token_id]
-                elif "eos_token" in elem and tokenizer.eos_token_id:
+                elif "eos_token" in elem and tokenizer.eos_token_id is not None:
                     token_ids += [tokenizer.eos_token_id]
             else:
                 raise ValueError("Input must be string, set[str] or dict[str, str], got {}".format(type(elem)))
@@ -144,10 +144,10 @@ class Template:
                 max_len=(cutoff_len - total_length),
                 reserved_label_len=reserved_label_len,
             )
-            encoded_messages[i] = encoded_messages[i][:max_source_len]
-            encoded_messages[i + 1] = encoded_messages[i + 1][:max_target_len]
-            total_length += len(encoded_messages[i]) + len(encoded_messages[i + 1])
-            encoded_pairs.append((encoded_messages[i], encoded_messages[i + 1]))
+            source_ids = encoded_messages[i][:max_source_len]
+            target_ids = encoded_messages[i + 1][:max_target_len]
+            total_length += len(source_ids) + len(target_ids)
+            encoded_pairs.append((source_ids, target_ids))
 
         return encoded_pairs
 
@@ -179,16 +179,16 @@ class Llama2Template(Template):
             elif i > 0 and i % 2 == 0:
                 elements += self.format_separator.apply()
 
-            if message["role"] == Role.USER:
+            if message["role"] == Role.USER.value:
                 elements += self.format_user.apply(content=system_text + message["content"])
-            elif message["role"] == Role.ASSISTANT:
+            elif message["role"] == Role.ASSISTANT.value:
                 elements += self.format_assistant.apply(content=message["content"])
-            elif message["role"] == Role.OBSERVATION:
+            elif message["role"] == Role.OBSERVATION.value:
                 elements += self.format_observation.apply(content=message["content"])
-            elif message["role"] == Role.FUNCTION:
+            elif message["role"] == Role.FUNCTION.value:
                 elements += self.format_function.apply(content=message["content"])
             else:
-                raise NotImplementedError
+                raise NotImplementedError("Unexpected role: {}".format(message["role"]))
 
             encoded_messages.append(self._convert_elements_to_ids(tokenizer, elements))
 
@@ -198,7 +198,7 @@ class Llama2Template(Template):
 templates: Dict[str, Template] = {}
 
 
-def register_template(
+def _register_template(
     name: str,
     format_user: Optional["Formatter"] = None,
     format_assistant: Optional["Formatter"] = None,
@@ -207,18 +207,44 @@ def register_template(
     format_observation: Optional["Formatter"] = None,
     format_tools: Optional["Formatter"] = None,
     format_separator: Optional["Formatter"] = None,
-    default_system: Optional[str] = "",
-    stop_words: Optional[List[str]] = [],
-    efficient_eos: Optional[bool] = False,
-    replace_eos: Optional[bool] = False,
-    force_system: Optional[bool] = False,
+    default_system: str = "",
+    stop_words: List[str] = [],
+    efficient_eos: bool = False,
+    replace_eos: bool = False,
+    force_system: bool = False,
 ) -> None:
+    r"""
+    Registers a chat template.
+
+    To add the following chat template:
+    ```
+    [HUMAN]:
+    user prompt here
+    [AI]:
+    model response here
+
+    [HUMAN]:
+    user prompt here
+    [AI]:
+    model response here
+    ```
+
+    The corresponding code should be:
+    ```
+    _register_template(
+        name="custom",
+        format_user=StringFormatter(slots=["[HUMAN]:\n{{content}}\n[AI]:\n"]),
+        format_separator=EmptyFormatter(slots=["\n\n"]),
+        efficient_eos=True,
+    )
+    ```
+    """
     eos_slots = [] if efficient_eos else [{"eos_token"}]
     template_class = Llama2Template if name.startswith("llama2") else Template
     default_user_formatter = StringFormatter(slots=["{{content}}"])
     default_assistant_formatter = StringFormatter(slots=["{{content}}"] + eos_slots)
     default_function_formatter = FunctionFormatter(slots=["Action: {{name}}\nAction Input: {{arguments}}"] + eos_slots)
-    default_tool_formatter = ToolFormatter(slots="default")
+    default_tool_formatter = ToolFormatter(tool_format="default")
     default_separator_formatter = EmptyFormatter()
     templates[name] = template_class(
         format_user=format_user or default_user_formatter,
@@ -236,40 +262,125 @@ def register_template(
     )
 
 
-def get_template_and_fix_tokenizer(name: str, tokenizer: "PreTrainedTokenizer") -> Template:
-    if tokenizer.eos_token_id is None:
-        tokenizer.eos_token = "<|endoftext|>"
+def _add_or_replace_eos_token(tokenizer: "PreTrainedTokenizer", eos_token: str) -> None:
+    is_added = tokenizer.eos_token_id is None
+    num_added_tokens = tokenizer.add_special_tokens({"eos_token": eos_token})
+
+    if is_added:
         logger.info("Add eos token: {}".format(tokenizer.eos_token))
+    else:
+        logger.info("Replace eos token: {}".format(tokenizer.eos_token))
 
-    if tokenizer.pad_token_id is None:
-        tokenizer.pad_token = tokenizer.eos_token
-        logger.info("Add pad token: {}".format(tokenizer.pad_token))
+    if num_added_tokens > 0:
+        logger.warning("New tokens have been added, make sure `resize_vocab` is True.")
 
-    if name is None:  # for pre-training
-        return None
 
-    template = templates.get(name, None)
-    assert template is not None, "Template {} does not exist.".format(name)
+def _jinja_escape(content: str) -> str:
+    return content.replace("\n", r"\n").replace("'", r"\'")
+
+
+def _convert_slots_to_jinja(slots: "SLOTS", tokenizer: "PreTrainedTokenizer", placeholder: str = "content") -> str:
+    slot_items = []
+    for slot in slots:
+        if isinstance(slot, str):
+            slot_pieces = slot.split("{{content}}")
+            if slot_pieces[0]:
+                slot_items.append("'" + _jinja_escape(slot_pieces[0]) + "'")
+            if len(slot_pieces) > 1:
+                slot_items.append(placeholder)
+                if slot_pieces[1]:
+                    slot_items.append("'" + _jinja_escape(slot_pieces[1]) + "'")
+        elif isinstance(slot, set):
+            if "bos_token" in slot:
+                slot_items.append("'" + tokenizer.bos_token + "'")
+            elif "eos_token" in slot:  # do not use {{ eos_token }} since it may be replaced
+                slot_items.append("'" + tokenizer.eos_token + "'")
+        elif isinstance(slot, dict):
+            raise ValueError("Dict is not supported.")
+
+    return " + ".join(slot_items)
+
+
+def _get_jinja_template(template: "Template", tokenizer: "PreTrainedTokenizer") -> str:
+    jinja_template = ""
+
+    if template.default_system:
+        jinja_template += "{% set system_message = '" + _jinja_escape(template.default_system) + "' %}"
+
+    jinja_template += (
+        "{% if messages[0]['role'] == 'system' %}" "{% set system_message = messages[0]['content'] %}" "{% endif %}"
+    )
+
+    system_message = _convert_slots_to_jinja(template.format_system.apply(), tokenizer, placeholder="system_message")
+    if isinstance(template, Llama2Template):
+        pass
+    elif template.force_system:
+        jinja_template += "{{ " + system_message + " }}"
+    else:
+        jinja_template += "{% if system_message is defined %}{{ " + system_message + " }}{% endif %}"
+
+    jinja_template += "{% for message in messages %}"
+    jinja_template += "{% set content = message['content'] %}"
+    if isinstance(template, Llama2Template):
+        jinja_template += "{% if loop.index0 == 0 and system_message is defined %}"
+        jinja_template += "{% set content = " + system_message + " + message['content'] %}"
+        jinja_template += "{% endif %}"
+    jinja_template += "{% if message['role'] == 'user' %}"
+    user_message = _convert_slots_to_jinja(template.format_user.apply(), tokenizer)
+    jinja_template += "{{ " + user_message + " }}"
+    jinja_template += "{% elif message['role'] == 'assistant' %}"
+    assistant_message = _convert_slots_to_jinja(
+        template.format_assistant.apply() + template.format_separator.apply(), tokenizer
+    )
+    jinja_template += "{{ " + assistant_message + " }}"
+    jinja_template += "{% endif %}"
+    jinja_template += "{% endfor %}"
+    return jinja_template
+
+
+def get_template_and_fix_tokenizer(
+    tokenizer: "PreTrainedTokenizer",
+    name: Optional[str] = None,
+) -> Template:
+    if name is None:
+        template = templates["vanilla"]  # placeholder
+    else:
+        template = templates.get(name, None)
+        if template is None:
+            raise ValueError("Template {} does not exist.".format(name))
 
     stop_words = template.stop_words
     if template.replace_eos:
         if not stop_words:
             raise ValueError("Stop words are required to replace the EOS token.")
 
-        tokenizer.eos_token = stop_words[0]
+        _add_or_replace_eos_token(tokenizer, eos_token=stop_words[0])
         stop_words = stop_words[1:]
-        logger.info("Replace eos token: {}".format(tokenizer.eos_token))
+
+    if tokenizer.eos_token_id is None:
+        _add_or_replace_eos_token(tokenizer, eos_token="<|endoftext|>")
+
+    if tokenizer.pad_token_id is None:
+        tokenizer.pad_token = tokenizer.eos_token
+        logger.info("Add pad token: {}".format(tokenizer.pad_token))
 
     if stop_words:
-        tokenizer.add_special_tokens(
+        num_added_tokens = tokenizer.add_special_tokens(
             dict(additional_special_tokens=stop_words), replace_additional_special_tokens=False
         )
         logger.info("Add {} to stop words.".format(",".join(stop_words)))
+        if num_added_tokens > 0:
+            logger.warning("New tokens have been added, make sure `resize_vocab` is True.")
+
+    try:
+        tokenizer.chat_template = _get_jinja_template(template, tokenizer)
+    except ValueError:
+        logger.info("Cannot add this chat template to tokenizer.")
 
     return template
 
 
-register_template(
+_register_template(
     name="alpaca",
     format_user=StringFormatter(slots=["### Instruction:\n{{content}}\n\n### Response:\n"]),
     format_separator=EmptyFormatter(slots=["\n\n"]),
@@ -279,7 +390,7 @@ register_template(
 )
 
 
-register_template(
+_register_template(
     name="aquila",
     format_user=StringFormatter(slots=["Human: {{content}}###Assistant:"]),
     format_separator=EmptyFormatter(slots=["###"]),
@@ -292,21 +403,30 @@ register_template(
 )
 
 
-register_template(
+_register_template(
+    name="atom",
+    format_user=StringFormatter(
+        slots=[{"bos_token"}, "Human: {{content}}\n", {"eos_token"}, {"bos_token"}, "Assistant:"]
+    ),
+    format_assistant=StringFormatter(slots=["{{content}}\n", {"eos_token"}]),
+)
+
+
+_register_template(
     name="baichuan",
-    format_user=StringFormatter(slots=[{"token": "<reserved_102>"}, "{{content}}", {"token": "<reserved_103>"}]),
+    format_user=StringFormatter(slots=["<reserved_102>{{content}}<reserved_103>"]),
     efficient_eos=True,
 )
 
 
-register_template(
+_register_template(
     name="baichuan2",
-    format_user=StringFormatter(slots=[{"token": "<reserved_106>"}, "{{content}}", {"token": "<reserved_107>"}]),
+    format_user=StringFormatter(slots=["<reserved_106>{{content}}<reserved_107>"]),
     efficient_eos=True,
 )
 
 
-register_template(
+_register_template(
     name="belle",
     format_user=StringFormatter(slots=["Human: {{content}}\n\nBelle: "]),
     format_system=StringFormatter(slots=[{"bos_token"}, "{{content}}"]),
@@ -315,13 +435,13 @@ register_template(
 )
 
 
-register_template(
+_register_template(
     name="bluelm",
     format_user=StringFormatter(slots=[{"token": "[|Human|]:"}, "{{content}}", {"token": "[|AI|]:"}]),
 )
 
 
-register_template(
+_register_template(
     name="chatglm2",
     format_user=StringFormatter(slots=["[Round {{idx}}]\n\n问：{{content}}\n\n答："]),
     format_system=StringFormatter(slots=[{"token": "[gMASK]"}, {"token": "sop"}, "{{content}}"]),
@@ -331,15 +451,32 @@ register_template(
 )
 
 
-register_template(
+_register_template(
     name="chatglm3",
     format_user=StringFormatter(slots=[{"token": "<|user|>"}, "\n", "{{content}}", {"token": "<|assistant|>"}]),
     format_assistant=StringFormatter(slots=["\n", "{{content}}"]),
+    format_system=StringFormatter(slots=[{"token": "[gMASK]"}, {"token": "sop"}, "{{content}}"]),
+    format_function=FunctionFormatter(slots=["{{name}}\n{{arguments}}"]),
+    format_observation=StringFormatter(
+        slots=[{"token": "<|observation|>"}, "\n", "{{content}}", {"token": "<|assistant|>"}]
+    ),
+    stop_words=["<|user|>", "<|observation|>"],
+    efficient_eos=True,
+    force_system=True,
+)
+
+
+_register_template(
+    name="chatglm3_system",
+    format_user=StringFormatter(slots=[{"token": "<|user|>"}, "\n", "{{content}}", {"token": "<|assistant|>"}]),
+    format_assistant=StringFormatter(slots=["\n", "{{content}}"]),
     format_system=StringFormatter(
         slots=[{"token": "[gMASK]"}, {"token": "sop"}, {"token": "<|system|>"}, "\n", "{{content}}"]
     ),
     format_function=FunctionFormatter(slots=["{{name}}\n{{arguments}}"]),
-    format_observation=StringFormatter(slots=[{"token": "<|observation|>"}, "\n", "{{content}}"]),
+    format_observation=StringFormatter(
+        slots=[{"token": "<|observation|>"}, "\n", "{{content}}", {"token": "<|assistant|>"}]
+    ),
     default_system=(
         "You are ChatGLM3, a large language model trained by Zhipu.AI. "
         "Follow the user's instructions carefully. Respond using markdown."
@@ -349,14 +486,43 @@ register_template(
 )
 
 
-register_template(
+_register_template(
+    name="chatml",
+    format_user=StringFormatter(slots=["<|im_start|>user\n{{content}}<|im_end|>\n<|im_start|>assistant\n"]),
+    format_system=StringFormatter(slots=["<|im_start|>system\n{{content}}<|im_end|>\n"]),
+    format_separator=EmptyFormatter(slots=["\n"]),
+    stop_words=["<|im_end|>", "<|im_start|>"],
+    replace_eos=True,
+)
+
+
+_register_template(
+    name="chatml_de",
+    format_user=StringFormatter(slots=["<|im_start|>user\n{{content}}<|im_end|>\n<|im_start|>assistant\n"]),
+    format_system=StringFormatter(slots=["<|im_start|>system\n{{content}}<|im_end|>\n"]),
+    format_separator=EmptyFormatter(slots=["\n"]),
+    default_system="Du bist ein freundlicher und hilfsbereiter KI-Assistent.",
+    stop_words=["<|im_end|>", "<|im_start|>"],
+    replace_eos=True,
+)
+
+
+_register_template(
     name="codegeex2",
     format_system=StringFormatter(slots=[{"token": "[gMASK]"}, {"token": "sop"}, "{{content}}"]),
     force_system=True,
 )
 
 
-register_template(
+_register_template(
+    name="cpm",
+    format_user=StringFormatter(slots=["<用户>{{content}}<AI>"]),
+    format_system=StringFormatter(slots=[{"bos_token"}, "{{content}}"]),
+    force_system=True,
+)
+
+
+_register_template(
     name="deepseek",
     format_user=StringFormatter(slots=["User: {{content}}\n\nAssistant:"]),
     format_system=StringFormatter(slots=[{"bos_token"}, "{{content}}"]),
@@ -364,10 +530,11 @@ register_template(
 )
 
 
-register_template(
+_register_template(
     name="deepseekcoder",
-    format_user=StringFormatter(slots=["### Instruction:\n{{content}}\n### Response:\n"]),
-    format_separator=EmptyFormatter(slots=["\n", {"token": "<|EOT|>"}, "\n"]),
+    format_user=StringFormatter(slots=["### Instruction:\n{{content}}\n### Response:"]),
+    format_assistant=StringFormatter(slots=["\n", "{{content}}"]),
+    format_separator=EmptyFormatter(slots=["\n<|EOT|>\n"]),
     default_system=(
         "You are an AI programming assistant, utilizing the Deepseek Coder model, "
         "developed by Deepseek Company, and you only answer questions related to computer science. "
@@ -379,14 +546,15 @@ register_template(
 )
 
 
-register_template(
+_register_template(
     name="default",
     format_user=StringFormatter(slots=["Human: {{content}}\nAssistant: "]),
+    format_system=StringFormatter(slots=["{{content}}\n"]),
     format_separator=EmptyFormatter(slots=["\n"]),
 )
 
 
-register_template(
+_register_template(
     name="falcon",
     format_user=StringFormatter(slots=["User: {{content}}\nFalcon:"]),
     format_separator=EmptyFormatter(slots=["\n"]),
@@ -394,7 +562,17 @@ register_template(
 )
 
 
-register_template(
+_register_template(
+    name="gemma",
+    format_user=StringFormatter(slots=["<start_of_turn>user\n{{content}}<end_of_turn>\n<start_of_turn>model\n"]),
+    format_system=StringFormatter(slots=[{"bos_token"}, "{{content}}"]),
+    format_separator=EmptyFormatter(slots=["<end_of_turn>\n"]),
+    efficient_eos=True,
+    force_system=True,
+)
+
+
+_register_template(
     name="intern",
     format_user=StringFormatter(slots=["<|User|>:{{content}}", {"token": "<eoh>"}, "\n<|Bot|>:"]),
     format_separator=EmptyFormatter(slots=[{"token": "<eoa>"}, "\n"]),
@@ -403,7 +581,7 @@ register_template(
 )
 
 
-register_template(
+_register_template(
     name="intern2",
     format_user=StringFormatter(slots=["<|im_start|>user\n{{content}}<|im_end|>\n<|im_start|>assistant\n"]),
     format_system=StringFormatter(slots=[{"bos_token"}, "<|im_start|>system\n{{content}}<|im_end|>\n"]),
@@ -420,7 +598,7 @@ register_template(
 )
 
 
-register_template(
+_register_template(
     name="llama2",
     format_user=StringFormatter(slots=[{"bos_token"}, "[INST] {{content}} [/INST]"]),
     format_system=StringFormatter(slots=["<<SYS>>\n{{content}}\n<</SYS>>\n\n"]),
@@ -437,7 +615,7 @@ register_template(
 )
 
 
-register_template(
+_register_template(
     name="llama2_zh",
     format_user=StringFormatter(slots=[{"bos_token"}, "[INST] {{content}} [/INST]"]),
     format_system=StringFormatter(slots=["<<SYS>>\n{{content}}\n<</SYS>>\n\n"]),
@@ -445,7 +623,7 @@ register_template(
 )
 
