Update quick_start.md

quick_start_clean/convert_hf_cpm.py

@@ -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!')

quick_start_clean/readmes/quick_start.md

@@ -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(