Gemma谷歌(google)开源大模型微调实战(fintune gemma-2b/7b)
Gemma-SFT(谷歌, Google), gemma-2b/gemma-7b微调(transformers)/LORA(peft)/推理。
文章共3,593字 · 阅读需要大约12分钟
一键AI生成摘要，助你高效阅读
问答
·
Gemma-SFT
Gemma-SFT(谷歌, Google), gemma-2b/gemma-7b微调(transformers)/LORA(peft)/推理
日志
- v1, 20240222, transformers==4.38.0时候, 微调只跑通了gemma-2b-it(因为只计算了output的loss, 且使用的是fp16), 同时该版本transformers实现有些问题, 比如说1.tokenizer要加bos, 2.RoPE计算精度问题(float32), 3.激活函数gelu_pytorch_tanh;
- v2, 20240329, 微调跑通了gemma-7b-it, 精度至少得bf16, 且loss必须计算input输入的;
项目地址
- https://github.com/yongzhuo/gemma-sft;
- 全部weights要用bf16/fp32/tf32, 使用fp16微调十几或几十的步数后大概率loss=nan;(即便layer-norm是fp32也不行, Gemma在这点上就有点娇气了);
- 强烈建议微调SFT的时候同预训练PT, 要计算inputh和output的损失函数Loss, ADVGEN数据集中如果只计算Output的Loss, 会学不到东西;
备注
0. gemma-7b的loss计算要包括input和output, 否则可能不收敛(gemma-7b-it模型);
1. 非常重要: 全部weights要用bf16/fp32/tf32, 使用fp16微调十几或几十的步数后大概率loss=nan;(即便layer-norm是fp32也不行, LLaMA就没有这个问题, 原因暂时未知)
2. transformers需要4.38及以上;
3. gemma词典大小为25w,多语言版本,包含繁/简体;
4. gemma网络架构同Llama, gemma-2b为18层网络, gemma-7b为28层网络;
5. prompt:
5.1 标准格式为:
bos + input + eos + bos + output + eos
5.2 prompt格式为:
<start_of_turn>user
input<end_of_turn>
<start_of_turn>model
output<end_of_turn>
6 微调输入输出:
输入："<start_of_turn>user\n{问题}<end_of_turn>\n"
输出："<start_of_turn>model\n{答案}<end_of_turn>"
7 推理输入输出(assistant\n放置位置不同):
输入："<start_of_turn>user\n{问题}<end_of_turn>\n<start_of_turn>model\n"
输出："{答案}<end_of_turn>"
8. 网络各层名称
('model.embed_tokens.weight', torch.bfloat16, True)
......
('model.layers.17.self_attn.q_proj.weight', torch.bfloat16, True)
('model.layers.17.self_attn.k_proj.weight', torch.bfloat16, True)
('model.layers.17.self_attn.v_proj.weight', torch.bfloat16, True)
('model.layers.17.self_attn.o_proj.weight', torch.bfloat16, True)
('model.layers.17.mlp.gate_proj.weight', torch.bfloat16, True)
('model.layers.17.mlp.up_proj.weight', torch.bfloat16, True)
('model.layers.17.mlp.down_proj.weight', torch.bfloat16, True)
('model.layers.17.input_layernorm.weight', torch.bfloat16, True)
('model.layers.17.post_attention_layernorm.weight', torch.bfloat16, True)
......
('model.norm.weight', torch.bfloat16, True)
