p12475398
/
LLaMA-Factory-Mirror
forked from p04798526/LLaMA-Factory-Mirror
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

support badam for all stages

This commit is contained in:
hiyouga 2024-04-16 17:44:48 +08:00
parent 4d660c5ade
commit e3d8fc75eb
9 changed files with 61 additions and 28 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
README.md
View File

@ -46,7 +46,7 @@ Choose your path:
- **Various models**: LLaMA, Mistral, Mixtral-MoE, Qwen, Yi, Gemma, Baichuan, ChatGLM, Phi, etc.
- **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.
- **Advanced algorithms**: GaLore, BAdam, DoRA, LongLoRA, LLaMA Pro, LoRA+, LoftQ and Agent tuning.
- **Practical tricks**: FlashAttention-2, Unsloth, RoPE scaling, NEFTune and rsLoRA.
- **Experiment monitors**: LlamaBoard, TensorBoard, Wandb, MLflow, etc.
- **Faster inference**: OpenAI-style API, Gradio UI and CLI with vLLM worker.
@ -68,14 +68,16 @@ Compared to ChatGLM's [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/
## Changelog
[24/04/16] We supported **[BAdam](https://arxiv.org/abs/2404.02827)**. See `examples/extras/badam` for usage.
[24/04/16] We supported **[unsloth](https://github.com/unslothai/unsloth)**'s long-sequence training (Llama-2-7B-56k within 24GB). It achieves **117%** speed and **50%** memory compared with FlashAttention-2, more benchmarks can be found in [this page](https://github.com/hiyouga/LLaMA-Factory/wiki/Performance-comparison).
[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!
<details><summary>Full Changelog</summary>
[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/extras/fsdp_qlora` for usage.
[24/03/13] We supported **[LoRA+](https://arxiv.org/abs/2402.12354)**. See `examples/extras/loraplus` for usage.
@ -279,12 +281,11 @@ huggingface-cli login
\* *estimated*
| Method | Bits | 7B | 13B | 30B | 70B | 8x7B |
| ------ | ---- | ----- | ----- | ----- | ------ | ------ |
| ----------------- | ---- | ----- | ----- | ----- | ------ | ------ |
| Full | AMP | 120GB | 240GB | 600GB | 1200GB | 900GB |
| Full | 16 | 60GB | 120GB | 300GB | 600GB | 400GB |
| GaLore | 16 | 16GB | 32GB | 64GB | 160GB | 120GB |
| Freeze | 16 | 20GB | 40GB | 80GB | 200GB | 160GB |
| LoRA | 16 | 16GB | 32GB | 64GB | 160GB | 120GB |
| LoRA/GaLore/BAdam | 16 | 16GB | 32GB | 64GB | 160GB | 120GB |
| QLoRA | 8 | 10GB | 20GB | 40GB | 80GB | 60GB |
| QLoRA | 4 | 6GB | 12GB | 24GB | 48GB | 30GB |
| QLoRA | 2 | 4GB | 8GB | 16GB | 24GB | 18GB |

19
README_zh.md
View File

@ -46,7 +46,7 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/ec36a9dd-37f4-4f72-81bd
- **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、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。
- **先进算法**：GaLore、BAdam、DoRA、LongLoRA、LLaMA Pro、LoRA+、LoftQ 和 Agent 微调。
- **实用技巧**：FlashAttention-2、Unsloth、RoPE scaling、NEFTune 和 rsLoRA。
- **实验监控**：LlamaBoard、TensorBoard、Wandb、MLflow 等等。
- **极速推理**：基于 vLLM 的 OpenAI 风格 API、浏览器界面和命令行接口。
@ -68,14 +68,16 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/ec36a9dd-37f4-4f72-81bd
## 更新日志
[24/04/16] 我们支持了 **[BAdam](https://arxiv.org/abs/2404.02827)**。详细用法请参照 `examples/extras/badam`。
[24/04/16] 我们支持了 **[unsloth](https://github.com/unslothai/unsloth)** 的长序列训练（24GB 可训练 Llama-2-7B-56k）。该方法相比 FlashAttention-2 提供了 **117%** 的训练速度和 **50%** 的显存节约。更多数据请见[此页面](https://github.com/hiyouga/LLaMA-Factory/wiki/Performance-comparison)。
[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！
<details><summary>展开日志</summary>
[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/extras/fsdp_qlora`。
[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`。
@ -279,12 +281,11 @@ huggingface-cli login
\* *估算值*
| 训练方法 | 精度 | 7B | 13B | 30B | 70B | 8x7B |
| ------- | ---- | ----- | ----- | ----- | ------ | ------ |
| 全参数 | AMP | 120GB | 240GB | 600GB | 1200GB | 900GB |
| 全参数 | 16 | 60GB | 120GB | 300GB | 600GB | 400GB |
| GaLore | 16 | 16GB | 32GB | 64GB | 160GB | 120GB |
| 部分参数 | 16 | 20GB | 40GB | 80GB | 200GB | 160GB |
| LoRA | 16 | 16GB | 32GB | 64GB | 160GB | 120GB |
| ----------------- | ---- | ----- | ----- | ----- | ------ | ------ |
| Full | AMP | 120GB | 240GB | 600GB | 1200GB | 900GB |
| Full | 16 | 60GB | 120GB | 300GB | 600GB | 400GB |
| Freeze | 16 | 20GB | 40GB | 80GB | 200GB | 160GB |
| LoRA/GaLore/BAdam | 16 | 16GB | 32GB | 64GB | 160GB | 120GB |
| QLoRA | 8 | 10GB | 20GB | 40GB | 80GB | 60GB |
| QLoRA | 4 | 6GB | 12GB | 24GB | 48GB | 30GB |
| QLoRA | 2 | 4GB | 8GB | 16GB | 24GB | 18GB |

4
examples/README.md
View File

@ -3,7 +3,7 @@ We provide diverse examples about fine-tuning LLMs.
```
examples/
├── lora_single_gpu/
│ ├── pretrain.sh: Do pre-training using LoRA
│ ├── pretrain.sh: Do continuous pre-training using LoRA
│ ├── sft.sh: Do supervised fine-tuning using LoRA
│ ├── reward.sh: Do reward modeling using LoRA
│ ├── ppo.sh: Do PPO training using LoRA
@ -34,6 +34,8 @@ examples/
└── extras/
├── galore/
│ └── sft.sh: Fine-tune model with GaLore
├── badam/
│ └── sft.sh: Fine-tune model with BAdam
├── loraplus/
│ └── sft.sh: Fine-tune model using LoRA+
├── llama_pro/

4
examples/README_zh.md
View File

@ -3,7 +3,7 @@
```
examples/
├── lora_single_gpu/
│ ├── pretrain.sh: 基于 LoRA 进行预训练
│ ├── pretrain.sh: 基于 LoRA 进行增量预训练
│ ├── sft.sh: 基于 LoRA 进行指令监督微调
│ ├── reward.sh: 基于 LoRA 进行奖励模型训练
│ ├── ppo.sh: 基于 LoRA 进行 PPO 训练
@ -34,6 +34,8 @@ examples/
└── extras/
├── galore/
│ └── sft.sh: 使用 GaLore 训练模型
├── badam/
│ └── sft.sh: 使用 BAdam 训练模型
├── loraplus/
│ └── sft.sh: 使用 LoRA+ 训练模型
├── llama_pro/

6
src/llmtuner/train/dpo/trainer.py
View File

@ -1,5 +1,6 @@
from collections import defaultdict
from contextlib import nullcontext
from types import MethodType
from typing import TYPE_CHECKING, Dict, Literal, Optional, Tuple, Union
import torch
@ -63,6 +64,11 @@ class CustomDPOTrainer(DPOTrainer):
else:
self.ref_model = self.accelerator.prepare_model(self.ref_model, evaluation_mode=True)
if finetuning_args.use_badam:
from badam import clip_grad_norm_for_sparse_tensor
self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_for_sparse_tensor, self.accelerator)
def create_optimizer(self) -> "torch.optim.Optimizer":
if self.optimizer is None:
self.optimizer = create_custom_optimzer(self.model, self.args, self.finetuning_args)

5
src/llmtuner/train/orpo/trainer.py
View File

@ -1,4 +1,5 @@
from collections import defaultdict
from types import MethodType
from typing import TYPE_CHECKING, Dict, Literal, Optional, Tuple, Union
import torch
@ -44,6 +45,10 @@ class CustomORPOTrainer(DPOTrainer):
self._stored_metrics = defaultdict(lambda: defaultdict(list))
Trainer.__init__(self, model=model, **kwargs)
if finetuning_args.use_badam:
from badam import clip_grad_norm_for_sparse_tensor
self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_for_sparse_tensor, self.accelerator)
def create_optimizer(self) -> "torch.optim.Optimizer":
if self.optimizer is None:

6
src/llmtuner/train/ppo/trainer.py
View File

@ -1,6 +1,7 @@
import math
import os
import sys
from types import MethodType
from typing import TYPE_CHECKING, Dict, List, Optional, Tuple
import torch
@ -124,6 +125,11 @@ class CustomPPOTrainer(PPOTrainer, Trainer):
else:
self.reward_model = self.accelerator.prepare_model(self.reward_model, evaluation_mode=True)
if finetuning_args.use_badam:
from badam import clip_grad_norm_for_sparse_tensor
self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_for_sparse_tensor, self.accelerator)
def ppo_train(self, resume_from_checkpoint: Optional[str] = None) -> None:
r"""
Implements training loop for the PPO stage, like _inner_training_loop() in Huggingface's Trainer.

5
src/llmtuner/train/pt/trainer.py
View File

@ -1,3 +1,4 @@
from types import MethodType
from typing import TYPE_CHECKING, Optional
from transformers import Trainer
@ -23,6 +24,10 @@ class CustomTrainer(Trainer):
def __init__(self, finetuning_args: "FinetuningArguments", **kwargs) -> None:
super().__init__(**kwargs)
self.finetuning_args = finetuning_args
if finetuning_args.use_badam:
from badam import clip_grad_norm_for_sparse_tensor
self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_for_sparse_tensor, self.accelerator)
def create_optimizer(self) -> "torch.optim.Optimizer":
if self.optimizer is None:

5
src/llmtuner/train/rm/trainer.py
View File

@ -1,5 +1,6 @@
import json
import os
from types import MethodType
from typing import TYPE_CHECKING, Dict, List, Optional, Tuple, Union
import torch
@ -28,6 +29,10 @@ class PairwiseTrainer(Trainer):
super().__init__(**kwargs)
self.finetuning_args = finetuning_args
self.can_return_loss = True # override property to return eval_loss
if finetuning_args.use_badam:
from badam import clip_grad_norm_for_sparse_tensor
self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_for_sparse_tensor, self.accelerator)
def create_optimizer(self) -> "torch.optim.Optimizer":
if self.optimizer is None: