update parser

Former-commit-id: 5262c8702382ff8bc36a172387bc4c8949f326ea

.dockerignore

@@ -4,10 +4,10 @@
 .venv
 cache
 data
+docker
+saves
 hf_cache
 output
-examples
 .dockerignore
 .gitattributes
 .gitignore
-Dockerfile

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -1,13 +1,19 @@
 name: "\U0001F41B Bug / Help"
 description: Create a report to help us improve the LLaMA Factory
 body:
+  - type: markdown
+    attributes:
+      value: |
+        Issues included in **FAQs** or those with **insufficient** information may be closed without a response.
+        包含在**常见问题**内或提供信息**不完整**的 issues 可能不会被回复。
+
   - type: checkboxes
     id: reminder
     attributes:
       label: Reminder
       description: |
-        Please ensure you have read the README carefully and searched the existing issues.
-        请确保您已经认真阅读了 README 并且搜索过现有的 Issue。
+        Please ensure you have read the README carefully and searched the existing issues (including FAQs).
+        请确保您已经认真阅读了 README 并且搜索过现有的 issues（包括常见问题）。
 
       options:
         - label: I have read the README and searched the existing issues.
@@ -38,7 +44,9 @@ body:
         请合理使用 Markdown 标签来格式化您的文本。
 
       placeholder: |
+        ```bash
         llamafactory-cli train ...
+        ```
 
   - type: textarea
     id: expected-behavior

.github/PULL_REQUEST_TEMPLATE.md

@@ -5,3 +5,4 @@ Fixes # (issue)
 ## Before submitting
 
 - [ ] Did you read the [contributor guideline](https://github.com/hiyouga/LLaMA-Factory/blob/main/.github/CONTRIBUTING.md)?
+- [ ] Did you write any new necessary tests?

.github/workflows/label_issue.yml

@@ -0,0 +1,30 @@
+name: label_issue
+
+on:
+  issues:
+    types:
+      - opened
+
+jobs:
+  label_issue:
+    runs-on: ubuntu-latest
+
+    permissions:
+      issues: write
+
+    steps:
+      - env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          ISSUE_URL: ${{ github.event.issue.html_url }}
+          ISSUE_TITLE: ${{ github.event.issue.title }}
+        run: |
+          LABEL=pending
+          NPU_KEYWORDS=(npu huawei ascend 华为 昇腾)
+          ISSUE_TITLE_LOWER=$(echo $ISSUE_TITLE | tr '[:upper:]' '[:lower:]')
+          for KEYWORD in ${NPU_KEYWORDS[@]}; do
+            if [[ $ISSUE_TITLE_LOWER == *$KEYWORD* ]] && [[ $ISSUE_TITLE_LOWER != *input* ]]; then
+              LABEL=pending,npu
+              break
+            fi
+          done
+          gh issue edit $ISSUE_URL --add-label $LABEL

.github/workflows/publish.yml

@@ -0,0 +1,40 @@
+name: publish
+
+on:
+  release:
+    types:
+      - published
+
+jobs:
+  publish:
+    name: Upload release to PyPI
+
+    runs-on: ubuntu-latest
+
+    environment:
+      name: release
+      url: https://pypi.org/p/llamafactory
+
+    permissions:
+      id-token: write
+
+    steps:
+      - name: Checkout
+        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 build
+      
+      - name: Build package
+        run: |
+          python -m build
+      
+      - name: Publish package
+        uses: pypa/gh-action-pypi-publish@release/v1

.github/workflows/tests.yml

@@ -19,21 +19,33 @@ on:
 jobs:
   tests:
     runs-on: ubuntu-latest
+
+    environment:
+      name: tests
+
+    env:
+      HF_TOKEN: ${{ secrets.HF_TOKEN }}
+
     steps:
-      - uses: actions/checkout@v4
+      - name: Checkout
+        uses: actions/checkout@v4
+
       - name: Set up Python
         uses: actions/setup-python@v5
         with:
           python-version: "3.8"
           cache: "pip"
           cache-dependency-path: "setup.py"
+
       - name: Install dependencies
         run: |
           python -m pip install --upgrade pip
-          python -m pip install .[torch,dev]
+          python -m pip install ".[torch,dev]"
+
       - name: Check quality
         run: |
           make style && make quality
+
       - name: Test with pytest
         run: |
           make test

.gitignore

@@ -160,6 +160,8 @@ cython_debug/
 .idea/
 
 # custom .gitignore
-user.config
-saves/
 cache/
+config/
+saves/
+output/
+wandb/

CITATION.cff

@@ -12,12 +12,16 @@ authors:
   given-names: "Yanhan"
 - family-names: "Luo"
   given-names: "Zheyan"
+- family-names: "Feng"
+  given-names: "Zhangchi"
 - 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
+  type: conference-paper
+  conference:
+    name: "Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations)"
   authors:
     - family-names: "Zheng"
       given-names: "Yaowei"
@@ -29,9 +33,12 @@ preferred-citation:
       given-names: "Yanhan"
     - family-names: "Luo"
       given-names: "Zheyan"
+    - family-names: "Feng"
+      given-names: "Zhangchi"
     - 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
+  publisher: "Association for Computational Linguistics"
+  address: "Bangkok, Thailand"

Dockerfile

@@ -1,14 +0,0 @@
-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 .[metrics,bitsandbytes,qwen]
-
-VOLUME [ "/root/.cache/huggingface/", "/app/data", "/app/output" ]
-EXPOSE 7860
-
-CMD [ "llamafactory-cli", "webui" ]

MANIFEST.in

@@ -0,0 +1 @@
+include LICENSE requirements.txt

Makefile

@@ -11,4 +11,4 @@ style:
 	ruff format $(check_dirs)
 
 test:
-	pytest tests/
+	CUDA_VISIBLE_DEVICES= pytest tests/

README.md

@@ -4,7 +4,7 @@
 [![GitHub Code License](https://img.shields.io/github/license/hiyouga/LLaMA-Factory)](LICENSE)
 [![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/llamafactory)](https://pypi.org/project/llamafactory/)
-[![Citation](https://img.shields.io/badge/citation-44-green)](#projects-using-llama-factory)
+[![Citation](https://img.shields.io/badge/citation-72-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)
 [![Twitter](https://img.shields.io/twitter/follow/llamafactory_ai)](https://twitter.com/llamafactory_ai)
@@ -15,7 +15,7 @@
 
 [![GitHub Tread](https://trendshift.io/api/badge/repositories/4535)](https://trendshift.io/repositories/4535)
 
-👋 Join our [WeChat](assets/wechat.jpg).
+👋 Join our [WeChat](assets/wechat.jpg) or [NPU user group](assets/wechat_npu.jpg).
 
 \[ English | [中文](README_zh.md) \]
 
@@ -48,8 +48,8 @@ Choose your path:
 
 - **Various models**: LLaMA, LLaVA, Mistral, Mixtral-MoE, Qwen, Yi, Gemma, Baichuan, ChatGLM, Phi, etc.
 - **Integrated methods**: (Continuous) pre-training, (multimodal) supervised fine-tuning, reward modeling, PPO, DPO, KTO, ORPO, etc.
-- **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, BAdam, DoRA, LongLoRA, LLaMA Pro, Mixture-of-Depths, LoRA+, LoftQ and Agent tuning.
+- **Scalable resources**: 16-bit full-tuning, freeze-tuning, LoRA and 2/3/4/5/6/8-bit QLoRA via AQLM/AWQ/GPTQ/LLM.int8/HQQ/EETQ.
+- **Advanced algorithms**: GaLore, BAdam, DoRA, LongLoRA, LLaMA Pro, Mixture-of-Depths, LoRA+, LoftQ, PiSSA 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.
@@ -71,9 +71,9 @@ Compared to ChatGLM's [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/
 
 ## Changelog
 
-[24/06/07] We supported fine-tuning the **[Qwen-2](https://qwenlm.github.io/blog/qwen2/)** series models.
+[24/06/16] We support **[PiSSA](https://arxiv.org/abs/2404.02948)** algorithm. See [examples](examples/README.md) for usage.
 
-[24/06/05] We supported fine-tuning the **[GLM-4-9B/GLM-4-9B-Chat](https://github.com/THUDM/GLM-4)** models.
+[24/06/07] We supported fine-tuning the **[Qwen2](https://qwenlm.github.io/blog/qwen2/)** and **[GLM-4](https://github.com/THUDM/GLM-4)** models.
 
 [24/05/26] We supported **[SimPO](https://arxiv.org/abs/2405.14734)** algorithm for preference learning. See [examples](examples/README.md) for usage.
 
@@ -151,35 +151,32 @@ Compared to ChatGLM's [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/
 
 ## Supported Models
 
-| Model                                                    | Model size                       | Template  |
-| -------------------------------------------------------- | -------------------------------- | --------- |
-| [Baichuan2](https://huggingface.co/baichuan-inc)         | 7B/13B                           | baichuan2 |
-| [BLOOM](https://huggingface.co/bigscience)               | 560M/1.1B/1.7B/3B/7.1B/176B      | -         |
-| [BLOOMZ](https://huggingface.co/bigscience)              | 560M/1.1B/1.7B/3B/7.1B/176B      | -         |
-| [ChatGLM3](https://huggingface.co/THUDM)                 | 6B                               | chatglm3  |
-| [Command-R](https://huggingface.co/CohereForAI)          | 35B/104B                         | cohere    |
-| [DeepSeek (MoE)](https://huggingface.co/deepseek-ai)     | 7B/16B/67B/236B                  | deepseek  |
-| [Falcon](https://huggingface.co/tiiuae)                  | 7B/11B/40B/180B                  | falcon    |
-| [Gemma/CodeGemma](https://huggingface.co/google)         | 2B/7B                            | gemma     |
-| [GLM4](https://huggingface.co/THUDM)                     | 9B                               | glm4      |
-| [InternLM2](https://huggingface.co/internlm)             | 7B/20B                           | intern2   |
-| [LLaMA](https://github.com/facebookresearch/llama)       | 7B/13B/33B/65B                   | -         |
-| [LLaMA-2](https://huggingface.co/meta-llama)             | 7B/13B/70B                       | llama2    |
-| [LLaMA-3](https://huggingface.co/meta-llama)             | 8B/70B                           | llama3    |
-| [LLaVA-1.5](https://huggingface.co/llava-hf)             | 7B/13B                           | vicuna    |
-| [Mistral/Mixtral](https://huggingface.co/mistralai)      | 7B/8x7B/8x22B                    | mistral   |
-| [OLMo](https://huggingface.co/allenai)                   | 1B/7B                            | -         |
-| [PaliGemma](https://huggingface.co/google)               | 3B                               | gemma     |
-| [Phi-1.5/2](https://huggingface.co/microsoft)            | 1.3B/2.7B                        | -         |
-| [Phi-3](https://huggingface.co/microsoft)                | 4B/7B/14B                        | phi       |
-| [Qwen](https://huggingface.co/Qwen)                      | 1.8B/7B/14B/72B                  | qwen      |
-| [Qwen1.5 (Code/MoE)](https://huggingface.co/Qwen)        | 0.5B/1.8B/4B/7B/14B/32B/72B/110B | qwen      |
-| [Qwen2 (MoE)](https://huggingface.co/Qwen)               | 0.5B/1.5B/7B/57B/72B             | qwen      |
-| [StarCoder2](https://huggingface.co/bigcode)             | 3B/7B/15B                        | -         |
-| [XVERSE](https://huggingface.co/xverse)                  | 7B/13B/65B                       | xverse    |
-| [Yi (1/1.5)](https://huggingface.co/01-ai)               | 6B/9B/34B                        | yi        |
-| [Yi-VL](https://huggingface.co/01-ai)                    | 6B/34B                           | yi_vl     |
-| [Yuan](https://huggingface.co/IEITYuan)                  | 2B/51B/102B                      | yuan      |
+| Model                                                        | Model size                       | Template  |
+| ------------------------------------------------------------ | -------------------------------- | --------- |
+| [Baichuan 2](https://huggingface.co/baichuan-inc)            | 7B/13B                           | baichuan2 |
+| [BLOOM/BLOOMZ](https://huggingface.co/bigscience)            | 560M/1.1B/1.7B/3B/7.1B/176B      | -         |
+| [ChatGLM3](https://huggingface.co/THUDM)                     | 6B                               | chatglm3  |
+| [Command R](https://huggingface.co/CohereForAI)              | 35B/104B                         | cohere    |
+| [DeepSeek (Code/MoE)](https://huggingface.co/deepseek-ai)    | 7B/16B/67B/236B                  | deepseek  |
+| [Falcon](https://huggingface.co/tiiuae)                      | 7B/11B/40B/180B                  | falcon    |
+| [Gemma/Gemma 2/CodeGemma](https://huggingface.co/google)     | 2B/7B/9B/27B                     | gemma     |
+| [GLM-4](https://huggingface.co/THUDM)                        | 9B                               | glm4      |
+| [InternLM2](https://huggingface.co/internlm)                 | 7B/20B                           | intern2   |
+| [Llama](https://github.com/facebookresearch/llama)           | 7B/13B/33B/65B                   | -         |
+| [Llama 2](https://huggingface.co/meta-llama)                 | 7B/13B/70B                       | llama2    |
+| [Llama 3](https://huggingface.co/meta-llama)                 | 8B/70B                           | llama3    |
+| [LLaVA-1.5](https://huggingface.co/llava-hf)                 | 7B/13B                           | vicuna    |
+| [Mistral/Mixtral](https://huggingface.co/mistralai)          | 7B/8x7B/8x22B                    | mistral   |
+| [OLMo](https://huggingface.co/allenai)                       | 1B/7B                            | -         |
+| [PaliGemma](https://huggingface.co/google)                   | 3B                               | gemma     |
+| [Phi-1.5/Phi-2](https://huggingface.co/microsoft)            | 1.3B/2.7B                        | -         |
+| [Phi-3](https://huggingface.co/microsoft)                    | 4B/7B/14B                        | phi       |
+| [Qwen/Qwen1.5/Qwen2 (Code/MoE)](https://huggingface.co/Qwen) | 0.5B/1.5B/4B/7B/14B/32B/72B/110B | qwen      |
+| [StarCoder 2](https://huggingface.co/bigcode)                | 3B/7B/15B                        | -         |
+| [XVERSE](https://huggingface.co/xverse)                      | 7B/13B/65B                       | xverse    |
+| [Yi/Yi-1.5](https://huggingface.co/01-ai)                    | 6B/9B/34B                        | yi        |
+| [Yi-VL](https://huggingface.co/01-ai)                        | 6B/34B                           | yi_vl     |
+| [Yuan 2](https://huggingface.co/IEITYuan)                    | 2B/51B/102B                      | yuan      |
 
 > [!NOTE]
 > For the "base" models, the `template` argument can be chosen from `default`, `alpaca`, `vicuna` etc. But make sure to use the **corresponding template** for the "instruct/chat" models.
@@ -259,6 +256,9 @@ You also can add a custom chat template to [template.py](src/llamafactory/data/t
 - [Cosmopedia (en)](https://huggingface.co/datasets/HuggingFaceTB/cosmopedia)
 - [STEM (zh)](https://huggingface.co/datasets/hfl/stem_zh_instruction)
 - [Ruozhiba (zh)](https://huggingface.co/datasets/hfl/ruozhiba_gpt4_turbo)
+- [Neo-sft (zh)](https://huggingface.co/datasets/m-a-p/neo_sft_phase2)
+- [WebInstructSub (en)](https://huggingface.co/datasets/TIGER-Lab/WebInstructSub)
+- [Magpie-Pro-300K-Filtered (en)](https://huggingface.co/datasets/Magpie-Align/Magpie-Pro-300K-Filtered)
 - [LLaVA mixed (en&zh)](https://huggingface.co/datasets/BUAADreamer/llava-en-zh-300k)
 - [Open Assistant (de)](https://huggingface.co/datasets/mayflowergmbh/oasst_de)
 - [Dolly 15k (de)](https://huggingface.co/datasets/mayflowergmbh/dolly-15k_de)
@@ -335,10 +335,10 @@ huggingface-cli login
 ```bash
 git clone --depth 1 https://github.com/hiyouga/LLaMA-Factory.git
 cd LLaMA-Factory
-pip install -e '.[torch,metrics]'
+pip install -e ".[torch,metrics]"
 ```
 
-Extra dependencies available: torch, torch_npu, metrics, deepspeed, bitsandbytes, vllm, galore, badam, gptq, awq, aqlm, qwen, modelscope, quality
+Extra dependencies available: torch, torch-npu, metrics, deepspeed, bitsandbytes, hqq, eetq, gptq, awq, aqlm, vllm, galore, badam, qwen, modelscope, quality
 
 > [!TIP]
 > Use `pip install --no-deps -e .` to resolve package conflicts.
@@ -357,9 +357,7 @@ To enable FlashAttention-2 on the Windows platform, you need to install the prec
 
 <details><summary>For Ascend NPU users</summary>
 
-Join [NPU user group](assets/wechat_npu.jpg).
-
-To install LLaMA Factory on Ascend NPU devices, please specify extra dependencies: `pip install -e '.[torch-npu,metrics]'`. Additionally, you need to install the **[Ascend CANN Toolkit and Kernels](https://www.hiascend.com/developer/download/community/result?module=cann)**. Please follow the [installation tutorial](https://www.hiascend.com/document/detail/en/CANNCommunityEdition/600alphaX/softwareinstall/instg/atlasdeploy_03_0031.html) or use the following commands:
+To install LLaMA Factory on Ascend NPU devices, please specify extra dependencies: `pip install -e ".[torch-npu,metrics]"`. Additionally, you need to install the **[Ascend CANN Toolkit and Kernels](https://www.hiascend.com/developer/download/community/result?module=cann)**. Please follow the [installation tutorial](https://www.hiascend.com/document/detail/en/CANNCommunityEdition/600alphaX/softwareinstall/instg/atlasdeploy_03_0031.html) or use the following commands:
 
 ```bash
 # replace the url according to your CANN version and devices
@@ -382,15 +380,12 @@ source /usr/local/Ascend/ascend-toolkit/set_env.sh
 | torch-npu    | 2.1.0   | 2.1.0.post3 |
 | deepspeed    | 0.13.2  | 0.13.2      |
 
-Docker image:
-
-- 32GB: [Download page](http://mirrors.cn-central-221.ovaijisuan.com/detail/130.html)
-- 64GB: [Download page](http://mirrors.cn-central-221.ovaijisuan.com/detail/131.html)
-
 Remember to use `ASCEND_RT_VISIBLE_DEVICES` instead of `CUDA_VISIBLE_DEVICES` to specify the device to use.
 
 If you cannot infer model on NPU devices, try setting `do_sample: false` in the configurations.
 
+Download the pre-built Docker images: [32GB](http://mirrors.cn-central-221.ovaijisuan.com/detail/130.html) | [64GB](http://mirrors.cn-central-221.ovaijisuan.com/detail/131.html)
+
 </details>
 
 ### Data Preparation
@@ -405,9 +400,9 @@ Please refer to [data/README.md](data/README.md) for checking the details about
 Use the following 3 commands to run LoRA **fine-tuning**, **inference** and **merging** of the Llama3-8B-Instruct model, respectively.
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_sft.yaml
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli chat examples/inference/llama3_lora_sft.yaml
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli export examples/merge_lora/llama3_lora_sft.yaml
+llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
+llamafactory-cli chat examples/inference/llama3_lora_sft.yaml
+llamafactory-cli export examples/merge_lora/llama3_lora_sft.yaml
 ```
 
 See [examples/README.md](examples/README.md) for advanced usage (including distributed training).
@@ -417,34 +412,89 @@ See [examples/README.md](examples/README.md) for advanced usage (including distr
 
 ### Fine-Tuning with LLaMA Board GUI (powered by [Gradio](https://github.com/gradio-app/gradio))
 
-#### Use local environment
-
 ```bash
-CUDA_VISIBLE_DEVICES=0 GRADIO_SHARE=1 llamafactory-cli webui
+llamafactory-cli webui
 ```
 
-</details>
+### Build Docker
 
-#### Use Docker
+For CUDA users:
 
 ```bash
-docker build -f ./Dockerfile -t llama-factory:latest .
-docker run --gpus=all \
-    -v ./hf_cache:/root/.cache/huggingface/ \
+cd docker/docker-cuda/
+docker-compose up -d
+docker-compose exec llamafactory bash
+```
+
+For Ascend NPU users:
+
+```bash
+cd docker/docker-npu/
+docker-compose up -d
+docker-compose exec llamafactory bash
+```
+
+<details><summary>Build without Docker Compose</summary>
+
+For CUDA users:
+
+```bash
+docker build -f ./docker/docker-cuda/Dockerfile \
+    --build-arg INSTALL_BNB=false \
+    --build-arg INSTALL_VLLM=false \
+    --build-arg INSTALL_DEEPSPEED=false \
+    --build-arg INSTALL_FLASHATTN=false \
+    --build-arg PIP_INDEX=https://pypi.org/simple \
+    -t llamafactory:latest .
+
+docker run -dit --gpus=all \
+    -v ./hf_cache:/root/.cache/huggingface \
+    -v ./ms_cache:/root/.cache/modelscope \
     -v ./data:/app/data \
     -v ./output:/app/output \
     -p 7860:7860 \
+    -p 8000:8000 \
     --shm-size 16G \
-    --name llama_factory \
-    -d llama-factory:latest
+    --name llamafactory \
+    llamafactory:latest
+
+docker exec -it llamafactory bash
 ```
 
-#### Use Docker Compose
+For Ascend NPU users:
 
 ```bash
-docker compose -f ./docker-compose.yml up -d
+# Choose docker image upon your environment
+docker build -f ./docker/docker-npu/Dockerfile \
+    --build-arg INSTALL_DEEPSPEED=false \
+    --build-arg PIP_INDEX=https://pypi.org/simple \
+    -t llamafactory:latest .
+
+# Change `device` upon your resources
+docker run -dit \
+    -v ./hf_cache:/root/.cache/huggingface \
+    -v ./ms_cache:/root/.cache/modelscope \
+    -v ./data:/app/data \
+    -v ./output:/app/output \
+    -v /usr/local/dcmi:/usr/local/dcmi \
+    -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
+    -v /usr/local/Ascend/driver:/usr/local/Ascend/driver \
+    -v /etc/ascend_install.info:/etc/ascend_install.info \
+    -p 7860:7860 \
+    -p 8000:8000 \
+    --device /dev/davinci0 \
+    --device /dev/davinci_manager \
+    --device /dev/devmm_svm \
+    --device /dev/hisi_hdc \
+    --shm-size 16G \
+    --name llamafactory \
+    llamafactory:latest
+
+docker exec -it llamafactory bash
 ```
 
+</details>
+
 <details><summary>Details about volume</summary>
 
 - hf_cache: Utilize Hugging Face cache on the host machine. Reassignable if a cache already exists in a different directory.
@@ -456,7 +506,7 @@ docker compose -f ./docker-compose.yml up -d
 ### Deploy with OpenAI-style API and vLLM
 
 ```bash
-CUDA_VISIBLE_DEVICES=0,1 API_PORT=8000 llamafactory-cli api examples/inference/llama3_vllm.yaml
+API_PORT=8000 llamafactory-cli api examples/inference/llama3_vllm.yaml
 ```
 
 > [!TIP]
@@ -474,7 +524,7 @@ Train the model by specifying a model ID of the ModelScope Hub as the `model_nam
 
