release v0.6.2

Former-commit-id: f92ad0a62d957b595f6a76a5403216b163eb3d17

.github/SECURITY.md

@@ -1,6 +1,6 @@
 # Reporting Security Issues
 
-To report a security issue, please use the GitHub Security Advisory ["Report a Vulnerability"](https://github.com/electron/electron/security/advisories/new) tab.
+To report a security issue, please use the GitHub Security Advisory ["Report a Vulnerability"](https://github.com/hiyouga/LLaMA-Factory/security/advisories/new) tab.
 
 We will send a response indicating the next steps in handling your report. After the initial reply to your report, the security team will keep you informed of the progress towards a fix and full announcement, and may ask for additional information or guidance.
 

README.md

@@ -5,7 +5,7 @@
 [![GitHub last commit](https://img.shields.io/github/last-commit/hiyouga/LLaMA-Factory)](https://github.com/hiyouga/LLaMA-Factory/commits/main)
 [![PyPI](https://img.shields.io/pypi/v/llmtuner)](https://pypi.org/project/llmtuner/)
 [![Downloads](https://static.pepy.tech/badge/llmtuner)](https://pypi.org/project/llmtuner/)
-[![Citation](https://img.shields.io/badge/citation-27-green)](#projects-using-llama-factory)
+[![Citation](https://img.shields.io/badge/citation-28-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)
@@ -44,7 +44,7 @@ Choose your path:
 ## Features
 
 - **Various models**: LLaMA, Mistral, Mixtral-MoE, Qwen, Yi, Gemma, Baichuan, ChatGLM, Phi, etc.
-- **Integrated methods**: (Continuous) pre-training, supervised fine-tuning, reward modeling, PPO and DPO.
+- **Integrated methods**: (Continuous) pre-training, supervised fine-tuning, reward modeling, PPO, DPO and ORPO.
 - **Scalable resources**: 32-bit full-tuning, 16-bit freeze-tuning, 16-bit LoRA and 2/4/8-bit QLoRA via AQLM/AWQ/GPTQ/LLM.int8.
 - **Advanced algorithms**: GaLore, DoRA, LongLoRA, LLaMA Pro, LoRA+, LoftQ and Agent tuning.
 - **Practical tricks**: FlashAttention-2, Unsloth, RoPE scaling, NEFTune and rsLoRA.
@@ -53,7 +53,7 @@ Choose your path:
 
 ## Benchmark
 
-Compared to ChatGLM's [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/ptuning), LLaMA-Factory's LoRA tuning offers up to **3.7 times faster** training speed with a better Rouge score on the advertising text generation task. By leveraging 4-bit quantization technique, LLaMA-Factory's QLoRA further improves the efficiency regarding the GPU memory.
+Compared to ChatGLM's [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/ptuning), LLaMA Factory's LoRA tuning offers up to **3.7 times faster** training speed with a better Rouge score on the advertising text generation task. By leveraging 4-bit quantization technique, LLaMA Factory's QLoRA further improves the efficiency regarding the GPU memory.
 
 ![benchmark](assets/benchmark.svg)
 
@@ -62,22 +62,24 @@ Compared to ChatGLM's [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/
 - **Training Speed**: the number of training samples processed per second during the training. (bs=4, cutoff_len=1024)
 - **Rouge Score**: Rouge-2 score on the development set of the [advertising text generation](https://aclanthology.org/D19-1321.pdf) task. (bs=4, cutoff_len=1024)
 - **GPU Memory**: Peak GPU memory usage in 4-bit quantized training. (bs=1, cutoff_len=1024)
-- We adopt `pre_seq_len=128` for ChatGLM's P-Tuning and `lora_rank=32` for LLaMA-Factory's LoRA tuning.
+- We adopt `pre_seq_len=128` for ChatGLM's P-Tuning and `lora_rank=32` for LLaMA Factory's LoRA tuning.
 
 </details>
 
 ## Changelog
 
+[24/03/31] We supported **[ORPO](https://arxiv.org/abs/2403.07691)**. See `examples/lora_single_gpu` for usage.
+
 [24/03/21] Our paper "[LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models](https://arxiv.org/abs/2403.13372)" is available at arXiv!
 
-[24/03/20] We supported **FSDP+QLoRA** that fine-tunes a 70B model on 2x24GB GPUs. See `examples/fsdp_qlora` for usage.
+[24/03/20] We supported **FSDP+QLoRA** that fine-tunes a 70B model on 2x24GB GPUs. See `examples/extras/fsdp_qlora` for usage.
+
+<details><summary>Full Changelog</summary>
 
 [24/03/13] We supported **[LoRA+](https://arxiv.org/abs/2402.12354)**. See `examples/extras/loraplus` for usage.
 
 [24/03/07] We supported gradient low-rank projection (**[GaLore](https://arxiv.org/abs/2403.03507)**) algorithm. See `examples/extras/galore` for usage.
 
-<details><summary>Full Changelog</summary>
-
 [24/03/07] We integrated **[vLLM](https://github.com/vllm-project/vllm)** for faster and concurrent inference. Try `--infer_backend vllm` to enjoy **270%** inference speed. (LoRA is not yet supported, merge it first.)
 
 [24/02/28] We supported weight-decomposed LoRA (**[DoRA](https://arxiv.org/abs/2402.09353)**). Try `--use_dora` to activate DoRA training.
@@ -136,12 +138,11 @@ Compared to ChatGLM's [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/
 | [InternLM2](https://huggingface.co/internlm)             | 7B/20B                      | wqkv              | intern2   |
 | [LLaMA](https://github.com/facebookresearch/llama)       | 7B/13B/33B/65B              | q_proj,v_proj     | -         |
 | [LLaMA-2](https://huggingface.co/meta-llama)             | 7B/13B/70B                  | q_proj,v_proj     | llama2    |
-| [Mistral](https://huggingface.co/mistralai)              | 7B                          | q_proj,v_proj     | mistral   |
-| [Mixtral](https://huggingface.co/mistralai)              | 8x7B                        | q_proj,v_proj     | mistral   |
+| [Mistral/Mixtral](https://huggingface.co/mistralai)      | 7B/8x7B                     | q_proj,v_proj     | mistral   |
 | [OLMo](https://huggingface.co/allenai)                   | 1B/7B                       | att_proj          | olmo      |
 | [Phi-1.5/2](https://huggingface.co/microsoft)            | 1.3B/2.7B                   | q_proj,v_proj     | -         |
 | [Qwen](https://huggingface.co/Qwen)                      | 1.8B/7B/14B/72B             | c_attn            | qwen      |
-| [Qwen1.5](https://huggingface.co/Qwen)                   | 0.5B/1.8B/4B/7B/14B/72B     | q_proj,v_proj     | qwen      |
+| [Qwen1.5 (MoE)](https://huggingface.co/Qwen)             | 0.5B/1.8B/4B/7B/14B/32B/72B | q_proj,v_proj     | qwen      |
 | [StarCoder2](https://huggingface.co/bigcode)             | 3B/7B/15B                   | q_proj,v_proj     | -         |
 | [XVERSE](https://huggingface.co/xverse)                  | 7B/13B/65B                  | q_proj,v_proj     | xverse    |
 | [Yi](https://huggingface.co/01-ai)                       | 6B/9B/34B                   | q_proj,v_proj     | yi        |
@@ -165,9 +166,7 @@ You also can add a custom chat template to [template.py](src/llmtuner/data/templ
 | Reward Modeling        | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
 | PPO Training           | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
 | DPO Training           | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
-
-> [!NOTE]
-> Use `--quantization_bit 4` argument to enable QLoRA.
+| ORPO Training          | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
 
 ## Provided Datasets
 
@@ -246,8 +245,6 @@ You also can add a custom chat template to [template.py](src/llmtuner/data/templ
 
 </details>
 
-Please refer to [data/README.md](data/README.md) for details.
-
 Some datasets require confirmation before using them, so we recommend logging in with your Hugging Face account using these commands.
 
 ```bash
@@ -261,8 +258,8 @@ huggingface-cli login
 | ------------ | ------- | --------- |
 | python       | 3.8     | 3.10      |
 | torch        | 1.13.1  | 2.2.0     |
-| transformers | 4.37.2  | 4.39.1    |
-| datasets     | 2.14.3  | 2.17.1    |
+| transformers | 4.37.2  | 4.39.3    |
+| datasets     | 2.14.3  | 2.18.0    |
 | accelerate   | 0.27.2  | 0.28.0    |
 | peft         | 0.9.0   | 0.10.0    |
 | trl          | 0.8.1   | 0.8.1     |
@@ -291,23 +288,27 @@ huggingface-cli login
 
 ## Getting Started
 
-### Data Preparation (optional)
+### Data Preparation
 
-Please refer to [data/README.md](data/README.md) for checking the details about the format of dataset files. You can either use a single `.json` file or a [dataset loading script](https://huggingface.co/docs/datasets/dataset_script) with multiple files to create a custom dataset.
+Please refer to [data/README.md](data/README.md) for checking the details about the format of dataset files. You can either use datasets on HuggingFace / ModelScope hub or load the dataset in local disk.
 
 > [!NOTE]
-> Please update `data/dataset_info.json` to use your custom dataset. About the format of this file, please refer to `data/README.md`.
+> Please update `data/dataset_info.json` to use your custom dataset.
 
-### Dependence Installation (optional)
+### Dependence Installation
 
 ```bash
 git clone https://github.com/hiyouga/LLaMA-Factory.git
 conda create -n llama_factory python=3.10
 conda activate llama_factory
 cd LLaMA-Factory
-pip install -r requirements.txt
+pip install -e .[metrics]
 ```
 
+Extra dependencies available: deepspeed, metrics, unsloth, galore, vllm, bitsandbytes, gptq, awq, aqlm, qwen, modelscope, quality
+
+<details><summary>For Windows users</summary>
+
 If you want to enable the quantized LoRA (QLoRA) on the Windows platform, you will be required to install a pre-built version of `bitsandbytes` library, which supports CUDA 11.1 to 12.2, please select the appropriate [release version](https://github.com/jllllll/bitsandbytes-windows-webui/releases/tag/wheels) based on your CUDA version.
 
 ```bash
@@ -316,357 +317,24 @@ pip install https://github.com/jllllll/bitsandbytes-windows-webui/releases/downl
 
 To enable FlashAttention-2 on the Windows platform, you need to install the precompiled `flash-attn` library, which supports CUDA 12.1 to 12.2. Please download the corresponding version from [flash-attention](https://github.com/bdashore3/flash-attention/releases) based on your requirements.
 
-### Use ModelScope Hub (optional)
+</details>
 
-If you have trouble with downloading models and datasets from Hugging Face, you can use LLaMA-Factory together with ModelScope in the following manner.
-
-```bash
-export USE_MODELSCOPE_HUB=1 # `set USE_MODELSCOPE_HUB=1` for Windows
-```
-
-Then you can train the corresponding model by specifying a model ID of the ModelScope Hub. (find a full list of model IDs at [ModelScope Hub](https://modelscope.cn/models))
-
-```bash
-CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
-    --model_name_or_path modelscope/Llama-2-7b-ms \
-    ... # arguments (same as below)
-```
-
-LLaMA Board also supports using the models and datasets on the ModelScope Hub.
-
-```bash
-CUDA_VISIBLE_DEVICES=0 USE_MODELSCOPE_HUB=1 python src/train_web.py
-```
-
-### Train on a single GPU
+### LLaMA Board GUI
 
 > [!IMPORTANT]
-> If you want to train models on multiple GPUs, please refer to [Distributed Training](#distributed-training).
+> LLaMA Board GUI only supports training on a single GPU, please use [CLI](#command-line-interface) for distributed training.
 
-
-#### LLaMA Board GUI
+#### Use local environment
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 python src/train_web.py
+export CUDA_VISIBLE_DEVICES=0 # `set CUDA_VISIBLE_DEVICES=0` for Windows
+python src/train_web.py # or python -m llmtuner.webui.interface
 ```
 
-#### Pre-Training
-
-```bash
-CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
-    --stage pt \
-    --do_train \
-    --model_name_or_path path_to_llama_model \
-    --dataset wiki_demo \
-    --finetuning_type lora \
-    --lora_target q_proj,v_proj \
-    --output_dir path_to_pt_checkpoint \
-    --overwrite_cache \
-    --per_device_train_batch_size 4 \
-    --gradient_accumulation_steps 4 \
-    --lr_scheduler_type cosine \
-    --logging_steps 10 \
-    --save_steps 1000 \
-    --learning_rate 5e-5 \
-    --num_train_epochs 3.0 \
-    --plot_loss \
-    --fp16
-```
-
-#### Supervised Fine-Tuning
-
-```bash
-CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
-    --stage sft \
-    --do_train \
-    --model_name_or_path path_to_llama_model \
-    --dataset alpaca_gpt4_en \
-    --template default \
-    --finetuning_type lora \
-    --lora_target q_proj,v_proj \
-    --output_dir path_to_sft_checkpoint \
-    --overwrite_cache \
-    --per_device_train_batch_size 4 \
-    --gradient_accumulation_steps 4 \
-    --lr_scheduler_type cosine \
-    --logging_steps 10 \
-    --save_steps 1000 \
-    --learning_rate 5e-5 \
-    --num_train_epochs 3.0 \
-    --plot_loss \
-    --fp16
-```
-
-#### Reward Modeling
-
-```bash
-CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
-    --stage rm \
-    --do_train \
-    --model_name_or_path path_to_llama_model \
-    --adapter_name_or_path path_to_sft_checkpoint \
-    --create_new_adapter \
-    --dataset comparison_gpt4_en \
-    --template default \
-    --finetuning_type lora \
-    --lora_target q_proj,v_proj \
-    --output_dir path_to_rm_checkpoint \
-    --per_device_train_batch_size 2 \
-    --gradient_accumulation_steps 4 \
-    --lr_scheduler_type cosine \
-    --logging_steps 10 \
-    --save_steps 1000 \
-    --learning_rate 1e-5 \
-    --num_train_epochs 1.0 \
-    --plot_loss \
-    --fp16
-```
-
-#### PPO Training
-
-```bash
-CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
-    --stage ppo \
-    --do_train \
-    --model_name_or_path path_to_llama_model \
-    --adapter_name_or_path path_to_sft_checkpoint \
-    --create_new_adapter \
-    --dataset alpaca_gpt4_en \
-    --template default \
-    --finetuning_type lora \
-    --lora_target q_proj,v_proj \
-    --reward_model path_to_rm_checkpoint \
-    --output_dir path_to_ppo_checkpoint \
-    --per_device_train_batch_size 2 \
-    --gradient_accumulation_steps 4 \
-    --lr_scheduler_type cosine \
-    --top_k 0 \
-    --top_p 0.9 \
-    --logging_steps 10 \
-    --save_steps 1000 \
-    --learning_rate 1e-5 \
-    --num_train_epochs 1.0 \
-    --plot_loss \
-    --fp16
-```
-
-> [!TIP]
-> Use `--adapter_name_or_path path_to_sft_checkpoint,path_to_ppo_checkpoint` to infer the fine-tuned model if `--create_new_adapter` was enabled.
-
-> [!WARNING]
-> Use `--per_device_train_batch_size=1` for LLaMA-2 models in fp16 PPO training.
-
-#### DPO Training
-
-```bash
-CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
-    --stage dpo \
-    --do_train \
-    --model_name_or_path path_to_llama_model \
-    --adapter_name_or_path path_to_sft_checkpoint \
-    --create_new_adapter \
-    --dataset comparison_gpt4_en \
-    --template default \
-    --finetuning_type lora \
-    --lora_target q_proj,v_proj \
-    --output_dir path_to_dpo_checkpoint \
-    --per_device_train_batch_size 2 \
-    --gradient_accumulation_steps 4 \
-    --lr_scheduler_type cosine \
-    --logging_steps 10 \
-    --save_steps 1000 \
-    --learning_rate 1e-5 \
-    --num_train_epochs 1.0 \
-    --plot_loss \
-    --fp16
-```
-
-> [!TIP]
-> Use `--adapter_name_or_path path_to_sft_checkpoint,path_to_dpo_checkpoint` to infer the fine-tuned model if `--create_new_adapter` was enabled.
-
-### Distributed Training
-
-#### Use Huggingface Accelerate
-
-```bash
-accelerate launch --config_file config.yaml src/train_bash.py \
-    --ddp_timeout 180000000 \
-    ... # arguments (same as above)
-```
-
-<details><summary>Example config.yaml for LoRA training</summary>
-
-```yaml
-compute_environment: LOCAL_MACHINE
-debug: false
-distributed_type: MULTI_GPU
-downcast_bf16: 'no'
-gpu_ids: all
-machine_rank: 0
-main_training_function: main
-mixed_precision: fp16
-num_machines: 1
-num_processes: 4
-rdzv_backend: static
-same_network: true
-tpu_env: []
-tpu_use_cluster: false
-tpu_use_sudo: false
-use_cpu: false
-```
-
-</details>
-
-> [!TIP]
-> We commend using Accelerate for LoRA tuning.
-
-#### Use DeepSpeed
-
-```bash
-deepspeed --num_gpus 8 src/train_bash.py \
-    --deepspeed ds_config.json \
-    --ddp_timeout 180000000 \
-    ... # arguments (same as above)
-```
-
-<details><summary>Example ds_config.json for full-parameter training with DeepSpeed ZeRO-2</summary>
-
-```json
-{
-  "train_batch_size": "auto",
-  "train_micro_batch_size_per_gpu": "auto",
-  "gradient_accumulation_steps": "auto",
-  "gradient_clipping": "auto",
-  "zero_allow_untested_optimizer": true,
-  "fp16": {
-    "enabled": "auto",
-    "loss_scale": 0,
-    "loss_scale_window": 1000,
-    "initial_scale_power": 16,
-    "hysteresis": 2,
-    "min_loss_scale": 1
-  },
-  "bf16": {
-    "enabled": "auto"
-  },
-  "zero_optimization": {
-    "stage": 2,
-    "allgather_partitions": true,
-    "allgather_bucket_size": 5e8,
-    "overlap_comm": true,
-    "reduce_scatter": true,
-    "reduce_bucket_size": 5e8,
-    "contiguous_gradients": true,
-    "round_robin_gradients": true
-  }
-}
-```
-
-</details>
-
-> [!TIP]
-> Refer to [examples](examples) for more training scripts.
-
-### Merge LoRA weights and export model
-
-```bash
-CUDA_VISIBLE_DEVICES= python src/export_model.py \
-    --model_name_or_path path_to_llama_model \
-    --adapter_name_or_path path_to_checkpoint \
-    --template default \
-    --finetuning_type lora \
-    --export_dir path_to_export \
-    --export_size 2 \
-    --export_legacy_format False
-```
-
-> [!WARNING]
-> Merging LoRA weights into a quantized model is not supported.
-
-> [!TIP]
-> Use `--model_name_or_path path_to_export` solely to use the exported model.
-> 
-> Use `CUDA_VISIBLE_DEVICES=0`, `--export_quantization_bit 4` and `--export_quantization_dataset data/c4_demo.json` to quantize the model with AutoGPTQ after merging the LoRA weights.
-
-### Inference with OpenAI-style API
-
-```bash
-CUDA_VISIBLE_DEVICES=0 API_PORT=8000 python src/api_demo.py \
-    --model_name_or_path path_to_llama_model \
-    --adapter_name_or_path path_to_checkpoint \
-    --template default \
-    --finetuning_type lora
-```
-
-> [!TIP]
-> Visit `http://localhost:8000/docs` for API documentation.
-
-### Inference with command line
-
-```bash
-CUDA_VISIBLE_DEVICES=0 python src/cli_demo.py \
-    --model_name_or_path path_to_llama_model \
-    --adapter_name_or_path path_to_checkpoint \
-    --template default \
-    --finetuning_type lora
-```
-
-### Inference with web browser
-
-```bash
-CUDA_VISIBLE_DEVICES=0 python src/web_demo.py \
-    --model_name_or_path path_to_llama_model \
-    --adapter_name_or_path path_to_checkpoint \
-    --template default \
-    --finetuning_type lora
-```
-
-### Evaluation
-
-```bash
-CUDA_VISIBLE_DEVICES=0 python src/evaluate.py \
-    --model_name_or_path path_to_llama_model \
-    --adapter_name_or_path path_to_checkpoint \
-    --template vanilla \
-    --finetuning_type lora \
-    --task mmlu \
-    --split test \
-    --lang en \
-    --n_shot 5 \
-    --batch_size 4
-```
-
-### Predict
-
-```bash
-CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
-    --stage sft \
-    --do_predict \
-    --model_name_or_path path_to_llama_model \
-    --adapter_name_or_path path_to_checkpoint \
-    --dataset alpaca_gpt4_en \
-    --template default \
-    --finetuning_type lora \
-    --output_dir path_to_predict_result \
-    --per_device_eval_batch_size 1 \
-    --max_samples 100 \
-    --predict_with_generate \
-    --fp16
-```
-
-> [!WARNING]
-> Use `--per_device_train_batch_size=1` for LLaMA-2 models in fp16 predict.
-
-> [!TIP]
-> We recommend using `--per_device_eval_batch_size=1` and `--max_target_length 128` at 4/8-bit predict.
-
-### Dockerize Training
-
 #### Use Docker
 