 
-register_template(
+_register_template(
     name="mistral",
     format_user=StringFormatter(slots=["[INST] {{content}} [/INST]"]),
     format_system=StringFormatter(slots=[{"bos_token"}, "{{content}}"]),
@@ -453,16 +631,33 @@ register_template(
 )
 
 
-register_template(
+_register_template(
+    name="olmo",
+    format_user=StringFormatter(slots=["<|user|>\n{{content}}<|assistant|>"]),
+    format_assistant=StringFormatter(slots=["{{content}}", {"eos_token"}]),
+    format_system=StringFormatter(slots=[{"eos_token"}, "{{content}}"]),
+    force_system=True,
+)
+
+
+_register_template(
     name="openchat",
     format_user=StringFormatter(slots=["GPT4 Correct User: {{content}}", {"eos_token"}, "GPT4 Correct Assistant:"]),
-    format_assistant=StringFormatter(slots=["{{content}}"]),
+    format_assistant=StringFormatter(slots=["{{content}}", {"eos_token"}]),
     format_system=StringFormatter(slots=[{"bos_token"}, "{{content}}"]),
     force_system=True,
 )
 
 
-register_template(
+_register_template(
+    name="orion",
+    format_user=StringFormatter(slots=["Human: {{content}}\n\nAssistant: ", {"eos_token"}]),
+    format_system=StringFormatter(slots=[{"bos_token"}, "{{content}}"]),
+    force_system=True,
+)
+
+
+_register_template(
     name="qwen",
     format_user=StringFormatter(slots=["<|im_start|>user\n{{content}}<|im_end|>\n<|im_start|>assistant\n"]),
     format_system=StringFormatter(slots=["<|im_start|>system\n{{content}}<|im_end|>\n"]),
@@ -473,7 +668,7 @@ register_template(
 )
 
 
-register_template(
+_register_template(
     name="solar",
     format_user=StringFormatter(slots=["### User:\n{{content}}\n\n### Assistant:\n"]),
     format_system=StringFormatter(slots=["### System:\n{{content}}\n\n"]),
@@ -481,12 +676,10 @@ register_template(
 )
 
 
-register_template(
+_register_template(
     name="starchat",
-    format_user=StringFormatter(
-        slots=[{"token": "<|user|>"}, "\n{{content}}", {"token": "<|end|>"}, "\n", {"token": "<|assistant|>"}]
-    ),
-    format_system=StringFormatter(slots=[{"token": "<|system|>"}, "\n{{content}}", {"token": "<|end|>"}, "\n"]),
+    format_user=StringFormatter(slots=["<|user|>\n{{content}}<|end|>\n<|assistant|>"]),
+    format_system=StringFormatter(slots=["<|system|>\n{{content}}<|end|>\n"]),
     format_separator=EmptyFormatter(slots=["\n"]),
     stop_words=["<|end|>"],
     replace_eos=True,
@@ -494,10 +687,12 @@ register_template(
 )
 
 
-register_template(name="vanilla")
+_register_template(
+    name="vanilla",
+)
 
 
-register_template(
+_register_template(
     name="vicuna",
     format_user=StringFormatter(slots=["USER: {{content}} ASSISTANT:"]),
     default_system=(
@@ -507,7 +702,7 @@ register_template(
 )
 
 
-register_template(
+_register_template(
     name="xuanyuan",
     format_user=StringFormatter(slots=["Human: {{content}} Assistant:"]),
     default_system=(
@@ -518,10 +713,13 @@ register_template(
 )
 
 
-register_template(name="xverse", format_user=StringFormatter(slots=["Human: {{content}}\n\nAssistant: "]))
+_register_template(
+    name="xverse",
+    format_user=StringFormatter(slots=["Human: {{content}}\n\nAssistant: "]),
+)
 
 
-register_template(
+_register_template(
     name="yayi",
     format_user=StringFormatter(slots=[{"token": "<|Human|>"}, ":\n{{content}}\n\n", {"token": "<|YaYi|>"}, ":"]),
     format_system=StringFormatter(slots=[{"token": "<|System|>"}, ":\n{{content}}\n\n"]),
@@ -541,7 +739,7 @@ register_template(
 )
 
 
-register_template(
+_register_template(
     name="yi",
     format_user=StringFormatter(slots=["<|im_start|>user\n{{content}}<|im_end|>\n<|im_start|>assistant\n"]),
     format_separator=EmptyFormatter(slots=["\n"]),
@@ -550,7 +748,7 @@ register_template(
 )
 
 
-register_template(
+_register_template(
     name="yuan",
     format_user=StringFormatter(slots=["{{content}}", {"token": "<sep>"}]),
     format_separator=EmptyFormatter(slots=["\n"]),
@@ -559,16 +757,17 @@ register_template(
 )
 
 
-register_template(
+_register_template(
     name="zephyr",
     format_user=StringFormatter(slots=["<|user|>\n{{content}}", {"eos_token"}, "<|assistant|>"]),
+    format_assistant=StringFormatter(slots=["\n{{content}}", {"eos_token"}]),
     format_system=StringFormatter(slots=["<|system|>\n{{content}}", {"eos_token"}]),
     default_system="You are a friendly chatbot who always responds in the style of a pirate",
 )
 
 
-register_template(
+_register_template(
     name="ziya",
-    format_user=StringFormatter(slots=[{"token": "<human>"}, ":{{content}}\n", {"token": "<bot>"}, ":"]),
+    format_user=StringFormatter(slots=["<human>:{{content}}\n<bot>:"]),
     format_separator=EmptyFormatter(slots=["\n"]),
 )

src/llmtuner/data/utils.py

@@ -2,12 +2,14 @@ import hashlib
 from enum import Enum, unique
 from typing import TYPE_CHECKING, Dict, List, Optional, Tuple, Union
 
+from datasets import concatenate_datasets, interleave_datasets
+
 from ..extras.logging import get_logger
 
 
 if TYPE_CHECKING:
     from datasets import Dataset, IterableDataset
-    from transformers import TrainingArguments
+    from transformers import Seq2SeqTrainingArguments
 
     from llmtuner.hparams import DataArguments
 
@@ -19,8 +21,9 @@ logger = get_logger(__name__)
 class Role(str, Enum):
     USER = "user"
     ASSISTANT = "assistant"
-    OBSERVATION = "observation"
+    SYSTEM = "system"
     FUNCTION = "function"
+    OBSERVATION = "observation"
 
 
 def checksum(data_files: List[str], file_sha1: Optional[str] = None) -> None:
@@ -45,8 +48,32 @@ def infer_max_len(source_len: int, target_len: int, max_len: int, reserved_label
     return max_source_len, max_target_len
 
 
+def merge_dataset(
+    all_datasets: List[Union["Dataset", "IterableDataset"]],
+    data_args: "DataArguments",
+    training_args: "Seq2SeqTrainingArguments",
+) -> Union["Dataset", "IterableDataset"]:
+    if len(all_datasets) == 1:
+        return all_datasets[0]
+    elif data_args.mix_strategy == "concat":
+        if data_args.streaming:
+            logger.warning("The samples between different datasets will not be mixed in streaming mode.")
+        return concatenate_datasets(all_datasets)
+    elif data_args.mix_strategy.startswith("interleave"):
+        if not data_args.streaming:
+            logger.warning("We recommend using `mix_strategy=concat` in non-streaming mode.")
+        return interleave_datasets(
+            datasets=all_datasets,
+            probabilities=data_args.interleave_probs,
+            seed=training_args.seed,
+            stopping_strategy="first_exhausted" if data_args.mix_strategy.endswith("under") else "all_exhausted",
+        )
+    else:
+        raise ValueError("Unknown mixing strategy.")
+
+
 def split_dataset(
-    dataset: Union["Dataset", "IterableDataset"], data_args: "DataArguments", training_args: "TrainingArguments"
+    dataset: Union["Dataset", "IterableDataset"], data_args: "DataArguments", training_args: "Seq2SeqTrainingArguments"
 ) -> Dict[str, "Dataset"]:
     if training_args.do_train:
         if data_args.val_size > 1e-6:  # Split the dataset

src/llmtuner/eval/evaluator.py

@@ -14,7 +14,7 @@ from transformers.utils import cached_file
 from ..data import get_template_and_fix_tokenizer
 from ..extras.constants import CHOICES, SUBJECTS
 from ..hparams import get_eval_args
-from ..model import dispatch_model, load_model_and_tokenizer
+from ..model import load_model_and_tokenizer
 from .template import get_eval_template
 
 
@@ -23,8 +23,7 @@ class Evaluator:
         self.model_args, self.data_args, self.eval_args, finetuning_args = get_eval_args(args)
         self.model, self.tokenizer = load_model_and_tokenizer(self.model_args, finetuning_args)
         self.tokenizer.padding_side = "right"  # avoid overflow issue in batched inference for llama2
-        self.model = dispatch_model(self.model)
-        self.template = get_template_and_fix_tokenizer(self.data_args.template, self.tokenizer)
+        self.template = get_template_and_fix_tokenizer(self.tokenizer, self.data_args.template)
         self.eval_template = get_eval_template(self.eval_args.lang)
         self.choice_inputs = [
             self.tokenizer.encode(self.eval_template.prefix + ch, add_special_tokens=False)[-1] for ch in CHOICES

src/llmtuner/extras/constants.py

@@ -11,7 +11,14 @@ DEFAULT_MODULE = defaultdict(str)
 
 DEFAULT_TEMPLATE = defaultdict(str)
 
-FILEEXT2TYPE = {"arrow": "arrow", "csv": "csv", "json": "json", "jsonl": "json", "parquet": "parquet", "txt": "text"}
+FILEEXT2TYPE = {
+    "arrow": "arrow",
+    "csv": "csv",
+    "json": "json",
+    "jsonl": "json",
+    "parquet": "parquet",
+    "txt": "text",
+}
 
 IGNORE_INDEX = -100
 
@@ -46,7 +53,9 @@ class DownloadSource(str, Enum):
 
 
 def register_model_group(
-    models: Dict[str, Dict[DownloadSource, str]], module: Optional[str] = None, template: Optional[str] = None
+    models: Dict[str, Dict[DownloadSource, str]],
+    module: Optional[str] = None,
+    template: Optional[str] = None,
 ) -> None:
     prefix = None
     for name, path in models.items():
@@ -219,22 +228,36 @@ register_model_group(
 
 register_model_group(
     models={
-        "DeepSeekLLM-7B-Base": {
+        "DeepSeek-LLM-7B-Base": {
             DownloadSource.DEFAULT: "deepseek-ai/deepseek-llm-7b-base",
             DownloadSource.MODELSCOPE: "deepseek-ai/deepseek-llm-7b-base",
         },
-        "DeepSeekLLM-67B-Base": {
+        "DeepSeek-LLM-67B-Base": {
             DownloadSource.DEFAULT: "deepseek-ai/deepseek-llm-67b-base",
             DownloadSource.MODELSCOPE: "deepseek-ai/deepseek-llm-67b-base",
         },
-        "DeepSeekLLM-7B-Chat": {
+        "DeepSeek-LLM-7B-Chat": {
             DownloadSource.DEFAULT: "deepseek-ai/deepseek-llm-7b-chat",
             DownloadSource.MODELSCOPE: "deepseek-ai/deepseek-llm-7b-chat",
         },
-        "DeepSeekLLM-67B-Chat": {
+        "DeepSeek-LLM-67B-Chat": {
             DownloadSource.DEFAULT: "deepseek-ai/deepseek-llm-67b-chat",
             DownloadSource.MODELSCOPE: "deepseek-ai/deepseek-llm-67b-chat",
         },
+        "DeepSeek-Math-7B-Base": {
+            DownloadSource.DEFAULT: "deepseek-ai/deepseek-math-7b-base",
+        },
+        "DeepSeek-Math-7B-Chat": {
+            DownloadSource.DEFAULT: "deepseek-ai/deepseek-math-7b-instruct",
+        },
+        "DeepSeek-MoE-16B-Base": {
+            DownloadSource.DEFAULT: "deepseek-ai/deepseek-moe-16b-base",
+            DownloadSource.MODELSCOPE: "deepseek-ai/deepseek-moe-16b-base",
+        },
+        "DeepSeek-MoE-16B-Chat": {
+            DownloadSource.DEFAULT: "deepseek-ai/deepseek-moe-16b-chat",
+            DownloadSource.MODELSCOPE: "deepseek-ai/deepseek-moe-16b-chat",
+        },
     },
     template="deepseek",
 )
@@ -246,6 +269,9 @@ register_model_group(
             DownloadSource.DEFAULT: "deepseek-ai/deepseek-coder-6.7b-base",
             DownloadSource.MODELSCOPE: "deepseek-ai/deepseek-coder-6.7b-base",
         },
+        "DeepSeekCoder-7B-Base": {
+            DownloadSource.DEFAULT: "deepseek-ai/deepseek-coder-7b-base-v1.5",
+        },
         "DeepSeekCoder-33B-Base": {
             DownloadSource.DEFAULT: "deepseek-ai/deepseek-coder-33b-base",
             DownloadSource.MODELSCOPE: "deepseek-ai/deepseek-coder-33b-base",
@@ -254,6 +280,9 @@ register_model_group(
             DownloadSource.DEFAULT: "deepseek-ai/deepseek-coder-6.7b-instruct",
             DownloadSource.MODELSCOPE: "deepseek-ai/deepseek-coder-6.7b-instruct",
         },
+        "DeepSeekCoder-7B-Chat": {
+            DownloadSource.DEFAULT: "deepseek-ai/deepseek-coder-7b-instruct-v1.5",
+        },
         "DeepSeekCoder-33B-Chat": {
             DownloadSource.DEFAULT: "deepseek-ai/deepseek-coder-33b-instruct",
             DownloadSource.MODELSCOPE: "deepseek-ai/deepseek-coder-33b-instruct",
@@ -263,21 +292,6 @@ register_model_group(
 )
 
 
-register_model_group(
-    models={
-        "DeepSeekMoE-16B-Base": {
-            DownloadSource.DEFAULT: "deepseek-ai/deepseek-moe-16b-base",
-            DownloadSource.MODELSCOPE: "deepseek-ai/deepseek-moe-16b-base",
-        },
-        "DeepSeekMoE-16B-Chat": {
-            DownloadSource.DEFAULT: "deepseek-ai/deepseek-moe-16b-chat",
-            DownloadSource.MODELSCOPE: "deepseek-ai/deepseek-moe-16b-chat",
-        },
-    },
-    template="deepseek",
-)
-
-
 register_model_group(
     models={
         "Falcon-7B": {
@@ -310,6 +324,29 @@ register_model_group(
 )
 
 
+register_model_group(
+    models={
+        "Gemma-2B": {
+            DownloadSource.DEFAULT: "google/gemma-2b",
+            DownloadSource.MODELSCOPE: "AI-ModelScope/gemma-2b",
+        },
+        "Gemma-7B": {
+            DownloadSource.DEFAULT: "google/gemma-7b",
+            DownloadSource.MODELSCOPE: "AI-ModelScope/gemma-2b-it",
+        },
+        "Gemma-2B-Chat": {
+            DownloadSource.DEFAULT: "google/gemma-2b-it",
+            DownloadSource.MODELSCOPE: "AI-ModelScope/gemma-7b",
+        },
+        "Gemma-7B-Chat": {
+            DownloadSource.DEFAULT: "google/gemma-7b-it",
+            DownloadSource.MODELSCOPE: "AI-ModelScope/gemma-7b-it",
+        },
+    },
+    template="gemma",
+)
+
+
 register_model_group(
     models={
         "InternLM-7B": {
@@ -370,7 +407,10 @@ register_model_group(
 
 register_model_group(
     models={
-        "LLaMA-7B": {DownloadSource.DEFAULT: "huggyllama/llama-7b", DownloadSource.MODELSCOPE: "skyline2006/llama-7b"},
+        "LLaMA-7B": {
+            DownloadSource.DEFAULT: "huggyllama/llama-7b",
+            DownloadSource.MODELSCOPE: "skyline2006/llama-7b",
+        },
         "LLaMA-13B": {
             DownloadSource.DEFAULT: "huggyllama/llama-13b",
             DownloadSource.MODELSCOPE: "skyline2006/llama-13b",
@@ -452,10 +492,28 @@ register_model_group(
 )
 
 
+register_model_group(
+    models={
+        "OLMo-1B": {
+            DownloadSource.DEFAULT: "allenai/OLMo-1B",
+        },
+        "OLMo-7B": {
+            DownloadSource.DEFAULT: "allenai/OLMo-7B",
+            DownloadSource.MODELSCOPE: "AI-ModelScope/OLMo-7B",
+        },
+        "OLMo-7B-Chat": {
+            DownloadSource.DEFAULT: "allenai/OLMo-7B-Instruct",
+        },
+    },
+    module="att_proj",
+    template="olmo",
+)
+
+
 register_model_group(
     models={
         "OpenChat3.5-7B-Chat": {
-            DownloadSource.DEFAULT: "openchat/openchat_3.5",
+            DownloadSource.DEFAULT: "openchat/openchat-3.5-0106",
             DownloadSource.MODELSCOPE: "myxiongmodel/openchat_3.5",
         }
     },
@@ -465,23 +523,71 @@ register_model_group(
 
 register_model_group(
     models={
-        "Phi-1.5-1.3B": {DownloadSource.DEFAULT: "microsoft/phi-1_5", DownloadSource.MODELSCOPE: "allspace/PHI_1-5"},
-        "Phi-2-2.7B": {DownloadSource.DEFAULT: "microsoft/phi-2", DownloadSource.MODELSCOPE: "AI-ModelScope/phi-2"},
+        "Orion-14B-Base": {
+            DownloadSource.DEFAULT: "OrionStarAI/Orion-14B-Base",
+            DownloadSource.MODELSCOPE: "OrionStarAI/Orion-14B-Base",
+        },
+        "Orion-14B-Chat": {
+            DownloadSource.DEFAULT: "OrionStarAI/Orion-14B-Chat",
+            DownloadSource.MODELSCOPE: "OrionStarAI/Orion-14B-Chat",
+        },
+        "Orion-14B-Long-Chat": {
+            DownloadSource.DEFAULT: "OrionStarAI/Orion-14B-LongChat",
+            DownloadSource.MODELSCOPE: "OrionStarAI/Orion-14B-LongChat",
+        },
+        "Orion-14B-RAG-Chat": {
+            DownloadSource.DEFAULT: "OrionStarAI/Orion-14B-Chat-RAG",
+            DownloadSource.MODELSCOPE: "OrionStarAI/Orion-14B-Chat-RAG",
+        },
+        "Orion-14B-Plugin-Chat": {
+            DownloadSource.DEFAULT: "OrionStarAI/Orion-14B-Chat-Plugin",
+            DownloadSource.MODELSCOPE: "OrionStarAI/Orion-14B-Chat-Plugin",
+        },
+    },
+    template="orion",
+)
+
+
+register_model_group(
+    models={
+        "Phi-1.5-1.3B": {
+            DownloadSource.DEFAULT: "microsoft/phi-1_5",
+            DownloadSource.MODELSCOPE: "allspace/PHI_1-5",
+        },
+        "Phi-2-2.7B": {
+            DownloadSource.DEFAULT: "microsoft/phi-2",
+            DownloadSource.MODELSCOPE: "AI-ModelScope/phi-2",
+        },
     }
 )
 
 
 register_model_group(
     models={
-        "Qwen-1.8B": {DownloadSource.DEFAULT: "Qwen/Qwen-1_8B", DownloadSource.MODELSCOPE: "qwen/Qwen-1_8B"},
-        "Qwen-7B": {DownloadSource.DEFAULT: "Qwen/Qwen-7B", DownloadSource.MODELSCOPE: "qwen/Qwen-7B"},
-        "Qwen-14B": {DownloadSource.DEFAULT: "Qwen/Qwen-14B", DownloadSource.MODELSCOPE: "qwen/Qwen-14B"},
-        "Qwen-72B": {DownloadSource.DEFAULT: "Qwen/Qwen-72B", DownloadSource.MODELSCOPE: "qwen/Qwen-72B"},
+        "Qwen-1.8B": {
+            DownloadSource.DEFAULT: "Qwen/Qwen-1_8B",
+            DownloadSource.MODELSCOPE: "qwen/Qwen-1_8B",
+        },
+        "Qwen-7B": {
+            DownloadSource.DEFAULT: "Qwen/Qwen-7B",
+            DownloadSource.MODELSCOPE: "qwen/Qwen-7B",
+        },
+        "Qwen-14B": {
+            DownloadSource.DEFAULT: "Qwen/Qwen-14B",
+            DownloadSource.MODELSCOPE: "qwen/Qwen-14B",
+        },
+        "Qwen-72B": {
+            DownloadSource.DEFAULT: "Qwen/Qwen-72B",
+            DownloadSource.MODELSCOPE: "qwen/Qwen-72B",
+        },
         "Qwen-1.8B-Chat": {
             DownloadSource.DEFAULT: "Qwen/Qwen-1_8B-Chat",
             DownloadSource.MODELSCOPE: "qwen/Qwen-1_8B-Chat",
         },
-        "Qwen-7B-Chat": {DownloadSource.DEFAULT: "Qwen/Qwen-7B-Chat", DownloadSource.MODELSCOPE: "qwen/Qwen-7B-Chat"},
+        "Qwen-7B-Chat": {
+            DownloadSource.DEFAULT: "Qwen/Qwen-7B-Chat",
+            DownloadSource.MODELSCOPE: "qwen/Qwen-7B-Chat",
+        },
         "Qwen-14B-Chat": {
             DownloadSource.DEFAULT: "Qwen/Qwen-14B-Chat",
             DownloadSource.MODELSCOPE: "qwen/Qwen-14B-Chat",
@@ -530,7 +636,112 @@ register_model_group(
 
 register_model_group(
     models={
-        "SOLAR-10.7B": {DownloadSource.DEFAULT: "upstage/SOLAR-10.7B-v1.0"},
+        "Qwen1.5-0.5B": {
+            DownloadSource.DEFAULT: "Qwen/Qwen1.5-0.5B",
+            DownloadSource.MODELSCOPE: "qwen/Qwen1.5-0.5B",
+        },
+        "Qwen1.5-1.8B": {
+            DownloadSource.DEFAULT: "Qwen/Qwen1.5-1.8B",
+            DownloadSource.MODELSCOPE: "qwen/Qwen1.5-1.8B",
+        },
+        "Qwen1.5-4B": {
+            DownloadSource.DEFAULT: "Qwen/Qwen1.5-4B",
+            DownloadSource.MODELSCOPE: "qwen/Qwen1.5-4B",
+        },
+        "Qwen1.5-7B": {
+            DownloadSource.DEFAULT: "Qwen/Qwen1.5-7B",
+            DownloadSource.MODELSCOPE: "qwen/Qwen1.5-7B",
+        },
+        "Qwen1.5-14B": {
+            DownloadSource.DEFAULT: "Qwen/Qwen1.5-14B",
+            DownloadSource.MODELSCOPE: "qwen/Qwen1.5-14B",
+        },
+        "Qwen1.5-72B": {
+            DownloadSource.DEFAULT: "Qwen/Qwen1.5-72B",
+            DownloadSource.MODELSCOPE: "qwen/Qwen1.5-72B",
+        },
+        "Qwen1.5-0.5B-Chat": {
+            DownloadSource.DEFAULT: "Qwen/Qwen1.5-0.5B-Chat",
+            DownloadSource.MODELSCOPE: "qwen/Qwen1.5-0.5B-Chat",
+        },
+        "Qwen1.5-1.8B-Chat": {
+            DownloadSource.DEFAULT: "Qwen/Qwen1.5-1.8B-Chat",
+            DownloadSource.MODELSCOPE: "qwen/Qwen1.5-1.8B-Chat",
+        },
+        "Qwen1.5-4B-Chat": {
+            DownloadSource.DEFAULT: "Qwen/Qwen1.5-4B-Chat",
+            DownloadSource.MODELSCOPE: "qwen/Qwen1.5-4B-Chat",
+        },
+        "Qwen1.5-7B-Chat": {
+            DownloadSource.DEFAULT: "Qwen/Qwen1.5-7B-Chat",
+            DownloadSource.MODELSCOPE: "qwen/Qwen1.5-7B-Chat",
+        },
+        "Qwen1.5-14B-Chat": {
+            DownloadSource.DEFAULT: "Qwen/Qwen1.5-14B-Chat",
+            DownloadSource.MODELSCOPE: "qwen/Qwen1.5-14B-Chat",
+        },
+        "Qwen1.5-72B-Chat": {
+            DownloadSource.DEFAULT: "Qwen/Qwen1.5-72B-Chat",
+            DownloadSource.MODELSCOPE: "qwen/Qwen1.5-72B-Chat",
+        },
+        "Qwen1.5-0.5B-int8-Chat": {
+            DownloadSource.DEFAULT: "Qwen/Qwen1.5-0.5B-Chat-GPTQ-Int8",
+            DownloadSource.MODELSCOPE: "qwen/Qwen1.5-0.5B-Chat-GPTQ-Int8",
+        },
+        "Qwen1.5-0.5B-int4-Chat": {
+            DownloadSource.DEFAULT: "Qwen/Qwen1.5-0.5B-Chat-AWQ",
+            DownloadSource.MODELSCOPE: "qwen/Qwen1.5-0.5B-Chat-AWQ",
+        },
+        "Qwen1.5-1.8B-int8-Chat": {
+            DownloadSource.DEFAULT: "Qwen/Qwen1.5-1.8B-Chat-GPTQ-Int8",
+            DownloadSource.MODELSCOPE: "qwen/Qwen1.5-1.8B-Chat-GPTQ-Int8",
+        },
+        "Qwen1.5-1.8B-int4-Chat": {
+            DownloadSource.DEFAULT: "Qwen/Qwen1.5-1.8B-Chat-AWQ",
+            DownloadSource.MODELSCOPE: "qwen/Qwen1.5-1.8B-Chat-AWQ",
+        },
+        "Qwen1.5-4B-int8-Chat": {
+            DownloadSource.DEFAULT: "Qwen/Qwen1.5-4B-Chat-GPTQ-Int8",
+            DownloadSource.MODELSCOPE: "qwen/Qwen1.5-4B-Chat-GPTQ-Int8",
+        },
+        "Qwen1.5-4B-int4-Chat": {
+            DownloadSource.DEFAULT: "Qwen/Qwen1.5-4B-Chat-AWQ",
+            DownloadSource.MODELSCOPE: "qwen/Qwen1.5-4B-Chat-AWQ",
+        },
+        "Qwen1.5-7B-int8-Chat": {
+            DownloadSource.DEFAULT: "Qwen/Qwen1.5-7B-Chat-GPTQ-Int8",
+            DownloadSource.MODELSCOPE: "qwen/Qwen1.5-7B-Chat-GPTQ-Int8",
+        },
+        "Qwen1.5-7B-int4-Chat": {
+            DownloadSource.DEFAULT: "Qwen/Qwen1.5-7B-Chat-AWQ",
+            DownloadSource.MODELSCOPE: "qwen/Qwen1.5-7B-Chat-AWQ",
+        },
+        "Qwen1.5-14B-int8-Chat": {
+            DownloadSource.DEFAULT: "Qwen/Qwen1.5-14B-Chat-GPTQ-Int8",
+            DownloadSource.MODELSCOPE: "qwen/Qwen1.5-14B-Chat-GPTQ-Int8",
+        },
+        "Qwen1.5-14B-int4-Chat": {
+            DownloadSource.DEFAULT: "Qwen/Qwen1.5-14B-Chat-AWQ",
+            DownloadSource.MODELSCOPE: "qwen/Qwen1.5-14B-Chat-AWQ",
+        },
+        "Qwen1.5-72B-int8-Chat": {
+            DownloadSource.DEFAULT: "Qwen/Qwen1.5-72B-Chat-GPTQ-Int8",
+            DownloadSource.MODELSCOPE: "qwen/Qwen1.5-72B-Chat-GPTQ-Int8",
+        },
+        "Qwen1.5-72B-int4-Chat": {
+            DownloadSource.DEFAULT: "Qwen/Qwen1.5-72B-Chat-AWQ",
+            DownloadSource.MODELSCOPE: "qwen/Qwen1.5-72B-Chat-AWQ",
+        },
+    },
+    template="qwen",
+)
+
+
+register_model_group(
+    models={
+        "SOLAR-10.7B": {
+            DownloadSource.DEFAULT: "upstage/SOLAR-10.7B-v1.0",
+        },
         "SOLAR-10.7B-Chat": {
             DownloadSource.DEFAULT: "upstage/SOLAR-10.7B-Instruct-v1.0",
             DownloadSource.MODELSCOPE: "AI-ModelScope/SOLAR-10.7B-Instruct-v1.0",
@@ -550,6 +761,21 @@ register_model_group(
 )
 
 
+register_model_group(
+    models={
+        "StarCoder2-3B": {
+            DownloadSource.DEFAULT: "bigcode/starcoder2-3b",
+        },
+        "StarCoder2-7B": {
+            DownloadSource.DEFAULT: "bigcode/starcoder2-7b",
+        },
+        "StarCoder2-15B": {
+            DownloadSource.DEFAULT: "bigcode/starcoder2-15b",
+        },
+    }
+)
+
+
 register_model_group(
     models={
         "Vicuna1.5-7B-Chat": {
@@ -567,10 +793,18 @@ register_model_group(
 
 register_model_group(
     models={
-        "XuanYuan-70B": {DownloadSource.DEFAULT: "Duxiaoman-DI/XuanYuan-70B"},
-        "XuanYuan-70B-Chat": {DownloadSource.DEFAULT: "Duxiaoman-DI/XuanYuan-70B-Chat"},
-        "XuanYuan-70B-int8-Chat": {DownloadSource.DEFAULT: "Duxiaoman-DI/XuanYuan-70B-Chat-8bit"},
-        "XuanYuan-70B-int4-Chat": {DownloadSource.DEFAULT: "Duxiaoman-DI/XuanYuan-70B-Chat-4bit"},
+        "XuanYuan-70B": {
+            DownloadSource.DEFAULT: "Duxiaoman-DI/XuanYuan-70B",
+        },
+        "XuanYuan-70B-Chat": {
+            DownloadSource.DEFAULT: "Duxiaoman-DI/XuanYuan-70B-Chat",
+        },
+        "XuanYuan-70B-int8-Chat": {
+            DownloadSource.DEFAULT: "Duxiaoman-DI/XuanYuan-70B-Chat-8bit",
+        },
+        "XuanYuan-70B-int4-Chat": {
+            DownloadSource.DEFAULT: "Duxiaoman-DI/XuanYuan-70B-Chat-4bit",
+        },
     },
     template="xuanyuan",
 )
@@ -578,9 +812,18 @@ register_model_group(
 
 register_model_group(
     models={
-        "XVERSE-7B": {DownloadSource.DEFAULT: "xverse/XVERSE-7B", DownloadSource.MODELSCOPE: "xverse/XVERSE-7B"},
-        "XVERSE-13B": {DownloadSource.DEFAULT: "xverse/XVERSE-13B", DownloadSource.MODELSCOPE: "xverse/XVERSE-13B"},
-        "XVERSE-65B": {DownloadSource.DEFAULT: "xverse/XVERSE-65B", DownloadSource.MODELSCOPE: "xverse/XVERSE-65B"},
+        "XVERSE-7B": {
+            DownloadSource.DEFAULT: "xverse/XVERSE-7B",
+            DownloadSource.MODELSCOPE: "xverse/XVERSE-7B",
+        },
+        "XVERSE-13B": {
+            DownloadSource.DEFAULT: "xverse/XVERSE-13B",
+            DownloadSource.MODELSCOPE: "xverse/XVERSE-13B",
+        },
+        "XVERSE-65B": {
+            DownloadSource.DEFAULT: "xverse/XVERSE-65B",
+            DownloadSource.MODELSCOPE: "xverse/XVERSE-65B",
+        },
         "XVERSE-65B-2": {
             DownloadSource.DEFAULT: "xverse/XVERSE-65B-2",
             DownloadSource.MODELSCOPE: "xverse/XVERSE-65B-2",
@@ -619,18 +862,42 @@ register_model_group(
 
 register_model_group(
     models={
-        "Yi-6B": {DownloadSource.DEFAULT: "01-ai/Yi-6B", DownloadSource.MODELSCOPE: "01ai/Yi-6B"},
-        "Yi-34B": {DownloadSource.DEFAULT: "01-ai/Yi-34B", DownloadSource.MODELSCOPE: "01ai/Yi-34B"},
-        "Yi-6B-Chat": {DownloadSource.DEFAULT: "01-ai/Yi-6B-Chat", DownloadSource.MODELSCOPE: "01ai/Yi-6B-Chat"},
-        "Yi-34B-Chat": {DownloadSource.DEFAULT: "01-ai/Yi-34B-Chat", DownloadSource.MODELSCOPE: "01ai/Yi-34B-Chat"},
+        "Yi-6B": {
+            DownloadSource.DEFAULT: "01-ai/Yi-6B",
+            DownloadSource.MODELSCOPE: "01ai/Yi-6B",
+        },
+        "Yi-9B": {
+            DownloadSource.DEFAULT: "01-ai/Yi-9B",
+            DownloadSource.MODELSCOPE: "01ai/Yi-9B",
+        },
+        "Yi-34B": {
+            DownloadSource.DEFAULT: "01-ai/Yi-34B",
+            DownloadSource.MODELSCOPE: "01ai/Yi-34B",
+        },
+        "Yi-6B-Chat": {
+            DownloadSource.DEFAULT: "01-ai/Yi-6B-Chat",
+            DownloadSource.MODELSCOPE: "01ai/Yi-6B-Chat",
+        },
+        "Yi-34B-Chat": {
+            DownloadSource.DEFAULT: "01-ai/Yi-34B-Chat",
+            DownloadSource.MODELSCOPE: "01ai/Yi-34B-Chat",
+        },
         "Yi-6B-int8-Chat": {
             DownloadSource.DEFAULT: "01-ai/Yi-6B-Chat-8bits",
             DownloadSource.MODELSCOPE: "01ai/Yi-6B-Chat-8bits",
         },
+        "Yi-6B-int4-Chat": {
+            DownloadSource.DEFAULT: "01-ai/Yi-6B-Chat-4bits",
+            DownloadSource.MODELSCOPE: "01ai/Yi-6B-Chat-4bits",
+        },
         "Yi-34B-int8-Chat": {
             DownloadSource.DEFAULT: "01-ai/Yi-34B-Chat-8bits",
             DownloadSource.MODELSCOPE: "01ai/Yi-34B-Chat-8bits",
         },
+        "Yi-34B-int4-Chat": {
+            DownloadSource.DEFAULT: "01-ai/Yi-34B-Chat-4bits",
+            DownloadSource.MODELSCOPE: "01ai/Yi-34B-Chat-4bits",
+        },
     },
     template="yi",
 )
@@ -668,3 +935,18 @@ register_model_group(
     },
     template="zephyr",
 )
+
+
+register_model_group(
+    models={
+        "Atom-7B": {
+            DownloadSource.DEFAULT: "FlagAlpha/Atom-7B",
+            DownloadSource.MODELSCOPE: "FlagAlpha/Atom-7B",
+        },
+        "Atom-7B-Chat": {
+            DownloadSource.DEFAULT: "FlagAlpha/Atom-7B-Chat",
+            DownloadSource.MODELSCOPE: "FlagAlpha/Atom-7B-Chat",
+        },
+    },
+    template="atom",
+)

src/llmtuner/extras/misc.py

@@ -10,9 +10,11 @@ from transformers.utils import (
     WEIGHTS_NAME,
     is_torch_bf16_gpu_available,
     is_torch_cuda_available,
+    is_torch_mps_available,
     is_torch_npu_available,
     is_torch_xpu_available,
 )
+from transformers.utils.versions import require_version
 
 from .constants import V_HEAD_SAFE_WEIGHTS_NAME, V_HEAD_WEIGHTS_NAME
 from .logging import get_logger
@@ -55,6 +57,17 @@ class AverageMeter:
         self.avg = self.sum / self.count
 
 
+def check_dependencies() -> None:
+    if int(os.environ.get("DISABLE_VERSION_CHECK", "0")):
+        logger.warning("Version checking has been disabled, may lead to unexpected behaviors.")
+    else:
+        require_version("transformers>=4.37.2", "To fix: pip install transformers>=4.37.2")
+        require_version("datasets>=2.14.3", "To fix: pip install datasets>=2.14.3")
+        require_version("accelerate>=0.27.2", "To fix: pip install accelerate>=0.27.2")
+        require_version("peft>=0.9.0", "To fix: pip install peft>=0.9.0")
+        require_version("trl>=0.8.1", "To fix: pip install trl>=0.8.1")
+
+
 def count_parameters(model: torch.nn.Module) -> Tuple[int, int]:
     r"""
     Returns the number of trainable parameters and number of all parameters in the model.
@@ -68,7 +81,12 @@ def count_parameters(model: torch.nn.Module) -> Tuple[int, int]:
 