 ### Use W&B Logger
 
-To use [Weights & Biases](https://wandb.ai) for logging experimental results, you need to add the following arguments.
+To use [Weights & Biases](https://wandb.ai) for logging experimental results, you need to add the following arguments to yaml files.
 
 ```yaml
 report_to: wandb
@@ -494,38 +544,63 @@ If you have a project that should be incorporated, please contact via email or c
 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. Wang et al. Know Your Needs Better: Towards Structured Understanding of Marketer Demands with Analogical Reasoning Augmented LLMs. KDD 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. ACL 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. Yi et al. Generation Meets Verification: Accelerating Large Language Model Inference with Smart Parallel Auto-Correct Decoding. ACL 2024 Findings. [[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. Yu et al. KIEval: A Knowledge-grounded Interactive Evaluation Framework for Large Language Models. ACL 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. Wu et al. Large Language Models are Parallel Multilingual Learners. 2024. [[arxiv]](https://arxiv.org/abs/2403.09073)
 1. Zhang et al. EDT: Improving Large Language Models' Generation by Entropy-based Dynamic Temperature Sampling. 2024. [[arxiv]](https://arxiv.org/abs/2403.14541)
 1. Weller et al. FollowIR: Evaluating and Teaching Information Retrieval Models to Follow Instructions. 2024. [[arxiv]](https://arxiv.org/abs/2403.15246)
-1. Hongbin Na. CBT-LLM: A Chinese Large Language Model for Cognitive Behavioral Therapy-based Mental Health Question Answering. 2024. [[arxiv]](https://arxiv.org/abs/2403.16008)
+1. Hongbin Na. CBT-LLM: A Chinese Large Language Model for Cognitive Behavioral Therapy-based Mental Health Question Answering. COLING 2024. [[arxiv]](https://arxiv.org/abs/2403.16008)
 1. Zan et al. CodeS: Natural Language to Code Repository via Multi-Layer Sketch. 2024. [[arxiv]](https://arxiv.org/abs/2403.16443)
 1. Liu et al. Extensive Self-Contrast Enables Feedback-Free Language Model Alignment. 2024. [[arxiv]](https://arxiv.org/abs/2404.00604)
 1. Luo et al. BAdam: A Memory Efficient Full Parameter Training Method for Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2404.02827)
 1. Du et al. Chinese Tiny LLM: Pretraining a Chinese-Centric Large Language Model. 2024. [[arxiv]](https://arxiv.org/abs/2404.04167)
-1. Ma et al. Parameter Efficient Quasi-Orthogonal Fine-Tuning via Givens Rotation. 2024. [[arxiv]](https://arxiv.org/abs/2404.04316)
+1. Ma et al. Parameter Efficient Quasi-Orthogonal Fine-Tuning via Givens Rotation. ICML 2024. [[arxiv]](https://arxiv.org/abs/2404.04316)
 1. Liu et al. Dynamic Generation of Personalities with Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2404.07084)
 1. Shang et al. How Far Have We Gone in Stripped Binary Code Understanding Using Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2404.09836)
 1. Huang et al. LLMTune: Accelerate Database Knob Tuning with Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2404.11581)
 1. Deng et al. Text-Tuple-Table: Towards Information Integration in Text-to-Table Generation via Global Tuple Extraction. 2024. [[arxiv]](https://arxiv.org/abs/2404.14215)
 1. Acikgoz et al. Hippocrates: An Open-Source Framework for Advancing Large Language Models in Healthcare. 2024. [[arxiv]](https://arxiv.org/abs/2404.16621)
-1. Zhang et al. Small Language Models Need Strong Verifiers to Self-Correct Reasoning. 2024. [[arxiv]](https://arxiv.org/abs/2404.17140)
-1. Zhou et al. FREB-TQA: A Fine-Grained Robustness Evaluation Benchmark for Table Question Answering. 2024. [[arxiv]](https://arxiv.org/abs/2404.18585)
+1. Zhang et al. Small Language Models Need Strong Verifiers to Self-Correct Reasoning. ACL 2024 Findings. [[arxiv]](https://arxiv.org/abs/2404.17140)
+1. Zhou et al. FREB-TQA: A Fine-Grained Robustness Evaluation Benchmark for Table Question Answering. NAACL 2024. [[arxiv]](https://arxiv.org/abs/2404.18585)
+1. Xu et al. Large Language Models for Cyber Security: A Systematic Literature Review. 2024. [[arxiv]](https://arxiv.org/abs/2405.04760)
+1. Dammu et al. "They are uncultured": Unveiling Covert Harms and Social Threats in LLM Generated Conversations. 2024. [[arxiv]](https://arxiv.org/abs/2405.05378)
+1. Yi et al. A safety realignment framework via subspace-oriented model fusion for large language models. 2024. [[arxiv]](https://arxiv.org/abs/2405.09055)
+1. Lou et al. SPO: Multi-Dimensional Preference Sequential Alignment With Implicit Reward Modeling. 2024. [[arxiv]](https://arxiv.org/abs/2405.12739)
+1. Zhang et al. Getting More from Less: Large Language Models are Good Spontaneous Multilingual Learners. 2024. [[arxiv]](https://arxiv.org/abs/2405.13816)
+1. Zhang et al. TS-Align: A Teacher-Student Collaborative Framework for Scalable Iterative Finetuning of Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2405.20215)
+1. Zihong Chen. Sentence Segmentation and Sentence Punctuation Based on XunziALLM. 2024. [[paper]](https://aclanthology.org/2024.lt4hala-1.30)
+1. Gao et al. The Best of Both Worlds: Toward an Honest and Helpful Large Language Model. 2024. [[arxiv]](https://arxiv.org/abs/2406.00380)
+1. Wang and Song. MARS: Benchmarking the Metaphysical Reasoning Abilities of Language Models with a Multi-task Evaluation Dataset. 2024. [[arxiv]](https://arxiv.org/abs/2406.02106)
+1. Hu et al. Computational Limits of Low-Rank Adaptation (LoRA) for Transformer-Based Models. 2024. [[arxiv]](https://arxiv.org/abs/2406.03136)
+1. Ge et al. Time Sensitive Knowledge Editing through Efficient Finetuning. ACL 2024. [[arxiv]](https://arxiv.org/abs/2406.04496)
+1. Tan et al. Peer Review as A Multi-Turn and Long-Context Dialogue with Role-Based Interactions. 2024. [[arxiv]](https://arxiv.org/abs/2406.05688)
+1. Song et al. Turbo Sparse: Achieving LLM SOTA Performance with Minimal Activated Parameters. 2024. [[arxiv]](https://arxiv.org/abs/2406.05955)
+1. Gu et al. RWKV-CLIP: A Robust Vision-Language Representation Learner. 2024. [[arxiv]](https://arxiv.org/abs/2406.06973)
+1. Chen et al. Advancing Tool-Augmented Large Language Models: Integrating Insights from Errors in Inference Trees. 2024. [[arxiv]](https://arxiv.org/abs/2406.07115)
+1. Zhu et al. Are Large Language Models Good Statisticians?. 2024. [[arxiv]](https://arxiv.org/abs/2406.07815)
+1. Li et al. Know the Unknown: An Uncertainty-Sensitive Method for LLM Instruction Tuning. 2024. [[arxiv]](https://arxiv.org/abs/2406.10099)
+1. Ding et al. IntentionQA: A Benchmark for Evaluating Purchase Intention Comprehension Abilities of Language Models in E-commerce. 2024. [[arxiv]](https://arxiv.org/abs/2406.10173)
+1. He et al. COMMUNITY-CROSS-INSTRUCT: Unsupervised Instruction Generation for Aligning Large Language Models to Online Communities. 2024. [[arxiv]](https://arxiv.org/abs/2406.12074)
+1. Lin et al. FVEL: Interactive Formal Verification Environment with Large Language Models via Theorem Proving. 2024. [[arxiv]](https://arxiv.org/abs/2406.14408)
+1. Treutlein et al. Connecting the Dots: LLMs can Infer and Verbalize Latent Structure from Disparate Training Data. 2024. [[arxiv]](https://arxiv.org/abs/2406.14546)
+1. Feng et al. SS-Bench: A Benchmark for Social Story Generation and Evaluation. 2024. [[arxiv]](https://arxiv.org/abs/2406.15695)
+1. Feng et al. Self-Constructed Context Decompilation with Fined-grained Alignment Enhancement. 2024. [[arxiv]](https://arxiv.org/abs/2406.17233)
+1. Liu et al. Large Language Models for Cuffless Blood Pressure Measurement From Wearable Biosignals. 2024. [[arxiv]](https://arxiv.org/abs/2406.18069)
+1. Iyer et al. Exploring Very Low-Resource Translation with LLMs: The University of Edinburgh’s Submission to AmericasNLP 2024 Translation Task. AmericasNLP 2024. [[paper]](https://aclanthology.org/2024.americasnlp-1.25)
 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/X-D-Lab/Sunsimiao)**: A large language model specialized in Chinese medical domain, based on Baichuan-7B and ChatGLM-6B.
@@ -533,6 +608,9 @@ If you have a project that should be incorporated, please contact via email or c
 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.
 1. **[Luminia-13B-v3](https://huggingface.co/Nekochu/Luminia-13B-v3)**: A large language model specialized in generate metadata for stable diffusion. [[🤗Demo]](https://huggingface.co/spaces/Nekochu/Luminia-13B_SD_Prompt)
 1. **[Chinese-LLaVA-Med](https://github.com/BUAADreamer/Chinese-LLaVA-Med)**: A multimodal large language model specialized in Chinese medical domain, based on LLaVA-1.5-7B.
+1. **[AutoRE](https://github.com/THUDM/AutoRE)**: A document-level relation extraction system based on large language models.
+1. **[NVIDIA RTX AI Toolkit](https://github.com/NVIDIA/RTX-AI-Toolkit)**: SDKs for fine-tuning LLMs on Windows PC for NVIDIA RTX.
+1. **[LazyLLM](https://github.com/LazyAGI/LazyLLM)**: An easy and lazy way for building multi-agent LLMs applications and supports model fine-tuning via LLaMA Factory.
 
 </details>
 
@@ -540,17 +618,19 @@ If you have a project that should be incorporated, please contact via email or c
 
 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) / [Command-R](https://cohere.com/c4ai-cc-by-nc-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) / [GLM4](https://huggingface.co/THUDM/glm-4-9b/blob/main/LICENSE) / [InternLM2](https://github.com/InternLM/InternLM#license) / [LLaMA](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md) / [LLaMA-2 (LLaVA-1.5)](https://ai.meta.com/llama/license/) / [LLaMA-3](https://llama.meta.com/llama3/license/) / [Mistral](LICENSE) / [OLMo](LICENSE) / [Phi-1.5/2](https://huggingface.co/microsoft/phi-1_5/resolve/main/Research%20License.docx) / [Phi-3](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/blob/main/LICENSE) / [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) / [Yi-1.5](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: [Baichuan 2](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) / [Command R](https://cohere.com/c4ai-cc-by-nc-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) / [GLM-4](https://huggingface.co/THUDM/glm-4-9b/blob/main/LICENSE) / [InternLM2](https://github.com/InternLM/InternLM#license) / [Llama](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md) / [Llama 2 (LLaVA-1.5)](https://ai.meta.com/llama/license/) / [Llama 3](https://llama.meta.com/llama3/license/) / [Mistral](LICENSE) / [OLMo](LICENSE) / [Phi-1.5/Phi-2](https://huggingface.co/microsoft/phi-1_5/resolve/main/Research%20License.docx) / [Phi-3](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/blob/main/LICENSE) / [Qwen](https://github.com/QwenLM/Qwen/blob/main/Tongyi%20Qianwen%20LICENSE%20AGREEMENT) / [StarCoder 2](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) / [Yi-1.5](LICENSE) / [Yuan 2](https://github.com/IEIT-Yuan/Yuan-2.0/blob/main/LICENSE-Yuan)
 
 ## Citation
 
 If this work is helpful, please kindly cite as:
 
 ```bibtex
-@article{zheng2024llamafactory,
+@inproceedings{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},
+  author={Yaowei Zheng and Richong Zhang and Junhao Zhang and Yanhan Ye and Zheyan Luo and Zhangchi Feng and Yongqiang Ma},
+  booktitle={Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations)},
+  address={Bangkok, Thailand},
+  publisher={Association for Computational Linguistics},
   year={2024},
   url={http://arxiv.org/abs/2403.13372}
 }

README_zh.md

@@ -4,7 +4,7 @@
 [![GitHub Code License](https://img.shields.io/github/license/hiyouga/LLaMA-Factory)](LICENSE)
 [![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/llamafactory)](https://pypi.org/project/llamafactory/)
-[![Citation](https://img.shields.io/badge/citation-44-green)](#使用了-llama-factory-的项目)
+[![Citation](https://img.shields.io/badge/citation-72-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)
 [![Twitter](https://img.shields.io/twitter/follow/llamafactory_ai)](https://twitter.com/llamafactory_ai)
@@ -15,7 +15,7 @@
 
 [![GitHub Tread](https://trendshift.io/api/badge/repositories/4535)](https://trendshift.io/repositories/4535)
 
-👋 加入我们的[微信群](assets/wechat.jpg)。
+👋 加入我们的[微信群](assets/wechat.jpg)或 [NPU 用户群](assets/wechat_npu.jpg)。
 
 \[ [English](README.md) | 中文 \]
 
@@ -48,8 +48,8 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/ec36a9dd-37f4-4f72-81bd
 
 - **多种模型**：LLaMA、LLaVA、Mistral、Mixtral-MoE、Qwen、Yi、Gemma、Baichuan、ChatGLM、Phi 等等。
 - **集成方法**：（增量）预训练、（多模态）指令监督微调、奖励模型训练、PPO 训练、DPO 训练、KTO 训练、ORPO 训练等等。
-- **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、BAdam、DoRA、LongLoRA、LLaMA Pro、Mixture-of-Depths、LoRA+、LoftQ 和 Agent 微调。
+- **多种精度**：16 比特全参数微调、冻结微调、LoRA 微调和基于 AQLM/AWQ/GPTQ/LLM.int8/HQQ/EETQ 的 2/3/4/5/6/8 比特 QLoRA 微调。
+- **先进算法**：GaLore、BAdam、DoRA、LongLoRA、LLaMA Pro、Mixture-of-Depths、LoRA+、LoftQ、PiSSA 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。
@@ -71,9 +71,9 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/ec36a9dd-37f4-4f72-81bd
 
 ## 更新日志
 
-[24/06/07] 我们支持了 **[Qwen-2](https://qwenlm.github.io/blog/qwen2/)** 系列模型的微调。
+[24/06/16] 我们支持了 **[PiSSA](https://arxiv.org/abs/2404.02948)** 算法。详细用法请参照 [examples](examples/README_zh.md)。
 
-[24/06/05] 我们支持了 **[GLM-4-9B/GLM-4-9B-Chat](https://github.com/THUDM/GLM-4)** 模型的微调。
+[24/06/07] 我们支持了 **[Qwen2](https://qwenlm.github.io/blog/qwen2/)** 和 **[GLM-4](https://github.com/THUDM/GLM-4)** 模型的微调。
 
 [24/05/26] 我们支持了 **[SimPO](https://arxiv.org/abs/2405.14734)** 偏好对齐算法。详细用法请参照 [examples](examples/README_zh.md)。
 
@@ -151,35 +151,32 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/ec36a9dd-37f4-4f72-81bd
 
 ## 模型
 
-| 模型名                                                   | 模型大小                          | Template  |
-| -------------------------------------------------------- | -------------------------------- | --------- |
-| [Baichuan2](https://huggingface.co/baichuan-inc)         | 7B/13B                           | baichuan2 |
-| [BLOOM](https://huggingface.co/bigscience)               | 560M/1.1B/1.7B/3B/7.1B/176B      | -         |
-| [BLOOMZ](https://huggingface.co/bigscience)              | 560M/1.1B/1.7B/3B/7.1B/176B      | -         |
-| [ChatGLM3](https://huggingface.co/THUDM)                 | 6B                               | chatglm3  |
-| [Command-R](https://huggingface.co/CohereForAI)          | 35B/104B                         | cohere    |
-| [DeepSeek (MoE)](https://huggingface.co/deepseek-ai)     | 7B/16B/67B/236B                  | deepseek  |
-| [Falcon](https://huggingface.co/tiiuae)                  | 7B/11B/40B/180B                  | falcon    |
-| [Gemma/CodeGemma](https://huggingface.co/google)         | 2B/7B                            | gemma     |
-| [GLM4](https://huggingface.co/THUDM)                     | 9B                               | glm4      |
-| [InternLM2](https://huggingface.co/internlm)             | 7B/20B                           | intern2   |
-| [LLaMA](https://github.com/facebookresearch/llama)       | 7B/13B/33B/65B                   | -         |
-| [LLaMA-2](https://huggingface.co/meta-llama)             | 7B/13B/70B                       | llama2    |
-| [LLaMA-3](https://huggingface.co/meta-llama)             | 8B/70B                           | llama3    |
-| [LLaVA-1.5](https://huggingface.co/llava-hf)             | 7B/13B                           | vicuna    |
-| [Mistral/Mixtral](https://huggingface.co/mistralai)      | 7B/8x7B/8x22B                    | mistral   |
-| [OLMo](https://huggingface.co/allenai)                   | 1B/7B                            | -         |
-| [PaliGemma](https://huggingface.co/google)               | 3B                               | gemma     |
-| [Phi-1.5/2](https://huggingface.co/microsoft)            | 1.3B/2.7B                        | -         |
-| [Phi-3](https://huggingface.co/microsoft)                | 4B/7B/14B                        | phi       |
-| [Qwen](https://huggingface.co/Qwen)                      | 1.8B/7B/14B/72B                  | qwen      |
-| [Qwen1.5 (Code/MoE)](https://huggingface.co/Qwen)        | 0.5B/1.8B/4B/7B/14B/32B/72B/110B | qwen      |
-| [Qwen2 (MoE)](https://huggingface.co/Qwen)               | 0.5B/1.5B/7B/57B/72B             | qwen      |
-| [StarCoder2](https://huggingface.co/bigcode)             | 3B/7B/15B                        | -         |
-| [XVERSE](https://huggingface.co/xverse)                  | 7B/13B/65B                       | xverse    |
-| [Yi (1/1.5)](https://huggingface.co/01-ai)               | 6B/9B/34B                        | yi        |
-| [Yi-VL](https://huggingface.co/01-ai)                    | 6B/34B                           | yi_vl     |
-| [Yuan](https://huggingface.co/IEITYuan)                  | 2B/51B/102B                      | yuan      |
+| 模型名                                                       | 模型大小                          | Template  |
+| ------------------------------------------------------------ | -------------------------------- | --------- |
+| [Baichuan 2](https://huggingface.co/baichuan-inc)            | 7B/13B                           | baichuan2 |
+| [BLOOM/BLOOMZ](https://huggingface.co/bigscience)            | 560M/1.1B/1.7B/3B/7.1B/176B      | -         |
+| [ChatGLM3](https://huggingface.co/THUDM)                     | 6B                               | chatglm3  |
+| [Command R](https://huggingface.co/CohereForAI)              | 35B/104B                         | cohere    |
+| [DeepSeek (Code/MoE)](https://huggingface.co/deepseek-ai)    | 7B/16B/67B/236B                  | deepseek  |
+| [Falcon](https://huggingface.co/tiiuae)                      | 7B/11B/40B/180B                  | falcon    |
+| [Gemma/Gemma 2/CodeGemma](https://huggingface.co/google)     | 2B/7B/9B/27B                     | gemma     |
+| [GLM-4](https://huggingface.co/THUDM)                        | 9B                               | glm4      |
+| [InternLM2](https://huggingface.co/internlm)                 | 7B/20B                           | intern2   |
+| [Llama](https://github.com/facebookresearch/llama)           | 7B/13B/33B/65B                   | -         |
+| [Llama 2](https://huggingface.co/meta-llama)                 | 7B/13B/70B                       | llama2    |
+| [Llama 3](https://huggingface.co/meta-llama)                 | 8B/70B                           | llama3    |
+| [LLaVA-1.5](https://huggingface.co/llava-hf)                 | 7B/13B                           | vicuna    |
+| [Mistral/Mixtral](https://huggingface.co/mistralai)          | 7B/8x7B/8x22B                    | mistral   |
+| [OLMo](https://huggingface.co/allenai)                       | 1B/7B                            | -         |
+| [PaliGemma](https://huggingface.co/google)                   | 3B                               | gemma     |
+| [Phi-1.5/Phi-2](https://huggingface.co/microsoft)            | 1.3B/2.7B                        | -         |
+| [Phi-3](https://huggingface.co/microsoft)                    | 4B/7B/14B                        | phi       |
+| [Qwen/Qwen1.5/Qwen2 (Code/MoE)](https://huggingface.co/Qwen) | 0.5B/1.5B/4B/7B/14B/32B/72B/110B | qwen      |
+| [StarCoder 2](https://huggingface.co/bigcode)                | 3B/7B/15B                        | -         |
+| [XVERSE](https://huggingface.co/xverse)                      | 7B/13B/65B                       | xverse    |
+| [Yi/Yi-1.5](https://huggingface.co/01-ai)                    | 6B/9B/34B                        | yi        |
+| [Yi-VL](https://huggingface.co/01-ai)                        | 6B/34B                           | yi_vl     |
+| [Yuan 2](https://huggingface.co/IEITYuan)                    | 2B/51B/102B                      | yuan      |
 
 > [!NOTE]
 > 对于所有“基座”（Base）模型，`template` 参数可以是 `default`, `alpaca`, `vicuna` 等任意值。但“对话”（Instruct/Chat）模型请务必使用**对应的模板**。
@@ -259,6 +256,9 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/ec36a9dd-37f4-4f72-81bd
 - [Cosmopedia (en)](https://huggingface.co/datasets/HuggingFaceTB/cosmopedia)
 - [STEM (zh)](https://huggingface.co/datasets/hfl/stem_zh_instruction)
 - [Ruozhiba (zh)](https://huggingface.co/datasets/hfl/ruozhiba_gpt4_turbo)
+- [Neo-sft (zh)](https://huggingface.co/datasets/m-a-p/neo_sft_phase2)
+- [WebInstructSub (en)](https://huggingface.co/datasets/TIGER-Lab/WebInstructSub)
+- [Magpie-Pro-300K-Filtered (en)](https://huggingface.co/datasets/Magpie-Align/Magpie-Pro-300K-Filtered)
 - [LLaVA mixed (en&zh)](https://huggingface.co/datasets/BUAADreamer/llava-en-zh-300k)
 - [Open Assistant (de)](https://huggingface.co/datasets/mayflowergmbh/oasst_de)
 - [Dolly 15k (de)](https://huggingface.co/datasets/mayflowergmbh/dolly-15k_de)
@@ -335,10 +335,10 @@ huggingface-cli login
 ```bash
 git clone --depth 1 https://github.com/hiyouga/LLaMA-Factory.git
 cd LLaMA-Factory
-pip install -e '.[torch,metrics]'
+pip install -e ".[torch,metrics]"
 ```
 
-可选的额外依赖项：torch、torch_npu、metrics、deepspeed、bitsandbytes、vllm、galore、badam、gptq、awq、aqlm、qwen、modelscope、quality
+可选的额外依赖项：torch、torch-npu、metrics、deepspeed、bitsandbytes、hqq、eetq、gptq、awq、aqlm、vllm、galore、badam、qwen、modelscope、quality
 
 > [!TIP]
 > 遇到包冲突时，可使用 `pip install --no-deps -e .` 解决。
@@ -357,9 +357,7 @@ pip install https://github.com/jllllll/bitsandbytes-windows-webui/releases/downl
 
 <details><summary>昇腾 NPU 用户指南</summary>
 
-加入 [NPU 用户群](assets/wechat_npu.jpg)。
-
-在昇腾 NPU 设备上安装 LLaMA Factory 时，需要指定额外依赖项，使用 `pip install -e '.[torch-npu,metrics]'` 命令安装。此外，还需要安装 **[Ascend CANN Toolkit and Kernels](https://www.hiascend.com/developer/download/community/result?module=cann)**，安装方法请参考[安装教程](https://www.hiascend.com/document/detail/zh/CANNCommunityEdition/80RC2alpha002/quickstart/quickstart/quickstart_18_0004.html)或使用以下命令：
+在昇腾 NPU 设备上安装 LLaMA Factory 时，需要指定额外依赖项，使用 `pip install -e ".[torch-npu,metrics]"` 命令安装。此外，还需要安装 **[Ascend CANN Toolkit 与 Kernels](https://www.hiascend.com/developer/download/community/result?module=cann)**，安装方法请参考[安装教程](https://www.hiascend.com/document/detail/zh/CANNCommunityEdition/80RC2alpha002/quickstart/quickstart/quickstart_18_0004.html)或使用以下命令：
 
 ```bash
 # 请替换 URL 为 CANN 版本和设备型号对应的 URL
@@ -382,15 +380,12 @@ source /usr/local/Ascend/ascend-toolkit/set_env.sh
 | torch-npu    | 2.1.0   | 2.1.0.post3 |
 | deepspeed    | 0.13.2  | 0.13.2      |
 
-Docker 镜像：
-
-- 32GB：[下载地址](http://mirrors.cn-central-221.ovaijisuan.com/detail/130.html)
-- 64GB：[下载地址](http://mirrors.cn-central-221.ovaijisuan.com/detail/131.html)
-
 请使用 `ASCEND_RT_VISIBLE_DEVICES` 而非 `CUDA_VISIBLE_DEVICES` 来指定运算设备。
 
 如果遇到无法正常推理的情况，请尝试设置 `do_sample: false`。
 
+下载预构建 Docker 镜像：[32GB](http://mirrors.cn-central-221.ovaijisuan.com/detail/130.html) | [64GB](http://mirrors.cn-central-221.ovaijisuan.com/detail/131.html)
+
 </details>
 
 ### 数据准备
@@ -405,9 +400,9 @@ Docker 镜像：
 下面三行命令分别对 Llama3-8B-Instruct 模型进行 LoRA **微调**、**推理**和**合并**。
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_sft.yaml
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli chat examples/inference/llama3_lora_sft.yaml
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli export examples/merge_lora/llama3_lora_sft.yaml
+llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
+llamafactory-cli chat examples/inference/llama3_lora_sft.yaml
+llamafactory-cli export examples/merge_lora/llama3_lora_sft.yaml
 ```
 
 高级用法请参考 [examples/README_zh.md](examples/README_zh.md)（包括多 GPU 微调）。
@@ -417,32 +412,89 @@ CUDA_VISIBLE_DEVICES=0 llamafactory-cli export examples/merge_lora/llama3_lora_s
 
 ### LLaMA Board 可视化微调（由 [Gradio](https://github.com/gradio-app/gradio) 驱动）
 
-#### 使用本地环境
-
 ```bash
-CUDA_VISIBLE_DEVICES=0 GRADIO_SHARE=1 llamafactory-cli webui
+llamafactory-cli webui
 ```
 
-#### 使用 Docker
+### 构建 Docker
+
+CUDA 用户：
 
 ```bash
-docker build -f ./Dockerfile -t llama-factory:latest .
-docker run --gpus=all \
-    -v ./hf_cache:/root/.cache/huggingface/ \
+cd docker/docker-cuda/
+docker-compose up -d
+docker-compose exec llamafactory bash
+```
+
+昇腾 NPU 用户：
+
+```bash
+cd docker/docker-npu/
+docker-compose up -d
+docker-compose exec llamafactory bash
+```
+
+<details><summary>不使用 Docker Compose 构建</summary>
+
+CUDA 用户：
+
+```bash
+docker build -f ./docker/docker-cuda/Dockerfile \
+    --build-arg INSTALL_BNB=false \
+    --build-arg INSTALL_VLLM=false \
+    --build-arg INSTALL_DEEPSPEED=false \
+    --build-arg INSTALL_FLASHATTN=false \
+    --build-arg PIP_INDEX=https://pypi.org/simple \
+    -t llamafactory:latest .
+
+docker run -dit --gpus=all \
+    -v ./hf_cache:/root/.cache/huggingface \
+    -v ./ms_cache:/root/.cache/modelscope \
     -v ./data:/app/data \
     -v ./output:/app/output \
     -p 7860:7860 \
+    -p 8000:8000 \
     --shm-size 16G \
-    --name llama_factory \
-    -d llama-factory:latest
+    --name llamafactory \
+    llamafactory:latest
+
+docker exec -it llamafactory bash
 ```
 
-#### 使用 Docker Compose
+昇腾 NPU 用户：
 
 ```bash
-docker compose -f ./docker-compose.yml up -d
+# 根据您的环境选择镜像
+docker build -f ./docker/docker-npu/Dockerfile \
+    --build-arg INSTALL_DEEPSPEED=false \
+    --build-arg PIP_INDEX=https://pypi.org/simple \
+    -t llamafactory:latest .
+
+# 根据您的资源更改 `device`
+docker run -dit \
+    -v ./hf_cache:/root/.cache/huggingface \
+    -v ./ms_cache:/root/.cache/modelscope \
+    -v ./data:/app/data \
+    -v ./output:/app/output \
+    -v /usr/local/dcmi:/usr/local/dcmi \
+    -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
+    -v /usr/local/Ascend/driver:/usr/local/Ascend/driver \
+    -v /etc/ascend_install.info:/etc/ascend_install.info \
+    -p 7860:7860 \
+    -p 8000:8000 \
+    --device /dev/davinci0 \
+    --device /dev/davinci_manager \
+    --device /dev/devmm_svm \
+    --device /dev/hisi_hdc \
+    --shm-size 16G \
+    --name llamafactory \
+    llamafactory:latest
+
+docker exec -it llamafactory bash
 ```
 
+</details>
+
 <details><summary>数据卷详情</summary>
 
 - hf_cache：使用宿主机的 Hugging Face 缓存文件夹，允许更改为新的目录。
@@ -454,7 +506,7 @@ docker compose -f ./docker-compose.yml up -d
 ### 利用 vLLM 部署 OpenAI API
 
 ```bash
-CUDA_VISIBLE_DEVICES=0,1 API_PORT=8000 llamafactory-cli api examples/inference/llama3_vllm.yaml
+API_PORT=8000 llamafactory-cli api examples/inference/llama3_vllm.yaml
 ```
 
 > [!TIP]
@@ -472,7 +524,7 @@ export USE_MODELSCOPE_HUB=1 # Windows 使用 `set USE_MODELSCOPE_HUB=1`
 