 ```bash
 docker build -f ./Dockerfile -t llama-factory:latest .
-
 docker run --gpus=all \
     -v ./hf_cache:/root/.cache/huggingface/ \
     -v ./data:/app/data \
@@ -684,14 +352,46 @@ docker run --gpus=all \
 docker compose -f ./docker-compose.yml up -d
 ```
 
-> [!TIP]
-> Details about volume:
-> * hf_cache: Utilize Hugging Face cache on the host machine. Reassignable if a cache already exists in a different directory.
-> * data: Place datasets on this dir of the host machine so that they can be selected on LLaMA Board GUI.
-> * output: Set export dir to this location so that the merged result can be accessed directly on the host machine.
+<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.
+- data: Place datasets on this dir of the host machine so that they can be selected on LLaMA Board GUI.
+- output: Set export dir to this location so that the merged result can be accessed directly on the host machine.
+
+</details>
+
+### Command Line Interface
+
+See [examples/README.md](examples/README.md) for usage.
+
+Use `python src/train_bash.py -h` to display arguments description.
+
+### Deploy with OpenAI-style API and vLLM
+
+```bash
+CUDA_VISIBLE_DEVICES=0,1 API_PORT=8000 python src/api_demo.py \
+    --model_name_or_path mistralai/Mistral-7B-Instruct-v0.2 \
+    --template mistral \
+    --infer_backend vllm \
+    --vllm_enforce_eager
+```
+
+### Use ModelScope Hub
+
+If you have trouble with downloading models and datasets from Hugging Face, you can use ModelScope.
+
+```bash
+export USE_MODELSCOPE_HUB=1 # `set USE_MODELSCOPE_HUB=1` for Windows
+```
+
+Train the model by specifying a model ID of the ModelScope Hub as the `--model_name_or_path`. You can find a full list of model IDs at [ModelScope Hub](https://modelscope.cn/models), e.g., `modelscope/Llama-2-7b-ms`.
 
 ## Projects using LLaMA Factory
 
+If you have a project that should be incorporated, please contact via email or create a pull request.
+
+<details><summary>Click to show</summary>
+
 1. Wang et al. ESRL: Efficient Sampling-based Reinforcement Learning for Sequence Generation. 2023. [[arxiv]](https://arxiv.org/abs/2308.02223)
 1. Yu et al. Open, Closed, or Small Language Models for Text Classification? 2023. [[arxiv]](https://arxiv.org/abs/2308.10092)
 1. Wang et al. UbiPhysio: Support Daily Functioning, Fitness, and Rehabilitation with Action Understanding and Feedback in Natural Language. 2023. [[arxiv]](https://arxiv.org/abs/2308.10526)
@@ -713,6 +413,7 @@ docker compose -f ./docker-compose.yml up -d
 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. 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. **[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.
@@ -720,8 +421,7 @@ docker compose -f ./docker-compose.yml up -d
 1. **[CareGPT](https://github.com/WangRongsheng/CareGPT)**: A series of large language models for Chinese medical domain, based on LLaMA2-7B and Baichuan-13B.
 1. **[MachineMindset](https://github.com/PKU-YuanGroup/Machine-Mindset/)**: A series of MBTI Personality large language models, capable of giving any LLM 16 different personality types based on different datasets and training methods.
 
-> [!TIP]
-> If you have a project that should be incorporated, please contact via email or create a pull request.
+</details>
 
 ## License
 
@@ -745,7 +445,7 @@ If this work is helpful, please kindly cite as:
 
 ## Acknowledgement
 
-This repo benefits from [PEFT](https://github.com/huggingface/peft), [QLoRA](https://github.com/artidoro/qlora) and [FastChat](https://github.com/lm-sys/FastChat). Thanks for their wonderful works.
+This repo benefits from [PEFT](https://github.com/huggingface/peft), [TRL](https://github.com/huggingface/trl), [QLoRA](https://github.com/artidoro/qlora) and [FastChat](https://github.com/lm-sys/FastChat). Thanks for their wonderful works.
 
 ## Star History
 

README_zh.md

@@ -5,7 +5,7 @@
 [![GitHub last commit](https://img.shields.io/github/last-commit/hiyouga/LLaMA-Factory)](https://github.com/hiyouga/LLaMA-Factory/commits/main)
 [![PyPI](https://img.shields.io/pypi/v/llmtuner)](https://pypi.org/project/llmtuner/)
 [![Downloads](https://static.pepy.tech/badge/llmtuner)](https://pypi.org/project/llmtuner/)
-[![Citation](https://img.shields.io/badge/citation-27-green)](#使用了-llama-factory-的项目)
+[![Citation](https://img.shields.io/badge/citation-28-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)
@@ -44,7 +44,7 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/ec36a9dd-37f4-4f72-81bd
 ## 项目特色
 
 - **多种模型**：LLaMA、Mistral、Mixtral-MoE、Qwen、Yi、Gemma、Baichuan、ChatGLM、Phi 等等。
-- **集成方法**：（增量）预训练、指令监督微调、奖励模型训练、PPO 训练和 DPO 训练。
+- **集成方法**：（增量）预训练、指令监督微调、奖励模型训练、PPO 训练、DPO 训练和 ORPO 训练。
 - **多种精度**：32 比特全参数微调、16 比特冻结微调、16 比特 LoRA 微调和基于 AQLM/AWQ/GPTQ/LLM.int8 的 2/4/8 比特 QLoRA 微调。
 - **先进算法**：GaLore、DoRA、LongLoRA、LLaMA Pro、LoRA+、LoftQ 和 Agent 微调。
 - **实用技巧**：FlashAttention-2、Unsloth、RoPE scaling、NEFTune 和 rsLoRA。
@@ -53,7 +53,7 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/ec36a9dd-37f4-4f72-81bd
 
 ## 性能指标
 
-与 ChatGLM 官方的 [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/ptuning) 微调相比，LLaMA-Factory 的 LoRA 微调提供了 **3.7 倍**的加速比，同时在广告文案生成任务上取得了更高的 Rouge 分数。结合 4 比特量化技术，LLaMA-Factory 的 QLoRA 微调进一步降低了 GPU 显存消耗。
+与 ChatGLM 官方的 [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/ptuning) 微调相比，LLaMA Factory 的 LoRA 微调提供了 **3.7 倍**的加速比，同时在广告文案生成任务上取得了更高的 Rouge 分数。结合 4 比特量化技术，LLaMA Factory 的 QLoRA 微调进一步降低了 GPU 显存消耗。
 
 ![benchmark](assets/benchmark.svg)
 
@@ -62,22 +62,24 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/ec36a9dd-37f4-4f72-81bd
 - **Training Speed**: 训练阶段每秒处理的样本数量。（批处理大小=4，截断长度=1024）
 - **Rouge Score**: [广告文案生成](https://aclanthology.org/D19-1321.pdf)任务验证集上的 Rouge-2 分数。（批处理大小=4，截断长度=1024）
 - **GPU Memory**: 4 比特量化训练的 GPU 显存峰值。（批处理大小=1，截断长度=1024）
-- 我们在 ChatGLM 的 P-Tuning 中采用 `pre_seq_len=128`，在 LLaMA-Factory 的 LoRA 微调中采用 `lora_rank=32`。
+- 我们在 ChatGLM 的 P-Tuning 中采用 `pre_seq_len=128`，在 LLaMA Factory 的 LoRA 微调中采用 `lora_rank=32`。
 
 </details>
 
 ## 更新日志
 
+[24/03/31] 我们支持了 **[ORPO](https://arxiv.org/abs/2403.07691)**。详细用法请参照 `examples/lora_single_gpu`。
+
 [24/03/21] 我们的论文 "[LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models](https://arxiv.org/abs/2403.13372)" 可在 arXiv 上查看！
 
-[24/03/20] 我们支持了能在 2x24GB GPU 上微调 70B 模型的 **FSDP+QLoRA**。详细用法请参照 `examples/fsdp_qlora`。
+[24/03/20] 我们支持了能在 2x24GB GPU 上微调 70B 模型的 **FSDP+QLoRA**。详细用法请参照 `examples/extras/fsdp_qlora`。
+
+<details><summary>展开日志</summary>
 
 [24/03/13] 我们支持了 **[LoRA+](https://arxiv.org/abs/2402.12354)**。详细用法请参照 `examples/extras/loraplus`。
 
 [24/03/07] 我们支持了梯度低秩投影（**[GaLore](https://arxiv.org/abs/2403.03507)**）算法。详细用法请参照 `examples/extras/galore`。
 
-<details><summary>展开日志</summary>
-
 [24/03/07] 我们集成了 **[vLLM](https://github.com/vllm-project/vllm)** 以实现极速并发推理。请使用 `--infer_backend vllm` 来获得 **270%** 的推理速度。（尚不支持 LoRA，请先合并权重。）
 
 [24/02/28] 我们支持了 **[DoRA](https://arxiv.org/abs/2402.09353)** 微调。请使用 `--use_dora` 参数进行 DoRA 微调。
@@ -136,12 +138,11 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/ec36a9dd-37f4-4f72-81bd
 | [InternLM2](https://huggingface.co/internlm)             | 7B/20B                      | wqkv              | intern2   |
 | [LLaMA](https://github.com/facebookresearch/llama)       | 7B/13B/33B/65B              | q_proj,v_proj     | -         |
 | [LLaMA-2](https://huggingface.co/meta-llama)             | 7B/13B/70B                  | q_proj,v_proj     | llama2    |
-| [Mistral](https://huggingface.co/mistralai)              | 7B                          | q_proj,v_proj     | mistral   |
-| [Mixtral](https://huggingface.co/mistralai)              | 8x7B                        | q_proj,v_proj     | mistral   |
+| [Mistral/Mixtral](https://huggingface.co/mistralai)      | 7B/8x7B                     | q_proj,v_proj     | mistral   |
 | [OLMo](https://huggingface.co/allenai)                   | 1B/7B                       | att_proj          | olmo      |
 | [Phi-1.5/2](https://huggingface.co/microsoft)            | 1.3B/2.7B                   | q_proj,v_proj     | -         |
 | [Qwen](https://huggingface.co/Qwen)                      | 1.8B/7B/14B/72B             | c_attn            | qwen      |
-| [Qwen1.5](https://huggingface.co/Qwen)                   | 0.5B/1.8B/4B/7B/14B/72B     | q_proj,v_proj     | qwen      |
+| [Qwen1.5 (MoE)](https://huggingface.co/Qwen)             | 0.5B/1.8B/4B/7B/14B/32B/72B | q_proj,v_proj     | qwen      |
 | [StarCoder2](https://huggingface.co/bigcode)             | 3B/7B/15B                   | q_proj,v_proj     | -         |
 | [XVERSE](https://huggingface.co/xverse)                  | 7B/13B/65B                  | q_proj,v_proj     | xverse    |
 | [Yi](https://huggingface.co/01-ai)                       | 6B/9B/34B                   | q_proj,v_proj     | yi        |
@@ -165,9 +166,7 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/ec36a9dd-37f4-4f72-81bd
 | 奖励模型训练            | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
 | PPO 训练               | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
 | DPO 训练               | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
-
-> [!NOTE]
-> 请使用 `--quantization_bit 4` 参数来启用 QLoRA 训练。
+| ORPO 训练              | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
 
 ## 数据集
 
@@ -246,8 +245,6 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/ec36a9dd-37f4-4f72-81bd
 
 </details>
 
-使用方法请参考 [data/README_zh.md](data/README_zh.md) 文件。
-
 部分数据集的使用需要确认，我们推荐使用下述命令登录您的 Hugging Face 账户。
 
 ```bash
@@ -261,8 +258,8 @@ huggingface-cli login
 | ------------ | ------- | --------- |
 | python       | 3.8     | 3.10      |
 | torch        | 1.13.1  | 2.2.0     |
-| transformers | 4.37.2  | 4.39.1    |
-| datasets     | 2.14.3  | 2.17.1    |
+| transformers | 4.37.2  | 4.39.3    |
+| datasets     | 2.14.3  | 2.18.0    |
 | accelerate   | 0.27.2  | 0.28.0    |
 | peft         | 0.9.0   | 0.10.0    |
 | trl          | 0.8.1   | 0.8.1     |
@@ -291,23 +288,27 @@ huggingface-cli login
 
 ## 如何使用
 
-### 数据准备（可跳过）
+### 数据准备
 
-关于数据集文件的格式，请参考 [data/README_zh.md](data/README_zh.md) 的内容。构建自定义数据集时，既可以使用单个 `.json` 文件，也可以使用一个[数据加载脚本](https://huggingface.co/docs/datasets/dataset_script)和多个文件。
+关于数据集文件的格式，请参考 [data/README_zh.md](data/README_zh.md) 的内容。你可以使用 HuggingFace / ModelScope 上的数据集或加载本地数据集。
 
 > [!NOTE]
-> 使用自定义数据集时，请更新 `data/dataset_info.json` 文件，该文件的格式请参考 `data/README_zh.md`。
+> 使用自定义数据集时，请更新 `data/dataset_info.json` 文件。
 
-### 环境搭建（可跳过）
+### 安装依赖
 
 ```bash
 git clone https://github.com/hiyouga/LLaMA-Factory.git
 conda create -n llama_factory python=3.10
 conda activate llama_factory
 cd LLaMA-Factory
-pip install -r requirements.txt
+pip install -e .[metrics]
 ```
 
+可选的额外依赖项：deepspeed、metrics、unsloth、galore、vllm、bitsandbytes、gptq、awq、aqlm、qwen、modelscope、quality
+
+<details><summary>Windows 用户指南</summary>
+
 如果要在 Windows 平台上开启量化 LoRA（QLoRA），需要安装预编译的 `bitsandbytes` 库, 支持 CUDA 11.1 到 12.2, 请根据您的 CUDA 版本情况选择适合的[发布版本](https://github.com/jllllll/bitsandbytes-windows-webui/releases/tag/wheels)。
 
 ```bash
@@ -316,356 +317,24 @@ pip install https://github.com/jllllll/bitsandbytes-windows-webui/releases/downl
 
 如果要在 Windows 平台上开启 FlashAttention-2，需要安装预编译的 `flash-attn` 库，支持 CUDA 12.1 到 12.2，请根据需求到 [flash-attention](https://github.com/bdashore3/flash-attention/releases) 下载对应版本安装。
 
-### 使用魔搭社区（可跳过）
+</details>
 
-如果您在 Hugging Face 模型和数据集的下载中遇到了问题，可以通过下述方法使用魔搭社区。
-
-```bash
-export USE_MODELSCOPE_HUB=1 # Windows 使用 `set USE_MODELSCOPE_HUB=1`
-```
-
-接着即可通过指定模型名称来训练对应的模型。（在[魔搭社区](https://modelscope.cn/models)查看所有可用的模型）
-
-```bash
-CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
-    --model_name_or_path modelscope/Llama-2-7b-ms \
-    ... # 参数同下
-```
-
-LLaMA Board 同样支持魔搭社区的模型和数据集下载。
-
-```bash
-CUDA_VISIBLE_DEVICES=0 USE_MODELSCOPE_HUB=1 python src/train_web.py
-```
-
-### 单 GPU 训练
+### LLaMA Board 可视化界面
 
 > [!IMPORTANT]
-> 如果您使用多张 GPU 训练模型，请移步[多 GPU 分布式训练](#多-gpu-分布式训练)部分。
+> LLaMA Board 可视化界面目前仅支持单 GPU 训练，请使用[命令行接口](#命令行接口)来进行分布式训练。
 
-#### LLaMA Board GUI
+#### 使用本地环境
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 python src/train_web.py
+export CUDA_VISIBLE_DEVICES=0 # Windows 使用 `set CUDA_VISIBLE_DEVICES=0`
+python src/train_web.py # 或 python -m llmtuner.webui.interface
 ```
 
-#### 预训练
-
-```bash
-CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
-    --stage pt \
-    --do_train \
-    --model_name_or_path path_to_llama_model \
-    --dataset wiki_demo \
-    --finetuning_type lora \
-    --lora_target q_proj,v_proj \
-    --output_dir path_to_pt_checkpoint \
-    --overwrite_cache \
-    --per_device_train_batch_size 4 \
-    --gradient_accumulation_steps 4 \
-    --lr_scheduler_type cosine \
-    --logging_steps 10 \
-    --save_steps 1000 \
-    --learning_rate 5e-5 \
-    --num_train_epochs 3.0 \
-    --plot_loss \
-    --fp16
-```
-
-#### 指令监督微调
-
-```bash
-CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
-    --stage sft \
-    --do_train \
-    --model_name_or_path path_to_llama_model \
-    --dataset alpaca_gpt4_zh \
-    --template default \
-    --finetuning_type lora \
-    --lora_target q_proj,v_proj \
-    --output_dir path_to_sft_checkpoint \
-    --overwrite_cache \
-    --per_device_train_batch_size 4 \
-    --gradient_accumulation_steps 4 \
-    --lr_scheduler_type cosine \
-    --logging_steps 10 \
-    --save_steps 1000 \
-    --learning_rate 5e-5 \
-    --num_train_epochs 3.0 \
-    --plot_loss \
-    --fp16
-```
-
-#### 奖励模型训练
-
-```bash
-CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
-    --stage rm \
-    --do_train \
-    --model_name_or_path path_to_llama_model \
-    --adapter_name_or_path path_to_sft_checkpoint \
-    --create_new_adapter \
-    --dataset comparison_gpt4_zh \
-    --template default \
-    --finetuning_type lora \
-    --lora_target q_proj,v_proj \
-    --output_dir path_to_rm_checkpoint \
-    --per_device_train_batch_size 2 \
-    --gradient_accumulation_steps 4 \
-    --lr_scheduler_type cosine \
-    --logging_steps 10 \
-    --save_steps 1000 \
-    --learning_rate 1e-5 \
-    --num_train_epochs 1.0 \
-    --plot_loss \
-    --fp16
-```
-
-#### PPO 训练
-
-```bash
-CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
-    --stage ppo \
-    --do_train \
-    --model_name_or_path path_to_llama_model \
-    --adapter_name_or_path path_to_sft_checkpoint \
-    --create_new_adapter \
-    --dataset alpaca_gpt4_zh \
-    --template default \
-    --finetuning_type lora \
-    --lora_target q_proj,v_proj \
-    --reward_model path_to_rm_checkpoint \
-    --output_dir path_to_ppo_checkpoint \
-    --per_device_train_batch_size 2 \
-    --gradient_accumulation_steps 4 \
-    --lr_scheduler_type cosine \
-    --top_k 0 \
-    --top_p 0.9 \
-    --logging_steps 10 \
-    --save_steps 1000 \
-    --learning_rate 1e-5 \
-    --num_train_epochs 1.0 \
-    --plot_loss \
-    --fp16
-```
-
-> [!TIP]
-> 如果开启了 `--create_new_adapter`，则使用 `--adapter_name_or_path path_to_sft_checkpoint,path_to_ppo_checkpoint` 来进行微调模型的推理。
-
-> [!WARNING]
-> 如果使用 fp16 精度进行 LLaMA-2 模型的 PPO 训练，请使用 `--per_device_train_batch_size=1`。
-
-#### DPO 训练
-
-```bash
-CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
-    --stage dpo \
-    --do_train \
-    --model_name_or_path path_to_llama_model \
-    --adapter_name_or_path path_to_sft_checkpoint \
-    --create_new_adapter \
-    --dataset comparison_gpt4_zh \
-    --template default \
-    --finetuning_type lora \
-    --lora_target q_proj,v_proj \
-    --output_dir path_to_dpo_checkpoint \
-    --per_device_train_batch_size 2 \
-    --gradient_accumulation_steps 4 \
-    --lr_scheduler_type cosine \
-    --logging_steps 10 \
-    --save_steps 1000 \
-    --learning_rate 1e-5 \
-    --num_train_epochs 1.0 \
-    --plot_loss \
-    --fp16
-```
-
-> [!TIP]
-> 如果开启了 `--create_new_adapter`，则使用 `--adapter_name_or_path path_to_sft_checkpoint,path_to_dpo_checkpoint` 来进行微调模型的推理。
-
-### 多 GPU 分布式训练
-
-#### 使用 Huggingface Accelerate
-
-```bash
-accelerate launch --config_file config.yaml src/train_bash.py \
-    --ddp_timeout 180000000 \
-    ... # 参数同上
-```
-
-<details><summary>使用 Accelerate 进行 LoRA 训练的 config.yaml 示例</summary>
-
-```yaml
-compute_environment: LOCAL_MACHINE
-debug: false
-distributed_type: MULTI_GPU
-downcast_bf16: 'no'
-gpu_ids: all
-machine_rank: 0
-main_training_function: main
-mixed_precision: fp16
-num_machines: 1
-num_processes: 4
-rdzv_backend: static
-same_network: true
-tpu_env: []
-tpu_use_cluster: false
-tpu_use_sudo: false
-use_cpu: false
-```
-
-</details>
-
-> [!TIP]
-> 我们推荐使用 Accelerate 进行 LoRA 训练。
-
-#### 使用 DeepSpeed
-
-```bash
-deepspeed --num_gpus 8 src/train_bash.py \
-    --deepspeed ds_config.json \
-    --ddp_timeout 180000000 \
-    ... # 参数同上
-```
-
-<details><summary>使用 DeepSpeed ZeRO-2 进行全参数训练的 ds_config.json 示例</summary>
-
-```json
-{
-  "train_batch_size": "auto",
-  "train_micro_batch_size_per_gpu": "auto",
-  "gradient_accumulation_steps": "auto",
-  "gradient_clipping": "auto",
-  "zero_allow_untested_optimizer": true,
-  "fp16": {
-    "enabled": "auto",
-    "loss_scale": 0,
-    "loss_scale_window": 1000,
-    "initial_scale_power": 16,
-    "hysteresis": 2,
-    "min_loss_scale": 1
-  },
-  "bf16": {
-    "enabled": "auto"
-  },
-  "zero_optimization": {
-    "stage": 2,
-    "allgather_partitions": true,
-    "allgather_bucket_size": 5e8,
-    "overlap_comm": true,
-    "reduce_scatter": true,
-    "reduce_bucket_size": 5e8,
-    "contiguous_gradients": true,
-    "round_robin_gradients": true
-  }
-}
-```
-
-</details>
-
-> [!TIP]
-> 更多训练脚本请查看 [examples](examples)。
-
-### 合并 LoRA 权重并导出模型
-
-```bash
-CUDA_VISIBLE_DEVICES= python src/export_model.py \
-    --model_name_or_path path_to_llama_model \
-    --adapter_name_or_path path_to_checkpoint \
-    --template default \
-    --finetuning_type lora \
-    --export_dir path_to_export \
-    --export_size 2 \
-    --export_legacy_format False
-```
-
-> [!WARNING]
-> 尚不支持量化模型的 LoRA 权重合并及导出。
-
-> [!TIP]
-> 仅使用 `--model_name_or_path path_to_export` 来加载导出后的模型。
-> 
-> 合并 LoRA 权重之后可再次使用 `CUDA_VISIBLE_DEVICES=0`、`--export_quantization_bit 4` 和 `--export_quantization_dataset data/c4_demo.json` 基于 AutoGPTQ 量化模型。
-
-### 使用 OpenAI 风格 API 推理
-
-```bash
-CUDA_VISIBLE_DEVICES=0 API_PORT=8000 python src/api_demo.py \
-    --model_name_or_path path_to_llama_model \
-    --adapter_name_or_path path_to_checkpoint \
-    --template default \
-    --finetuning_type lora
-```
-
-> [!TIP]
-> 关于 API 文档请见 `http://localhost:8000/docs`。
-
-### 使用命令行推理
-
-```bash
-CUDA_VISIBLE_DEVICES=0 python src/cli_demo.py \
-    --model_name_or_path path_to_llama_model \
-    --adapter_name_or_path path_to_checkpoint \
-    --template default \
-    --finetuning_type lora
-```
-
-### 使用浏览器推理
-
-```bash
-CUDA_VISIBLE_DEVICES=0 python src/web_demo.py \
-    --model_name_or_path path_to_llama_model \
-    --adapter_name_or_path path_to_checkpoint \
-    --template default \
-    --finetuning_type lora
-```
-
-### 模型评估
-
-```bash
-CUDA_VISIBLE_DEVICES=0 python src/evaluate.py \
-    --model_name_or_path path_to_llama_model \
-    --adapter_name_or_path path_to_checkpoint \
-    --template vanilla \
-    --finetuning_type lora \
-    --task ceval \
-    --split validation \
-    --lang zh \
-    --n_shot 5 \
-    --batch_size 4
-```
-
-### 模型预测
-
-```bash
-CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
-    --stage sft \
-    --do_predict \
-    --model_name_or_path path_to_llama_model \
-    --adapter_name_or_path path_to_checkpoint \
-    --dataset alpaca_gpt4_zh \
-    --template default \
-    --finetuning_type lora \
-    --output_dir path_to_predict_result \
-    --per_device_eval_batch_size 1 \
-    --max_samples 100 \
-    --predict_with_generate \
-    --fp16
-```
-
-> [!WARNING]
-> 如果使用 fp16 精度进行 LLaMA-2 模型的预测，请使用 `--per_device_eval_batch_size=1`。
-
-> [!TIP]
-> 我们建议在量化模型的预测中使用 `--per_device_eval_batch_size=1` 和 `--max_target_length 128`。
-
-### 使用容器
-
 #### 使用 Docker
 
 ```bash
 docker build -f ./Dockerfile -t llama-factory:latest .
-
 docker run --gpus=all \
     -v ./hf_cache:/root/.cache/huggingface/ \
     -v ./data:/app/data \
@@ -683,14 +352,46 @@ docker run --gpus=all \
 docker compose -f ./docker-compose.yml up -d
 ```
 
-> [!TIP]
-> 数据卷详情：
-> * hf_cache：使用宿主机的 Hugging Face 缓存文件夹，允许更改为新的目录。
-> * data：宿主机中存放数据集的文件夹路径。
-> * output：将导出目录设置为该路径后，即可在宿主机中访问导出后的模型。
+<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。
+- data：宿主机中存放数据集的文件夹路径。
+- output：将导出目录设置为该路径后，即可在宿主机中访问导出后的模型。
+
+</details>
+
+### 命令行接口
+
+使用方法请参考 [examples/README_zh.md](examples/README_zh.md)。
+
+使用 `python src/train_bash.py -h` 查看参数文档。
+
+### 使用 OpenAI 风格 API 和 vLLM 部署
+
+```bash
+CUDA_VISIBLE_DEVICES=0,1 API_PORT=8000 python src/api_demo.py \
+    --model_name_or_path mistralai/Mistral-7B-Instruct-v0.2 \
+    --template mistral \
+    --infer_backend vllm \
+    --vllm_enforce_eager
+```
+
+### 使用魔搭社区
+
+如果您在 Hugging Face 模型和数据集的下载中遇到了问题，可以通过下述方法使用魔搭社区。
+
+```bash
+export USE_MODELSCOPE_HUB=1 # Windows 使用 `set USE_MODELSCOPE_HUB=1`
+```
+
+将 `--model_name_or_path` 设置为模型 ID 来加载对应的模型。在[魔搭社区](https://modelscope.cn/models)查看所有可用的模型，例如 `modelscope/Llama-2-7b-ms`。
 
 ## 使用了 LLaMA Factory 的项目
 
+如果您有项目希望添加至上述列表，请通过邮件联系或者创建一个 PR。
+
+<details><summary>点击显示</summary>
+
 1. Wang et al. ESRL: Efficient Sampling-based Reinforcement Learning for Sequence Generation. 2023. [[arxiv]](https://arxiv.org/abs/2308.02223)
 1. Yu et al. Open, Closed, or Small Language Models for Text Classification? 2023. [[arxiv]](https://arxiv.org/abs/2308.10092)
 1. Wang et al. UbiPhysio: Support Daily Functioning, Fitness, and Rehabilitation with Action Understanding and Feedback in Natural Language. 2023. [[arxiv]](https://arxiv.org/abs/2308.10526)
@@ -712,6 +413,7 @@ docker compose -f ./docker-compose.yml up -d
 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. 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. **[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。
@@ -719,8 +421,7 @@ docker compose -f ./docker-compose.yml up -d
 1. **[CareGPT](https://github.com/WangRongsheng/CareGPT)**: 医疗大模型项目 CareGPT，基于 LLaMA2-7B 和 Baichuan-13B 在中文医疗数据上微调而得。
 1. **[MachineMindset](https://github.com/PKU-YuanGroup/Machine-Mindset/)**：MBTI性格大模型项目，根据数据集与训练方式让任意 LLM 拥有 16 个不同的性格类型。
 
-> [!TIP]
-> 如果您有项目希望添加至上述列表，请通过邮件联系或者创建一个 PR。
+</details>
 
 ## 协议
 
@@ -744,7 +445,7 @@ docker compose -f ./docker-compose.yml up -d
 
 ## 致谢
 
-本项目受益于 [PEFT](https://github.com/huggingface/peft)、[QLoRA](https://github.com/artidoro/qlora) 和 [FastChat](https://github.com/lm-sys/FastChat)，感谢以上诸位作者的付出。
+本项目受益于 [PEFT](https://github.com/huggingface/peft)、[TRL](https://github.com/huggingface/trl)、[QLoRA](https://github.com/artidoro/qlora) 和 [FastChat](https://github.com/lm-sys/FastChat)，感谢以上诸位作者的付出。
 
 ## Star History
 

data/README.md

@@ -34,6 +34,8 @@ If you are using a custom dataset, please provide your dataset definition in the
 
 Given above, you can use the custom dataset via specifying `--dataset dataset_name`.
 
