shuxueslpi / chatGLM-6B-QLoRA

使用peft库,对chatGLM-6B/chatGLM2-6B实现4bit的QLoRA高效微调,并做lora model和base model的merge及4bit的量化(quantize)。
356 stars 46 forks source link

chatglm2-6b输入不跟官方的保持统一吗? #24

Open SCAUapc opened 1 year ago

SCAUapc commented 1 year ago

我看chatglm2的官方代码里build_inputs的函数如下:

def build_inputs(self, tokenizer, query: str, history: List[Tuple[str, str]] = None):
      prompt = ""
      for i, (old_query, response) in enumerate(history):
          prompt += "[Round {}]\n\n问:{}\n\n答:{}\n\n".format(i + 1, old_query, response)
      prompt += "[Round {}]\n\n问:{}\n\n答:".format(len(history) + 1, query)
      inputs = tokenizer([prompt], return_tensors="pt")
      inputs = inputs.to(self.device)
      return inputs

也就是说无论有没有多轮,

[Round {}]\n\n问:{}\n\n答:

这个格式都是有的。但是我看训练代码里的处理数据方式如下:

def tokenize_func(example, tokenizer, global_args, ignore_label_id=-100):
    """单样本tokenize处理"""
    question = global_args.prompt_text + example['instruction']
    if example.get('input', None):
        if example['input'].strip():
            question += f'''\n{example['input']}'''
    answer = example['output']
    q_ids = tokenizer.encode(text=question, add_special_tokens=False)
    a_ids = tokenizer.encode(text=answer, add_special_tokens=False)
    if len(q_ids) > global_args.max_input_length - 2:  # 2 - gmask, bos
        q_ids = q_ids[: global_args.max_input_length - 2]
    if len(a_ids) > global_args.max_output_length - 1:  # 1 - eos
        a_ids = a_ids[: global_args.max_output_length - 1]
    input_ids = tokenizer.build_inputs_with_special_tokens(q_ids, a_ids)
    # question_length = input_ids.index(tokenizer.bos_token_id)
    question_length = len(q_ids) + 2  # chatglm1 - gmask, bos, chatglm2 - gmask, sop
    labels = [ignore_label_id] * question_length + input_ids[question_length:]
    return {'input_ids': input_ids, 'labels': labels}

这里并没有加入如上的模板。是不是加上跟chatglm2对齐的话,会更合适一些?

shuxueslpi commented 1 year ago

@SCAUapc 确实是这样,空的时候我验证下,然后更新

SCAUapc commented 1 year ago

@SCAUapc 确实是这样,空的时候我验证下,然后更新

辛苦大佬~

SCAUapc commented 1 year ago

老哥我根据官方的改了下,你看看对你哟没有帮助。

chatglm2中官方ptuing处理的方法是:

    def preprocess_function_train(examples):
        max_seq_length = data_args.max_source_length + data_args.max_target_length + 1

        model_inputs = {
            "input_ids": [],
            "labels": [],
        }
        for i in range(len(examples[prompt_column])):
            if examples[prompt_column][i] and examples[response_column][i]:
                query, answer = examples[prompt_column][i], examples[response_column][i]

                history = examples[history_column][i] if history_column is not None else None
                prompt = tokenizer.build_prompt(query, history)

                prompt = prefix + prompt
                a_ids = tokenizer.encode(text=prompt, add_special_tokens=True, truncation=True,
                                         max_length=data_args.max_source_length)
                b_ids = tokenizer.encode(text=answer, add_special_tokens=False, truncation=True,
                                         max_length=data_args.max_target_length)

                context_length = len(a_ids)
                input_ids = a_ids + b_ids + [tokenizer.eos_token_id]
                labels = [tokenizer.pad_token_id] * context_length + b_ids + [tokenizer.eos_token_id]

                pad_len = max_seq_length - len(input_ids)
                input_ids = input_ids + [tokenizer.pad_token_id] * pad_len
                labels = labels + [tokenizer.pad_token_id] * pad_len
                if data_args.ignore_pad_token_for_loss:
                    labels = [(l if l != tokenizer.pad_token_id else -100) for l in labels]

                model_inputs["input_ids"].append(input_ids)
                model_inputs["labels"].append(labels)

        return model_inputs

修改之后:

def tokenize_func_v2(example, tokenizer, global_args, ignore_label_id=-100):
    query = example['instruction']
    if example.get('input', None):
        if example['input'].strip():
            query += f'''\n{example['input']}'''

    prompt = tokenizer.build_prompt(query)
    prompt = global_args.prompt_text + prompt
    answer = example['output']

    a_ids = tokenizer.encode(text=prompt, add_special_tokens=True, truncation=True,
                             max_length=global_args.max_input_length)
    b_ids = tokenizer.encode(text=answer, add_special_tokens=False, truncation=True,
                             max_length=global_args.max_target_length)

    context_length = len(a_ids)
    input_ids = a_ids + b_ids + [tokenizer.eos_token_id]
    # prompt需要使用pad_token_id,后面不需要学习。
    labels = [tokenizer.pad_token_id] * context_length + b_ids + [tokenizer.eos_token_id]

    max_seq_length = global_args.max_input_length + global_args.max_output_length + 1

    pad_len = max_seq_length - len(input_ids)
    input_ids = input_ids + [tokenizer.pad_token_id] * pad_len
    labels = labels + [tokenizer.pad_token_id] * pad_len

    # ignore_pad_token_for_loss (包括pad和prompt的)
    labels = [(l if l != tokenizer.pad_token_id else ignore_label_id) for l in labels]

    return {'input_ids': input_ids, 'labels': labels}
SCAUapc commented 1 year ago

官方的prefix是在query加了ROUND模板之后,然后在最前面加的。我这里也是按它的来

shuxueslpi commented 1 year ago

👍 我回头用这个跑一个完整的对比下

SCAUapc commented 1 year ago
    b_ids = tokenizer.encode(text=answer, add_special_tokens=False, truncation=True,
                             max_length=global_args.max_target_length)

中的max_target_length写错了,应该是max_output_length。目前我能跑起来。 想问下大佬,这边训练速度快吗?我1W条薯条2个epoch要70+小时...1080卡单张,batch 1,感觉好慢...我之前用Lora我记得蛮快的,2、3小时就好了

调了下input_max_len 和output_max_len 还有batch_size 这样可以4小时左右好了

shuxueslpi commented 1 year ago

@SCAUapc 我在RTX3090上,例子里的那个数据集,11W数据,1个epoch,bz=32,input_max_len+output_max_len=2048,这样的配置大概跑7个小时

darvsum commented 1 year ago

参照你的修改,能正常跑起来,但loss始终为0.0 ,你知道是什么原因吗?

SCAUapc commented 1 year ago

参照你的修改,能正常跑起来,但loss始终为0.0 ,你知道是什么原因吗?

请问你代码改动的是哪部分?能把改动的那一版面都发一下不。如果按照原本仓库作者的代码,loss是正常的吗?

darvsum commented 1 year ago

@SCAUapc 原官方chatglm2-6b中的ptuning 的main.py中就修改了这个函数 def preprocess_function_train(examples): max_seq_length = data_args.max_source_length + data_args.max_target_length + 1

    model_inputs = {
        "input_ids": [],
        "labels": [],
    }

    #print(examples['text'])
    examples = examples['text']
    for i in range(len(examples)):
        example = json.loads(examples[i])
        #print(example)
        if example[prompt_column] and example[response_column]:
            query, answer = example[prompt_column], example[response_column]

            #history = examples[history_column][i] if history_column is not None else None
            prompt = tokenizer.build_prompt(query)
            prompt = prefix + prompt

            a_ids = tokenizer.encode(text=prompt, add_special_tokens=True, truncation=True,
                                     max_length=data_args.max_source_length)
            b_ids = tokenizer.encode(text=answer, add_special_tokens=False, truncation=True,
                                     max_length=data_args.max_target_length)

            context_length = len(a_ids)
            input_ids = a_ids + b_ids + [tokenizer.eos_token_id]
            labels = [tokenizer.pad_token_id] * context_length + b_ids + [tokenizer.eos_token_id]

            pad_len = max_seq_length - len(input_ids)
            input_ids = input_ids + [tokenizer.pad_token_id] * pad_len
            labels = labels + [tokenizer.pad_token_id] * pad_len

            if data_args.ignore_pad_token_for_loss:
                labels = [(l if l != tokenizer.pad_token_id else -100) for l in labels]

            model_inputs["input_ids"].append(input_ids)
            model_inputs["labels"].append(labels)

    return model_inputs
Parker0000 commented 1 year ago

如果要训练多轮,是不是改为:prompt = tokenizer.build_prompt(query, history)就可以训练多轮对话呢?