 ### 使用 W&B 面板
 
-若要使用 [Weights & Biases](https://wandb.ai) 记录实验数据，请添加下面的参数。
+若要使用 [Weights & Biases](https://wandb.ai) 记录实验数据，请在 yaml 文件中添加下面的参数。
 
 ```yaml
 report_to: wandb
@@ -492,38 +544,63 @@ run_name: test_run # 可选
 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. Wang et al. Know Your Needs Better: Towards Structured Understanding of Marketer Demands with Analogical Reasoning Augmented LLMs. KDD 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. ACL 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. Yi et al. Generation Meets Verification: Accelerating Large Language Model Inference with Smart Parallel Auto-Correct Decoding. ACL 2024 Findings. [[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. Yu et al. KIEval: A Knowledge-grounded Interactive Evaluation Framework for Large Language Models. ACL 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. Wu et al. Large Language Models are Parallel Multilingual Learners. 2024. [[arxiv]](https://arxiv.org/abs/2403.09073)
 1. Zhang et al. EDT: Improving Large Language Models' Generation by Entropy-based Dynamic Temperature Sampling. 2024. [[arxiv]](https://arxiv.org/abs/2403.14541)
 1. Weller et al. FollowIR: Evaluating and Teaching Information Retrieval Models to Follow Instructions. 2024. [[arxiv]](https://arxiv.org/abs/2403.15246)
-1. Hongbin Na. CBT-LLM: A Chinese Large Language Model for Cognitive Behavioral Therapy-based Mental Health Question Answering. 2024. [[arxiv]](https://arxiv.org/abs/2403.16008)
+1. Hongbin Na. CBT-LLM: A Chinese Large Language Model for Cognitive Behavioral Therapy-based Mental Health Question Answering. COLING 2024. [[arxiv]](https://arxiv.org/abs/2403.16008)
 1. Zan et al. CodeS: Natural Language to Code Repository via Multi-Layer Sketch. 2024. [[arxiv]](https://arxiv.org/abs/2403.16443)
 1. Liu et al. Extensive Self-Contrast Enables Feedback-Free Language Model Alignment. 2024. [[arxiv]](https://arxiv.org/abs/2404.00604)
 1. Luo et al. BAdam: A Memory Efficient Full Parameter Training Method for Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2404.02827)
 1. Du et al. Chinese Tiny LLM: Pretraining a Chinese-Centric Large Language Model. 2024. [[arxiv]](https://arxiv.org/abs/2404.04167)
-1. Ma et al. Parameter Efficient Quasi-Orthogonal Fine-Tuning via Givens Rotation. 2024. [[arxiv]](https://arxiv.org/abs/2404.04316)
+1. Ma et al. Parameter Efficient Quasi-Orthogonal Fine-Tuning via Givens Rotation. ICML 2024. [[arxiv]](https://arxiv.org/abs/2404.04316)
 1. Liu et al. Dynamic Generation of Personalities with Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2404.07084)
 1. Shang et al. How Far Have We Gone in Stripped Binary Code Understanding Using Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2404.09836)
 1. Huang et al. LLMTune: Accelerate Database Knob Tuning with Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2404.11581)
 1. Deng et al. Text-Tuple-Table: Towards Information Integration in Text-to-Table Generation via Global Tuple Extraction. 2024. [[arxiv]](https://arxiv.org/abs/2404.14215)
 1. Acikgoz et al. Hippocrates: An Open-Source Framework for Advancing Large Language Models in Healthcare. 2024. [[arxiv]](https://arxiv.org/abs/2404.16621)
-1. Zhang et al. Small Language Models Need Strong Verifiers to Self-Correct Reasoning. 2024. [[arxiv]](https://arxiv.org/abs/2404.17140)
-1. Zhou et al. FREB-TQA: A Fine-Grained Robustness Evaluation Benchmark for Table Question Answering. 2024. [[arxiv]](https://arxiv.org/abs/2404.18585)
+1. Zhang et al. Small Language Models Need Strong Verifiers to Self-Correct Reasoning. ACL 2024 Findings. [[arxiv]](https://arxiv.org/abs/2404.17140)
+1. Zhou et al. FREB-TQA: A Fine-Grained Robustness Evaluation Benchmark for Table Question Answering. NAACL 2024. [[arxiv]](https://arxiv.org/abs/2404.18585)
+1. Xu et al. Large Language Models for Cyber Security: A Systematic Literature Review. 2024. [[arxiv]](https://arxiv.org/abs/2405.04760)
+1. Dammu et al. "They are uncultured": Unveiling Covert Harms and Social Threats in LLM Generated Conversations. 2024. [[arxiv]](https://arxiv.org/abs/2405.05378)
+1. Yi et al. A safety realignment framework via subspace-oriented model fusion for large language models. 2024. [[arxiv]](https://arxiv.org/abs/2405.09055)
+1. Lou et al. SPO: Multi-Dimensional Preference Sequential Alignment With Implicit Reward Modeling. 2024. [[arxiv]](https://arxiv.org/abs/2405.12739)
+1. Zhang et al. Getting More from Less: Large Language Models are Good Spontaneous Multilingual Learners. 2024. [[arxiv]](https://arxiv.org/abs/2405.13816)
+1. Zhang et al. TS-Align: A Teacher-Student Collaborative Framework for Scalable Iterative Finetuning of Large Language Models. 2024. [[arxiv]](https://arxiv.org/abs/2405.20215)
+1. Zihong Chen. Sentence Segmentation and Sentence Punctuation Based on XunziALLM. 2024. [[paper]](https://aclanthology.org/2024.lt4hala-1.30)
+1. Gao et al. The Best of Both Worlds: Toward an Honest and Helpful Large Language Model. 2024. [[arxiv]](https://arxiv.org/abs/2406.00380)
+1. Wang and Song. MARS: Benchmarking the Metaphysical Reasoning Abilities of Language Models with a Multi-task Evaluation Dataset. 2024. [[arxiv]](https://arxiv.org/abs/2406.02106)
+1. Hu et al. Computational Limits of Low-Rank Adaptation (LoRA) for Transformer-Based Models. 2024. [[arxiv]](https://arxiv.org/abs/2406.03136)
+1. Ge et al. Time Sensitive Knowledge Editing through Efficient Finetuning. ACL 2024. [[arxiv]](https://arxiv.org/abs/2406.04496)
+1. Tan et al. Peer Review as A Multi-Turn and Long-Context Dialogue with Role-Based Interactions. 2024. [[arxiv]](https://arxiv.org/abs/2406.05688)
+1. Song et al. Turbo Sparse: Achieving LLM SOTA Performance with Minimal Activated Parameters. 2024. [[arxiv]](https://arxiv.org/abs/2406.05955)
+1. Gu et al. RWKV-CLIP: A Robust Vision-Language Representation Learner. 2024. [[arxiv]](https://arxiv.org/abs/2406.06973)
+1. Chen et al. Advancing Tool-Augmented Large Language Models: Integrating Insights from Errors in Inference Trees. 2024. [[arxiv]](https://arxiv.org/abs/2406.07115)
+1. Zhu et al. Are Large Language Models Good Statisticians?. 2024. [[arxiv]](https://arxiv.org/abs/2406.07815)
+1. Li et al. Know the Unknown: An Uncertainty-Sensitive Method for LLM Instruction Tuning. 2024. [[arxiv]](https://arxiv.org/abs/2406.10099)
+1. Ding et al. IntentionQA: A Benchmark for Evaluating Purchase Intention Comprehension Abilities of Language Models in E-commerce. 2024. [[arxiv]](https://arxiv.org/abs/2406.10173)
+1. He et al. COMMUNITY-CROSS-INSTRUCT: Unsupervised Instruction Generation for Aligning Large Language Models to Online Communities. 2024. [[arxiv]](https://arxiv.org/abs/2406.12074)
+1. Lin et al. FVEL: Interactive Formal Verification Environment with Large Language Models via Theorem Proving. 2024. [[arxiv]](https://arxiv.org/abs/2406.14408)
+1. Treutlein et al. Connecting the Dots: LLMs can Infer and Verbalize Latent Structure from Disparate Training Data. 2024. [[arxiv]](https://arxiv.org/abs/2406.14546)
+1. Feng et al. SS-Bench: A Benchmark for Social Story Generation and Evaluation. 2024. [[arxiv]](https://arxiv.org/abs/2406.15695)
+1. Feng et al. Self-Constructed Context Decompilation with Fined-grained Alignment Enhancement. 2024. [[arxiv]](https://arxiv.org/abs/2406.17233)
+1. Liu et al. Large Language Models for Cuffless Blood Pressure Measurement From Wearable Biosignals. 2024. [[arxiv]](https://arxiv.org/abs/2406.18069)
+1. Iyer et al. Exploring Very Low-Resource Translation with LLMs: The University of Edinburgh’s Submission to AmericasNLP 2024 Translation Task. AmericasNLP 2024. [[paper]](https://aclanthology.org/2024.americasnlp-1.25)
 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/X-D-Lab/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。
@@ -531,6 +608,9 @@ run_name: test_run # 可选
 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。
 1. **[Luminia-13B-v3](https://huggingface.co/Nekochu/Luminia-13B-v3)**：一个用于生成 Stable Diffusion 提示词的大型语言模型。[[🤗Demo]](https://huggingface.co/spaces/Nekochu/Luminia-13B_SD_Prompt)
 1. **[Chinese-LLaVA-Med](https://github.com/BUAADreamer/Chinese-LLaVA-Med)**：中文多模态医学大模型，基于 LLaVA-1.5-7B 在中文多模态医疗数据上微调而得。
+1. **[AutoRE](https://github.com/THUDM/AutoRE)**：基于大语言模型的文档级关系抽取系统。
+1. **[NVIDIA RTX AI Toolkit](https://github.com/NVIDIA/RTX-AI-Toolkit)**：在 Windows 主机上利用英伟达 RTX 设备进行大型语言模型微调的开发包。
+1. **[LazyLLM](https://github.com/LazyAGI/LazyLLM)**：一个低代码构建多 Agent 大模型应用的开发工具，支持基于 LLaMA Factory 的模型微调.
 
 </details>
 
@@ -538,17 +618,19 @@ run_name: test_run # 可选
 
 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) / [Command-R](https://cohere.com/c4ai-cc-by-nc-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) / [GLM4](https://huggingface.co/THUDM/glm-4-9b/blob/main/LICENSE) / [InternLM2](https://github.com/InternLM/InternLM#license) / [LLaMA](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md) / [LLaMA-2 (LLaVA-1.5)](https://ai.meta.com/llama/license/) / [LLaMA-3](https://llama.meta.com/llama3/license/) / [Mistral](LICENSE) / [OLMo](LICENSE) / [Phi-1.5/2](https://huggingface.co/microsoft/phi-1_5/resolve/main/Research%20License.docx) / [Phi-3](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/blob/main/LICENSE) / [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) / [Yi-1.5](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：[Baichuan 2](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) / [Command R](https://cohere.com/c4ai-cc-by-nc-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) / [GLM-4](https://huggingface.co/THUDM/glm-4-9b/blob/main/LICENSE) / [InternLM2](https://github.com/InternLM/InternLM#license) / [Llama](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md) / [Llama 2 (LLaVA-1.5)](https://ai.meta.com/llama/license/) / [Llama 3](https://llama.meta.com/llama3/license/) / [Mistral](LICENSE) / [OLMo](LICENSE) / [Phi-1.5/Phi-2](https://huggingface.co/microsoft/phi-1_5/resolve/main/Research%20License.docx) / [Phi-3](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/blob/main/LICENSE) / [Qwen](https://github.com/QwenLM/Qwen/blob/main/Tongyi%20Qianwen%20LICENSE%20AGREEMENT) / [StarCoder 2](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) / [Yi-1.5](LICENSE) / [Yuan 2](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
-@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},
+@inproceedings{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 Zhangchi Feng and Yongqiang Ma},
+  booktitle={Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations)},
+  address={Bangkok, Thailand},
+  publisher={Association for Computational Linguistics},
   year={2024},
   url={http://arxiv.org/abs/2403.13372}
 }

data/README.md

@@ -11,8 +11,9 @@ Currently we support datasets in **alpaca** and **sharegpt** format.
   "formatting": "the format of the dataset. (optional, default: alpaca, can be chosen from {alpaca, sharegpt})",
   "ranking": "whether the dataset is a preference dataset or not. (default: False)",
   "subset": "the name of the subset. (optional, default: None)",
+  "split": "the name of dataset split to be used. (optional, default: train)",
   "folder": "the name of the folder of the dataset repository on the Hugging Face hub. (optional, default: None)",
-  "num_samples": "the number of samples in the dataset used for training. (optional, default: None)",
+  "num_samples": "the number of samples in the dataset to be used. (optional, default: None)",
   "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)",

data/README_zh.md

@@ -11,8 +11,9 @@
   "formatting": "数据集格式（可选，默认：alpaca，可以为 alpaca 或 sharegpt）",
   "ranking": "是否为偏好数据集（可选，默认：False）",
   "subset": "数据集子集的名称（可选，默认：None）",
+  "split": "所使用的数据集切分（可选，默认：train）",
   "folder": "Hugging Face 仓库的文件夹名称（可选，默认：None）",
-  "num_samples": "该数据集中用于训练的样本数量。（可选，默认：None）",
+  "num_samples": "该数据集所使用的样本数量。（可选，默认：None）",
   "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）",

docker-compose.yml

@@ -1,23 +0,0 @@
-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

docker/docker-cuda/Dockerfile

@@ -0,0 +1,59 @@
+# Use the NVIDIA official image with PyTorch 2.3.0
+# https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/rel-24-02.html
+FROM nvcr.io/nvidia/pytorch:24.02-py3
+
+# Define environments
+ENV MAX_JOBS=4
+ENV FLASH_ATTENTION_FORCE_BUILD=TRUE
+ENV VLLM_WORKER_MULTIPROC_METHOD=spawn
+
+# Define installation arguments
+ARG INSTALL_BNB=false
+ARG INSTALL_VLLM=false
+ARG INSTALL_DEEPSPEED=false
+ARG INSTALL_FLASHATTN=false
+ARG PIP_INDEX=https://pypi.org/simple
+
+# Set the working directory
+WORKDIR /app
+
+# Install the requirements
+COPY requirements.txt /app
+RUN pip config set global.index-url "$PIP_INDEX" && \
+    pip config set global.extra-index-url "$PIP_INDEX" && \
+    python -m pip install --upgrade pip && \
+    python -m pip install -r requirements.txt
+
+# Copy the rest of the application into the image
+COPY . /app
+
+# Install the LLaMA Factory
+RUN EXTRA_PACKAGES="metrics"; \
+    if [ "$INSTALL_BNB" == "true" ]; then \
+        EXTRA_PACKAGES="${EXTRA_PACKAGES},bitsandbytes"; \
+    fi; \
+    if [ "$INSTALL_VLLM" == "true" ]; then \
+        EXTRA_PACKAGES="${EXTRA_PACKAGES},vllm"; \
+    fi; \
+    if [ "$INSTALL_DEEPSPEED" == "true" ]; then \
+        EXTRA_PACKAGES="${EXTRA_PACKAGES},deepspeed"; \
+    fi; \
+    pip install -e ".[$EXTRA_PACKAGES]"
+
+# Rebuild flash attention
+RUN pip uninstall -y transformer-engine flash-attn && \
+    if [ "$INSTALL_FLASHATTN" == "true" ]; then \
+        pip uninstall -y ninja && pip install ninja && \
+        pip install --no-cache-dir flash-attn --no-build-isolation; \
+    fi
+
+# Set up volumes
+VOLUME [ "/root/.cache/huggingface", "/root/.cache/modelscope", "/app/data", "/app/output" ]
+
+# Expose port 7860 for the LLaMA Board
+ENV GRADIO_SERVER_PORT 7860
+EXPOSE 7860
+
+# Expose port 8000 for the API service
+ENV API_PORT 8000
+EXPOSE 8000

docker/docker-cuda/docker-compose.yml

@@ -0,0 +1,32 @@
+services:
+  llamafactory:
+    build:
+      dockerfile: ./docker/docker-cuda/Dockerfile
+      context: ../..
+      args:
+        INSTALL_BNB: false
+        INSTALL_VLLM: false
+        INSTALL_DEEPSPEED: false
+        INSTALL_FLASHATTN: false
+        PIP_INDEX: https://pypi.org/simple
+    container_name: llamafactory
+    volumes:
+      - ../../hf_cache:/root/.cache/huggingface
+      - ../../ms_cache:/root/.cache/modelscope
+      - ../../data:/app/data
+      - ../../output:/app/output
+    ports:
+      - "7860:7860"
+      - "8000:8000"
+    ipc: host
+    tty: true
+    stdin_open: true
+    command: bash
+    deploy:
+      resources:
+        reservations:
+          devices:
+          - driver: nvidia
+            count: "all"
+            capabilities: [gpu]
+    restart: unless-stopped

docker/docker-npu/Dockerfile

@@ -0,0 +1,45 @@
+# Use the Ubuntu 22.04 image with CANN 8.0.rc1
+# More versions can be found at https://hub.docker.com/r/cosdt/cann/tags
+# FROM cosdt/cann:8.0.rc1-910-ubuntu22.04
+FROM cosdt/cann:8.0.rc1-910b-ubuntu22.04
+# FROM cosdt/cann:8.0.rc1-910-openeuler22.03
+# FROM cosdt/cann:8.0.rc1-910b-openeuler22.03
+
+# Define environments
+ENV DEBIAN_FRONTEND=noninteractive
+
+# Define installation arguments
+ARG INSTALL_DEEPSPEED=false
+ARG PIP_INDEX=https://pypi.org/simple
+ARG TORCH_INDEX=https://download.pytorch.org/whl/cpu
+
+# Set the working directory
+WORKDIR /app
+
+# Install the requirements
+COPY requirements.txt /app
+RUN pip config set global.index-url "$PIP_INDEX" && \
+    pip config set global.extra-index-url "$TORCH_INDEX" && \
+    python -m pip install --upgrade pip && \
+    python -m pip install -r requirements.txt
+
+# Copy the rest of the application into the image
+COPY . /app
+
+# Install the LLaMA Factory
+RUN EXTRA_PACKAGES="torch-npu,metrics"; \
+    if [ "$INSTALL_DEEPSPEED" == "true" ]; then \
+        EXTRA_PACKAGES="${EXTRA_PACKAGES},deepspeed"; \
+    fi; \
+    pip install -e ".[$EXTRA_PACKAGES]"
+
+# Set up volumes
+VOLUME [ "/root/.cache/huggingface", "/root/.cache/modelscope", "/app/data", "/app/output" ]
+
+# Expose port 7860 for the LLaMA Board
+ENV GRADIO_SERVER_PORT 7860
+EXPOSE 7860
+
+# Expose port 8000 for the API service
+ENV API_PORT 8000
+EXPOSE 8000

docker/docker-npu/docker-compose.yml

@@ -0,0 +1,31 @@
+services:
+  llamafactory:
+    build:
+      dockerfile: ./docker/docker-npu/Dockerfile
+      context: ../..
+      args:
+        INSTALL_DEEPSPEED: false
+        PIP_INDEX: https://pypi.org/simple
+    container_name: llamafactory
+    volumes:
+      - ../../hf_cache:/root/.cache/huggingface
+      - ../../ms_cache:/root/.cache/modelscope
+      - ../../data:/app/data
+      - ../../output:/app/output
+      - /usr/local/dcmi:/usr/local/dcmi
+      - /usr/local/bin/npu-smi:/usr/local/bin/npu-smi
+      - /usr/local/Ascend/driver:/usr/local/Ascend/driver
+      - /etc/ascend_install.info:/etc/ascend_install.info
+    ports:
+      - "7860:7860"
+      - "8000:8000"
+    ipc: host
+    tty: true
+    stdin_open: true
+    command: bash
+    devices:
+      - /dev/davinci0
+      - /dev/davinci_manager
+      - /dev/devmm_svm
+      - /dev/hisi_hdc
+    restart: unless-stopped

evaluation/ceval/ceval.py

@@ -11,6 +11,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+
 import os
 
 import datasets

evaluation/cmmlu/cmmlu.py

@@ -11,6 +11,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+
 import os
 
 import datasets

evaluation/mmlu/mmlu.py

@@ -11,6 +11,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+
 import os
 
 import datasets

examples/README.md

@@ -4,59 +4,59 @@ Make sure to execute these commands in the `LLaMA-Factory` directory.
 
 ## Table of Contents
 
-- [LoRA Fine-Tuning on A Single GPU](#lora-fine-tuning-on-a-single-gpu)
-- [QLoRA Fine-Tuning on a Single GPU](#qlora-fine-tuning-on-a-single-gpu)
-- [LoRA Fine-Tuning on Multiple GPUs](#lora-fine-tuning-on-multiple-gpus)
-- [LoRA Fine-Tuning on Multiple NPUs](#lora-fine-tuning-on-multiple-npus)
-- [Full-Parameter Fine-Tuning on Multiple GPUs](#full-parameter-fine-tuning-on-multiple-gpus)
+- [LoRA Fine-Tuning](#lora-fine-tuning)
+- [QLoRA Fine-Tuning](#qlora-fine-tuning)
+- [Full-Parameter Fine-Tuning](#full-parameter-fine-tuning)
 - [Merging LoRA Adapters and Quantization](#merging-lora-adapters-and-quantization)
 - [Inferring LoRA Fine-Tuned Models](#inferring-lora-fine-tuned-models)
 - [Extras](#extras)
 
+Use `CUDA_VISIBLE_DEVICES` (GPU) or `ASCEND_RT_VISIBLE_DEVICES` (NPU) to choose computing devices.
+
 ## Examples
 
-### LoRA Fine-Tuning on A Single GPU
+### LoRA Fine-Tuning
 
 #### (Continuous) Pre-Training
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_pretrain.yaml
+llamafactory-cli train examples/train_lora/llama3_lora_pretrain.yaml
 ```
 
 #### Supervised Fine-Tuning
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_sft.yaml
+llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
 ```
 
 #### Multimodal Supervised Fine-Tuning
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llava1_5_lora_sft.yaml
+llamafactory-cli train examples/train_lora/llava1_5_lora_sft.yaml
 ```
 
 #### Reward Modeling
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_reward.yaml
+llamafactory-cli train examples/train_lora/llama3_lora_reward.yaml
 ```
 
 #### PPO Training
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_ppo.yaml
+llamafactory-cli train examples/train_lora/llama3_lora_ppo.yaml
 ```
 
 #### DPO/ORPO/SimPO Training
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_dpo.yaml
+llamafactory-cli train examples/train_lora/llama3_lora_dpo.yaml
 ```
 
 #### KTO Training
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_kto.yaml
+llamafactory-cli train examples/train_lora/llama3_lora_kto.yaml
 ```
 
 #### Preprocess Dataset
@@ -64,95 +64,79 @@ CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lo
 It is useful for large dataset, use `tokenized_path` in config to load the preprocessed dataset.
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_preprocess.yaml
+llamafactory-cli train examples/train_lora/llama3_preprocess.yaml
 ```
 
 #### Evaluating on MMLU/CMMLU/C-Eval Benchmarks
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli eval examples/lora_single_gpu/llama3_lora_eval.yaml
+llamafactory-cli eval examples/train_lora/llama3_lora_eval.yaml
 ```
 
 #### Batch Predicting and Computing BLEU and ROUGE Scores
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_predict.yaml
-```
-
-### QLoRA Fine-Tuning on a Single GPU
-
-#### Supervised Fine-Tuning with 4/8-bit Bitsandbytes Quantization (Recommended)
-
-```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/qlora_single_gpu/llama3_lora_sft_bitsandbytes.yaml
-```
-
-#### Supervised Fine-Tuning with 4/8-bit GPTQ Quantization
-
-```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/qlora_single_gpu/llama3_lora_sft_gptq.yaml
-```
-
-#### Supervised Fine-Tuning with 4-bit AWQ Quantization
-
-```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/qlora_single_gpu/llama3_lora_sft_awq.yaml
-```
-
-#### Supervised Fine-Tuning with 2-bit AQLM Quantization
-
-```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/qlora_single_gpu/llama3_lora_sft_aqlm.yaml
-```
-
-### LoRA Fine-Tuning on Multiple GPUs
-
-#### Supervised Fine-Tuning on Single Node
-
-```bash
-CUDA_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/lora_multi_gpu/llama3_lora_sft.yaml
+llamafactory-cli train examples/train_lora/llama3_lora_predict.yaml
 ```
 
 #### Supervised Fine-Tuning on Multiple Nodes
 
 ```bash
-CUDA_VISIBLE_DEVICES=0,1,2,3 NNODES=2 RANK=0 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/lora_multi_gpu/llama3_lora_sft.yaml
-CUDA_VISIBLE_DEVICES=0,1,2,3 NNODES=2 RANK=1 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/lora_multi_gpu/llama3_lora_sft.yaml
+FORCE_TORCHRUN=1 NNODES=2 RANK=0 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
+FORCE_TORCHRUN=1 NNODES=2 RANK=1 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
 ```
 
 #### Supervised Fine-Tuning with DeepSpeed ZeRO-3 (Weight Sharding)
 
 ```bash
-CUDA_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/lora_multi_gpu/llama3_lora_sft_ds.yaml
+FORCE_TORCHRUN=1 llamafactory-cli train examples/train_lora/llama3_lora_sft_ds3.yaml
 ```
 
-### LoRA Fine-Tuning on Multiple NPUs
+### QLoRA Fine-Tuning
 
-#### Supervised Fine-Tuning with DeepSpeed ZeRO-0
+#### Supervised Fine-Tuning with 4/8-bit Bitsandbytes/HQQ/EETQ Quantization (Recommended)
 
 ```bash
-ASCEND_RT_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/lora_multi_npu/llama3_lora_sft_ds.yaml
+llamafactory-cli train examples/train_qlora/llama3_lora_sft_otfq.yaml
 ```
 
-### Full-Parameter Fine-Tuning on Multiple GPUs
+#### Supervised Fine-Tuning with 4/8-bit GPTQ Quantization
+
+```bash
+llamafactory-cli train examples/train_qlora/llama3_lora_sft_gptq.yaml
+```
+
+#### Supervised Fine-Tuning with 4-bit AWQ Quantization
+
+```bash
+llamafactory-cli train examples/train_qlora/llama3_lora_sft_awq.yaml
+```
+
+#### Supervised Fine-Tuning with 2-bit AQLM Quantization
+
+```bash
+llamafactory-cli train examples/train_qlora/llama3_lora_sft_aqlm.yaml
+```
+
+### Full-Parameter Fine-Tuning
 
 #### Supervised Fine-Tuning on Single Node
 
 ```bash
-CUDA_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/full_multi_gpu/llama3_full_sft.yaml
+FORCE_TORCHRUN=1 llamafactory-cli train examples/train_full/llama3_full_sft_ds3.yaml
 ```
 
 #### Supervised Fine-Tuning on Multiple Nodes
 
 ```bash
-CUDA_VISIBLE_DEVICES=0,1,2,3 NNODES=2 RANK=0 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/full_multi_gpu/llama3_full_sft.yaml
-CUDA_VISIBLE_DEVICES=0,1,2,3 NNODES=2 RANK=1 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/full_multi_gpu/llama3_full_sft.yaml
+FORCE_TORCHRUN=1 NNODES=2 RANK=0 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/train_full/llama3_full_sft_ds3.yaml
+FORCE_TORCHRUN=1 NNODES=2 RANK=1 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/train_full/llama3_full_sft_ds3.yaml
 ```
 
 #### Batch Predicting and Computing BLEU and ROUGE Scores
 
 ```bash
-CUDA_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/full_multi_gpu/llama3_full_predict.yaml
+llamafactory-cli train examples/train_full/llama3_full_predict.yaml
 ```
 
 ### Merging LoRA Adapters and Quantization
@@ -162,35 +146,33 @@ CUDA_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/full_multi_gpu/llam
 Note: DO NOT use quantized model or `quantization_bit` when merging LoRA adapters.
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli export examples/merge_lora/llama3_lora_sft.yaml
+llamafactory-cli export examples/merge_lora/llama3_lora_sft.yaml
 ```
 
 #### Quantizing Model using AutoGPTQ
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli export examples/merge_lora/llama3_gptq.yaml
+llamafactory-cli export examples/merge_lora/llama3_gptq.yaml
 ```
 
 ### Inferring LoRA Fine-Tuned Models
 
-Use `CUDA_VISIBLE_DEVICES=0,1` to infer models on multiple devices.
-
 #### Use CLI
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli chat examples/inference/llama3_lora_sft.yaml
+llamafactory-cli chat examples/inference/llama3_lora_sft.yaml
 ```
 
 #### Use Web UI
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli webchat examples/inference/llama3_lora_sft.yaml
+llamafactory-cli webchat examples/inference/llama3_lora_sft.yaml
 ```
 
 #### Launch OpenAI-style API
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli api examples/inference/llama3_lora_sft.yaml
+llamafactory-cli api examples/inference/llama3_lora_sft.yaml
 ```
 
 ### Extras
@@ -198,36 +180,42 @@ CUDA_VISIBLE_DEVICES=0 llamafactory-cli api examples/inference/llama3_lora_sft.y
 #### Full-Parameter Fine-Tuning using GaLore
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/galore/llama3_full_sft.yaml
+llamafactory-cli train examples/extras/galore/llama3_full_sft.yaml
 ```
 
 #### Full-Parameter Fine-Tuning using BAdam
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/badam/llama3_full_sft.yaml
+llamafactory-cli train examples/extras/badam/llama3_full_sft.yaml
 ```
 
 #### LoRA+ Fine-Tuning
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/loraplus/llama3_lora_sft.yaml
+llamafactory-cli train examples/extras/loraplus/llama3_lora_sft.yaml
+```
+
+#### PiSSA Fine-Tuning
+
+```bash
+llamafactory-cli train examples/extras/pissa/llama3_lora_sft.yaml
 ```
 
 #### Mixture-of-Depths Fine-Tuning
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/mod/llama3_full_sft.yaml
+llamafactory-cli train examples/extras/mod/llama3_full_sft.yaml
 ```
 
 #### LLaMA-Pro Fine-Tuning
 
 ```bash
 bash examples/extras/llama_pro/expand.sh
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/llama_pro/llama3_freeze_sft.yaml
+llamafactory-cli train examples/extras/llama_pro/llama3_freeze_sft.yaml
 ```
 
 #### FSDP+QLoRA Fine-Tuning
 
 ```bash
-bash examples/extras/fsdp_qlora/single_node.sh
+bash examples/extras/fsdp_qlora/train.sh
 ```

examples/README_zh.md

@@ -4,59 +4,59 @@
 
 ## 目录
 
-- [单 GPU LoRA 微调](#单-gpu-lora-微调)
-- [单 GPU QLoRA 微调](#单-gpu-qlora-微调)
-- [多 GPU LoRA 微调](#多-gpu-lora-微调)
-- [多 NPU LoRA 微调](#多-npu-lora-微调)
-- [多 GPU 全参数微调](#多-gpu-全参数微调)
+- [LoRA 微调](#lora-微调)
+- [QLoRA 微调](#qlora-微调)
+- [全参数微调](#全参数微调)
 - [合并 LoRA 适配器与模型量化](#合并-lora-适配器与模型量化)
 - [推理 LoRA 模型](#推理-lora-模型)
 - [杂项](#杂项)
 
+使用 `CUDA_VISIBLE_DEVICES`（GPU）或 `ASCEND_RT_VISIBLE_DEVICES`（NPU）选择计算设备。
+
 ## 示例
 
-### 单 GPU LoRA 微调
+### LoRA 微调
 
 #### （增量）预训练
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_pretrain.yaml
+llamafactory-cli train examples/train_lora/llama3_lora_pretrain.yaml
 ```
 
 #### 指令监督微调
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_sft.yaml
+llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
 ```
 
 #### 多模态指令监督微调
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llava1_5_lora_sft.yaml
+llamafactory-cli train examples/train_lora/llava1_5_lora_sft.yaml
 ```
 
 #### 奖励模型训练
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_reward.yaml
+llamafactory-cli train examples/train_lora/llama3_lora_reward.yaml
 ```
 
 #### PPO 训练
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_ppo.yaml
+llamafactory-cli train examples/train_lora/llama3_lora_ppo.yaml
 ```
 
 #### DPO/ORPO/SimPO 训练
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_dpo.yaml
+llamafactory-cli train examples/train_lora/llama3_lora_dpo.yaml
 ```
 
 #### KTO 训练
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_kto.yaml
+llamafactory-cli train examples/train_lora/llama3_lora_kto.yaml
 ```
 
 #### 预处理数据集
@@ -64,95 +64,79 @@ CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lo
 对于大数据集有帮助，在配置中使用 `tokenized_path` 以加载预处理后的数据集。
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_preprocess.yaml
+llamafactory-cli train examples/train_lora/llama3_preprocess.yaml
 ```
 
 #### 在 MMLU/CMMLU/C-Eval 上评估
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli eval examples/lora_single_gpu/llama3_lora_eval.yaml
+llamafactory-cli eval examples/train_lora/llama3_lora_eval.yaml
 ```
 
 #### 批量预测并计算 BLEU 和 ROUGE 分数
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_predict.yaml
+llamafactory-cli train examples/train_lora/llama3_lora_predict.yaml
 ```
 
-### 单 GPU QLoRA 微调
-
-#### 基于 4/8 比特 Bitsandbytes 量化进行指令监督微调（推荐）
+#### 多机指令监督微调
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/qlora_single_gpu/llama3_lora_sft_bitsandbytes.yaml
-```
-
-#### 基于 4/8 比特 GPTQ 量化进行指令监督微调
-
-```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/qlora_single_gpu/llama3_lora_sft_gptq.yaml
-```
-
-#### 基于 4 比特 AWQ 量化进行指令监督微调
-
-```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/qlora_single_gpu/llama3_lora_sft_awq.yaml
-```
-
-#### 基于 2 比特 AQLM 量化进行指令监督微调
-
-```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/qlora_single_gpu/llama3_lora_sft_aqlm.yaml
-```
-
-### 多 GPU LoRA 微调
-
-#### 在单机上进行指令监督微调
-
-```bash
-CUDA_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/lora_multi_gpu/llama3_lora_sft.yaml
-```
-
-#### 在多机上进行指令监督微调
-
-```bash
-CUDA_VISIBLE_DEVICES=0,1,2,3 NNODES=2 RANK=0 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/lora_multi_gpu/llama3_lora_sft.yaml
-CUDA_VISIBLE_DEVICES=0,1,2,3 NNODES=2 RANK=1 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/lora_multi_gpu/llama3_lora_sft.yaml
+FORCE_TORCHRUN=1 NNODES=2 RANK=0 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
+FORCE_TORCHRUN=1 NNODES=2 RANK=1 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
 ```
 
 #### 使用 DeepSpeed ZeRO-3 平均分配显存
 
 ```bash
-CUDA_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/lora_multi_gpu/llama3_lora_sft_ds.yaml
+FORCE_TORCHRUN=1 llamafactory-cli train examples/train_lora/llama3_lora_sft_ds3.yaml
 ```
 
-### 多 NPU LoRA 微调
+### QLoRA 微调
 
-#### 使用 DeepSpeed ZeRO-0 进行指令监督微调
+#### 基于 4/8 比特 Bitsandbytes/HQQ/EETQ 量化进行指令监督微调（推荐）
 
 ```bash
-ASCEND_RT_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/lora_multi_npu/llama3_lora_sft_ds.yaml
+llamafactory-cli train examples/train_qlora/llama3_lora_sft_otfq.yaml
 ```
 
-### 多 GPU 全参数微调
+#### 基于 4/8 比特 GPTQ 量化进行指令监督微调
+
+```bash
+llamafactory-cli train examples/train_qlora/llama3_lora_sft_gptq.yaml
+```
+
+#### 基于 4 比特 AWQ 量化进行指令监督微调
+
+```bash
+llamafactory-cli train examples/train_qlora/llama3_lora_sft_awq.yaml
+```
+
+#### 基于 2 比特 AQLM 量化进行指令监督微调
+
+```bash
+llamafactory-cli train examples/train_qlora/llama3_lora_sft_aqlm.yaml
+```
+
+### 全参数微调
 
 #### 在单机上进行指令监督微调
 
 ```bash
-CUDA_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/full_multi_gpu/llama3_full_sft.yaml
+FORCE_TORCHRUN=1 llamafactory-cli train examples/train_full/llama3_full_sft_ds3.yaml
 ```
 
 #### 在多机上进行指令监督微调
 
 ```bash
-CUDA_VISIBLE_DEVICES=0,1,2,3 NNODES=2 RANK=0 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/full_multi_gpu/llama3_full_sft.yaml
-CUDA_VISIBLE_DEVICES=0,1,2,3 NNODES=2 RANK=1 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/full_multi_gpu/llama3_full_sft.yaml
+FORCE_TORCHRUN=1 NNODES=2 RANK=0 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/train_full/llama3_full_sft_ds3.yaml
+FORCE_TORCHRUN=1 NNODES=2 RANK=1 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/train_full/llama3_full_sft_ds3.yaml
 ```
 
 #### 批量预测并计算 BLEU 和 ROUGE 分数
 
 ```bash
-CUDA_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/full_multi_gpu/llama3_full_predict.yaml
+llamafactory-cli train examples/train_full/llama3_full_predict.yaml
 ```
 
 ### 合并 LoRA 适配器与模型量化
@@ -162,35 +146,33 @@ CUDA_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/full_multi_gpu/llam
 注：请勿使用量化后的模型或 `quantization_bit` 参数来合并 LoRA 适配器。
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli export examples/merge_lora/llama3_lora_sft.yaml
+llamafactory-cli export examples/merge_lora/llama3_lora_sft.yaml
 ```
 
 #### 使用 AutoGPTQ 量化模型
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli export examples/merge_lora/llama3_gptq.yaml
+llamafactory-cli export examples/merge_lora/llama3_gptq.yaml
 ```
 
 ### 推理 LoRA 模型
 
-使用 `CUDA_VISIBLE_DEVICES=0,1` 进行多卡推理。
-
 #### 使用命令行接口
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli chat examples/inference/llama3_lora_sft.yaml
+llamafactory-cli chat examples/inference/llama3_lora_sft.yaml
 ```
 
 #### 使用浏览器界面
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli webchat examples/inference/llama3_lora_sft.yaml
+llamafactory-cli webchat examples/inference/llama3_lora_sft.yaml
 ```
 
 #### 启动 OpenAI 风格 API
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli api examples/inference/llama3_lora_sft.yaml
+llamafactory-cli api examples/inference/llama3_lora_sft.yaml
 ```
 
 ### 杂项
@@ -198,36 +180,42 @@ CUDA_VISIBLE_DEVICES=0 llamafactory-cli api examples/inference/llama3_lora_sft.y
 #### 使用 GaLore 进行全参数训练
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/galore/llama3_full_sft.yaml
+llamafactory-cli train examples/extras/galore/llama3_full_sft.yaml
 ```
 
 #### 使用 BAdam 进行全参数训练
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/badam/llama3_full_sft.yaml
+llamafactory-cli train examples/extras/badam/llama3_full_sft.yaml
 ```
 
 #### LoRA+ 微调
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/loraplus/llama3_lora_sft.yaml
+llamafactory-cli train examples/extras/loraplus/llama3_lora_sft.yaml
+```
+
+#### PiSSA 微调
+
+```bash
+llamafactory-cli train examples/extras/pissa/llama3_lora_sft.yaml
 ```
 
 #### 深度混合微调
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/mod/llama3_full_sft.yaml
+llamafactory-cli train examples/extras/mod/llama3_full_sft.yaml
 ```
 
 #### LLaMA-Pro 微调
 
 ```bash
 bash examples/extras/llama_pro/expand.sh
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/llama_pro/llama3_freeze_sft.yaml
+llamafactory-cli train examples/extras/llama_pro/llama3_freeze_sft.yaml
 ```
 
 #### FSDP+QLoRA 微调
 
 ```bash
-bash examples/extras/fsdp_qlora/single_node.sh
+bash examples/extras/fsdp_qlora/train.sh
 ```

examples/extras/badam/llama3_full_sft.yaml

@@ -5,10 +5,11 @@ model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
 stage: sft
 do_train: true
 finetuning_type: full
-
-### ddp
-ddp_timeout: 180000000
-deepspeed: examples/deepspeed/ds_z3_config.json
+use_badam: true
+badam_mode: layer
+badam_switch_mode: ascending
+badam_switch_interval: 50
+badam_verbose: 2
 
 ### dataset
 dataset: identity,alpaca_en_demo
@@ -27,12 +28,11 @@ overwrite_output_dir: true
 
 ### train
 per_device_train_batch_size: 1
-gradient_accumulation_steps: 2
+gradient_accumulation_steps: 8
 learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
 
 ### eval
 val_size: 0.1

examples/extras/badam/llama3_full_sft_ds3.yaml

@@ -6,9 +6,11 @@ stage: sft
 do_train: true
 finetuning_type: full
 use_badam: true
+badam_mode: layer
 badam_switch_mode: ascending
 badam_switch_interval: 50
 badam_verbose: 2
+deepspeed: examples/deepspeed/ds_z3_config.json
 
 ### dataset
 dataset: identity,alpaca_en_demo
@@ -32,7 +34,6 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-pure_bf16: true
 
 ### eval
 val_size: 0.1

examples/extras/fsdp_qlora/llama3_lora_sft.yaml

@@ -8,9 +8,6 @@ do_train: true
 finetuning_type: lora
 lora_target: all
 
-### ddp
-ddp_timeout: 180000000
-
 ### dataset
 dataset: identity,alpaca_en_demo
 template: llama3
@@ -33,7 +30,8 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/extras/llama_pro/llama3_freeze_sft.yaml

@@ -31,7 +31,8 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/extras/loraplus/llama3_lora_sft.yaml

@@ -30,7 +30,8 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/extras/mod/llama3_full_sft.yaml

@@ -31,6 +31,7 @@ num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
 pure_bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/extras/pissa/llama3_lora_sft.yaml

@@ -1,12 +1,14 @@
 ### model
 model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
-quantization_bit: 4
 
 ### method
 stage: sft
 do_train: true
 finetuning_type: lora
 lora_target: all
+pissa_init: true
+pissa_iter: 16
+pissa_convert: true
 
 ### dataset
 dataset: identity,alpaca_en_demo
@@ -30,7 +32,8 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/inference/llava1_5.yaml

@@ -0,0 +1,3 @@
+model_name_or_path: llava-hf/llava-1.5-7b-hf
+template: vicuna
+visual_inputs: true

examples/train_full/llama3_full_sft_ds3.yaml

@@ -4,11 +4,8 @@ model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
 ### method
 stage: sft
 do_train: true
-finetuning_type: lora
-lora_target: all
-
-### ddp
-ddp_timeout: 180000000
+finetuning_type: full
+deepspeed: examples/deepspeed/ds_z3_config.json
 
 ### dataset
 dataset: identity,alpaca_en_demo
@@ -19,7 +16,7 @@ overwrite_cache: true
 preprocessing_num_workers: 16
 
 ### output
-output_dir: saves/llama3-8b/lora/sft
+output_dir: saves/llama3-8b/full/sft
 logging_steps: 10
 save_steps: 500
 plot_loss: true
@@ -32,7 +29,8 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/train_lora/llama3_lora_dpo.yaml

@@ -7,7 +7,7 @@ do_train: true
 finetuning_type: lora
 lora_target: all
 pref_beta: 0.1
-pref_loss: sigmoid  # [sigmoid (dpo), orpo, simpo]
+pref_loss: sigmoid  # choices: [sigmoid (dpo), orpo, simpo]
 
 ### dataset
 dataset: dpo_en_demo
@@ -31,7 +31,8 @@ learning_rate: 5.0e-6
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/train_lora/llama3_lora_eval.yaml

@@ -6,8 +6,7 @@ adapter_name_or_path: saves/llama3-8b/lora/sft
 finetuning_type: lora
 
 ### dataset
-task: mmlu
-split: test
+task: mmlu_test  # choices: [mmlu_test, ceval_validation, cmmlu_test]
 template: fewshot
 lang: en
 n_shot: 5

examples/train_lora/llama3_lora_kto.yaml

@@ -6,6 +6,7 @@ stage: kto
 do_train: true
 finetuning_type: lora
 lora_target: all
+pref_beta: 0.1
 
 ### dataset
 dataset: kto_en_demo
@@ -29,7 +30,8 @@ learning_rate: 5.0e-6
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/train_lora/llama3_lora_ppo.yaml

@@ -30,7 +30,8 @@ learning_rate: 1.0e-5
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
+ddp_timeout: 180000000
 
 ### generate
 max_new_tokens: 512

examples/train_lora/llama3_lora_predict.yaml

@@ -8,7 +8,7 @@ do_predict: true
 finetuning_type: lora
 
 ### dataset
-dataset: identity,alpaca_en_demo
+eval_dataset: identity,alpaca_en_demo
 template: llama3
 cutoff_len: 1024
 max_samples: 50
@@ -22,3 +22,4 @@ overwrite_output_dir: true
 ### eval
 per_device_eval_batch_size: 1
 predict_with_generate: true
+ddp_timeout: 180000000

examples/train_lora/llama3_lora_pretrain.yaml

@@ -15,7 +15,7 @@ overwrite_cache: true
 preprocessing_num_workers: 16
 
 ### output
-output_dir: saves/llama3-8b/lora/sft
+output_dir: saves/llama3-8b/lora/pretrain
 logging_steps: 10
 save_steps: 500
 plot_loss: true
@@ -28,7 +28,8 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/train_lora/llama3_lora_reward.yaml

@@ -25,11 +25,12 @@ overwrite_output_dir: true
 ### train
 per_device_train_batch_size: 1
 gradient_accumulation_steps: 8
-learning_rate: 1.0e-5
+learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/train_lora/llama3_lora_sft.yaml

@@ -29,7 +29,8 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/train_lora/llama3_lora_sft_ds0.yaml

@@ -6,9 +6,6 @@ stage: sft
 do_train: true
 finetuning_type: lora
 lora_target: all
-
-### ddp
-ddp_timeout: 180000000
 deepspeed: examples/deepspeed/ds_z0_config.json
 
 ### dataset
@@ -33,7 +30,8 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/train_lora/llama3_lora_sft_ds3.yaml

@@ -6,9 +6,6 @@ stage: sft
 do_train: true
 finetuning_type: lora
 lora_target: all
-
-### ddp
-ddp_timeout: 180000000
 deepspeed: examples/deepspeed/ds_z3_config.json
 
 ### dataset
@@ -33,7 +30,8 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/train_lora/llava1_5_lora_sft.yaml

@@ -30,7 +30,8 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/train_qlora/llama3_lora_sft_aqlm.yaml

@@ -29,7 +29,8 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/train_qlora/llama3_lora_sft_awq.yaml

@@ -29,7 +29,8 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/train_qlora/llama3_lora_sft_gptq.yaml

@@ -29,7 +29,8 @@ learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-fp16: true
+bf16: true
+ddp_timeout: 180000000
 
 ### eval
 val_size: 0.1

examples/train_qlora/llama3_lora_sft_otfq.yaml

@@ -0,0 +1,41 @@
+### model
+model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
+quantization_bit: 4
+quantization_method: bitsandbytes  # choices: [bitsandbytes (4/8), hqq (2/3/4/5/6/8), eetq (8)]
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_target: all
+
+### dataset
+dataset: identity,alpaca_en_demo
+template: llama3
+cutoff_len: 1024
+max_samples: 1000
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: saves/llama3-8b/lora/sft
+logging_steps: 10
+save_steps: 500
+plot_loss: true
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 1
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 3.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 1
+eval_strategy: steps
+eval_steps: 500

requirements.txt

@@ -4,6 +4,7 @@ accelerate>=0.30.1
 peft>=0.11.1
 trl>=0.8.6
 gradio>=4.0.0
+pandas>=2.0.0
 scipy
 einops
 sentencepiece
@@ -17,3 +18,4 @@ matplotlib>=3.7.0
 fire
 packaging
 pyyaml
+numpy<2.0.0

scripts/cal_flops.py

@@ -1,7 +1,20 @@
 # coding=utf-8
-# Calculates the flops of pre-trained models.
-# Usage: python cal_flops.py --model_name_or_path path_to_model --batch_size 1 --seq_length 512
-# Inspired by: https://www.deepspeed.ai/tutorials/flops-profiler/
+# Copyright 2024 Microsoft Corporation and the LlamaFactory team.
+#
+# This code is inspired by the Microsoft's DeepSpeed library.
+# https://www.deepspeed.ai/tutorials/flops-profiler/
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
 
 import fire
 import torch
@@ -17,6 +30,10 @@ def calculate_flops(
     seq_length: int = 256,
     flash_attn: str = "auto",
 ):
+    r"""
+    Calculates the flops of pre-trained models.
+    Usage: python cal_flops.py --model_name_or_path path_to_model --batch_size 1 --seq_length 512
+    """
     with get_accelerator().device(0):
         chat_model = ChatModel(dict(model_name_or_path=model_name_or_path, template="empty", flash_attn=flash_attn))
         fake_input = torch.ones((batch_size, seq_length), dtype=torch.long, device=chat_model.model.device)

scripts/cal_lr.py

@@ -1,7 +1,20 @@
 # coding=utf-8
-# Calculates the optimal learning rate for 7B/13B models using LLaMA's hyper-parameters.
-# Usage: python cal_lr.py --model_name_or_path path_to_model --dataset alpaca_en --cutoff_len 1024 --batch_size 16
-# Inspired by: https://github.com/imoneoi/openchat/blob/master/ochat/training_deepspeed/train.py
+# Copyright 2024 imoneoi and the LlamaFactory team.
+#
+# This code is inspired by the imoneoi's OpenChat library.
+# https://github.com/imoneoi/openchat/blob/3.6.0/ochat/training_deepspeed/train.py
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
 
 import math
 from typing import Literal
@@ -31,7 +44,12 @@ def calculate_lr(
     template: str = "default",
     cutoff_len: int = 1024,  # i.e. maximum input length during training
     is_mistral: bool = False,  # mistral model uses a smaller learning rate,
+    packing: bool = False,
 ):
+    r"""
+    Calculates the optimal learning rate for 7B/13B models using LLaMA's hyper-parameters.
+    Usage: python cal_lr.py --model_name_or_path path_to_model --dataset alpaca_en --cutoff_len 1024 --batch_size 16
+    """
     model_args, data_args, training_args, _, _ = get_train_args(
         dict(
             stage=stage,
@@ -40,19 +58,21 @@ def calculate_lr(
             dataset_dir=dataset_dir,
             template=template,
             cutoff_len=cutoff_len,
+            packing=packing,
             output_dir="dummy_dir",
             overwrite_cache=True,
+            do_train=True,
         )
     )
     tokenizer_module = load_tokenizer(model_args)
     tokenizer = tokenizer_module["tokenizer"]
-    trainset = get_dataset(model_args, data_args, training_args, stage, **tokenizer_module)
+    trainset = get_dataset(model_args, data_args, training_args, stage, **tokenizer_module)["train_dataset"]
     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
+        raise NotImplementedError("Stage does not supported: {}.".format(stage))
 
     dataloader = DataLoader(trainset, batch_size, shuffle=False, collate_fn=data_collator, pin_memory=True)
     valid_tokens, total_tokens = 0, 0

scripts/cal_ppl.py

@@ -1,6 +1,17 @@
 # coding=utf-8
-# Calculates the ppl on the dataset of the pre-trained models.
-# Usage: python cal_ppl.py --model_name_or_path path_to_model --save_name ppl.json
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
 
 import json
 from dataclasses import dataclass
@@ -56,6 +67,10 @@ def cal_ppl(
     max_samples: Optional[int] = None,
     train_on_prompt: bool = False,
 ):
+    r"""
+    Calculates the ppl on the dataset of the pre-trained models.
+    Usage: python cal_ppl.py --model_name_or_path path_to_model --save_name ppl.json
+    """
     model_args, data_args, training_args, finetuning_args, _ = get_train_args(
         dict(
             stage=stage,
@@ -68,11 +83,12 @@ def cal_ppl(
             train_on_prompt=train_on_prompt,
             output_dir="dummy_dir",
             overwrite_cache=True,
+            do_train=True,
         )
     )
     tokenizer_module = load_tokenizer(model_args)
     tokenizer = tokenizer_module["tokenizer"]
-    trainset = get_dataset(model_args, data_args, training_args, stage, **tokenizer_module)
+    trainset = get_dataset(model_args, data_args, training_args, stage, **tokenizer_module)["train_dataset"]
     model = load_model(tokenizer, model_args, finetuning_args, is_trainable=False)
     if stage == "pt":
         data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)
@@ -83,7 +99,7 @@ def cal_ppl(
             tokenizer=tokenizer, label_pad_token_id=IGNORE_INDEX, train_on_prompt=train_on_prompt
         )
     else:
-        raise NotImplementedError
+        raise NotImplementedError("Stage does not supported: {}.".format(stage))
 
     dataloader = DataLoader(trainset, batch_size, shuffle=False, collate_fn=data_collator, pin_memory=True)
     criterion = torch.nn.CrossEntropyLoss(reduction="none")

scripts/length_cdf.py

@@ -1,6 +1,17 @@
 # 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
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
 
 from collections import defaultdict
 
@@ -19,6 +30,10 @@ def length_cdf(
     template: str = "default",
     interval: int = 1000,
 ):
+    r"""
+    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
+    """
     model_args, data_args, training_args, _, _ = get_train_args(
         dict(
             stage="sft",
@@ -29,10 +44,11 @@ def length_cdf(
             cutoff_len=1_000_000,
             output_dir="dummy_dir",
             overwrite_cache=True,
+            do_train=True,
         )
     )
     tokenizer_module = load_tokenizer(model_args)
-    trainset = get_dataset(model_args, data_args, training_args, stage="sft", **tokenizer_module)
+    trainset = get_dataset(model_args, data_args, training_args, stage="sft", **tokenizer_module)["train_dataset"]
     total_num = len(trainset)
     length_dict = defaultdict(int)
     for sample in tqdm(trainset["input_ids"]):

scripts/llama_pro.py

@@ -1,7 +1,20 @@
 # coding=utf-8
-# Performs block expansion for LLaMA, Mistral, Qwen1.5 or Yi 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
+# Copyright 2024 Tencent Inc. and the LlamaFactory team.
+#
+# This code is inspired by the Tencent's LLaMA-Pro library.
+# https://github.com/TencentARC/LLaMA-Pro/blob/main/scripts/block_expansion.py
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
 
 import json
 import os
@@ -37,6 +50,10 @@ def block_expansion(
     shard_size: Optional[str] = "2GB",
     save_safetensors: Optional[bool] = False,
 ):
+    r"""
+    Performs block expansion for LLaMA, Mistral, Qwen1.5 or Yi models.
+    Usage: python llama_pro.py --model_name_or_path meta-llama/Llama-2-7b-hf --output_dir llama2_pro --num_expand 8
+    """
     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)
@@ -103,7 +120,7 @@ def block_expansion(
             json.dump(index, f, indent=2, sort_keys=True)
         print("Model weights saved in {}".format(output_dir))
 
-    print("Fine-tune this model with:")
+    print("- Fine-tune this model with:")
     print("model_name_or_path: {}".format(output_dir))
     print("finetuning_type: freeze")
     print("freeze_trainable_layers: {}".format(num_expand))

scripts/llamafy_baichuan2.py

@@ -1,8 +1,17 @@
 # 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
-# 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
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
 
 import json
 import os
@@ -79,6 +88,11 @@ def save_config(input_dir: str, output_dir: str):
 def llamafy_baichuan2(
     input_dir: str, output_dir: str, shard_size: Optional[str] = "2GB", save_safetensors: Optional[bool] = False
 ):
+    r"""
+    Converts the Baichuan2-7B model in the same format as LLaMA2-7B.
+    Usage: python llamafy_baichuan2.py --input_dir input --output_dir output
+    Converted model: https://huggingface.co/hiyouga/Baichuan2-7B-Base-LLaMAfied
+    """
     try:
         os.makedirs(output_dir, exist_ok=False)
     except Exception as e:

scripts/llamafy_qwen.py

@@ -1,7 +1,17 @@
 # coding=utf-8
-# Converts the Qwen models in the same format as LLaMA2.
-# Usage: python llamafy_qwen.py --input_dir input --output_dir output
-# Converted model: https://huggingface.co/hiyouga/Qwen-14B-Chat-LLaMAfied
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
 
 import json
 import os
@@ -131,6 +141,11 @@ def save_config(input_dir: str, output_dir: str, torch_dtype: str):
 def llamafy_qwen(
     input_dir: str, output_dir: str, shard_size: Optional[str] = "2GB", save_safetensors: Optional[bool] = False
 ):
+    r"""
+    Converts the Qwen models in the same format as LLaMA2.
+    Usage: python llamafy_qwen.py --input_dir input --output_dir output
+    Converted model: https://huggingface.co/hiyouga/Qwen-14B-Chat-LLaMAfied
+    """
     try:
         os.makedirs(output_dir, exist_ok=False)
     except Exception as e:

scripts/loftq_init.py

@@ -1,14 +1,25 @@
 # coding=utf-8
-# Initializes LoRA weights with LoRA-fine-tuning-aware Quantization (LoftQ)
-# Usage: python loftq_init.py --model_name_or_path path_to_model --save_dir output_dir
-# Inspired by: https://github.com/huggingface/peft/blob/main/examples/loftq_finetuning/quantize_save_load.py
+# Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
+#
+# This code is based on the HuggingFace's PEFT library.
+# https://github.com/huggingface/peft/blob/v0.10.0/examples/loftq_finetuning/quantize_save_load.py
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
 
 import os
-from typing import TYPE_CHECKING, Optional
+from typing import TYPE_CHECKING
 
 import fire
-import torch
-import torch.nn as nn
 from peft import LoftQConfig, LoraConfig, TaskType, get_peft_model
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
@@ -17,65 +28,61 @@ if TYPE_CHECKING:
     from transformers import PreTrainedModel
 
 
-class Shell(nn.Module):
-    def __init__(self, weight: torch.Tensor, bias: Optional[torch.Tensor] = None):
-        super().__init__()
-        self.weight = nn.Parameter(weight, requires_grad=False)
-        if bias is not None:
-            self.bias = nn.Parameter(bias, requires_grad=False)
-
-
-def unwrap_model(model: nn.Module, pattern=".base_layer") -> None:
-    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)
-        child_module = getattr(parent_module, child_name)
-        base_layer = getattr(child_module, "base_layer")
-        weight = getattr(base_layer, "weight", None)
-        bias = getattr(base_layer, "bias", None)
-        setattr(parent_module, child_name, Shell(weight, bias))
-
-    print("Model unwrapped.")
-
-
 def quantize_loftq(
     model_name_or_path: str,
-    save_dir: str,
-    loftq_bits: Optional[int] = 4,
-    loftq_iter: Optional[int] = 1,
-    lora_alpha: Optional[int] = None,
-    lora_rank: Optional[int] = 16,
-    lora_target: Optional[str] = "q_proj,v_proj",
-    save_safetensors: Optional[bool] = False,
+    output_dir: str,
+    loftq_bits: int = 4,
+    loftq_iter: int = 4,
+    lora_alpha: int = None,
+    lora_rank: int = 16,
+    lora_dropout: float = 0,
+    lora_target: tuple = ("q_proj", "v_proj"),
+    save_safetensors: bool = True,
 ):
+    r"""
+    Initializes LoRA weights with LoRA-fine-tuning-aware Quantization (LoftQ)
+    Usage: python loftq_init.py --model_name_or_path path_to_model --output_dir output_dir
+    """
+    if isinstance(lora_target, str):
+        lora_target = [name.strip() for name in lora_target.split(",")]
+
     tokenizer = AutoTokenizer.from_pretrained(model_name_or_path, trust_remote_code=True)
     model = AutoModelForCausalLM.from_pretrained(model_name_or_path, trust_remote_code=True, torch_dtype="auto")
+
     loftq_config = LoftQConfig(loftq_bits=loftq_bits, loftq_iter=loftq_iter)
     lora_config = LoraConfig(
         task_type=TaskType.CAUSAL_LM,
         inference_mode=True,
         r=lora_rank,
         lora_alpha=lora_alpha if lora_alpha is not None else lora_rank * 2,
-        lora_dropout=0.1,
-        target_modules=[name.strip() for name in lora_target.split(",")],
+        lora_dropout=lora_dropout,
+        target_modules=lora_target,
         init_lora_weights="loftq",
         loftq_config=loftq_config,
     )
 
     # Init LoftQ model
-    lora_model = get_peft_model(model, lora_config)
-    base_model: "PreTrainedModel" = lora_model.get_base_model()
+    print("Initializing LoftQ weights, it may be take several minutes, wait patiently.")
+    peft_model = get_peft_model(model, lora_config)
+    loftq_dir = os.path.join(output_dir, "loftq_init")
 
     # Save LoftQ model
-    setattr(lora_model.base_model.peft_config["default"], "base_model_name_or_path", save_dir)
-    setattr(lora_model.base_model.peft_config["default"], "init_lora_weights", True)
-    lora_model.save_pretrained(os.path.join(save_dir, "adapters"), safe_serialization=save_safetensors)
+    setattr(peft_model.peft_config["default"], "base_model_name_or_path", output_dir)
+    setattr(peft_model.peft_config["default"], "init_lora_weights", True)  # don't apply loftq again
+    peft_model.save_pretrained(loftq_dir, safe_serialization=save_safetensors)
+    print("Adapter weights saved in {}".format(loftq_dir))
 