         # Due to the design of 4bit linear layers from bitsandbytes, multiply the number of parameters by 2
         if param.__class__.__name__ == "Params4bit":
-            num_params = num_params * 2
+            if hasattr(param, "quant_storage") and hasattr(param.quant_storage, "itemsize"):
+                num_bytes = param.quant_storage.itemsize
+            else:
+                num_bytes = 1
+
+            num_params = num_params * 2 * num_bytes
 
         all_param += num_params
         if param.requires_grad:
@@ -133,6 +151,8 @@ def get_current_device() -> torch.device:
         device = "xpu:{}".format(os.environ.get("LOCAL_RANK", "0"))
     elif is_torch_npu_available():
         device = "npu:{}".format(os.environ.get("LOCAL_RANK", "0"))
+    elif is_torch_mps_available():
+        device = "mps:{}".format(os.environ.get("LOCAL_RANK", "0"))
     elif is_torch_cuda_available():
         device = "cuda:{}".format(os.environ.get("LOCAL_RANK", "0"))
     else:
@@ -142,6 +162,12 @@ def get_current_device() -> torch.device:
 
 
 def get_device_count() -> int:
+    r"""
+    Gets the number of available GPU devices.
+    """
+    if not torch.cuda.is_available():
+        return 0
+
     return torch.cuda.device_count()
 
 

src/llmtuner/extras/packages.py

@@ -2,11 +2,11 @@ import importlib.metadata
 import importlib.util
 
 
-def is_package_available(name: str) -> bool:
+def _is_package_available(name: str) -> bool:
     return importlib.util.find_spec(name) is not None
 
 
-def get_package_version(name: str) -> str:
+def _get_package_version(name: str) -> str:
     try:
         return importlib.metadata.version(name)
     except Exception:
@@ -14,36 +14,48 @@ def get_package_version(name: str) -> str:
 
 
 def is_fastapi_availble():
-    return is_package_available("fastapi")
+    return _is_package_available("fastapi")
 
 
 def is_flash_attn2_available():
-    return is_package_available("flash_attn") and get_package_version("flash_attn").startswith("2")
+    return _is_package_available("flash_attn") and _get_package_version("flash_attn").startswith("2")
+
+
+def is_galore_available():
+    return _is_package_available("galore_torch")
 
 
 def is_jieba_available():
-    return is_package_available("jieba")
+    return _is_package_available("jieba")
 
 
 def is_matplotlib_available():
-    return is_package_available("matplotlib")
+    return _is_package_available("matplotlib")
 
 
 def is_nltk_available():
-    return is_package_available("nltk")
+    return _is_package_available("nltk")
 
 
 def is_requests_available():
-    return is_package_available("requests")
+    return _is_package_available("requests")
 
 
 def is_rouge_available():
-    return is_package_available("rouge_chinese")
+    return _is_package_available("rouge_chinese")
 
 
 def is_starlette_available():
-    return is_package_available("sse_starlette")
+    return _is_package_available("sse_starlette")
+
+
+def is_unsloth_available():
+    return _is_package_available("unsloth")
 
 
 def is_uvicorn_available():
-    return is_package_available("uvicorn")
+    return _is_package_available("uvicorn")
+
+
+def is_vllm_available():
+    return _is_package_available("vllm")

src/llmtuner/extras/patches/llama_patch.py

@@ -11,12 +11,14 @@ from transformers.models.llama.modeling_llama import (
     repeat_kv,
 )
 from transformers.utils import logging
+from transformers.utils.versions import require_version
 
 
 logger = logging.get_logger(__name__)
 
 
-# Modified from: https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/modeling_llama.py
+# Modified from:
+# https://github.com/huggingface/transformers/blob/v4.39.1/src/transformers/models/llama/modeling_llama.py
 def llama_torch_attn_forward(
     self: "LlamaAttention",
     hidden_states: torch.Tensor,
@@ -24,6 +26,7 @@ def llama_torch_attn_forward(
     position_ids: Optional[torch.LongTensor] = None,
     past_key_value: Optional["Cache"] = None,
     output_attentions: bool = False,
+    cache_position: Optional[torch.LongTensor] = None,
     **kwargs,
 ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
     bsz, q_len, _ = hidden_states.size()
@@ -36,15 +39,12 @@ def llama_torch_attn_forward(
     key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
     value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
 
-    kv_seq_len = key_states.shape[-2]
-    if past_key_value is not None:
-        kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
-
-    cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
-    query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
+    past_key_value = getattr(self, "past_key_value", past_key_value)
+    cos, sin = self.rotary_emb(value_states, position_ids)
+    query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
 
     if past_key_value is not None:
-        cache_kwargs = {"sin": sin, "cos": cos}  # Specific to RoPE models
+        cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
         key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
 
     key_states = repeat_kv(key_states, self.num_key_value_groups)
@@ -96,14 +96,16 @@ def llama_torch_attn_forward(
     return attn_output, attn_weights, past_key_value
 
 
-# Modified from: https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/modeling_llama.py
+# Modified from:
+# https://github.com/huggingface/transformers/blob/v4.39.1/src/transformers/models/llama/modeling_llama.py
 def llama_flash_attn_forward(
     self: "LlamaFlashAttention2",
     hidden_states: torch.Tensor,
     attention_mask: Optional[torch.Tensor] = None,
     position_ids: Optional[torch.LongTensor] = None,
-    past_key_value: Optional[Tuple[torch.Tensor]] = None,
+    past_key_value: Optional["Cache"] = None,
     output_attentions: bool = False,
+    cache_position: Optional[torch.LongTensor] = None,
     **kwargs,
 ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
     # LlamaFlashAttention2 attention does not support output_attentions
@@ -120,15 +122,13 @@ def llama_flash_attn_forward(
     key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
     value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
 
-    kv_seq_len = key_states.shape[-2]
-    if past_key_value is not None:
-        kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+    cos, sin = self.rotary_emb(value_states, position_ids)
+    query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
 
-    cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
-    query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
+    past_key_value = getattr(self, "past_key_value", past_key_value)
 
     if past_key_value is not None:
-        cache_kwargs = {"sin": sin, "cos": cos}  # Specific to RoPE models
+        cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
         key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
 
     key_states = repeat_kv(key_states, self.num_key_value_groups)
@@ -193,5 +193,6 @@ def llama_flash_attn_forward(
 
 
 def apply_llama_patch() -> None:
+    require_version("transformers==4.39.1", "To fix: pip install transformers==4.39.1")
     LlamaAttention.forward = llama_torch_attn_forward
     LlamaFlashAttention2.forward = llama_flash_attn_forward

src/llmtuner/extras/patches/mixtral_patch.py

@@ -0,0 +1,38 @@
+import torch
+import torch.nn.functional as F
+from transformers.models.mixtral.modeling_mixtral import MixtralBLockSparseTop2MLP, MixtralSparseMoeBlock
+
+
+def mlp_forward(self: "MixtralBLockSparseTop2MLP", hidden_states: torch.Tensor) -> torch.Tensor:
+    current_hidden_states = self.act_fn(self.w1(hidden_states)) * self.w3(hidden_states)
+    current_hidden_states = self.w2(current_hidden_states)
+    return current_hidden_states
+
+
+# Modified from: https://huggingface.co/deepseek-ai/deepseek-moe-16b-base/blob/main/modeling_deepseek.py
+def moe_forward(self: "MixtralSparseMoeBlock", hidden_states: torch.Tensor) -> torch.Tensor:
+    batch_size, sequence_length, hidden_dim = hidden_states.shape
+    hidden_states = hidden_states.view(-1, hidden_dim)
+    # router_logits: (batch * sequence_length, n_experts)
+    router_logits = self.gate(hidden_states)
+
+    routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
+    topk_weight, topk_idx = torch.topk(routing_weights, self.top_k, dim=-1, sorted=False)
+    topk_weight /= topk_weight.sum(dim=-1, keepdim=True)
+    # we cast back to the input dtype
+    topk_weight = topk_weight.to(hidden_states.dtype)
+
+    hidden_states = hidden_states.repeat_interleave(self.top_k, dim=0)
+    y = torch.empty_like(hidden_states)
+    flat_topk_idx = topk_idx.view(-1)
+    for i in range(self.num_experts):
+        expert = self.experts[i]
+        y[flat_topk_idx == i] = expert(hidden_states[flat_topk_idx == i])
+    y = (y.view(*topk_weight.shape, -1) * topk_weight.unsqueeze(-1)).sum(dim=1)
+    final_hidden_states = y.reshape(batch_size, sequence_length, hidden_dim)
+    return final_hidden_states, router_logits
+
+
+def patch_mixtral_replace_moe_impl() -> None:
+    MixtralBLockSparseTop2MLP.forward = mlp_forward
+    MixtralSparseMoeBlock.forward = moe_forward

src/llmtuner/extras/ploting.py

@@ -1,7 +1,7 @@
 import json
 import math
 import os
-from typing import List, Optional
+from typing import List
 
 from transformers.trainer import TRAINER_STATE_NAME
 
@@ -30,7 +30,7 @@ def smooth(scalars: List[float]) -> List[float]:
     return smoothed
 
 
-def plot_loss(save_dictionary: os.PathLike, keys: Optional[List[str]] = ["loss"]) -> None:
+def plot_loss(save_dictionary: os.PathLike, keys: List[str] = ["loss"]) -> None:
     with open(os.path.join(save_dictionary, TRAINER_STATE_NAME), "r", encoding="utf-8") as f:
         data = json.load(f)
 
@@ -46,11 +46,12 @@ def plot_loss(save_dictionary: os.PathLike, keys: Optional[List[str]] = ["loss"]
             continue
 
         plt.figure()
-        plt.plot(steps, metrics, alpha=0.4, label="original")
-        plt.plot(steps, smooth(metrics), label="smoothed")
+        plt.plot(steps, metrics, color="#1f77b4", alpha=0.4, label="original")
+        plt.plot(steps, smooth(metrics), color="#1f77b4", label="smoothed")
         plt.title("training {} of {}".format(key, save_dictionary))
         plt.xlabel("step")
         plt.ylabel(key)
         plt.legend()
-        plt.savefig(os.path.join(save_dictionary, "training_{}.png".format(key)), format="png", dpi=100)
-        print("Figure saved:", os.path.join(save_dictionary, "training_{}.png".format(key)))
+        figure_path = os.path.join(save_dictionary, "training_{}.png".format(key.replace(os.path.sep, "_")))
+        plt.savefig(figure_path, format="png", dpi=100)
+        print("Figure saved at:", figure_path)

src/llmtuner/hparams/data_args.py

@@ -7,33 +7,44 @@ class DataArguments:
     r"""
     Arguments pertaining to what data we are going to input our model for training and evaluation.
     """
+
     template: Optional[str] = field(
-        default=None, metadata={"help": "Which template to use for constructing prompts in training and inference."}
+        default=None,
+        metadata={"help": "Which template to use for constructing prompts in training and inference."},
     )
     dataset: Optional[str] = field(
         default=None,
         metadata={"help": "The name of provided dataset(s) to use. Use commas to separate multiple datasets."},
     )
-    dataset_dir: Optional[str] = field(
-        default="data", metadata={"help": "Path to the folder containing the datasets."}
+    dataset_dir: str = field(
+        default="data",
+        metadata={"help": "Path to the folder containing the datasets."},
     )
-    split: Optional[str] = field(
-        default="train", metadata={"help": "Which dataset split to use for training and evaluation."}
+    split: str = field(
+        default="train",
+        metadata={"help": "Which dataset split to use for training and evaluation."},
     )
-    cutoff_len: Optional[int] = field(
-        default=1024, metadata={"help": "The maximum length of the model inputs after tokenization."}
+    cutoff_len: int = field(
+        default=1024,
+        metadata={"help": "The cutoff length of the model inputs after tokenization."},
     )
-    reserved_label_len: Optional[int] = field(
-        default=1, metadata={"help": "The maximum length reserved for label after tokenization."}
+    reserved_label_len: int = field(
+        default=1,
+        metadata={"help": "The minimum cutoff length reserved for label after tokenization."},
     )
-    train_on_prompt: Optional[bool] = field(
-        default=False, metadata={"help": "Whether to disable the mask on the prompt or not."}
+    train_on_prompt: bool = field(
+        default=False,
+        metadata={"help": "Whether to disable the mask on the prompt or not."},
     )
-    streaming: Optional[bool] = field(default=False, metadata={"help": "Enable dataset streaming."})
-    buffer_size: Optional[int] = field(
-        default=16384, metadata={"help": "Size of the buffer to randomly sample examples from in dataset streaming."}
+    streaming: bool = field(
+        default=False,
+        metadata={"help": "Enable dataset streaming."},
     )
-    mix_strategy: Optional[Literal["concat", "interleave_under", "interleave_over"]] = field(
+    buffer_size: int = field(
+        default=16384,
+        metadata={"help": "Size of the buffer to randomly sample examples from in dataset streaming."},
+    )
+    mix_strategy: Literal["concat", "interleave_under", "interleave_over"] = field(
         default="concat",
         metadata={"help": "Strategy to use in dataset mixing (concat/interleave) (undersampling/oversampling)."},
     )
@@ -41,33 +52,41 @@ class DataArguments:
         default=None,
         metadata={"help": "Probabilities to sample data from datasets. Use commas to separate multiple datasets."},
     )
-    overwrite_cache: Optional[bool] = field(
-        default=False, metadata={"help": "Overwrite the cached training and evaluation sets."}
+    overwrite_cache: bool = field(
+        default=False,
+        metadata={"help": "Overwrite the cached training and evaluation sets."},
     )
     preprocessing_num_workers: Optional[int] = field(
-        default=None, metadata={"help": "The number of processes to use for the preprocessing."}
+        default=None,
+        metadata={"help": "The number of processes to use for the pre-processing."},
     )
     max_samples: Optional[int] = field(
-        default=None, metadata={"help": "For debugging purposes, truncate the number of examples for each dataset."}
+        default=None,
+        metadata={"help": "For debugging purposes, truncate the number of examples for each dataset."},
     )
     eval_num_beams: Optional[int] = field(
         default=None,
         metadata={"help": "Number of beams to use for evaluation. This argument will be passed to `model.generate`"},
     )
-    ignore_pad_token_for_loss: Optional[bool] = field(
+    ignore_pad_token_for_loss: bool = field(
         default=True,
         metadata={
-            "help": "Whether to ignore the tokens corresponding to padded labels in the loss computation or not."
+            "help": "Whether or not to ignore the tokens corresponding to padded labels in the loss computation."
         },
     )
-    val_size: Optional[float] = field(
-        default=0, metadata={"help": "Size of the development set, should be an integer or a float in range `[0,1)`."}
+    val_size: float = field(
+        default=0.0,
+        metadata={"help": "Size of the development set, should be an integer or a float in range `[0,1)`."},
     )
-    sft_packing: Optional[bool] = field(
-        default=False, metadata={"help": "Packing the questions and answers in the supervised fine-tuning stage."}
+    packing: Optional[bool] = field(
+        default=None,
+        metadata={
+            "help": "Whether or not to pack the sequences in training. Will automatically enable in pre-training."
+        },
     )
     cache_path: Optional[str] = field(
-        default=None, metadata={"help": "Path to save or load the preprocessed datasets."}
+        default=None,
+        metadata={"help": "Path to save or load the pre-processed datasets."},
     )
 
     def __post_init__(self):

src/llmtuner/hparams/evaluation_args.py

@@ -10,16 +10,35 @@ class EvaluationArguments:
     r"""
     Arguments pertaining to specify the evaluation parameters.
     """
-    task: str = field(metadata={"help": "Name of the evaluation task."})
-    task_dir: Optional[str] = field(
-        default="evaluation", metadata={"help": "Path to the folder containing the evaluation datasets."}
+
+    task: str = field(
+        metadata={"help": "Name of the evaluation task."},
     )
-    batch_size: Optional[int] = field(default=4, metadata={"help": "The batch size per GPU for evaluation."})
-    seed: Optional[int] = field(default=42, metadata={"help": "Random seed to be used with data loaders."})
-    lang: Optional[Literal["en", "zh"]] = field(default="en", metadata={"help": "Language used at evaluation."})
-    n_shot: Optional[int] = field(default=5, metadata={"help": "Number of examplars for few-shot learning."})
-    save_dir: Optional[str] = field(default=None, metadata={"help": "Path to save the evaluation results."})
-    download_mode: Optional[DownloadMode] = field(
+    task_dir: str = field(
+        default="evaluation",
+        metadata={"help": "Path to the folder containing the evaluation datasets."},
+    )
+    batch_size: int = field(
+        default=4,
+        metadata={"help": "The batch size per GPU for evaluation."},
+    )
+    seed: int = field(
+        default=42,
+        metadata={"help": "Random seed to be used with data loaders."},
+    )
+    lang: Literal["en", "zh"] = field(
+        default="en",
+        metadata={"help": "Language used at evaluation."},
+    )
+    n_shot: int = field(
+        default=5,
+        metadata={"help": "Number of examplars for few-shot learning."},
+    )
+    save_dir: Optional[str] = field(
+        default=None,
+        metadata={"help": "Path to save the evaluation results."},
+    )
+    download_mode: DownloadMode = field(
         default=DownloadMode.REUSE_DATASET_IF_EXISTS,
         metadata={"help": "Download mode used for the evaluation datasets."},
     )

src/llmtuner/hparams/finetuning_args.py

@@ -8,20 +8,23 @@ class FreezeArguments:
     r"""
     Arguments pertaining to the freeze (partial-parameter) training.
     """
-    name_module_trainable: Optional[str] = field(
-        default="mlp",
+
+    name_module_trainable: str = field(
+        default="all",
         metadata={
-            "help": 'Name of trainable modules for partial-parameter (freeze) fine-tuning. \
-                  Use commas to separate multiple modules. \
-                  LLaMA choices: ["mlp", "self_attn"], \
-                  BLOOM & Falcon & ChatGLM choices: ["mlp", "self_attention"], \
-                  Qwen choices: ["mlp", "attn"], \
-                  Phi choices: ["mlp", "mixer"], \
-                  Others choices: the same as LLaMA.'
+            "help": """Name of trainable modules for partial-parameter (freeze) fine-tuning. \
+                    Use commas to separate multiple modules. \
+                    Use "all" to specify all the available modules. \
+                    LLaMA choices: ["mlp", "self_attn"], \
+                    BLOOM & Falcon & ChatGLM choices: ["mlp", "self_attention"], \
+                    Qwen choices: ["mlp", "attn"], \
+                    InternLM2 choices: ["feed_forward", "attention"], \
+                    Others choices: the same as LLaMA."""
         },
     )
-    num_layer_trainable: Optional[int] = field(
-        default=3, metadata={"help": "The number of trainable layers for partial-parameter (freeze) fine-tuning."}
+    num_layer_trainable: int = field(
+        default=2,
+        metadata={"help": "The number of trainable layers for partial-parameter (freeze) fine-tuning."},
     )
 
