请问训练结束还要准备什么

[fangyifei222]

我对13B模型lora微调之后，只获得了如下的文件，我将.pt文件转成了.bin模型，但是其他的configuration_baichuan.py，generation_config.json， modeling_baichuan.py都使用原来的好像不可以，请问是要针对微调后的模型自行修改吗

[ssbuild]

1. 正常情况下是需要通过对应infer 模块下的文件进行推理。
2. lora 权重 不能用peft直接加载 ， 可以把lora 权重合并下， 这样权重就跟官方权重一样了，可以直接加载推理

[fangyifei222]

我训练之后的模型文件就是27G，我的数据集是问答对所以对训练脚本的数据处理做了一点修改，训练完得到的这个pt文件我转成了.bin格式的权重。我加载这个模型就会提示缺少configuration_baichuan.py这种文件，我把baichuan-2-13B的对应文件复制过来加载模型会提示部分权重未使用
-rw-rw-r-- 1 fyf fyf 4918199 Mar 6 10:00 bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
-rw-rw-r-- 1 fyf fyf 4918199 Mar 6 10:00 bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
-rw-rw-r-- 1 fyf fyf 13896988349 Mar 6 10:00 zero_pp_rank_0_mp_rank_00_model_states.pt
-rw-rw-r-- 1 fyf fyf 13896988349 Mar 6 10:00 zero_pp_rank_1_mp_rank_00_model_states.pt

[ssbuild]

我训练之后的模型文件就是27G，我的数据集是问答对所以对训练脚本的数据处理做了一点修改，训练完得到的这个pt文件我转成了.bin格式的权重。我加载这个模型就会提示缺少configuration_baichuan.py这种文件，我把baichuan-2-13B的对应文件复制过来加载模型会提示部分权重未使用 -rw-rw-r-- 1 fyf fyf 4918199 Mar 6 10:00 bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt -rw-rw-r-- 1 fyf fyf 4918199 Mar 6 10:00 bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt -rw-rw-r-- 1 fyf fyf 13896988349 Mar 6 10:00 zero_pp_rank_0_mp_rank_00_model_states.pt -rw-rw-r-- 1 fyf fyf 13896988349 Mar 6 10:00 zero_pp_rank_1_mp_rank_00_model_states.pt

不能直接改后缀 ， deepspeed 权重需要通过下面转换下
cd best_ckpt/last
python zero_to_fp32.py . ../last.ckpt

[ssbuild]

baichuan_finetuning/infer/infer_finetuning.py

Line 31 in 089367d

# deepspeed 权重使用转换脚本命令

[fangyifei222]

我好像知道问题出在哪了，我想请问一下我用的是自制的数据集，是不是数据集的格式得跟data文件夹的示例数据格式相同，我的数据集格式是这种：{
"Question": "请问8端口OC-12c/STM-4c POS-SFP灵活插卡的尺寸(宽×深×高)分别是多少？",
"Answer": "8端口OC-12c/STM-4c POS-SFP灵活插卡的尺寸(宽×深×高)分别是169mm × 189.9mm × 18.4mm。"
},
{
"Question": "请问4端口OC-3c/STM-1c POS-SFP 灵活插卡的典型散热值是多少？",
"Answer": "4端口OC-3c/STM-1c POS-SFP 灵活插卡的典型散热值是275.8 BTU/hour。"
},

[ssbuild]

可以看下readme 的 datasample ， 或者 data文件夹下的例子。

[fangyifei222]

好的，谢谢您

[fangyifei222]

您好我想请问一下，我修改了fine-tune.py中的peft_config = LoraConfig(
task_type=TaskType.CAUSAL_LM,
target_modules=["W_pack"],
inference_mode=False,
r=1,
lora_alpha=32,
lora_dropout=0.1,
)调整了r=8和 lora_alpha=16，但是训练结束后的模型加载报错
RuntimeError: Error(s) in loading state_dict for PeftModelForCausalLM: size mismatch for base_model.model.model.layers.0.self_attn.W_pack.lora_B.default.weight: copying a param with shape torch.Size([0]) from checkpoint, the shape in current model is torch.Size([15360, 8]).
size mismatch for base_model.model.model.layers.1.self_attn.W_pack.lora_B.default.weight: copying a param with shape torch.Size([0]) from checkpoint, the shape in current model is torch.Size([15360, 8]).
size mismatch for base_model.model.model.layers.2.self_attn.W_pack.lora_B.default.weight: copying a param with shape torch.Size([0]) from checkpoint, the shape in current model is torch.Size([15360, 8]).
size mismatch for base_model.model.model.layers.3.self_attn.W_pack.lora_B.default.weight: copying a param with shape torch.Size([0]) from checkpoint, the shape in current model is torch.Size([15360, 8]).
size mismatch for base_model.model.model.layers.4.self_attn.W_pack.lora_B.default.weight: copying a param with shape torch.Size([0]) from checkpoint, the shape in current model is torch.Size([15360, 8]).您知道怎么解决吗

[ssbuild]

您好我想请问一下，我修改了fine-tune.py中的peft_config = LoraConfig( task_type=TaskType.CAUSAL_LM, target_modules=["W_pack"], inference_mode=False, r=1, lora_alpha=32, lora_dropout=0.1, )调整了r=8和 lora_alpha=16，但是训练结束后的模型加载报错 RuntimeError: Error(s) in loading state_dict for PeftModelForCausalLM: size mismatch for base_model.model.model.layers.0.self_attn.W_pack.lora_B.default.weight: copying a param with shape torch.Size([0]) from checkpoint, the shape in current model is torch.Size([15360, 8]). size mismatch for base_model.model.model.layers.1.self_attn.W_pack.lora_B.default.weight: copying a param with shape torch.Size([0]) from checkpoint, the shape in current model is torch.Size([15360, 8]). size mismatch for base_model.model.model.layers.2.self_attn.W_pack.lora_B.default.weight: copying a param with shape torch.Size([0]) from checkpoint, the shape in current model is torch.Size([15360, 8]). size mismatch for base_model.model.model.layers.3.self_attn.W_pack.lora_B.default.weight: copying a param with shape torch.Size([0]) from checkpoint, the shape in current model is torch.Size([15360, 8]). size mismatch for base_model.model.model.layers.4.self_attn.W_pack.lora_B.default.weight: copying a param with shape torch.Size([0]) from checkpoint, the shape in current model is torch.Size([15360, 8]).您知道怎么解决吗

你用的哪个脚本做的推理？

[fangyifei222]

我用的fine-tune.py做的训练，cli_demo.py做的加载和推理，我用了AutoPeftModelForCausalLM来加载模型，之前默认的r=1 lora_alpha=32训练之后的模型可以正常加载，修改了这两个参数训练之后的模型就出现了尺寸问题

[ssbuild]

我用的fine-tune.py做的训练，cli_demo.py做的加载和推理，我用了AutoPeftModelForCausalLM来加载模型，之前默认的r=1 lora_alpha=32训练之后的模型可以正常加载，修改了这两个参数训练之后的模型就出现了尺寸问题

推理脚本代码 贴一下，我看看

[fangyifei222]

from transformers import AutoTokenizer
from peft import AutoPeftModelForCausalLM
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "/data/fyf/Baichuan2-main/fine-tune/outputr8a16-1"
device = "cuda:0"
model = AutoPeftModelForCausalLM.from_pretrained(model_id, trust_remote_code=True)

model = model.to(device)
model.eval()
tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=False, trust_remote_code=True)

while True:

user_input = input("请输入您的问题（或输入'退出'来结束对话）：")
if user_input == "退出":
    break


messages = [{"role": "user", "content": user_input}]

response = model.chat(tokenizer, messages)


print("模型回复：", response)

[ssbuild]

from transformers import AutoTokenizer from peft import AutoPeftModelForCausalLM from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "/data/fyf/Baichuan2-main/fine-tune/outputr8a16-1" device = "cuda:0" model = AutoPeftModelForCausalLM.from_pretrained(model_id, trust_remote_code=True)

model = model.to(device) model.eval() tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=False, trust_remote_code=True)

while True:

user_input = input("请输入您的问题（或输入'退出'来结束对话）：")
if user_input == "退出":
    break


messages = [{"role": "user", "content": user_input}]

response = model.chat(tokenizer, messages)


print("模型回复：", response)

不能直接加载，名字有点区别， 回头加一个权重转换脚本

[fangyifei222]

from transformers import AutoTokenizer from peft import AutoPeftModelForCausalLM from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "/data/fyf/Baichuan2-main/fine-tune/outputr8a16-1" device = "cuda:0" model = AutoPeftModelForCausalLM.from_pretrained(model_id, trust_remote_code=True)
model = model.to(device) model.eval() tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=False, trust_remote_code=True)
while True:

user_input = input("请输入您的问题（或输入'退出'来结束对话）：")
if user_input == "退出":
    break


messages = [{"role": "user", "content": user_input}]

response = model.chat(tokenizer, messages)


print("模型回复：", response)

不能直接加载，名字有点区别， 回头加一个权重转换脚本

要转换为.bin格式嘛，那为什么r=1 ，lora_alpha=32的时候直接就能加载

[ssbuild]

from transformers import AutoTokenizer from peft import AutoPeftModelForCausalLM from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "/data/fyf/Baichuan2-main/fine-tune/outputr8a16-1" device = "cuda:0" model = AutoPeftModelForCausalLM.from_pretrained(model_id, trust_remote_code=True)
model = model.to(device) model.eval() tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=False, trust_remote_code=True)
while True:

user_input = input("请输入您的问题（或输入'退出'来结束对话）：")
if user_input == "退出":
    break


messages = [{"role": "user", "content": user_input}]

response = model.chat(tokenizer, messages)


print("模型回复：", response)

不能直接加载，名字有点区别， 回头加一个权重转换脚本

要转换为.bin格式嘛，那为什么r=1 ，lora_alpha=32的时候直接就能加载

不是不是， 就是权重key ，不是文件名字 ， 有时间我加一下。

[fangyifei222]

我用的是https://github.com/baichuan-inc/Baichuan2这个项目，您是百川的开发人员嘛QvQ

[ssbuild]

deep_export --mode=hf --src adapter_model.bin --dst=.
@fangyifei222 使用以上方式可以转换成huggingface 权重