     # Save base model
-    unwrap_model(base_model)
-    base_model.save_pretrained(save_dir, safe_serialization=save_safetensors)
-    tokenizer.save_pretrained(save_dir)
+    base_model: "PreTrainedModel" = peft_model.unload()
+    base_model.save_pretrained(output_dir, safe_serialization=save_safetensors)
+    tokenizer.save_pretrained(output_dir)
+    print("Model weights saved in {}".format(output_dir))
+
+    print("- Fine-tune this model with:")
+    print("model_name_or_path: {}".format(output_dir))
+    print("adapter_name_or_path: {}".format(loftq_dir))
+    print("finetuning_type: lora")
+    print("quantization_bit: {}".format(loftq_bits))
 
 
 if __name__ == "__main__":

scripts/pissa_init.py

@@ -0,0 +1,86 @@
+# coding=utf-8
+# Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
+#
+# This code is based on the HuggingFace's PEFT library.
+# https://github.com/huggingface/peft/blob/v0.11.0/examples/pissa_finetuning/preprocess.py
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+from typing import TYPE_CHECKING
+
+import fire
+from peft import LoraConfig, TaskType, get_peft_model
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+
+if TYPE_CHECKING:
+    from transformers import PreTrainedModel
+
+
+def quantize_pissa(
+    model_name_or_path: str,
+    output_dir: str,
+    pissa_iter: int = 4,
+    lora_alpha: int = None,
+    lora_rank: int = 16,
+    lora_dropout: float = 0,
+    lora_target: tuple = ("q_proj", "v_proj"),
+    save_safetensors: bool = True,
+):
+    r"""
+    Initializes LoRA weights with Principal Singular values and Singular vectors Adaptation (PiSSA)
+    Usage: python pissa_init.py --model_name_or_path path_to_model --output_dir output_dir
+    """
+    if isinstance(lora_target, str):
+        lora_target = [name.strip() for name in lora_target.split(",")]
+
+    tokenizer = AutoTokenizer.from_pretrained(model_name_or_path, trust_remote_code=True)
+    model = AutoModelForCausalLM.from_pretrained(model_name_or_path, trust_remote_code=True, torch_dtype="auto")
+
+    lora_config = LoraConfig(
+        task_type=TaskType.CAUSAL_LM,
+        r=lora_rank,
+        lora_alpha=lora_alpha if lora_alpha is not None else lora_rank * 2,
+        lora_dropout=lora_dropout,
+        target_modules=lora_target,
+        init_lora_weights="pissa" if pissa_iter == -1 else "pissa_niter_{}".format(pissa_iter),
+    )
+
+    # Init PiSSA model
+    peft_model = get_peft_model(model, lora_config)
+    pissa_dir = os.path.join(output_dir, "pissa_init")
+
+    # Save PiSSA model
+    setattr(peft_model.peft_config["default"], "init_lora_weights", True)  # don't apply pissa again
+    peft_model.save_pretrained(pissa_dir, safe_serialization=save_safetensors)
+    print("Adapter weights saved in {}".format(pissa_dir))
+
+    # Save base model
+    base_model: "PreTrainedModel" = peft_model.unload()
+    base_model.save_pretrained(output_dir, safe_serialization=save_safetensors)
+    tokenizer.save_pretrained(output_dir)
+    print("Model weights saved in {}".format(output_dir))
+
+    print("- Fine-tune this model with:")
+    print("model_name_or_path: {}".format(output_dir))
+    print("adapter_name_or_path: {}".format(pissa_dir))
+    print("finetuning_type: lora")
+    print("pissa_init: false")
+    print("pissa_convert: true")
+    print("- and optionally with:")
+    print("quantization_bit: 4")
+
+
+if __name__ == "__main__":
+    fire.Fire(quantize_pissa)

scripts/test_toolcall.py

@@ -1,3 +1,18 @@
+# coding=utf-8
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import json
 import os
 from typing import Sequence

setup.py

@@ -1,3 +1,17 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import os
 import re
 
@@ -23,14 +37,16 @@ extra_require = {
     "torch": ["torch>=1.13.1"],
     "torch-npu": ["torch==2.1.0", "torch-npu==2.1.0.post3", "decorator"],
     "metrics": ["nltk", "jieba", "rouge-chinese"],
-    "deepspeed": ["deepspeed>=0.10.0,<=0.14.0"],
+    "deepspeed": ["deepspeed>=0.10.0"],
     "bitsandbytes": ["bitsandbytes>=0.39.0"],
-    "vllm": ["vllm>=0.4.3"],
-    "galore": ["galore-torch"],
-    "badam": ["badam"],
-    "gptq": ["optimum>=1.16.0", "auto-gptq>=0.5.0"],
+    "hqq": ["hqq"],
+    "eetq": ["eetq"],
+    "gptq": ["optimum>=1.17.0", "auto-gptq>=0.5.0"],
     "awq": ["autoawq"],
     "aqlm": ["aqlm[gpu]>=1.1.0"],
+    "vllm": ["vllm>=0.4.3"],
+    "galore": ["galore-torch"],
+    "badam": ["badam>=1.2.1"],
     "qwen": ["transformers_stream_generator"],
     "modelscope": ["modelscope"],
     "dev": ["ruff", "pytest"],

src/api.py

@@ -1,3 +1,17 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import os
 
 import uvicorn

src/llamafactory/__init__.py

@@ -1,4 +1,39 @@
-# Level: api, webui > chat, eval, train > data, model > hparams > extras
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+r"""
+Efficient fine-tuning of large language models.
+
+Level:
+  api, webui > chat, eval, train > data, model > hparams > extras
+
+Dependency graph:
+  main:
+    transformers>=4.41.2
+    datasets>=2.16.0
+    accelerate>=0.30.1
+    peft>=0.11.1
+    trl>=0.8.6
+  attention:
+    transformers>=4.42.4 (gemma+fa2)
+  longlora:
+    transformers>=4.41.2,<=4.42.4
+  packing:
+    transformers>=4.41.2,<=4.42.4
+  patcher:
+    transformers==4.41.2 (chatglm)
+"""
 
 from .cli import VERSION
 

src/llamafactory/api/app.py

@@ -1,3 +1,17 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import os
 from contextlib import asynccontextmanager
 from typing import Optional

src/llamafactory/api/chat.py

@@ -1,3 +1,17 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import base64
 import io
 import json
@@ -78,9 +92,11 @@ def _process_request(
             raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="Invalid role")
 
         if message.role == Role.ASSISTANT and isinstance(message.tool_calls, list) and len(message.tool_calls):
-            name = message.tool_calls[0].function.name
-            arguments = message.tool_calls[0].function.arguments
-            content = json.dumps({"name": name, "argument": arguments}, ensure_ascii=False)
+            tool_calls = [
+                {"name": tool_call.function.name, "arguments": tool_call.function.arguments}
+                for tool_call in message.tool_calls
+            ]
+            content = json.dumps(tool_calls, ensure_ascii=False)
             input_messages.append({"role": ROLE_MAPPING[Role.FUNCTION], "content": content})
         elif isinstance(message.content, list):
             for input_item in message.content:
@@ -104,7 +120,7 @@ def _process_request(
     if isinstance(tool_list, list) and len(tool_list):
         try:
             tools = json.dumps([dictify(tool.function) for tool in tool_list], ensure_ascii=False)
-        except Exception:
+        except json.JSONDecodeError:
             raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="Invalid tools")
     else:
         tools = None
@@ -146,15 +162,17 @@ async def create_chat_completion_response(
     choices = []
     for i, response in enumerate(responses):
         if tools:
-            result = chat_model.engine.template.format_tools.extract(response.response_text)
+            result = chat_model.engine.template.extract_tool(response.response_text)
         else:
             result = response.response_text
 
-        if isinstance(result, tuple):
-            name, arguments = result
-            function = Function(name=name, arguments=arguments)
-            tool_call = FunctionCall(id="call_{}".format(uuid.uuid4().hex), function=function)
-            response_message = ChatCompletionMessage(role=Role.ASSISTANT, tool_calls=[tool_call])
+        if isinstance(result, list):
+            tool_calls = []
+            for tool in result:
+                function = Function(name=tool[0], arguments=tool[1])
+                tool_calls.append(FunctionCall(id="call_{}".format(uuid.uuid4().hex), function=function))
+
+            response_message = ChatCompletionMessage(role=Role.ASSISTANT, tool_calls=tool_calls)
             finish_reason = Finish.TOOL
         else:
             response_message = ChatCompletionMessage(role=Role.ASSISTANT, content=result)

src/llamafactory/api/common.py

@@ -1,3 +1,17 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import json
 from typing import TYPE_CHECKING, Any, Dict
 

src/llamafactory/api/protocol.py

@@ -1,3 +1,17 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import time
 from enum import Enum, unique
 from typing import Any, Dict, List, Optional, Union
@@ -82,7 +96,7 @@ class ChatCompletionRequest(BaseModel):
     model: str
     messages: List[ChatMessage]
     tools: Optional[List[FunctionAvailable]] = None
-    do_sample: bool = True
+    do_sample: Optional[bool] = None
     temperature: Optional[float] = None
     top_p: Optional[float] = None
     n: int = 1

src/llamafactory/chat/__init__.py

@@ -1,3 +1,17 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 from .base_engine import BaseEngine
 from .chat_model import ChatModel
 

src/llamafactory/chat/base_engine.py

@@ -1,3 +1,17 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 from abc import ABC, abstractmethod
 from dataclasses import dataclass
 from typing import TYPE_CHECKING, Any, AsyncGenerator, Dict, List, Literal, Optional, Sequence, Union
@@ -36,11 +50,6 @@ class BaseEngine(ABC):
         generating_args: "GeneratingArguments",
     ) -> None: ...
 
-    @abstractmethod
-    async def start(
-        self,
-    ) -> None: ...
-
     @abstractmethod
     async def chat(
         self,

src/llamafactory/chat/chat_model.py

@@ -1,4 +1,22 @@
+# Copyright 2024 THUDM and the LlamaFactory team.
+#
+# This code is inspired by the THUDM's ChatGLM implementation.
+# https://github.com/THUDM/ChatGLM-6B/blob/main/cli_demo.py
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import asyncio
+import os
 from threading import Thread
 from typing import TYPE_CHECKING, Any, AsyncGenerator, Dict, Generator, List, Optional, Sequence
 
@@ -14,7 +32,7 @@ if TYPE_CHECKING:
     from .base_engine import BaseEngine, Response
 
 
-def _start_background_loop(loop: asyncio.AbstractEventLoop) -> None:
+def _start_background_loop(loop: "asyncio.AbstractEventLoop") -> None:
     asyncio.set_event_loop(loop)
     loop.run_forever()
 
@@ -32,7 +50,6 @@ class ChatModel:
         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)
 
     def chat(
         self,
@@ -99,13 +116,11 @@ class ChatModel:
 
 
 def run_chat() -> None:
-    try:
-        import platform
-
-        if platform.system() != "Windows":
+    if os.name != "nt":
+        try:
             import readline  # noqa: F401
-    except ImportError:
-        print("Install `readline` for a better experience.")
+        except ImportError:
+            print("Install `readline` for a better experience.")
 
     chat_model = ChatModel()
     messages = []

src/llamafactory/chat/hf_engine.py

@@ -1,3 +1,17 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import asyncio
 import concurrent.futures
 import os
@@ -40,11 +54,19 @@ class HuggingfaceEngine(BaseEngine):
         self.tokenizer = tokenizer_module["tokenizer"]
         self.processor = tokenizer_module["processor"]
         self.tokenizer.padding_side = "left" if self.can_generate else "right"
-        self.template = get_template_and_fix_tokenizer(self.tokenizer, data_args.template)
+        self.template = get_template_and_fix_tokenizer(self.tokenizer, data_args.template, data_args.tool_format)
         self.model = load_model(
             self.tokenizer, model_args, finetuning_args, is_trainable=False, add_valuehead=(not self.can_generate)
         )  # must after fixing tokenizer to resize vocab
         self.generating_args = generating_args.to_dict()
+        try:
+            asyncio.get_event_loop()
+        except RuntimeError:
+            logger.warning("There is no current event loop, creating a new one.")
+            loop = asyncio.new_event_loop()
+            asyncio.set_event_loop(loop)
+
+        self.semaphore = asyncio.Semaphore(int(os.environ.get("MAX_CONCURRENT", "1")))
 
     @staticmethod
     def _process_args(
@@ -97,7 +119,7 @@ class HuggingfaceEngine(BaseEngine):
         stop: Optional[Union[str, List[str]]] = input_kwargs.pop("stop", None)
 
         if stop is not None:
-            logger.warning("Stop parameter is not supported in Huggingface engine yet.")
+            logger.warning("Stop parameter is not supported by the huggingface engine yet.")
 
         generating_args = generating_args.copy()
         generating_args.update(
@@ -245,9 +267,6 @@ class HuggingfaceEngine(BaseEngine):
 
         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]],
@@ -272,7 +291,7 @@ class HuggingfaceEngine(BaseEngine):
             image,
             input_kwargs,
         )
-        async with self._semaphore:
+        async with self.semaphore:
             with concurrent.futures.ThreadPoolExecutor() as pool:
                 return await loop.run_in_executor(pool, self._chat, *input_args)
 
@@ -300,7 +319,7 @@ class HuggingfaceEngine(BaseEngine):
             image,
             input_kwargs,
         )
-        async with self._semaphore:
+        async with self.semaphore:
             with concurrent.futures.ThreadPoolExecutor() as pool:
                 stream = self._stream_chat(*input_args)
                 while True:
@@ -319,6 +338,6 @@ class HuggingfaceEngine(BaseEngine):
 
         loop = asyncio.get_running_loop()
         input_args = (self.model, self.tokenizer, batch_input, input_kwargs)
-        async with self._semaphore:
+        async with self.semaphore:
             with concurrent.futures.ThreadPoolExecutor() as pool:
                 return await loop.run_in_executor(pool, self._get_scores, *input_args)

src/llamafactory/chat/vllm_engine.py

@@ -1,11 +1,26 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import uuid
-from typing import TYPE_CHECKING, AsyncGenerator, AsyncIterator, Dict, List, Optional, Sequence, Union
+from typing import TYPE_CHECKING, Any, AsyncGenerator, AsyncIterator, Dict, List, Optional, Sequence, Union
 
 from ..data import get_template_and_fix_tokenizer
 from ..extras.logging import get_logger
 from ..extras.misc import get_device_count
-from ..extras.packages import is_vllm_available
+from ..extras.packages import is_vllm_available, is_vllm_version_greater_than_0_5, is_vllm_version_greater_than_0_5_1
 from ..model import load_config, load_tokenizer
+from ..model.model_utils.quantization import QuantizationMethod
 from ..model.model_utils.visual import LlavaMultiModalProjectorForYiVLForVLLM
 from .base_engine import BaseEngine, Response
 
@@ -13,7 +28,13 @@ from .base_engine import BaseEngine, Response
 if is_vllm_available():
     from vllm import AsyncEngineArgs, AsyncLLMEngine, RequestOutput, SamplingParams
     from vllm.lora.request import LoRARequest
-    from vllm.sequence import MultiModalData
+
+    if is_vllm_version_greater_than_0_5_1():
+        pass
+    elif is_vllm_version_greater_than_0_5():
+        from vllm.multimodal.image import ImagePixelData
+    else:
+        from vllm.sequence import MultiModalData
 
 
 if TYPE_CHECKING:
@@ -35,20 +56,25 @@ class VllmEngine(BaseEngine):
         generating_args: "GeneratingArguments",
     ) -> None:
         config = load_config(model_args)  # may download model from ms hub
+        if getattr(config, "quantization_config", None):  # gptq models should use float16
+            quantization_config: Dict[str, Any] = getattr(config, "quantization_config", None)
+            quant_method = quantization_config.get("quant_method", "")
+            if quant_method == QuantizationMethod.GPTQ and model_args.infer_dtype == "auto":
+                model_args.infer_dtype = "float16"
 
         self.can_generate = finetuning_args.stage == "sft"
         tokenizer_module = load_tokenizer(model_args)
         self.tokenizer = tokenizer_module["tokenizer"]
         self.processor = tokenizer_module["processor"]
         self.tokenizer.padding_side = "left"
-        self.template = get_template_and_fix_tokenizer(self.tokenizer, data_args.template)
+        self.template = get_template_and_fix_tokenizer(self.tokenizer, data_args.template, data_args.tool_format)
         self.generating_args = generating_args.to_dict()
 
         engine_args = {
             "model": model_args.model_name_or_path,
             "trust_remote_code": True,
             "download_dir": model_args.cache_dir,
-            "dtype": model_args.vllm_dtype,
+            "dtype": model_args.infer_dtype,
             "max_model_len": model_args.vllm_maxlen,
             "tensor_parallel_size": get_device_count() or 1,
             "gpu_memory_utilization": model_args.vllm_gpu_util,
@@ -106,7 +132,12 @@ class VllmEngine(BaseEngine):
         if self.processor is not None and image is not None:  # add image features
             image_processor: "BaseImageProcessor" = getattr(self.processor, "image_processor")
             pixel_values = image_processor(image, return_tensors="pt")["pixel_values"]
-            multi_modal_data = MultiModalData(type=MultiModalData.Type.IMAGE, data=pixel_values)
+            if is_vllm_version_greater_than_0_5_1():
+                multi_modal_data = {"image": pixel_values}
+            elif is_vllm_version_greater_than_0_5():
+                multi_modal_data = ImagePixelData(image=pixel_values)
+            else:  # TODO: remove vllm 0.4.3 support
+                multi_modal_data = MultiModalData(type=MultiModalData.Type.IMAGE, data=pixel_values)
         else:
             multi_modal_data = None
 
@@ -162,9 +193,6 @@ class VllmEngine(BaseEngine):
         )
         return result_generator
 
-    async def start(self) -> None:
-        pass
-
     async def chat(
         self,
         messages: Sequence[Dict[str, str]],

src/llamafactory/cli.py

@@ -1,3 +1,17 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import os
 import random
 import subprocess
@@ -60,7 +74,7 @@ class Command(str, Enum):
 
 
 def main():
-    command = sys.argv.pop(1)
+    command = sys.argv.pop(1) if len(sys.argv) != 1 else Command.HELP
     if command == Command.API:
         run_api()
     elif command == Command.CHAT:
@@ -77,7 +91,7 @@ def main():
             master_addr = os.environ.get("MASTER_ADDR", "127.0.0.1")
             master_port = os.environ.get("MASTER_PORT", str(random.randint(20001, 29999)))
             logger.info("Initializing distributed tasks at: {}:{}".format(master_addr, master_port))
-            subprocess.run(
+            process = subprocess.run(
                 (
                     "torchrun --nnodes {nnodes} --node_rank {node_rank} --nproc_per_node {nproc_per_node} "
                     "--master_addr {master_addr} --master_port {master_port} {file_name} {args}"
@@ -92,6 +106,7 @@ def main():
                 ),
                 shell=True,
             )
+            sys.exit(process.returncode)
         else:
             run_exp()
     elif command == Command.WEBDEMO:

src/llamafactory/data/__init__.py

@@ -1,4 +1,18 @@
-from .collator import KTODataCollatorWithPadding, PairwiseDataCollatorWithPadding
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .collator import KTODataCollatorWithPadding, PairwiseDataCollatorWithPadding, SFTDataCollatorWith4DAttentionMask
 from .data_utils import Role, split_dataset
 from .loader import get_dataset
 from .template import TEMPLATES, Template, get_template_and_fix_tokenizer
@@ -7,6 +21,7 @@ from .template import TEMPLATES, Template, get_template_and_fix_tokenizer
 __all__ = [
     "KTODataCollatorWithPadding",
     "PairwiseDataCollatorWithPadding",
+    "SFTDataCollatorWith4DAttentionMask",
     "Role",
     "split_dataset",
     "get_dataset",

src/llamafactory/data/aligner.py

@@ -1,3 +1,17 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import os
 from functools import partial
 from typing import TYPE_CHECKING, Any, Dict, List, Union
@@ -10,6 +24,7 @@ from .data_utils import Role
 
 if TYPE_CHECKING:
     from datasets import Dataset, IterableDataset
+    from transformers import Seq2SeqTrainingArguments
 
     from ..hparams import DataArguments
     from .parser import DatasetAttr
@@ -175,7 +190,10 @@ def convert_sharegpt(
 
 
 def align_dataset(
-    dataset: Union["Dataset", "IterableDataset"], dataset_attr: "DatasetAttr", data_args: "DataArguments"
+    dataset: Union["Dataset", "IterableDataset"],
+    dataset_attr: "DatasetAttr",
+    data_args: "DataArguments",
+    training_args: "Seq2SeqTrainingArguments",
 ) -> Union["Dataset", "IterableDataset"]:
     r"""
     Aligned dataset:
@@ -208,7 +226,7 @@ def align_dataset(
     if not data_args.streaming:
         kwargs = dict(
             num_proc=data_args.preprocessing_num_workers,
-            load_from_cache_file=(not data_args.overwrite_cache),
+            load_from_cache_file=(not data_args.overwrite_cache) or (training_args.local_process_index != 0),
             desc="Converting format of dataset",
         )
 

src/llamafactory/data/collator.py

@@ -1,17 +1,91 @@
+# Copyright 2024 OpenAccess AI Collective and the LlamaFactory team.
+#
+# This code is inspired by the OpenAccess AI Collective's axolotl library.
+# https://github.com/OpenAccess-AI-Collective/axolotl/blob/main/src/axolotl/monkeypatch/utils.py
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 from dataclasses import dataclass
-from typing import Any, Dict, Sequence
+from typing import Any, Dict, Literal, Sequence
 
 import torch
 from transformers import DataCollatorForSeq2Seq
 
 
+def prepare_4d_attention_mask(attention_mask_with_indices: "torch.Tensor", dtype: "torch.dtype") -> "torch.Tensor":
+    r"""
+    Expands the attention mask with indices from (batch_size, seq_len) to (batch_size, 1, seq_len, seq_len),
+    while handles packed sequences and transforms the mask to lower triangular form to prevent future peeking.
+
+    e.g.
+    ```python
+    # input
+    [[1, 1, 2, 2, 2, 0]]
+    # output
+    [
+        [
+            [
+                [o, x, x, x, x, x],
+                [o, o, x, x, x, x],
+                [x, x, o, x, x, x],
+                [x, x, o, o, x, x],
+                [x, x, o, o, o, x],
+                [x, x, x, x, x, x],
+            ]
+        ]
+    ]
+    ```
+    where `o` equals to `0.0`, `x` equals to `min_dtype`.
+    """
+    bsz, seq_len = attention_mask_with_indices.size()
+    min_dtype = torch.finfo(dtype).min
+    expanded_mask = attention_mask_with_indices[:, None, None, :].expand(bsz, 1, seq_len, seq_len)
+    # Create a binary mask from the original mask where zeros remain zeros and all other values are set to one
+    padding_mask = torch.where(expanded_mask != 0, 1, 0)
+    # Create a block-diagonal mask.
+    attention_mask_4d = torch.eq(expanded_mask, expanded_mask.transpose(-1, -2)).int() * padding_mask
+    # Use the lower triangular mask to zero out the upper triangular part
+    attention_mask_4d *= torch.tril(torch.ones((seq_len, seq_len), dtype=torch.long))
+    # Invert the attention mask.
+    attention_mask_4d = torch.where(attention_mask_4d != 0, torch.tensor(0, dtype=dtype), min_dtype)
+    return attention_mask_4d
+
+
+@dataclass
+class SFTDataCollatorWith4DAttentionMask(DataCollatorForSeq2Seq):
+    r"""
+    Data collator for 4d attention mask.
+    """
+
+    block_diag_attn: bool = False
+    attn_implementation: Literal["eager", "sdpa", "flash_attention_2"] = "eager"
+    compute_dtype: "torch.dtype" = torch.float32
+
+    def __call__(self, features: Sequence[Dict[str, Any]]) -> Dict[str, "torch.Tensor"]:
+        features = super().__call__(features)
+        if self.block_diag_attn and self.attn_implementation != "flash_attention_2":
+            features["attention_mask"] = prepare_4d_attention_mask(features["attention_mask"], self.compute_dtype)
+
+        return features
+
+
 @dataclass
 class PairwiseDataCollatorWithPadding(DataCollatorForSeq2Seq):
     r"""
     Data collator for pairwise data.
     """
 
-    def __call__(self, features: Sequence[Dict[str, Any]]) -> Dict[str, torch.Tensor]:
+    def __call__(self, features: Sequence[Dict[str, Any]]) -> Dict[str, "torch.Tensor"]:
         r"""
         Pads batched data to the longest sequence in the batch.
 
@@ -43,7 +117,7 @@ class KTODataCollatorWithPadding(DataCollatorForSeq2Seq):
     Data collator for KTO data.
     """
 
-    def __call__(self, features: Sequence[Dict[str, Any]]) -> Dict[str, torch.Tensor]:
+    def __call__(self, features: Sequence[Dict[str, Any]]) -> Dict[str, "torch.Tensor"]:
         target_features = []
         kl_features = []
         kto_tags = []

src/llamafactory/data/data_utils.py

@@ -1,14 +1,27 @@
-from enum import Enum, unique
-from typing import TYPE_CHECKING, Dict, List, Tuple, Union
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
 
-from datasets import concatenate_datasets, interleave_datasets
+from enum import Enum, unique
+from typing import TYPE_CHECKING, Dict, List, Optional, Sequence, Set, TypedDict, Union
+
+from datasets import DatasetDict, concatenate_datasets, interleave_datasets
 
 from ..extras.logging import get_logger
 
 
 if TYPE_CHECKING:
     from datasets import Dataset, IterableDataset
-    from transformers import Seq2SeqTrainingArguments
 
     from ..hparams import DataArguments
 
@@ -16,6 +29,9 @@ if TYPE_CHECKING:
 logger = get_logger(__name__)
 
 
+SLOTS = Sequence[Union[str, Set[str], Dict[str, str]]]
+
+
 @unique
 class Role(str, Enum):
     USER = "user"
@@ -25,31 +41,29 @@ class Role(str, Enum):
     OBSERVATION = "observation"
 
 
-def infer_max_len(source_len: int, target_len: int, max_len: int, reserved_label_len: int) -> Tuple[int, int]:
-    max_target_len = int(max_len * (target_len / (source_len + target_len)))
-    max_target_len = max(max_target_len, reserved_label_len)
-    max_source_len = max_len - min(max_target_len, target_len)
-    return max_source_len, max_target_len
+class DatasetModule(TypedDict):
+    train_dataset: Optional[Union["Dataset", "IterableDataset"]]
+    eval_dataset: Optional[Union["Dataset", "IterableDataset"]]
 
 
 def merge_dataset(
-    all_datasets: List[Union["Dataset", "IterableDataset"]],
-    data_args: "DataArguments",
-    training_args: "Seq2SeqTrainingArguments",
+    all_datasets: List[Union["Dataset", "IterableDataset"]], data_args: "DataArguments", seed: int
 ) -> 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,
+            seed=seed,
             stopping_strategy="first_exhausted" if data_args.mix_strategy.endswith("under") else "all_exhausted",
         )
     else:
@@ -57,22 +71,17 @@ def merge_dataset(
 
 
 def split_dataset(
-    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
-            if data_args.streaming:
-                dataset = dataset.shuffle(buffer_size=data_args.buffer_size, seed=training_args.seed)
-                val_set = dataset.take(int(data_args.val_size))
-                train_set = dataset.skip(int(data_args.val_size))
-                return {"train_dataset": train_set, "eval_dataset": val_set}
-            else:
-                val_size = int(data_args.val_size) if data_args.val_size > 1 else data_args.val_size
-                dataset = dataset.train_test_split(test_size=val_size, seed=training_args.seed)
-                return {"train_dataset": dataset["train"], "eval_dataset": dataset["test"]}
-        else:
-            if data_args.streaming:
-                dataset = dataset.shuffle(buffer_size=data_args.buffer_size, seed=training_args.seed)
-            return {"train_dataset": dataset}
-    else:  # do_eval or do_predict
-        return {"eval_dataset": dataset}
+    dataset: Union["Dataset", "IterableDataset"], data_args: "DataArguments", seed: int
+) -> "DatasetDict":
+    r"""
+    Splits the dataset and returns a dataset dict containing train set (required) and validation set (optional).
+    """
+    if data_args.streaming:
+        dataset = dataset.shuffle(buffer_size=data_args.buffer_size, seed=seed)
+        val_set = dataset.take(int(data_args.val_size))
+        train_set = dataset.skip(int(data_args.val_size))
+        return DatasetDict({"train": train_set, "validation": val_set})
+    else:
+        val_size = int(data_args.val_size) if data_args.val_size > 1 else data_args.val_size
+        dataset = dataset.train_test_split(test_size=val_size, seed=seed)
+        return DatasetDict({"train": dataset["train"], "validation": dataset["test"]})

src/llamafactory/data/formatter.py

@@ -1,83 +1,36 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import json
 import re
 from abc import ABC, abstractmethod
 from dataclasses import dataclass, field
-from typing import Any, Dict, List, Literal, Optional, Sequence, Set, Tuple, Union
+from typing import List, Literal, Optional, Tuple, Union
 
-
-SLOTS = Sequence[Union[str, Set[str], Dict[str, str]]]
-
-
-JSON_FORMAT_PROMPT = (
-    """, in a JSON format representing the kwargs (e.g. ```{"input": "hello world", "num_beams": 5}```)"""
-)
-
-
-TOOL_SYSTEM_PROMPT = (
-    "You have access to the following tools:\n{tool_text}"
-    "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"
-)
-
-
-def default_tool_formatter(tools: List[Dict[str, Any]]) -> str:
-    tool_text = ""
-    tool_names = []
-    for tool in tools:
-        param_text = ""
-        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 ""
-            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(
-            name=tool["name"], desc=tool.get("description", ""), args=param_text
-        )
-        tool_names.append(tool["name"])
-
-    return TOOL_SYSTEM_PROMPT.format(
-        tool_text=tool_text, tool_names=", ".join(tool_names), format_prompt=JSON_FORMAT_PROMPT
-    )
-
-
-def default_tool_extractor(content: str) -> Union[str, Tuple[str, str]]:
-    regex = re.compile(r"Action:\s*([a-zA-Z0-9_]+).*?Action Input:\s*(.*)", re.DOTALL)
-    action_match = re.search(regex, content)
-    if not action_match:
-        return content
-
-    tool_name = action_match.group(1).strip()
-    tool_input = action_match.group(2).strip().strip('"').strip("```")
-    try:
-        arguments = json.loads(tool_input)
-    except json.JSONDecodeError:
-        return content
-
-    return tool_name, json.dumps(arguments, ensure_ascii=False)
+from .data_utils import SLOTS
+from .tool_utils import DefaultToolUtils, GLM4ToolUtils
 
 
 @dataclass
 class Formatter(ABC):
     slots: SLOTS = field(default_factory=list)
-    tool_format: Optional[Literal["default"]] = None
+    tool_format: Optional[Literal["default", "glm4"]] = None
 
     @abstractmethod
     def apply(self, **kwargs) -> SLOTS: ...
 