9. RuntimeError: unscale_() has already been called on this optimizer since the last update().
微调语料太少导致的
环境配置
transformers>=4.38.1
torch>=1.13.1
safetensors>=0.4.1
accelerate==0.27.1
fsspec==2023.9.2
rouge==1.0.1
nltk==3.6.6
peft>=0.2.0
numpy
tqdm
微调
地址: gemma_sft/ft_gemma
配置: gemma_sft/ft_gemma/config.py
训练: python train.py
推理: python predict.py
验证: python evaluation.py
接口: python post_api.py
数据集-中文
- https://huggingface.co/datasets/JosephusCheung/GuanacoDataset
- https://huggingface.co/datasets/shareAI/shareGPT_cn
- https://huggingface.co/datasets/Mutonix/RefGPT-Fact
- https://huggingface.co/datasets/BAAI/COIG
- https://github.com/Instruction-Tuning-with-GPT-4/GPT-4-LLM
- https://github.com/carbonz0/alpaca-chinese-dataset
- https://github.com/LianjiaTech/BELLE
- https://github.com/PhoebusSi/Alpaca-CoT
- https://github.com/Hello-SimpleAI/chatgpt-comparison-detection
- https://github.com/yangjianxin1/Firefly
- https://github.com/XueFuzhao/InstructionWild
- https://github.com/OpenLMLab/MOSS
- https://github.com/thu-coai/Safety-Prompts
- https://github.com/LAION-AI/Open-Assistant
- https://github.com/TigerResearch/TigerBot
参考/感谢
- https://github.com/google/gemma_pytorch
- https://huggingface.co/google/gemma-2b-it
- https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard
- https://github.com/THUDM/ChatGLM-6B
- https://github.com/THUDM/GLM
- https://github.com/tatsu-lab/stanford_alpaca
- https://github.com/LianjiaTech/BELLE
- https://github.com/huggingface/peft
- https://github.com/mymusise/ChatGLM-Tuning
- https://github.com/bojone/bert4keras
- trl
- math23k
微调日志-advgen(gemma-7b)
微调日志-advgen(gemma-2b)
推理日志-advgen(gemma-7b)
......
('base_model.model.base_model.model.model.layers.27.self_attn.v_proj.weight', torch.bfloat16, False, tensor([[-6.2866e-03, 3.0670e-03, -2.0599e-03, ..., 2.7847e-04,
-4.3335e-03, 3.3875e-03],
[ 3.5286e-04, -8.6670e-03, -2.3804e-03, ..., -7.9346e-03,
-3.4027e-03, 6.5918e-03],
[-1.5068e-04, -5.0049e-03, 2.6398e-03, ..., -2.9144e-03,
-1.9836e-03, 4.3030e-03],
...,
[ 2.0385e-05, -7.3853e-03, 1.0681e-02, ..., 8.3618e-03,
2.5330e-03, -3.1128e-03],
[ 6.5613e-03, 1.5381e-02, 4.0894e-03, ..., 1.9897e-02,
9.0408e-04, 8.4839e-03],
[-2.3484e-05, 8.1253e-04, 5.9814e-03, ..., -7.6599e-03,
-2.0142e-03, -3.2196e-03]], device='cuda:0', dtype=torch.bfloat16))
('base_model.model.base_model.model.model.layers.27.self_attn.v_proj.lora_A.default.weight', torch.float32, False, tensor([[-9.4881e-03, -1.3139e-02, 1.1027e-02, ..., -6.6568e-03,
-1.1245e-02, -1.4258e-02],
[-6.3376e-04, -1.2741e-02, -1.0051e-02, ..., -9.7185e-03,
8.1461e-03, -5.5879e-04],
[-9.0759e-03, -6.7125e-03, -8.2137e-03, ..., -4.1291e-03,
4.1570e-03, 5.2024e-03],
...,
[-3.9784e-03, 8.0388e-03, -7.3589e-03, ..., 6.9465e-03,
-2.8870e-03, -9.8036e-03],
[ 7.3561e-03, -2.0375e-03, 1.3235e-02, ..., 8.6937e-03,
1.7849e-02, 1.4610e-02],
[ 1.4926e-02, -2.1681e-03, 2.9391e-03, ..., -6.4769e-05,
1.3550e-02, 1.0751e-02]], device='cuda:0'))
('base_model.model.base_model.model.model.layers.27.self_attn.v_proj.lora_B.default.weight', torch.float32, False, tensor([[ 0.0030, -0.0038, -0.0011, ..., 0.0005, -0.0009, -0.0021],
[-0.0058, 0.0004, 0.0030, ..., -0.0041, 0.0035, -0.0032],
[-0.0013, -0.0022, 0.0016, ..., -0.0022, 0.0015, 0.0029],
...,
[-0.0057, 0.0047, 0.0029, ..., 0.0006, -0.0003, 0.0024],
[ 0.0023, -0.0038, 0.0006, ..., -0.0016, 0.0011, -0.0014],
[ 0.0062, -0.0020, -0.0025, ..., 0.0012, -0.0009, -0.0006]],
device='cuda:0'))
......
