p04798526
/
LLaMA-Factory-Mirror
1
0
Fork 1
Code Issues Pull Requests Packages Projects Releases Wiki Activity

update examples

This commit is contained in:
hiyouga 2024-06-13 03:15:06 +08:00
parent 6baafd4eb3
commit b6e008c152
27 changed files with 128 additions and 199 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
README.md
View File

@ -406,7 +406,7 @@ Please refer to [data/README.md](data/README.md) for checking the details about
Use the following 3 commands to run LoRA **fine-tuning**, **inference** and **merging** of the Llama3-8B-Instruct model, respectively. Use the following 3 commands to run LoRA **fine-tuning**, **inference** and **merging** of the Llama3-8B-Instruct model, respectively.
```bash ```bash
llamafactory-cli train examples/lora_single_gpu/llama3_lora_sft.yaml llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
llamafactory-cli chat examples/inference/llama3_lora_sft.yaml llamafactory-cli chat examples/inference/llama3_lora_sft.yaml
llamafactory-cli export examples/merge_lora/llama3_lora_sft.yaml llamafactory-cli export examples/merge_lora/llama3_lora_sft.yaml
``` ```

2
README_zh.md
View File

@ -406,7 +406,7 @@ Docker 镜像：
下面三行命令分别对 Llama3-8B-Instruct 模型进行 LoRA **微调**、**推理**和**合并**。 下面三行命令分别对 Llama3-8B-Instruct 模型进行 LoRA **微调**、**推理**和**合并**。
```bash ```bash
llamafactory-cli train examples/lora_single_gpu/llama3_lora_sft.yaml llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
llamafactory-cli chat examples/inference/llama3_lora_sft.yaml llamafactory-cli chat examples/inference/llama3_lora_sft.yaml
llamafactory-cli export examples/merge_lora/llama3_lora_sft.yaml llamafactory-cli export examples/merge_lora/llama3_lora_sft.yaml
``` ```