-    def extract(self, content: str) -> Union[str, Tuple[str, str]]:
+    def extract(self, content: str) -> Union[str, List[Tuple[str, str]]]:
         raise NotImplementedError
 
 
@@ -128,34 +81,37 @@ class StringFormatter(Formatter):
 @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.")
+        if self.tool_format == "default":
+            self.slots = DefaultToolUtils.get_function_slots() + self.slots
+        elif self.tool_format == "glm4":
+            self.slots = GLM4ToolUtils.get_function_slots() + self.slots
+        else:
+            raise NotImplementedError("Tool format {} was not found.".format(self.tool_format))
 
     def apply(self, **kwargs) -> SLOTS:
         content = kwargs.pop("content")
+        functions: List[Tuple[str, str]] = []
         try:
-            function = json.loads(content)
-            name = function["name"]
-            arguments = json.dumps(function["arguments"], ensure_ascii=False)
-        except Exception:
-            name, arguments = "", ""
+            tool_calls = json.loads(content)
+            if not isinstance(tool_calls, list):  # parallel function call
+                tool_calls = [tool_calls]
+
+            for tool_call in tool_calls:
+                functions.append((tool_call["name"], json.dumps(tool_call["arguments"], ensure_ascii=False)))
+
+        except json.JSONDecodeError:
+            functions = []
 
         elements = []
-        for slot in self.slots:
-            if isinstance(slot, str):
-                slot = slot.replace("{{name}}", name).replace("{{arguments}}", arguments)
-                elements.append(slot)
-            elif isinstance(slot, (dict, set)):
-                elements.append(slot)
-            else:
-                raise RuntimeError("Input must be string, set[str] or dict[str, str], got {}".format(type(slot)))
+        for name, arguments in functions:
+            for slot in self.slots:
+                if isinstance(slot, str):
+                    slot = slot.replace("{{name}}", name).replace("{{arguments}}", arguments)
+                    elements.append(slot)
+                elif isinstance(slot, (dict, set)):
+                    elements.append(slot)
+                else:
+                    raise RuntimeError("Input must be string, set[str] or dict[str, str], got {}".format(type(slot)))
 
         return elements
 
@@ -163,25 +119,22 @@ class FunctionFormatter(Formatter):
 @dataclass
 class ToolFormatter(Formatter):
     def __post_init__(self):
-        if self.tool_format is None:
-            raise ValueError("Tool format was not found.")
+        if self.tool_format == "default":
+            self._tool_formatter = DefaultToolUtils.tool_formatter
+            self._tool_extractor = DefaultToolUtils.tool_extractor
+        elif self.tool_format == "glm4":
+            self._tool_formatter = GLM4ToolUtils.tool_formatter
+            self._tool_extractor = GLM4ToolUtils.tool_extractor
+        else:
+            raise NotImplementedError("Tool format {} was not found.".format(self.tool_format))
 
     def apply(self, **kwargs) -> SLOTS:
         content = kwargs.pop("content")
         try:
             tools = json.loads(content)
-            if not len(tools):
-                return [""]
-
-            if self.tool_format == "default":
-                return [default_tool_formatter(tools)]
-            else:
-                raise NotImplementedError
-        except Exception:
+            return [self._tool_formatter(tools) if len(tools) != 0 else ""]
+        except json.JSONDecodeError:
             return [""]
 
-    def extract(self, content: str) -> Union[str, Tuple[str, str]]:
-        if self.tool_format == "default":
-            return default_tool_extractor(content)
-        else:
-            raise NotImplementedError
+    def extract(self, content: str) -> Union[str, List[Tuple[str, str]]]:
+        return self._tool_extractor(content)

src/llamafactory/data/loader.py

@@ -1,16 +1,30 @@
-import inspect
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import os
 import sys
-from typing import TYPE_CHECKING, Literal, Optional, Union
+from typing import TYPE_CHECKING, Dict, Literal, Optional, Sequence, Union
 
 import numpy as np
-from datasets import load_dataset, load_from_disk
+from datasets import DatasetDict, load_dataset, load_from_disk
+from transformers.utils.versions import require_version
 
 from ..extras.constants import FILEEXT2TYPE
 from ..extras.logging import get_logger
 from ..extras.misc import has_tokenized_data
 from .aligner import align_dataset
-from .data_utils import merge_dataset
+from .data_utils import merge_dataset, split_dataset
 from .parser import get_dataset_list
 from .preprocess import get_preprocess_and_print_func
 from .template import get_template_and_fix_tokenizer
@@ -18,20 +32,22 @@ from .template import get_template_and_fix_tokenizer
 
 if TYPE_CHECKING:
     from datasets import Dataset, IterableDataset
-    from transformers import ProcessorMixin, Seq2SeqTrainingArguments
-    from transformers.tokenization_utils import PreTrainedTokenizer
+    from transformers import PreTrainedTokenizer, ProcessorMixin, Seq2SeqTrainingArguments
 
     from ..hparams import DataArguments, ModelArguments
+    from .data_utils import DatasetModule
     from .parser import DatasetAttr
+    from .template import Template
 
 
 logger = get_logger(__name__)
 
 
-def load_single_dataset(
+def _load_single_dataset(
     dataset_attr: "DatasetAttr",
     model_args: "ModelArguments",
     data_args: "DataArguments",
+    training_args: "Seq2SeqTrainingArguments",
 ) -> Union["Dataset", "IterableDataset"]:
     logger.info("Loading dataset {}...".format(dataset_attr))
     data_path, data_name, data_dir, data_files = None, None, None, None
@@ -67,41 +83,34 @@ def load_single_dataset(
         raise NotImplementedError("Unknown load type: {}.".format(dataset_attr.load_from))
 
     if dataset_attr.load_from == "ms_hub":
-        try:
-            from modelscope import MsDataset
-            from modelscope.utils.config_ds import MS_DATASETS_CACHE
+        require_version("modelscope>=1.11.0", "To fix: pip install modelscope>=1.11.0")
+        from modelscope import MsDataset
+        from modelscope.utils.config_ds import MS_DATASETS_CACHE
 
-            cache_dir = model_args.cache_dir or MS_DATASETS_CACHE
-            dataset = MsDataset.load(
-                dataset_name=data_path,
-                subset_name=data_name,
-                data_dir=data_dir,
-                data_files=data_files,
-                split=data_args.split,
-                cache_dir=cache_dir,
-                token=model_args.ms_hub_token,
-                use_streaming=(data_args.streaming and (dataset_attr.load_from != "file")),
-            )
-            if isinstance(dataset, MsDataset):
-                dataset = dataset.to_hf_dataset()
-        except ImportError:
-            raise ImportError("Please install modelscope via `pip install modelscope -U`")
+        cache_dir = model_args.cache_dir or MS_DATASETS_CACHE
+        dataset = MsDataset.load(
+            dataset_name=data_path,
+            subset_name=data_name,
+            data_dir=data_dir,
+            data_files=data_files,
+            split=dataset_attr.split,
+            cache_dir=cache_dir,
+            token=model_args.ms_hub_token,
+            use_streaming=(data_args.streaming and (dataset_attr.load_from != "file")),
+        )
+        if isinstance(dataset, MsDataset):
+            dataset = dataset.to_hf_dataset()
     else:
-        if "trust_remote_code" in inspect.signature(load_dataset).parameters:  # for datasets==2.16.0
-            kwargs = {"trust_remote_code": True}
-        else:
-            kwargs = {}
-
         dataset = load_dataset(
             path=data_path,
             name=data_name,
             data_dir=data_dir,
             data_files=data_files,
-            split=data_args.split,
+            split=dataset_attr.split,
             cache_dir=model_args.cache_dir,
             token=model_args.hf_hub_token,
             streaming=(data_args.streaming and (dataset_attr.load_from != "file")),
-            **kwargs,
+            trust_remote_code=True,
         )
 
     if data_args.streaming and (dataset_attr.load_from == "file"):  # faster than specifying streaming=True
@@ -123,7 +132,67 @@ def load_single_dataset(
         max_samples = min(data_args.max_samples, len(dataset))
         dataset = dataset.select(range(max_samples))
 
-    return align_dataset(dataset, dataset_attr, data_args)
+    return align_dataset(dataset, dataset_attr, data_args, training_args)
+
+
+def _get_merged_dataset(
+    dataset_names: Optional[Sequence[str]],
+    model_args: "ModelArguments",
+    data_args: "DataArguments",
+    training_args: "Seq2SeqTrainingArguments",
+    stage: Literal["pt", "sft", "rm", "ppo", "kto"],
+) -> Optional[Union["Dataset", "IterableDataset"]]:
+    if dataset_names is None:
+        return None
+
+    datasets = []
+    for dataset_attr in get_dataset_list(dataset_names, data_args.dataset_dir):
+        if (stage == "rm" and dataset_attr.ranking is False) or (stage != "rm" and dataset_attr.ranking is True):
+            raise ValueError("The dataset is not applicable in the current training stage.")
+
+        datasets.append(_load_single_dataset(dataset_attr, model_args, data_args, training_args))
+
+    return merge_dataset(datasets, data_args, seed=training_args.seed)
+
+
+def _get_preprocessed_dataset(
+    dataset: Optional[Union["Dataset", "IterableDataset"]],
+    data_args: "DataArguments",
+    training_args: "Seq2SeqTrainingArguments",
+    stage: Literal["pt", "sft", "rm", "ppo", "kto"],
+    template: "Template",
+    tokenizer: "PreTrainedTokenizer",
+    processor: Optional["ProcessorMixin"] = None,
+    is_eval: bool = False,
+) -> Optional[Union["Dataset", "IterableDataset"]]:
+    if dataset is None:
+        return None
+
+    preprocess_func, print_function = get_preprocess_and_print_func(
+        data_args, stage, template, tokenizer, processor, do_generate=(training_args.predict_with_generate and is_eval)
+    )
+    column_names = list(next(iter(dataset)).keys())
+    kwargs = {}
+    if not data_args.streaming:
+        kwargs = dict(
+            num_proc=data_args.preprocessing_num_workers,
+            load_from_cache_file=(not data_args.overwrite_cache) or (training_args.local_process_index != 0),
+            desc="Running tokenizer on dataset",
+        )
+
+    dataset = dataset.map(preprocess_func, batched=True, remove_columns=column_names, **kwargs)
+
+    if training_args.should_log:
+        try:
+            print("eval example:" if is_eval else "training example:")
+            print_function(next(iter(dataset)))
+        except StopIteration:
+            if stage == "pt":
+                raise RuntimeError("Cannot find sufficient samples, consider increasing dataset size.")
+            else:
+                raise RuntimeError("Cannot find valid samples, check `data/README.md` for the data format.")
+
+    return dataset
 
 
 def get_dataset(
@@ -133,8 +202,8 @@ def get_dataset(
     stage: Literal["pt", "sft", "rm", "ppo", "kto"],
     tokenizer: "PreTrainedTokenizer",
     processor: Optional["ProcessorMixin"] = None,
-) -> Union["Dataset", "IterableDataset"]:
-    template = get_template_and_fix_tokenizer(tokenizer, data_args.template)
+) -> "DatasetModule":
+    template = get_template_and_fix_tokenizer(tokenizer, data_args.template, data_args.tool_format)
     if data_args.train_on_prompt and template.efficient_eos:
         raise ValueError("Current template does not support `train_on_prompt`.")
 
@@ -142,54 +211,66 @@ def get_dataset(
     if data_args.tokenized_path is not None:
         if has_tokenized_data(data_args.tokenized_path):
             logger.warning("Loading dataset from disk will ignore other data arguments.")
-            dataset = load_from_disk(data_args.tokenized_path)
+            dataset_dict: "DatasetDict" = load_from_disk(data_args.tokenized_path)
             logger.info("Loaded tokenized dataset from {}.".format(data_args.tokenized_path))
+
+            dataset_module: Dict[str, "Dataset"] = {}
+            if "train" in dataset_dict:
+                dataset_module["train_dataset"] = dataset_dict["train"]
+            if "validation" in dataset_dict:
+                dataset_module["eval_dataset"] = dataset_dict["validation"]
+
             if data_args.streaming:
-                dataset = dataset.to_iterable_dataset()
-            return dataset
+                dataset_module = {k: v.to_iterable_dataset() for k, v in dataset_module.items()}
+
+            return dataset_module
 
         if data_args.streaming:
             raise ValueError("Turn off `streaming` when saving dataset to disk.")
 
+    # Load and preprocess dataset
     with training_args.main_process_first(desc="load dataset"):
-        all_datasets = []
-        for dataset_attr in get_dataset_list(data_args):
-            if (stage == "rm" and dataset_attr.ranking is False) or (stage != "rm" and dataset_attr.ranking is True):
-                raise ValueError("The dataset is not applicable in the current training stage.")
-
-            all_datasets.append(load_single_dataset(dataset_attr, model_args, data_args))
-        dataset = merge_dataset(all_datasets, data_args, training_args)
+        dataset = _get_merged_dataset(data_args.dataset, model_args, data_args, training_args, stage)
+        eval_dataset = _get_merged_dataset(data_args.eval_dataset, model_args, data_args, training_args, stage)
 
     with training_args.main_process_first(desc="pre-process dataset"):
-        preprocess_func, print_function = get_preprocess_and_print_func(
-            data_args, training_args, stage, template, tokenizer, processor
+        dataset = _get_preprocessed_dataset(
+            dataset, data_args, training_args, stage, template, tokenizer, processor, is_eval=False
+        )
+        eval_dataset = _get_preprocessed_dataset(
+            eval_dataset, data_args, training_args, stage, template, tokenizer, processor, is_eval=True
         )
-        column_names = list(next(iter(dataset)).keys())
-        kwargs = {}
-        if not data_args.streaming:
-            kwargs = dict(
-                num_proc=data_args.preprocessing_num_workers,
-                load_from_cache_file=(not data_args.overwrite_cache),
-                desc="Running tokenizer on dataset",
-            )
 
-        dataset = dataset.map(preprocess_func, batched=True, remove_columns=column_names, **kwargs)
+        if data_args.val_size > 1e-6:
+            dataset_dict = split_dataset(dataset, data_args, seed=training_args.seed)
+        else:
+            dataset_dict = {}
+            if dataset is not None:
+                if data_args.streaming:
+                    dataset = dataset.shuffle(buffer_size=data_args.buffer_size, seed=training_args.seed)
+
+                dataset_dict["train"] = dataset
+
+            if eval_dataset is not None:
+                if data_args.streaming:
+                    eval_dataset = eval_dataset.shuffle(buffer_size=data_args.buffer_size, seed=training_args.seed)
+
+                dataset_dict["validation"] = eval_dataset
+
+            dataset_dict = DatasetDict(dataset_dict)
 
         if data_args.tokenized_path is not None:
             if training_args.should_save:
-                dataset.save_to_disk(data_args.tokenized_path)
+                dataset_dict.save_to_disk(data_args.tokenized_path)
                 logger.info("Tokenized dataset saved at {}.".format(data_args.tokenized_path))
                 logger.info("Please restart the training with `tokenized_path: {}`.".format(data_args.tokenized_path))
 
             sys.exit(0)
 
-        if training_args.should_log:
-            try:
-                print_function(next(iter(dataset)))
-            except StopIteration:
-                if stage == "pt":
-                    raise RuntimeError("Cannot find sufficient samples, consider increasing dataset size.")
-                else:
-                    raise RuntimeError("Cannot find valid samples, check `data/README.md` for the data format.")
+        dataset_module = {}
+        if "train" in dataset_dict:
+            dataset_module["train_dataset"] = dataset_dict["train"]
+        if "validation" in dataset_dict:
+            dataset_module["eval_dataset"] = dataset_dict["validation"]
 
-        return dataset
+        return dataset_module

src/llamafactory/data/parser.py

@@ -1,47 +1,60 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import json
 import os
 from dataclasses import dataclass
-from typing import TYPE_CHECKING, Any, Dict, List, Literal, Optional
+from typing import Any, Dict, List, Literal, Optional, Sequence
+
+from transformers.utils import cached_file
 
 from ..extras.constants import DATA_CONFIG
 from ..extras.misc import use_modelscope
 
 
-if TYPE_CHECKING:
-    from ..hparams import DataArguments
-
-
 @dataclass
 class DatasetAttr:
     r"""
     Dataset attributes.
     """
 
-    """ basic configs """
+    # basic configs
     load_from: Literal["hf_hub", "ms_hub", "script", "file"]
     dataset_name: str
     formatting: Literal["alpaca", "sharegpt"] = "alpaca"
     ranking: bool = False
-    """ extra configs """
+    # extra configs
     subset: Optional[str] = None
+    split: str = "train"
     folder: Optional[str] = None
     num_samples: Optional[int] = None
-    """ common columns """
+    # common columns
     system: Optional[str] = None
     tools: Optional[str] = None
     images: Optional[str] = None
-    """ rlhf columns """
+    # rlhf columns
     chosen: Optional[str] = None
     rejected: Optional[str] = None
     kto_tag: Optional[str] = None
-    """ alpaca columns """
+    # alpaca columns
     prompt: Optional[str] = "instruction"
     query: Optional[str] = "input"
     response: Optional[str] = "output"
     history: Optional[str] = None
-    """ sharegpt columns """
+    # sharegpt columns
     messages: Optional[str] = "conversations"
-    """ sharegpt tags """
+    # sharegpt tags
     role_tag: Optional[str] = "from"
     content_tag: Optional[str] = "value"
     user_tag: Optional[str] = "human"
@@ -57,31 +70,33 @@ class DatasetAttr:
         setattr(self, key, obj.get(key, default))
 
 
-def get_dataset_list(data_args: "DataArguments") -> List["DatasetAttr"]:
-    if data_args.dataset is not None:
-        dataset_names = [ds.strip() for ds in data_args.dataset.split(",")]
-    else:
+def get_dataset_list(dataset_names: Optional[Sequence[str]], dataset_dir: str) -> List["DatasetAttr"]:
+    r"""
+    Gets the attributes of the datasets.
+    """
+    if dataset_names is None:
         dataset_names = []
 
-    if data_args.dataset_dir == "ONLINE":
+    if dataset_dir == "ONLINE":
         dataset_info = None
     else:
+        if dataset_dir.startswith("REMOTE:"):
+            config_path = cached_file(path_or_repo_id=dataset_dir[7:], filename=DATA_CONFIG, repo_type="dataset")
+        else:
+            config_path = os.path.join(dataset_dir, DATA_CONFIG)
+
         try:
-            with open(os.path.join(data_args.dataset_dir, DATA_CONFIG), "r") as f:
+            with open(config_path, "r") as f:
                 dataset_info = json.load(f)
         except Exception as err:
             if len(dataset_names) != 0:
-                raise ValueError(
-                    "Cannot open {} due to {}.".format(os.path.join(data_args.dataset_dir, DATA_CONFIG), str(err))
-                )
+                raise ValueError("Cannot open {} due to {}.".format(config_path, str(err)))
+
             dataset_info = None
 
-    if data_args.interleave_probs is not None:
-        data_args.interleave_probs = [float(prob.strip()) for prob in data_args.interleave_probs.split(",")]
-
-    dataset_list: List[DatasetAttr] = []
+    dataset_list: List["DatasetAttr"] = []
     for name in dataset_names:
-        if dataset_info is None:
+        if dataset_info is None:  # dataset_dir is ONLINE
             load_from = "ms_hub" if use_modelscope() else "hf_hub"
             dataset_attr = DatasetAttr(load_from, dataset_name=name)
             dataset_list.append(dataset_attr)
@@ -106,6 +121,7 @@ def get_dataset_list(data_args: "DataArguments") -> List["DatasetAttr"]:
         dataset_attr.set_attr("formatting", dataset_info[name], default="alpaca")
         dataset_attr.set_attr("ranking", dataset_info[name], default=False)
         dataset_attr.set_attr("subset", dataset_info[name])
+        dataset_attr.set_attr("split", dataset_info[name], default="train")
         dataset_attr.set_attr("folder", dataset_info[name])
         dataset_attr.set_attr("num_samples", dataset_info[name])
 

src/llamafactory/data/preprocess.py

@@ -1,3 +1,17 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 from functools import partial
 from typing import TYPE_CHECKING, Callable, Literal, Optional, Tuple
 
@@ -13,8 +27,7 @@ from .processors.unsupervised import preprocess_unsupervised_dataset, print_unsu
 
 
 if TYPE_CHECKING:
-    from transformers import ProcessorMixin, Seq2SeqTrainingArguments
-    from transformers.tokenization_utils import PreTrainedTokenizer
+    from transformers import PreTrainedTokenizer, ProcessorMixin
 
     from ..hparams import DataArguments
     from .template import Template
@@ -22,11 +35,11 @@ if TYPE_CHECKING:
 
 def get_preprocess_and_print_func(
     data_args: "DataArguments",
-    training_args: "Seq2SeqTrainingArguments",
     stage: Literal["pt", "sft", "rm", "ppo", "kto"],
     template: "Template",
     tokenizer: "PreTrainedTokenizer",
     processor: Optional["ProcessorMixin"],
+    do_generate: bool = False,
 ) -> Tuple[Callable, Callable]:
     if stage == "pt":
         preprocess_func = partial(
@@ -35,8 +48,21 @@ def get_preprocess_and_print_func(
             data_args=data_args,
         )
         print_function = partial(print_unsupervised_dataset_example, tokenizer=tokenizer)
-    elif stage == "sft" and not training_args.predict_with_generate:
+    elif stage == "sft" and not do_generate:
         if data_args.packing:
+            if data_args.neat_packing:
+                from datasets.arrow_writer import OptimizedTypedSequence, TypedSequence
+
+                def __init__(self, data, **kwargs):
+                    return TypedSequence.__init__(
+                        self,
+                        data,
+                        type=kwargs.pop("type", None),
+                        try_type=kwargs.pop("try_type", None),
+                        optimized_int_type=kwargs.pop("optimized_int_type", None),
+                    )
+
+                OptimizedTypedSequence.__init__ = __init__
             preprocess_func = partial(
                 preprocess_packed_supervised_dataset,
                 template=template,

src/llamafactory/data/processors/feedback.py

@@ -1,13 +1,26 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 from typing import TYPE_CHECKING, Any, Dict, List, Optional, Sequence, Tuple
 
 from ...extras.constants import IGNORE_INDEX
 from ...extras.logging import get_logger
-from .processor_utils import get_paligemma_token_type_ids, get_pixel_values
+from .processor_utils import get_paligemma_token_type_ids, get_pixel_values, infer_seqlen
 
 
 if TYPE_CHECKING:
-    from transformers import ProcessorMixin
-    from transformers.tokenization_utils import PreTrainedTokenizer
+    from transformers import PreTrainedTokenizer, ProcessorMixin
 
     from ...hparams import DataArguments
     from ..template import Template
@@ -42,12 +55,8 @@ def _encode_feedback_example(
     else:
         kl_messages = prompt + [kl_response[1]]
 
-    prompt_ids, response_ids = template.encode_oneturn(
-        tokenizer, messages, system, tools, data_args.cutoff_len, data_args.reserved_label_len
-    )
-    _, kl_response_ids = template.encode_oneturn(
-        tokenizer, kl_messages, system, tools, data_args.cutoff_len, data_args.reserved_label_len
-    )
+    prompt_ids, response_ids = template.encode_oneturn(tokenizer, messages, system, tools)
+    kl_prompt_ids, kl_response_ids = template.encode_oneturn(tokenizer, kl_messages, system, tools)
 
     if template.efficient_eos:
         response_ids += [tokenizer.eos_token_id]
@@ -56,11 +65,19 @@ def _encode_feedback_example(
     if processor is not None and hasattr(processor, "image_seq_length"):  # paligemma models
         image_token_id = tokenizer.convert_tokens_to_ids(template.image_token)
         prompt_ids = [image_token_id] * getattr(processor, "image_seq_length") + prompt_ids
+        kl_prompt_ids = [image_token_id] * getattr(processor, "image_seq_length") + kl_prompt_ids
+
+    source_len, target_len = infer_seqlen(len(prompt_ids), len(response_ids), data_args.cutoff_len)
+    prompt_ids = prompt_ids[:source_len]
+    response_ids = response_ids[:target_len]
+    kl_source_len, kl_target_len = infer_seqlen(len(kl_prompt_ids), len(kl_response_ids), data_args.cutoff_len)
+    kl_prompt_ids = kl_prompt_ids[:kl_source_len]
+    kl_response_ids = kl_response_ids[:kl_target_len]
 
     input_ids = prompt_ids + response_ids
-    labels = [IGNORE_INDEX] * len(prompt_ids) + response_ids
-    kl_input_ids = prompt_ids + kl_response_ids
-    kl_labels = [IGNORE_INDEX] * len(prompt_ids) + kl_response_ids
+    labels = [IGNORE_INDEX] * source_len + response_ids
+    kl_input_ids = kl_prompt_ids + kl_response_ids
+    kl_labels = [IGNORE_INDEX] * kl_source_len + kl_response_ids
 
     return input_ids, labels, kl_input_ids, kl_labels, kto_tag
 

src/llamafactory/data/processors/pairwise.py

@@ -1,13 +1,26 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 from typing import TYPE_CHECKING, Any, Dict, List, Optional, Sequence, Tuple
 
 from ...extras.constants import IGNORE_INDEX
 from ...extras.logging import get_logger
-from .processor_utils import get_paligemma_token_type_ids, get_pixel_values
+from .processor_utils import get_paligemma_token_type_ids, get_pixel_values, infer_seqlen
 
 
 if TYPE_CHECKING:
-    from transformers import ProcessorMixin
-    from transformers.tokenization_utils import PreTrainedTokenizer
+    from transformers import PreTrainedTokenizer, ProcessorMixin
 
     from ...hparams import DataArguments
     from ..template import Template
@@ -31,12 +44,8 @@ def _encode_pairwise_example(
 
     chosen_messages = prompt + [response[0]]
     rejected_messages = prompt + [response[1]]
-    prompt_ids, chosen_ids = template.encode_oneturn(
-        tokenizer, chosen_messages, system, tools, data_args.cutoff_len, data_args.reserved_label_len
-    )
-    _, rejected_ids = template.encode_oneturn(
-        tokenizer, rejected_messages, system, tools, data_args.cutoff_len, data_args.reserved_label_len
-    )
+    prompt_ids, chosen_ids = template.encode_oneturn(tokenizer, chosen_messages, system, tools)
+    _, rejected_ids = template.encode_oneturn(tokenizer, rejected_messages, system, tools)
 
     if template.efficient_eos:
         chosen_ids += [tokenizer.eos_token_id]
@@ -46,10 +55,17 @@ def _encode_pairwise_example(
         image_token_id = tokenizer.convert_tokens_to_ids(template.image_token)
         prompt_ids = [image_token_id] * getattr(processor, "image_seq_length") + prompt_ids
 
+    source_len, target_len = infer_seqlen(
+        len(prompt_ids), max(len(chosen_ids), len(rejected_ids)), data_args.cutoff_len
+    )  # consider the response is more important
+    prompt_ids = prompt_ids[:source_len]
+    chosen_ids = chosen_ids[:target_len]
+    rejected_ids = rejected_ids[:target_len]
+
     chosen_input_ids = prompt_ids + chosen_ids
-    chosen_labels = [IGNORE_INDEX] * len(prompt_ids) + chosen_ids
+    chosen_labels = [IGNORE_INDEX] * source_len + chosen_ids
     rejected_input_ids = prompt_ids + rejected_ids
-    rejected_labels = [IGNORE_INDEX] * len(prompt_ids) + rejected_ids
+    rejected_labels = [IGNORE_INDEX] * source_len + rejected_ids
 
     return chosen_input_ids, chosen_labels, rejected_input_ids, rejected_labels
 

src/llamafactory/data/processors/pretrain.py

@@ -1,9 +1,26 @@
+# Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
+#
+# This code is inspired by the HuggingFace's transformers library.
+# https://github.com/huggingface/transformers/blob/v4.40.0/examples/pytorch/language-modeling/run_clm.py
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 from itertools import chain
 from typing import TYPE_CHECKING, Any, Dict, List
 
 
 if TYPE_CHECKING:
-    from transformers.tokenization_utils import PreTrainedTokenizer
+    from transformers import PreTrainedTokenizer
 
     from ...hparams import DataArguments
 
@@ -12,7 +29,8 @@ 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 ...` if packing is enabled
-    text_examples = [messages[0]["content"] + tokenizer.eos_token for messages in examples["prompt"]]
+    eos_token = "<|end_of_text|>" if data_args.template == "llama3" else tokenizer.eos_token
+    text_examples = [messages[0]["content"] + eos_token for messages in examples["prompt"]]
 
     if not data_args.packing:
         if data_args.template == "gemma":

src/llamafactory/data/processors/processor_utils.py

@@ -1,5 +1,19 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import bisect
-from typing import TYPE_CHECKING, List, Sequence
+from typing import TYPE_CHECKING, List, Sequence, Tuple
 
 from ...extras.packages import is_pillow_available
 
@@ -62,3 +76,20 @@ def get_paligemma_token_type_ids(input_len: int, processor: "ProcessorMixin") ->
     """
     image_seq_length = getattr(processor, "image_seq_length")
     return [0] * image_seq_length + [1] * (input_len - image_seq_length)
+
+
+def infer_seqlen(source_len: int, target_len: int, cutoff_len: int) -> Tuple[int, int]:
+    r"""
+    Computes the real sequence length after truncation by the cutoff_len.
+    """
+    if target_len * 2 < cutoff_len:  # truncate source
+        max_target_len = cutoff_len
+    elif source_len * 2 < cutoff_len:  # truncate target
+        max_target_len = cutoff_len - source_len
+    else:  # truncate both
+        max_target_len = int(cutoff_len * (target_len / (source_len + target_len)))
+
+    new_target_len = min(max_target_len, target_len)
+    max_source_len = max(cutoff_len - new_target_len, 0)
+    new_source_len = min(max_source_len, source_len)
+    return new_source_len, new_target_len

src/llamafactory/data/processors/supervised.py

@@ -1,14 +1,27 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 from collections import defaultdict
 from typing import TYPE_CHECKING, Any, Dict, List, Optional, Sequence, Tuple
 
 from ...extras.constants import IGNORE_INDEX
 from ...extras.logging import get_logger
-from .processor_utils import get_paligemma_token_type_ids, get_pixel_values, greedy_knapsack
+from .processor_utils import get_paligemma_token_type_ids, get_pixel_values, greedy_knapsack, infer_seqlen
 
 
 if TYPE_CHECKING:
-    from transformers import ProcessorMixin
-    from transformers.tokenization_utils import PreTrainedTokenizer
+    from transformers import PreTrainedTokenizer, ProcessorMixin
 
     from ...hparams import DataArguments
     from ..template import Template
@@ -38,19 +51,31 @@ def _encode_supervised_example(
         input_ids += [image_token_id] * getattr(processor, "image_seq_length")
         labels += [IGNORE_INDEX] * getattr(processor, "image_seq_length")
 
-    encoded_pairs = template.encode_multiturn(
-        tokenizer, messages, system, tools, data_args.cutoff_len, data_args.reserved_label_len
-    )
+    encoded_pairs = template.encode_multiturn(tokenizer, messages, system, tools)
+    total_length = 1 if template.efficient_eos else 0
     for turn_idx, (source_ids, target_ids) in enumerate(encoded_pairs):
+        if total_length >= data_args.cutoff_len:
+            break
+
+        source_len, target_len = infer_seqlen(len(source_ids), len(target_ids), data_args.cutoff_len - total_length)
+        source_ids = source_ids[:source_len]
+        target_ids = target_ids[:target_len]
+        total_length += source_len + target_len
+
         if data_args.train_on_prompt:
-            source_mask = source_ids
+            source_label = source_ids
         elif turn_idx != 0 and template.efficient_eos:
-            source_mask = [tokenizer.eos_token_id] + [IGNORE_INDEX] * (len(source_ids) - 1)
+            source_label = [tokenizer.eos_token_id] + [IGNORE_INDEX] * (source_len - 1)
         else:
-            source_mask = [IGNORE_INDEX] * len(source_ids)
+            source_label = [IGNORE_INDEX] * source_len
+
+        if data_args.mask_history and turn_idx != len(encoded_pairs) - 1:
+            target_label = [IGNORE_INDEX] * target_len
+        else:
+            target_label = target_ids
 
         input_ids += source_ids + target_ids
-        labels += source_mask + target_ids
+        labels += source_label + target_label
 
     if template.efficient_eos:
         input_ids += [tokenizer.eos_token_id]
@@ -140,22 +165,30 @@ def preprocess_packed_supervised_dataset(
     model_inputs = {"input_ids": [], "attention_mask": [], "labels": []}
     knapsacks = greedy_knapsack(lengths, data_args.cutoff_len)
     for knapsack in knapsacks:
-        packed_input_ids, packed_labels = [], []
-        for length in knapsack:
+        packed_input_ids, packed_attention_masks, packed_labels = [], [], []
+        for i, length in enumerate(knapsack):
             index = length2indexes[length].pop()
             packed_input_ids += batch_input_ids[index]
             packed_labels += batch_labels[index]
+            if data_args.neat_packing:
+                packed_attention_masks += [i + 1] * len(batch_input_ids[index])  # start from 1
+            else:
+                packed_attention_masks += [1] * len(batch_input_ids[index])
 
         if len(packed_input_ids) < data_args.cutoff_len:
             pad_length = data_args.cutoff_len - len(packed_input_ids)
             packed_input_ids += [tokenizer.pad_token_id] * pad_length
             packed_labels += [IGNORE_INDEX] * pad_length
+            if data_args.neat_packing:
+                packed_attention_masks += [0] * pad_length
+            else:
+                packed_attention_masks += [1] * pad_length  # more efficient flash_attn
 
         if len(packed_input_ids) != data_args.cutoff_len:
             raise ValueError("The length of packed example should be identical to the cutoff length.")
 
         model_inputs["input_ids"].append(packed_input_ids)
-        model_inputs["attention_mask"].append([1] * data_args.cutoff_len)
+        model_inputs["attention_mask"].append(packed_attention_masks)
         model_inputs["labels"].append(packed_labels)
 
     return model_inputs

src/llamafactory/data/processors/unsupervised.py

@@ -1,13 +1,26 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 from typing import TYPE_CHECKING, Any, Dict, List, Optional, Sequence, Tuple
 
 from ...extras.logging import get_logger
 from ..data_utils import Role
-from .processor_utils import get_paligemma_token_type_ids, get_pixel_values
+from .processor_utils import get_paligemma_token_type_ids, get_pixel_values, infer_seqlen
 
 
 if TYPE_CHECKING:
-    from transformers import ProcessorMixin
-    from transformers.tokenization_utils import PreTrainedTokenizer
+    from transformers import PreTrainedTokenizer, ProcessorMixin
 
     from ...hparams import DataArguments
     from ..template import Template
@@ -34,9 +47,7 @@ def _encode_unsupervised_example(
     else:
         messages = prompt + [{"role": Role.ASSISTANT.value, "content": ""}]
 
-    input_ids, labels = template.encode_oneturn(
-        tokenizer, messages, system, tools, data_args.cutoff_len, data_args.reserved_label_len
-    )
+    input_ids, labels = template.encode_oneturn(tokenizer, messages, system, tools)
     if template.efficient_eos:
         labels += [tokenizer.eos_token_id]
 
@@ -44,6 +55,9 @@ def _encode_unsupervised_example(
         image_token_id = tokenizer.convert_tokens_to_ids(template.image_token)
         input_ids = [image_token_id] * getattr(processor, "image_seq_length") + input_ids
 
+    source_len, target_len = infer_seqlen(len(input_ids), len(labels), data_args.cutoff_len)
+    input_ids = input_ids[:source_len]
+    labels = labels[:target_len]
     return input_ids, labels
 
 

src/llamafactory/data/template.py

@@ -1,8 +1,22 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 from dataclasses import dataclass
 from typing import TYPE_CHECKING, Dict, List, Optional, Sequence, Tuple, Union
 
 from ..extras.logging import get_logger
-from .data_utils import Role, infer_max_len
+from .data_utils import Role
 from .formatter import EmptyFormatter, FunctionFormatter, StringFormatter, ToolFormatter
 
 
@@ -24,69 +38,74 @@ class Template:
     format_observation: "Formatter"
     format_tools: "Formatter"
     format_separator: "Formatter"
+    format_prefix: "Formatter"
     default_system: str
     stop_words: List[str]
     image_token: str
     efficient_eos: bool
     replace_eos: bool
-    force_system: bool
 
     def encode_oneturn(
         self,
         tokenizer: "PreTrainedTokenizer",
-        messages: List[Dict[str, str]],
+        messages: Sequence[Dict[str, str]],
         system: Optional[str] = None,
         tools: Optional[str] = None,
-        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.
         """
-        encoded_pairs = self._encode(tokenizer, messages, system, tools, cutoff_len, reserved_label_len)
+        encoded_messages = self._encode(tokenizer, messages, system, tools)
         prompt_ids = []
-        for query_ids, resp_ids in encoded_pairs[:-1]:
-            prompt_ids += query_ids + resp_ids
-        prompt_ids = prompt_ids + encoded_pairs[-1][0]
-        answer_ids = encoded_pairs[-1][1]
+        for encoded_ids in encoded_messages[:-1]:
+            prompt_ids += encoded_ids
+
+        answer_ids = encoded_messages[-1]
         return prompt_ids, answer_ids
 
     def encode_multiturn(
         self,
         tokenizer: "PreTrainedTokenizer",
-        messages: List[Dict[str, str]],
+        messages: Sequence[Dict[str, str]],
         system: Optional[str] = None,
         tools: Optional[str] = None,
-        cutoff_len: int = 1_000_000,
-        reserved_label_len: int = 1,
-    ) -> Sequence[Tuple[List[int], List[int]]]:
+    ) -> List[Tuple[List[int], List[int]]]:
         r"""
         Returns multiple pairs of token ids representing prompts and responses respectively.
         """
-        return self._encode(tokenizer, messages, system, tools, cutoff_len, reserved_label_len)
+        encoded_messages = self._encode(tokenizer, messages, system, tools)
+        return [(encoded_messages[i], encoded_messages[i + 1]) for i in range(0, len(encoded_messages), 2)]
+
+    def extract_tool(self, content: str) -> Union[str, List[Tuple[str, str]]]:
+        r"""
+        Extracts tool message.
+        """
+        return self.format_tools.extract(content)
 
     def _encode(
         self,
         tokenizer: "PreTrainedTokenizer",
-        messages: List[Dict[str, str]],
+        messages: Sequence[Dict[str, str]],
         system: Optional[str],
         tools: Optional[str],
-        cutoff_len: int,
-        reserved_label_len: int,
-    ) -> Sequence[Tuple[List[int], List[int]]]:
+    ) -> List[List[int]]:
         r"""
         Encodes formatted inputs to pairs of token ids.
-        Turn 0: system + query        resp
-        Turn t: sep + query           resp
+        Turn 0: prefix + system + query        resp
+        Turn t: sep + query                    resp
         """
         system = system or self.default_system
         encoded_messages = []
         for i, message in enumerate(messages):
             elements = []
-            if i == 0 and (system or tools or self.force_system):
-                tool_text = self.format_tools.apply(content=tools)[0] if tools else ""
-                elements += self.format_system.apply(content=(system + tool_text))
-            elif i > 0 and i % 2 == 0:
+
+            if i == 0:
+                elements += self.format_prefix.apply()
+                if system or tools:
+                    tool_text = self.format_tools.apply(content=tools)[0] if tools else ""
+                    elements += self.format_system.apply(content=(system + tool_text))
+
+            if i > 0 and i % 2 == 0:
                 elements += self.format_separator.apply()
 
             if message["role"] == Role.USER.value:
@@ -102,11 +121,9 @@ class Template:
 
             encoded_messages.append(self._convert_elements_to_ids(tokenizer, elements))
 
-        return self._make_pairs(encoded_messages, cutoff_len, reserved_label_len)
+        return encoded_messages
 
-    def _convert_elements_to_ids(
-        self, tokenizer: "PreTrainedTokenizer", elements: List[Union[str, Dict[str, str]]]
-    ) -> List[int]:
+    def _convert_elements_to_ids(self, tokenizer: "PreTrainedTokenizer", elements: "SLOTS") -> List[int]:
         r"""
         Converts elements to token ids.
         """
@@ -127,57 +144,34 @@ class Template:
 
         return token_ids
 
-    def _make_pairs(
-        self,
-        encoded_messages: Sequence[List[int]],
-        cutoff_len: int,
-        reserved_label_len: int,
-    ) -> Sequence[Tuple[List[int], List[int]]]:
-        encoded_pairs = []
-        total_length = 0
-        for i in range(0, len(encoded_messages), 2):
-            if total_length >= cutoff_len:
-                break
-
-            max_source_len, max_target_len = infer_max_len(
-                source_len=len(encoded_messages[i]),
-                target_len=len(encoded_messages[i + 1]),
-                max_len=(cutoff_len - total_length),
-                reserved_label_len=reserved_label_len,
-            )
-            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
-
 
 @dataclass
 class Llama2Template(Template):
     def _encode(
         self,
         tokenizer: "PreTrainedTokenizer",
-        messages: List[Dict[str, str]],
+        messages: Sequence[Dict[str, str]],
         system: str,
         tools: str,
-        cutoff_len: int,
-        reserved_label_len: int,
-    ) -> Sequence[Tuple[List[int], List[int]]]:
+    ) -> List[List[int]]:
         r"""
         Encodes formatted inputs to pairs of token ids.
-        Turn 0: system + query        resp
-        Turn t: sep + query           resp
+        Turn 0: prefix + system + query        resp
+        Turn t: sep + query                    resp
         """
         system = system or self.default_system
         encoded_messages = []
         for i, message in enumerate(messages):
             elements = []
+
             system_text = ""
-            if i == 0 and (system or tools or self.force_system):
-                tool_text = self.format_tools.apply(content=tools)[0] if tools else ""
-                system_text = self.format_system.apply(content=(system + tool_text))[0]
-            elif i > 0 and i % 2 == 0:
+            if i == 0:
+                elements += self.format_prefix.apply()
+                if system or tools:
+                    tool_text = self.format_tools.apply(content=tools)[0] if tools else ""
+                    system_text = self.format_system.apply(content=(system + tool_text))[0]
+
+            if i > 0 and i % 2 == 0:
                 elements += self.format_separator.apply()
 
             if message["role"] == Role.USER.value:
@@ -193,7 +187,7 @@ class Llama2Template(Template):
 
             encoded_messages.append(self._convert_elements_to_ids(tokenizer, elements))
 
-        return self._make_pairs(encoded_messages, cutoff_len, reserved_label_len)
+        return encoded_messages
 
 
 TEMPLATES: Dict[str, Template] = {}
@@ -208,12 +202,12 @@ def _register_template(
     format_observation: Optional["Formatter"] = None,
     format_tools: Optional["Formatter"] = None,
     format_separator: Optional["Formatter"] = None,
+    format_prefix: Optional["Formatter"] = None,
     default_system: str = "",
-    stop_words: List[str] = [],
+    stop_words: Sequence[str] = [],
     image_token: str = "<image>",
     efficient_eos: bool = False,
     replace_eos: bool = False,
-    force_system: bool = False,
 ) -> None:
     r"""
     Registers a chat template.
@@ -245,9 +239,10 @@ def _register_template(
     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_function_formatter = FunctionFormatter(slots=eos_slots, tool_format="default")
     default_tool_formatter = ToolFormatter(tool_format="default")
     default_separator_formatter = EmptyFormatter()
+    default_prefix_formatter = EmptyFormatter()
     TEMPLATES[name] = template_class(
         format_user=format_user or default_user_formatter,
         format_assistant=format_assistant or default_assistant_formatter,
@@ -256,12 +251,12 @@ def _register_template(
         format_observation=format_observation or format_user or default_user_formatter,
         format_tools=format_tools or default_tool_formatter,
         format_separator=format_separator or default_separator_formatter,
+        format_prefix=format_prefix or default_prefix_formatter,
         default_system=default_system,
         stop_words=stop_words,
         image_token=image_token,
         efficient_eos=efficient_eos,
         replace_eos=replace_eos,
-        force_system=force_system,
     )
 
 
@@ -307,6 +302,10 @@ def _convert_slots_to_jinja(slots: "SLOTS", tokenizer: "PreTrainedTokenizer", pl
 def _get_jinja_template(template: "Template", tokenizer: "PreTrainedTokenizer") -> str:
     jinja_template = ""
 
+    prefix = _convert_slots_to_jinja(template.format_prefix.apply(), tokenizer)
+    if prefix:
+        jinja_template += "{{ " + prefix + " }}"
+
     if template.default_system:
         jinja_template += "{% set system_message = '" + _jinja_escape(template.default_system) + "' %}"
 
@@ -315,11 +314,7 @@ def _get_jinja_template(template: "Template", tokenizer: "PreTrainedTokenizer")
     )
 
     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:
+    if not isinstance(template, Llama2Template):
         jinja_template += "{% if system_message is defined %}{{ " + system_message + " }}{% endif %}"
 
     jinja_template += "{% for message in messages %}"
@@ -346,6 +341,7 @@ def _get_jinja_template(template: "Template", tokenizer: "PreTrainedTokenizer")
 def get_template_and_fix_tokenizer(
     tokenizer: "PreTrainedTokenizer",
     name: Optional[str] = None,
+    tool_format: Optional[str] = None,
 ) -> Template:
     if name is None:
         template = TEMPLATES["empty"]  # placeholder
@@ -354,6 +350,12 @@ def get_template_and_fix_tokenizer(
         if template is None:
             raise ValueError("Template {} does not exist.".format(name))
 
+    if tool_format is not None:
+        logger.info("Using tool format: {}.".format(tool_format))
+        eos_slots = [] if template.efficient_eos else [{"eos_token"}]
+        template.format_tools = ToolFormatter(tool_format=tool_format)
+        template.format_function = FunctionFormatter(slots=eos_slots, tool_format=tool_format)
+
     stop_words = template.stop_words
     if template.replace_eos:
         if not stop_words:
@@ -435,9 +437,8 @@ _register_template(
 _register_template(
     name="belle",
     format_user=StringFormatter(slots=["Human: {{content}}\n\nBelle: "]),
-    format_system=StringFormatter(slots=[{"bos_token"}, "{{content}}"]),
     format_separator=EmptyFormatter(slots=["\n\n"]),
-    force_system=True,
+    format_prefix=EmptyFormatter(slots=[{"bos_token"}]),
 )
 
 
@@ -450,11 +451,7 @@ _register_template(
 _register_template(
     name="breeze",
     format_user=StringFormatter(slots=["[INST] {{content}} [/INST] "]),
-    format_system=StringFormatter(slots=[{"bos_token"}, "{{content}}"]),
-    default_system=(
-        "You are a helpful AI assistant built by MediaTek Research. "
-        "The user you are helping speaks Traditional Chinese and comes from Taiwan."
-    ),
+    format_prefix=EmptyFormatter(slots=[{"bos_token"}]),
     efficient_eos=True,
 )
 
@@ -462,10 +459,9 @@ _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}}"]),
     format_separator=EmptyFormatter(slots=["\n\n"]),
+    format_prefix=EmptyFormatter(slots=[{"token": "[gMASK]"}, {"token": "sop"}]),
     efficient_eos=True,
-    force_system=True,
 )
 
 
@@ -473,32 +469,13 @@ _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_system=StringFormatter(slots=[{"token": "<|system|>"}, "\n", "{{content}}"]),
+    format_function=FunctionFormatter(slots=[], tool_format="glm4"),
     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}}", {"token": "<|assistant|>"}]
-    ),
-    default_system=(
-        "You are ChatGLM3, a large language model trained by Zhipu.AI. "
-        "Follow the user's instructions carefully. Respond using markdown."
-    ),
+    format_tools=ToolFormatter(tool_format="glm4"),
+    format_prefix=EmptyFormatter(slots=[{"token": "[gMASK]"}, {"token": "sop"}]),
     stop_words=["<|user|>", "<|observation|>"],
     efficient_eos=True,
 )
@@ -529,8 +506,24 @@ _register_template(
 
 _register_template(
     name="codegeex2",
-    format_system=StringFormatter(slots=[{"token": "[gMASK]"}, {"token": "sop"}, "{{content}}"]),
-    force_system=True,
+    format_prefix=EmptyFormatter(slots=[{"token": "[gMASK]"}, {"token": "sop"}]),
+)
+
+
+_register_template(
+    name="codegeex4",
+    format_user=StringFormatter(slots=["<|user|>\n{{content}}<|assistant|>\n"]),
+    format_system=StringFormatter(slots=["<|system|>\n{{content}}"]),
+    format_function=FunctionFormatter(slots=[], tool_format="glm4"),
+    format_observation=StringFormatter(slots=["<|observation|>\n{{content}}<|assistant|>\n"]),
+    format_tools=ToolFormatter(tool_format="glm4"),
+    format_prefix=EmptyFormatter(slots=["[gMASK]<sop>"]),
+    default_system=(
+        "你是一位智能编程助手，你叫CodeGeeX。你会为用户回答关于编程、代码、计算机方面的任何问题，"
+        "并提供格式规范、可以执行、准确安全的代码，并在必要时提供详细的解释。"
+    ),
+    stop_words=["<|user|>", "<|observation|>"],
+    efficient_eos=True,
 )
 
 
@@ -544,21 +537,15 @@ _register_template(
             )
         ]
     ),
-    format_system=StringFormatter(
-        slots=[{"bos_token"}, "<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>{{content}}<|END_OF_TURN_TOKEN|>"]
-    ),
-    default_system=(
-        "You are Command-R, a brilliant, sophisticated, AI-assistant trained to assist human users "
-        "by providing thorough responses. You are trained by Cohere."
-    ),
+    format_system=StringFormatter(slots=["<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>{{content}}<|END_OF_TURN_TOKEN|>"]),
+    format_prefix=EmptyFormatter(slots=[{"bos_token"}]),
 )
 
 
 _register_template(
     name="cpm",
     format_user=StringFormatter(slots=["<用户>{{content}}<AI>"]),
-    format_system=StringFormatter(slots=[{"bos_token"}, "{{content}}"]),
-    force_system=True,
+    format_prefix=EmptyFormatter(slots=[{"bos_token"}]),
 )
 
 
@@ -591,30 +578,28 @@ _register_template(
 _register_template(
     name="deepseek",
     format_user=StringFormatter(slots=["User: {{content}}\n\nAssistant:"]),
-    format_system=StringFormatter(slots=[{"bos_token"}, "{{content}}"]),
-    force_system=True,
+    format_prefix=EmptyFormatter(slots=[{"bos_token"}]),
 )
 
 
 _register_template(
     name="deepseekcoder",
     format_user=StringFormatter(slots=["### Instruction:\n{{content}}\n### Response:"]),
-    format_assistant=StringFormatter(slots=["\n", "{{content}}"]),
-    format_separator=EmptyFormatter(slots=["\n<|EOT|>\n"]),
+    format_assistant=StringFormatter(slots=["\n{{content}}\n"]),
+    format_separator=EmptyFormatter(slots=["\n"]),
+    format_prefix=EmptyFormatter(slots=[{"bos_token"}]),
     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. "
         "For politically sensitive questions, security and privacy issues, "
         "and other non-computer science questions, you will refuse to answer\n"
     ),
-    stop_words=["<|EOT|>"],
-    efficient_eos=True,
 )
 
 
 _register_template(
     name="default",
-    format_user=StringFormatter(slots=["Human: {{content}}\nAssistant: "]),
+    format_user=StringFormatter(slots=["Human: {{content}}\nAssistant:"]),
     format_system=StringFormatter(slots=["{{content}}\n"]),
     format_separator=EmptyFormatter(slots=["\n"]),
 )
@@ -622,11 +607,7 @@ _register_template(
 
 _register_template(
     name="empty",
-    format_user=StringFormatter(slots=["{{content}}"]),
-    format_assistant=StringFormatter(slots=["{{content}}"]),
-    format_system=StringFormatter(slots=[{"bos_token"}, "{{content}}"]),
     efficient_eos=True,
-    force_system=True,
 )
 
 
@@ -648,13 +629,12 @@ _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_observation=StringFormatter(
         slots=["<start_of_turn>tool\n{{content}}<end_of_turn>\n<start_of_turn>model\n"]
     ),
     format_separator=EmptyFormatter(slots=["<end_of_turn>\n"]),
+    format_prefix=EmptyFormatter(slots=[{"bos_token"}]),
     efficient_eos=True,
-    force_system=True,
 )
 
 
@@ -662,36 +642,33 @@ _register_template(
     name="glm4",
     format_user=StringFormatter(slots=["<|user|>\n{{content}}<|assistant|>"]),
     format_assistant=StringFormatter(slots=["\n{{content}}"]),
-    format_system=StringFormatter(slots=["[gMASK]<sop>{{content}}"]),
-    format_function=FunctionFormatter(slots=["{{name}}\n{{arguments}}"]),
+    format_system=StringFormatter(slots=["<|system|>\n{{content}}"]),
+    format_function=FunctionFormatter(slots=[], tool_format="glm4"),
     format_observation=StringFormatter(slots=["<|observation|>\n{{content}}<|assistant|>"]),
+    format_tools=ToolFormatter(tool_format="glm4"),
+    format_prefix=EmptyFormatter(slots=["[gMASK]<sop>"]),
     stop_words=["<|user|>", "<|observation|>"],
     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"]),
+    format_user=StringFormatter(slots=["<|User|>:{{content}}\n<|Bot|>:"]),
+    format_system=StringFormatter(slots=["<|System|>:{{content}}\n"]),
+    format_separator=EmptyFormatter(slots=["<eoa>\n"]),
+    format_prefix=EmptyFormatter(slots=[{"bos_token"}]),
     stop_words=["<eoa>"],
-    efficient_eos=True,
+    efficient_eos=True,  # internlm tokenizer cannot set eos_token_id
 )
 
 
 _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"]),
-    format_separator=EmptyFormatter(slots=["\n"]),
-    default_system=(
-        "You are an AI assistant whose name is InternLM (书生·浦语).\n"
-        "- InternLM (书生·浦语) is a conversational language model that is developed "
-        "by Shanghai AI Laboratory (上海人工智能实验室). It is designed to be helpful, honest, and harmless.\n"
-        "- InternLM (书生·浦语) can understand and communicate fluently in the language chosen "
-        "by the user such as English and 中文."
-    ),
+    format_system=StringFormatter(slots=["<|im_start|>system\n{{content}}<|im_end|>\n"]),
+    format_separator=EmptyFormatter(slots=["<|im_end|>\n"]),
+    format_prefix=EmptyFormatter(slots=[{"bos_token"}]),
     stop_words=["<|im_end|>"],
     efficient_eos=True,  # internlm2 tokenizer cannot set eos_token_id
 )
@@ -700,7 +677,6 @@ _register_template(
 _register_template(
     name="llama2",
     format_user=StringFormatter(slots=[{"bos_token"}, "[INST] {{content}} [/INST]"]),
-    format_assistant=StringFormatter(slots=[" {{content}} ", {"eos_token"}]),
     format_system=StringFormatter(slots=["<<SYS>>\n{{content}}\n<</SYS>>\n\n"]),
 )
 
