Merge pull request #1 from ChuangLee/main

多卡版本

README.md

@@ -1,5 +1,11 @@
 # ChatGLM-6B
 
+## 修改介绍
+将模型加载到多张gpu卡中，根据gpu的数量自动分配平均的显存占用，需要安装accelerate
+```shell
+python -m pip install accelerate
+```
+请注意，仍然需要24GB的内存， 后续优化 TODO
 ## 介绍
 
 ChatGLM-6B 是一个开源的、支持中英双语的对话语言模型，基于 [General Language Model (GLM)](https://github.com/THUDM/GLM) 架构，具有 62 亿参数。结合模型量化技术，用户可以在消费级的显卡上进行本地部署（INT4 量化级别下最低只需 6GB 显存）。

README_en.md

@@ -1,5 +1,12 @@
 # ChatGLM-6B
 
+## Modification 
+Load the model into multiple GPUs and automatically allocate the average memory usage according to the number of GPUs.
+```shell
+python -m pip install accelerate
+```
+Please note that 24GB of cpu memory is still required. TODO optimization.”
+
 ## Introduction
 
 ChatGLM-6B is an open bilingual language model based on [General Language Model (GLM)](https://github.com/THUDM/GLM) framework, with 6.2 billion parameters. With the quantization technique, users can deploy locally on consumer-grade graphics cards (only 6GB of GPU memory is required at the INT4 quantization level).

chatglm_parallel.py

@@ -0,0 +1,34 @@
+'''
+Author: lichuang
+Date: 2023-03-23 09:18:13
+Description: 将模型加载到多张GPU卡中，根据gpu的数量自动分配平均的显存占用 
+'''
+
+from transformers import AutoModel, AutoTokenizer
+from accelerate import load_checkpoint_and_dispatch
+
+
+def load_model_on_gpus(checkpoint_path, num_gpus=2):
+    # 总共占用13GB显存,28层transformer每层0.39GB左右
+    # 第一层 word_embeddings和最后一层 lm_head 层各占用1.2GB左右
+    num_trans_layers = 28
+    vram_per_layer = 0.39
+    average = 13/num_gpus
+    used = 1.2
+    device_map = {'transformer.word_embeddings': 0,
+                  'transformer.final_layernorm': num_gpus-1, 'lm_head': num_gpus-1}
+    gpu_target = 0
+    for i in range(num_trans_layers):
+        if used > average-vram_per_layer/2 and gpu_target < num_gpus:
+            gpu_target += 1
+            used = 0
+        else:
+            used += vram_per_layer
+        device_map['transformer.layers.%d' % i] = gpu_target
+
+    model = AutoModel.from_pretrained(
+        checkpoint_path, trust_remote_code=True)
+    model = model.eval()
+    model = load_checkpoint_and_dispatch(
+        model, checkpoint_path, device_map=device_map, offload_folder="offload", offload_state_dict=True).half()
+    return model

cli_demo.py

@@ -1,10 +1,10 @@
 import os
 import platform
 from transformers import AutoTokenizer, AutoModel
+from chatglm_parallel import load_model_on_gpus
 
 tokenizer = AutoTokenizer.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True)
-model = AutoModel.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True).half().cuda()
-model = model.eval()
+model = load_model_on_gpus("THUDM/chatglm-6b", num_gpus=2)
 
 os_name = platform.system()
 clear_command = 'cls' if os_name == 'Windows' else 'clear'

web_demo.py

@@ -1,9 +1,9 @@
 from transformers import AutoModel, AutoTokenizer
 import gradio as gr
+from chatglm_parallel import load_model_on_gpus
 
 tokenizer = AutoTokenizer.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True)
-model = AutoModel.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True).half().cuda()
-model = model.eval()
+model = load_model_on_gpus("THUDM/chatglm-6b", num_gpus=2)
 
 MAX_TURNS = 20
 MAX_BOXES = MAX_TURNS * 2

web_demo2.py

@@ -1,6 +1,7 @@
 from transformers import AutoModel, AutoTokenizer
 import streamlit as st
 from streamlit_chat import message
+from chatglm_parallel import load_model_on_gpus
 
 
 st.set_page_config(
@@ -12,8 +13,7 @@
 @st.cache_resource
 def get_model():
     tokenizer = AutoTokenizer.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True)
-    model = AutoModel.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True).half().cuda()
-    model = model.eval()
+    model = load_model_on_gpus("THUDM/chatglm-6b", num_gpus=2)
     return tokenizer, model