+----
+
 Currently we support dataset in **alpaca** or **sharegpt** format, the dataset in alpaca format should follow the below format:
 
 ```json
@@ -84,6 +86,10 @@ For the preference datasets, the `response` column should be a string list whose
 }
 ```
 
+Remember to set `"ranking": true` for the preference datasets.
+
+----
+
 The dataset in sharegpt format should follow the below format:
 
 ```json

data/README_zh.md

@@ -34,6 +34,8 @@
 
 添加后可通过指定 `--dataset 数据集名称` 参数使用自定义数据集。
 
+----
+
 该项目目前支持两种格式的数据集：**alpaca** 和 **sharegpt**，其中 alpaca 格式的数据集按照以下方式组织：
 
 ```json
@@ -84,6 +86,10 @@
 }
 ```
 
+添加偏好数据集需要额外指定 `"ranking": true`。
+
+----
+
 而 sharegpt 格式的数据集按照以下方式组织：
 
 ```json

examples/README.md

@@ -0,0 +1,43 @@
+We provide diverse examples about fine-tuning LLMs.
+
+```
+examples/
+├── lora_single_gpu/
+│   ├── pretrain.sh: Do pre-training
+│   ├── sft.sh: Do supervised fine-tuning
+│   ├── reward.sh: Do reward modeling
+│   ├── ppo.sh: Do PPO training
+│   ├── dpo.sh: Do DPO training
+│   ├── orpo.sh: Do ORPO training
+│   ├── prepare.sh: Save tokenized dataset
+│   └── predict.sh: Do batch predict
+├── qlora_single_gpu/
+│   ├── bitsandbytes.sh: Fine-tune 4/8-bit BNB models
+│   ├── gptq.sh: Fine-tune 4/8-bit GPTQ models
+│   ├── awq.sh: Fine-tune 4-bit AWQ models
+│   └── aqlm.sh: Fine-tune 2-bit AQLM models
+├── lora_multi_gpu/
+│   ├── single_node.sh: Fine-tune model with Accelerate on single node
+│   └── multi_node.sh: Fine-tune model with Accelerate on multiple nodes
+├── full_multi_gpu/
+│   ├── single_node.sh: Fine-tune model with DeepSpeed on single node
+│   └── multi_node.sh: Fine-tune model with DeepSpeed on multiple nodes
+├── merge_lora/
+│   ├── merge.sh: Merge LoRA weights into the pre-trained models
+│   └── quantize.sh: Quantize fine-tuned model with AutoGPTQ
+├── inference/
+│   ├── cli_demo.sh: Launch a command line interface
+│   ├── api_demo.sh: Launch an OpenAI-style API
+│   ├── web_demo.sh: Launch a web interface
+│   └── evaluate.sh: Evaluate model on the MMLU benchmark
+└── extras/
+    ├── galore/
+    │   └── sft.sh: Fine-tune model with GaLore
+    ├── loraplus/
+    │   └── sft.sh: Fine-tune model with LoRA+
+    ├── llama_pro/
+    │   ├── expand.sh: Expand layers in the model
+    │   └── sft.sh: Fine-tune expanded model
+    └── fsdp_qlora/
+        └── sft.sh: Fine-tune quantized model with FSDP
+```

examples/README_zh.md

@@ -0,0 +1,43 @@
+我们提供了多样化的示例脚本。
+
+```
+examples/
+├── lora_single_gpu/
+│   ├── pretrain.sh: 进行预训练
+│   ├── sft.sh: 进行指令监督微调
+│   ├── reward.sh: 进行奖励模型训练
+│   ├── ppo.sh: 进行 PPO 训练
+│   ├── dpo.sh: 进行 DPO 训练
+│   ├── orpo.sh: 进行 ORPO 训练
+│   ├── prepare.sh: 保存预处理后的数据集
+│   └── predict.sh: 进行批量预测
+├── qlora_single_gpu/
+│   ├── bitsandbytes.sh: 微调 4/8 比特 BNB 模型
+│   ├── gptq.sh: 微调 4/8 比特 GPTQ 模型
+│   ├── awq.sh: 微调 4 比特 AWQ 模型
+│   └── aqlm.sh: 微调 2 比特 AQLM 模型
+├── lora_multi_gpu/
+│   ├── single_node.sh: 使用 Accelerate 进行单节点训练
+│   └── multi_node.sh: 使用 Accelerate 进行多节点训练
+├── full_multi_gpu/
+│   ├── single_node.sh: 使用 DeepSpeed 进行单节点训练
+│   └── multi_node.sh: 使用 DeepSpeed 进行多节点训练
+├── merge_lora/
+│   ├── merge.sh: 将 LoRA 权重合并到预训练模型中
+│   └── quantize.sh: 使用 AutoGPTQ 量化模型
+├── inference/
+│   ├── cli_demo.sh: 启动命令行推理接口
+│   ├── api_demo.sh: 启动 OpenAI 风格 API
+│   ├── web_demo.sh: 启动浏览器推理接口
+│   └── evaluate.sh: 在 MMLU 数据集上评测模型
+└── extras/
+    ├── galore/
+    │   └── sft.sh: 使用 GaLore 训练模型
+    ├── loraplus/
+    │   └── sft.sh: 使用 LoRA+ 训练模型
+    ├── llama_pro/
+    │   ├── expand.sh: 扩展模型中的层
+    │   └── sft.sh: 训练扩展后的模型
+    └── fsdp_qlora/
+        └── sft.sh: 使用 FSDP 微调量化模型
+```

examples/accelerate/fsdp_config.yaml

@@ -15,8 +15,8 @@ fsdp_config:
 machine_rank: 0
 main_training_function: main
 mixed_precision: fp16
-num_machines: 1
-num_processes: 2
+num_machines: 1 # the number of nodes
+num_processes: 2 # the number of GPUs in all nodes
 rdzv_backend: static
 same_network: true
 tpu_env: []

examples/accelerate/master_config.yaml

@@ -8,8 +8,8 @@ main_process_ip: 192.168.0.1
 main_process_port: 29555
 main_training_function: main
 mixed_precision: fp16
-num_machines: 2
-num_processes: 16
+num_machines: 2 # the number of nodes
+num_processes: 16 # the number of GPUs in all nodes
 rdzv_backend: static
 same_network: true
 tpu_env: []

examples/accelerate/single_config.yaml

@@ -6,8 +6,8 @@ gpu_ids: all
 machine_rank: 0
 main_training_function: main
 mixed_precision: fp16
-num_machines: 1
-num_processes: 4
+num_machines: 1 # the number of nodes
+num_processes: 4 # the number of GPUs in all nodes
 rdzv_backend: static
 same_network: true
 tpu_env: []

examples/accelerate/slave_config.yaml

@@ -8,8 +8,8 @@ main_process_ip: 192.168.0.1
 main_process_port: 29555
 main_training_function: main
 mixed_precision: fp16
-num_machines: 2
-num_processes: 16
+num_machines: 2 # the number of nodes
+num_processes: 16 # the number of GPUs in all nodes
 rdzv_backend: static
 same_network: true
 tpu_env: []

examples/extras/fsdp_qlora/sft.sh

@@ -1,25 +1,31 @@
 #!/bin/bash
 
+pip install "transformers>=4.39.1"
+pip install "accelerate>=0.28.0"
+pip install "bitsandbytes>=0.43.0"
+
 CUDA_VISIBLE_DEVICES=0,1 accelerate launch \
-    --config_file ../accelerate/fsdp_config.yaml \
-    ../../src/train_bash.py \
+    --config_file ../../accelerate/fsdp_config.yaml \
+    ../../../src/train_bash.py \
     --stage sft \
     --do_train \
     --model_name_or_path meta-llama/Llama-2-70b-hf \
     --dataset alpaca_gpt4_en,glaive_toolcall \
-    --dataset_dir ../../data \
+    --dataset_dir ../../../data \
     --template default \
     --finetuning_type lora \
     --lora_target q_proj,v_proj \
-    --output_dir ../../saves/LLaMA2-70B/lora/sft \
+    --output_dir ../../../saves/LLaMA2-70B/lora/sft \
     --overwrite_cache \
     --overwrite_output_dir \
     --cutoff_len 1024 \
+    --preprocessing_num_workers 16 \
     --per_device_train_batch_size 1 \
     --per_device_eval_batch_size 1 \
-    --gradient_accumulation_steps 8 \
+    --gradient_accumulation_steps 4 \
     --lr_scheduler_type cosine \
     --logging_steps 10 \
+    --warmup_steps 20 \
     --save_steps 100 \
     --eval_steps 100 \
     --evaluation_strategy steps \
@@ -28,6 +34,7 @@ CUDA_VISIBLE_DEVICES=0,1 accelerate launch \
     --num_train_epochs 3.0 \
     --max_samples 3000 \
     --val_size 0.1 \
+    --ddp_timeout 180000000 \
     --quantization_bit 4 \
     --plot_loss \
     --fp16

examples/fsdp_qlora/README.md

@@ -1,5 +0,0 @@
-```bash
-pip install "transformers>=4.39.1"
-pip install "accelerate>=0.28.0"
-pip install "bitsandbytes>=0.43.0"
-```

examples/full_multi_gpu/multi_node.sh

@@ -33,6 +33,6 @@ python -m torch.distributed.run \
     --num_train_epochs 3.0 \
     --max_samples 3000 \
     --val_size 0.1 \
-    --ddp_timeout 1800000 \
+    --ddp_timeout 180000000 \
     --plot_loss \
     --fp16

examples/full_multi_gpu/single_node.sh

@@ -27,6 +27,6 @@ deepspeed --num_gpus 4 ../../src/train_bash.py \
     --num_train_epochs 3.0 \
     --max_samples 3000 \
     --val_size 0.1 \
-    --ddp_timeout 1800000 \
+    --ddp_timeout 180000000 \
     --plot_loss \
     --fp16

examples/inference/api_demo.sh

@@ -0,0 +1,7 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 API_PORT=8000 python ../../src/api_demo.py \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --adapter_name_or_path ../../saves/LLaMA2-7B/lora/sft \
+    --template default \
+    --finetuning_type lora

examples/inference/cli_demo.sh

@@ -0,0 +1,7 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../src/cli_demo.py \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --adapter_name_or_path ../../saves/LLaMA2-7B/lora/sft \
+    --template default \
+    --finetuning_type lora

examples/inference/evaluate.sh

@@ -0,0 +1,12 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../src/evaluate.py \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --adapter_name_or_path ../../saves/LLaMA2-7B/lora/sft \
+    --template vanilla \
+    --finetuning_type lora \
+    --task mmlu \
+    --split test \
+    --lang en \
+    --n_shot 5 \
+    --batch_size 4

examples/inference/web_demo.sh

@@ -0,0 +1,7 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../src/web_demo.py \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --adapter_name_or_path ../../saves/LLaMA2-7B/lora/sft \
+    --template default \
+    --finetuning_type lora

examples/lora_multi_gpu/multi_node.sh

@@ -30,6 +30,6 @@ CUDA_VISIBLE_DEVICES=0,1,2,3 accelerate launch \
     --num_train_epochs 3.0 \
     --max_samples 3000 \
     --val_size 0.1 \
-    --ddp_timeout 1800000 \
+    --ddp_timeout 180000000 \
     --plot_loss \
     --fp16

examples/lora_multi_gpu/single_node.sh

@@ -30,6 +30,6 @@ CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 accelerate launch \
     --num_train_epochs 3.0 \
     --max_samples 3000 \
     --val_size 0.1 \
-    --ddp_timeout 1800000 \
+    --ddp_timeout 180000000 \
     --plot_loss \
     --fp16

examples/lora_single_gpu/README.md

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

examples/lora_single_gpu/dpo.sh

@@ -6,7 +6,7 @@ CUDA_VISIBLE_DEVICES=0 python ../../src/train_bash.py \
     --model_name_or_path meta-llama/Llama-2-7b-hf \
     --adapter_name_or_path ../../saves/LLaMA2-7B/lora/sft \
     --create_new_adapter \
-    --dataset comparison_gpt4_en \
+    --dataset orca_rlhf \
     --dataset_dir ../../data \
     --template default \
     --finetuning_type lora \

examples/lora_single_gpu/orpo.sh

@@ -0,0 +1,32 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../src/train_bash.py \
+    --stage orpo \
+    --do_train \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --dataset orca_rlhf \
+    --dataset_dir ../../data \
+    --template default \
+    --finetuning_type lora \
+    --lora_target q_proj,v_proj \
+    --output_dir ../../saves/LLaMA2-7B/lora/orpo \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --preprocessing_num_workers 16 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 8 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --warmup_steps 20 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --load_best_model_at_end \
+    --learning_rate 1e-5 \
+    --num_train_epochs 1.0 \
+    --max_samples 1000 \
+    --val_size 0.1 \
+    --plot_loss \
+    --fp16

examples/lora_single_gpu/prepare.sh

@@ -0,0 +1,18 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES= python ../../src/train_bash.py \
+    --stage sft \
+    --do_train \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --dataset alpaca_gpt4_en,glaive_toolcall \
+    --dataset_dir ../../data \
+    --template default \
+    --finetuning_type lora \
+    --lora_target q_proj,v_proj \
+    --output_dir ../../saves/LLaMA2-7B/lora/sft \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --preprocessing_num_workers 16 \
+    --max_samples 3000 \
+    --tokenized_path ../../saves/datasets/sft

examples/lora_single_gpu/reward.sh

@@ -6,7 +6,7 @@ CUDA_VISIBLE_DEVICES=0 python ../../src/train_bash.py \
     --model_name_or_path meta-llama/Llama-2-7b-hf \
     --adapter_name_or_path ../../saves/LLaMA2-7B/lora/sft \
     --create_new_adapter \
-    --dataset comparison_gpt4_en \
+    --dataset orca_rlhf \
     --dataset_dir ../../data \
     --template default \
     --finetuning_type lora \

examples/merge_lora/README.md

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

examples/merge_lora/merge.sh

@@ -1,6 +1,7 @@
 #!/bin/bash
+# DO NOT use quantized model or quantization_bit when merging lora weights
 
-CUDA_VISIBLE_DEVICES=0 python ../../src/export_model.py \
+CUDA_VISIBLE_DEVICES= python ../../src/export_model.py \
     --model_name_or_path meta-llama/Llama-2-7b-hf \
     --adapter_name_or_path ../../saves/LLaMA2-7B/lora/sft \
     --template default \

requirements.txt

@@ -4,7 +4,7 @@ datasets>=2.14.3
 accelerate>=0.27.2
 peft>=0.10.0
 trl>=0.8.1
-gradio>=3.38.0,<4.0.0
+gradio>=4.0.0,<=4.21.0
 scipy
 einops
 sentencepiece
@@ -15,4 +15,3 @@ fastapi
 sse-starlette
 matplotlib
 fire
-galore-torch

scripts/cal_lr.py

@@ -15,7 +15,7 @@ from transformers import DataCollatorForLanguageModeling, DataCollatorForSeq2Seq
 from llmtuner.data import get_dataset
 from llmtuner.extras.constants import IGNORE_INDEX
 from llmtuner.hparams import get_train_args
-from llmtuner.model import load_model_and_tokenizer
+from llmtuner.model import load_tokenizer
 
 
 BASE_LR = 3e-4  # 1.5e-4 for 30B-70B models
@@ -32,7 +32,7 @@ def calculate_lr(
     cutoff_len: Optional[int] = 1024,  # i.e. maximum input length during training
     is_mistral: Optional[bool] = False,  # mistral model uses a smaller learning rate,
 ):
-    model_args, data_args, training_args, finetuning_args, _ = get_train_args(
+    model_args, data_args, training_args, _, _ = get_train_args(
         dict(
             stage=stage,
             model_name_or_path=model_name_or_path,
@@ -44,8 +44,8 @@ def calculate_lr(
             overwrite_cache=True,
         )
     )
-    _, tokenizer = load_model_and_tokenizer(model_args, finetuning_args, is_trainable=False, add_valuehead=False)
-    trainset = get_dataset(tokenizer, model_args, data_args, training_args, stage=stage)
+    tokenizer = load_tokenizer(model_args)
+    trainset = get_dataset(tokenizer, model_args, data_args, training_args, stage)
     if stage == "pt":
         data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)
     elif stage == "sft":

scripts/length_cdf.py

@@ -10,7 +10,7 @@ from tqdm import tqdm
 
 from llmtuner.data import get_dataset
 from llmtuner.hparams import get_train_args
-from llmtuner.model import load_model_and_tokenizer
+from llmtuner.model import load_tokenizer
 
 
 def length_cdf(
@@ -20,7 +20,7 @@ def length_cdf(
     template: Optional[str] = "default",
     interval: Optional[int] = 1000,
 ):
-    model_args, data_args, training_args, finetuning_args, _ = get_train_args(
+    model_args, data_args, training_args, _, _ = get_train_args(
         dict(
             stage="sft",
             model_name_or_path=model_name_or_path,
@@ -32,7 +32,7 @@ def length_cdf(
             overwrite_cache=True,
         )
     )
-    _, tokenizer = load_model_and_tokenizer(model_args, finetuning_args, is_trainable=False, add_valuehead=False)
+    tokenizer = load_tokenizer(model_args)
     trainset = get_dataset(tokenizer, model_args, data_args, training_args, stage="sft")
     total_num = len(trainset)
     length_dict = defaultdict(int)

scripts/llamafy_internlm2.py

@@ -1,114 +0,0 @@
-# coding=utf-8
-# Converts the InternLM2 model in the same format as LLaMA2.
-# Usage: python llamafy_internlm2.py --input_dir input --output_dir output
-# Warning: We have found that the converted model cannot infer correctly. It will be fixed later.
-
-import json
-import os
-from collections import OrderedDict
-from typing import Any, Dict, Optional
-
-import fire
-import torch
-from safetensors.torch import save_file
-from tqdm import tqdm
-from transformers.modeling_utils import (
-    SAFE_WEIGHTS_INDEX_NAME,
-    SAFE_WEIGHTS_NAME,
-    WEIGHTS_INDEX_NAME,
-    WEIGHTS_NAME,
-    shard_checkpoint,
-)
-
-
-CONFIG_NAME = "config.json"
-
-
-def save_weight(input_dir: str, output_dir: str, shard_size: str, save_safetensors: bool):
-    with open(os.path.join(input_dir, CONFIG_NAME), "r", encoding="utf-8") as f:
-        internlm2_config_dict: Dict[str, Any] = json.load(f)
-
-    internlm2_state_dict: Dict[str, torch.Tensor] = OrderedDict()
-    for filepath in tqdm(os.listdir(input_dir), desc="Load weights"):
-        if os.path.isfile(os.path.join(input_dir, filepath)) and filepath.endswith(".bin"):
-            shard_weight = torch.load(os.path.join(input_dir, filepath), map_location="cpu")
-            internlm2_state_dict.update(shard_weight)
-
-    llama2_state_dict: Dict[str, torch.Tensor] = OrderedDict()
-    for key, value in tqdm(internlm2_state_dict.items(), desc="Convert format"):
-        if "output" in key:
-            llama2_state_dict[key.replace("output", "lm_head")] = value
-        elif "tok_embeddings" in key:
-            llama2_state_dict[key.replace("tok_embeddings", "embed_tokens")] = value
-        elif "wqkv" in key:
-            num_q_heads = internlm2_config_dict["num_attention_heads"]
-            num_kv_heads = internlm2_config_dict["num_key_value_heads"]
-            q_size = value.size(0) // (num_q_heads + 2 * num_kv_heads) * num_q_heads
-            kv_size = value.size(0) // (num_q_heads + 2 * num_kv_heads) * num_kv_heads
-            llama2_state_dict[key.replace("attention.wqkv", "self_attn.q_proj")] = value[:q_size, ...]
-            llama2_state_dict[key.replace("attention.wqkv", "self_attn.k_proj")] = value[
-                q_size : q_size + kv_size, ...
-            ]
-            llama2_state_dict[key.replace("attention.wqkv", "self_attn.v_proj")] = value[q_size + kv_size :, ...]
-        elif "wo" in key:
-            llama2_state_dict[key.replace("attention.wo", "self_attn.o_proj")] = value
-        elif "attention_norm" in key:
-            llama2_state_dict[key.replace("attention_norm", "input_layernorm")] = value
-        elif "ffn_norm" in key:
-            llama2_state_dict[key.replace("ffn_norm", "post_attention_layernorm")] = value
-        elif "w1" in key:
-            llama2_state_dict[key.replace("feed_forward.w1", "mlp.gate_proj")] = value
-        elif "w2" in key:
-            llama2_state_dict[key.replace("feed_forward.w2", "mlp.down_proj")] = value
-        elif "w3" in key:
-            llama2_state_dict[key.replace("feed_forward.w3", "mlp.up_proj")] = value
-        else:
-            llama2_state_dict[key] = value
-
-    weights_name = SAFE_WEIGHTS_NAME if save_safetensors else WEIGHTS_NAME
-    shards, index = shard_checkpoint(llama2_state_dict, max_shard_size=shard_size, weights_name=weights_name)
-
-    for shard_file, shard in tqdm(shards.items(), desc="Save weights"):
-        if save_safetensors:
-            save_file(shard, os.path.join(output_dir, shard_file), metadata={"format": "pt"})
-        else:
-            torch.save(shard, os.path.join(output_dir, shard_file))
-
-    if index is None:
-        print("Model weights saved in {}".format(os.path.join(output_dir, WEIGHTS_NAME)))
-    else:
-        index_name = SAFE_WEIGHTS_INDEX_NAME if save_safetensors else WEIGHTS_INDEX_NAME
-        with open(os.path.join(output_dir, index_name), "w", encoding="utf-8") as f:
-            json.dump(index, f, indent=2, sort_keys=True)
-        print("Model weights saved in {}".format(output_dir))
-
-
-def save_config(input_dir: str, output_dir: str):
-    with open(os.path.join(input_dir, CONFIG_NAME), "r", encoding="utf-8") as f:
-        llama2_config_dict: Dict[str, Any] = json.load(f)
-
-    llama2_config_dict["architectures"] = ["LlamaForCausalLM"]
-    llama2_config_dict.pop("auto_map", None)
-    llama2_config_dict.pop("bias", None)
-    llama2_config_dict.pop("rope_scaling", None)
-    llama2_config_dict["model_type"] = "llama"
-
-    with open(os.path.join(output_dir, CONFIG_NAME), "w", encoding="utf-8") as f:
-        json.dump(llama2_config_dict, f, indent=2)
-    print("Model config saved in {}".format(os.path.join(output_dir, CONFIG_NAME)))
-
-
-def llamafy_internlm2(
-    input_dir: str, output_dir: str, shard_size: Optional[str] = "2GB", save_safetensors: Optional[bool] = False
-):
-    try:
-        os.makedirs(output_dir, exist_ok=False)
-    except Exception as e:
-        raise print("Output dir already exists", e)
-
-    save_weight(input_dir, output_dir, shard_size, save_safetensors)
-    save_config(input_dir, output_dir)
-
-
-if __name__ == "__main__":
-    fire.Fire(llamafy_internlm2)

setup.py

@@ -20,15 +20,17 @@ def get_requires():
 
 
 extra_require = {
-    "deepspeed": ["deepspeed"],
+    "deepspeed": ["deepspeed>=0.10.0"],
     "metrics": ["nltk", "jieba", "rouge-chinese"],
     "unsloth": ["torch==2.2.0", "unsloth[cu121-ampere-torch220]"],
+    "galore": ["galore-torch"],
     "vllm": ["vllm>=0.3.3"],
     "bitsandbytes": ["bitsandbytes>=0.39.0"],
     "gptq": ["optimum>=1.16.0", "auto-gptq>=0.5.0"],
     "awq": ["autoawq"],
     "aqlm": ["aqlm[gpu]>=1.1.0"],
     "qwen": ["tiktoken", "transformers_stream_generator"],
+    "modelscope": ["modelscope"],
     "quality": ["ruff"],
 }
 

src/evaluate.py

@@ -2,8 +2,7 @@ from llmtuner import Evaluator
 
 
 def main():
-    evaluator = Evaluator()
-    evaluator.eval()
+    Evaluator().eval()
 
 
 if __name__ == "__main__":

src/llmtuner/__init__.py

@@ -7,5 +7,5 @@ from .train import export_model, run_exp
 from .webui import create_ui, create_web_demo
 
 
-__version__ = "0.6.1"
+__version__ = "0.6.2"
 __all__ = ["create_app", "ChatModel", "Evaluator", "export_model", "run_exp", "create_ui", "create_web_demo"]

src/llmtuner/api/app.py

@@ -108,12 +108,18 @@ def create_app(chat_model: "ChatModel") -> "FastAPI":
             elif i % 2 == 1 and message.role not in [Role.ASSISTANT, Role.FUNCTION]:
                 raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="Invalid role")
 
+            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)
+                input_messages.append({"role": role_mapping[Role.FUNCTION], "content": content})
+            else:
                 input_messages.append({"role": role_mapping[message.role], "content": message.content})
 
         tool_list = request.tools
         if isinstance(tool_list, list) and len(tool_list):
             try:
-                tools = json.dumps([tool["function"] for tool in tool_list], ensure_ascii=False)
+                tools = json.dumps([dictify(tool.function) for tool in tool_list], ensure_ascii=False)
             except Exception:
                 raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="Invalid tools")
         else:

src/llmtuner/api/protocol.py

@@ -1,6 +1,6 @@
 import time
 from enum import Enum, unique
-from typing import List, Optional
+from typing import Any, Dict, List, Optional
 
 from pydantic import BaseModel, Field
 from typing_extensions import Literal
@@ -39,6 +39,17 @@ class Function(BaseModel):
     arguments: str
 
 
+class FunctionDefinition(BaseModel):
+    name: str
+    description: str
+    parameters: Dict[str, Any]
+
+
+class FunctionAvailable(BaseModel):
+    type: Literal["function", "code_interpreter"] = "function"
+    function: Optional[FunctionDefinition] = None
+
+
 class FunctionCall(BaseModel):
     id: Literal["call_default"] = "call_default"
     type: Literal["function"] = "function"
@@ -47,7 +58,8 @@ class FunctionCall(BaseModel):
 
 class ChatMessage(BaseModel):
     role: Role
-    content: str
+    content: Optional[str] = None
+    tool_calls: Optional[List[FunctionCall]] = None
 
 
 class ChatCompletionMessage(BaseModel):
@@ -59,7 +71,7 @@ class ChatCompletionMessage(BaseModel):
 class ChatCompletionRequest(BaseModel):
     model: str
     messages: List[ChatMessage]
-    tools: list = []
+    tools: Optional[List[FunctionAvailable]] = None
     do_sample: bool = True
     temperature: Optional[float] = None
     top_p: Optional[float] = None

src/llmtuner/chat/hf_engine.py

@@ -9,7 +9,7 @@ from transformers import GenerationConfig, TextIteratorStreamer
 
 from ..data import get_template_and_fix_tokenizer
 from ..extras.misc import get_logits_processor
-from ..model import load_model_and_tokenizer
+from ..model import load_model, load_tokenizer
 from .base_engine import BaseEngine, Response
 
 
@@ -30,11 +30,12 @@ class HuggingfaceEngine(BaseEngine):
         generating_args: "GeneratingArguments",
     ) -> None:
         self.can_generate = finetuning_args.stage == "sft"
-        self.model, self.tokenizer = load_model_and_tokenizer(
-            model_args, finetuning_args, is_trainable=False, add_valuehead=(not self.can_generate)
-        )
+        self.tokenizer = load_tokenizer(model_args)
         self.tokenizer.padding_side = "left" if self.can_generate else "right"
         self.template = get_template_and_fix_tokenizer(self.tokenizer, data_args.template)
+        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()
 
     @staticmethod

src/llmtuner/data/__init__.py

@@ -1,6 +1,15 @@
+from .collator import PairwiseDataCollatorWithPadding
 from .loader import get_dataset
 from .template import Template, get_template_and_fix_tokenizer, templates
 from .utils import Role, split_dataset
 
 
-__all__ = ["get_dataset", "Template", "get_template_and_fix_tokenizer", "templates", "Role", "split_dataset"]
+__all__ = [
+    "PairwiseDataCollatorWithPadding",
+    "get_dataset",
+    "Template",
+    "get_template_and_fix_tokenizer",
+    "templates",
+    "Role",
+    "split_dataset",
+]

src/llmtuner/data/collator.py

@@ -6,12 +6,15 @@ from transformers import DataCollatorForSeq2Seq
 
 
 @dataclass
-class DPODataCollatorWithPadding(DataCollatorForSeq2Seq):
+class PairwiseDataCollatorWithPadding(DataCollatorForSeq2Seq):
     r"""
     Data collator for pairwise data.
     """
 
     def _pad_labels(self, batch: torch.Tensor, positions: List[Tuple[int, int]]) -> torch.Tensor:
+        r"""
+        Masks out the input ids except for the responses.
+        """
         padded_labels = []
         for feature, (prompt_len, answer_len) in zip(batch, positions):
             if self.tokenizer.padding_side == "left":
@@ -43,12 +46,6 @@ class DPODataCollatorWithPadding(DataCollatorForSeq2Seq):
                 )
                 label_positions.append((prompt_len, answer_len))
 
-        batch = self.tokenizer.pad(
-            concatenated_features,
-            padding=self.padding,
-            max_length=self.max_length,
-            pad_to_multiple_of=self.pad_to_multiple_of,
-            return_tensors=self.return_tensors,
-        )
+        batch = super().__call__(concatenated_features)
         batch["labels"] = self._pad_labels(batch["input_ids"], label_positions)
         return batch

src/llmtuner/data/loader.py

@@ -6,6 +6,7 @@ from datasets import load_dataset, load_from_disk
 
 from ..extras.constants import FILEEXT2TYPE
 from ..extras.logging import get_logger
+from ..extras.misc import has_tokenized_data
 from .aligner import align_dataset
 from .parser import get_dataset_list
 from .preprocess import get_preprocess_and_print_func
@@ -80,7 +81,9 @@ def load_single_dataset(
                 cache_dir=cache_dir,
                 token=model_args.ms_hub_token,
                 use_streaming=(data_args.streaming and (dataset_attr.load_from != "file")),
-            ).to_hf_dataset()
+            )
+            if isinstance(dataset, MsDataset):
+                dataset = dataset.to_hf_dataset()
         except ImportError:
             raise ImportError("Please install modelscope via `pip install modelscope -U`")
     else:
@@ -117,17 +120,17 @@ def get_dataset(
     data_args: "DataArguments",
     training_args: "Seq2SeqTrainingArguments",
     stage: Literal["pt", "sft", "rm", "ppo"],
-    # split: Optional[str] = "train", # TODO: add split
 ) -> Union["Dataset", "IterableDataset"]:
     template = get_template_and_fix_tokenizer(tokenizer, data_args.template)
     if data_args.train_on_prompt and template.efficient_eos:
         raise ValueError("Current template does not support `train_on_prompt`.")
 
-    # Load from cache
-    if data_args.cache_path is not None:
-        if os.path.exists(data_args.cache_path):
+    # Load tokenized 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.cache_path)
+            dataset = load_from_disk(data_args.tokenized_path)
+            logger.info("Loaded tokenized dataset from {}.".format(data_args.tokenized_path))
             if data_args.streaming:
                 dataset = dataset.to_iterable_dataset()
             return dataset
@@ -138,6 +141,9 @@ def get_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)
 
@@ -156,10 +162,13 @@ def get_dataset(
 
         dataset = dataset.map(preprocess_func, batched=True, remove_columns=column_names, **kwargs)
 
-        if data_args.cache_path is not None and not os.path.exists(data_args.cache_path):
+        if data_args.tokenized_path is not None:
             if training_args.should_save:
-                dataset.save_to_disk(data_args.cache_path)
-                logger.info("Dataset cache saved at {}.".format(data_args.cache_path))
+                dataset.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))
+
+            exit(0)
 
         if training_args.should_log:
             try:

src/llmtuner/data/parser.py

@@ -53,12 +53,19 @@ class DatasetAttr:
 
 
 def get_dataset_list(data_args: "DataArguments") -> List["DatasetAttr"]:
-    dataset_names = [ds.strip() for ds in data_args.dataset.split(",")] if data_args.dataset is not None else []
+    if data_args.dataset is not None:
+        dataset_names = [ds.strip() for ds in data_args.dataset.split(",")]
+    else:
+        dataset_names = []
+
+    if data_args.dataset_dir == "ONLINE":
+        dataset_info = None
+    else:
         try:
             with open(os.path.join(data_args.dataset_dir, DATA_CONFIG), "r") as f:
                 dataset_info = json.load(f)
         except Exception as err:
-        if data_args.dataset is not None:
+            if len(dataset_names) != 0:
                 raise ValueError(
                     "Cannot open {} due to {}.".format(os.path.join(data_args.dataset_dir, DATA_CONFIG), str(err))
                 )
@@ -69,6 +76,12 @@ def get_dataset_list(data_args: "DataArguments") -> List["DatasetAttr"]:
 
     dataset_list: List[DatasetAttr] = []
     for name in dataset_names:
+        if dataset_info is None:
+            load_from = "ms_hub" if use_modelscope() else "hf_hub"
+            dataset_attr = DatasetAttr(load_from, dataset_name=name)
+            dataset_list.append(dataset_attr)
+            continue
+
         if name not in dataset_info:
             raise ValueError("Undefined dataset {} in {}.".format(name, DATA_CONFIG))
 

src/llmtuner/data/preprocess.py

@@ -23,16 +23,18 @@ def preprocess_pretrain_dataset(
 ) -> 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"]]
-    if not data_args.packing:
-        return tokenizer(text_examples, add_special_tokens=False, max_length=data_args.cutoff_len)
 
+    if not data_args.packing:
+        if data_args.template == "gemma":
+            text_examples = [tokenizer.bos_token + example for example in text_examples]
+
+        result = tokenizer(text_examples, add_special_tokens=False, max_length=data_args.cutoff_len)
+    else:
         tokenized_examples = tokenizer(text_examples, add_special_tokens=False)
         concatenated_examples = {k: list(chain(*tokenized_examples[k])) for k in tokenized_examples.keys()}
         total_length = len(concatenated_examples[list(concatenated_examples.keys())[0]])
         block_size = data_args.cutoff_len
-    # we drop the small remainder, and if the total_length < block_size, we exclude this batch
         total_length = (total_length // block_size) * block_size
-    # split by chunks of cutoff_len
         result = {
             k: [t[i : i + block_size] for i in range(0, total_length, block_size)]
             for k, t in concatenated_examples.items()

src/llmtuner/data/template.py

@@ -414,7 +414,7 @@ _register_template(
 
 _register_template(
     name="baichuan",
-    format_user=StringFormatter(slots=["<reserved_102>{{content}}<reserved_103>"]),
+    format_user=StringFormatter(slots=[{"token": "<reserved_102>"}, "{{content}}", {"token": "<reserved_103>"}]),
     efficient_eos=True,
 )
 
@@ -441,6 +441,18 @@ _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."
+    ),
+    efficient_eos=True,
+)
+
+
 _register_template(
     name="chatglm2",
     format_user=StringFormatter(slots=["[Round {{idx}}]\n\n问：{{content}}\n\n答："]),

src/llmtuner/data/utils.py

@@ -44,7 +44,7 @@ def checksum(data_files: List[str], file_sha1: Optional[str] = None) -> None:
 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 - max_target_len
+    max_source_len = max_len - min(max_target_len, target_len)
     return max_source_len, max_target_len
 
 

src/llmtuner/eval/evaluator.py

@@ -14,16 +14,17 @@ from transformers.utils import cached_file
 from ..data import get_template_and_fix_tokenizer
 from ..extras.constants import CHOICES, SUBJECTS
 from ..hparams import get_eval_args
-from ..model import load_model_and_tokenizer
+from ..model import load_model, load_tokenizer
 from .template import get_eval_template
 
 
 class Evaluator:
     def __init__(self, args: Optional[Dict[str, Any]] = None) -> None:
         self.model_args, self.data_args, self.eval_args, finetuning_args = get_eval_args(args)
-        self.model, self.tokenizer = load_model_and_tokenizer(self.model_args, finetuning_args)
+        self.tokenizer = load_tokenizer(self.model_args)
         self.tokenizer.padding_side = "right"  # avoid overflow issue in batched inference for llama2
         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

src/llmtuner/extras/callbacks.py

@@ -134,6 +134,7 @@ class LogCallback(TrainerCallback):
             eval_loss=state.log_history[-1].get("eval_loss", None),
             predict_loss=state.log_history[-1].get("predict_loss", None),
             reward=state.log_history[-1].get("reward", None),
+            accuracy=state.log_history[-1].get("rewards/accuracies", None),
             learning_rate=state.log_history[-1].get("learning_rate", None),
             epoch=state.log_history[-1].get("epoch", None),
             percentage=round(self.cur_steps / self.max_steps * 100, 2) if self.max_steps != 0 else 100,

src/llmtuner/extras/constants.py

@@ -39,9 +39,12 @@ TRAINING_STAGES = {
     "Reward Modeling": "rm",
     "PPO": "ppo",
     "DPO": "dpo",
+    "ORPO": "orpo",
     "Pre-Training": "pt",
 }
 
+STAGES_USE_PAIR_DATA = ["rm", "dpo", "orpo"]
+
 V_HEAD_WEIGHTS_NAME = "value_head.bin"
 
 V_HEAD_SAFE_WEIGHTS_NAME = "value_head.safetensors"
@@ -167,6 +170,19 @@ register_model_group(
 )
 
 
+register_model_group(
+    models={
+        "Breeze-7B": {
+            DownloadSource.DEFAULT: "MediaTek-Research/Breeze-7B-Base-v1_0",
+        },
+        "Breeze-7B-Chat": {
+            DownloadSource.DEFAULT: "MediaTek-Research/Breeze-7B-Instruct-v1_0",
+        },
+    },
+    template="breeze",
+)
+
+
 register_model_group(
     models={
         "ChatGLM2-6B-Chat": {
@@ -460,14 +476,18 @@ register_model_group(
 
 register_model_group(
     models={
-        "Mistral-7B": {
+        "Mistral-7B-v0.1": {
             DownloadSource.DEFAULT: "mistralai/Mistral-7B-v0.1",
             DownloadSource.MODELSCOPE: "AI-ModelScope/Mistral-7B-v0.1",
         },
-        "Mistral-7B-Chat": {
+        "Mistral-7B-v0.1-Chat": {
             DownloadSource.DEFAULT: "mistralai/Mistral-7B-Instruct-v0.1",
             DownloadSource.MODELSCOPE: "AI-ModelScope/Mistral-7B-Instruct-v0.1",
         },
+        "Mistral-7B-v0.2": {
+            DownloadSource.DEFAULT: "alpindale/Mistral-7B-v0.2-hf",
+            DownloadSource.MODELSCOPE: "AI-ModelScope/Mistral-7B-v0.2-hf",
+        },
         "Mistral-7B-v0.2-Chat": {
             DownloadSource.DEFAULT: "mistralai/Mistral-7B-Instruct-v0.2",
             DownloadSource.MODELSCOPE: "AI-ModelScope/Mistral-7B-Instruct-v0.2",
@@ -656,10 +676,18 @@ register_model_group(
             DownloadSource.DEFAULT: "Qwen/Qwen1.5-14B",
             DownloadSource.MODELSCOPE: "qwen/Qwen1.5-14B",
         },
+        "Qwen1.5-32B": {
+            DownloadSource.DEFAULT: "Qwen/Qwen1.5-32B",
+            DownloadSource.MODELSCOPE: "qwen/Qwen1.5-32B",
+        },
         "Qwen1.5-72B": {
             DownloadSource.DEFAULT: "Qwen/Qwen1.5-72B",
             DownloadSource.MODELSCOPE: "qwen/Qwen1.5-72B",
         },
+        "Qwen1.5-MoE-A2.7B": {
+            DownloadSource.DEFAULT: "Qwen/Qwen1.5-MoE-A2.7B",
+            DownloadSource.MODELSCOPE: "qwen/Qwen1.5-MoE-A2.7B",
+        },
         "Qwen1.5-0.5B-Chat": {
             DownloadSource.DEFAULT: "Qwen/Qwen1.5-0.5B-Chat",
             DownloadSource.MODELSCOPE: "qwen/Qwen1.5-0.5B-Chat",
@@ -680,10 +708,18 @@ register_model_group(
             DownloadSource.DEFAULT: "Qwen/Qwen1.5-14B-Chat",
             DownloadSource.MODELSCOPE: "qwen/Qwen1.5-14B-Chat",
         },
+        "Qwen1.5-32B-Chat": {
+            DownloadSource.DEFAULT: "Qwen/Qwen1.5-32B-Chat",
+            DownloadSource.MODELSCOPE: "qwen/Qwen1.5-32B-Chat",
+        },
         "Qwen1.5-72B-Chat": {
             DownloadSource.DEFAULT: "Qwen/Qwen1.5-72B-Chat",
             DownloadSource.MODELSCOPE: "qwen/Qwen1.5-72B-Chat",
         },
+        "Qwen1.5-MoE-A2.7B-Chat": {
+            DownloadSource.DEFAULT: "Qwen/Qwen1.5-MoE-A2.7B-Chat",
+            DownloadSource.MODELSCOPE: "qwen/Qwen1.5-MoE-A2.7B-Chat",
+        },
         "Qwen1.5-0.5B-int8-Chat": {
             DownloadSource.DEFAULT: "Qwen/Qwen1.5-0.5B-Chat-GPTQ-Int8",
             DownloadSource.MODELSCOPE: "qwen/Qwen1.5-0.5B-Chat-GPTQ-Int8",
@@ -724,6 +760,10 @@ register_model_group(
             DownloadSource.DEFAULT: "Qwen/Qwen1.5-14B-Chat-AWQ",
             DownloadSource.MODELSCOPE: "qwen/Qwen1.5-14B-Chat-AWQ",
         },
+        "Qwen1.5-32B-int4-Chat": {
+            DownloadSource.DEFAULT: "Qwen/Qwen1.5-32B-Chat-AWQ",
+            DownloadSource.MODELSCOPE: "qwen/Qwen1.5-32B-Chat-AWQ",
+        },
         "Qwen1.5-72B-int8-Chat": {
             DownloadSource.DEFAULT: "Qwen/Qwen1.5-72B-Chat-GPTQ-Int8",
             DownloadSource.MODELSCOPE: "qwen/Qwen1.5-72B-Chat-GPTQ-Int8",
@@ -732,6 +772,10 @@ register_model_group(
             DownloadSource.DEFAULT: "Qwen/Qwen1.5-72B-Chat-AWQ",
             DownloadSource.MODELSCOPE: "qwen/Qwen1.5-72B-Chat-AWQ",
         },
+        "Qwen1.5-MoE-A2.7B-int4-Chat": {
+            DownloadSource.DEFAULT: "Qwen/Qwen1.5-MoE-A2.7B-Chat-GPTQ-Int4",
+            DownloadSource.MODELSCOPE: "qwen/Qwen1.5-MoE-A2.7B-Chat-GPTQ-Int4",
+        },
     },
     template="qwen",
 )

src/llmtuner/extras/misc.py

@@ -66,6 +66,7 @@ def check_dependencies() -> None:
         require_version("accelerate>=0.27.2", "To fix: pip install accelerate>=0.27.2")
         require_version("peft>=0.10.0", "To fix: pip install peft>=0.10.0")
         require_version("trl>=0.8.1", "To fix: pip install trl>=0.8.1")
+        require_version("gradio>=4.0.0,<=4.21.0", "To fix: pip install gradio==4.21.0")
 
 
 def count_parameters(model: torch.nn.Module) -> Tuple[int, int]:
@@ -192,6 +193,13 @@ def infer_optim_dtype(model_dtype: torch.dtype) -> torch.dtype:
         return torch.float32
 
 
+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 torch_gc() -> None:
     r"""
     Collects GPU memory.
@@ -202,17 +210,15 @@ def torch_gc() -> None:
         torch.cuda.ipc_collect()
 
 
-def try_download_model_from_ms(model_args: "ModelArguments") -> None:
+def try_download_model_from_ms(model_args: "ModelArguments") -> str:
     if not use_modelscope() or os.path.exists(model_args.model_name_or_path):
-        return
+        return model_args.model_name_or_path
 
     try:
         from modelscope import snapshot_download
 
         revision = "master" if model_args.model_revision == "main" else model_args.model_revision
-        model_args.model_name_or_path = snapshot_download(
-            model_args.model_name_or_path, revision=revision, cache_dir=model_args.cache_dir
-        )
+        return snapshot_download(model_args.model_name_or_path, revision=revision, cache_dir=model_args.cache_dir)
     except ImportError:
         raise ImportError("Please install modelscope via `pip install modelscope -U`")
 

src/llmtuner/extras/patches/llama_patch.py

@@ -193,6 +193,6 @@ def llama_flash_attn_forward(
 
 
 def apply_llama_patch() -> None:
-    require_version("transformers==4.39.1", "To fix: pip install transformers==4.39.1")
+    require_version("transformers==4.39.3", "To fix: pip install transformers==4.39.3")
     LlamaAttention.forward = llama_torch_attn_forward
     LlamaFlashAttention2.forward = llama_flash_attn_forward

src/llmtuner/extras/patches/mixtral_patch.py

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

src/llmtuner/extras/ploting.py

@@ -52,6 +52,6 @@ def plot_loss(save_dictionary: os.PathLike, keys: List[str] = ["loss"]) -> None:
         plt.xlabel("step")
         plt.ylabel(key)
         plt.legend()
-        figure_path = os.path.join(save_dictionary, "training_{}.png".format(key.replace(os.path.sep, "_")))
+        figure_path = os.path.join(save_dictionary, "training_{}.png".format(key.replace("/", "_")))
         plt.savefig(figure_path, format="png", dpi=100)
         print("Figure saved at:", figure_path)

src/llmtuner/hparams/data_args.py

@@ -84,9 +84,9 @@ class DataArguments:
             "help": "Whether or not to pack the sequences in training. Will automatically enable in pre-training."
         },
     )
-    cache_path: Optional[str] = field(
+    tokenized_path: Optional[str] = field(
         default=None,
-        metadata={"help": "Path to save or load the pre-processed datasets."},
+        metadata={"help": "Path to save or load the tokenized datasets."},
     )
 
     def __post_init__(self):

src/llmtuner/hparams/finetuning_args.py

@@ -110,6 +110,10 @@ class RLHFArguments:
         default=0.0,
         metadata={"help": "The supervised fine-tuning loss coefficient in DPO training."},
     )
+    orpo_beta: float = field(
+        default=0.1,
+        metadata={"help": "The beta (lambda) parameter in ORPO loss representing the weight of the SFT loss."},
+    )
     ppo_buffer_size: int = field(
         default=1,
         metadata={"help": "The number of mini-batches to make experience buffer in a PPO optimization step."},
@@ -209,7 +213,7 @@ class FinetuningArguments(FreezeArguments, LoraArguments, RLHFArguments, GaloreA
         default=False,
         metadata={"help": "Whether or not to train model in purely bf16 precision (without AMP)."},
     )
-    stage: Literal["pt", "sft", "rm", "ppo", "dpo"] = field(
+    stage: Literal["pt", "sft", "rm", "ppo", "dpo", "orpo"] = field(
         default="sft",
         metadata={"help": "Which stage will be performed in training."},
     )

src/llmtuner/hparams/model_args.py

@@ -53,6 +53,10 @@ class ModelArguments:
         default=True,
         metadata={"help": "Whether or not to use double quantization in int4 training."},
     )
+    quantization_device_map: Optional[Literal["auto"]] = field(
+        default=None,
+        metadata={"help": "Device map used for loading the 4-bit quantized model, needs bitsandbytes>=0.43.0."},
+    )
     rope_scaling: Optional[Literal["linear", "dynamic"]] = field(
         default=None,
         metadata={"help": "Which scaling strategy should be adopted for the RoPE embeddings."},
@@ -69,6 +73,10 @@ class ModelArguments:
         default=False,
         metadata={"help": "Whether or not to use unsloth's optimization for the LoRA training."},
     )
+    moe_aux_loss_coef: Optional[float] = field(
+        default=None,
+        metadata={"help": "Coefficient of the auxiliary router loss in mixture-of-experts model."},
+    )
     disable_gradient_checkpointing: bool = field(
         default=False,
         metadata={"help": "Whether or not to disable gradient checkpointing."},

src/llmtuner/hparams/parser.py

@@ -120,7 +120,7 @@ def get_train_args(args: Optional[Dict[str, Any]] = None) -> _TRAIN_CLS:
 
     if (
         finetuning_args.stage == "ppo"
-        and training_args.report_to is not None
+        and training_args.report_to
         and training_args.report_to[0] not in ["wandb", "tensorboard"]
     ):
         raise ValueError("PPO only accepts wandb or tensorboard logger.")

src/llmtuner/model/__init__.py

@@ -1,10 +1,9 @@
-from .loader import load_model, load_model_and_tokenizer, load_tokenizer
+from .loader import load_model, load_tokenizer
 from .utils import find_all_linear_modules, load_valuehead_params
 
 
 __all__ = [
     "load_model",
-    "load_model_and_tokenizer",
     "load_tokenizer",
     "load_valuehead_params",
     "find_all_linear_modules",

src/llmtuner/model/loader.py

@@ -1,4 +1,4 @@
-from typing import TYPE_CHECKING, Any, Dict, Tuple
+from typing import TYPE_CHECKING, Any, Dict
 
 from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer
 from trl import AutoModelForCausalLMWithValueHead
@@ -20,6 +20,7 @@ logger = get_logger(__name__)
 
 
 def _get_init_kwargs(model_args: "ModelArguments") -> Dict[str, Any]:
+    model_args.model_name_or_path = try_download_model_from_ms(model_args)
     return {
         "trust_remote_code": True,
         "cache_dir": model_args.cache_dir,
@@ -34,9 +35,7 @@ def load_tokenizer(model_args: "ModelArguments") -> "PreTrainedTokenizer":
 
     Note: including inplace operation of model_args.
     """
-    try_download_model_from_ms(model_args)
     init_kwargs = _get_init_kwargs(model_args)
-
     tokenizer = AutoTokenizer.from_pretrained(
         model_args.model_name_or_path,
         use_fast=model_args.use_fast_tokenizer,
@@ -110,9 +109,6 @@ def load_model(
     if not is_trainable:
         model.requires_grad_(False)
         model.eval()
-        for param in model.parameters():
-            if param.device.type == "cuda":
-                param.data = param.data.to(model_args.compute_dtype)
     else:
         model.train()
 
@@ -134,17 +130,3 @@ def load_model(
             )
 
     return model
-
-
-def load_model_and_tokenizer(
-    model_args: "ModelArguments",
-    finetuning_args: "FinetuningArguments",
-    is_trainable: bool = False,
-    add_valuehead: bool = False,
-) -> Tuple["PreTrainedModel", "PreTrainedTokenizer"]:
-    r"""
-    Loads pretrained model and tokenizer.
-    """
-    tokenizer = load_tokenizer(model_args)
-    model = load_model(tokenizer, model_args, finetuning_args, is_trainable, add_valuehead)
-    return model, tokenizer

src/llmtuner/model/patcher.py

@@ -17,8 +17,7 @@ from ..extras.logging import get_logger
 from ..extras.misc import get_current_device, infer_optim_dtype
 from ..extras.packages import is_flash_attn2_available
 from ..extras.patches.llama_patch import apply_llama_patch
-from ..extras.patches.mixtral_patch import patch_mixtral_replace_moe_impl
-from .utils import QuantizationMethod
+from .utils import QuantizationMethod, add_z3_leaf_module
 
 
 if TYPE_CHECKING:
@@ -32,47 +31,6 @@ logger = get_logger(__name__)
 SUPPORTED_CLASS_FOR_S2ATTN = ["llama"]
 
 
-def _noisy_mean_initialization(embed_weight: torch.Tensor, num_new_tokens: int):
-    embedding_dim = embed_weight.size(1)
-    avg_weight = embed_weight[:-num_new_tokens].mean(dim=0, keepdim=True)
-    noise_weight = torch.empty_like(embed_weight[-num_new_tokens:])
-    noise_weight.normal_(mean=0, std=(1.0 / math.sqrt(embedding_dim)))
-    embed_weight[-num_new_tokens:] = avg_weight + noise_weight
-
-
-def _resize_embedding_layer(model: "PreTrainedModel", tokenizer: "PreTrainedTokenizer") -> None:
-    r"""
-    Resize token embeddings.
-    """
-    if is_deepspeed_zero3_enabled():
-        import deepspeed  # type: ignore
-
-        params = [model.get_input_embeddings().weight]
-        if model.get_output_embeddings() is not None and not model.config.tie_word_embeddings:
-            params.append(model.get_output_embeddings().weight)
-
-        context_maybe_zero3 = deepspeed.zero.GatheredParameters(params, modifier_rank=0)
-    else:
-        context_maybe_zero3 = nullcontext()
-
-    with context_maybe_zero3:
-        current_embedding_size = model.get_input_embeddings().weight.size(0)
-
-    if len(tokenizer) > current_embedding_size:
-        if not isinstance(model.get_output_embeddings(), torch.nn.Linear):
-            logger.warning("Current model does not support resizing token embeddings.")
-            return
-
-        model.resize_token_embeddings(len(tokenizer), pad_to_multiple_of=64)
-        with context_maybe_zero3:
-            new_embedding_size = model.get_input_embeddings().weight.size(0)
-            num_new_tokens = new_embedding_size - current_embedding_size
-            _noisy_mean_initialization(model.get_input_embeddings().weight.data, num_new_tokens)
-            _noisy_mean_initialization(model.get_output_embeddings().weight.data, num_new_tokens)
-
-        logger.info("Resized token embeddings from {} to {}.".format(current_embedding_size, new_embedding_size))
-
-
 def _get_quantization_dataset(tokenizer: "PreTrainedTokenizer", model_args: "ModelArguments") -> List[str]:
     r"""
     Inspired by: https://github.com/huggingface/optimum/blob/v1.16.0/optimum/gptq/data.py#L133
@@ -180,8 +138,12 @@ def _configure_quantization(
         quant_method = quantization_config.get("quant_method", "")
 
         if quant_method == QuantizationMethod.GPTQ:
+            require_version("auto_gptq>=0.5.0", "To fix: pip install auto_gptq>=0.5.0")
             quantization_config["use_exllama"] = False  # disable exllama
 