@@ -723,9 +699,7 @@ _register_template(
             )
         ]
     ),
-    format_system=StringFormatter(
-        slots=[{"bos_token"}, "<|start_header_id|>system<|end_header_id|>\n\n{{content}}<|eot_id|>"]
-    ),
+    format_system=StringFormatter(slots=["<|start_header_id|>system<|end_header_id|>\n\n{{content}}<|eot_id|>"]),
     format_observation=StringFormatter(
         slots=[
             (
@@ -734,7 +708,7 @@ _register_template(
             )
         ]
     ),
-    default_system="You are a helpful assistant.",
+    format_prefix=EmptyFormatter(slots=[{"bos_token"}]),
     stop_words=["<|eot_id|>"],
     replace_eos=True,
 )
@@ -743,24 +717,21 @@ _register_template(
 _register_template(
     name="mistral",
     format_user=StringFormatter(slots=["[INST] {{content}} [/INST]"]),
-    format_system=StringFormatter(slots=[{"bos_token"}, "{{content}}"]),
-    force_system=True,
+    format_prefix=EmptyFormatter(slots=[{"bos_token"}]),
 )
 
 
 _register_template(
     name="olmo",
     format_user=StringFormatter(slots=["<|user|>\n{{content}}<|assistant|>\n"]),
-    format_system=StringFormatter(slots=[{"eos_token"}, "{{content}}"]),
-    force_system=True,
+    format_prefix=EmptyFormatter(slots=[{"eos_token"}]),
 )
 
 
 _register_template(
     name="openchat",
     format_user=StringFormatter(slots=["GPT4 Correct User: {{content}}", {"eos_token"}, "GPT4 Correct Assistant:"]),
-    format_system=StringFormatter(slots=[{"bos_token"}, "{{content}}"]),
-    force_system=True,
+    format_prefix=EmptyFormatter(slots=[{"bos_token"}]),
 )
 
 
@@ -774,27 +745,25 @@ _register_template(
             )
         ]
     ),
-    format_system=StringFormatter(slots=[{"bos_token"}, "{{content}}"]),
+    format_prefix=EmptyFormatter(slots=[{"bos_token"}]),
     stop_words=["<|eot_id|>"],
     replace_eos=True,
-    force_system=True,
 )
 
 
 _register_template(
     name="orion",
     format_user=StringFormatter(slots=["Human: {{content}}\n\nAssistant: ", {"eos_token"}]),
-    format_system=StringFormatter(slots=[{"bos_token"}, "{{content}}"]),
-    force_system=True,
+    format_prefix=EmptyFormatter(slots=[{"bos_token"}]),
 )
 
 
 _register_template(
     name="phi",
     format_user=StringFormatter(slots=["<|user|>\n{{content}}<|end|>\n<|assistant|>\n"]),
-    format_system=StringFormatter(slots=[{"bos_token"}, "<|system|>\n{{content}}<|end|>\n"]),
+    format_system=StringFormatter(slots=["<|system|>\n{{content}}<|end|>\n"]),
     format_separator=EmptyFormatter(slots=["\n"]),
-    default_system="You are a helpful AI assistant.",
+    format_prefix=EmptyFormatter(slots=[{"bos_token"}]),
     stop_words=["<|end|>"],
     replace_eos=True,
 )
@@ -827,7 +796,6 @@ _register_template(
     format_separator=EmptyFormatter(slots=["\n"]),
     stop_words=["<|end|>"],
     replace_eos=True,
-    force_system=True,
 )
 
 
@@ -924,8 +892,7 @@ _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_user=StringFormatter(slots=["<|user|>\n{{content}}", {"eos_token"}, "<|assistant|>\n"]),
     format_system=StringFormatter(slots=["<|system|>\n{{content}}", {"eos_token"}]),
     default_system="You are Zephyr, a helpful assistant.",
 )

src/llamafactory/data/tool_utils.py

@@ -0,0 +1,140 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import json
+import re
+from abc import ABC, abstractmethod
+from dataclasses import dataclass
+from typing import Any, Dict, List, Tuple, Union
+
+from .data_utils import SLOTS
+
+
+DEFAULT_TOOL_PROMPT = (
+    "You have access to the following tools:\n{tool_text}"
+    "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, in a JSON format representing the kwargs "
+    """(e.g. ```{{"input": "hello world", "num_beams": 5}}```)\n"""
+    "```\n"
+)
+
+
+GLM4_TOOL_PROMPT = (
+    "你是一个名为 ChatGLM 的人工智能助手。你是基于智谱AI训练的语言模型 GLM-4 模型开发的，"
+    "你的任务是针对用户的问题和要求提供适当的答复和支持。# 可用工具{tool_text}"
+)
+
+
+@dataclass
+class ToolUtils(ABC):
+    @staticmethod
+    @abstractmethod
+    def get_function_slots() -> SLOTS: ...
+
+    @staticmethod
+    @abstractmethod
+    def tool_formatter(tools: List[Dict[str, Any]]) -> str: ...
+
+    @staticmethod
+    @abstractmethod
+    def tool_extractor(content: str) -> Union[str, List[Tuple[str, str]]]: ...
+
+
+class DefaultToolUtils(ToolUtils):
+    @staticmethod
+    def get_function_slots() -> SLOTS:
+        return ["Action: {{name}}\nAction Input: {{arguments}}\n"]
+
+    @staticmethod
+    def tool_formatter(tools: List[Dict[str, Any]]) -> str:
+        tool_text = ""
+        tool_names = []
+        for tool in tools:
+            param_text = ""
+            for name, param in tool["parameters"]["properties"].items():
+                required, enum, items = "", "", ""
+                if name in tool["parameters"].get("required", []):
+                    required = ", required"
+
+                if param.get("enum", None):
+                    enum = ", should be one of [{}]".format(", ".join(param["enum"]))
+
+                if param.get("items", None):
+                    items = ", where each item should be {}".format(param["items"].get("type", ""))
+
+                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(
+                name=tool["name"], desc=tool.get("description", ""), args=param_text
+            )
+            tool_names.append(tool["name"])
+
+        return DEFAULT_TOOL_PROMPT.format(tool_text=tool_text, tool_names=", ".join(tool_names))
+
+    @staticmethod
+    def tool_extractor(content: str) -> Union[str, List[Tuple[str, str]]]:
+        regex = re.compile(r"Action:\s*([a-zA-Z0-9_]+)\s*Action Input:\s*(.+?)(?=\s*Action:|\s*$)", re.DOTALL)
+        action_match: List[Tuple[str, str]] = re.findall(regex, content)
+        if not action_match:
+            return content
+
+        results = []
+        for match in action_match:
+            tool_name = match[0].strip()
+            tool_input = match[1].strip().strip('"').strip("```")
+            try:
+                arguments = json.loads(tool_input)
+                results.append((tool_name, json.dumps(arguments, ensure_ascii=False)))
+            except json.JSONDecodeError:
+                return content
+
+        return results
+
+
+class GLM4ToolUtils(ToolUtils):
+    @staticmethod
+    def get_function_slots() -> SLOTS:
+        return ["{{name}}\n{{arguments}}"]
+
+    @staticmethod
+    def tool_formatter(tools: List[Dict[str, Any]]) -> str:
+        tool_text = ""
+        for tool in tools:
+            tool_text += "\n\n## {name}\n\n{body}\n在调用上述函数时，请使用 Json 格式表示调用的参数。".format(
+                name=tool["name"], body=json.dumps(tool, indent=4, ensure_ascii=False)
+            )
+
+        return GLM4_TOOL_PROMPT.format(tool_text=tool_text)
+
+    @staticmethod
+    def tool_extractor(content: str) -> Union[str, List[Tuple[str, str]]]:
+        if "\n" not in content:
+            return content
+
+        tool_name, tool_input = content.split("\n", maxsplit=1)
+        try:
+            arguments = json.loads(tool_input)
+        except json.JSONDecodeError:
+            return content
+
+        return [(tool_name, json.dumps(arguments, ensure_ascii=False))]

src/llamafactory/eval/evaluator.py

@@ -1,6 +1,42 @@
-# Inspired by: https://github.com/hendrycks/test/blob/master/evaluate_flan.py
+# Copyright 2024 the LlamaFactory team.
+#
+# This code is inspired by the Dan's test library.
+# https://github.com/hendrycks/test/blob/master/evaluate_flan.py
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+# MIT License
+#
+# Copyright (c) 2020 Dan Hendrycks
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
 
-import inspect
 import json
 import os
 from typing import Any, Dict, List, Optional
@@ -26,9 +62,7 @@ class Evaluator:
         self.template = get_template_and_fix_tokenizer(self.tokenizer, self.data_args.template)
         self.model = load_model(self.tokenizer, self.model_args, finetuning_args)
         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
-        ]
+        self.choice_inputs = [self.tokenizer.encode(ch, add_special_tokens=False)[-1] for ch in CHOICES]
 
     @torch.inference_mode()
     def batch_inference(self, batch_input: Dict[str, torch.Tensor]) -> List[str]:
@@ -39,8 +73,11 @@ class Evaluator:
         return [chr(ord("A") + offset.item()) for offset in torch.argmax(choice_probs, dim=-1)]
 
     def eval(self) -> None:
+        eval_task = self.eval_args.task.split("_")[0]
+        eval_split = self.eval_args.task.split("_")[1]
+
         mapping = cached_file(
-            path_or_repo_id=os.path.join(self.eval_args.task_dir, self.eval_args.task),
+            path_or_repo_id=os.path.join(self.eval_args.task_dir, eval_task),
             filename="mapping.json",
             cache_dir=self.model_args.cache_dir,
             token=self.model_args.hf_hub_token,
@@ -53,27 +90,22 @@ class Evaluator:
         pbar = tqdm(categorys.keys(), desc="Processing subjects", position=0)
         results = {}
         for subject in pbar:
-            if "trust_remote_code" in inspect.signature(load_dataset).parameters:  # for datasets==2.16.0
-                kwargs = {"trust_remote_code": True}
-            else:
-                kwargs = {}
-
             dataset = load_dataset(
-                path=os.path.join(self.eval_args.task_dir, self.eval_args.task),
+                path=os.path.join(self.eval_args.task_dir, eval_task),
                 name=subject,
                 cache_dir=self.model_args.cache_dir,
                 download_mode=self.eval_args.download_mode,
                 token=self.model_args.hf_hub_token,
-                **kwargs,
+                trust_remote_code=True,
             )
             pbar.set_postfix_str(categorys[subject]["name"])
             inputs, outputs, labels = [], [], []
-            for i in trange(len(dataset[self.data_args.split]), desc="Formatting batches", position=1, leave=False):
+            for i in trange(len(dataset[eval_split]), desc="Formatting batches", position=1, leave=False):
                 support_set = (
                     dataset["train"].shuffle().select(range(min(self.eval_args.n_shot, len(dataset["train"]))))
                 )
                 messages = self.eval_template.format_example(
-                    target_data=dataset[self.data_args.split][i],
+                    target_data=dataset[eval_split][i],
                     support_set=support_set,
                     subject_name=categorys[subject]["name"],
                 )

src/llamafactory/eval/template.py

@@ -1,3 +1,17 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 from dataclasses import dataclass
 from typing import Dict, List, Sequence, Tuple
 
@@ -10,7 +24,6 @@ class EvalTemplate:
     system: str
     choice: str
     answer: str
-    prefix: str
 
     def _parse_example(self, example: Dict[str, str]) -> Tuple[str, str]:
         r"""
@@ -42,8 +55,8 @@ class EvalTemplate:
 eval_templates: Dict[str, "EvalTemplate"] = {}
 
 
-def _register_eval_template(name: str, system: str, choice: str, answer: str, prefix: str) -> None:
-    eval_templates[name] = EvalTemplate(system=system, choice=choice, answer=answer, prefix=prefix)
+def _register_eval_template(name: str, system: str, choice: str, answer: str) -> None:
+    eval_templates[name] = EvalTemplate(system=system, choice=choice, answer=answer)
 
 
 def get_eval_template(name: str) -> "EvalTemplate":
@@ -56,8 +69,7 @@ _register_eval_template(
     name="en",
     system="The following are multiple choice questions (with answers) about {subject}.\n\n",
     choice="\n{choice}. {content}",
-    answer="\nAnswer: ",
-    prefix=" ",
+    answer="\nAnswer:",
 )
 
 
@@ -66,5 +78,4 @@ _register_eval_template(
     system="以下是中国关于{subject}考试的单项选择题，请选出其中的正确答案。\n\n",
     choice="\n{choice}. {content}",
     answer="\n答案：",
-    prefix=" ",
 )

src/llamafactory/extras/constants.py

@@ -1,3 +1,17 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 from collections import OrderedDict, defaultdict
 from enum import Enum
 from typing import Dict, Optional
@@ -64,6 +78,19 @@ TRAINING_STAGES = {
 
 STAGES_USE_PAIR_DATA = {"rm", "dpo"}
 
+SUPPORTED_CLASS_FOR_BLOCK_DIAG_ATTN = {
+    "cohere",
+    "falcon",
+    "gemma",
+    "gemma2",
+    "llama",
+    "mistral",
+    "phi",
+    "phi3",
+    "qwen2",
+    "starcoder2",
+}
+
 SUPPORTED_CLASS_FOR_S2ATTN = {"llama"}
 
 V_HEAD_WEIGHTS_NAME = "value_head.bin"
@@ -272,6 +299,17 @@ register_model_group(
 )
 
 
+register_model_group(
+    models={
+        "CodeGeeX4-9B-Chat": {
+            DownloadSource.DEFAULT: "THUDM/codegeex4-all-9b",
+            DownloadSource.MODELSCOPE: "ZhipuAI/codegeex4-all-9b",
+        },
+    },
+    template="codegeex4",
+)
+
+
 register_model_group(
     models={
         "CodeGemma-7B": {
@@ -389,6 +427,18 @@ register_model_group(
             DownloadSource.DEFAULT: "deepseek-ai/DeepSeek-V2-Chat",
             DownloadSource.MODELSCOPE: "deepseek-ai/DeepSeek-V2-Chat",
         },
+        "DeepSeek-MoE-Coder-16B-Base": {
+            DownloadSource.DEFAULT: "deepseek-ai/DeepSeek-Coder-V2-Lite-Base",
+        },
+        "DeepSeek-MoE-Coder-236B-Base": {
+            DownloadSource.DEFAULT: "deepseek-ai/DeepSeek-Coder-V2-Base",
+        },
+        "DeepSeek-MoE-Coder-16B-Chat": {
+            DownloadSource.DEFAULT: "deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct",
+        },
+        "DeepSeek-MoE-Coder-236B-Chat": {
+            DownloadSource.DEFAULT: "deepseek-ai/DeepSeek-Coder-V2-Instruct",
+        },
     },
     template="deepseek",
 )
@@ -481,6 +531,22 @@ register_model_group(
         "Gemma-1.1-7B-Chat": {
             DownloadSource.DEFAULT: "google/gemma-1.1-7b-it",
         },
+        "Gemma-2-9B": {
+            DownloadSource.DEFAULT: "google/gemma-2-9b",
+            DownloadSource.MODELSCOPE: "LLM-Research/gemma-2-9b",
+        },
+        "Gemma-2-27B": {
+            DownloadSource.DEFAULT: "google/gemma-2-27b",
+            DownloadSource.MODELSCOPE: "LLM-Research/gemma-2-27b",
+        },
+        "Gemma-2-9B-Chat": {
+            DownloadSource.DEFAULT: "google/gemma-2-9b-it",
+            DownloadSource.MODELSCOPE: "LLM-Research/gemma-2-9b-it",
+        },
+        "Gemma-2-27B-Chat": {
+            DownloadSource.DEFAULT: "google/gemma-2-27b-it",
+            DownloadSource.MODELSCOPE: "LLM-Research/gemma-2-27b-it",
+        },
     },
     template="gemma",
 )
@@ -553,7 +619,26 @@ register_model_group(
 
 register_model_group(
     models={
-        "Jambda-v0.1": {
+        "InternLM2.5-7B": {
+            DownloadSource.DEFAULT: "internlm/internlm2_5-7b",
+            DownloadSource.MODELSCOPE: "Shanghai_AI_Laboratory/internlm2_5-7b",
+        },
+        "InternLM2.5-7B-Chat": {
+            DownloadSource.DEFAULT: "internlm/internlm2_5-7b-chat",
+            DownloadSource.MODELSCOPE: "Shanghai_AI_Laboratory/internlm2_5-7b-chat",
+        },
+        "InternLM2.5-7B-1M-Chat": {
+            DownloadSource.DEFAULT: "internlm/internlm2_5-7b-chat-1m",
+            DownloadSource.MODELSCOPE: "Shanghai_AI_Laboratory/internlm2_5-7b-chat-1m",
+        },
+    },
+    template="intern2",
+)
+
+
+register_model_group(
+    models={
+        "Jamba-v0.1": {
             DownloadSource.DEFAULT: "ai21labs/Jamba-v0.1",
             DownloadSource.MODELSCOPE: "AI-ModelScope/Jamba-v0.1",
         }
@@ -668,6 +753,21 @@ register_model_group(
 )
 
 
+register_model_group(
+    models={
+        "MiniCPM-2B-SFT-Chat": {
+            DownloadSource.DEFAULT: "openbmb/MiniCPM-2B-sft-bf16",
+            DownloadSource.MODELSCOPE: "OpenBMB/miniCPM-bf16",
+        },
+        "MiniCPM-2B-DPO-Chat": {
+            DownloadSource.DEFAULT: "openbmb/MiniCPM-2B-dpo-bf16",
+            DownloadSource.MODELSCOPE: "OpenBMB/MiniCPM-2B-dpo-bf16",
+        },
+    },
+    template="cpm",
+)
+
+
 register_model_group(
     models={
         "Mistral-7B-v0.1": {
@@ -1207,6 +1307,10 @@ register_model_group(
 
 register_model_group(
     models={
+        "TeleChat-1B-Chat": {
+            DownloadSource.DEFAULT: "Tele-AI/TeleChat-1B",
+            DownloadSource.MODELSCOPE: "TeleAI/TeleChat-1B",
+        },
         "TeleChat-7B-Chat": {
             DownloadSource.DEFAULT: "Tele-AI/telechat-7B",
             DownloadSource.MODELSCOPE: "TeleAI/telechat-7B",

src/llamafactory/extras/env.py

@@ -1,3 +1,20 @@
+# Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
+#
+# This code is inspired by the HuggingFace's transformers library.
+# https://github.com/huggingface/transformers/blob/v4.40.0/src/transformers/commands/env.py
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import platform
 
 import accelerate
@@ -6,13 +23,10 @@ import peft
 import torch
 import transformers
 import trl
-from transformers.integrations import is_deepspeed_available
-from transformers.utils import is_bitsandbytes_available, is_torch_cuda_available, is_torch_npu_available
-
-from .packages import is_vllm_available
+from transformers.utils import is_torch_cuda_available, is_torch_npu_available
 
 
-VERSION = "0.8.0"
+VERSION = "0.8.3"
 
 
 def print_env() -> None:
@@ -37,19 +51,25 @@ def print_env() -> None:
         info["NPU type"] = torch.npu.get_device_name()
         info["CANN version"] = torch.version.cann
 
-    if is_deepspeed_available():
+    try:
         import deepspeed  # type: ignore
 
         info["DeepSpeed version"] = deepspeed.__version__
+    except Exception:
+        pass
 
-    if is_bitsandbytes_available():
+    try:
         import bitsandbytes
 
         info["Bitsandbytes version"] = bitsandbytes.__version__
+    except Exception:
+        pass
 
-    if is_vllm_available():
+    try:
         import vllm
 
         info["vLLM version"] = vllm.__version__
+    except Exception:
+        pass
 
     print("\n" + "\n".join(["- {}: {}".format(key, value) for key, value in info.items()]) + "\n")

src/llamafactory/extras/logging.py

@@ -1,3 +1,17 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import logging
 import os
 import sys

src/llamafactory/extras/misc.py

@@ -1,13 +1,29 @@
+# Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
+#
+# This code is inspired by the HuggingFace's PEFT library.
+# https://github.com/huggingface/peft/blob/v0.10.0/src/peft/peft_model.py
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import gc
 import os
-from typing import TYPE_CHECKING, Dict, Tuple
+from typing import TYPE_CHECKING, Tuple, Union
 
 import torch
-from peft import PeftModel
-from transformers import InfNanRemoveLogitsProcessor, LogitsProcessorList, PreTrainedModel
+import transformers.dynamic_module_utils
+from transformers import InfNanRemoveLogitsProcessor, LogitsProcessorList
+from transformers.dynamic_module_utils import get_relative_imports
 from transformers.utils import (
-    SAFE_WEIGHTS_NAME,
-    WEIGHTS_NAME,
     is_torch_bf16_gpu_available,
     is_torch_cuda_available,
     is_torch_mps_available,
@@ -16,7 +32,6 @@ from transformers.utils import (
 )
 from transformers.utils.versions import require_version
 
-from .constants import V_HEAD_SAFE_WEIGHTS_NAME, V_HEAD_WEIGHTS_NAME
 from .logging import get_logger
 
 
@@ -28,7 +43,7 @@ except Exception:
 
 
 if TYPE_CHECKING:
-    from trl import AutoModelForCausalLMWithValueHead
+    from numpy.typing import NDArray
 
     from ..hparams import ModelArguments
 
@@ -58,6 +73,9 @@ class AverageMeter:
 
 
 def check_dependencies() -> None:
+    r"""
+    Checks the version of the required packages.
+    """
     if os.environ.get("DISABLE_VERSION_CHECK", "0").lower() in ["true", "1"]:
         logger.warning("Version checking has been disabled, may lead to unexpected behaviors.")
     else:
@@ -68,7 +86,7 @@ def check_dependencies() -> None:
         require_version("trl>=0.8.6", "To fix: pip install trl>=0.8.6")
 
 
-def count_parameters(model: torch.nn.Module) -> Tuple[int, int]:
+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.
     """
@@ -79,7 +97,7 @@ def count_parameters(model: torch.nn.Module) -> Tuple[int, int]:
         if num_params == 0 and hasattr(param, "ds_numel"):
             num_params = param.ds_numel
 
-        # Due to the design of 4bit linear layers from bitsandbytes, multiply the number of parameters by 2
+        # Due to the design of 4bit linear layers from bitsandbytes, multiply the number of parameters by itemsize
         if param.__class__.__name__ == "Params4bit":
             if hasattr(param, "quant_storage") and hasattr(param.quant_storage, "itemsize"):
                 num_bytes = param.quant_storage.itemsize
@@ -97,55 +115,7 @@ def count_parameters(model: torch.nn.Module) -> Tuple[int, int]:
     return trainable_params, all_param
 
 
-def fix_valuehead_checkpoint(
-    model: "AutoModelForCausalLMWithValueHead", output_dir: str, safe_serialization: bool
-) -> None:
-    r"""
-    The model is already unwrapped.
-
-    There are three cases:
-    1. full tuning without ds_zero3: state_dict = {"model.layers.*": ..., "v_head.summary.*": ...}
-    2. lora tuning without ds_zero3: state_dict = {"v_head.summary.*": ...}
-    3. under deepspeed zero3: state_dict = {"pretrained_model.model.layers.*": ..., "v_head.summary.*": ...}
-
-    We assume `stage3_gather_16bit_weights_on_model_save=true`.
-    """
-    if not isinstance(model.pretrained_model, (PreTrainedModel, PeftModel)):
-        return
-
-    if safe_serialization:
-        from safetensors import safe_open
-        from safetensors.torch import save_file
-
-        path_to_checkpoint = os.path.join(output_dir, SAFE_WEIGHTS_NAME)
-        with safe_open(path_to_checkpoint, framework="pt", device="cpu") as f:
-            state_dict: Dict[str, torch.Tensor] = {key: f.get_tensor(key) for key in f.keys()}
-    else:
-        path_to_checkpoint = os.path.join(output_dir, WEIGHTS_NAME)
-        state_dict: Dict[str, torch.Tensor] = torch.load(path_to_checkpoint, map_location="cpu")
-
-    decoder_state_dict = {}
-    v_head_state_dict = {}
-    for name, param in state_dict.items():
-        if name.startswith("v_head."):
-            v_head_state_dict[name] = param
-        else:
-            decoder_state_dict[name.replace("pretrained_model.", "")] = param
-
-    os.remove(path_to_checkpoint)
-    model.pretrained_model.save_pretrained(
-        output_dir, state_dict=decoder_state_dict or None, safe_serialization=safe_serialization
-    )
-
-    if safe_serialization:
-        save_file(v_head_state_dict, os.path.join(output_dir, V_HEAD_SAFE_WEIGHTS_NAME), metadata={"format": "pt"})
-    else:
-        torch.save(v_head_state_dict, os.path.join(output_dir, V_HEAD_WEIGHTS_NAME))
-
-    logger.info("Value head model saved at: {}".format(output_dir))
-
-
-def get_current_device() -> torch.device:
+def get_current_device() -> "torch.device":
     r"""
     Gets the current available device.
     """
@@ -184,7 +154,14 @@ def get_logits_processor() -> "LogitsProcessorList":
     return logits_processor
 
 
-def infer_optim_dtype(model_dtype: torch.dtype) -> torch.dtype:
+def has_tokenized_data(path: "os.PathLike") -> bool:
+    r"""
+    Checks if the path has a tokenized dataset.
+    """
+    return os.path.isdir(path) and len(os.listdir(path)) > 0
+
+
+def infer_optim_dtype(model_dtype: "torch.dtype") -> "torch.dtype":
     r"""
     Infers the optimal dtype according to the model_dtype and device compatibility.
     """
@@ -203,11 +180,20 @@ def is_gpu_or_npu_available() -> bool:
     return is_torch_npu_available() or is_torch_cuda_available()
 
 
-def has_tokenized_data(path: os.PathLike) -> bool:
-    r"""
-    Checks if the path has a tokenized dataset.
-    """
-    return os.path.isdir(path) and len(os.listdir(path)) > 0
+def numpify(inputs: Union["NDArray", "torch.Tensor"]) -> "NDArray":
+    if isinstance(inputs, torch.Tensor):
+        inputs = inputs.cpu()
+        if inputs.dtype == torch.bfloat16:  # numpy does not support bfloat16 until 1.21.4
+            inputs = inputs.to(torch.float32)
+
+        inputs = inputs.numpy()
+
+    return inputs
+
+
+def skip_check_imports() -> None:
+    if os.environ.get("FORCE_CHECK_IMPORTS", "0").lower() not in ["true", "1"]:
+        transformers.dynamic_module_utils.check_imports = get_relative_imports
 
 
 def torch_gc() -> None:

src/llamafactory/extras/packages.py

@@ -1,5 +1,23 @@
+# Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
+#
+# This code is inspired by the HuggingFace's transformers library.
+# https://github.com/huggingface/transformers/blob/v4.40.0/src/transformers/utils/import_utils.py
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import importlib.metadata
 import importlib.util
+from functools import lru_cache
 from typing import TYPE_CHECKING
 
 from packaging import version
@@ -24,10 +42,6 @@ def is_fastapi_available():
     return _is_package_available("fastapi")
 
 
-def is_flash_attn2_available():
-    return _is_package_available("flash_attn") and _get_package_version("flash_attn") > version.parse("2.0.0")
-
-
 def is_galore_available():
     return _is_package_available("galore_torch")
 
@@ -36,18 +50,10 @@ def is_gradio_available():
     return _is_package_available("gradio")
 
 
-def is_jieba_available():
-    return _is_package_available("jieba")
-
-
 def is_matplotlib_available():
     return _is_package_available("matplotlib")
 
 
-def is_nltk_available():
-    return _is_package_available("nltk")
-
-
 def is_pillow_available():
     return _is_package_available("PIL")
 
@@ -60,10 +66,6 @@ def is_rouge_available():
     return _is_package_available("rouge_chinese")
 
 
-def is_sdpa_available():
-    return _get_package_version("torch") > version.parse("2.1.1")
-
-
 def is_starlette_available():
     return _is_package_available("sse_starlette")
 
@@ -74,3 +76,13 @@ def is_uvicorn_available():
 
 def is_vllm_available():
     return _is_package_available("vllm")
+
+
+@lru_cache
+def is_vllm_version_greater_than_0_5():
+    return _get_package_version("vllm") >= version.parse("0.5.0")
+
+
+@lru_cache
+def is_vllm_version_greater_than_0_5_1():
+    return _get_package_version("vllm") >= version.parse("0.5.1")

src/llamafactory/extras/ploting.py

@@ -1,3 +1,17 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import json
 import math
 import os