 
@@ -30,6 +33,7 @@ class LoraArguments:
     r"""
     Arguments pertaining to the LoRA training.
     """
+
     additional_target: Optional[str] = field(
         default=None,
         metadata={
@@ -37,27 +41,50 @@ class LoraArguments:
         },
     )
     lora_alpha: Optional[int] = field(
-        default=None, metadata={"help": "The scale factor for LoRA fine-tuning (default: lora_rank * 2)."}
-    )
-    lora_dropout: Optional[float] = field(default=0.0, metadata={"help": "Dropout rate for the LoRA fine-tuning."})
-    lora_rank: Optional[int] = field(default=8, metadata={"help": "The intrinsic dimension for LoRA fine-tuning."})
-    lora_target: Optional[str] = field(
         default=None,
+        metadata={"help": "The scale factor for LoRA fine-tuning (default: lora_rank * 2)."},
+    )
+    lora_dropout: float = field(
+        default=0.0,
+        metadata={"help": "Dropout rate for the LoRA fine-tuning."},
+    )
+    lora_rank: int = field(
+        default=8,
+        metadata={"help": "The intrinsic dimension for LoRA fine-tuning."},
+    )
+    lora_target: str = field(
+        default="all",
         metadata={
-            "help": 'Name(s) of target modules to apply LoRA. Use commas to separate multiple modules. \
-                  LLaMA choices: ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"], \
-                  BLOOM & Falcon & ChatGLM choices: ["query_key_value", "dense", "dense_h_to_4h", "dense_4h_to_h"], \
-                  Baichuan choices: ["W_pack", "o_proj", "gate_proj", "up_proj", "down_proj"], \
-                  Qwen choices: ["c_attn", "attn.c_proj", "w1", "w2", "mlp.c_proj"], \
-                  Phi choices: ["Wqkv", "out_proj", "fc1", "fc2"], \
-                  Others choices: the same as LLaMA.'
+            "help": """Name(s) of target modules to apply LoRA. \
+                    Use commas to separate multiple modules. \
+                    Use "all" to specify all the linear modules. \
+                    LLaMA choices: ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"], \
+                    BLOOM & Falcon & ChatGLM choices: ["query_key_value", "dense", "dense_h_to_4h", "dense_4h_to_h"], \
+                    Baichuan choices: ["W_pack", "o_proj", "gate_proj", "up_proj", "down_proj"], \
+                    Qwen choices: ["c_attn", "attn.c_proj", "w1", "w2", "mlp.c_proj"], \
+                    InternLM2 choices: ["wqkv", "wo", "w1", "w2", "w3"], \
+                    Others choices: the same as LLaMA."""
         },
     )
-    lora_bf16_mode: Optional[bool] = field(
-        default=False, metadata={"help": "Whether or not to train lora adapters in bf16 precision."}
+    loraplus_lr_ratio: Optional[float] = field(
+        default=None,
+        metadata={"help": "LoRA plus learning rate ratio (lr_B / lr_A)."},
     )
-    create_new_adapter: Optional[bool] = field(
-        default=False, metadata={"help": "Whether or not to create a new adapter with randomly initialized weight."}
+    loraplus_lr_embedding: float = field(
+        default=1e-6,
+        metadata={"help": "LoRA plus learning rate for lora embedding layers."},
+    )
+    use_rslora: bool = field(
+        default=False,
+        metadata={"help": "Whether or not to use the rank stabilization scaling factor for LoRA layer."},
+    )
+    use_dora: bool = field(
+        default=False,
+        metadata={"help": "Whether or not to use the weight-decomposed lora method (DoRA)."},
+    )
+    create_new_adapter: bool = field(
+        default=False,
+        metadata={"help": "Whether or not to create a new adapter with randomly initialized weight."},
     )
 
 
@@ -66,69 +93,137 @@ class RLHFArguments:
     r"""
     Arguments pertaining to the PPO and DPO training.
     """
-    dpo_beta: Optional[float] = field(default=0.1, metadata={"help": "The beta parameter for the DPO loss."})
-    dpo_loss: Optional[Literal["sigmoid", "hinge", "ipo", "kto"]] = field(
-        default="sigmoid", metadata={"help": "The type of DPO loss to use."}
+
+    dpo_beta: float = field(
+        default=0.1,
+        metadata={"help": "The beta parameter for the DPO loss."},
     )
-    dpo_ftx: Optional[float] = field(
-        default=0, metadata={"help": "The supervised fine-tuning loss coefficient in DPO training."}
+    dpo_loss: Literal["sigmoid", "hinge", "ipo", "kto_pair"] = field(
+        default="sigmoid",
+        metadata={"help": "The type of DPO loss to use."},
     )
-    ppo_buffer_size: Optional[int] = field(
+    dpo_ftx: float = field(
+        default=0.0,
+        metadata={"help": "The supervised fine-tuning loss coefficient in DPO training."},
+    )
+    ppo_buffer_size: int = field(
         default=1,
         metadata={"help": "The number of mini-batches to make experience buffer in a PPO optimization step."},
     )
-    ppo_epochs: Optional[int] = field(
-        default=4, metadata={"help": "The number of epochs to perform in a PPO optimization step."}
+    ppo_epochs: int = field(
+        default=4,
+        metadata={"help": "The number of epochs to perform in a PPO optimization step."},
     )
     ppo_logger: Optional[str] = field(
-        default=None, metadata={"help": 'Log with either "wandb" or "tensorboard" in PPO training.'}
+        default=None,
+        metadata={"help": 'Log with either "wandb" or "tensorboard" in PPO training.'},
     )
-    ppo_score_norm: Optional[bool] = field(
-        default=False, metadata={"help": "Use score normalization in PPO training."}
+    ppo_score_norm: bool = field(
+        default=False,
+        metadata={"help": "Use score normalization in PPO training."},
     )
-    ppo_target: Optional[float] = field(
-        default=6.0, metadata={"help": "Target KL value for adaptive KL control in PPO training."}
+    ppo_target: float = field(
+        default=6.0,
+        metadata={"help": "Target KL value for adaptive KL control in PPO training."},
     )
-    ppo_whiten_rewards: Optional[bool] = field(
-        default=False, metadata={"help": "Whiten the rewards before compute advantages in PPO training."}
+    ppo_whiten_rewards: bool = field(
+        default=False,
+        metadata={"help": "Whiten the rewards before compute advantages in PPO training."},
     )
     ref_model: Optional[str] = field(
-        default=None, metadata={"help": "Path to the reference model used for the PPO or DPO training."}
+        default=None,
+        metadata={"help": "Path to the reference model used for the PPO or DPO training."},
     )
     ref_model_adapters: Optional[str] = field(
-        default=None, metadata={"help": "Path to the adapters of the reference model."}
+        default=None,
+        metadata={"help": "Path to the adapters of the reference model."},
     )
     ref_model_quantization_bit: Optional[int] = field(
-        default=None, metadata={"help": "The number of bits to quantize the reference model."}
+        default=None,
+        metadata={"help": "The number of bits to quantize the reference model."},
     )
     reward_model: Optional[str] = field(
-        default=None, metadata={"help": "Path to the reward model used for the PPO training."}
+        default=None,
+        metadata={"help": "Path to the reward model used for the PPO training."},
     )
     reward_model_adapters: Optional[str] = field(
-        default=None, metadata={"help": "Path to the adapters of the reward model."}
+        default=None,
+        metadata={"help": "Path to the adapters of the reward model."},
     )
     reward_model_quantization_bit: Optional[int] = field(
-        default=None, metadata={"help": "The number of bits to quantize the reward model."}
+        default=None,
+        metadata={"help": "The number of bits to quantize the reward model."},
     )
-    reward_model_type: Optional[Literal["lora", "full", "api"]] = field(
+    reward_model_type: Literal["lora", "full", "api"] = field(
         default="lora",
         metadata={"help": "The type of the reward model in PPO training. Lora model only supports lora training."},
     )
 
 
 @dataclass
-class FinetuningArguments(FreezeArguments, LoraArguments, RLHFArguments):
+class GaloreArguments:
+    r"""
+    Arguments pertaining to the GaLore algorithm.
+    """
+
+    use_galore: bool = field(
+        default=False,
+        metadata={"help": "Whether or not to use gradient low-Rank projection."},
+    )
+    galore_target: str = field(
+        default="all",
+        metadata={
+            "help": """Name(s) of modules to apply GaLore. Use commas to separate multiple modules. \
+                    Use "all" to specify all the linear modules."""
+        },
+    )
+    galore_rank: int = field(
+        default=16,
+        metadata={"help": "The rank of GaLore gradients."},
+    )
+    galore_update_interval: int = field(
+        default=200,
+        metadata={"help": "Number of steps to update the GaLore projection."},
+    )
+    galore_scale: float = field(
+        default=0.25,
+        metadata={"help": "GaLore scaling coefficient."},
+    )
+    galore_proj_type: Literal["std", "reverse_std", "right", "left", "full"] = field(
+        default="std",
+        metadata={"help": "Type of GaLore projection."},
+    )
+    galore_layerwise: bool = field(
+        default=False,
+        metadata={"help": "Whether or not to enable layer-wise update to further save memory."},
+    )
+
+
+@dataclass
+class FinetuningArguments(FreezeArguments, LoraArguments, RLHFArguments, GaloreArguments):
     r"""
     Arguments pertaining to which techniques we are going to fine-tuning with.
     """
-    stage: Optional[Literal["pt", "sft", "rm", "ppo", "dpo"]] = field(
-        default="sft", metadata={"help": "Which stage will be performed in training."}
+
+    pure_bf16: bool = field(
+        default=False,
+        metadata={"help": "Whether or not to train model in purely bf16 precision (without AMP)."},
     )
-    finetuning_type: Optional[Literal["lora", "freeze", "full"]] = field(
-        default="lora", metadata={"help": "Which fine-tuning method to use."}
+    stage: Literal["pt", "sft", "rm", "ppo", "dpo"] = field(
+        default="sft",
+        metadata={"help": "Which stage will be performed in training."},
     )
-    plot_loss: Optional[bool] = field(
-        default=False, metadata={"help": "Whether or not to save the training loss curves."}
+    finetuning_type: Literal["lora", "freeze", "full"] = field(
+        default="lora",
+        metadata={"help": "Which fine-tuning method to use."},
+    )
+    use_llama_pro: bool = field(
+        default=False,
+        metadata={"help": "Whether or not to make only the parameters in the expanded blocks trainable."},
+    )
+    plot_loss: bool = field(
+        default=False,
+        metadata={"help": "Whether or not to save the training loss curves."},
     )
 
     def __post_init__(self):
@@ -141,16 +236,23 @@ class FinetuningArguments(FreezeArguments, LoraArguments, RLHFArguments):
         self.lora_alpha = self.lora_alpha or self.lora_rank * 2
         self.lora_target = split_arg(self.lora_target)
         self.additional_target = split_arg(self.additional_target)
+        self.galore_target = split_arg(self.galore_target)
 
         assert self.finetuning_type in ["lora", "freeze", "full"], "Invalid fine-tuning method."
         assert self.ref_model_quantization_bit in [None, 8, 4], "We only accept 4-bit or 8-bit quantization."
         assert self.reward_model_quantization_bit in [None, 8, 4], "We only accept 4-bit or 8-bit quantization."
 
         if self.stage == "ppo" and self.reward_model is None:
-            raise ValueError("Reward model is necessary for PPO training.")
+            raise ValueError("`reward_model` is necessary for PPO training.")
 
         if self.stage == "ppo" and self.reward_model_type == "lora" and self.finetuning_type != "lora":
-            raise ValueError("Freeze/Full PPO training needs `reward_model_type=full`.")
+            raise ValueError("`reward_model_type` cannot be lora for Freeze/Full PPO training.")
+
+        if self.use_llama_pro and self.finetuning_type == "full":
+            raise ValueError("`use_llama_pro` is only valid for the Freeze or LoRA method.")
+
+        if self.use_galore and self.finetuning_type == "lora":
+            raise ValueError("Cannot use LoRA with GaLore together.")
 
     def save_to_json(self, json_path: str):
         r"""Saves the content of this instance in JSON format inside `json_path`."""

src/llmtuner/hparams/generating_args.py

@@ -1,5 +1,5 @@
 from dataclasses import asdict, dataclass, field
-from typing import Any, Dict, Optional
+from typing import Any, Dict
 
 
 @dataclass
@@ -7,38 +7,44 @@ class GeneratingArguments:
     r"""
     Arguments pertaining to specify the decoding parameters.
     """
-    do_sample: Optional[bool] = field(
-        default=True, metadata={"help": "Whether or not to use sampling, use greedy decoding otherwise."}
+
+    do_sample: bool = field(
+        default=True,
+        metadata={"help": "Whether or not to use sampling, use greedy decoding otherwise."},
     )
-    temperature: Optional[float] = field(
-        default=0.95, metadata={"help": "The value used to modulate the next token probabilities."}
+    temperature: float = field(
+        default=0.95,
+        metadata={"help": "The value used to modulate the next token probabilities."},
     )
-    top_p: Optional[float] = field(
+    top_p: float = field(
         default=0.7,
         metadata={
             "help": "The smallest set of most probable tokens with probabilities that add up to top_p or higher are kept."
         },
     )
-    top_k: Optional[int] = field(
+    top_k: int = field(
         default=50,
         metadata={"help": "The number of highest probability vocabulary tokens to keep for top-k filtering."},
     )
-    num_beams: Optional[int] = field(
-        default=1, metadata={"help": "Number of beams for beam search. 1 means no beam search."}
+    num_beams: int = field(
+        default=1,
+        metadata={"help": "Number of beams for beam search. 1 means no beam search."},
     )
-    max_length: Optional[int] = field(
+    max_length: int = field(
         default=512,
         metadata={"help": "The maximum length the generated tokens can have. It can be overridden by max_new_tokens."},
     )
-    max_new_tokens: Optional[int] = field(
+    max_new_tokens: int = field(
         default=512,
         metadata={"help": "The maximum numbers of tokens to generate, ignoring the number of tokens in the prompt."},
     )
-    repetition_penalty: Optional[float] = field(
-        default=1.0, metadata={"help": "The parameter for repetition penalty. 1.0 means no penalty."}
+    repetition_penalty: float = field(
+        default=1.0,
+        metadata={"help": "The parameter for repetition penalty. 1.0 means no penalty."},
     )
-    length_penalty: Optional[float] = field(
-        default=1.0, metadata={"help": "Exponential penalty to the length that is used with beam-based generation."}
+    length_penalty: float = field(
+        default=1.0,
+        metadata={"help": "Exponential penalty to the length that is used with beam-based generation."},
     )
 
     def to_dict(self) -> Dict[str, Any]:

src/llmtuner/hparams/model_args.py

@@ -5,92 +5,154 @@ from typing import Any, Dict, Literal, Optional
 @dataclass
 class ModelArguments:
     r"""
-    Arguments pertaining to which model/config/tokenizer we are going to fine-tune.
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune or infer.
     """
+
     model_name_or_path: str = field(
-        metadata={"help": "Path to the model weight or identifier from huggingface.co/models or modelscope.cn/models."}
+        metadata={
+            "help": "Path to the model weight or identifier from huggingface.co/models or modelscope.cn/models."
+        },
     )
     adapter_name_or_path: Optional[str] = field(
-        default=None, metadata={"help": "Path to the adapter weight or identifier from huggingface.co/models."}
+        default=None,
+        metadata={"help": "Path to the adapter weight or identifier from huggingface.co/models."},
     )
     cache_dir: Optional[str] = field(
         default=None,
         metadata={"help": "Where to store the pre-trained models downloaded from huggingface.co or modelscope.cn."},
     )
-    use_fast_tokenizer: Optional[bool] = field(
+    use_fast_tokenizer: bool = field(
         default=False,
         metadata={"help": "Whether or not to use one of the fast tokenizer (backed by the tokenizers library)."},
     )
-    resize_vocab: Optional[bool] = field(
-        default=False, metadata={"help": "Whether or not to resize the tokenizer vocab and the embedding layers."}
+    resize_vocab: bool = field(
+        default=False,
+        metadata={"help": "Whether or not to resize the tokenizer vocab and the embedding layers."},
     )
-    split_special_tokens: Optional[bool] = field(
+    split_special_tokens: bool = field(
         default=False,
         metadata={"help": "Whether or not the special tokens should be split during the tokenization process."},
     )
-    model_revision: Optional[str] = field(
+    model_revision: str = field(
         default="main",
         metadata={"help": "The specific model version to use (can be a branch name, tag name or commit id)."},
     )
+    low_cpu_mem_usage: bool = field(
+        default=True,
+        metadata={"help": "Whether or not to use memory-efficient model loading."},
+    )
     quantization_bit: Optional[int] = field(
-        default=None, metadata={"help": "The number of bits to quantize the model."}
+        default=None,
+        metadata={"help": "The number of bits to quantize the model using bitsandbytes."},
     )
-    quantization_type: Optional[Literal["fp4", "nf4"]] = field(
-        default="nf4", metadata={"help": "Quantization data type to use in int4 training."}
+    quantization_type: Literal["fp4", "nf4"] = field(
+        default="nf4",
+        metadata={"help": "Quantization data type to use in int4 training."},
     )
-    double_quantization: Optional[bool] = field(
-        default=True, metadata={"help": "Whether or not to use double quantization in int4 training."}
+    double_quantization: bool = field(
+        default=True,
+        metadata={"help": "Whether or not to use double quantization in int4 training."},
     )
     rope_scaling: Optional[Literal["linear", "dynamic"]] = field(
-        default=None, metadata={"help": "Which scaling strategy should be adopted for the RoPE embeddings."}
+        default=None,
+        metadata={"help": "Which scaling strategy should be adopted for the RoPE embeddings."},
     )
-    flash_attn: Optional[bool] = field(
-        default=False, metadata={"help": "Enable FlashAttention-2 for faster training."}
+    flash_attn: bool = field(
+        default=False,
+        metadata={"help": "Enable FlashAttention-2 for faster training."},
     )
-    shift_attn: Optional[bool] = field(
-        default=False, metadata={"help": "Enable shift short attention (S^2-Attn) proposed by LongLoRA."}
+    shift_attn: bool = field(
+        default=False,
+        metadata={"help": "Enable shift short attention (S^2-Attn) proposed by LongLoRA."},
     )
-    use_unsloth: Optional[bool] = field(
-        default=False, metadata={"help": "Whether or not to use unsloth's optimization for the LoRA training."}
+    use_unsloth: bool = field(
+        default=False,
+        metadata={"help": "Whether or not to use unsloth's optimization for the LoRA training."},
     )
-    disable_gradient_checkpointing: Optional[bool] = field(
-        default=False, metadata={"help": "Whether or not to disable gradient checkpointing."}
+    disable_gradient_checkpointing: bool = field(
+        default=False,
+        metadata={"help": "Whether or not to disable gradient checkpointing."},
     )
-    upcast_layernorm: Optional[bool] = field(
-        default=False, metadata={"help": "Whether or not to upcast the layernorm weights in fp32."}
+    upcast_layernorm: bool = field(
+        default=False,
+        metadata={"help": "Whether or not to upcast the layernorm weights in fp32."},
     )
-    upcast_lmhead_output: Optional[bool] = field(
-        default=False, metadata={"help": "Whether or not to upcast the output of lm_head in fp32."}
+    upcast_lmhead_output: bool = field(
+        default=False,
+        metadata={"help": "Whether or not to upcast the output of lm_head in fp32."},
+    )
+    infer_backend: Literal["huggingface", "vllm"] = field(
+        default="huggingface",
+        metadata={"help": "Backend engine used at inference."},
+    )
+    vllm_maxlen: int = field(
+        default=2048,
+        metadata={"help": "Maximum input length of the vLLM engine."},
+    )
+    vllm_gpu_util: float = field(
+        default=0.9,
+        metadata={"help": "The fraction of GPU memory in (0,1) to be used for the vLLM engine."},
+    )
+    vllm_enforce_eager: bool = field(
+        default=False,
+        metadata={"help": "Whether or not to disable CUDA graph in the vLLM engine."},
+    )
+    offload_folder: str = field(
+        default="offload",
+        metadata={"help": "Path to offload model weights."},
+    )
+    use_cache: bool = field(
+        default=True,
+        metadata={"help": "Whether or not to use KV cache in generation."},
+    )
+    hf_hub_token: Optional[str] = field(
+        default=None,
+        metadata={"help": "Auth token to log in with Hugging Face Hub."},
+    )
+    ms_hub_token: Optional[str] = field(
+        default=None,
+        metadata={"help": "Auth token to log in with ModelScope Hub."},
     )
-    hf_hub_token: Optional[str] = field(default=None, metadata={"help": "Auth token to log in with Hugging Face Hub."})
-    ms_hub_token: Optional[str] = field(default=None, metadata={"help": "Auth token to log in with ModelScope Hub."})
     export_dir: Optional[str] = field(
-        default=None, metadata={"help": "Path to the directory to save the exported model."}
+        default=None,
+        metadata={"help": "Path to the directory to save the exported model."},
     )
-    export_size: Optional[int] = field(
-        default=1, metadata={"help": "The file shard size (in GB) of the exported model."}
+    export_size: int = field(
+        default=1,
+        metadata={"help": "The file shard size (in GB) of the exported model."},
     )
     export_quantization_bit: Optional[int] = field(
-        default=None, metadata={"help": "The number of bits to quantize the exported model."}
+        default=None,
+        metadata={"help": "The number of bits to quantize the exported model."},
     )
     export_quantization_dataset: Optional[str] = field(
-        default=None, metadata={"help": "Path to the dataset or dataset name to use in quantizing the exported model."}
+        default=None,
+        metadata={"help": "Path to the dataset or dataset name to use in quantizing the exported model."},
     )
-    export_quantization_nsamples: Optional[int] = field(
-        default=128, metadata={"help": "The number of samples used for quantization."}
+    export_quantization_nsamples: int = field(
+        default=128,
+        metadata={"help": "The number of samples used for quantization."},
     )
-    export_quantization_maxlen: Optional[int] = field(
-        default=1024, metadata={"help": "The maximum length of the model inputs used for quantization."}
+    export_quantization_maxlen: int = field(
+        default=1024,
+        metadata={"help": "The maximum length of the model inputs used for quantization."},
     )
-    export_legacy_format: Optional[bool] = field(
-        default=False, metadata={"help": "Whether or not to save the `.bin` files instead of `.safetensors`."}
+    export_legacy_format: bool = field(
+        default=False,
+        metadata={"help": "Whether or not to save the `.bin` files instead of `.safetensors`."},
     )
     export_hub_model_id: Optional[str] = field(
-        default=None, metadata={"help": "The name of the repository if push the model to the Hugging Face hub."}
+        default=None,
+        metadata={"help": "The name of the repository if push the model to the Hugging Face hub."},
+    )
+    print_param_status: bool = field(
+        default=False,
+        metadata={"help": "For debugging purposes, print the status of the parameters in the model."},
     )
 
     def __post_init__(self):
         self.compute_dtype = None
+        self.device_map = None
         self.model_max_length = None
 
         if self.split_special_tokens and self.use_fast_tokenizer:

src/llmtuner/hparams/parser.py

@@ -3,13 +3,16 @@ import os
 import sys
 from typing import Any, Dict, Optional, Tuple
 
-import datasets
 import torch
 import transformers
 from transformers import HfArgumentParser, Seq2SeqTrainingArguments
 from transformers.trainer_utils import get_last_checkpoint
+from transformers.utils import is_torch_bf16_gpu_available
+from transformers.utils.versions import require_version
 
 from ..extras.logging import get_logger
+from ..extras.misc import check_dependencies
+from ..extras.packages import is_unsloth_available
 from .data_args import DataArguments
 from .evaluation_args import EvaluationArguments
 from .finetuning_args import FinetuningArguments
@@ -20,6 +23,9 @@ from .model_args import ModelArguments
 logger = get_logger(__name__)
 
 
+check_dependencies()
+
+
 _TRAIN_ARGS = [ModelArguments, DataArguments, Seq2SeqTrainingArguments, FinetuningArguments, GeneratingArguments]
 _TRAIN_CLS = Tuple[ModelArguments, DataArguments, Seq2SeqTrainingArguments, FinetuningArguments, GeneratingArguments]
 _INFER_ARGS = [ModelArguments, DataArguments, FinetuningArguments, GeneratingArguments]
@@ -49,25 +55,23 @@ def _parse_args(parser: "HfArgumentParser", args: Optional[Dict[str, Any]] = Non
 
 
 def _set_transformers_logging(log_level: Optional[int] = logging.INFO) -> None:
-    datasets.utils.logging.set_verbosity(log_level)
     transformers.utils.logging.set_verbosity(log_level)
     transformers.utils.logging.enable_default_handler()
     transformers.utils.logging.enable_explicit_format()
 
 
 def _verify_model_args(model_args: "ModelArguments", finetuning_args: "FinetuningArguments") -> None:
+    if model_args.adapter_name_or_path is not None and finetuning_args.finetuning_type != "lora":
+        raise ValueError("Adapter is only valid for the LoRA method.")
+
     if model_args.quantization_bit is not None:
         if finetuning_args.finetuning_type != "lora":
             raise ValueError("Quantization is only compatible with the LoRA method.")
 
-        if finetuning_args.create_new_adapter:
+        if model_args.adapter_name_or_path is not None and finetuning_args.create_new_adapter:
             raise ValueError("Cannot create new adapter upon a quantized model.")
 
-    if model_args.adapter_name_or_path is not None and len(model_args.adapter_name_or_path) != 1:
-        if finetuning_args.finetuning_type != "lora":
-            raise ValueError("Multiple adapters are only available for LoRA tuning.")
-
-        if model_args.quantization_bit is not None:
+        if model_args.adapter_name_or_path is not None and len(model_args.adapter_name_or_path) != 1:
             raise ValueError("Quantized model only accepts a single adapter. Merge them first.")
 
 
@@ -121,37 +125,75 @@ def get_train_args(args: Optional[Dict[str, Any]] = None) -> _TRAIN_CLS:
     if training_args.do_train and training_args.predict_with_generate:
         raise ValueError("`predict_with_generate` cannot be set as True while training.")
 
-    if training_args.do_train and finetuning_args.finetuning_type == "lora" and finetuning_args.lora_target is None:
-        raise ValueError("Please specify `lora_target` in LoRA training.")
+    if training_args.do_train and model_args.use_unsloth and not is_unsloth_available():
+        raise ValueError("Unsloth was not installed: https://github.com/unslothai/unsloth")
+
+    if finetuning_args.use_dora:
+        if model_args.quantization_bit is not None:
+            require_version("peft>=0.10.0", "To fix: pip install peft>=0.10.0")
+
+        if model_args.use_unsloth:
+            raise ValueError("Unsloth does not support DoRA.")
+
+    if finetuning_args.pure_bf16:
+        if not is_torch_bf16_gpu_available():
+            raise ValueError("This device does not support `pure_bf16`.")
+
+        if training_args.fp16 or training_args.bf16:
+            raise ValueError("Turn off mixed precision training when using `pure_bf16`.")
+
+    if (
+        finetuning_args.use_galore
+        and finetuning_args.galore_layerwise
+        and training_args.parallel_mode.value == "distributed"
+    ):
+        raise ValueError("Distributed training does not support layer-wise GaLore.")
+
+    if finetuning_args.use_galore and training_args.deepspeed is not None:
+        raise ValueError("GaLore is incompatible with DeepSpeed.")
+
+    if model_args.infer_backend == "vllm":
+        raise ValueError("vLLM backend is only available for API, CLI and Web.")
 
     _verify_model_args(model_args, finetuning_args)
 
+    if (
+        training_args.do_train
+        and finetuning_args.finetuning_type == "lora"
+        and model_args.resize_vocab
+        and finetuning_args.additional_target is None
+    ):
+        logger.warning("Add token embeddings to `additional_target` to make the added tokens trainable.")
+
     if training_args.do_train and model_args.quantization_bit is not None and (not model_args.upcast_layernorm):
         logger.warning("We recommend enable `upcast_layernorm` in quantized training.")
 
     if training_args.do_train and (not training_args.fp16) and (not training_args.bf16):
         logger.warning("We recommend enable mixed precision training.")
 
+    if training_args.do_train and finetuning_args.use_galore and not finetuning_args.pure_bf16:
+        logger.warning("Using GaLore with mixed precision training may significantly increases GPU memory usage.")
+
     if (not training_args.do_train) and model_args.quantization_bit is not None:
         logger.warning("Evaluating model in 4/8-bit mode may cause lower scores.")
 
     if (not training_args.do_train) and finetuning_args.stage == "dpo" and finetuning_args.ref_model is None:
         logger.warning("Specify `ref_model` for computing rewards at evaluation.")
 
-    # postprocess training_args
+    # Post-process training arguments
     if (
-        training_args.local_rank != -1
+        training_args.parallel_mode.value == "distributed"
         and training_args.ddp_find_unused_parameters is None
         and finetuning_args.finetuning_type == "lora"
     ):
         logger.warning("`ddp_find_unused_parameters` needs to be set as False for LoRA in DDP training.")
-        training_args_dict = training_args.to_dict()
-        training_args_dict.update(dict(ddp_find_unused_parameters=False))
-        training_args = Seq2SeqTrainingArguments(**training_args_dict)
+        training_args.ddp_find_unused_parameters = False
 
     if finetuning_args.stage in ["rm", "ppo"] and finetuning_args.finetuning_type in ["full", "freeze"]:
         can_resume_from_checkpoint = False
-        training_args.resume_from_checkpoint = None
+        if training_args.resume_from_checkpoint is not None:
+            logger.warning("Cannot resume from checkpoint in current stage.")
+            training_args.resume_from_checkpoint = None
     else:
         can_resume_from_checkpoint = True
 
@@ -167,9 +209,7 @@ def get_train_args(args: Optional[Dict[str, Any]] = None) -> _TRAIN_CLS:
             raise ValueError("Output directory already exists and is not empty. Please set `overwrite_output_dir`.")
 
         if last_checkpoint is not None:
-            training_args_dict = training_args.to_dict()
-            training_args_dict.update(dict(resume_from_checkpoint=last_checkpoint))
-            training_args = Seq2SeqTrainingArguments(**training_args_dict)
+            training_args.resume_from_checkpoint = last_checkpoint
             logger.info(
                 "Resuming training from {}. Change `output_dir` or use `overwrite_output_dir` to avoid.".format(
                     training_args.resume_from_checkpoint
@@ -187,25 +227,26 @@ def get_train_args(args: Optional[Dict[str, Any]] = None) -> _TRAIN_CLS:
             )
         )
 
-    # postprocess model_args
-    model_args.compute_dtype = (
-        torch.bfloat16 if training_args.bf16 else (torch.float16 if training_args.fp16 else None)
-    )
+    # Post-process model arguments
+    if training_args.bf16 or finetuning_args.pure_bf16:
+        model_args.compute_dtype = torch.bfloat16
+    elif training_args.fp16:
+        model_args.compute_dtype = torch.float16
+
     model_args.model_max_length = data_args.cutoff_len
+    data_args.packing = data_args.packing if data_args.packing is not None else finetuning_args.stage == "pt"
 
     # Log on each process the small summary:
     logger.info(
-        "Process rank: {}, device: {}, n_gpu: {}\n  distributed training: {}, compute dtype: {}".format(
+        "Process rank: {}, device: {}, n_gpu: {}, distributed training: {}, compute dtype: {}".format(
             training_args.local_rank,
             training_args.device,
             training_args.n_gpu,
-            bool(training_args.local_rank != -1),
+            training_args.parallel_mode.value == "distributed",
             str(model_args.compute_dtype),
         )
     )
-    logger.info(f"Training/evaluation parameters {training_args}")
 
-    # Set seed before initializing model.
     transformers.set_seed(training_args.seed)
 
     return model_args, data_args, training_args, finetuning_args, generating_args
@@ -213,25 +254,47 @@ def get_train_args(args: Optional[Dict[str, Any]] = None) -> _TRAIN_CLS:
 
 def get_infer_args(args: Optional[Dict[str, Any]] = None) -> _INFER_CLS:
     model_args, data_args, finetuning_args, generating_args = _parse_infer_args(args)
+
     _set_transformers_logging()
 
     if data_args.template is None:
         raise ValueError("Please specify which `template` to use.")
 
+    if model_args.infer_backend == "vllm":
+        if finetuning_args.stage != "sft":
+            raise ValueError("vLLM engine only supports auto-regressive models.")
+
+        if model_args.adapter_name_or_path is not None:
+            raise ValueError("vLLM engine does not support LoRA adapters. Merge them first.")
+
+        if model_args.quantization_bit is not None:
+            raise ValueError("vLLM engine does not support quantization.")
+
+        if model_args.rope_scaling is not None:
+            raise ValueError("vLLM engine does not support RoPE scaling.")
+
     _verify_model_args(model_args, finetuning_args)
 
+    model_args.device_map = "auto"
+
     return model_args, data_args, finetuning_args, generating_args
 
 
 def get_eval_args(args: Optional[Dict[str, Any]] = None) -> _EVAL_CLS:
     model_args, data_args, eval_args, finetuning_args = _parse_eval_args(args)
+
     _set_transformers_logging()
 
     if data_args.template is None:
         raise ValueError("Please specify which `template` to use.")
 
+    if model_args.infer_backend == "vllm":
+        raise ValueError("vLLM backend is only available for API, CLI and Web.")
+
     _verify_model_args(model_args, finetuning_args)
 
+    model_args.device_map = "auto"
+
     transformers.set_seed(eval_args.seed)
 
     return model_args, data_args, eval_args, finetuning_args

src/llmtuner/model/__init__.py

@@ -1,5 +1,11 @@
-from .loader import load_model_and_tokenizer
-from .utils import dispatch_model, get_modelcard_args, load_valuehead_params
+from .loader import load_model, load_model_and_tokenizer, load_tokenizer
+from .utils import find_all_linear_modules, load_valuehead_params
 
 
-__all__ = ["load_model_and_tokenizer", "dispatch_model", "get_modelcard_args", "load_valuehead_params"]
+__all__ = [
+    "load_model",
+    "load_model_and_tokenizer",
+    "load_tokenizer",
+    "load_valuehead_params",
+    "find_all_linear_modules",
+]

src/llmtuner/model/adapter.py

@@ -1,12 +1,11 @@
-import inspect
 from typing import TYPE_CHECKING
 
 import torch
-from peft import LoraConfig, PeftModel, TaskType, get_peft_model
+from peft import LoraConfig, LoraModel, PeftModel, TaskType, get_peft_model
 from transformers.integrations import is_deepspeed_zero3_enabled
 
 from ..extras.logging import get_logger
-from .utils import find_all_linear_modules
+from .utils import QuantizationMethod, find_all_linear_modules, find_expanded_modules
 
 
 if TYPE_CHECKING:
@@ -35,7 +34,8 @@ def init_adapter(
 
     if finetuning_args.finetuning_type == "full" and is_trainable:
         logger.info("Fine-tuning method: Full")
-        model = model.float()
+        if not finetuning_args.pure_bf16:
+            model = model.float()
 
     if finetuning_args.finetuning_type == "freeze" and is_trainable:
         logger.info("Fine-tuning method: Freeze")
@@ -47,24 +47,47 @@ def init_adapter(
         if not num_layers:
             raise ValueError("Current model does not support freeze tuning.")
 