126
examples/README.md
View File

@ -4,59 +4,57 @@ Make sure to execute these commands in the `LLaMA-Factory` directory.
## Table of Contents ## Table of Contents
- [LoRA Fine-Tuning on A Single GPU](#lora-fine-tuning-on-a-single-gpu) - [LoRA Fine-Tuning](#lora-fine-tuning)
- [QLoRA Fine-Tuning on a Single GPU](#qlora-fine-tuning-on-a-single-gpu) - [QLoRA Fine-Tuning](#qlora-fine-tuning)
- [LoRA Fine-Tuning on Multiple GPUs](#lora-fine-tuning-on-multiple-gpus) - [Full-Parameter Fine-Tuning](#full-parameter-fine-tuning)
- [LoRA Fine-Tuning on Multiple NPUs](#lora-fine-tuning-on-multiple-npus)
- [Full-Parameter Fine-Tuning on Multiple GPUs](#full-parameter-fine-tuning-on-multiple-gpus)
- [Merging LoRA Adapters and Quantization](#merging-lora-adapters-and-quantization) - [Merging LoRA Adapters and Quantization](#merging-lora-adapters-and-quantization)
- [Inferring LoRA Fine-Tuned Models](#inferring-lora-fine-tuned-models) - [Inferring LoRA Fine-Tuned Models](#inferring-lora-fine-tuned-models)
- [Extras](#extras) - [Extras](#extras)
## Examples ## Examples
### LoRA Fine-Tuning on A Single GPU ### LoRA Fine-Tuning
#### (Continuous) Pre-Training #### (Continuous) Pre-Training
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_pretrain.yaml llamafactory-cli train examples/train_lora/llama3_lora_pretrain.yaml
``` ```
#### Supervised Fine-Tuning #### Supervised Fine-Tuning
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_sft.yaml llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
``` ```
#### Multimodal Supervised Fine-Tuning #### Multimodal Supervised Fine-Tuning
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llava1_5_lora_sft.yaml llamafactory-cli train examples/train_lora/llava1_5_lora_sft.yaml
``` ```
#### Reward Modeling #### Reward Modeling
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_reward.yaml llamafactory-cli train examples/train_lora/llama3_lora_reward.yaml
``` ```
#### PPO Training #### PPO Training
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_ppo.yaml llamafactory-cli train examples/train_lora/llama3_lora_ppo.yaml
``` ```
#### DPO/ORPO/SimPO Training #### DPO/ORPO/SimPO Training
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_dpo.yaml llamafactory-cli train examples/train_lora/llama3_lora_dpo.yaml
``` ```
#### KTO Training #### KTO Training
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_kto.yaml llamafactory-cli train examples/train_lora/llama3_lora_kto.yaml
``` ```
#### Preprocess Dataset #### Preprocess Dataset
@ -64,95 +62,79 @@ CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lo
It is useful for large dataset, use `tokenized_path` in config to load the preprocessed dataset. It is useful for large dataset, use `tokenized_path` in config to load the preprocessed dataset.
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_preprocess.yaml llamafactory-cli train examples/train_lora/llama3_preprocess.yaml
``` ```
#### Evaluating on MMLU/CMMLU/C-Eval Benchmarks #### Evaluating on MMLU/CMMLU/C-Eval Benchmarks
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli eval examples/lora_single_gpu/llama3_lora_eval.yaml llamafactory-cli eval examples/train_lora/llama3_lora_eval.yaml
``` ```
#### Batch Predicting and Computing BLEU and ROUGE Scores #### Batch Predicting and Computing BLEU and ROUGE Scores
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_predict.yaml llamafactory-cli train examples/train_lora/llama3_lora_predict.yaml
```
### QLoRA Fine-Tuning on a Single GPU
#### Supervised Fine-Tuning with 4/8-bit Bitsandbytes Quantization (Recommended)
```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/qlora_single_gpu/llama3_lora_sft_bitsandbytes.yaml
```
#### Supervised Fine-Tuning with 4/8-bit GPTQ Quantization
```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/qlora_single_gpu/llama3_lora_sft_gptq.yaml
```
#### Supervised Fine-Tuning with 4-bit AWQ Quantization
```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/qlora_single_gpu/llama3_lora_sft_awq.yaml
```
#### Supervised Fine-Tuning with 2-bit AQLM Quantization
```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/qlora_single_gpu/llama3_lora_sft_aqlm.yaml
```
### LoRA Fine-Tuning on Multiple GPUs
#### Supervised Fine-Tuning on Single Node
```bash
CUDA_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/lora_multi_gpu/llama3_lora_sft.yaml
``` ```
#### Supervised Fine-Tuning on Multiple Nodes #### Supervised Fine-Tuning on Multiple Nodes
```bash ```bash
CUDA_VISIBLE_DEVICES=0,1,2,3 NNODES=2 RANK=0 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/lora_multi_gpu/llama3_lora_sft.yaml FORCE_TORCHRUN=1 NNODES=2 RANK=0 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
CUDA_VISIBLE_DEVICES=0,1,2,3 NNODES=2 RANK=1 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/lora_multi_gpu/llama3_lora_sft.yaml FORCE_TORCHRUN=1 NNODES=2 RANK=1 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
``` ```
#### Supervised Fine-Tuning with DeepSpeed ZeRO-3 (Weight Sharding) #### Supervised Fine-Tuning with DeepSpeed ZeRO-3 (Weight Sharding)
```bash ```bash
CUDA_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/lora_multi_gpu/llama3_lora_sft_ds.yaml FORCE_TORCHRUN=1 llamafactory-cli train examples/train_lora/llama3_lora_sft_ds.yaml
``` ```
### LoRA Fine-Tuning on Multiple NPUs ### QLoRA Fine-Tuning
#### Supervised Fine-Tuning with DeepSpeed ZeRO-0 #### Supervised Fine-Tuning with 4/8-bit Bitsandbytes Quantization (Recommended)
```bash ```bash
ASCEND_RT_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/lora_multi_npu/llama3_lora_sft_ds.yaml CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/train_qlora/llama3_lora_sft_bitsandbytes.yaml
``` ```
### Full-Parameter Fine-Tuning on Multiple GPUs #### Supervised Fine-Tuning with 4/8-bit GPTQ Quantization
```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/train_qlora/llama3_lora_sft_gptq.yaml
```
#### Supervised Fine-Tuning with 4-bit AWQ Quantization
```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/train_qlora/llama3_lora_sft_awq.yaml
```
#### Supervised Fine-Tuning with 2-bit AQLM Quantization
```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/train_qlora/llama3_lora_sft_aqlm.yaml
```
### Full-Parameter Fine-Tuning
#### Supervised Fine-Tuning on Single Node #### Supervised Fine-Tuning on Single Node
```bash ```bash
CUDA_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/full_multi_gpu/llama3_full_sft.yaml FORCE_TORCHRUN=1 llamafactory-cli train examples/train_full/llama3_full_sft.yaml
``` ```
#### Supervised Fine-Tuning on Multiple Nodes #### Supervised Fine-Tuning on Multiple Nodes
```bash ```bash
CUDA_VISIBLE_DEVICES=0,1,2,3 NNODES=2 RANK=0 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/full_multi_gpu/llama3_full_sft.yaml FORCE_TORCHRUN=1 NNODES=2 RANK=0 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/train_full/llama3_full_sft.yaml
CUDA_VISIBLE_DEVICES=0,1,2,3 NNODES=2 RANK=1 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/full_multi_gpu/llama3_full_sft.yaml FORCE_TORCHRUN=1 NNODES=2 RANK=1 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/train_full/llama3_full_sft.yaml
``` ```
#### Batch Predicting and Computing BLEU and ROUGE Scores #### Batch Predicting and Computing BLEU and ROUGE Scores
```bash ```bash
CUDA_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/full_multi_gpu/llama3_full_predict.yaml llamafactory-cli train examples/train_full/llama3_full_predict.yaml
``` ```
### Merging LoRA Adapters and Quantization ### Merging LoRA Adapters and Quantization
@ -162,35 +144,33 @@ CUDA_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/full_multi_gpu/llam
Note: DO NOT use quantized model or `quantization_bit` when merging LoRA adapters. Note: DO NOT use quantized model or `quantization_bit` when merging LoRA adapters.
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli export examples/merge_lora/llama3_lora_sft.yaml llamafactory-cli export examples/merge_lora/llama3_lora_sft.yaml
``` ```
#### Quantizing Model using AutoGPTQ #### Quantizing Model using AutoGPTQ
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli export examples/merge_lora/llama3_gptq.yaml llamafactory-cli export examples/merge_lora/llama3_gptq.yaml
``` ```
### Inferring LoRA Fine-Tuned Models ### Inferring LoRA Fine-Tuned Models
Use `CUDA_VISIBLE_DEVICES=0,1` to infer models on multiple devices.
#### Use CLI #### Use CLI
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli chat examples/inference/llama3_lora_sft.yaml llamafactory-cli chat examples/inference/llama3_lora_sft.yaml
``` ```
#### Use Web UI #### Use Web UI
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli webchat examples/inference/llama3_lora_sft.yaml llamafactory-cli webchat examples/inference/llama3_lora_sft.yaml
``` ```
#### Launch OpenAI-style API #### Launch OpenAI-style API
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli api examples/inference/llama3_lora_sft.yaml llamafactory-cli api examples/inference/llama3_lora_sft.yaml
``` ```
### Extras ### Extras
@ -198,32 +178,32 @@ CUDA_VISIBLE_DEVICES=0 llamafactory-cli api examples/inference/llama3_lora_sft.y
#### Full-Parameter Fine-Tuning using GaLore #### Full-Parameter Fine-Tuning using GaLore
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/galore/llama3_full_sft.yaml llamafactory-cli train examples/extras/galore/llama3_full_sft.yaml
