From b6e008c152421db668c971b0828cbee6a80b16bc Mon Sep 17 00:00:00 2001 From: hiyouga <467089858@qq.com> Date: Thu, 13 Jun 2024 03:15:06 +0800 Subject: [PATCH] update examples --- README.md | 2 +- README_zh.md | 2 +- examples/README.md | 126 ++++++++--------- examples/README_zh.md | 128 ++++++++---------- .../extras/fsdp_qlora/llama3_lora_sft.yaml | 4 +- .../extras/llama_pro/llama3_freeze_sft.yaml | 1 + examples/extras/loraplus/llama3_lora_sft.yaml | 1 + examples/extras/mod/llama3_full_sft.yaml | 1 + examples/lora_multi_gpu/llama3_lora_sft.yaml | 41 ------ .../llama3_full_predict.yaml | 0 .../llama3_full_sft_ds3.yaml} | 0 .../llama3_lora_dpo.yaml | 1 + .../llama3_lora_eval.yaml | 0 .../llama3_lora_kto.yaml | 2 + .../llama3_lora_ppo.yaml | 1 + .../llama3_lora_predict.yaml | 1 + .../llama3_lora_pretrain.yaml | 1 + .../llama3_lora_reward.yaml | 1 + .../llama3_lora_sft.yaml | 1 + .../llama3_lora_sft_ds0.yaml} | 4 +- .../llama3_lora_sft_ds3.yaml} | 4 +- .../llama3_preprocess.yaml | 0 .../llava1_5_lora_sft.yaml | 1 + .../llama3_lora_sft_aqlm.yaml | 1 + .../llama3_lora_sft_awq.yaml | 1 + .../llama3_lora_sft_bitsandbytes.yaml | 1 + .../llama3_lora_sft_gptq.yaml | 1 + 27 files changed, 128 insertions(+), 199 deletions(-) delete mode 100644 examples/lora_multi_gpu/llama3_lora_sft.yaml rename examples/{full_multi_gpu => train_full}/llama3_full_predict.yaml (100%) rename examples/{full_multi_gpu/llama3_full_sft.yaml => train_full/llama3_full_sft_ds3.yaml} (100%) rename examples/{lora_single_gpu => train_lora}/llama3_lora_dpo.yaml (96%) rename examples/{lora_single_gpu => train_lora}/llama3_lora_eval.yaml (100%) rename examples/{lora_single_gpu => train_lora}/llama3_lora_kto.yaml (94%) rename examples/{lora_single_gpu => train_lora}/llama3_lora_ppo.yaml (96%) rename examples/{lora_single_gpu => train_lora}/llama3_lora_predict.yaml (95%) rename examples/{lora_single_gpu => train_lora}/llama3_lora_pretrain.yaml (96%) rename examples/{lora_single_gpu => train_lora}/llama3_lora_reward.yaml (96%) rename examples/{lora_single_gpu => train_lora}/llama3_lora_sft.yaml (96%) rename examples/{lora_multi_npu/llama3_lora_sft_ds.yaml => train_lora/llama3_lora_sft_ds0.yaml} (98%) rename examples/{lora_multi_gpu/llama3_lora_sft_ds.yaml => train_lora/llama3_lora_sft_ds3.yaml} (98%) rename examples/{lora_single_gpu => train_lora}/llama3_preprocess.yaml (100%) rename examples/{lora_single_gpu => train_lora}/llava1_5_lora_sft.yaml (96%) rename examples/{qlora_single_gpu => train_qlora}/llama3_lora_sft_aqlm.yaml (96%) rename examples/{qlora_single_gpu => train_qlora}/llama3_lora_sft_awq.yaml (96%) rename examples/{qlora_single_gpu => train_qlora}/llama3_lora_sft_bitsandbytes.yaml (96%) rename examples/{qlora_single_gpu => train_qlora}/llama3_lora_sft_gptq.yaml (96%) diff --git a/README.md b/README.md index 5bbaf2d7..5dd10d5a 100644 --- a/README.md +++ b/README.md @@ -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. ```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 export examples/merge_lora/llama3_lora_sft.yaml ``` diff --git a/README_zh.md b/README_zh.md index fb616909..76bd2d89 100644 --- a/README_zh.md +++ b/README_zh.md @@ -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**。 ```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 export examples/merge_lora/llama3_lora_sft.yaml ``` diff --git a/examples/README.md b/examples/README.md index f985d552..3372afb9 100644 --- a/examples/README.md +++ b/examples/README.md @@ -4,59 +4,57 @@ Make sure to execute these commands in the `LLaMA-Factory` directory. ## Table of Contents -- [LoRA Fine-Tuning on A Single GPU](#lora-fine-tuning-on-a-single-gpu) -- [QLoRA Fine-Tuning on a Single GPU](#qlora-fine-tuning-on-a-single-gpu) -- [LoRA Fine-Tuning on Multiple GPUs](#lora-fine-tuning-on-multiple-gpus) -- [LoRA Fine-Tuning on Multiple NPUs](#lora-fine-tuning-on-multiple-npus) -- [Full-Parameter Fine-Tuning on Multiple GPUs](#full-parameter-fine-tuning-on-multiple-gpus) +- [LoRA Fine-Tuning](#lora-fine-tuning) +- [QLoRA Fine-Tuning](#qlora-fine-tuning) +- [Full-Parameter Fine-Tuning](#full-parameter-fine-tuning) - [Merging LoRA Adapters and Quantization](#merging-lora-adapters-and-quantization) - [Inferring LoRA Fine-Tuned Models](#inferring-lora-fine-tuned-models) - [Extras](#extras) ## Examples -### LoRA Fine-Tuning on A Single GPU +### LoRA Fine-Tuning #### (Continuous) Pre-Training ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_pretrain.yaml +llamafactory-cli train examples/train_lora/llama3_lora_pretrain.yaml ``` #### Supervised Fine-Tuning ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_sft.yaml +llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml ``` #### Multimodal Supervised Fine-Tuning ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llava1_5_lora_sft.yaml +llamafactory-cli train examples/train_lora/llava1_5_lora_sft.yaml ``` #### Reward Modeling ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_reward.yaml +llamafactory-cli train examples/train_lora/llama3_lora_reward.yaml ``` #### PPO Training ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_ppo.yaml +llamafactory-cli train examples/train_lora/llama3_lora_ppo.yaml ``` #### DPO/ORPO/SimPO Training ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_dpo.yaml +llamafactory-cli train examples/train_lora/llama3_lora_dpo.yaml ``` #### KTO Training ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_kto.yaml +llamafactory-cli train examples/train_lora/llama3_lora_kto.yaml ``` #### Preprocess Dataset @@ -64,95 +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. ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_preprocess.yaml +llamafactory-cli train examples/train_lora/llama3_preprocess.yaml ``` #### Evaluating on MMLU/CMMLU/C-Eval Benchmarks ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli eval examples/lora_single_gpu/llama3_lora_eval.yaml +llamafactory-cli eval examples/train_lora/llama3_lora_eval.yaml ``` #### Batch Predicting and Computing BLEU and ROUGE Scores ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_predict.yaml -``` - -### QLoRA Fine-Tuning on a Single GPU - -#### Supervised Fine-Tuning with 4/8-bit Bitsandbytes Quantization (Recommended) - -```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/qlora_single_gpu/llama3_lora_sft_bitsandbytes.yaml -``` - -#### Supervised Fine-Tuning with 4/8-bit GPTQ Quantization - -```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/qlora_single_gpu/llama3_lora_sft_gptq.yaml -``` - -#### Supervised Fine-Tuning with 4-bit AWQ Quantization - -```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/qlora_single_gpu/llama3_lora_sft_awq.yaml -``` - -#### Supervised Fine-Tuning with 2-bit AQLM Quantization - -```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/qlora_single_gpu/llama3_lora_sft_aqlm.yaml -``` - -### LoRA Fine-Tuning on Multiple GPUs - -#### Supervised Fine-Tuning on Single Node - -```bash -CUDA_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/lora_multi_gpu/llama3_lora_sft.yaml +llamafactory-cli train examples/train_lora/llama3_lora_predict.yaml ``` #### Supervised Fine-Tuning on Multiple Nodes ```bash -CUDA_VISIBLE_DEVICES=0,1,2,3 NNODES=2 RANK=0 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/lora_multi_gpu/llama3_lora_sft.yaml -CUDA_VISIBLE_DEVICES=0,1,2,3 NNODES=2 RANK=1 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/lora_multi_gpu/llama3_lora_sft.yaml +FORCE_TORCHRUN=1 NNODES=2 RANK=0 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml +FORCE_TORCHRUN=1 NNODES=2 RANK=1 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml ``` #### Supervised Fine-Tuning with DeepSpeed ZeRO-3 (Weight Sharding) ```bash -CUDA_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/lora_multi_gpu/llama3_lora_sft_ds.yaml +FORCE_TORCHRUN=1 llamafactory-cli train examples/train_lora/llama3_lora_sft_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 -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 ```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 ```bash -CUDA_VISIBLE_DEVICES=0,1,2,3 NNODES=2 RANK=0 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/full_multi_gpu/llama3_full_sft.yaml -CUDA_VISIBLE_DEVICES=0,1,2,3 NNODES=2 RANK=1 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/full_multi_gpu/llama3_full_sft.yaml +FORCE_TORCHRUN=1 NNODES=2 RANK=0 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/train_full/llama3_full_sft.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 ```bash -CUDA_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/full_multi_gpu/llama3_full_predict.yaml +llamafactory-cli train examples/train_full/llama3_full_predict.yaml ``` ### Merging LoRA Adapters and Quantization @@ -162,35 +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. ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli export examples/merge_lora/llama3_lora_sft.yaml +llamafactory-cli export examples/merge_lora/llama3_lora_sft.yaml ``` #### Quantizing Model using AutoGPTQ ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli export examples/merge_lora/llama3_gptq.yaml +llamafactory-cli export examples/merge_lora/llama3_gptq.yaml ``` ### Inferring LoRA Fine-Tuned Models -Use `CUDA_VISIBLE_DEVICES=0,1` to infer models on multiple devices. - #### Use CLI ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli chat examples/inference/llama3_lora_sft.yaml +llamafactory-cli chat examples/inference/llama3_lora_sft.yaml ``` #### Use Web UI ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli webchat examples/inference/llama3_lora_sft.yaml +llamafactory-cli webchat examples/inference/llama3_lora_sft.yaml ``` #### Launch OpenAI-style API ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli api examples/inference/llama3_lora_sft.yaml +llamafactory-cli api examples/inference/llama3_lora_sft.yaml ``` ### Extras @@ -198,32 +178,32 @@ CUDA_VISIBLE_DEVICES=0 llamafactory-cli api examples/inference/llama3_lora_sft.y #### Full-Parameter Fine-Tuning using GaLore ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/galore/llama3_full_sft.yaml +llamafactory-cli train examples/extras/galore/llama3_full_sft.yaml ``` #### Full-Parameter Fine-Tuning using BAdam ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/badam/llama3_full_sft.yaml +llamafactory-cli train examples/extras/badam/llama3_full_sft.yaml ``` #### LoRA+ Fine-Tuning ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/loraplus/llama3_lora_sft.yaml +llamafactory-cli train examples/extras/loraplus/llama3_lora_sft.yaml ``` #### Mixture-of-Depths Fine-Tuning ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/mod/llama3_full_sft.yaml +llamafactory-cli