+        if quant_method == QuantizationMethod.AWQ:
+            require_version("autoawq", "To fix: pip install autoawq")
+
         if quant_method == QuantizationMethod.AQLM:
             require_version("transformers>=4.39.0", "To fix: pip install transformers>=4.39.0")
             require_version("aqlm>=1.1.0", "To fix: pip install aqlm[gpu]>=1.1.0")
@@ -208,11 +170,6 @@ def _configure_quantization(
         logger.info("Quantizing model to {} bit.".format(model_args.export_quantization_bit))
 
     elif model_args.quantization_bit is not None:  # bnb
-        if is_deepspeed_zero3_enabled():
-            require_version("transformers>=4.39.0", "To fix: pip install transformers>=4.39.0")
-            require_version("accelerate>=0.28.0", "To fix: pip install accelerate>=0.28.0")
-            require_version("bitsandbytes>=0.43.0", "To fix: pip install bitsandbytes>=0.43.0")
-
         if model_args.quantization_bit == 8:
             require_version("bitsandbytes>=0.37.0", "To fix: pip install bitsandbytes>=0.37.0")
             init_kwargs["quantization_config"] = BitsAndBytesConfig(load_in_8bit=True)
@@ -227,10 +184,66 @@ def _configure_quantization(
                 bnb_4bit_quant_storage=model_args.compute_dtype,  # crucial for fsdp qlora
             )
 
+        if is_deepspeed_zero3_enabled() or model_args.quantization_device_map == "auto":
+            if model_args.quantization_bit != 4:
+                raise ValueError("Only 4-bit quantized model can use auto device map.")
+
+            require_version("transformers>=4.39.0", "To fix: pip install transformers>=4.39.0")
+            require_version("accelerate>=0.28.0", "To fix: pip install accelerate>=0.28.0")
+            require_version("bitsandbytes>=0.43.0", "To fix: pip install bitsandbytes>=0.43.0")
+        else:
             init_kwargs["device_map"] = {"": get_current_device()}
+
         logger.info("Quantizing model to {} bit.".format(model_args.quantization_bit))
 
 
+def _noisy_mean_initialization(embed_weight: torch.Tensor, num_new_tokens: int):
+    embedding_dim = embed_weight.size(1)
+    avg_weight = embed_weight[:-num_new_tokens].mean(dim=0, keepdim=True)
+    noise_weight = torch.empty_like(embed_weight[-num_new_tokens:])
+    noise_weight.normal_(mean=0, std=(1.0 / math.sqrt(embedding_dim)))
+    embed_weight[-num_new_tokens:] = avg_weight + noise_weight
+
+
+def _resize_embedding_layer(model: "PreTrainedModel", tokenizer: "PreTrainedTokenizer") -> None:
+    r"""
+    Resize token embeddings.
+    """
+    if is_deepspeed_zero3_enabled():
+        import deepspeed  # type: ignore
+
+        params = [model.get_input_embeddings().weight]
+        if model.get_output_embeddings() is not None and not model.config.tie_word_embeddings:
+            params.append(model.get_output_embeddings().weight)
+
+        context_maybe_zero3 = deepspeed.zero.GatheredParameters(params, modifier_rank=0)
+    else:
+        context_maybe_zero3 = nullcontext()
+
+    with context_maybe_zero3:
+        current_embedding_size = model.get_input_embeddings().weight.size(0)
+
+    if len(tokenizer) > current_embedding_size:
+        if not isinstance(model.get_output_embeddings(), torch.nn.Linear):
+            logger.warning("Current model does not support resizing token embeddings.")
+            return
+
+        model.resize_token_embeddings(len(tokenizer), pad_to_multiple_of=64)
+        with context_maybe_zero3:
+            new_embedding_size = model.get_input_embeddings().weight.size(0)
+            num_new_tokens = new_embedding_size - current_embedding_size
+            _noisy_mean_initialization(model.get_input_embeddings().weight.data, num_new_tokens)
+            _noisy_mean_initialization(model.get_output_embeddings().weight.data, num_new_tokens)
+
+        logger.info("Resized token embeddings from {} to {}.".format(current_embedding_size, new_embedding_size))
+
+
+def _fp32_forward_post_hook(
+    module: "torch.nn.Module", args: Tuple["torch.Tensor"], output: "torch.Tensor"
+) -> "torch.Tensor":
+    return output.to(torch.float32)
+
+
 def _prepare_model_for_training(
     model: "PreTrainedModel", model_args: "ModelArguments", output_layer_name: str = "lm_head"
 ) -> None:
@@ -259,14 +272,10 @@ def _prepare_model_for_training(
             logger.info("Gradient checkpointing enabled.")
 
     if hasattr(model, output_layer_name) and model_args.upcast_lmhead_output:
-
-        def fp32_forward_post_hook(module: torch.nn.Module, args: Tuple[torch.Tensor], output: torch.Tensor):
-            return output.to(torch.float32)
-
         logger.info("Upcasting lm_head outputs in float32.")
         output_layer = getattr(model, output_layer_name)
         if isinstance(output_layer, torch.nn.Linear) and output_layer.weight.dtype != torch.float32:
-            output_layer.register_forward_hook(fp32_forward_post_hook)
+            output_layer.register_forward_hook(_fp32_forward_post_hook)
 
 
 def patch_tokenizer(tokenizer: "PreTrainedTokenizer") -> None:
@@ -284,11 +293,6 @@ def patch_config(
     if model_args.compute_dtype is None:  # priority: bf16 > fp16 > fp32
         model_args.compute_dtype = infer_optim_dtype(model_dtype=getattr(config, "torch_dtype", None))
 
-    if getattr(config, "model_type", None) == "qwen":
-        setattr(config, "use_flash_attn", model_args.flash_attn)
-        for dtype_name, dtype in [("fp16", torch.float16), ("bf16", torch.bfloat16), ("fp32", torch.float32)]:
-            setattr(config, dtype_name, model_args.compute_dtype == dtype)
-
     _configure_attn_implementation(config, model_args, init_kwargs)
     _configure_rope(config, model_args, is_trainable)
     _configure_longlora(config, model_args, is_trainable)
@@ -298,11 +302,28 @@ def patch_config(
         setattr(config, "use_cache", True)
         logger.info("Using KV cache for faster generation.")
 
+    if model_args.moe_aux_loss_coef is not None:
+        if getattr(config, "model_type", None) in ["mixtral", "qwen2_moe"]:
+            setattr(config, "router_aux_loss_coef", model_args.moe_aux_loss_coef)
+        elif getattr(config, "model_type", None) == "deepseek":
+            setattr(config, "aux_loss_alpha", model_args.moe_aux_loss_coef)
+
+    if getattr(config, "model_type", None) == "qwen":
+        setattr(config, "use_flash_attn", model_args.flash_attn)
+        for dtype_name, dtype in [("fp16", torch.float16), ("bf16", torch.bfloat16), ("fp32", torch.float32)]:
+            setattr(config, dtype_name, model_args.compute_dtype == dtype)
+
+    if getattr(config, "model_type", None) == "qwen2" and is_trainable and model_args.flash_attn:
+        setattr(config, "use_cache", False)  # qwen2 does not support use_cache when using flashattn
+
+    if getattr(config, "model_type", None) == "qwen2_moe" and is_trainable:
+        setattr(config, "output_router_logits", True)
+
     init_kwargs["torch_dtype"] = model_args.compute_dtype
     if not is_deepspeed_zero3_enabled():
         init_kwargs["low_cpu_mem_usage"] = model_args.low_cpu_mem_usage
         if init_kwargs["low_cpu_mem_usage"]:
-            if "device_map" not in init_kwargs:  # quant models cannot use auto device map
+            if "device_map" not in init_kwargs:
                 init_kwargs["device_map"] = model_args.device_map or {"": get_current_device()}
 
             if init_kwargs["device_map"] == "auto":
@@ -312,10 +333,18 @@ def patch_config(
 def patch_model(
     model: "PreTrainedModel", tokenizer: "PreTrainedTokenizer", model_args: "ModelArguments", is_trainable: bool
 ) -> None:
+    gen_config = model.generation_config  # check and fix generation config
+    if not gen_config.do_sample and (
+        (gen_config.temperature is not None and gen_config.temperature != 1.0)
+        or (gen_config.top_p is not None and gen_config.top_p != 1.0)
+        or (gen_config.typical_p is not None and gen_config.typical_p != 1.0)
+    ):
+        gen_config.do_sample = True
+
     if "GenerationMixin" not in str(model.generate.__func__):
         model.generate = MethodType(PreTrainedModel.generate, model)
 
-    if getattr(model.config, "model_type", None) == "chatglm":
+    if is_trainable and getattr(model.config, "model_type", None) == "chatglm":
         setattr(model, "lm_head", model.transformer.output_layer)
         setattr(model, "_keys_to_ignore_on_save", ["lm_head.weight"])
 
@@ -325,15 +354,15 @@ def patch_model(
     if is_trainable:
         _prepare_model_for_training(model, model_args)
 
-    if getattr(model.config, "model_type", None) == "mixtral" and is_deepspeed_zero3_enabled():
-        require_version("deepspeed>=0.13.0", "To fix: pip install deepspeed>=0.13.0")
-        from deepspeed.utils import set_z3_leaf_modules  # type: ignore
+    if getattr(model.config, "model_type", None) == "mixtral":
         from transformers.models.mixtral.modeling_mixtral import MixtralSparseMoeBlock
 
-        set_z3_leaf_modules(model, [MixtralSparseMoeBlock])
+        add_z3_leaf_module(model, MixtralSparseMoeBlock)
 
-        if is_trainable:
-            patch_mixtral_replace_moe_impl()
+    if getattr(model.config, "model_type", None) == "qwen2moe":
+        from transformers.models.qwen2_moe.modeling_qwen2_moe import Qwen2MoeSparseMoeBlock
+
+        add_z3_leaf_module(model, Qwen2MoeSparseMoeBlock)
 
     try:
         model.add_model_tags(["llama-factory"])

src/llmtuner/model/utils.py

@@ -3,7 +3,9 @@ from typing import TYPE_CHECKING, Dict, List
 
 import torch
 from transformers import PreTrainedModel
+from transformers.integrations import is_deepspeed_zero3_enabled
 from transformers.utils import cached_file
+from transformers.utils.versions import require_version
 
 from ..extras.constants import V_HEAD_SAFE_WEIGHTS_NAME, V_HEAD_WEIGHTS_NAME
 from ..extras.logging import get_logger
@@ -28,11 +30,23 @@ class QuantizationMethod(str, Enum):
     GPTQ = "gptq"
     AWQ = "awq"
     AQLM = "aqlm"
+    QUANTO = "quanto"
+
+
+def add_z3_leaf_module(model: "PreTrainedModel", module: "torch.nn.Module") -> None:
+    r"""
+    Sets module as a leaf module to skip partitioning in deepspeed zero3.
+    """
+    if is_deepspeed_zero3_enabled():
+        require_version("deepspeed>=0.13.0", "To fix: pip install deepspeed>=0.13.0")
+        from deepspeed.utils import set_z3_leaf_modules  # type: ignore
+
+        set_z3_leaf_modules(model, [module])
 
 
 def find_all_linear_modules(model: "PreTrainedModel") -> List[str]:
     r"""
-    Finds all available modules to apply lora.
+    Finds all available modules to apply lora or galore.
     """
     quantization_method = getattr(model, "quantization_method", None)
     if quantization_method is None:

src/llmtuner/train/dpo/trainer.py

@@ -74,7 +74,7 @@ class CustomDPOTrainer(DPOTrainer):
         create_custom_scheduler(self.args, num_training_steps, optimizer)
         return super().create_scheduler(num_training_steps, optimizer)
 
-    def sft_loss(self, chosen_logits: torch.FloatTensor, chosen_labels: torch.LongTensor) -> torch.Tensor:
+    def sft_loss(self, chosen_logits: "torch.FloatTensor", chosen_labels: "torch.LongTensor") -> "torch.Tensor":
         r"""
         Computes supervised cross-entropy loss of given labels under the given logits.
 
@@ -85,18 +85,27 @@ class CustomDPOTrainer(DPOTrainer):
         return -all_logps
 
     def concatenated_forward(
-        self, model: "PreTrainedModel", batch: Dict[str, torch.Tensor]
-    ) -> Tuple[torch.FloatTensor, torch.FloatTensor, torch.FloatTensor, torch.FloatTensor]:
+        self, model: "PreTrainedModel", batch: Dict[str, "torch.Tensor"]
+    ) -> Tuple["torch.Tensor", "torch.Tensor", "torch.Tensor", "torch.Tensor"]:
+        r"""
+        Computes the sum log probabilities of the labels under the given logits if loss_type != IPO.
+
+        Otherwise the average log probabilities.
+        """
         batch_copied = BatchEncoding({k: v.detach().clone() for k, v in batch.items()})  # avoid error
 
-        all_logits = model(
-            input_ids=batch_copied["input_ids"], attention_mask=batch_copied["attention_mask"], return_dict=True
+        all_logits: "torch.Tensor" = model(
+            input_ids=batch_copied["input_ids"],
+            attention_mask=batch_copied["attention_mask"],
+            return_dict=True,
+            use_cache=False,
         ).logits.to(torch.float32)
 
         all_logps = self.get_batch_logps(
-            all_logits,
-            batch["labels"],
-            average_log_prob=False,
+            logits=all_logits,
+            labels=batch_copied["labels"],
+            average_log_prob=(self.loss_type == "ipo"),
+            is_encoder_decoder=self.is_encoder_decoder,
             label_pad_token_id=self.label_pad_token_id,
         )
         batch_size = batch["input_ids"].size(0) // 2
@@ -107,9 +116,9 @@ class CustomDPOTrainer(DPOTrainer):
     def get_batch_loss_metrics(
         self,
         model: "PreTrainedModel",
-        batch: Dict[str, torch.Tensor],
+        batch: Dict[str, "torch.Tensor"],
         train_eval: Literal["train", "eval"] = "train",
-    ) -> Tuple[torch.Tensor, Dict[str, torch.Tensor]]:
+    ) -> Tuple["torch.Tensor", Dict[str, "torch.Tensor"]]:
         r"""
         Computes the DPO loss and other metrics for the given batch of inputs for train or test.
         """
@@ -150,13 +159,13 @@ class CustomDPOTrainer(DPOTrainer):
         reward_accuracies = (chosen_rewards > rejected_rewards).float()
 
         prefix = "eval_" if train_eval == "eval" else ""
-        metrics[f"{prefix}rewards/chosen"] = chosen_rewards.cpu().mean()
-        metrics[f"{prefix}rewards/rejected"] = rejected_rewards.cpu().mean()
-        metrics[f"{prefix}rewards/accuracies"] = reward_accuracies.cpu().mean()
-        metrics[f"{prefix}rewards/margins"] = (chosen_rewards - rejected_rewards).cpu().mean()
-        metrics[f"{prefix}logps/rejected"] = policy_rejected_logps.detach().cpu().mean()
-        metrics[f"{prefix}logps/chosen"] = policy_chosen_logps.detach().cpu().mean()
-        metrics[f"{prefix}logits/rejected"] = policy_rejected_logits.detach().cpu().mean()
-        metrics[f"{prefix}logits/chosen"] = policy_chosen_logits.detach().cpu().mean()
+        metrics["{}rewards/chosen".format(prefix)] = chosen_rewards.cpu().mean()
+        metrics["{}rewards/rejected".format(prefix)] = rejected_rewards.cpu().mean()
+        metrics["{}rewards/accuracies".format(prefix)] = reward_accuracies.cpu().mean()
+        metrics["{}rewards/margins".format(prefix)] = (chosen_rewards - rejected_rewards).cpu().mean()
+        metrics["{}logps/rejected".format(prefix)] = policy_rejected_logps.detach().cpu().mean()
+        metrics["{}logps/chosen".format(prefix)] = policy_chosen_logps.detach().cpu().mean()
+        metrics["{}logits/rejected".format(prefix)] = policy_rejected_logits.detach().cpu().mean()
+        metrics["{}logits/chosen".format(prefix)] = policy_chosen_logits.detach().cpu().mean()
 
         return losses.mean(), metrics

src/llmtuner/train/dpo/workflow.py

@@ -2,13 +2,12 @@
 
 from typing import TYPE_CHECKING, List, Optional
 
-from ...data import get_dataset, split_dataset
+from ...data import PairwiseDataCollatorWithPadding, get_dataset, split_dataset
 from ...extras.constants import IGNORE_INDEX
 from ...extras.ploting import plot_loss
 from ...hparams import ModelArguments
 from ...model import load_model, load_tokenizer
 from ..utils import create_modelcard_and_push, create_ref_model
-from .collator import DPODataCollatorWithPadding
 from .trainer import CustomDPOTrainer
 
 
@@ -29,7 +28,7 @@ def run_dpo(
     dataset = get_dataset(tokenizer, model_args, data_args, training_args, stage="rm")
     model = load_model(tokenizer, model_args, finetuning_args, training_args.do_train)
 
-    data_collator = DPODataCollatorWithPadding(
+    data_collator = PairwiseDataCollatorWithPadding(
         tokenizer=tokenizer,
         pad_to_multiple_of=8,
         label_pad_token_id=IGNORE_INDEX if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id,
@@ -64,7 +63,7 @@ def run_dpo(
         trainer.save_metrics("train", train_result.metrics)
         trainer.save_state()
         if trainer.is_world_process_zero() and finetuning_args.plot_loss:
-            plot_loss(training_args.output_dir, keys=["loss", "eval_loss"])
+            plot_loss(training_args.output_dir, keys=["loss", "eval_loss", "rewards/accuracies"])
 
     # Evaluation
     if training_args.do_eval:

src/llmtuner/train/orpo/__init__.py

@@ -0,0 +1,4 @@
+from .workflow import run_orpo
+
+
+__all__ = ["run_orpo"]

src/llmtuner/train/orpo/trainer.py

@@ -0,0 +1,122 @@
+from collections import defaultdict
+from typing import TYPE_CHECKING, Dict, Literal, Optional, Tuple, Union
+
+import torch
+import torch.nn.functional as F
+from transformers import Trainer
+from trl import DPOTrainer
+from trl.trainer.utils import disable_dropout_in_model
+
+from ...extras.constants import IGNORE_INDEX
+from ..utils import create_custom_optimzer, create_custom_scheduler
+
+
+if TYPE_CHECKING:
+    from transformers import PreTrainedModel
+
+    from ...hparams import FinetuningArguments
+
+
+class CustomORPOTrainer(DPOTrainer):
+    def __init__(
+        self,
+        model: Union["PreTrainedModel", "torch.nn.Module"],
+        finetuning_args: "FinetuningArguments",
+        disable_dropout: bool = True,
+        **kwargs,
+    ):
+        if disable_dropout:
+            disable_dropout_in_model(model)
+
+        self.finetuning_args = finetuning_args
+        self.reference_free = False
+        self.use_dpo_data_collator = True  # hack to avoid warning
+        self.generate_during_eval = False  # disable at evaluation
+        self.label_pad_token_id = IGNORE_INDEX
+        self.padding_value = 0
+        self.is_encoder_decoder = model.config.is_encoder_decoder
+        self.precompute_ref_log_probs = False
+        self._precomputed_train_ref_log_probs = False
+        self._precomputed_eval_ref_log_probs = False
+        self._peft_has_been_casted_to_bf16 = False
+
+        self.beta = finetuning_args.orpo_beta
+        self._stored_metrics = defaultdict(lambda: defaultdict(list))
+
+        Trainer.__init__(self, model=model, **kwargs)
+
+    def create_optimizer(self) -> "torch.optim.Optimizer":
+        if self.optimizer is None:
+            self.optimizer = create_custom_optimzer(self.model, self.args, self.finetuning_args)
+        return super().create_optimizer()
+
+    def create_scheduler(
+        self, num_training_steps: int, optimizer: Optional["torch.optim.Optimizer"] = None
+    ) -> "torch.optim.lr_scheduler.LRScheduler":
+        create_custom_scheduler(self.args, num_training_steps, optimizer)
+        return super().create_scheduler(num_training_steps, optimizer)
+
+    def odds_ratio_loss(self, chosen_logps: "torch.Tensor", rejected_logps: "torch.Tensor") -> "torch.Tensor":
+        r"""
+        Computes ORPO's odds ratio (OR) loss.
+        """
+        log_odds = (chosen_logps - rejected_logps) - (
+            torch.log1p(-torch.exp(chosen_logps)) - torch.log1p(-torch.exp(rejected_logps))
+        )
+        odds_ratio_loss = -F.logsigmoid(log_odds)
+        return odds_ratio_loss
+
+    def concatenated_forward(
+        self, model: "PreTrainedModel", batch: Dict[str, "torch.Tensor"]
+    ) -> Tuple["torch.Tensor", "torch.Tensor", "torch.Tensor", "torch.Tensor"]:
+        r"""
+        Computes the average log probabilities of the labels under the given logits.
+        """
+        all_logits: "torch.Tensor" = model(
+            input_ids=batch["input_ids"], attention_mask=batch["attention_mask"], return_dict=True, use_cache=False
+        ).logits.to(torch.float32)
+
+        all_logps = self.get_batch_logps(
+            logits=all_logits,
+            labels=batch["labels"],
+            average_log_prob=True,
+            is_encoder_decoder=self.is_encoder_decoder,
+            label_pad_token_id=self.label_pad_token_id,
+        )
+        batch_size = batch["input_ids"].size(0) // 2
+        chosen_logps, rejected_logps = all_logps.split(batch_size, dim=0)
+        chosen_logits, rejected_logits = all_logits.split(batch_size, dim=0)
+        return chosen_logps, rejected_logps, chosen_logits, rejected_logits
+
+    def get_batch_loss_metrics(
+        self,
+        model: "PreTrainedModel",
+        batch: Dict[str, "torch.Tensor"],
+        train_eval: Literal["train", "eval"] = "train",
+    ) -> Tuple["torch.Tensor", Dict[str, "torch.Tensor"]]:
+        r"""
+        Computes the ORPO loss and other metrics for the given batch of inputs for train or test.
+        """
+        metrics = {}
+        chosen_logps, rejected_logps, chosen_logits, rejected_logits = self.concatenated_forward(model, batch)
+        sft_loss = -chosen_logps
+        odds_ratio_loss = self.odds_ratio_loss(chosen_logps, rejected_logps)
+        batch_loss = (sft_loss + self.beta * odds_ratio_loss).mean()
+
+        chosen_rewards = self.beta * chosen_logps.detach()
+        rejected_rewards = self.beta * rejected_logps.detach()
+        reward_accuracies = (chosen_rewards > rejected_rewards).float()
+
+        prefix = "eval_" if train_eval == "eval" else ""
+        metrics["{}rewards/chosen".format(prefix)] = chosen_rewards.cpu().mean()
+        metrics["{}rewards/rejected".format(prefix)] = rejected_rewards.cpu().mean()
+        metrics["{}rewards/accuracies".format(prefix)] = reward_accuracies.cpu().mean()
+        metrics["{}rewards/margins".format(prefix)] = (chosen_rewards - rejected_rewards).cpu().mean()
+        metrics["{}logps/rejected".format(prefix)] = rejected_logps.detach().cpu().mean()
+        metrics["{}logps/chosen".format(prefix)] = chosen_logps.detach().cpu().mean()
+        metrics["{}logits/rejected".format(prefix)] = rejected_logits.detach().cpu().mean()
+        metrics["{}logits/chosen".format(prefix)] = chosen_logits.detach().cpu().mean()
+        metrics["{}sft_loss".format(prefix)] = sft_loss.detach().cpu().mean()
+        metrics["{}odds_ratio_loss".format(prefix)] = odds_ratio_loss.detach().cpu().mean()
+
+        return batch_loss, metrics

src/llmtuner/train/orpo/workflow.py

@@ -0,0 +1,68 @@
+# Inspired by: https://github.com/huggingface/trl/blob/main/examples/research_projects/stack_llama_2/scripts/dpo_llama2.py
+
+from typing import TYPE_CHECKING, List, Optional
+
+from ...data import PairwiseDataCollatorWithPadding, get_dataset, split_dataset
+from ...extras.constants import IGNORE_INDEX
+from ...extras.ploting import plot_loss
+from ...hparams import ModelArguments
+from ...model import load_model, load_tokenizer
+from ..utils import create_modelcard_and_push
+from .trainer import CustomORPOTrainer
+
+
+if TYPE_CHECKING:
+    from transformers import Seq2SeqTrainingArguments, TrainerCallback
+
+    from ...hparams import DataArguments, FinetuningArguments
+
+
+def run_orpo(
+    model_args: "ModelArguments",
+    data_args: "DataArguments",
+    training_args: "Seq2SeqTrainingArguments",
+    finetuning_args: "FinetuningArguments",
+    callbacks: Optional[List["TrainerCallback"]] = None,
+):
+    tokenizer = load_tokenizer(model_args)
+    dataset = get_dataset(tokenizer, model_args, data_args, training_args, stage="rm")
+    model = load_model(tokenizer, model_args, finetuning_args, training_args.do_train)
+
+    data_collator = PairwiseDataCollatorWithPadding(
+        tokenizer=tokenizer,
+        pad_to_multiple_of=8,
+        label_pad_token_id=IGNORE_INDEX if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id,
+    )
+
+    # Update arguments
+    training_args.remove_unused_columns = False  # important for pairwise dataset
+
+    # Initialize our Trainer
+    trainer = CustomORPOTrainer(
+        model=model,
+        args=training_args,
+        finetuning_args=finetuning_args,
+        tokenizer=tokenizer,
+        data_collator=data_collator,
+        callbacks=callbacks,
+        **split_dataset(dataset, data_args, training_args),
+    )
+
+    # Training
+    if training_args.do_train:
+        train_result = trainer.train(resume_from_checkpoint=training_args.resume_from_checkpoint)
+        trainer.save_model()
+        trainer.log_metrics("train", train_result.metrics)
+        trainer.save_metrics("train", train_result.metrics)
+        trainer.save_state()
+        if trainer.is_world_process_zero() and finetuning_args.plot_loss:
+            plot_loss(training_args.output_dir, keys=["loss", "eval_loss", "rewards/accuracies", "sft_loss"])
+
+    # Evaluation
+    if training_args.do_eval:
+        metrics = trainer.evaluate(metric_key_prefix="eval")
+        trainer.log_metrics("eval", metrics)
+        trainer.save_metrics("eval", metrics)
+
+    # Create model card
+    create_modelcard_and_push(trainer, model_args, data_args, training_args, finetuning_args)

src/llmtuner/train/ppo/trainer.py

@@ -66,7 +66,7 @@ class CustomPPOTrainer(PPOTrainer, Trainer):
             use_score_norm=finetuning_args.ppo_score_norm,
             whiten_rewards=finetuning_args.ppo_whiten_rewards,
             accelerator_kwargs={"step_scheduler_with_optimizer": False},
-            log_with=training_args.report_to[0] if training_args.report_to is not None else None,
+            log_with=training_args.report_to[0] if training_args.report_to else None,
             project_kwargs={"logging_dir": training_args.logging_dir},
         )
 