``` ```
#### Full-Parameter Fine-Tuning using BAdam #### Full-Parameter Fine-Tuning using BAdam
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/badam/llama3_full_sft.yaml llamafactory-cli train examples/extras/badam/llama3_full_sft.yaml
``` ```
#### LoRA+ Fine-Tuning #### LoRA+ Fine-Tuning
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/loraplus/llama3_lora_sft.yaml llamafactory-cli train examples/extras/loraplus/llama3_lora_sft.yaml
``` ```
#### Mixture-of-Depths Fine-Tuning #### Mixture-of-Depths Fine-Tuning
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/mod/llama3_full_sft.yaml llamafactory-cli train examples/extras/mod/llama3_full_sft.yaml
``` ```
#### LLaMA-Pro Fine-Tuning #### LLaMA-Pro Fine-Tuning
```bash ```bash
bash examples/extras/llama_pro/expand.sh bash examples/extras/llama_pro/expand.sh
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/llama_pro/llama3_freeze_sft.yaml llamafactory-cli train examples/extras/llama_pro/llama3_freeze_sft.yaml
``` ```
#### FSDP+QLoRA Fine-Tuning #### FSDP+QLoRA Fine-Tuning

128
examples/README_zh.md
View File

@ -4,59 +4,57 @@
## 目录 ## 目录
- [单 GPU LoRA 微调](#单-gpu-lora-微调) - [LoRA 微调](#lora-微调)
- [单 GPU QLoRA 微调](#单-gpu-qlora-微调) - [QLoRA 微调](#qlora-微调)
- [多 GPU LoRA 微调](#多-gpu-lora-微调) - [全参数微调](#全参数微调)
- [多 NPU LoRA 微调](#多-npu-lora-微调)
- [多 GPU 全参数微调](#多-gpu-全参数微调)
- [合并 LoRA 适配器与模型量化](#合并-lora-适配器与模型量化) - [合并 LoRA 适配器与模型量化](#合并-lora-适配器与模型量化)
- [推理 LoRA 模型](#推理-lora-模型) - [推理 LoRA 模型](#推理-lora-模型)
- [杂项](#杂项) - [杂项](#杂项)
## 示例 ## 示例
### 单 GPU LoRA 微调 ### LoRA 微调
#### （增量）预训练 #### （增量）预训练
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_pretrain.yaml llamafactory-cli train examples/train_lora/llama3_lora_pretrain.yaml
``` ```
#### 指令监督微调 #### 指令监督微调
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_sft.yaml llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
``` ```
#### 多模态指令监督微调 #### 多模态指令监督微调
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llava1_5_lora_sft.yaml llamafactory-cli train examples/train_lora/llava1_5_lora_sft.yaml
``` ```
#### 奖励模型训练 #### 奖励模型训练
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_reward.yaml llamafactory-cli train examples/train_lora/llama3_lora_reward.yaml
``` ```
#### PPO 训练 #### PPO 训练
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_ppo.yaml llamafactory-cli train examples/train_lora/llama3_lora_ppo.yaml
``` ```
#### DPO/ORPO/SimPO 训练 #### DPO/ORPO/SimPO 训练
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_dpo.yaml llamafactory-cli train examples/train_lora/llama3_lora_dpo.yaml
``` ```
#### KTO 训练 #### KTO 训练
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_kto.yaml llamafactory-cli train examples/train_lora/llama3_lora_kto.yaml
``` ```
#### 预处理数据集 #### 预处理数据集
@ -64,95 +62,79 @@ CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lo
对于大数据集有帮助，在配置中使用 `tokenized_path` 以加载预处理后的数据集。 对于大数据集有帮助，在配置中使用 `tokenized_path` 以加载预处理后的数据集。
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_preprocess.yaml llamafactory-cli train examples/train_lora/llama3_preprocess.yaml
``` ```
#### 在 MMLU/CMMLU/C-Eval 上评估 #### 在 MMLU/CMMLU/C-Eval 上评估
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli eval examples/lora_single_gpu/llama3_lora_eval.yaml llamafactory-cli eval examples/train_lora/llama3_lora_eval.yaml
``` ```
#### 批量预测并计算 BLEU 和 ROUGE 分数 #### 批量预测并计算 BLEU 和 ROUGE 分数
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_predict.yaml llamafactory-cli train examples/train_lora/llama3_lora_predict.yaml
``` ```
### 单 GPU QLoRA 微调 #### 多机指令监督微调
#### 基于 4/8 比特 Bitsandbytes 量化进行指令监督微调（推荐）
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/qlora_single_gpu/llama3_lora_sft_bitsandbytes.yaml FORCE_TORCHRUN=1 NNODES=2 RANK=0 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
``` FORCE_TORCHRUN=1 NNODES=2 RANK=1 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
#### 基于 4/8 比特 GPTQ 量化进行指令监督微调
```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/qlora_single_gpu/llama3_lora_sft_gptq.yaml
```
#### 基于 4 比特 AWQ 量化进行指令监督微调
```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/qlora_single_gpu/llama3_lora_sft_awq.yaml
```
#### 基于 2 比特 AQLM 量化进行指令监督微调
```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/qlora_single_gpu/llama3_lora_sft_aqlm.yaml
```
### 多 GPU LoRA 微调
#### 在单机上进行指令监督微调
```bash
CUDA_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/lora_multi_gpu/llama3_lora_sft.yaml
```
#### 在多机上进行指令监督微调
```bash
CUDA_VISIBLE_DEVICES=0,1,2,3 NNODES=2 RANK=0 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/lora_multi_gpu/llama3_lora_sft.yaml
CUDA_VISIBLE_DEVICES=0,1,2,3 NNODES=2 RANK=1 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/lora_multi_gpu/llama3_lora_sft.yaml
``` ```
#### 使用 DeepSpeed ZeRO-3 平均分配显存 #### 使用 DeepSpeed ZeRO-3 平均分配显存
```bash ```bash
CUDA_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/lora_multi_gpu/llama3_lora_sft_ds.yaml FORCE_TORCHRUN=1 llamafactory-cli train examples/train_lora/llama3_lora_sft_ds.yaml
``` ```
### 多 NPU LoRA 微调 ### QLoRA 微调
#### 使用 DeepSpeed ZeRO-0 进行指令监督微调 #### 基于 4/8 比特 Bitsandbytes 量化进行指令监督微调（推荐）
```bash ```bash
ASCEND_RT_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/lora_multi_npu/llama3_lora_sft_ds.yaml llamafactory-cli train examples/train_qlora/llama3_lora_sft_bitsandbytes.yaml
``` ```
### 多 GPU 全参数微调 #### 基于 4/8 比特 GPTQ 量化进行指令监督微调
```bash
llamafactory-cli train examples/train_qlora/llama3_lora_sft_gptq.yaml
```
#### 基于 4 比特 AWQ 量化进行指令监督微调
```bash
llamafactory-cli train examples/train_qlora/llama3_lora_sft_awq.yaml
```
#### 基于 2 比特 AQLM 量化进行指令监督微调
```bash
llamafactory-cli train examples/train_qlora/llama3_lora_sft_aqlm.yaml
```
### 全参数微调
#### 在单机上进行指令监督微调 #### 在单机上进行指令监督微调
```bash ```bash
CUDA_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/full_multi_gpu/llama3_full_sft.yaml FORCE_TORCHRUN=1 llamafactory-cli train examples/train_full/llama3_full_sft.yaml
``` ```
#### 在多机上进行指令监督微调 #### 在多机上进行指令监督微调
```bash ```bash
CUDA_VISIBLE_DEVICES=0,1,2,3 NNODES=2 RANK=0 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/full_multi_gpu/llama3_full_sft.yaml FORCE_TORCHRUN=1 NNODES=2 RANK=0 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/train_full/llama3_full_sft.yaml
CUDA_VISIBLE_DEVICES=0,1,2,3 NNODES=2 RANK=1 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/full_multi_gpu/llama3_full_sft.yaml FORCE_TORCHRUN=1 NNODES=2 RANK=1 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/train_full/llama3_full_sft.yaml
``` ```
#### 批量预测并计算 BLEU 和 ROUGE 分数 #### 批量预测并计算 BLEU 和 ROUGE 分数
```bash ```bash
CUDA_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/full_multi_gpu/llama3_full_predict.yaml llamafactory-cli train examples/train_full/llama3_full_predict.yaml
``` ```
### 合并 LoRA 适配器与模型量化 ### 合并 LoRA 适配器与模型量化
@ -162,35 +144,33 @@ CUDA_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/full_multi_gpu/llam
注：请勿使用量化后的模型或 `quantization_bit` 参数来合并 LoRA 适配器。 注：请勿使用量化后的模型或 `quantization_bit` 参数来合并 LoRA 适配器。
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli export examples/merge_lora/llama3_lora_sft.yaml llamafactory-cli export examples/merge_lora/llama3_lora_sft.yaml
``` ```
#### 使用 AutoGPTQ 量化模型 #### 使用 AutoGPTQ 量化模型
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli export examples/merge_lora/llama3_gptq.yaml llamafactory-cli export examples/merge_lora/llama3_gptq.yaml
``` ```
### 推理 LoRA 模型 ### 推理 LoRA 模型
使用 `CUDA_VISIBLE_DEVICES=0,1` 进行多卡推理。
#### 使用命令行接口 #### 使用命令行接口
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli chat examples/inference/llama3_lora_sft.yaml llamafactory-cli chat examples/inference/llama3_lora_sft.yaml
``` ```
#### 使用浏览器界面 #### 使用浏览器界面
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli webchat examples/inference/llama3_lora_sft.yaml llamafactory-cli webchat examples/inference/llama3_lora_sft.yaml
``` ```
#### 启动 OpenAI 风格 API #### 启动 OpenAI 风格 API
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli api examples/inference/llama3_lora_sft.yaml llamafactory-cli api examples/inference/llama3_lora_sft.yaml
``` ```
### 杂项 ### 杂项
@ -198,32 +178,32 @@ CUDA_VISIBLE_DEVICES=0 llamafactory-cli api examples/inference/llama3_lora_sft.y
#### 使用 GaLore 进行全参数训练 #### 使用 GaLore 进行全参数训练
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/galore/llama3_full_sft.yaml llamafactory-cli train examples/extras/galore/llama3_full_sft.yaml
``` ```
#### 使用 BAdam 进行全参数训练 #### 使用 BAdam 进行全参数训练
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/badam/llama3_full_sft.yaml llamafactory-cli train examples/extras/badam/llama3_full_sft.yaml
``` ```
#### LoRA+ 微调 #### LoRA+ 微调
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/loraplus/llama3_lora_sft.yaml llamafactory-cli train examples/extras/loraplus/llama3_lora_sft.yaml
``` ```
#### 深度混合微调 #### 深度混合微调
```bash ```bash
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/mod/llama3_full_sft.yaml llamafactory-cli train examples/extras/mod/llama3_full_sft.yaml
``` ```
#### LLaMA-Pro 微调 #### LLaMA-Pro 微调
```bash ```bash
bash examples/extras/llama_pro/expand.sh bash examples/extras/llama_pro/expand.sh
CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/llama_pro/llama3_freeze_sft.yaml llamafactory-cli train examples/extras/llama_pro/llama3_freeze_sft.yaml
``` ```
#### FSDP+QLoRA 微调 #### FSDP+QLoRA 微调

4
examples/extras/fsdp_qlora/llama3_lora_sft.yaml
View File

@ -8,9 +8,6 @@ do_train: true
finetuning_type: lora finetuning_type: lora
lora_target: all lora_target: all
### ddp
ddp_timeout: 180000000
### dataset ### dataset
dataset: identity,alpaca_en_demo dataset: identity,alpaca_en_demo
template: llama3 template: llama3
@ -34,6 +31,7 @@ num_train_epochs: 3.0
lr_scheduler_type: cosine lr_scheduler_type: cosine
warmup_ratio: 0.1 warmup_ratio: 0.1
fp16: true fp16: true
ddp_timeout: 180000000
### eval ### eval
val_size: 0.1 val_size: 0.1

1
examples/extras/llama_pro/llama3_freeze_sft.yaml
View File

@ -32,6 +32,7 @@ num_train_epochs: 3.0
lr_scheduler_type: cosine lr_scheduler_type: cosine
warmup_ratio: 0.1 warmup_ratio: 0.1
fp16: true fp16: true
ddp_timeout: 180000000
### eval ### eval
val_size: 0.1 val_size: 0.1

1
examples/extras/loraplus/llama3_lora_sft.yaml
View File

@ -31,6 +31,7 @@ num_train_epochs: 3.0
lr_scheduler_type: cosine lr_scheduler_type: cosine
warmup_ratio: 0.1 warmup_ratio: 0.1
fp16: true fp16: true
ddp_timeout: 180000000
### eval ### eval
val_size: 0.1 val_size: 0.1

1
examples/extras/mod/llama3_full_sft.yaml
View File

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

41
examples/lora_multi_gpu/llama3_lora_sft.yaml
View File

@ -1,41 +0,0 @@
### model
model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
### method
stage: sft
do_train: true
finetuning_type: lora
lora_target: all
### ddp
ddp_timeout: 180000000
### dataset
dataset: identity,alpaca_en_demo
template: llama3
cutoff_len: 1024
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16
### output
output_dir: saves/llama3-8b/lora/sft
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true
### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 2
learning_rate: 1.0e-4
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
fp16: true
### eval
val_size: 0.1
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500

0
examples/full_multi_gpu/llama3_full_predict.yaml → examples/train_full/llama3_full_predict.yaml
View File

0
examples/full_multi_gpu/llama3_full_sft.yaml → examples/train_full/llama3_full_sft_ds3.yaml
View File

1
examples/lora_single_gpu/llama3_lora_dpo.yaml → examples/train_lora/llama3_lora_dpo.yaml
View File

@ -32,6 +32,7 @@ num_train_epochs: 3.0
lr_scheduler_type: cosine lr_scheduler_type: cosine
warmup_ratio: 0.1 warmup_ratio: 0.1
fp16: true fp16: true
ddp_timeout: 180000000
### eval ### eval
val_size: 0.1 val_size: 0.1

0
examples/lora_single_gpu/llama3_lora_eval.yaml → examples/train_lora/llama3_lora_eval.yaml
View File

2
examples/lora_single_gpu/llama3_lora_kto.yaml → examples/train_lora/llama3_lora_kto.yaml
View File

@ -6,6 +6,7 @@ stage: kto
do_train: true do_train: true
finetuning_type: lora finetuning_type: lora
lora_target: all lora_target: all
pref_beta: 0.1
### dataset ### dataset
dataset: kto_en_demo dataset: kto_en_demo
@ -30,6 +31,7 @@ num_train_epochs: 3.0
lr_scheduler_type: cosine lr_scheduler_type: cosine
warmup_ratio: 0.1 warmup_ratio: 0.1
fp16: true fp16: true
ddp_timeout: 180000000
### eval ### eval
val_size: 0.1 val_size: 0.1

1
examples/lora_single_gpu/llama3_lora_ppo.yaml → examples/train_lora/llama3_lora_ppo.yaml
View File

@ -31,6 +31,7 @@ num_train_epochs: 3.0
lr_scheduler_type: cosine lr_scheduler_type: cosine
warmup_ratio: 0.1 warmup_ratio: 0.1
fp16: true fp16: true
ddp_timeout: 180000000
### generate ### generate
max_new_tokens: 512 max_new_tokens: 512

1
examples/lora_single_gpu/llama3_lora_predict.yaml → examples/train_lora/llama3_lora_predict.yaml
View File

@ -22,3 +22,4 @@ overwrite_output_dir: true
### eval ### eval
per_device_eval_batch_size: 1 per_device_eval_batch_size: 1
predict_with_generate: true predict_with_generate: true
ddp_timeout: 180000000

1
examples/lora_single_gpu/llama3_lora_pretrain.yaml → examples/train_lora/llama3_lora_pretrain.yaml
View File

@ -29,6 +29,7 @@ num_train_epochs: 3.0
lr_scheduler_type: cosine lr_scheduler_type: cosine
warmup_ratio: 0.1 warmup_ratio: 0.1
fp16: true fp16: true
ddp_timeout: 180000000
### eval ### eval
val_size: 0.1 val_size: 0.1

1
examples/lora_single_gpu/llama3_lora_reward.yaml → examples/train_lora/llama3_lora_reward.yaml
View File

@ -30,6 +30,7 @@ num_train_epochs: 3.0
lr_scheduler_type: cosine lr_scheduler_type: cosine
warmup_ratio: 0.1 warmup_ratio: 0.1
fp16: true fp16: true
ddp_timeout: 180000000
### eval ### eval
val_size: 0.1 val_size: 0.1

1
examples/lora_single_gpu/llama3_lora_sft.yaml → examples/train_lora/llama3_lora_sft.yaml
View File

@ -30,6 +30,7 @@ num_train_epochs: 3.0
lr_scheduler_type: cosine lr_scheduler_type: cosine
warmup_ratio: 0.1 warmup_ratio: 0.1
fp16: true fp16: true
ddp_timeout: 180000000
### eval ### eval
val_size: 0.1 val_size: 0.1

4
examples/lora_multi_npu/llama3_lora_sft_ds.yaml → examples/train_lora/llama3_lora_sft_ds0.yaml
View File

@ -6,9 +6,6 @@ stage: sft
do_train: true do_train: true
finetuning_type: lora finetuning_type: lora
lora_target: all lora_target: all
### ddp
ddp_timeout: 180000000
deepspeed: examples/deepspeed/ds_z0_config.json deepspeed: examples/deepspeed/ds_z0_config.json
### dataset ### dataset
@ -34,6 +31,7 @@ num_train_epochs: 3.0
lr_scheduler_type: cosine lr_scheduler_type: cosine
warmup_ratio: 0.1 warmup_ratio: 0.1
fp16: true fp16: true
ddp_timeout: 180000000
### eval ### eval
val_size: 0.1 val_size: 0.1

4
examples/lora_multi_gpu/llama3_lora_sft_ds.yaml → examples/train_lora/llama3_lora_sft_ds3.yaml
View File

@ -6,9 +6,6 @@ stage: sft
do_train: true do_train: true
finetuning_type: lora finetuning_type: lora
lora_target: all lora_target: all
### ddp
ddp_timeout: 180000000
deepspeed: examples/deepspeed/ds_z3_config.json deepspeed: examples/deepspeed/ds_z3_config.json
### dataset ### dataset
@ -34,6 +31,7 @@ num_train_epochs: 3.0
lr_scheduler_type: cosine lr_scheduler_type: cosine
warmup_ratio: 0.1 warmup_ratio: 0.1
fp16: true fp16: true
ddp_timeout: 180000000
### eval ### eval
val_size: 0.1 val_size: 0.1

0
examples/lora_single_gpu/llama3_preprocess.yaml → examples/train_lora/llama3_preprocess.yaml
View File

1
examples/lora_single_gpu/llava1_5_lora_sft.yaml → examples/train_lora/llava1_5_lora_sft.yaml
View File

@ -31,6 +31,7 @@ num_train_epochs: 3.0
lr_scheduler_type: cosine lr_scheduler_type: cosine
warmup_ratio: 0.1 warmup_ratio: 0.1
fp16: true fp16: true
ddp_timeout: 180000000
### eval ### eval
val_size: 0.1 val_size: 0.1

1
examples/qlora_single_gpu/llama3_lora_sft_aqlm.yaml → examples/train_qlora/llama3_lora_sft_aqlm.yaml
View File

@ -30,6 +30,7 @@ num_train_epochs: 3.0
lr_scheduler_type: cosine lr_scheduler_type: cosine
warmup_ratio: 0.1 warmup_ratio: 0.1
fp16: true fp16: true
ddp_timeout: 180000000
### eval ### eval
val_size: 0.1 val_size: 0.1

1
examples/qlora_single_gpu/llama3_lora_sft_awq.yaml → examples/train_qlora/llama3_lora_sft_awq.yaml
View File

@ -30,6 +30,7 @@ num_train_epochs: 3.0
lr_scheduler_type: cosine lr_scheduler_type: cosine
warmup_ratio: 0.1 warmup_ratio: 0.1
fp16: true fp16: true
ddp_timeout: 180000000
### eval ### eval
val_size: 0.1 val_size: 0.1

1
examples/qlora_single_gpu/llama3_lora_sft_bitsandbytes.yaml → examples/train_qlora/llama3_lora_sft_bitsandbytes.yaml
View File

@ -31,6 +31,7 @@ num_train_epochs: 3.0
lr_scheduler_type: cosine lr_scheduler_type: cosine
warmup_ratio: 0.1 warmup_ratio: 0.1
fp16: true fp16: true
ddp_timeout: 180000000
### eval ### eval
val_size: 0.1 val_size: 0.1

1
examples/qlora_single_gpu/llama3_lora_sft_gptq.yaml → examples/train_qlora/llama3_lora_sft_gptq.yaml
View File

@ -30,6 +30,7 @@ num_train_epochs: 3.0
lr_scheduler_type: cosine lr_scheduler_type: cosine
warmup_ratio: 0.1 warmup_ratio: 0.1
fp16: true fp16: true
ddp_timeout: 180000000
### eval ### eval
val_size: 0.1 val_size: 0.1