-        if finetuning_args.num_layer_trainable > 0:  # fine-tuning the last n layers if num_layer_trainable > 0
-            trainable_layer_ids = [num_layers - k - 1 for k in range(finetuning_args.num_layer_trainable)]
+        if finetuning_args.use_llama_pro:
+            if num_layers % finetuning_args.num_layer_trainable != 0:
+                raise ValueError(
+                    "`num_layers` {} should be divisible by `num_layer_trainable` {}.".format(
+                        num_layers, finetuning_args.num_layer_trainable
+                    )
+                )
+
+            stride = num_layers // finetuning_args.num_layer_trainable
+            trainable_layer_ids = range(stride - 1, num_layers + stride - 1, stride)
+        elif finetuning_args.num_layer_trainable > 0:  # fine-tuning the last n layers if num_layer_trainable > 0
+            trainable_layer_ids = range(num_layers - finetuning_args.num_layer_trainable, num_layers)
         else:  # fine-tuning the first n layers if num_layer_trainable < 0
-            trainable_layer_ids = [k for k in range(-finetuning_args.num_layer_trainable)]  # noqa: C416
+            trainable_layer_ids = range(-finetuning_args.num_layer_trainable)
+
+        freeze_modules = {"all"}
+        for name, _ in model.named_modules():
+            if ".0." in name:
+                freeze_modules.add(name.split(".0.")[-1].split(".")[0])
 
         trainable_layers = []
         for module_name in finetuning_args.name_module_trainable:
+            if module_name not in freeze_modules:
+                raise ValueError(
+                    "Module {} is not found, please choose from {}".format(module_name, ", ".join(freeze_modules))
+                )
+
             for idx in trainable_layer_ids:
-                trainable_layers.append("{:d}.{}".format(idx, module_name))
+                trainable_layers.append(".{:d}.{}".format(idx, module_name if module_name != "all" else ""))
 
         for name, param in model.named_parameters():
-            if not any(trainable_layer in name for trainable_layer in trainable_layers):
-                param.requires_grad_(False)
+            if any(trainable_layer in name for trainable_layer in trainable_layers):
+                if not finetuning_args.pure_bf16:
+                    param.data = param.data.to(torch.float32)
             else:
-                param.data = param.data.to(torch.float32)
+                param.requires_grad_(False)
+
+        logger.info("Set trainable layers: {}".format(",".join(map(str, trainable_layer_ids))))
 
     if finetuning_args.finetuning_type == "lora":
-        logger.info("Fine-tuning method: LoRA")
+        logger.info("Fine-tuning method: {}".format("DoRA" if finetuning_args.use_dora else "LoRA"))
         adapter_to_resume = None
 
         if model_args.adapter_name_or_path is not None:
@@ -84,14 +107,18 @@ def init_adapter(
                 adapter_to_merge = model_args.adapter_name_or_path
 
             for adapter in adapter_to_merge:
-                model = PeftModel.from_pretrained(model, adapter)
+                model: "LoraModel" = PeftModel.from_pretrained(
+                    model, adapter, offload_folder=model_args.offload_folder
+                )
                 model = model.merge_and_unload()
 
             if len(adapter_to_merge) > 0:
                 logger.info("Merged {} adapter(s).".format(len(adapter_to_merge)))
 
             if adapter_to_resume is not None:  # resume lora training
-                model = PeftModel.from_pretrained(model, adapter_to_resume, is_trainable=is_trainable)
+                model = PeftModel.from_pretrained(
+                    model, adapter_to_resume, is_trainable=is_trainable, offload_folder=model_args.offload_folder
+                )
 
         if is_trainable and adapter_to_resume is None:  # create new lora weights while training
             if len(finetuning_args.lora_target) == 1 and finetuning_args.lora_target[0] == "all":
@@ -99,40 +126,41 @@ def init_adapter(
             else:
                 target_modules = finetuning_args.lora_target
 
+            if finetuning_args.use_llama_pro:
+                target_modules = find_expanded_modules(model, target_modules, finetuning_args.num_layer_trainable)
+
+            if finetuning_args.use_dora and getattr(model, "quantization_method", None) is not None:
+                if getattr(model, "quantization_method", None) != QuantizationMethod.BITS_AND_BYTES:
+                    raise ValueError("DoRA is not compatible with PTQ-quantized models.")
+
             peft_kwargs = {
                 "r": finetuning_args.lora_rank,
                 "target_modules": target_modules,
                 "lora_alpha": finetuning_args.lora_alpha,
                 "lora_dropout": finetuning_args.lora_dropout,
+                "use_rslora": finetuning_args.use_rslora,
             }
 
             if model_args.use_unsloth:
-                from unsloth import FastLlamaModel, FastMistralModel  # type: ignore
+                from unsloth import FastLanguageModel  # type: ignore
 
                 unsloth_peft_kwargs = {"model": model, "max_seq_length": model_args.model_max_length}
-                if "loftq_config" in inspect.signature(FastLlamaModel.get_peft_model).parameters:
-                    unsloth_peft_kwargs["loftq_config"] = {}
-
-                if getattr(model.config, "model_type", None) == "llama":
-                    model = FastLlamaModel.get_peft_model(**peft_kwargs, **unsloth_peft_kwargs)
-                elif getattr(model.config, "model_type", None) == "mistral":
-                    model = FastMistralModel.get_peft_model(**peft_kwargs, **unsloth_peft_kwargs)
-                else:
-                    raise NotImplementedError
-
+                model = FastLanguageModel.get_peft_model(**peft_kwargs, **unsloth_peft_kwargs)
             else:
                 lora_config = LoraConfig(
                     task_type=TaskType.CAUSAL_LM,
                     inference_mode=False,
                     modules_to_save=finetuning_args.additional_target,
+                    use_dora=finetuning_args.use_dora,
                     **peft_kwargs,
                 )
                 model = get_peft_model(model, lora_config)
 
-        for param in filter(lambda p: p.requires_grad, model.parameters()):
-            param.data = param.data.to(torch.bfloat16 if finetuning_args.lora_bf16_mode else torch.float32)
+        if not finetuning_args.pure_bf16:
+            for param in filter(lambda p: p.requires_grad, model.parameters()):
+                param.data = param.data.to(torch.float32)
 
-    if model_args.adapter_name_or_path is not None:
-        logger.info("Loaded adapter(s): {}".format(",".join(model_args.adapter_name_or_path)))
+        if model_args.adapter_name_or_path is not None:
+            logger.info("Loaded adapter(s): {}".format(",".join(model_args.adapter_name_or_path)))
 
     return model

src/llmtuner/model/loader.py

@@ -1,8 +1,6 @@
-from typing import TYPE_CHECKING, Optional, Tuple
+from typing import TYPE_CHECKING, Any, Dict, Tuple
 
 from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer
-from transformers.integrations import is_deepspeed_zero3_enabled
-from transformers.utils.versions import require_version
 from trl import AutoModelForCausalLMWithValueHead
 
 from ..extras.logging import get_logger
@@ -21,50 +19,52 @@ if TYPE_CHECKING:
 logger = get_logger(__name__)
 
 
-require_version("transformers>=4.36.2", "To fix: pip install transformers>=4.36.2")
-require_version("datasets>=2.14.3", "To fix: pip install datasets>=2.14.3")
-require_version("accelerate>=0.21.0", "To fix: pip install accelerate>=0.21.0")
-require_version("peft>=0.7.0", "To fix: pip install peft>=0.7.0")
-require_version("trl>=0.7.6", "To fix: pip install trl>=0.7.6")
-
-
-def load_model_and_tokenizer(
-    model_args: "ModelArguments",
-    finetuning_args: "FinetuningArguments",
-    is_trainable: Optional[bool] = False,
-    add_valuehead: Optional[bool] = False,
-) -> Tuple["PreTrainedModel", "PreTrainedTokenizer"]:
-    r"""
-    Loads pretrained model and tokenizer.
-
-    Support both training and inference.
-    """
-
-    try_download_model_from_ms(model_args)
-
-    config_kwargs = {
+def _get_init_kwargs(model_args: "ModelArguments") -> Dict[str, Any]:
+    return {
         "trust_remote_code": True,
         "cache_dir": model_args.cache_dir,
         "revision": model_args.model_revision,
         "token": model_args.hf_hub_token,
     }
 
+
+def load_tokenizer(model_args: "ModelArguments") -> "PreTrainedTokenizer":
+    r"""
+    Loads pretrained tokenizer. Must before load_model.
+
+    Note: including inplace operation of model_args.
+    """
+    try_download_model_from_ms(model_args)
+    init_kwargs = _get_init_kwargs(model_args)
+
     tokenizer = AutoTokenizer.from_pretrained(
         model_args.model_name_or_path,
         use_fast=model_args.use_fast_tokenizer,
         split_special_tokens=model_args.split_special_tokens,
         padding_side="right",
-        **config_kwargs,
+        **init_kwargs,
     )
     patch_tokenizer(tokenizer)
+    return tokenizer
 
-    config = AutoConfig.from_pretrained(model_args.model_name_or_path, **config_kwargs)
-    patch_config(config, tokenizer, model_args, config_kwargs, is_trainable)
+
+def load_model(
+    tokenizer: "PreTrainedTokenizer",
+    model_args: "ModelArguments",
+    finetuning_args: "FinetuningArguments",
+    is_trainable: bool = False,
+    add_valuehead: bool = False,
+) -> "PreTrainedModel":
+    r"""
+    Loads pretrained model. Must after load_tokenizer.
+    """
+    init_kwargs = _get_init_kwargs(model_args)
+    config = AutoConfig.from_pretrained(model_args.model_name_or_path, **init_kwargs)
+    patch_config(config, tokenizer, model_args, init_kwargs, is_trainable)
 
     model = None
     if is_trainable and model_args.use_unsloth:
-        require_version("unsloth", "Follow the instructions at: https://github.com/unslothai/unsloth")
-        from unsloth import FastLlamaModel, FastMistralModel  # type: ignore
+        from unsloth import FastLanguageModel  # type: ignore
 
         unsloth_kwargs = {
             "model_name": model_args.model_name_or_path,
@@ -72,14 +72,12 @@ def load_model_and_tokenizer(
             "dtype": model_args.compute_dtype,
             "load_in_4bit": model_args.quantization_bit == 4,
             "token": model_args.hf_hub_token,
-            "device_map": get_current_device(),
+            "device_map": {"": get_current_device()},
             "rope_scaling": getattr(config, "rope_scaling", None),
         }
-        if getattr(config, "model_type", None) == "llama":
-            model, _ = FastLlamaModel.from_pretrained(**unsloth_kwargs)
-        elif getattr(config, "model_type", None) == "mistral":
-            model, _ = FastMistralModel.from_pretrained(**unsloth_kwargs)
-        else:
+        try:
+            model, _ = FastLanguageModel.from_pretrained(**unsloth_kwargs)
+        except NotImplementedError:
             logger.warning("Unsloth does not support model type {}.".format(getattr(config, "model_type", None)))
             model_args.use_unsloth = False
 
@@ -88,13 +86,7 @@ def load_model_and_tokenizer(
             logger.warning("Unsloth does not support loading adapters.")
 
     if model is None:
-        model = AutoModelForCausalLM.from_pretrained(
-            model_args.model_name_or_path,
-            config=config,
-            torch_dtype=model_args.compute_dtype,
-            low_cpu_mem_usage=(not is_deepspeed_zero3_enabled()),
-            **config_kwargs,
-        )
+        model = AutoModelForCausalLM.from_pretrained(model_args.model_name_or_path, config=config, **init_kwargs)
 
     patch_model(model, tokenizer, model_args, is_trainable)
     register_autoclass(config, model, tokenizer)
@@ -117,19 +109,42 @@ def load_model_and_tokenizer(
 
     if not is_trainable:
         model.requires_grad_(False)
-        model = model.to(model_args.compute_dtype) if not getattr(model, "quantization_method", None) else model
         model.eval()
+        for param in model.parameters():
+            if param.device.type == "cuda":
+                param.data = param.data.to(model_args.compute_dtype)
     else:
         model.train()
 
     trainable_params, all_param = count_parameters(model)
-    logger.info(
-        "trainable params: {:d} || all params: {:d} || trainable%: {:.4f}".format(
+    if is_trainable:
+        param_stats = "trainable params: {:d} || all params: {:d} || trainable%: {:.4f}".format(
             trainable_params, all_param, 100 * trainable_params / all_param
         )
-    )
+    else:
+        param_stats = "all params: {:d}".format(all_param)
+    logger.info(param_stats)
 
-    if not is_trainable:
-        logger.info("This IS expected that the trainable params is 0 if you are using model for inference only.")
+    if model_args.print_param_status:
+        for name, param in model.named_parameters():
+            print(
+                "name: {}, dtype: {}, device: {}, trainable: {}".format(
+                    name, param.dtype, param.device, param.requires_grad
+                )
+            )
 
+    return model
+
+
+def load_model_and_tokenizer(
+    model_args: "ModelArguments",
+    finetuning_args: "FinetuningArguments",
+    is_trainable: bool = False,
+    add_valuehead: bool = False,
+) -> Tuple["PreTrainedModel", "PreTrainedTokenizer"]:
+    r"""
+    Loads pretrained model and tokenizer.
+    """
+    tokenizer = load_tokenizer(model_args)
+    model = load_model(tokenizer, model_args, finetuning_args, is_trainable, add_valuehead)
     return model, tokenizer

src/llmtuner/model/patcher.py

@@ -3,10 +3,11 @@ import os
 import random
 from contextlib import nullcontext
 from types import MethodType
-from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple
+from typing import TYPE_CHECKING, Any, Dict, List, Tuple
 
 import torch
 from datasets import load_dataset
+from peft import PeftModel
 from transformers import BitsAndBytesConfig, GPTQConfig, PreTrainedModel, PreTrainedTokenizerBase
 from transformers.integrations import is_deepspeed_zero3_enabled
 from transformers.utils.versions import require_version
@@ -16,6 +17,8 @@ from ..extras.logging import get_logger
 from ..extras.misc import get_current_device, infer_optim_dtype
 from ..extras.packages import is_flash_attn2_available
 from ..extras.patches.llama_patch import apply_llama_patch
+from ..extras.patches.mixtral_patch import patch_mixtral_replace_moe_impl
+from .utils import QuantizationMethod
 
 
 if TYPE_CHECKING:
@@ -100,7 +103,27 @@ def _get_quantization_dataset(tokenizer: "PreTrainedTokenizer", model_args: "Mod
     return samples
 
 
+def _configure_attn_implementation(
+    config: "PretrainedConfig", model_args: "ModelArguments", init_kwargs: Dict[str, Any]
+) -> None:
+    if model_args.flash_attn:
+        if not is_flash_attn2_available():
+            logger.warning("FlashAttention2 is not installed.")
+            return
+
+        logger.info("Using FlashAttention-2 for faster training and inference.")
+        if getattr(config, "model_type", None) == "internlm2":  # special case for custom models
+            setattr(config, "attn_implementation", "flash_attention_2")
+        else:
+            init_kwargs["attn_implementation"] = "flash_attention_2"
+    else:
+        init_kwargs["attn_implementation"] = "eager"
+
+
 def _configure_rope(config: "PretrainedConfig", model_args: "ModelArguments", is_trainable: bool) -> None:
+    if model_args.rope_scaling is None:
+        return
+
     if not hasattr(config, "rope_scaling"):
         logger.warning("Current model does not support RoPE scaling.")
         return
@@ -127,16 +150,10 @@ def _configure_rope(config: "PretrainedConfig", model_args: "ModelArguments", is
     )
 
 
-def _configure_flashattn(config_kwargs: Dict[str, Any]) -> None:
-    if not is_flash_attn2_available():
-        logger.warning("FlashAttention2 is not installed.")
+def _configure_longlora(config: "PretrainedConfig", model_args: "ModelArguments", is_trainable: bool) -> None:
+    if not is_trainable or not model_args.shift_attn:
         return
 
-    config_kwargs["use_flash_attention_2"] = True
-    logger.info("Using FlashAttention-2 for faster training and inference.")
-
-
-def _configure_longlora(config: "PretrainedConfig") -> None:
     if getattr(config, "model_type", None) in SUPPORTED_CLASS_FOR_S2ATTN:
         setattr(config, "group_size_ratio", 0.25)
         apply_llama_patch()
@@ -149,20 +166,29 @@ def _configure_quantization(
     config: "PretrainedConfig",
     tokenizer: "PreTrainedTokenizer",
     model_args: "ModelArguments",
-    config_kwargs: Dict[str, Any],
+    init_kwargs: Dict[str, Any],
 ) -> None:
     r"""
-    Priority: GPTQ-quantized (training) > AutoGPTQ (export) > Bitsandbytes (training)
+    Priority: PTQ-quantized (training) > AutoGPTQ (export) > Bitsandbytes (training)
     """
-    if getattr(config, "quantization_config", None):  # gptq
+    if getattr(config, "quantization_config", None):  # ptq
         if is_deepspeed_zero3_enabled():
             raise ValueError("DeepSpeed ZeRO-3 is incompatible with quantization.")
 
-        config_kwargs["device_map"] = {"": get_current_device()}
+        init_kwargs["device_map"] = {"": get_current_device()}
         quantization_config: Dict[str, Any] = getattr(config, "quantization_config", None)
-        if quantization_config.get("quant_method", None) == "gptq" and quantization_config.get("bits", -1) == 4:
+        quant_method = quantization_config.get("quant_method", "")
+
+        if quant_method == QuantizationMethod.GPTQ:
             quantization_config["use_exllama"] = False  # disable exllama
-        logger.info("Loading {}-bit GPTQ-quantized model.".format(quantization_config.get("bits", -1)))
+
+        if quant_method == QuantizationMethod.AQLM:
+            require_version("transformers>=4.39.0", "To fix: pip install transformers>=4.39.0")
+            require_version("aqlm>=1.1.0", "To fix: pip install aqlm[gpu]>=1.1.0")
+            quantization_config["bits"] = 2
+
+        quant_bits = quantization_config.get("bits", "?")
+        logger.info("Loading {}-bit {}-quantized model.".format(quant_bits, quant_method.upper()))
 
     elif model_args.export_quantization_bit is not None:  # auto-gptq
         require_version("optimum>=1.16.0", "To fix: pip install optimum>=1.16.0")
@@ -172,38 +198,41 @@ def _configure_quantization(
         if getattr(config, "model_type", None) == "chatglm":
             raise ValueError("ChatGLM model is not supported.")
 
-        config_kwargs["quantization_config"] = GPTQConfig(
+        init_kwargs["quantization_config"] = GPTQConfig(
             bits=model_args.export_quantization_bit,
             tokenizer=tokenizer,
             dataset=_get_quantization_dataset(tokenizer, model_args),
         )
-        config_kwargs["device_map"] = "auto"
-        config_kwargs["max_memory"] = get_max_memory()
+        init_kwargs["device_map"] = "auto"
+        init_kwargs["max_memory"] = get_max_memory()
         logger.info("Quantizing model to {} bit.".format(model_args.export_quantization_bit))
 
     elif model_args.quantization_bit is not None:  # bnb
         if is_deepspeed_zero3_enabled():
-            raise ValueError("DeepSpeed ZeRO-3 is incompatible with quantization.")
+            require_version("transformers>=4.39.0", "To fix: pip install transformers>=4.39.0")
+            require_version("accelerate>=0.28.0", "To fix: pip install accelerate>=0.28.0")
+            require_version("bitsandbytes>=0.43.0", "To fix: pip install bitsandbytes>=0.43.0")
 
         if model_args.quantization_bit == 8:
             require_version("bitsandbytes>=0.37.0", "To fix: pip install bitsandbytes>=0.37.0")
-            config_kwargs["quantization_config"] = BitsAndBytesConfig(load_in_8bit=True)
+            init_kwargs["quantization_config"] = BitsAndBytesConfig(load_in_8bit=True)
 
         elif model_args.quantization_bit == 4:
             require_version("bitsandbytes>=0.39.0", "To fix: pip install bitsandbytes>=0.39.0")
-            config_kwargs["quantization_config"] = BitsAndBytesConfig(
+            init_kwargs["quantization_config"] = BitsAndBytesConfig(
                 load_in_4bit=True,
                 bnb_4bit_compute_dtype=model_args.compute_dtype,
                 bnb_4bit_use_double_quant=model_args.double_quantization,
                 bnb_4bit_quant_type=model_args.quantization_type,
+                bnb_4bit_quant_storage=model_args.compute_dtype,  # crucial for fsdp qlora
             )
 
-        config_kwargs["device_map"] = {"": get_current_device()}
+        init_kwargs["device_map"] = {"": get_current_device()}
         logger.info("Quantizing model to {} bit.".format(model_args.quantization_bit))
 
 
 def _prepare_model_for_training(
-    model: "PreTrainedModel", model_args: "ModelArguments", output_layer_name: Optional[str] = "lm_head"
+    model: "PreTrainedModel", model_args: "ModelArguments", output_layer_name: str = "lm_head"
 ) -> None:
     r"""
     Includes:
@@ -213,17 +242,20 @@ def _prepare_model_for_training(
     Inspired by: https://github.com/huggingface/peft/blob/v0.7.1/src/peft/utils/other.py#L72
     """
     if model_args.upcast_layernorm:
+        logger.info("Upcasting layernorm weights in float32.")
         for name, param in model.named_parameters():
             if param.ndim == 1 and any(ln_name in name for ln_name in LAYERNORM_NAMES):
                 param.data = param.data.to(torch.float32)
-        logger.info("Upcasting layernorm weights in float32.")
 
     if not model_args.disable_gradient_checkpointing:
         if not getattr(model, "supports_gradient_checkpointing", False):
             logger.warning("Current model does not support gradient checkpointing.")
         else:
-            model.gradient_checkpointing_enable(gradient_checkpointing_kwargs={"use_reentrant": False})
-            model.config.use_cache = False  # turn off when gradient checkpointing is enabled
+            # use_reentrant=False might increase VRAM usage (have not been empirically verified yet)
+            # According to: https://github.com/huggingface/transformers/issues/28339
+            model.gradient_checkpointing_enable(gradient_checkpointing_kwargs={"use_reentrant": True})
+            model.enable_input_require_grads()
+            setattr(model.config, "use_cache", False)  # turn off when gradient checkpointing is enabled
             logger.info("Gradient checkpointing enabled.")
 
     if hasattr(model, output_layer_name) and model_args.upcast_lmhead_output:
@@ -231,6 +263,7 @@ def _prepare_model_for_training(
         def fp32_forward_post_hook(module: torch.nn.Module, args: Tuple[torch.Tensor], output: torch.Tensor):
             return output.to(torch.float32)
 
+        logger.info("Upcasting lm_head outputs in float32.")
         output_layer = getattr(model, output_layer_name)
         if isinstance(output_layer, torch.nn.Linear) and output_layer.weight.dtype != torch.float32:
             output_layer.register_forward_hook(fp32_forward_post_hook)
@@ -245,26 +278,35 @@ def patch_config(
     config: "PretrainedConfig",
     tokenizer: "PreTrainedTokenizer",
     model_args: "ModelArguments",
-    config_kwargs: Dict[str, Any],
+    init_kwargs: Dict[str, Any],
     is_trainable: bool,
 ) -> None:
     if model_args.compute_dtype is None:  # priority: bf16 > fp16 > fp32
         model_args.compute_dtype = infer_optim_dtype(model_dtype=getattr(config, "torch_dtype", None))
 
     if getattr(config, "model_type", None) == "qwen":
+        setattr(config, "use_flash_attn", model_args.flash_attn)
         for dtype_name, dtype in [("fp16", torch.float16), ("bf16", torch.bfloat16), ("fp32", torch.float32)]:
             setattr(config, dtype_name, model_args.compute_dtype == dtype)
 
-    if model_args.rope_scaling is not None:
-        _configure_rope(config, model_args, is_trainable)
+    _configure_attn_implementation(config, model_args, init_kwargs)
+    _configure_rope(config, model_args, is_trainable)
+    _configure_longlora(config, model_args, is_trainable)
+    _configure_quantization(config, tokenizer, model_args, init_kwargs)
 
-    if model_args.flash_attn:
-        _configure_flashattn(config_kwargs)
+    if model_args.use_cache and not is_trainable:
+        setattr(config, "use_cache", True)
+        logger.info("Using KV cache for faster generation.")
 
-    if is_trainable and model_args.shift_attn:
-        _configure_longlora(config)
+    init_kwargs["torch_dtype"] = model_args.compute_dtype
+    if not is_deepspeed_zero3_enabled():
+        init_kwargs["low_cpu_mem_usage"] = model_args.low_cpu_mem_usage
+        if init_kwargs["low_cpu_mem_usage"]:
+            if "device_map" not in init_kwargs:  # quant models cannot use auto device map
+                init_kwargs["device_map"] = model_args.device_map or {"": get_current_device()}
 
-    _configure_quantization(config, tokenizer, model_args, config_kwargs)
+            if init_kwargs["device_map"] == "auto":
+                init_kwargs["offload_folder"] = model_args.offload_folder
 
 
 def patch_model(
@@ -283,6 +325,21 @@ def patch_model(
     if is_trainable:
         _prepare_model_for_training(model, model_args)
 
+    if getattr(model.config, "model_type", None) == "mixtral" and is_deepspeed_zero3_enabled():
+        require_version("deepspeed>=0.13.0", "To fix: pip install deepspeed>=0.13.0")
+        from deepspeed.utils import set_z3_leaf_modules  # type: ignore
+        from transformers.models.mixtral.modeling_mixtral import MixtralSparseMoeBlock
+
+        set_z3_leaf_modules(model, [MixtralSparseMoeBlock])
+
+        if is_trainable:
+            patch_mixtral_replace_moe_impl()
+
+    try:
+        model.add_model_tags(["llama-factory"])
+    except Exception:
+        logger.warning("Cannot properly tag the model.")
+
 
 def patch_valuehead_model(model: "AutoModelForCausalLMWithValueHead") -> None:
     def tie_weights(self: "AutoModelForCausalLMWithValueHead") -> None:
@@ -293,7 +350,12 @@ def patch_valuehead_model(model: "AutoModelForCausalLMWithValueHead") -> None:
         if isinstance(self.pretrained_model, PreTrainedModel):
             return self.pretrained_model.get_input_embeddings()
 
+    def create_or_update_model_card(self: "AutoModelForCausalLMWithValueHead", output_dir: str) -> None:
+        if isinstance(self.pretrained_model, PeftModel):
+            self.pretrained_model.create_or_update_model_card(output_dir)
+
     ignore_modules = [name for name, _ in model.named_parameters() if "pretrained_model" in name]
     setattr(model, "_keys_to_ignore_on_save", ignore_modules)
     setattr(model, "tie_weights", MethodType(tie_weights, model))
     setattr(model, "get_input_embeddings", MethodType(get_input_embeddings, model))
+    setattr(model, "create_or_update_model_card", MethodType(create_or_update_model_card, model))

src/llmtuner/model/utils.py

@@ -1,5 +1,5 @@
-import inspect
-from typing import TYPE_CHECKING, Any, Dict, List
+from enum import Enum, unique
+from typing import TYPE_CHECKING, Dict, List
 
 import torch
 from transformers import PreTrainedModel
@@ -7,46 +7,27 @@ from transformers.utils import cached_file
 
 from ..extras.constants import V_HEAD_SAFE_WEIGHTS_NAME, V_HEAD_WEIGHTS_NAME
 from ..extras.logging import get_logger
-from ..extras.misc import get_current_device
 
 
 if TYPE_CHECKING:
     from transformers import PretrainedConfig, PreTrainedTokenizer
 
-    from ..hparams import DataArguments, FinetuningArguments, ModelArguments
+    from ..hparams import ModelArguments
 
 
 logger = get_logger(__name__)
 
 
-def dispatch_model(model: "PreTrainedModel") -> "PreTrainedModel":
+@unique
+class QuantizationMethod(str, Enum):
     r"""
-    Dispatches a pre-trained model to GPUs with balanced memory when the GPU is available.
-    Borrowed from: https://github.com/huggingface/transformers/blob/v4.36.2/src/transformers/modeling_utils.py#L3570
+    Borrowed from `transformers.utils.quantization_config.QuantizationMethod`.
     """
-    if getattr(model, "quantization_method", None):  # already set on current device
-        return model
 
-    if (
-        torch.cuda.device_count() > 1
-        and isinstance(model, PreTrainedModel)
-        and model._no_split_modules is not None
-        and model.config.model_type != "chatglm"
-    ):
-        from accelerate import dispatch_model
-        from accelerate.utils import get_balanced_memory, infer_auto_device_map
-
-        kwargs = {"dtype": model.dtype, "no_split_module_classes": model._get_no_split_modules("auto")}
-        max_memory = get_balanced_memory(model, **kwargs)
-        # Make sure tied weights are tied before creating the device map.
-        model.tie_weights()
-        device_map = infer_auto_device_map(model, max_memory=max_memory, **kwargs)
-        device_map_kwargs = {"device_map": device_map}
-        if "skip_keys" in inspect.signature(dispatch_model).parameters:
-            device_map_kwargs["skip_keys"] = model._skip_keys_device_placement
-        return dispatch_model(model, **device_map_kwargs)
-    else:
-        return model.to(device=get_current_device())
+    BITS_AND_BYTES = "bitsandbytes"
+    GPTQ = "gptq"
+    AWQ = "awq"
+    AQLM = "aqlm"
 
 
 def find_all_linear_modules(model: "PreTrainedModel") -> List[str]:
@@ -56,7 +37,7 @@ def find_all_linear_modules(model: "PreTrainedModel") -> List[str]:
     quantization_method = getattr(model, "quantization_method", None)
     if quantization_method is None:
         linear_cls = torch.nn.Linear
-    elif quantization_method == "bitsandbytes":
+    elif quantization_method == QuantizationMethod.BITS_AND_BYTES:
         import bitsandbytes as bnb
 
         linear_cls = bnb.nn.Linear4bit if getattr(model, "is_loaded_in_4bit", False) else bnb.nn.Linear8bitLt
@@ -76,16 +57,31 @@ def find_all_linear_modules(model: "PreTrainedModel") -> List[str]:
     return list(module_names)
 
 
-def get_modelcard_args(
-    model_args: "ModelArguments", data_args: "DataArguments", finetuning_args: "FinetuningArguments"
-) -> Dict[str, Any]:
-    return {
-        "tasks": "text-generation",
-        "license": "other",
-        "finetuned_from": model_args.model_name_or_path,
-        "dataset": [dataset.strip() for dataset in data_args.dataset.split(",")],
-        "tags": ["llama-factory"] + (["lora"] if finetuning_args.finetuning_type == "lora" else []),
-    }
+def find_expanded_modules(model: "PreTrainedModel", target_modules: List[str], num_layer_trainable: int) -> List[str]:
+    r"""
+    Finds the modules in the expanded blocks to apply lora.
+    """
+    num_layers = getattr(model.config, "num_hidden_layers", None)
+    if not num_layers:
+        raise ValueError("Model was not supported.")
+
+    if num_layers % num_layer_trainable != 0:
+        raise ValueError(
+            "`num_layers` {} should be divisible by `num_layer_trainable` {}.".format(num_layers, num_layer_trainable)
+        )
+
+    stride = num_layers // num_layer_trainable
+    trainable_layer_ids = range(stride - 1, num_layers + stride - 1, stride)
+    trainable_layers = [".{:d}.".format(idx) for idx in trainable_layer_ids]
+    module_names = []
+    for name, _ in model.named_modules():
+        if any(target_module in name for target_module in target_modules) and any(
+            trainable_layer in name for trainable_layer in trainable_layers
+        ):
+            module_names.append(name)
+
+    logger.info("Apply lora to layers: {}".format(",".join(map(str, trainable_layer_ids))))
+    return module_names
 
 
 def load_valuehead_params(path_or_repo_id: str, model_args: "ModelArguments") -> Dict[str, torch.Tensor]:

src/llmtuner/train/dpo/trainer.py

@@ -8,21 +8,25 @@ from trl import DPOTrainer
 from trl.trainer.utils import disable_dropout_in_model
 
 from ...extras.constants import IGNORE_INDEX
+from ..utils import create_custom_optimzer
 
 
 if TYPE_CHECKING:
     from transformers import PreTrainedModel
 
+    from ...hparams import FinetuningArguments
+
 
 class CustomDPOTrainer(DPOTrainer):
     def __init__(
         self,
         beta: float,
-        loss_type: Literal["sigmoid", "hinge", "ipo", "kto"],
+        loss_type: Literal["sigmoid", "hinge", "ipo", "kto_pair"],
         ftx_gamma: float,
         model: Union["PreTrainedModel", torch.nn.Module],
+        finetuning_args: "FinetuningArguments",
         ref_model: Optional[Union["PreTrainedModel", torch.nn.Module]] = None,
-        disable_dropout: Optional[bool] = True,
+        disable_dropout: bool = True,
         **kwargs,
     ):
         if disable_dropout:
@@ -30,6 +34,8 @@ class CustomDPOTrainer(DPOTrainer):
             if ref_model is not None:
                 disable_dropout_in_model(ref_model)
 
+        self.finetuning_args = finetuning_args
+        self.reference_free = False
         self.use_dpo_data_collator = True  # hack to avoid warning
         self.generate_during_eval = False  # disable at evaluation
         self.label_pad_token_id = IGNORE_INDEX
@@ -60,6 +66,13 @@ class CustomDPOTrainer(DPOTrainer):
             else:
                 self.ref_model = self.accelerator.prepare_model(self.ref_model, evaluation_mode=True)
 
+    def create_optimizer_and_scheduler(self, num_training_steps: int) -> None:
+        self.optimizer = create_custom_optimzer(self.model, self.args, self.finetuning_args, num_training_steps)
+        if self.optimizer is None:
+            self.create_optimizer()
+
+        self.create_scheduler(num_training_steps=num_training_steps, optimizer=self.optimizer)
+
     def sft_loss(self, chosen_logits: torch.FloatTensor, chosen_labels: torch.LongTensor) -> torch.Tensor:
         r"""
         Computes supervised cross-entropy loss of given labels under the given logits.
@@ -94,7 +107,7 @@ class CustomDPOTrainer(DPOTrainer):
         self,
         model: "PreTrainedModel",
         batch: Dict[str, torch.Tensor],
-        train_eval: Optional[Literal["train", "eval"]] = "train",
+        train_eval: Literal["train", "eval"] = "train",
     ) -> Tuple[torch.Tensor, Dict[str, torch.Tensor]]:
         r"""
         Computes the DPO loss and other metrics for the given batch of inputs for train or test.

src/llmtuner/train/dpo/workflow.py

@@ -2,20 +2,18 @@
 
 from typing import TYPE_CHECKING, List, Optional
 
-from transformers import Seq2SeqTrainingArguments
-
 from ...data import get_dataset, split_dataset
 from ...extras.constants import IGNORE_INDEX
 from ...extras.ploting import plot_loss
 from ...hparams import ModelArguments
-from ...model import load_model_and_tokenizer
-from ...train.dpo.collator import DPODataCollatorWithPadding
-from ...train.dpo.trainer import CustomDPOTrainer
-from ...train.utils import create_modelcard_and_push, create_ref_model
+from ...model import load_model, load_tokenizer
+from ..utils import create_modelcard_and_push, create_ref_model
+from .collator import DPODataCollatorWithPadding
+from .trainer import CustomDPOTrainer
 
 
 if TYPE_CHECKING:
-    from transformers import TrainerCallback
+    from transformers import Seq2SeqTrainingArguments, TrainerCallback
 
     from ...hparams import DataArguments, FinetuningArguments
 
@@ -27,8 +25,9 @@ def run_dpo(
     finetuning_args: "FinetuningArguments",
     callbacks: Optional[List["TrainerCallback"]] = None,
 ):
-    model, tokenizer = load_model_and_tokenizer(model_args, finetuning_args, training_args.do_train)
+    tokenizer = load_tokenizer(model_args)
     dataset = get_dataset(tokenizer, model_args, data_args, training_args, stage="rm")
+    model = load_model(tokenizer, model_args, finetuning_args, training_args.do_train)
     data_collator = DPODataCollatorWithPadding(
         tokenizer=tokenizer,
         pad_to_multiple_of=8,
@@ -42,15 +41,14 @@ def run_dpo(
         ref_model = create_ref_model(model_args, finetuning_args)
 
     # Update arguments
-    training_args_dict = training_args.to_dict()
-    training_args_dict.update(dict(remove_unused_columns=False))  # important for pairwise dataset
-    training_args = Seq2SeqTrainingArguments(**training_args_dict)
+    training_args.remove_unused_columns = False  # important for pairwise dataset
 
     # Initialize our Trainer
     trainer = CustomDPOTrainer(
         beta=finetuning_args.dpo_beta,
         loss_type=finetuning_args.dpo_loss,
         ftx_gamma=finetuning_args.dpo_ftx,
+        finetuning_args=finetuning_args,
         model=model,
         ref_model=ref_model,
         args=training_args,

src/llmtuner/train/ppo/trainer.py

@@ -14,7 +14,7 @@ from trl.core import PPODecorators, logprobs_from_logits
 
 from ...extras.callbacks import FixValueHeadModelCallback, LogCallback
 from ...extras.logging import get_logger
-from ...extras.misc import AverageMeter, count_parameters, get_logits_processor
+from ...extras.misc import AverageMeter, count_parameters, get_current_device, get_logits_processor
 from .utils import dump_layernorm, get_rewards_from_server, replace_model, restore_layernorm
 
 
@@ -49,6 +49,7 @@ class CustomPPOTrainer(PPOTrainer, Trainer):
         self.model_args = model_args
         self.finetuning_args = finetuning_args
         self.reward_model = reward_model
+        self.current_device = get_current_device()  # patch for deepspeed training
 
         self.generation_config = GenerationConfig(
             pad_token_id=self.tokenizer.pad_token_id,
@@ -291,7 +292,7 @@ class CustomPPOTrainer(PPOTrainer, Trainer):
         queries: torch.Tensor,
         responses: torch.Tensor,
         model_inputs: dict,
-        return_logits: Optional[bool] = False,
+        return_logits: bool = False,
         response_masks: Optional[torch.Tensor] = None,
     ):
         r"""

src/llmtuner/train/ppo/utils.py

@@ -1,7 +1,9 @@
 import json
+from contextlib import nullcontext
 from typing import TYPE_CHECKING, Dict, List, Literal, Optional
 
 import torch
+from transformers.integrations import is_deepspeed_zero3_enabled
 
 from ...extras.packages import is_requests_available
 
@@ -23,18 +25,22 @@ def get_rewards_from_server(server_url: str, messages: List[str]) -> List[torch.
 
 
 def replace_model(model: "AutoModelForCausalLMWithValueHead", target: Literal["default", "reward"]) -> None:
-    if target == "reward":  # save default head temporarily
-        valuehead_state_dict: Dict[str, torch.Tensor] = model.v_head.state_dict()
-        setattr(model, "default_head_weight", valuehead_state_dict["summary.weight"].detach().clone())
-        setattr(model, "default_head_bias", valuehead_state_dict["summary.bias"].detach().clone())
+    if is_deepspeed_zero3_enabled():
+        import deepspeed  # type: ignore
 
-    model.pretrained_model.set_adapter(target)  # set the LoRA adapter to be active
-    model.v_head.load_state_dict(
-        {
-            "summary.weight": model.get_buffer("{}_head_weight".format(target)).detach().clone(),
-            "summary.bias": model.get_buffer("{}_head_bias".format(target)).detach().clone(),
-        }
-    )
+        params = [model.v_head.summary.weight, model.v_head.summary.bias]
+        context_maybe_zero3 = deepspeed.zero.GatheredParameters(params, modifier_rank=0)
+    else:
+        context_maybe_zero3 = nullcontext()
+
+    with context_maybe_zero3:
+        if target == "reward":  # save default head temporarily
+            setattr(model, "default_head_weight", model.v_head.summary.weight.data.detach().clone())
+            setattr(model, "default_head_bias", model.v_head.summary.bias.data.detach().clone())
+
+        model.pretrained_model.set_adapter(target)  # set the LoRA adapter to be active
+        model.v_head.summary.weight.data = model.get_buffer("{}_head_weight".format(target)).detach().clone()
+        model.v_head.summary.bias.data = model.get_buffer("{}_head_bias".format(target)).detach().clone()
 
 
 def dump_layernorm(model: "PreTrainedModel") -> Dict[str, torch.Tensor]: