chatglm3-6b微调实战_chatglm3 you are chatglm3, a large language model

chatglm3-6b-sft

chatglm3-6b, chatglm-3b微调/LORA/推理/单机多卡/deepspeed/支持多轮对话
源码:https://github.com/yongzhuo/ChatGLM3-SFT

踩坑(截至20231031)

0.官方prompt, 支持chatml(system), system可以改,但是要注意<|user|>最好放第一位, 如果换成<|assistant|>直接预测、微调会变得很差

<|system|>
You are ChatGLM3, a large language model trained by Zhipu.AI. Follow the user's instructions carefully. Respond using markdown.
<|user|>
Hello
<|assistant|>
Hello, I'm ChatGLM3. What can I assist you today?
1
2
3
4
5
6

1. torch>=2.0, 否则微调会报很多错误(单纯推理可以用低版本);
2. tokenizer.encode输出为 [gMASK, sop, 真实文本token]

    "[MASK]": 64789,
    "[gMASK]": 64790,
    "[sMASK]": 64791,
    "sop": 64792,
    "eop": 64793,
    "<|system|>": 64794,
    "<|user|>": 64795,
    "<|assistant|>": 64796,
    "<|observation|>": 64797
1
2
3
4
5
6
7
8
9

3. modeling_chatglm.py自带get_masks()的代码full_attention_mask -= padding_mask.unsqueeze(-1) - 1改为
   full_attention_mask = full_attention_mask.long() - padding_mask.unsqueeze(-1).long() - 1
4. 不支持gradient_checkpointing, 修复的话需要modeling_chatglm.py新增get_input_embeddings, set_input_embeddings;
5. 官方prompt, chatglm3支持工具调用

<|system|>
Answer the following questions as best as you can. You have access to the following tools:
[
    {
        "name": "get_current_weather",
        "description": "Get the current weather in a given location",
        "parameters": {
            "type": "object",
            "properties": {
                "location": {
                    "type": "string",
                    "description": "The city and state, e.g. San Francisco, CA",
                },
                "unit": {"type": "string"},
            },
            "required": ["location"],
        },
    }
]
<|user|>
今天北京的天气怎么样？
<|assistant|>
好的，让我们来查看今天的天气
<|assistant|>get_current_weather
'''python
tool_call(location="beijing", unit="celsius")
'''
<|observation|>
{"temperature": 22}
<|assistant|>
根据查询结果，今天北京的气温为 22 摄氏度。
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31

6. 官方prompt构建chat实现代码:

def build_chat_input(self, query, history=None, role="user"):
    if history is None:
        history = []
    input_ids = []
    for item in history:
        content = item["content"]
        if item["role"] == "system" and "tools" in item:
            content = content + "\n" + json.dumps(item["tools"], indent=4, ensure_ascii=False)
        input_ids.extend(self.build_single_message(item["role"], item.get("metadata", ""), content))
    input_ids.extend(self.build_single_message(role, "", query))
    input_ids.extend([self.get_command("<|assistant|>")])
    return self.batch_encode_plus([input_ids], return_tensors="pt", is_split_into_words=True)
1
2
3
4
5
6
7
8
9
10
11
12

环境配置

transformers==4.27.1
torch>=2.0
sentencepiece
cpm_kernels
mdtex2html
accelerate
protobuf
gradio
1
2
3
4
5
6
7
8

微调样例

地址: chatglm3_6b/ft_chatglm3

配置: chatglm3_6b/ft_chatglm3/config.py
训练: python train.py
接口: python post_api.py
1
2
3
4
5

实验日志

微调日志(metalwoz数据集, 英文主题多轮对话)

推理样例(LoRA, R=8)

回答得比较短了, 也确实遵循了主题

参考/感谢

• https://github.com/THUDM/ChatGLM3-6B
• https://github.com/THUDM/ChatGLM2-6B
• https://github.com/THUDM/ChatGLM-6B
• https://github.com/huggingface/peft
• https://github.com/Instruction-Tuning-with-GPT-4/GPT-4-LLM
• math23k
• metalwoz

免责申明

本项目相关资源仅供学术研究之用，使用涉及第三方代码的部分时，请严格遵循相应的开源协议。模型生成的内容受模型计算、随机性和量化精度损失等因素影响，本项目不对其准确性作出保证。对于模型输出的任何内容，本项目不承担任何法律责任，亦不对因使用相关资源和输出结果而可能产生的任何损失承担责任。

• 大模型权重的详细协议见THUDM/chatglm3-6b

型输出的任何内容，本项目不承担任何法律责任，亦不对因使用相关资源和输出结果而可能产生的任何损失承担责任。

• 大模型权重的详细协议见THUDM/chatglm3-6b

其他-lora合并

可以直接model.merge_and_unload()，也可以参考https://github.com/ymcui/Chinese-LLaMA-Alpaca/blob/main/scripts/merge_llama_with_chinese_lora.py

代码样例

print_named_parameters(model, True)
model = load_model_state(model=model, model_save_dir=MODEL_SAVE_DIR)
model = prepare_model_for_half_training(model,
        use_gradient_checkpointing=False,
        output_embedding_layer_name="lm_head",
        layer_norm_names=["post_attention_layernorm",
                          "final_layernorm",
                          "input_layernorm",
                          ],
        )
print_named_parameters(model, True)
if USE_CUDA:
    model = model.half().cuda()
else:
    model = model.bfloat16()
print_named_parameters(model, True)
model = model.merge_and_unload()
print_named_parameters(model, True)

model.train(False)
model.save_pretrained("merge_and_unload")
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21

参数结果(已生效)

merge前:
('base_model.model.transformer.encoder.layers.0.self_attention.query_key_value.weight', torch.float16, False, tensor([
        [ 0.0064,  0.0006,  0.0198,  ..., -0.0052, -0.0012, -0.0023],
('base_model.model.transformer.encoder.layers.0.self_attention.query_key_value.lora_A.default.weight', torch.float16, False, tensor([
        [-0.0134,  0.0130,  0.0087,  ..., -0.0275, -0.0126, -0.0121],

merge后:
('transformer.encoder.layers.0.self_attention.query_key_value.weight', torch.bfloat16, False, tensor([
        [ 0.0069, -0.0003,  0.0182,  ..., -0.0044, -0.0007, -0.0022],
1
2
3
4
5
6
7
8
9

推理结果(已生效)

merge前:
gMASK sop<|system|>
You are a helpful assistant,10 words per answer.
<|user|>
What's the weather like today?
<|assistant|>
It's sunny and warm.
<|user|>

merge后:
gMASK sop<|system|>
You are a helpful assistant, 10 words per answer.
<|user|>
What's the weather like today?
<|assistant|>
It's sunny and warm.
<|user|>
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17