@@ -353,7 +353,7 @@ class CustomPPOTrainer(PPOTrainer, Trainer):
         batch = self.prepare_model_inputs(queries, responses)
 
         with torch.cuda.amp.autocast(dtype=self.model_args.compute_dtype):  # support bf16
-            _, _, values = reward_model(**batch, output_hidden_states=True, return_dict=True)
+            _, _, values = reward_model(**batch, output_hidden_states=True, return_dict=True, use_cache=False)
 
         if getattr(unwrapped_model.config, "model_type", None) == "chatglm":  # assume same architecture
             values = torch.transpose(values, 0, 1)

src/llmtuner/train/rm/collator.py

@@ -1,29 +0,0 @@
-from dataclasses import dataclass
-from typing import Any, Dict, Sequence
-
-import torch
-from transformers import DataCollatorWithPadding
-
-
-@dataclass
-class PairwiseDataCollatorWithPadding(DataCollatorWithPadding):
-    r"""
-    Data collator for pairwise data.
-    """
-
-    def __call__(self, features: Sequence[Dict[str, Any]]) -> Dict[str, torch.Tensor]:
-        r"""
-        Pads batched data to the longest sequence in the batch.
-
-        We generate 2 * n examples where the first n examples represent chosen examples and
-        the last n examples represent rejected examples.
-        """
-        features = [
-            {
-                "input_ids": feature["prompt_ids"] + feature[key],
-                "attention_mask": [1] * (len(feature["prompt_ids"]) + len(feature[key])),
-            }
-            for key in ("chosen_ids", "rejected_ids")
-            for feature in features
-        ]
-        return super().__call__(features)

src/llmtuner/train/rm/workflow.py

@@ -2,13 +2,12 @@
 
 from typing import TYPE_CHECKING, List, Optional
 
-from ...data import get_dataset, split_dataset
+from ...data import PairwiseDataCollatorWithPadding, get_dataset, split_dataset
 from ...extras.callbacks import FixValueHeadModelCallback
 from ...extras.misc import fix_valuehead_checkpoint
 from ...extras.ploting import plot_loss
 from ...model import load_model, load_tokenizer
 from ..utils import create_modelcard_and_push
-from .collator import PairwiseDataCollatorWithPadding
 from .metric import compute_accuracy
 from .trainer import PairwiseTrainer
 
@@ -56,7 +55,7 @@ def run_rm(
         trainer.save_metrics("train", train_result.metrics)
         trainer.save_state()
         if trainer.is_world_process_zero() and finetuning_args.plot_loss:
-            plot_loss(training_args.output_dir, keys=["loss", "eval_loss"])
+            plot_loss(training_args.output_dir, keys=["loss", "eval_loss", "eval_accuracy"])
 
     # Evaluation
     if training_args.do_eval:

src/llmtuner/train/tuner.py

@@ -7,8 +7,9 @@ from ..data import get_template_and_fix_tokenizer
 from ..extras.callbacks import LogCallback
 from ..extras.logging import get_logger
 from ..hparams import get_infer_args, get_train_args
-from ..model import load_model_and_tokenizer
+from ..model import load_model, load_tokenizer
 from .dpo import run_dpo
+from .orpo import run_orpo
 from .ppo import run_ppo
 from .pt import run_pt
 from .rm import run_rm
@@ -36,6 +37,8 @@ def run_exp(args: Optional[Dict[str, Any]] = None, callbacks: Optional[List["Tra
         run_ppo(model_args, data_args, training_args, finetuning_args, generating_args, callbacks)
     elif finetuning_args.stage == "dpo":
         run_dpo(model_args, data_args, training_args, finetuning_args, callbacks)
+    elif finetuning_args.stage == "orpo":
+        run_orpo(model_args, data_args, training_args, finetuning_args, callbacks)
     else:
         raise ValueError("Unknown task.")
 
@@ -49,8 +52,9 @@ def export_model(args: Optional[Dict[str, Any]] = None):
     if model_args.adapter_name_or_path is not None and model_args.export_quantization_bit is not None:
         raise ValueError("Please merge adapters before quantizing the model.")
 
-    model, tokenizer = load_model_and_tokenizer(model_args, finetuning_args)
+    tokenizer = load_tokenizer(model_args)
     get_template_and_fix_tokenizer(tokenizer, data_args.template)
+    model = load_model(tokenizer, model_args, finetuning_args)  # must after fixing tokenizer to resize vocab
 
     if getattr(model, "quantization_method", None) and model_args.adapter_name_or_path is not None:
         raise ValueError("Cannot merge adapters to a quantized model.")
@@ -64,14 +68,6 @@ def export_model(args: Optional[Dict[str, Any]] = None):
         for param in model.parameters():
             param.data = param.data.to(output_dtype)
 
-    gen_config = model.generation_config  # check and fix generation config
-    if not gen_config.do_sample and (
-        (gen_config.temperature is not None and gen_config.temperature != 1.0)
-        or (gen_config.top_p is not None and gen_config.top_p != 1.0)
-        or (gen_config.typical_p is not None and gen_config.typical_p != 1.0)
-    ):
-        gen_config.do_sample = True
-
     model.save_pretrained(
         save_directory=model_args.export_dir,
         max_shard_size="{}GB".format(model_args.export_size),

src/llmtuner/train/utils.py

@@ -10,7 +10,7 @@ from transformers.utils.versions import require_version
 from ..extras.logging import get_logger
 from ..extras.packages import is_galore_available
 from ..hparams import FinetuningArguments, ModelArguments
-from ..model import find_all_linear_modules, load_model_and_tokenizer, load_valuehead_params
+from ..model import find_all_linear_modules, load_model, load_tokenizer, load_valuehead_params
 
 
 if is_galore_available():
@@ -87,16 +87,18 @@ def create_ref_model(
         )
         ref_model_args = ModelArguments(**ref_model_args_dict)
         ref_finetuning_args = FinetuningArguments(finetuning_type="lora")
-        ref_model, _ = load_model_and_tokenizer(
-            ref_model_args, ref_finetuning_args, is_trainable=False, add_valuehead=add_valuehead
+        tokenizer = load_tokenizer(ref_model_args)
+        ref_model = load_model(
+            tokenizer, ref_model_args, ref_finetuning_args, is_trainable=False, add_valuehead=add_valuehead
         )
         logger.info("Created reference model from {}".format(finetuning_args.ref_model))
     else:
         if finetuning_args.finetuning_type == "lora":
             ref_model = None
         else:
-            ref_model, _ = load_model_and_tokenizer(
-                model_args, finetuning_args, is_trainable=False, add_valuehead=add_valuehead
+            tokenizer = load_tokenizer(model_args)
+            ref_model = load_model(
+                tokenizer, model_args, finetuning_args, is_trainable=False, add_valuehead=add_valuehead
             )
             logger.info("Created reference model from the model itself.")
 
@@ -141,8 +143,9 @@ def create_reward_model(
         )
         reward_model_args = ModelArguments(**reward_model_args_dict)
         reward_finetuning_args = FinetuningArguments(finetuning_type="lora")
-        reward_model, _ = load_model_and_tokenizer(
-            reward_model_args, reward_finetuning_args, is_trainable=False, add_valuehead=True
+        tokenizer = load_tokenizer(reward_model_args)
+        reward_model = load_model(
+            tokenizer, reward_model_args, reward_finetuning_args, is_trainable=False, add_valuehead=True
         )
         logger.info("Loaded full weights of reward model from {}".format(finetuning_args.reward_model))
         logger.warning("Please ensure the ppo model and reward model share SAME tokenizer and vocabulary.")

src/llmtuner/webui/chatter.py

@@ -36,7 +36,7 @@ class WebChatModel(ChatModel):
         return self.engine is not None
 
     def load_model(self, data: Dict[Component, Any]) -> Generator[str, None, None]:
-        get = lambda name: data[self.manager.get_elem_by_name(name)]
+        get = lambda elem_id: data[self.manager.get_elem_by_id(elem_id)]
         lang = get("top.lang")
         error = ""
         if self.loaded:
@@ -80,7 +80,7 @@ class WebChatModel(ChatModel):
         yield ALERTS["info_loaded"][lang]
 
     def unload_model(self, data: Dict[Component, Any]) -> Generator[str, None, None]:
-        lang = data[self.manager.get_elem_by_name("top.lang")]
+        lang = data[self.manager.get_elem_by_id("top.lang")]
 
         if self.demo_mode:
             gr.Warning(ALERTS["err_demo"][lang])
@@ -92,23 +92,29 @@ class WebChatModel(ChatModel):
         torch_gc()
         yield ALERTS["info_unloaded"][lang]
 
-    def predict(
+    def append(
         self,
-        chatbot: List[Tuple[str, str]],
+        chatbot: List[List[Optional[str]]],
+        messages: Sequence[Dict[str, str]],
         role: str,
         query: str,
-        messages: Sequence[Tuple[str, str]],
+    ) -> Tuple[List[List[Optional[str]]], List[Dict[str, str]], str]:
+        return chatbot + [[query, None]], messages + [{"role": role, "content": query}], ""
+
+    def stream(
+        self,
+        chatbot: List[List[Optional[str]]],
+        messages: Sequence[Dict[str, str]],
         system: str,
         tools: str,
         max_new_tokens: int,
         top_p: float,
         temperature: float,
-    ) -> Generator[Tuple[Sequence[Tuple[str, str]], Sequence[Tuple[str, str]]], None, None]:
-        chatbot.append([query, ""])
-        query_messages = messages + [{"role": role, "content": query}]
+    ) -> Generator[Tuple[List[List[Optional[str]]], List[Dict[str, str]]], None, None]:
+        chatbot[-1][1] = ""
         response = ""
         for new_text in self.stream_chat(
-            query_messages, system, tools, max_new_tokens=max_new_tokens, top_p=top_p, temperature=temperature
+            messages, system, tools, max_new_tokens=max_new_tokens, top_p=top_p, temperature=temperature
         ):
             response += new_text
             if tools:
@@ -120,18 +126,11 @@ class WebChatModel(ChatModel):
                 name, arguments = result
                 arguments = json.loads(arguments)
                 tool_call = json.dumps({"name": name, "arguments": arguments}, ensure_ascii=False)
-                output_messages = query_messages + [{"role": Role.FUNCTION.value, "content": tool_call}]
+                output_messages = messages + [{"role": Role.FUNCTION.value, "content": tool_call}]
                 bot_text = "```json\n" + tool_call + "\n```"
             else:
-                output_messages = query_messages + [{"role": Role.ASSISTANT.value, "content": result}]
+                output_messages = messages + [{"role": Role.ASSISTANT.value, "content": result}]
                 bot_text = result
 
-            chatbot[-1] = [query, self.postprocess(bot_text)]
+            chatbot[-1][1] = bot_text
             yield chatbot, output_messages
-
-    def postprocess(self, response: str) -> str:
-        blocks = response.split("```")
-        for i, block in enumerate(blocks):
-            if i % 2 == 0:
-                blocks[i] = block.replace("<", "&lt;").replace(">", "&gt;")
-        return "```".join(blocks)

src/llmtuner/webui/common.py

@@ -11,6 +11,7 @@ from ..extras.constants import (
     DEFAULT_MODULE,
     DEFAULT_TEMPLATE,
     PEFT_METHODS,
+    STAGES_USE_PAIR_DATA,
     SUPPORTED_MODELS,
     TRAINING_STAGES,
     DownloadSource,
@@ -20,6 +21,7 @@ from ..extras.misc import use_modelscope
 
 ADAPTER_NAMES = {WEIGHTS_NAME, SAFETENSORS_WEIGHTS_NAME}
 DEFAULT_CACHE_DIR = "cache"
+DEFAULT_CONFIG_DIR = "config"
 DEFAULT_DATA_DIR = "data"
 DEFAULT_SAVE_DIR = "saves"
 USER_CONFIG = "user.config"
@@ -33,6 +35,10 @@ def get_config_path() -> os.PathLike:
     return os.path.join(DEFAULT_CACHE_DIR, USER_CONFIG)
 
 
+def get_save_path(config_path: str) -> os.PathLike:
+    return os.path.join(DEFAULT_CONFIG_DIR, config_path)
+
+
 def load_config() -> Dict[str, Any]:
     try:
         with open(get_config_path(), "r", encoding="utf-8") as f:
@@ -52,6 +58,22 @@ def save_config(lang: str, model_name: Optional[str] = None, model_path: Optiona
         json.dump(user_config, f, indent=2, ensure_ascii=False)
 
 
+def load_args(config_path: str) -> Optional[Dict[str, Any]]:
+    try:
+        with open(get_save_path(config_path), "r", encoding="utf-8") as f:
+            return json.load(f)
+    except Exception:
+        return None
+
+
+def save_args(config_path: str, config_dict: Dict[str, Any]) -> str:
+    os.makedirs(DEFAULT_CONFIG_DIR, exist_ok=True)
+    with open(get_save_path(config_path), "w", encoding="utf-8") as f:
+        json.dump(config_dict, f, indent=2, ensure_ascii=False)
+
+    return str(get_save_path(config_path))
+
+
 def get_model_path(model_name: str) -> str:
     user_config = load_config()
     path_dict: Dict[DownloadSource, str] = SUPPORTED_MODELS.get(model_name, defaultdict(str))
@@ -79,9 +101,9 @@ def get_template(model_name: str) -> str:
     return "default"
 
 
-def list_adapters(model_name: str, finetuning_type: str) -> Dict[str, Any]:
+def list_adapters(model_name: str, finetuning_type: str) -> "gr.Dropdown":
     if finetuning_type not in PEFT_METHODS:
-        return gr.update(value=[], choices=[], interactive=False)
+        return gr.Dropdown(value=[], choices=[], interactive=False)
 
     adapters = []
     if model_name and finetuning_type == "lora":
@@ -92,7 +114,7 @@ def list_adapters(model_name: str, finetuning_type: str) -> Dict[str, Any]:
                     os.path.isfile(os.path.join(save_dir, adapter, name)) for name in ADAPTER_NAMES
                 ):
                     adapters.append(adapter)
-    return gr.update(value=[], choices=adapters, interactive=True)
+    return gr.Dropdown(value=[], choices=adapters, interactive=True)
 
 
 def load_dataset_info(dataset_dir: str) -> Dict[str, Dict[str, Any]]:
@@ -104,12 +126,12 @@ def load_dataset_info(dataset_dir: str) -> Dict[str, Dict[str, Any]]:
         return {}
 
 
-def list_dataset(dataset_dir: str = None, training_stage: str = list(TRAINING_STAGES.keys())[0]) -> Dict[str, Any]:
+def list_dataset(dataset_dir: str = None, training_stage: str = list(TRAINING_STAGES.keys())[0]) -> "gr.Dropdown":
     dataset_info = load_dataset_info(dataset_dir if dataset_dir is not None else DEFAULT_DATA_DIR)
-    ranking = TRAINING_STAGES[training_stage] in ["rm", "dpo"]
+    ranking = TRAINING_STAGES[training_stage] in STAGES_USE_PAIR_DATA
     datasets = [k for k, v in dataset_info.items() if v.get("ranking", False) == ranking]
-    return gr.update(value=[], choices=datasets)
+    return gr.Dropdown(value=[], choices=datasets)
 
 
-def autoset_packing(training_stage: str = list(TRAINING_STAGES.keys())[0]) -> Dict[str, Any]:
-    return gr.update(value=(TRAINING_STAGES[training_stage] == "pt"))
+def autoset_packing(training_stage: str = list(TRAINING_STAGES.keys())[0]) -> "gr.Button":
+    return gr.Button(value=(TRAINING_STAGES[training_stage] == "pt"))

src/llmtuner/webui/components/chatbot.py

@@ -7,7 +7,6 @@ from ..utils import check_json_schema
 
 
 if TYPE_CHECKING:
-    from gradio.blocks import Block
     from gradio.components import Component
 
     from ..engine import Engine
@@ -15,9 +14,9 @@ if TYPE_CHECKING:
 
 def create_chat_box(
     engine: "Engine", visible: bool = False
-) -> Tuple["Block", "Component", "Component", Dict[str, "Component"]]:
-    with gr.Box(visible=visible) as chat_box:
-        chatbot = gr.Chatbot()
+) -> Tuple["gr.Column", "Component", "Component", Dict[str, "Component"]]:
+    with gr.Column(visible=visible) as chat_box:
+        chatbot = gr.Chatbot(show_copy_button=True)
         messages = gr.State([])
         with gr.Row():
             with gr.Column(scale=4):
@@ -33,16 +32,18 @@ def create_chat_box(
                 temperature = gr.Slider(0.01, 1.5, value=0.95, step=0.01)
                 clear_btn = gr.Button()
 
-    tools.input(check_json_schema, [tools, engine.manager.get_elem_by_name("top.lang")])
+    tools.input(check_json_schema, inputs=[tools, engine.manager.get_elem_by_id("top.lang")])
 
     submit_btn.click(
-        engine.chatter.predict,
-        [chatbot, role, query, messages, system, tools, max_new_tokens, top_p, temperature],
+        engine.chatter.append,
+        [chatbot, messages, role, query],
+        [chatbot, messages, query],
+    ).then(
+        engine.chatter.stream,
+        [chatbot, messages, system, tools, max_new_tokens, top_p, temperature],
         [chatbot, messages],
-        show_progress=True,
-    ).then(lambda: gr.update(value=""), outputs=[query])
-
-    clear_btn.click(lambda: ([], []), outputs=[chatbot, messages], show_progress=True)
+    )
+    clear_btn.click(lambda: ([], []), outputs=[chatbot, messages])
 
     return (
         chat_box,

src/llmtuner/webui/components/data.py

@@ -1,6 +1,6 @@
 import json
 import os
-from typing import TYPE_CHECKING, Any, Dict, Tuple
+from typing import TYPE_CHECKING, Dict, Tuple
 
 import gradio as gr
 
@@ -22,24 +22,24 @@ def next_page(page_index: int, total_num: int) -> int:
     return page_index + 1 if (page_index + 1) * PAGE_SIZE < total_num else page_index
 
 
-def can_preview(dataset_dir: str, dataset: list) -> Dict[str, Any]:
+def can_preview(dataset_dir: str, dataset: list) -> "gr.Button":
     try:
         with open(os.path.join(dataset_dir, DATA_CONFIG), "r", encoding="utf-8") as f:
             dataset_info = json.load(f)
     except Exception:
-        return gr.update(interactive=False)
+        return gr.Button(interactive=False)
 
     if (
         len(dataset) > 0
         and "file_name" in dataset_info[dataset[0]]
         and os.path.isfile(os.path.join(dataset_dir, dataset_info[dataset[0]]["file_name"]))
     ):
-        return gr.update(interactive=True)
+        return gr.Button(interactive=True)
     else:
-        return gr.update(interactive=False)
+        return gr.Button(interactive=False)
 
 
-def get_preview(dataset_dir: str, dataset: list, page_index: int) -> Tuple[int, list, Dict[str, Any]]:
+def get_preview(dataset_dir: str, dataset: list, page_index: int) -> Tuple[int, list, "gr.Column"]:
     with open(os.path.join(dataset_dir, DATA_CONFIG), "r", encoding="utf-8") as f:
         dataset_info = json.load(f)
 
@@ -51,7 +51,7 @@ def get_preview(dataset_dir: str, dataset: list, page_index: int) -> Tuple[int,
             data = [json.loads(line) for line in f]
         else:
             data = [line for line in f]  # noqa: C416
-    return len(data), data[PAGE_SIZE * page_index : PAGE_SIZE * (page_index + 1)], gr.update(visible=True)
+    return len(data), data[PAGE_SIZE * page_index : PAGE_SIZE * (page_index + 1)], gr.Column(visible=True)
 
 
 def create_preview_box(dataset_dir: "gr.Textbox", dataset: "gr.Dropdown") -> Dict[str, "Component"]:
@@ -67,7 +67,7 @@ def create_preview_box(dataset_dir: "gr.Textbox", dataset: "gr.Dropdown") -> Dic
             close_btn = gr.Button()
 
         with gr.Row():
-            preview_samples = gr.JSON(interactive=False)
+            preview_samples = gr.JSON()
 
     dataset.change(can_preview, [dataset_dir, dataset], [data_preview_btn], queue=False).then(
         lambda: 0, outputs=[page_index], queue=False
@@ -81,7 +81,7 @@ def create_preview_box(dataset_dir: "gr.Textbox", dataset: "gr.Dropdown") -> Dic
     next_btn.click(next_page, [page_index, preview_count], [page_index], queue=False).then(
         get_preview, [dataset_dir, dataset, page_index], [preview_count, preview_samples, preview_box], queue=False
     )
-    close_btn.click(lambda: gr.update(visible=False), outputs=[preview_box], queue=False)
+    close_btn.click(lambda: gr.Column(visible=False), outputs=[preview_box], queue=False)
     return dict(
         data_preview_btn=data_preview_btn,
         preview_count=preview_count,

src/llmtuner/webui/components/eval.py

@@ -18,11 +18,9 @@ def create_eval_tab(engine: "Engine") -> Dict[str, "Component"]:
 
     with gr.Row():
         dataset_dir = gr.Textbox(value=DEFAULT_DATA_DIR, scale=2)
-        dataset = gr.Dropdown(multiselect=True, scale=4)
+        dataset = gr.Dropdown(multiselect=True, allow_custom_value=True, scale=4)
         preview_elems = create_preview_box(dataset_dir, dataset)
 
-    dataset_dir.change(list_dataset, [dataset_dir], [dataset], queue=False)
-
     input_elems.update({dataset_dir, dataset})
     elem_dict.update(dict(dataset_dir=dataset_dir, dataset=dataset, **preview_elems))
 
@@ -46,14 +44,14 @@ def create_eval_tab(engine: "Engine") -> Dict[str, "Component"]:
 
     with gr.Row():
         cmd_preview_btn = gr.Button()
-        start_btn = gr.Button()
-        stop_btn = gr.Button()
+        start_btn = gr.Button(variant="primary")
+        stop_btn = gr.Button(variant="stop")
 
     with gr.Row():
-        resume_btn = gr.Checkbox(visible=False, interactive=False, value=False)
+        resume_btn = gr.Checkbox(visible=False, interactive=False)
         process_bar = gr.Slider(visible=False, interactive=False)
 
-    with gr.Box():
+    with gr.Row():
         output_box = gr.Markdown()
 
     output_elems = [output_box, process_bar]
@@ -68,9 +66,11 @@ def create_eval_tab(engine: "Engine") -> Dict[str, "Component"]:
         )
     )
 
-    cmd_preview_btn.click(engine.runner.preview_eval, input_elems, output_elems)
+    cmd_preview_btn.click(engine.runner.preview_eval, input_elems, output_elems, concurrency_limit=None)
     start_btn.click(engine.runner.run_eval, input_elems, output_elems)
-    stop_btn.click(engine.runner.set_abort, queue=False)
-    resume_btn.change(engine.runner.monitor, outputs=output_elems)
+    stop_btn.click(engine.runner.set_abort)
+    resume_btn.change(engine.runner.monitor, outputs=output_elems, concurrency_limit=None)
+
+    dataset_dir.change(list_dataset, [dataset_dir], [dataset], queue=False)
 
     return elem_dict

src/llmtuner/webui/components/export.py

@@ -74,7 +74,7 @@ def save_model(
 
 def create_export_tab(engine: "Engine") -> Dict[str, "Component"]:
     with gr.Row():
-        max_shard_size = gr.Slider(value=1, minimum=1, maximum=100)
+        max_shard_size = gr.Slider(value=1, minimum=1, maximum=100, step=1)
         export_quantization_bit = gr.Dropdown(choices=["none", "8", "4", "3", "2"], value="none")
         export_quantization_dataset = gr.Textbox(value="data/c4_demo.json")
         export_legacy_format = gr.Checkbox()
@@ -89,12 +89,12 @@ def create_export_tab(engine: "Engine") -> Dict[str, "Component"]:
     export_btn.click(
         save_model,
         [
-            engine.manager.get_elem_by_name("top.lang"),
-            engine.manager.get_elem_by_name("top.model_name"),
-            engine.manager.get_elem_by_name("top.model_path"),
-            engine.manager.get_elem_by_name("top.adapter_path"),
-            engine.manager.get_elem_by_name("top.finetuning_type"),
-            engine.manager.get_elem_by_name("top.template"),
+            engine.manager.get_elem_by_id("top.lang"),
+            engine.manager.get_elem_by_id("top.model_name"),
+            engine.manager.get_elem_by_id("top.model_path"),
+            engine.manager.get_elem_by_id("top.adapter_path"),
+            engine.manager.get_elem_by_id("top.finetuning_type"),
+            engine.manager.get_elem_by_id("top.template"),
             max_shard_size,
             export_quantization_bit,
             export_quantization_dataset,

src/llmtuner/webui/components/infer.py

@@ -25,15 +25,15 @@ def create_infer_tab(engine: "Engine") -> Dict[str, "Component"]:
     input_elems.update({infer_backend})
     elem_dict.update(dict(infer_backend=infer_backend, load_btn=load_btn, unload_btn=unload_btn, info_box=info_box))
 
-    chat_box, chatbot, history, chat_elems = create_chat_box(engine, visible=False)
+    chat_box, chatbot, messages, chat_elems = create_chat_box(engine, visible=False)
     elem_dict.update(dict(chat_box=chat_box, **chat_elems))
 
     load_btn.click(engine.chatter.load_model, input_elems, [info_box]).then(
-        lambda: gr.update(visible=engine.chatter.loaded), outputs=[chat_box]
+        lambda: gr.Column(visible=engine.chatter.loaded), outputs=[chat_box]
     )
 
     unload_btn.click(engine.chatter.unload_model, input_elems, [info_box]).then(
-        lambda: ([], []), outputs=[chatbot, history]
-    ).then(lambda: gr.update(visible=engine.chatter.loaded), outputs=[chat_box])
+        lambda: ([], []), outputs=[chatbot, messages]
+    ).then(lambda: gr.Column(visible=engine.chatter.loaded), outputs=[chat_box])
 
     return elem_dict

src/llmtuner/webui/components/top.py

@@ -1,4 +1,4 @@
-from typing import TYPE_CHECKING, Dict, Tuple
+from typing import TYPE_CHECKING, Dict
 
 import gradio as gr
 
@@ -12,7 +12,7 @@ if TYPE_CHECKING:
     from gradio.components import Component
 
 
-def create_top() -> Tuple["gr.Dropdown", Dict[str, "Component"]]:
+def create_top() -> Dict[str, "Component"]:
     available_models = list(SUPPORTED_MODELS.keys()) + ["Custom"]
 
     with gr.Row():
@@ -25,7 +25,7 @@ def create_top() -> Tuple["gr.Dropdown", Dict[str, "Component"]]:
         adapter_path = gr.Dropdown(multiselect=True, allow_custom_value=True, scale=5)
         refresh_btn = gr.Button(scale=1)
 
-    with gr.Accordion(label="Advanced config", open=False) as advanced_tab:
+    with gr.Accordion(open=False) as advanced_tab:
         with gr.Row():
             quantization_bit = gr.Dropdown(choices=["none", "8", "4"], value="none")
             template = gr.Dropdown(choices=list(templates.keys()), value="default")
@@ -44,7 +44,7 @@ def create_top() -> Tuple["gr.Dropdown", Dict[str, "Component"]]:
 
     refresh_btn.click(list_adapters, [model_name, finetuning_type], [adapter_path], queue=False)
 
-    return lang, dict(
+    return dict(
         lang=lang,
         model_name=model_name,
         model_path=model_path,

src/llmtuner/webui/components/train.py

@@ -6,7 +6,6 @@ from transformers.trainer_utils import SchedulerType
 from ...extras.constants import TRAINING_STAGES
 from ..common import DEFAULT_DATA_DIR, autoset_packing, list_adapters, list_dataset
 from ..components.data import create_preview_box
-from ..utils import gen_plot
 
 
 if TYPE_CHECKING:
@@ -21,14 +20,12 @@ def create_train_tab(engine: "Engine") -> Dict[str, "Component"]:
 
     with gr.Row():
         training_stage = gr.Dropdown(
-            choices=list(TRAINING_STAGES.keys()), value=list(TRAINING_STAGES.keys())[0], scale=2
+            choices=list(TRAINING_STAGES.keys()), value=list(TRAINING_STAGES.keys())[0], scale=1
         )
-        dataset_dir = gr.Textbox(value=DEFAULT_DATA_DIR, scale=2)
-        dataset = gr.Dropdown(multiselect=True, scale=4)
+        dataset_dir = gr.Textbox(value=DEFAULT_DATA_DIR, scale=1)
+        dataset = gr.Dropdown(multiselect=True, allow_custom_value=True, scale=4)
         preview_elems = create_preview_box(dataset_dir, dataset)
 
-    dataset_dir.change(list_dataset, [dataset_dir, training_stage], [dataset], queue=False)
-
     input_elems.update({training_stage, dataset_dir, dataset})
     elem_dict.update(dict(training_stage=training_stage, dataset_dir=dataset_dir, dataset=dataset, **preview_elems))
 
@@ -68,7 +65,7 @@ def create_train_tab(engine: "Engine") -> Dict[str, "Component"]:
         )
     )
 
-    with gr.Accordion(label="Extra config", open=False) as extra_tab:
+    with gr.Accordion(open=False) as extra_tab:
         with gr.Row():
             logging_steps = gr.Slider(value=5, minimum=5, maximum=1000, step=5)
             save_steps = gr.Slider(value=100, minimum=10, maximum=5000, step=10)
@@ -77,11 +74,17 @@ def create_train_tab(engine: "Engine") -> Dict[str, "Component"]:
             optim = gr.Textbox(value="adamw_torch")
 
         with gr.Row():
+            with gr.Column():
                 resize_vocab = gr.Checkbox()
                 packing = gr.Checkbox()
+
+            with gr.Column():
                 upcast_layernorm = gr.Checkbox()
                 use_llama_pro = gr.Checkbox()
+
+            with gr.Column():
                 shift_attn = gr.Checkbox()
+                report_to = gr.Checkbox()
 
     input_elems.update(
         {
@@ -95,6 +98,7 @@ def create_train_tab(engine: "Engine") -> Dict[str, "Component"]:
             upcast_layernorm,
             use_llama_pro,
             shift_attn,
+            report_to,
         }
     )
     elem_dict.update(
@@ -110,13 +114,14 @@ def create_train_tab(engine: "Engine") -> Dict[str, "Component"]:
             upcast_layernorm=upcast_layernorm,
             use_llama_pro=use_llama_pro,
             shift_attn=shift_attn,
+            report_to=report_to,
         )
     )
 
-    with gr.Accordion(label="Freeze config", open=False) as freeze_tab:
+    with gr.Accordion(open=False) as freeze_tab:
         with gr.Row():
-            num_layer_trainable = gr.Slider(value=3, minimum=1, maximum=128, step=1, scale=2)
-            name_module_trainable = gr.Textbox(value="all", scale=3)
+            num_layer_trainable = gr.Slider(value=3, minimum=1, maximum=128, step=1)
+            name_module_trainable = gr.Textbox(value="all")
 
     input_elems.update({num_layer_trainable, name_module_trainable})
     elem_dict.update(
@@ -125,21 +130,34 @@ def create_train_tab(engine: "Engine") -> Dict[str, "Component"]:
         )
     )
 
-    with gr.Accordion(label="LoRA config", open=False) as lora_tab:
+    with gr.Accordion(open=False) as lora_tab:
         with gr.Row():
-            lora_rank = gr.Slider(value=8, minimum=1, maximum=1024, step=1, scale=1)
-            lora_alpha = gr.Slider(value=16, minimum=1, maximum=2048, step=1, scale=1)
-            lora_dropout = gr.Slider(value=0.1, minimum=0, maximum=1, step=0.01, scale=1)
-            lora_target = gr.Textbox(scale=2)
+            lora_rank = gr.Slider(value=8, minimum=1, maximum=1024, step=1)
+            lora_alpha = gr.Slider(value=16, minimum=1, maximum=2048, step=1)
+            lora_dropout = gr.Slider(value=0.1, minimum=0, maximum=1, step=0.01)
+            loraplus_lr_ratio = gr.Slider(value=0, minimum=0, maximum=64, step=0.01)
+            create_new_adapter = gr.Checkbox()
 
         with gr.Row():
-            use_rslora = gr.Checkbox(scale=1)
-            use_dora = gr.Checkbox(scale=1)
-            create_new_adapter = gr.Checkbox(scale=1)
+            with gr.Column(scale=1):
+                use_rslora = gr.Checkbox()
+                use_dora = gr.Checkbox()
+
+            lora_target = gr.Textbox(scale=2)
             additional_target = gr.Textbox(scale=2)
 
     input_elems.update(
-        {lora_rank, lora_alpha, lora_dropout, lora_target, use_rslora, use_dora, create_new_adapter, additional_target}
+        {
+            lora_rank,
+            lora_alpha,
+            lora_dropout,
+            loraplus_lr_ratio,
+            create_new_adapter,
+            use_rslora,
+            use_dora,
+            lora_target,
+            additional_target,
+        }
     )
     elem_dict.update(
         dict(
@@ -147,37 +165,34 @@ def create_train_tab(engine: "Engine") -> Dict[str, "Component"]:
             lora_rank=lora_rank,
             lora_alpha=lora_alpha,
             lora_dropout=lora_dropout,
-            lora_target=lora_target,
+            loraplus_lr_ratio=loraplus_lr_ratio,
+            create_new_adapter=create_new_adapter,
             use_rslora=use_rslora,
             use_dora=use_dora,
-            create_new_adapter=create_new_adapter,
+            lora_target=lora_target,
             additional_target=additional_target,
         )
     )
 
-    with gr.Accordion(label="RLHF config", open=False) as rlhf_tab:
+    with gr.Accordion(open=False) as rlhf_tab:
         with gr.Row():
-            dpo_beta = gr.Slider(value=0.1, minimum=0, maximum=1, step=0.01, scale=1)
-            dpo_ftx = gr.Slider(value=0, minimum=0, maximum=10, step=0.01, scale=1)
-            reward_model = gr.Dropdown(multiselect=True, allow_custom_value=True, scale=2)
+            dpo_beta = gr.Slider(value=0.1, minimum=0, maximum=1, step=0.01)
+            dpo_ftx = gr.Slider(value=0, minimum=0, maximum=10, step=0.01)
+            orpo_beta = gr.Slider(value=0.1, minimum=0, maximum=1, step=0.01)
+            reward_model = gr.Dropdown(multiselect=True, allow_custom_value=True)
 
-    training_stage.change(list_dataset, [dataset_dir, training_stage], [dataset], queue=False).then(
-        list_adapters,
-        [engine.manager.get_elem_by_name("top.model_name"), engine.manager.get_elem_by_name("top.finetuning_type")],
-        [reward_model],
-        queue=False,
-    ).then(autoset_packing, [training_stage], [packing], queue=False)
+    input_elems.update({dpo_beta, dpo_ftx, orpo_beta, reward_model})
+    elem_dict.update(
+        dict(rlhf_tab=rlhf_tab, dpo_beta=dpo_beta, dpo_ftx=dpo_ftx, orpo_beta=orpo_beta, reward_model=reward_model)
+    )
 
-    input_elems.update({dpo_beta, dpo_ftx, reward_model})
-    elem_dict.update(dict(rlhf_tab=rlhf_tab, dpo_beta=dpo_beta, dpo_ftx=dpo_ftx, reward_model=reward_model))
-
-    with gr.Accordion(label="GaLore config", open=False) as galore_tab:
+    with gr.Accordion(open=False) as galore_tab:
         with gr.Row():
-            use_galore = gr.Checkbox(scale=1)
-            galore_rank = gr.Slider(value=16, minimum=1, maximum=1024, step=1, scale=2)
-            galore_update_interval = gr.Slider(value=200, minimum=1, maximum=1024, step=1, scale=2)
-            galore_scale = gr.Slider(value=0.25, minimum=0, maximum=1, step=0.01, scale=2)
-            galore_target = gr.Textbox(value="mlp,attn", scale=3)
+            use_galore = gr.Checkbox()
+            galore_rank = gr.Slider(value=16, minimum=1, maximum=1024, step=1)
+            galore_update_interval = gr.Slider(value=200, minimum=1, maximum=1024, step=1)
+            galore_scale = gr.Slider(value=0.25, minimum=0, maximum=1, step=0.01)
+            galore_target = gr.Textbox(value="all")
 
     input_elems.update({use_galore, galore_rank, galore_update_interval, galore_scale, galore_target})
     elem_dict.update(
@@ -193,38 +208,36 @@ def create_train_tab(engine: "Engine") -> Dict[str, "Component"]:
 
     with gr.Row():
         cmd_preview_btn = gr.Button()
-        start_btn = gr.Button()
-        stop_btn = gr.Button()
+        arg_save_btn = gr.Button()
+        arg_load_btn = gr.Button()
+        start_btn = gr.Button(variant="primary")
+        stop_btn = gr.Button(variant="stop")
 
     with gr.Row():
         with gr.Column(scale=3):
             with gr.Row():
                 output_dir = gr.Textbox()
+                config_path = gr.Textbox()
 
             with gr.Row():
                 resume_btn = gr.Checkbox(visible=False, interactive=False)
                 process_bar = gr.Slider(visible=False, interactive=False)
 
-            with gr.Box():
+            with gr.Row():
                 output_box = gr.Markdown()
 
         with gr.Column(scale=1):
             loss_viewer = gr.Plot()
 
-    input_elems.add(output_dir)
-    output_elems = [output_box, process_bar]
-
-    cmd_preview_btn.click(engine.runner.preview_train, input_elems, output_elems)
-    start_btn.click(engine.runner.run_train, input_elems, output_elems)
-    stop_btn.click(engine.runner.set_abort, queue=False)
-    resume_btn.change(engine.runner.monitor, outputs=output_elems)
-
     elem_dict.update(
         dict(
             cmd_preview_btn=cmd_preview_btn,
+            arg_save_btn=arg_save_btn,
+            arg_load_btn=arg_load_btn,
             start_btn=start_btn,
             stop_btn=stop_btn,
             output_dir=output_dir,
+            config_path=config_path,
             resume_btn=resume_btn,
             process_bar=process_bar,
             output_box=output_box,
@@ -232,15 +245,27 @@ def create_train_tab(engine: "Engine") -> Dict[str, "Component"]:
         )
     )
 
-    output_box.change(
-        gen_plot,
-        [
-            engine.manager.get_elem_by_name("top.model_name"),
-            engine.manager.get_elem_by_name("top.finetuning_type"),
-            output_dir,
-        ],
-        loss_viewer,
-        queue=False,
+    input_elems.update({output_dir, config_path})
+    output_elems = [output_box, process_bar, loss_viewer]
+
+    cmd_preview_btn.click(engine.runner.preview_train, input_elems, output_elems, concurrency_limit=None)
+    arg_save_btn.click(engine.runner.save_args, input_elems, output_elems, concurrency_limit=None)
+    arg_load_btn.click(
+        engine.runner.load_args,
+        [engine.manager.get_elem_by_id("top.lang"), config_path],
+        list(input_elems) + [output_box],
+        concurrency_limit=None,
     )
+    start_btn.click(engine.runner.run_train, input_elems, output_elems)
+    stop_btn.click(engine.runner.set_abort)
+    resume_btn.change(engine.runner.monitor, outputs=output_elems, concurrency_limit=None)
+
+    dataset_dir.change(list_dataset, [dataset_dir, training_stage], [dataset], queue=False)
+    training_stage.change(list_dataset, [dataset_dir, training_stage], [dataset], queue=False).then(
+        list_adapters,
+        [engine.manager.get_elem_by_id("top.model_name"), engine.manager.get_elem_by_id("top.finetuning_type")],
+        [reward_model],
+        queue=False,
+    ).then(autoset_packing, [training_stage], [packing], queue=False)
 
     return elem_dict

src/llmtuner/webui/engine.py

@@ -1,6 +1,5 @@
 from typing import Any, Dict, Generator
 
-import gradio as gr
 from gradio.components import Component  # cannot use TYPE_CHECKING here
 
 from .chatter import WebChatModel
@@ -19,44 +18,46 @@ class Engine:
         self.runner = Runner(self.manager, demo_mode)
         self.chatter = WebChatModel(self.manager, demo_mode, lazy_init=(not pure_chat))
 
-    def _form_dict(self, resume_dict: Dict[str, Dict[str, Any]]):
-        return {self.manager.get_elem_by_name(k): gr.update(**v) for k, v in resume_dict.items()}
+    def _update_component(self, input_dict: Dict[str, Dict[str, Any]]) -> Dict["Component", "Component"]:
+        r"""
+        Gets the dict to update the components.
+        """
+        output_dict: Dict["Component", "Component"] = {}
+        for elem_id, elem_attr in input_dict.items():
+            elem = self.manager.get_elem_by_id(elem_id)
+            output_dict[elem] = elem.__class__(**elem_attr)
 
-    def resume(self) -> Generator[Dict[Component, Dict[str, Any]], None, None]:
+        return output_dict
+
+    def resume(self) -> Generator[Dict[Component, Component], None, None]:
         user_config = load_config() if not self.demo_mode else {}
         lang = user_config.get("lang", None) or "en"
 
         init_dict = {"top.lang": {"value": lang}, "infer.chat_box": {"visible": self.chatter.loaded}}
 
         if not self.pure_chat:
-            init_dict["train.dataset"] = {"choices": list_dataset()["choices"]}
-            init_dict["eval.dataset"] = {"choices": list_dataset()["choices"]}
+            init_dict["train.dataset"] = {"choices": list_dataset().choices}
+            init_dict["eval.dataset"] = {"choices": list_dataset().choices}
+            init_dict["train.output_dir"] = {"value": "train_{}".format(get_time())}
+            init_dict["train.config_path"] = {"value": "{}.json".format(get_time())}
+            init_dict["eval.output_dir"] = {"value": "eval_{}".format(get_time())}
 
             if user_config.get("last_model", None):
                 init_dict["top.model_name"] = {"value": user_config["last_model"]}
                 init_dict["top.model_path"] = {"value": get_model_path(user_config["last_model"])}
 
-        yield self._form_dict(init_dict)
+        yield self._update_component(init_dict)
 
-        if not self.pure_chat:
-            if self.runner.alive and not self.demo_mode:
-                yield {elem: gr.update(value=value) for elem, value in self.runner.running_data.items()}
+        if self.runner.alive and not self.demo_mode and not self.pure_chat:
+            yield {elem: elem.__class__(value=value) for elem, value in self.runner.running_data.items()}
             if self.runner.do_train:
-                    yield self._form_dict({"train.resume_btn": {"value": True}})
+                yield self._update_component({"train.resume_btn": {"value": True}})
             else:
-                    yield self._form_dict({"eval.resume_btn": {"value": True}})
-            else:
-                yield self._form_dict(
-                    {
-                        "train.output_dir": {"value": "train_" + get_time()},
-                        "eval.output_dir": {"value": "eval_" + get_time()},
-                    }
-                )
+                yield self._update_component({"eval.resume_btn": {"value": True}})
 
-    def change_lang(self, lang: str) -> Dict[Component, Dict[str, Any]]:
+    def change_lang(self, lang: str) -> Dict[Component, Component]:
         return {
-            component: gr.update(**LOCALES[name][lang])
-            for elems in self.manager.all_elems.values()
-            for name, component in elems.items()
-            if name in LOCALES
+            elem: elem.__class__(**LOCALES[elem_name][lang])
+            for elem_name, elem in self.manager.get_elem_iter()
+            if elem_name in LOCALES
         }

src/llmtuner/webui/interface.py

@@ -1,5 +1,4 @@
 import gradio as gr
-from transformers.utils.versions import require_version
 
 from .common import save_config
 from .components import (
@@ -14,9 +13,6 @@ from .css import CSS
 from .engine import Engine
 
 
-require_version("gradio>=3.38.0,<4.0.0", 'To fix: pip install "gradio>=3.38.0,<4.0.0"')
-
-
 def create_ui(demo_mode: bool = False) -> gr.Blocks:
     engine = Engine(demo_mode=demo_mode, pure_chat=False)
 
@@ -29,23 +25,24 @@ def create_ui(demo_mode: bool = False) -> gr.Blocks:
             )
             gr.DuplicateButton(value="Duplicate Space for private use", elem_classes="duplicate-button")
 
-        lang, engine.manager.all_elems["top"] = create_top()
+        engine.manager.add_elems("top", create_top())
+        lang: "gr.Dropdown" = engine.manager.get_elem_by_id("top.lang")
 
         with gr.Tab("Train"):
-            engine.manager.all_elems["train"] = create_train_tab(engine)
+            engine.manager.add_elems("train", create_train_tab(engine))
 
         with gr.Tab("Evaluate & Predict"):
-            engine.manager.all_elems["eval"] = create_eval_tab(engine)
+            engine.manager.add_elems("eval", create_eval_tab(engine))
 
         with gr.Tab("Chat"):
-            engine.manager.all_elems["infer"] = create_infer_tab(engine)
+            engine.manager.add_elems("infer", create_infer_tab(engine))
 
         if not demo_mode:
             with gr.Tab("Export"):
-                engine.manager.all_elems["export"] = create_export_tab(engine)
+                engine.manager.add_elems("export", create_export_tab(engine))
 
-        demo.load(engine.resume, outputs=engine.manager.list_elems())
-        lang.change(engine.change_lang, [lang], engine.manager.list_elems(), queue=False)
+        demo.load(engine.resume, outputs=engine.manager.get_elem_list(), concurrency_limit=None)
+        lang.change(engine.change_lang, [lang], engine.manager.get_elem_list(), queue=False)
         lang.input(save_config, inputs=[lang], queue=False)
 
     return demo
@@ -56,19 +53,17 @@ def create_web_demo() -> gr.Blocks:
 
     with gr.Blocks(title="Web Demo", css=CSS) as demo:
         lang = gr.Dropdown(choices=["en", "zh"])
-        engine.manager.all_elems["top"] = dict(lang=lang)
+        engine.manager.add_elems("top", dict(lang=lang))
 
         chat_box, _, _, chat_elems = create_chat_box(engine, visible=True)
-        engine.manager.all_elems["infer"] = dict(chat_box=chat_box, **chat_elems)
+        engine.manager.add_elems("infer", dict(chat_box=chat_box, **chat_elems))
 
-        demo.load(engine.resume, outputs=engine.manager.list_elems())
-        lang.change(engine.change_lang, [lang], engine.manager.list_elems(), queue=False)
+        demo.load(engine.resume, outputs=engine.manager.get_elem_list(), concurrency_limit=None)
+        lang.change(engine.change_lang, [lang], engine.manager.get_elem_list(), queue=False)
         lang.input(save_config, inputs=[lang], queue=False)
 
     return demo
 
 
 if __name__ == "__main__":
-    demo = create_ui()
-    demo.queue()
-    demo.launch(server_name="0.0.0.0", share=False, inbrowser=True)
+    create_ui().queue().launch(server_name="0.0.0.0", server_port=None, share=False, inbrowser=True)

src/llmtuner/webui/locales.py

@@ -536,6 +536,20 @@ LOCALES = {
             "info": "使用 LongLoRA 提出的 shift short attention。",
         },
     },
+    "report_to": {
+        "en": {
+            "label": "Enable external logger",
+            "info": "Use TensorBoard or wandb to log experiment.",
+        },
+        "ru": {
+            "label": "Включить внешний регистратор",
+            "info": "Использовать TensorBoard или wandb для ведения журнала экспериментов.",
+        },
+        "zh": {
+            "label": "启用外部记录面板",
+            "info": "使用 TensorBoard 或 wandb 记录实验。",
+        },
+    },
     "freeze_tab": {
         "en": {
             "label": "Freeze tuning configurations",
@@ -628,18 +642,32 @@ LOCALES = {
             "info": "LoRA 权重随机丢弃的概率。",
         },
     },
-    "lora_target": {
+    "loraplus_lr_ratio": {
         "en": {
-            "label": "LoRA modules (optional)",
-            "info": "Name(s) of modules to apply LoRA. Use commas to separate multiple modules.",
+            "label": "LoRA+ LR ratio",
+            "info": "The LR ratio of the B matrices in LoRA.",
         },
         "ru": {
-            "label": "Модули LoRA (опционально)",
-            "info": "Имена модулей для применения LoRA. Используйте запятые для разделения нескольких модулей.",
+            "label": "LoRA+ LR коэффициент",
+            "info": "Коэффициент LR матриц B в LoRA.",
         },
         "zh": {
-            "label": "LoRA 作用模块（非必填）",
-            "info": "应用 LoRA 的模块名称。使用英文逗号分隔多个名称。",
+            "label": "LoRA+ 学习率比例",
+            "info": "LoRA+ 中 B 矩阵的学习率倍数。",
+        },
+    },
+    "create_new_adapter": {
+        "en": {
+            "label": "Create new adapter",
+            "info": "Create a new adapter with randomly initialized weight upon the existing one.",
+        },
+        "ru": {
+            "label": "Создать новый адаптер",
+            "info": "Создать новый адаптер с случайной инициализацией веса на основе существующего.",
+        },
+        "zh": {
+            "label": "新建适配器",
+            "info": "在现有的适配器上创建一个随机初始化后的新适配器。",
         },
     },
     "use_rslora": {
@@ -670,18 +698,18 @@ LOCALES = {
             "info": "使用权重分解的 LoRA。",
         },
     },
-    "create_new_adapter": {
+    "lora_target": {
         "en": {
-            "label": "Create new adapter",
-            "info": "Create a new adapter with randomly initialized weight upon the existing one.",
+            "label": "LoRA modules (optional)",
+            "info": "Name(s) of modules to apply LoRA. Use commas to separate multiple modules.",
         },
         "ru": {
-            "label": "Создать новый адаптер",
-            "info": "Создать новый адаптер с случайной инициализацией веса на основе существующего.",
+            "label": "Модули LoRA (опционально)",
+            "info": "Имена модулей для применения LoRA. Используйте запятые для разделения нескольких модулей.",
         },
         "zh": {
-            "label": "新建适配器",
-            "info": "在现有的适配器上创建一个随机初始化后的新适配器。",
+            "label": "LoRA 作用模块（非必填）",
+            "info": "应用 LoRA 的模块名称。使用英文逗号分隔多个名称。",
         },
     },
     "additional_target": {
@@ -743,6 +771,20 @@ LOCALES = {
             "info": "DPO-ftx 中 SFT 损失的权重大小。",
         },
     },
+    "orpo_beta": {
+        "en": {
+            "label": "ORPO beta",
+            "info": "Value of the beta parameter in the ORPO loss.",
+        },
+        "ru": {
+            "label": "ORPO бета",
+            "info": "Значение параметра бета в функции потерь ORPO.",
+        },
+        "zh": {
+            "label": "ORPO beta 参数",
+            "info": "ORPO 损失函数中 beta 超参数大小。",
+        },
+    },
     "reward_model": {
         "en": {
             "label": "Reward model",
@@ -849,6 +891,28 @@ LOCALES = {
             "value": "预览命令",
         },
     },
+    "arg_save_btn": {
+        "en": {
+            "value": "Save arguments",
+        },
+        "ru": {
+            "value": "Сохранить аргументы",
+        },
+        "zh": {
+            "value": "保存训练参数",
+        },
+    },
+    "arg_load_btn": {
+        "en": {
+            "value": "Load arguments",
+        },
+        "ru": {
+            "value": "Загрузить аргументы",
+        },
+        "zh": {
+            "value": "载入训练参数",
+        },
+    },
     "start_btn": {
         "en": {
             "value": "Start",
@@ -885,6 +949,20 @@ LOCALES = {
             "info": "保存结果的路径。",
         },
     },
+    "config_path": {
+        "en": {
+            "label": "Config path",
+            "info": "Path to config saving arguments.",
+        },
+        "ru": {
+            "label": "Путь к конфигурации",
+            "info": "Путь для сохранения аргументов конфигурации.",
+        },
+        "zh": {
+            "label": "配置路径",
+            "info": "保存训练参数的配置文件路径。",
+        },
+    },
     "output_box": {
         "en": {
             "value": "Ready.",
@@ -1236,6 +1314,11 @@ ALERTS = {
         "ru": "Неверная схема JSON.",
         "zh": "Json 格式错误。",
     },
+    "err_config_not_found": {
+        "en": "Config file is not found.",
+        "ru": "Файл конфигурации не найден.",
+        "zh": "未找到配置文件。",
+    },
     "warn_no_cuda": {
         "en": "CUDA environment was not detected.",
         "ru": "Среда CUDA не обнаружена.",
@@ -1256,6 +1339,16 @@ ALERTS = {
         "ru": "Завершено.",
         "zh": "训练完毕。",
     },
+    "info_config_saved": {
+        "en": "Arguments have been saved at: ",
+        "ru": "Аргументы были сохранены по адресу: ",
+        "zh": "训练参数已保存至：",
+    },
+    "info_config_loaded": {
+        "en": "Arguments have been restored.",
+        "ru": "Аргументы были восстановлены.",
+        "zh": "训练参数已载入。",
+    },
     "info_loading": {
         "en": "Loading model...",
         "ru": "Загрузка модели...",

src/llmtuner/webui/manager.py

@@ -1,4 +1,4 @@
-from typing import TYPE_CHECKING, Dict, List, Set
+from typing import TYPE_CHECKING, Dict, Generator, List, Set, Tuple
 
 
 if TYPE_CHECKING:
@@ -7,27 +7,57 @@ if TYPE_CHECKING:
 
 class Manager:
     def __init__(self) -> None:
-        self.all_elems: Dict[str, Dict[str, "Component"]] = {}
+        self._id_to_elem: Dict[str, "Component"] = {}
+        self._elem_to_id: Dict["Component", str] = {}
 
-    def get_elem_by_name(self, name: str) -> "Component":
+    def add_elems(self, tab_name: str, elem_dict: Dict[str, "Component"]) -> None:
         r"""
+        Adds elements to manager.
+        """
+        for elem_name, elem in elem_dict.items():
+            elem_id = "{}.{}".format(tab_name, elem_name)
+            self._id_to_elem[elem_id] = elem
+            self._elem_to_id[elem] = elem_id
+
+    def get_elem_list(self) -> List["Component"]:
+        r"""
+        Returns the list of all elements.
+        """
+        return list(self._id_to_elem.values())
+
+    def get_elem_iter(self) -> Generator[Tuple[str, "Component"], None, None]:
+        r"""
+        Returns an iterator over all elements with their names.
+        """
+        for elem_id, elem in self._id_to_elem.items():
+            yield elem_id.split(".")[-1], elem
+
+    def get_elem_by_id(self, elem_id: str) -> "Component":
+        r"""
+        Gets element by id.
+
         Example: top.lang, train.dataset
         """
-        tab_name, elem_name = name.split(".")
-        return self.all_elems[tab_name][elem_name]
+        return self._id_to_elem[elem_id]
+
+    def get_id_by_elem(self, elem: "Component") -> str:
+        r"""
+        Gets id by element.
+        """
+        return self._elem_to_id[elem]
 
     def get_base_elems(self) -> Set["Component"]:
+        r"""
+        Gets the base elements that are commonly used.
+        """
         return {
-            self.all_elems["top"]["lang"],
-            self.all_elems["top"]["model_name"],
-            self.all_elems["top"]["model_path"],
-            self.all_elems["top"]["adapter_path"],
-            self.all_elems["top"]["finetuning_type"],
-            self.all_elems["top"]["quantization_bit"],
-            self.all_elems["top"]["template"],
-            self.all_elems["top"]["rope_scaling"],
-            self.all_elems["top"]["booster"],
+            self._id_to_elem["top.lang"],
+            self._id_to_elem["top.model_name"],
+            self._id_to_elem["top.model_path"],
+            self._id_to_elem["top.finetuning_type"],
+            self._id_to_elem["top.adapter_path"],
+            self._id_to_elem["top.quantization_bit"],
+            self._id_to_elem["top.template"],
+            self._id_to_elem["top.rope_scaling"],
+            self._id_to_elem["top.booster"],
         }
-
-    def list_elems(self) -> List["Component"]:
-        return [elem for elems in self.all_elems.values() for elem in elems.values()]

src/llmtuner/webui/runner.py

@@ -2,7 +2,7 @@ import logging
 import os
 import time
 from threading import Thread
-from typing import TYPE_CHECKING, Any, Dict, Generator, Tuple
+from typing import TYPE_CHECKING, Any, Dict, Generator
 
 import gradio as gr
 import transformers
@@ -15,9 +15,9 @@ from ..extras.constants import TRAINING_STAGES
 from ..extras.logging import LoggerHandler
 from ..extras.misc import get_device_count, torch_gc
 from ..train import run_exp
-from .common import get_module, get_save_dir, load_config
+from .common import get_module, get_save_dir, load_args, load_config, save_args
 from .locales import ALERTS
-from .utils import gen_cmd, get_eval_results, update_process_bar
+from .utils import gen_cmd, gen_plot, get_eval_results, update_process_bar
 
 
 if TYPE_CHECKING:
@@ -48,8 +48,8 @@ class Runner:
     def set_abort(self) -> None:
         self.aborted = True
 
-    def _initialize(self, data: Dict[Component, Any], do_train: bool, from_preview: bool) -> str:
-        get = lambda name: data[self.manager.get_elem_by_name(name)]
+    def _initialize(self, data: Dict["Component", Any], do_train: bool, from_preview: bool) -> str:
+        get = lambda elem_id: data[self.manager.get_elem_by_id(elem_id)]
         lang, model_name, model_path = get("top.lang"), get("top.model_name"), get("top.model_path")
         dataset = get("train.dataset") if do_train else get("eval.dataset")
 
@@ -80,23 +80,21 @@ class Runner:
         if not from_preview and not is_torch_cuda_available():
             gr.Warning(ALERTS["warn_no_cuda"][lang])
 
-        self.aborted = False
         self.logger_handler.reset()
         self.trainer_callback = LogCallback(self)
         return ""
 
     def _finalize(self, lang: str, finish_info: str) -> str:
+        finish_info = ALERTS["info_aborted"][lang] if self.aborted else finish_info
         self.thread = None
-        self.running_data = None
+        self.aborted = False
         self.running = False
+        self.running_data = None
         torch_gc()
-        if self.aborted:
-            return ALERTS["info_aborted"][lang]
-        else:
         return finish_info
 
-    def _parse_train_args(self, data: Dict[Component, Any]) -> Dict[str, Any]:
-        get = lambda name: data[self.manager.get_elem_by_name(name)]
+    def _parse_train_args(self, data: Dict["Component", Any]) -> Dict[str, Any]:
+        get = lambda elem_id: data[self.manager.get_elem_by_id(elem_id)]
         user_config = load_config()
 
         if get("top.adapter_path"):
@@ -141,6 +139,7 @@ class Runner:
             upcast_layernorm=get("train.upcast_layernorm"),
             use_llama_pro=get("train.use_llama_pro"),
             shift_attn=get("train.shift_attn"),
+            report_to="all" if get("train.report_to") else "none",
             use_galore=get("train.use_galore"),
             output_dir=get_save_dir(get("top.model_name"), get("top.finetuning_type"), get("train.output_dir")),
             fp16=(get("train.compute_type") == "fp16"),
@@ -150,23 +149,21 @@ class Runner:
         args["disable_tqdm"] = True
 
         if args["finetuning_type"] == "freeze":
-            args["num_layer_trainable"] = int(get("train.num_layer_trainable"))
+            args["num_layer_trainable"] = get("train.num_layer_trainable")
             args["name_module_trainable"] = get("train.name_module_trainable")
         elif args["finetuning_type"] == "lora":
-            args["lora_rank"] = int(get("train.lora_rank"))
-            args["lora_alpha"] = int(get("train.lora_alpha"))
-            args["lora_dropout"] = float(get("train.lora_dropout"))
-            args["lora_target"] = get("train.lora_target") or get_module(get("top.model_name"))
+            args["lora_rank"] = get("train.lora_rank")
+            args["lora_alpha"] = get("train.lora_alpha")
+            args["lora_dropout"] = get("train.lora_dropout")
+            args["loraplus_lr_ratio"] = get("train.loraplus_lr_ratio") or None
+            args["create_new_adapter"] = get("train.create_new_adapter")
             args["use_rslora"] = get("train.use_rslora")
             args["use_dora"] = get("train.use_dora")
+            args["lora_target"] = get("train.lora_target") or get_module(get("top.model_name"))
             args["additional_target"] = get("train.additional_target") or None
-            if args["stage"] in ["rm", "ppo", "dpo"]:
-                args["create_new_adapter"] = args["quantization_bit"] is None
-            else:
-                args["create_new_adapter"] = get("train.create_new_adapter")
 
             if args["use_llama_pro"]:
-                args["num_layer_trainable"] = int(get("train.num_layer_trainable"))
+                args["num_layer_trainable"] = get("train.num_layer_trainable")
 
         if args["stage"] == "ppo":
             args["reward_model"] = ",".join(
@@ -176,10 +173,11 @@ class Runner:
                 ]
             )
             args["reward_model_type"] = "lora" if args["finetuning_type"] == "lora" else "full"
-
-        if args["stage"] == "dpo":
+        elif args["stage"] == "dpo":
             args["dpo_beta"] = get("train.dpo_beta")
             args["dpo_ftx"] = get("train.dpo_ftx")
+        elif args["stage"] == "orpo":
+            args["orpo_beta"] = get("train.orpo_beta")
 
         if get("train.val_size") > 1e-6 and args["stage"] != "ppo":
             args["val_size"] = get("train.val_size")
@@ -196,8 +194,8 @@ class Runner:
 
         return args
 
-    def _parse_eval_args(self, data: Dict[Component, Any]) -> Dict[str, Any]:
-        get = lambda name: data[self.manager.get_elem_by_name(name)]
+    def _parse_eval_args(self, data: Dict["Component", Any]) -> Dict[str, Any]:
+        get = lambda elem_id: data[self.manager.get_elem_by_id(elem_id)]
         user_config = load_config()
 
         if get("top.adapter_path"):
@@ -232,6 +230,7 @@ class Runner:
             temperature=get("eval.temperature"),
             output_dir=get_save_dir(get("top.model_name"), get("top.finetuning_type"), get("eval.output_dir")),
         )
+        args["disable_tqdm"] = True
 
         if get("eval.predict"):
             args["do_predict"] = True
@@ -240,22 +239,22 @@ class Runner:
 
         return args
 
-    def _preview(
-        self, data: Dict[Component, Any], do_train: bool
-    ) -> Generator[Tuple[str, Dict[str, Any]], None, None]:
+    def _preview(self, data: Dict["Component", Any], do_train: bool) -> Generator[Dict[Component, str], None, None]:
+        output_box = self.manager.get_elem_by_id("{}.output_box".format("train" if do_train else "eval"))
         error = self._initialize(data, do_train, from_preview=True)
         if error:
             gr.Warning(error)
-            yield error, gr.update(visible=False)
+            yield {output_box: error}
         else:
             args = self._parse_train_args(data) if do_train else self._parse_eval_args(data)
-            yield gen_cmd(args), gr.update(visible=False)
+            yield {output_box: gen_cmd(args)}
 
-    def _launch(self, data: Dict[Component, Any], do_train: bool) -> Generator[Tuple[str, Dict[str, Any]], None, None]:
+    def _launch(self, data: Dict["Component", Any], do_train: bool) -> Generator[Dict[Component, Any], None, None]:
+        output_box = self.manager.get_elem_by_id("{}.output_box".format("train" if do_train else "eval"))
         error = self._initialize(data, do_train, from_preview=False)
         if error:
             gr.Warning(error)
-            yield error, gr.update(visible=False)
+            yield {output_box: error}
         else:
             args = self._parse_train_args(data) if do_train else self._parse_eval_args(data)
             run_kwargs = dict(args=args, callbacks=[self.trainer_callback])
@@ -264,44 +263,103 @@ class Runner:
             self.thread.start()
             yield from self.monitor()
 
-    def preview_train(self, data: Dict[Component, Any]) -> Generator[Tuple[str, Dict[str, Any]], None, None]:
+    def preview_train(self, data: Dict[Component, Any]) -> Generator[Dict[Component, str], None, None]:
         yield from self._preview(data, do_train=True)
 
-    def preview_eval(self, data: Dict[Component, Any]) -> Generator[Tuple[str, Dict[str, Any]], None, None]:
+    def preview_eval(self, data: Dict[Component, Any]) -> Generator[Dict[Component, str], None, None]:
         yield from self._preview(data, do_train=False)
 
-    def run_train(self, data: Dict[Component, Any]) -> Generator[Tuple[str, Dict[str, Any]], None, None]:
+    def run_train(self, data: Dict[Component, Any]) -> Generator[Dict[Component, Any], None, None]:
         yield from self._launch(data, do_train=True)
 
-    def run_eval(self, data: Dict[Component, Any]) -> Generator[Tuple[str, Dict[str, Any]], None, None]:
+    def run_eval(self, data: Dict[Component, Any]) -> Generator[Dict[Component, Any], None, None]:
         yield from self._launch(data, do_train=False)
 
-    def monitor(self) -> Generator[Tuple[str, Dict[str, Any]], None, None]:
-        get = lambda name: self.running_data[self.manager.get_elem_by_name(name)]
+    def monitor(self) -> Generator[Dict[Component, Any], None, None]:
+        get = lambda elem_id: self.running_data[self.manager.get_elem_by_id(elem_id)]
+        self.aborted = False
         self.running = True
-        lang = get("top.lang")
-        output_dir = get_save_dir(
-            get("top.model_name"),
-            get("top.finetuning_type"),
-            get("{}.output_dir".format("train" if self.do_train else "eval")),
-        )
 
-        while self.thread.is_alive():
-            time.sleep(2)
+        lang = get("top.lang")
+        model_name = get("top.model_name")
+        finetuning_type = get("top.finetuning_type")
+        output_dir = get("{}.output_dir".format("train" if self.do_train else "eval"))
+        output_path = get_save_dir(model_name, finetuning_type, output_dir)
+
+        output_box = self.manager.get_elem_by_id("{}.output_box".format("train" if self.do_train else "eval"))
+        process_bar = self.manager.get_elem_by_id("{}.process_bar".format("train" if self.do_train else "eval"))
+        loss_viewer = self.manager.get_elem_by_id("train.loss_viewer") if self.do_train else None
+
+        while self.thread is not None and self.thread.is_alive():
             if self.aborted:
-                yield ALERTS["info_aborting"][lang], gr.update(visible=False)
+                yield {
+                    output_box: ALERTS["info_aborting"][lang],
+                    process_bar: gr.Slider(visible=False),
+                }
             else:
-                yield self.logger_handler.log, update_process_bar(self.trainer_callback)
+                return_dict = {
+                    output_box: self.logger_handler.log,
+                    process_bar: update_process_bar(self.trainer_callback),
+                }
+                if self.do_train:
+                    plot = gen_plot(output_path)
+                    if plot is not None:
+                        return_dict[loss_viewer] = plot
+
+                yield return_dict
+
+            time.sleep(2)
 
         if self.do_train:
-            if os.path.exists(os.path.join(output_dir, TRAINING_ARGS_NAME)):
+            if os.path.exists(os.path.join(output_path, TRAINING_ARGS_NAME)):
                 finish_info = ALERTS["info_finished"][lang]
             else:
                 finish_info = ALERTS["err_failed"][lang]
         else:
-            if os.path.exists(os.path.join(output_dir, "all_results.json")):
-                finish_info = get_eval_results(os.path.join(output_dir, "all_results.json"))
+            if os.path.exists(os.path.join(output_path, "all_results.json")):
+                finish_info = get_eval_results(os.path.join(output_path, "all_results.json"))
             else:
                 finish_info = ALERTS["err_failed"][lang]
 
-        yield self._finalize(lang, finish_info), gr.update(visible=False)
+        return_dict = {
+            output_box: self._finalize(lang, finish_info),
+            process_bar: gr.Slider(visible=False),
+        }
+        if self.do_train:
+            plot = gen_plot(output_path)
+            if plot is not None:
+                return_dict[loss_viewer] = plot
+
+        yield return_dict
+
+    def save_args(self, data: Dict[Component, Any]) -> Dict[Component, str]:
+        output_box = self.manager.get_elem_by_id("train.output_box")
+        error = self._initialize(data, do_train=True, from_preview=True)
+        if error:
+            gr.Warning(error)
+            return {output_box: error}
+
+        config_dict: Dict[str, Any] = {}
+        lang = data[self.manager.get_elem_by_id("top.lang")]
+        config_path = data[self.manager.get_elem_by_id("train.config_path")]
+        skip_ids = ["top.lang", "top.model_path", "train.output_dir", "train.config_path"]
+        for elem, value in data.items():
+            elem_id = self.manager.get_id_by_elem(elem)
+            if elem_id not in skip_ids:
+                config_dict[elem_id] = value
+
+        save_path = save_args(config_path, config_dict)
+        return {output_box: ALERTS["info_config_saved"][lang] + save_path}
+
+    def load_args(self, lang: str, config_path: str) -> Dict[Component, Any]:
+        output_box = self.manager.get_elem_by_id("train.output_box")
+        config_dict = load_args(config_path)
+        if config_dict is None:
+            gr.Warning(ALERTS["err_config_not_found"][lang])
+            return {output_box: ALERTS["err_config_not_found"][lang]}
+
+        output_dict: Dict["Component", Any] = {output_box: ALERTS["info_config_loaded"][lang]}
+        for elem_id, value in config_dict.items():
+            output_dict[self.manager.get_elem_by_id(elem_id)] = value
+
+        return output_dict

src/llmtuner/webui/utils.py

@@ -1,13 +1,12 @@
 import json
 import os
 from datetime import datetime
-from typing import TYPE_CHECKING, Any, Dict
+from typing import TYPE_CHECKING, Any, Dict, Optional
 
 import gradio as gr
 
 from ..extras.packages import is_matplotlib_available
 from ..extras.ploting import smooth
-from .common import get_save_dir
 from .locales import ALERTS
 
 
@@ -19,26 +18,26 @@ if is_matplotlib_available():
     import matplotlib.pyplot as plt
 
 
-def update_process_bar(callback: "LogCallback") -> Dict[str, Any]:
+def update_process_bar(callback: "LogCallback") -> "gr.Slider":
     if not callback.max_steps:
-        return gr.update(visible=False)
+        return gr.Slider(visible=False)
 
     percentage = round(100 * callback.cur_steps / callback.max_steps, 0) if callback.max_steps != 0 else 100.0
     label = "Running {:d}/{:d}: {} < {}".format(
         callback.cur_steps, callback.max_steps, callback.elapsed_time, callback.remaining_time
     )
-    return gr.update(label=label, value=percentage, visible=True)
+    return gr.Slider(label=label, value=percentage, visible=True)
 
 
 def get_time() -> str:
-    return datetime.now().strftime("%Y-%m-%d-%H-%M-%S")
+    return datetime.now().strftime(r"%Y-%m-%d-%H-%M-%S")
 
 
-def can_quantize(finetuning_type: str) -> Dict[str, Any]:
+def can_quantize(finetuning_type: str) -> "gr.Dropdown":
     if finetuning_type != "lora":
-        return gr.update(value="None", interactive=False)
+        return gr.Dropdown(value="none", interactive=False)
     else:
-        return gr.update(interactive=True)
+        return gr.Dropdown(interactive=True)
 
 
 def check_json_schema(text: str, lang: str) -> None:
@@ -48,8 +47,8 @@ def check_json_schema(text: str, lang: str) -> None:
             assert isinstance(tools, list)
             for tool in tools:
                 if "name" not in tool:
-                    raise ValueError("Name not found.")
-    except ValueError:
+                    raise NotImplementedError("Name not found.")
+    except NotImplementedError:
         gr.Warning(ALERTS["err_tool_name"][lang])
     except Exception:
         gr.Warning(ALERTS["err_json_schema"][lang])
@@ -74,11 +73,9 @@ def get_eval_results(path: os.PathLike) -> str:
     return "```json\n{}\n```\n".format(result)
 
 
-def gen_plot(base_model: str, finetuning_type: str, output_dir: str) -> "matplotlib.figure.Figure":
-    if not base_model:
-        return
-    log_file = get_save_dir(base_model, finetuning_type, output_dir, "trainer_log.jsonl")
-    if not os.path.isfile(log_file):
+def gen_plot(output_path: str) -> Optional["matplotlib.figure.Figure"]:
+    log_file = os.path.join(output_path, "trainer_log.jsonl")
+    if not os.path.isfile(log_file) or not is_matplotlib_available():
         return
 
     plt.close("all")
@@ -88,13 +85,13 @@ def gen_plot(base_model: str, finetuning_type: str, output_dir: str) -> "matplot
     steps, losses = [], []
     with open(log_file, "r", encoding="utf-8") as f:
         for line in f:
-            log_info = json.loads(line)
+            log_info: Dict[str, Any] = json.loads(line)
             if log_info.get("loss", None):
                 steps.append(log_info["current_steps"])
                 losses.append(log_info["loss"])
 
     if len(losses) == 0:
-        return None
+        return
 
     ax.plot(steps, losses, color="#1f77b4", alpha=0.4, label="original")
     ax.plot(steps, smooth(losses), color="#1f77b4", label="smoothed")

src/train_web.py

@@ -2,9 +2,7 @@ from llmtuner import create_ui
 
 
 def main():
-    demo = create_ui()
-    demo.queue()
-    demo.launch(server_name="0.0.0.0", share=False, inbrowser=True)
+    create_ui().queue().launch(server_name="0.0.0.0", server_port=None, share=False, inbrowser=True)
 
 
 if __name__ == "__main__":

src/web_demo.py

@@ -2,9 +2,7 @@ from llmtuner import create_web_demo
 
 
 def main():
-    demo = create_web_demo()
-    demo.queue()
-    demo.launch(server_name="0.0.0.0", share=False, inbrowser=True)
+    create_web_demo().queue().launch(server_name="0.0.0.0", server_port=None, share=False, inbrowser=True)
 
 
 if __name__ == "__main__":

tests/test_toolcall.py

@@ -15,7 +15,7 @@ def calculate_gpa(grades: Sequence[str], hours: Sequence[int]) -> float:
     for grade, hour in zip(grades, hours):
         total_score += grade_to_score[grade] * hour
         total_hour += hour
-    return total_score / total_hour
+    return round(total_score / total_hour, 2)
 
 
 def main():
@@ -45,16 +45,19 @@ def main():
     messages = []
     messages.append({"role": "user", "content": "My grades are A, A, B, and C. The credit hours are 3, 4, 3, and 2."})
     result = client.chat.completions.create(messages=messages, model="test", tools=tools)
+    if result.choices[0].message.tool_calls is None:
+        raise ValueError("Cannot retrieve function call from the response.")
+
+    messages.append(result.choices[0].message)
     tool_call = result.choices[0].message.tool_calls[0].function
+    print(tool_call)
+    # Function(arguments='{"grades": ["A", "A", "B", "C"], "hours": [3, 4, 3, 2]}', name='calculate_gpa')
     name, arguments = tool_call.name, json.loads(tool_call.arguments)
-    messages.append(
-        {"role": "function", "content": json.dumps({"name": name, "argument": arguments}, ensure_ascii=False)}
-    )
     tool_result = tool_map[name](**arguments)
     messages.append({"role": "tool", "content": json.dumps({"gpa": tool_result}, ensure_ascii=False)})
     result = client.chat.completions.create(messages=messages, model="test", tools=tools)
     print(result.choices[0].message.content)
-    # Based on your grades and credit hours, your calculated Grade Point Average (GPA) is 3.4166666666666665.
+    # Based on the grades and credit hours you provided, your Grade Point Average (GPA) is 3.42.
 
 
 if __name__ == "__main__":