train examples/extras/mod/llama3_full_sft.yaml ``` #### LLaMA-Pro Fine-Tuning ```bash bash examples/extras/llama_pro/expand.sh -CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/llama_pro/llama3_freeze_sft.yaml +llamafactory-cli train examples/extras/llama_pro/llama3_freeze_sft.yaml ``` #### FSDP+QLoRA Fine-Tuning diff --git a/examples/README_zh.md b/examples/README_zh.md index cf5bbf49..64c31fbd 100644 --- a/examples/README_zh.md +++ b/examples/README_zh.md @@ -4,59 +4,57 @@ ## 目录 -- [单 GPU LoRA 微调](#单-gpu-lora-微调) -- [单 GPU QLoRA 微调](#单-gpu-qlora-微调) -- [多 GPU LoRA 微调](#多-gpu-lora-微调) -- [多 NPU LoRA 微调](#多-npu-lora-微调) -- [多 GPU 全参数微调](#多-gpu-全参数微调) +- [LoRA 微调](#lora-微调) +- [QLoRA 微调](#qlora-微调) +- [全参数微调](#全参数微调) - [合并 LoRA 适配器与模型量化](#合并-lora-适配器与模型量化) - [推理 LoRA 模型](#推理-lora-模型) - [杂项](#杂项) ## 示例 -### 单 GPU LoRA 微调 +### LoRA 微调 #### （增量）预训练 ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_pretrain.yaml +llamafactory-cli train examples/train_lora/llama3_lora_pretrain.yaml ``` #### 指令监督微调 ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_sft.yaml +llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml ``` #### 多模态指令监督微调 ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llava1_5_lora_sft.yaml +llamafactory-cli train examples/train_lora/llava1_5_lora_sft.yaml ``` #### 奖励模型训练 ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_reward.yaml +llamafactory-cli train examples/train_lora/llama3_lora_reward.yaml ``` #### PPO 训练 ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_ppo.yaml +llamafactory-cli train examples/train_lora/llama3_lora_ppo.yaml ``` #### DPO/ORPO/SimPO 训练 ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_dpo.yaml +llamafactory-cli train examples/train_lora/llama3_lora_dpo.yaml ``` #### KTO 训练 ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_kto.yaml +llamafactory-cli train examples/train_lora/llama3_lora_kto.yaml ``` #### 预处理数据集 @@ -64,95 +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。 ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_preprocess.yaml +llamafactory-cli train examples/train_lora/llama3_preprocess.yaml ``` #### 在 MMLU/CMMLU/C-Eval 上评估 ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli eval examples/lora_single_gpu/llama3_lora_eval.yaml +llamafactory-cli eval examples/train_lora/llama3_lora_eval.yaml ``` #### 批量预测并计算 BLEU 和 ROUGE 分数 ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/lora_single_gpu/llama3_lora_predict.yaml +llamafactory-cli train examples/train_lora/llama3_lora_predict.yaml ``` -### 单 GPU QLoRA 微调 - -#### 基于 4/8 比特 Bitsandbytes 量化进行指令监督微调（推荐） +#### 多机指令监督微调 ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/qlora_single_gpu/llama3_lora_sft_bitsandbytes.yaml -``` - -#### 基于 4/8 比特 GPTQ 量化进行指令监督微调 - -```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/qlora_single_gpu/llama3_lora_sft_gptq.yaml -``` - -#### 基于 4 比特 AWQ 量化进行指令监督微调 - -```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/qlora_single_gpu/llama3_lora_sft_awq.yaml -``` - -#### 基于 2 比特 AQLM 量化进行指令监督微调 - -```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/qlora_single_gpu/llama3_lora_sft_aqlm.yaml -``` - -### 多 GPU LoRA 微调 - -#### 在单机上进行指令监督微调 - -```bash -CUDA_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/lora_multi_gpu/llama3_lora_sft.yaml -``` - -#### 在多机上进行指令监督微调 - -```bash -CUDA_VISIBLE_DEVICES=0,1,2,3 NNODES=2 RANK=0 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/lora_multi_gpu/llama3_lora_sft.yaml -CUDA_VISIBLE_DEVICES=0,1,2,3 NNODES=2 RANK=1 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/lora_multi_gpu/llama3_lora_sft.yaml +FORCE_TORCHRUN=1 NNODES=2 RANK=0 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml +FORCE_TORCHRUN=1 NNODES=2 RANK=1 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml ``` #### 使用 DeepSpeed ZeRO-3 平均分配显存 ```bash -CUDA_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/lora_multi_gpu/llama3_lora_sft_ds.yaml +FORCE_TORCHRUN=1 llamafactory-cli train examples/train_lora/llama3_lora_sft_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 -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 ```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 ```bash -CUDA_VISIBLE_DEVICES=0,1,2,3 NNODES=2 RANK=0 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/full_multi_gpu/llama3_full_sft.yaml -CUDA_VISIBLE_DEVICES=0,1,2,3 NNODES=2 RANK=1 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/full_multi_gpu/llama3_full_sft.yaml +FORCE_TORCHRUN=1 NNODES=2 RANK=0 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500 llamafactory-cli train examples/train_full/llama3_full_sft.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 ```bash -CUDA_VISIBLE_DEVICES=0,1,2,3 llamafactory-cli train examples/full_multi_gpu/llama3_full_predict.yaml +llamafactory-cli train examples/train_full/llama3_full_predict.yaml ``` ### 合并 LoRA 适配器与模型量化 @@ -162,35 +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。 ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli export examples/merge_lora/llama3_lora_sft.yaml +llamafactory-cli export examples/merge_lora/llama3_lora_sft.yaml ``` #### 使用 AutoGPTQ 量化模型 ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli export examples/merge_lora/llama3_gptq.yaml +llamafactory-cli export examples/merge_lora/llama3_gptq.yaml ``` ### 推理 LoRA 模型 -使用 `CUDA_VISIBLE_DEVICES=0,1` 进行多卡推理。 - #### 使用命令行接口 ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli chat examples/inference/llama3_lora_sft.yaml +llamafactory-cli chat examples/inference/llama3_lora_sft.yaml ``` #### 使用浏览器界面 ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli webchat examples/inference/llama3_lora_sft.yaml +llamafactory-cli webchat examples/inference/llama3_lora_sft.yaml ``` #### 启动 OpenAI 风格 API ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli api examples/inference/llama3_lora_sft.yaml +llamafactory-cli api examples/inference/llama3_lora_sft.yaml ``` ### 杂项 @@ -198,32 +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 ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/galore/llama3_full_sft.yaml +llamafactory-cli train examples/extras/galore/llama3_full_sft.yaml ``` #### 使用 BAdam 进行全参数训练 ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/badam/llama3_full_sft.yaml +llamafactory-cli train examples/extras/badam/llama3_full_sft.yaml ``` #### LoRA+ 微调 ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/loraplus/llama3_lora_sft.yaml +llamafactory-cli train examples/extras/loraplus/llama3_lora_sft.yaml ``` #### 深度混合微调 ```bash -CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/mod/llama3_full_sft.yaml +llamafactory-cli train examples/extras/mod/llama3_full_sft.yaml ``` #### LLaMA-Pro 微调 ```bash bash examples/extras/llama_pro/expand.sh -CUDA_VISIBLE_DEVICES=0 llamafactory-cli train examples/extras/llama_pro/llama3_freeze_sft.yaml +llamafactory-cli train examples/extras/llama_pro/llama3_freeze_sft.yaml ``` #### FSDP+QLoRA 微调 diff --git a/examples/extras/fsdp_qlora/llama3_lora_sft.yaml b/examples/extras/fsdp_qlora/llama3_lora_sft.yaml index 084269ef..cc773991 100644 --- a/examples/extras/fsdp_qlora/llama3_lora_sft.yaml +++ b/examples/extras/fsdp_qlora/llama3_lora_sft.yaml @@ -8,9 +8,6 @@ do_train: true finetuning_type: lora lora_target: all -### ddp -ddp_timeout: 180000000 - ### dataset dataset: identity,alpaca_en_demo template: llama3 @@ -34,6 +31,7 @@ num_train_epochs: 3.0 lr_scheduler_type: cosine warmup_ratio: 0.1 fp16: true +ddp_timeout: 180000000 ### eval val_size: 0.1 diff --git a/examples/extras/llama_pro/llama3_freeze_sft.yaml b/examples/extras/llama_pro/llama3_freeze_sft.yaml index 444a1113..f92d6945 100644 --- a/examples/extras/llama_pro/llama3_freeze_sft.yaml +++ b/examples/extras/llama_pro/llama3_freeze_sft.yaml @@ -32,6 +32,7 @@ num_train_epochs: 3.0 lr_scheduler_type: cosine warmup_ratio: 0.1 fp16: true +ddp_timeout: 180000000 ### eval val_size: 0.1 diff --git a/examples/extras/loraplus/llama3_lora_sft.yaml b/examples/extras/loraplus/llama3_lora_sft.yaml index 1ba654ec..57383ae0 100644 --- a/examples/extras/loraplus/llama3_lora_sft.yaml +++ b/examples/extras/loraplus/llama3_lora_sft.yaml @@ -31,6 +31,7 @@ num_train_epochs: 3.0 lr_scheduler_type: cosine warmup_ratio: 0.1 fp16: true +ddp_timeout: 180000000 ### eval val_size: 0.1 diff --git a/examples/extras/mod/llama3_full_sft.yaml b/examples/extras/mod/llama3_full_sft.yaml index df03c1e0..085febfc 100644 --- a/examples/extras/mod/llama3_full_sft.yaml +++ b/examples/extras/mod/llama3_full_sft.yaml @@ -31,6 +31,7 @@ num_train_epochs: 3.0 lr_scheduler_type: cosine warmup_ratio: 0.1 pure_bf16: true +ddp_timeout: 180000000 ### eval val_size: 0.1 diff --git a/examples/lora_multi_gpu/llama3_lora_sft.yaml b/examples/lora_multi_gpu/llama3_lora_sft.yaml deleted file mode 100644 index 348e53b9..00000000 --- a/examples/lora_multi_gpu/llama3_lora_sft.yaml +++ /dev/null @@ -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 diff --git a/examples/full_multi_gpu/llama3_full_predict.yaml b/examples/train_full/llama3_full_predict.yaml similarity index 100% rename from examples/full_multi_gpu/llama3_full_predict.yaml rename to examples/train_full/llama3_full_predict.yaml diff --git a/examples/full_multi_gpu/llama3_full_sft.yaml b/examples/train_full/llama3_full_sft_ds3.yaml similarity index 100% rename from examples/full_multi_gpu/llama3_full_sft.yaml rename to examples/train_full/llama3_full_sft_ds3.yaml diff --git a/examples/lora_single_gpu/llama3_lora_dpo.yaml b/examples/train_lora/llama3_lora_dpo.yaml similarity index 96% rename from examples/lora_single_gpu/llama3_lora_dpo.yaml rename to examples/train_lora/llama3_lora_dpo.yaml index 78344330..db25fb51 100644 --- a/examples/lora_single_gpu/llama3_lora_dpo.yaml +++ b/examples/train_lora/llama3_lora_dpo.yaml @@ -32,6 +32,7 @@ num_train_epochs: 3.0 lr_scheduler_type: cosine warmup_ratio: 0.1 fp16: true +ddp_timeout: 180000000 ### eval val_size: 0.1 diff --git a/examples/lora_single_gpu/llama3_lora_eval.yaml b/examples/train_lora/llama3_lora_eval.yaml similarity index 100% rename from examples/lora_single_gpu/llama3_lora_eval.yaml rename to examples/train_lora/llama3_lora_eval.yaml diff --git a/examples/lora_single_gpu/llama3_lora_kto.yaml b/examples/train_lora/llama3_lora_kto.yaml similarity index 94% rename from examples/lora_single_gpu/llama3_lora_kto.yaml rename to examples/train_lora/llama3_lora_kto.yaml index d5234c0a..f730c82e 100644 --- a/examples/lora_single_gpu/llama3_lora_kto.yaml +++ b/examples/train_lora/llama3_lora_kto.yaml @@ -6,6 +6,7 @@ stage: kto do_train: true finetuning_type: lora lora_target: all +pref_beta: 0.1 ### dataset dataset: kto_en_demo @@ -30,6 +31,7 @@ num_train_epochs: 3.0 lr_scheduler_type: cosine warmup_ratio: 0.1 fp16: true +ddp_timeout: 180000000 ### eval val_size: 0.1 diff --git a/examples/lora_single_gpu/llama3_lora_ppo.yaml b/examples/train_lora/llama3_lora_ppo.yaml similarity index 96% rename from examples/lora_single_gpu/llama3_lora_ppo.yaml rename to examples/train_lora/llama3_lora_ppo.yaml index 98c842f9..e574014e 100644 --- a/examples/lora_single_gpu/llama3_lora_ppo.yaml +++ b/examples/train_lora/llama3_lora_ppo.yaml @@ -31,6 +31,7 @@ num_train_epochs: 3.0 lr_scheduler_type: cosine warmup_ratio: 0.1 fp16: true +ddp_timeout: 180000000 ### generate max_new_tokens: 512 diff --git a/examples/lora_single_gpu/llama3_lora_predict.yaml b/examples/train_lora/llama3_lora_predict.yaml similarity index 95% rename from examples/lora_single_gpu/llama3_lora_predict.yaml rename to examples/train_lora/llama3_lora_predict.yaml index a127d248..148c8635 100644 --- a/examples/lora_single_gpu/llama3_lora_predict.yaml +++ b/examples/train_lora/llama3_lora_predict.yaml @@ -22,3 +22,4 @@ overwrite_output_dir: true ### eval per_device_eval_batch_size: 1 predict_with_generate: true +ddp_timeout: 180000000 diff --git a/examples/lora_single_gpu/llama3_lora_pretrain.yaml b/examples/train_lora/llama3_lora_pretrain.yaml similarity index 96% rename from examples/lora_single_gpu/llama3_lora_pretrain.yaml rename to examples/train_lora/llama3_lora_pretrain.yaml index db435ca9..839b3e51 100644 --- a/examples/lora_single_gpu/llama3_lora_pretrain.yaml +++ b/examples/train_lora/llama3_lora_pretrain.yaml @@ -29,6 +29,7 @@ num_train_epochs: 3.0 lr_scheduler_type: cosine warmup_ratio: 0.1 fp16: true +ddp_timeout: 180000000 ### eval val_size: 0.1 diff --git a/examples/lora_single_gpu/llama3_lora_reward.yaml b/examples/train_lora/llama3_lora_reward.yaml similarity index 96% rename from examples/lora_single_gpu/llama3_lora_reward.yaml rename to examples/train_lora/llama3_lora_reward.yaml index 1ce42ea4..79559d19 100644 --- a/examples/lora_single_gpu/llama3_lora_reward.yaml +++ b/examples/train_lora/llama3_lora_reward.yaml @@ -30,6 +30,7 @@ num_train_epochs: 3.0 lr_scheduler_type: cosine warmup_ratio: 0.1 fp16: true +ddp_timeout: 180000000 ### eval val_size: 0.1 diff --git a/examples/lora_single_gpu/llama3_lora_sft.yaml b/examples/train_lora/llama3_lora_sft.yaml similarity index 96% rename from examples/lora_single_gpu/llama3_lora_sft.yaml rename to examples/train_lora/llama3_lora_sft.yaml index 651b636f..fe30c575 100644 --- a/examples/lora_single_gpu/llama3_lora_sft.yaml +++ b/examples/train_lora/llama3_lora_sft.yaml @@ -30,6 +30,7 @@ num_train_epochs: 3.0 lr_scheduler_type: cosine warmup_ratio: 0.1 fp16: true +ddp_timeout: 180000000 ### eval val_size: 0.1 diff --git a/examples/lora_multi_npu/llama3_lora_sft_ds.yaml b/examples/train_lora/llama3_lora_sft_ds0.yaml similarity index 98% rename from examples/lora_multi_npu/llama3_lora_sft_ds.yaml rename to examples/train_lora/llama3_lora_sft_ds0.yaml index a0ec8aa1..08b638e6 100644 --- a/examples/lora_multi_npu/llama3_lora_sft_ds.yaml +++ b/examples/train_lora/llama3_lora_sft_ds0.yaml @@ -6,9 +6,6 @@ stage: sft do_train: true finetuning_type: lora lora_target: all - -### ddp -ddp_timeout: 180000000 deepspeed: examples/deepspeed/ds_z0_config.json ### dataset @@ -34,6 +31,7 @@ num_train_epochs: 3.0 lr_scheduler_type: cosine warmup_ratio: 0.1 fp16: true +ddp_timeout: 180000000 ### eval val_size: 0.1 diff --git a/examples/lora_multi_gpu/llama3_lora_sft_ds.yaml b/examples/train_lora/llama3_lora_sft_ds3.yaml similarity index 98% rename from examples/lora_multi_gpu/llama3_lora_sft_ds.yaml rename to examples/train_lora/llama3_lora_sft_ds3.yaml index 1c432fa7..b7266d61 100644 --- a/examples/lora_multi_gpu/llama3_lora_sft_ds.yaml +++ b/examples/train_lora/llama3_lora_sft_ds3.yaml @@ -6,9 +6,6 @@ stage: sft do_train: true finetuning_type: lora lora_target: all - -### ddp -ddp_timeout: 180000000 deepspeed: examples/deepspeed/ds_z3_config.json ### dataset @@ -34,6 +31,7 @@ num_train_epochs: 3.0 lr_scheduler_type: cosine warmup_ratio: 0.1 fp16: true +ddp_timeout: 180000000 ### eval val_size: 0.1 diff --git a/examples/lora_single_gpu/llama3_preprocess.yaml b/examples/train_lora/llama3_preprocess.yaml similarity index 100% rename from examples/lora_single_gpu/llama3_preprocess.yaml rename to examples/train_lora/llama3_preprocess.yaml diff --git a/examples/lora_single_gpu/llava1_5_lora_sft.yaml b/examples/train_lora/llava1_5_lora_sft.yaml similarity index 96% rename from examples/lora_single_gpu/llava1_5_lora_sft.yaml rename to examples/train_lora/llava1_5_lora_sft.yaml index df510a93..55ac31fa 100644 --- a/examples/lora_single_gpu/llava1_5_lora_sft.yaml +++ b/examples/train_lora/llava1_5_lora_sft.yaml @@ -31,6 +31,7 @@ num_train_epochs: 3.0 lr_scheduler_type: cosine warmup_ratio: 0.1 fp16: true +ddp_timeout: 180000000 ### eval val_size: 0.1 diff --git a/examples/qlora_single_gpu/llama3_lora_sft_aqlm.yaml b/examples/train_qlora/llama3_lora_sft_aqlm.yaml similarity index 96% rename from examples/qlora_single_gpu/llama3_lora_sft_aqlm.yaml rename to examples/train_qlora/llama3_lora_sft_aqlm.yaml index d54d6af6..7b6767d5 100644 --- a/examples/qlora_single_gpu/llama3_lora_sft_aqlm.yaml +++ b/examples/train_qlora/llama3_lora_sft_aqlm.yaml @@ -30,6 +30,7 @@ num_train_epochs: 3.0 lr_scheduler_type: cosine warmup_ratio: 0.1 fp16: true +ddp_timeout: 180000000 ### eval val_size: 0.1 diff --git a/examples/qlora_single_gpu/llama3_lora_sft_awq.yaml b/examples/train_qlora/llama3_lora_sft_awq.yaml similarity index 96% rename from examples/qlora_single_gpu/llama3_lora_sft_awq.yaml rename to examples/train_qlora/llama3_lora_sft_awq.yaml index 5cef178a..a2a26e4b 100644 --- a/examples/qlora_single_gpu/llama3_lora_sft_awq.yaml +++ b/examples/train_qlora/llama3_lora_sft_awq.yaml @@ -30,6 +30,7 @@ num_train_epochs: 3.0 lr_scheduler_type: cosine warmup_ratio: 0.1 fp16: true +ddp_timeout: 180000000 ### eval val_size: 0.1 diff --git a/examples/qlora_single_gpu/llama3_lora_sft_bitsandbytes.yaml b/examples/train_qlora/llama3_lora_sft_bitsandbytes.yaml similarity index 96% rename from examples/qlora_single_gpu/llama3_lora_sft_bitsandbytes.yaml rename to examples/train_qlora/llama3_lora_sft_bitsandbytes.yaml index b308dcab..cc773991 100644 --- a/examples/qlora_single_gpu/llama3_lora_sft_bitsandbytes.yaml +++ b/examples/train_qlora/llama3_lora_sft_bitsandbytes.yaml @@ -31,6 +31,7 @@ num_train_epochs: 3.0 lr_scheduler_type: cosine warmup_ratio: 0.1 fp16: true +ddp_timeout: 180000000 ### eval val_size: 0.1 diff --git a/examples/qlora_single_gpu/llama3_lora_sft_gptq.yaml b/examples/train_qlora/llama3_lora_sft_gptq.yaml similarity index 96% rename from examples/qlora_single_gpu/llama3_lora_sft_gptq.yaml rename to examples/train_qlora/llama3_lora_sft_gptq.yaml index b950042e..ad3d854c 100644 --- a/examples/qlora_single_gpu/llama3_lora_sft_gptq.yaml +++ b/examples/train_qlora/llama3_lora_sft_gptq.yaml @@ -30,6 +30,7 @@ num_train_epochs: 3.0 lr_scheduler_type: cosine warmup_ratio: 0.1 fp16: true +ddp_timeout: 180000000 ### eval val_size: 0.1