Update quick_start.md
This commit is contained in:
p18457032
parent
394c3f96e7
commit
994a95f94d
|
|
@ -1,98 +0,0 @@
|
|||
# coding=utf-8
|
||||
# Copyright 2022 The OpenBMB team.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
|
||||
from collections import OrderedDict
|
||||
|
||||
import torch
|
||||
import argparse
|
||||
import os
|
||||
|
||||
parser = argparse.ArgumentParser(description='Load and save model weights with specified paths.')
|
||||
parser.add_argument('--model_path', type=str, required=True, help='Path to the model directory.')
|
||||
parser.add_argument('--output_path', type=str, required=True, help='Path to save the new weights.')
|
||||
parser.add_argument('--model_type',type=str,default='fm9g',help='The model type need to be one of "fm9g" or "9g-8b"')
|
||||
parser.add_argument('--task',type=str,default='pt2bin',help='The task need to be one of "pt2bin" or "bin2pt"')
|
||||
# parser.add_argument('--layer_num', type=int, required=True, help='The layers of model')
|
||||
|
||||
args = parser.parse_args()
|
||||
|
||||
src_path = args.model_path
|
||||
dst_path = args.output_path if args.output_path.endswith('/') else args.output_path + ('/')
|
||||
model_type = args.model_type
|
||||
task = args.task
|
||||
|
||||
assert model_type in ['fm9g'], 'The "model_type" must be one of "fm9g"!'
|
||||
assert task in ['pt2bin','bin2pt'], 'The task need to be one of "pt2bin" or "bin2pt"!'
|
||||
|
||||
if model_type == 'fm9g':
|
||||
layer_num = 40
|
||||
|
||||
if not os.path.exists(dst_path):
|
||||
os.makedirs(dst_path)
|
||||
|
||||
|
||||
def convert_hf_to_fm9g():
|
||||
# 2B模型转换bin2pt
|
||||
ckpt = torch.load(src_path)
|
||||
new_ckpt = OrderedDict()
|
||||
|
||||
new_ckpt['input_embedding.weight'] = ckpt['model.embed_tokens.weight']
|
||||
new_ckpt["encoder.output_layernorm.weight"] = ckpt['model.norm.weight']
|
||||
for i in range(layer_num):
|
||||
new_ckpt[f"encoder.layers.{i}.self_att.self_attention.project_q.weight"] = ckpt[f"model.layers.{i}.self_attn.q_proj.weight"]
|
||||
new_ckpt[f"encoder.layers.{i}.self_att.self_attention.project_k.weight"] = ckpt[f"model.layers.{i}.self_attn.k_proj.weight"]
|
||||
new_ckpt[f"encoder.layers.{i}.self_att.self_attention.project_v.weight"] = ckpt[f"model.layers.{i}.self_attn.v_proj.weight"]
|
||||
new_ckpt[f"encoder.layers.{i}.self_att.self_attention.attention_out.weight"] = ckpt[f"model.layers.{i}.self_attn.o_proj.weight"]
|
||||
new_ckpt[f"encoder.layers.{i}.self_att.layernorm_before_attention.weight"] = ckpt[f"model.layers.{i}.input_layernorm.weight"]
|
||||
new_ckpt[f"encoder.layers.{i}.ffn.layernorm_before_ffn.weight"] = ckpt[f"model.layers.{i}.post_attention_layernorm.weight"]
|
||||
|
||||
new_ckpt[f"encoder.layers.{i}.ffn.ffn.w_in.w_0.weight"] = ckpt[f'model.layers.{i}.mlp.gate_proj.weight']
|
||||
new_ckpt[f"encoder.layers.{i}.ffn.ffn.w_in.w_1.weight"] = ckpt[f'model.layers.{i}.mlp.up_proj.weight']
|
||||
new_ckpt[f"encoder.layers.{i}.ffn.ffn.w_out.weight"] = ckpt[f'model.layers.{i}.mlp.down_proj.weight']
|
||||
|
||||
torch.save(new_ckpt, f"{dst_path}fm9g.pt")
|
||||
|
||||
def convert_fm9g_to_hf():
|
||||
#2B模型转换pt2bin
|
||||
state = torch.load(src_path)
|
||||
|
||||
new_state = {}
|
||||
new_state["model.embed_tokens.weight"] = state["input_embedding.weight"]
|
||||
new_state["model.norm.weight"] = state["encoder.output_layernorm.weight"]
|
||||
for lid in range(layer_num):
|
||||
print(lid)
|
||||
new_state[f"model.layers.{lid}.self_attn.q_proj.weight"] = state[f"encoder.layers.{lid}.self_att.self_attention.project_q.weight"]
|
||||
new_state[f"model.layers.{lid}.self_attn.k_proj.weight"] = state[f"encoder.layers.{lid}.self_att.self_attention.project_k.weight"]
|
||||
new_state[f"model.layers.{lid}.self_attn.v_proj.weight"] = state[f"encoder.layers.{lid}.self_att.self_attention.project_v.weight"]
|
||||
|
||||
new_state[f"model.layers.{lid}.self_attn.o_proj.weight"] = state[f"encoder.layers.{lid}.self_att.self_attention.attention_out.weight"]
|
||||
new_state[f"model.layers.{lid}.mlp.gate_proj.weight"] = state[f"encoder.layers.{lid}.ffn.ffn.w_in.w_0.weight"]
|
||||
new_state[f"model.layers.{lid}.mlp.up_proj.weight"] = state[f"encoder.layers.{lid}.ffn.ffn.w_in.w_1.weight"]
|
||||
new_state[f"model.layers.{lid}.mlp.down_proj.weight"] = state[f"encoder.layers.{lid}.ffn.ffn.w_out.weight"]
|
||||
|
||||
new_state[f"model.layers.{lid}.input_layernorm.weight"] = state[f"encoder.layers.{lid}.self_att.layernorm_before_attention.weight"]
|
||||
new_state[f"model.layers.{lid}.post_attention_layernorm.weight"] = state[f"encoder.layers.{lid}.ffn.layernorm_before_ffn.weight"]
|
||||
del state
|
||||
state = None
|
||||
torch.save(new_state, f"{dst_path}fm9g.bin")
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
if model_type == 'fm9g' and task == 'bin2pt':
|
||||
convert_hf_to_fm9g()
|
||||
elif model_type == 'fm9g' and task == 'pt2bin':
|
||||
convert_fm9g_to_hf()
|
||||
else:
|
||||
raise ValueError('Please check the model type and task!')
|
||||
|
|
@ -52,6 +52,7 @@
|
|||
- [多机训练](#多机训练)
|
||||
- [参数详细介绍](#参数详细介绍)
|
||||
- [查看训练情况](#查看训练情况)
|
||||
- [模型格式转换](#模型格式转换)
|
||||
- [模型推理](#模型推理)
|
||||
- [常见问题](#常见问题)
|
||||
|
||||
|
|
@ -433,6 +434,21 @@ tensorboard –-logdir /apps/fm9g_2b/data/tensorboard/2b_0701 #存放.events文
|
|||
TypeError: MessageToJson() got an unexpected keyword argument 'including_default_value_fields'
|
||||
```
|
||||
|
||||
## 模型格式转换
|
||||
模型训练完成后，需将pt格式模型文件转换为bin格式模型文件用于模型推理。
|
||||
我们在本项目中提供了2B模型两种格式相互转换时所用到脚本，脚本位于./quick_start_clean/convert_hf_fm9g.py，应用方法如下：
|
||||
|
||||
```shell
|
||||
python convert_hf_fm9g.py \
|
||||
--model_path /the_path_to_pt_or_bin/ \ #需要转换模型的路径
|
||||
--output_path /the_path_to_target_directory/ \ #转换后新格式模型所存放路径
|
||||
--model_type fm9g \ #2B模型指定fm9g
|
||||
--task pt2bin #任务类型如果pt模型转换为bin模型指定为pt2bin，反之指定为bin2pt
|
||||
```
|
||||
|
||||
8B模型格式转换脚本需要切换至master分支，脚本位于本项目master分支下convert.py。
|
||||
|
||||
|
||||
## 模型推理
|
||||
模型推理列举了两种推理方法：离线批量推理和部署OpenAI API服务推理
|
||||
|
||||
|
|
@ -498,7 +514,7 @@ python -m vllm.entrypoints.openai.api_server \
|
|||
--tokenizer-mode auto \
|
||||
--dtype auto \
|
||||
--trust-remote-code \
|
||||
--api-key CPMAPI
|
||||
--api-key FM9GAPI
|
||||
#同样需注意模型加载的是.bin格式
|
||||
#与离线批量推理类似，使用端侧2B模型，tokenizer-mode为"auto"
|
||||
#dtype为模型数据类型，设置为"auto"即可
|
||||
|
|
@ -511,7 +527,7 @@ python -m vllm.entrypoints.openai.api_server \
|
|||
--model ../models/8b_sft_model/ \
|
||||
--tokenizer-mode cpm \
|
||||
--dtype auto \
|
||||
--api-key CPMAPI
|
||||
--api-key FM9GAPI
|
||||
#与离线批量推理类似，使用8B百亿SFT模型，tokenizer-mode为"cpm"
|
||||
```
|
||||
|
||||
|
|
@ -530,7 +546,7 @@ INFO: Uvicorn running on http://0.0.0.0:8000 (Press CTRL+C to quit)
|
|||
# client.py
|
||||
from openai import OpenAI
|
||||
# 如果启动服务时指定了api密钥，需要修改为对应的密钥，否则为"EMPTY"
|
||||
openai_api_key = "CPMAPI"
|
||||
openai_api_key = "FM9GAPI"
|
||||
openai_api_base = "http://localhost:8000/v1"
|
||||
client = OpenAI(
|
||||
api_key=openai_api_key,
|
||||
|
|
@ -549,7 +565,7 @@ print("Completion result:", completion)
|
|||
from openai import OpenAI
|
||||
client = OpenAI(
|
||||
base_url="http://localhost:8000/v1",
|
||||
api_key="CPMAPI",
|
||||
api_key="FM9GAPI",
|
||||
)
|
||||
#每次将上一轮的问题和答案拼接到本轮输入即可
|
||||
completion = client.chat.completions.create(
|
||||
|
|
|
|||
Loading…
Reference in New Issue