('base_model.model.base_model.model.model.layers.27.input_layernorm.weight', torch.bfloat16, False, tensor([1.1641, 0.5234, 0.2793, ..., 0.3809, 1.3828, 0.7461], device='cuda:0',
dtype=torch.bfloat16))
('base_model.model.base_model.model.model.layers.27.post_attention_layernorm.weight', torch.bfloat16, False, tensor([3.1875, 3.6875, 3.1250, ..., 3.4375, 3.3906, 2.8750], device='cuda:0',
dtype=torch.bfloat16))
('base_model.model.base_model.model.model.norm.weight', torch.bfloat16, False, tensor([3.0938, 2.0625, 1.6875, ..., 1.5938, 1.5547, 1.4766], device='cuda:0',
dtype=torch.bfloat16))
trainable params: 0 || all params: 8547314688 || trainable%: 0.0
normalizer.cc(51) LOG(INFO) precompiled_charsmap is empty. use identity normalization.
0
2
1
3
4
106
106
108
1645
2516
{'instruction': '类型#上衣*衣样式#风衣*衣样式#外套*衣领型#翻领', 'input': '', 'output': ''}
tensor([[ 2, 106, 1645, 108, 23055, 235345, 235502, 236524, 235287,
236524, 95243, 235345, 236229, 236524, 235287, 236524, 95243, 235345,
182735, 235287, 236524, 236933, 235954, 235345, 237157, 236933, 107,
108, 106, 2516, 108]], device='cuda:0')
<bos><start_of_turn>user
类型#上衣*衣样式#风衣*衣样式#外套*衣领型#翻领<end_of_turn>
<start_of_turn>model
这款风衣外套，采用轻薄透气的材质，穿着舒适透气，不会有闷热感，适合夏天穿。领口翻领，简单利落，造型百搭，上身穿起来很显气质。袖口和下摆采用收紧设计，穿着舒适，拉长身姿。<end_of_turn><eos>
请输入:
类型#裤*版型#宽松*材质#亚麻*裤长#九分裤*裤长#连体裤*裤款式#勾花镂空*裤腰型#松紧腰
请稍等...
################################################################################################################################
{'instruction': '类型#裤*版型#宽松*材质#亚麻*裤长#九分裤*裤长#连体裤*裤款式#勾花镂空*裤腰型#松紧腰', 'input': '', 'output': ''}
tensor([[ 2, 106, 1645, 108, 23055, 235345, 239032, 235287, 236059,
235954, 235345, 238246, 236546, 235287, 139677, 235345, 236971, 237080,
235287, 239032, 236045, 235345, 236496, 235588, 239032, 235287, 239032,
236045, 235345, 236557, 235784, 239032, 235287, 239032, 166242, 235345,
239132, 235894, 244200, 235977, 235287, 239032, 238090, 235954, 235345,
236546, 237220, 238090, 107, 108, 106, 2516, 108]],
device='cuda:0')
<bos><start_of_turn>user
类型#裤*版型#宽松*材质#亚麻*裤长#九分裤*裤长#连体裤*裤款式#勾花镂空*裤腰型#松紧腰<end_of_turn>
<start_of_turn>model
连体裤采用亚麻面料，亲肤舒适，手感柔软，穿着舒适。九分宽松的裤型，让腿部更显纤瘦。裙摆处采用镂空设计，让裙子更显浪漫。松紧腰设计，轻松穿脱。<end_of_turn><eos>
286.55721163749695
请输入:
类型#裙*版型#宽松*版型#显瘦*裙领型#圆领*裙款式#勾花镂空*裙款式#纽扣
请稍等...
################################################################################################################################
{'instruction': '类型#裙*版型#宽松*版型#显瘦*裙领型#圆领*裙款式#勾花镂空*裙款式#纽扣', 'input': '', 'output': ''}
tensor([[ 2, 106, 1645, 108, 23055, 235345, 238728, 235287, 236059,
235954, 235345, 238246, 236546, 235287, 236059, 235954, 235345, 236626,
239300, 235287, 238728, 236933, 235954, 235345, 237869, 236933, 235287,
238728, 166242, 235345, 239132, 235894, 244200, 235977, 235287, 238728,
166242, 235345, 240038, 238498, 107, 108, 106, 2516, 108]],
device='cuda:0')
<bos><start_of_turn>user
类型#裙*版型#宽松*版型#显瘦*裙领型#圆领*裙款式#勾花镂空*裙款式#纽扣<end_of_turn>
<start_of_turn>model
宽松的版型，让你的身材轻松的吸收到这件宽松的裙子，显瘦又显高。圆领的设计，尽显优雅的姿态，修饰颈部曲线。裙子本身是透明的，加上镂空的裙摆，裙子是半透明的，露出内衣的款式，透着淡淡的仙气。纽扣的装饰，美观又实用。<end_of_turn><eos>
13.377868413925171
请输入:
推理日志-advgen(gemma-2b)
请输入:
类型#上衣*颜色#黑色*颜色#紫色*风格#性感*图案#字母*图案#文字*图案#线条*图案#刺绣*衣样式#卫衣*衣长#短款*衣袖型#落肩袖*衣款式#连帽
请稍等...
################################################################################################################################
{'instruction': '类型#上衣*颜色#黑色*颜色#紫色*风格#性感*图案#字母*图案#文字*图案#线条*图案#刺绣*衣样式#卫衣*衣长#短款*衣袖型#落肩袖*衣款式#连帽', 'input': '', 'output': ''}
tensor([[ 2, 23055, 235345, 235502, 236524, 235287, 43972, 235345, 57988,
235287, 43972, 235345, 124945, 235287, 60525, 235345, 135994, 235287,
182148, 235345, 125156, 235287, 182148, 235345, 25047, 235287, 182148,
235345, 179958, 235287, 182148, 235345, 237164, 240872, 235287, 236524,
95243, 235345, 237587, 236524, 235287, 236524, 236045, 235345, 236809,
236604, 235287, 236524, 237785, 235954, 235345, 236362, 238047, 237785,
235287, 236524, 166242, 235345, 236557, 238229]])
一款个性吸睛的连帽服務卫衣，黑色系底色，增添了甜美小性感；经典的落肩短款版型，修饰出纤细的颈脖线条；个性时尚的连帽设计，搭配字母刺绣装饰，增添了甜美少女感；肩部的字母刺绣装饰，增添了时尚感，使整体更具特色；紫色的刺绣设计，丰富了视觉感，使整体更具个性。<eos>
请输入:
类型#上衣*风格#街头*图案#创意*衣样式#卫衣
请稍等...
################################################################################################################################
{'instruction': '类型#上衣*风格#街头*图案#创意*衣样式#卫衣', 'input': '', 'output': ''}
tensor([[ 2, 23055, 235345, 235502, 236524, 235287, 60525, 235345, 218295,
235287, 182148, 235345, 50259, 235287, 236524, 95243, 235345, 237587,
236524]])
这一款卫衣采用经典的领口设计，不拘一格的设计，展现出街头风。领口的设计，不仅能够修饰脸型，还能够打造出精致的小脸，而且还能够起到遮挡口型的效果，让脸型更加小巧。领口处采用了创意的圆环装饰，让整个卫衣更加丰富，视觉上更加亮眼。卫衣采用宽大的版型设计，不挑人穿，即使是身材不那么好的人也能轻松驾驭。<eos>
请输入:
旨在为数千万中国开发者提供一个无缝且高效的云端环境，以支持学习、使用和贡献开源项目。
更多推荐
21
0- 0
已为社区贡献7条内容
所有评论(0)