lmdeploy 升级为 0.6.2

README.md

@@ -258,7 +258,7 @@ streamlit run server_ui.py
 |    glm-4v-9b    |chatglm| × |  ×  |         ×         |        √        |
 |    InternVL2    |internvl2| × |  ×  |         √         |        √        |
 |    MiniCPM-V-2_6   |minicpmv | × |  √  |         √         |        ×        |
-|    Qwen2-VL   |qwen | × |  √  |         ×         |        ×        |
+|    Qwen2-VL   |qwen | × |  √  |         ×         |        √        |
 <br>
 
 ### Embedding模型

gpt_server/__init__.py

@@ -22,7 +22,7 @@ def check_lmdeploy_lib():
 lmdeploy_path = os.path.dirname(get_module_path(module_name))
 if not check_lmdeploy_lib():
     logger.warning("不存在lmdeploy的lib文件目录,系统将会自动下载！")
-    cmd = "pip install --force-reinstall lmdeploy==0.6.0 --no-deps"
+    cmd = "pip install --force-reinstall lmdeploy==0.6.2 --no-deps"
     logger.info(f"正在执行命令：{cmd}")
     os.system(cmd)
     logger.info("安装成功，请重新启动服务！")

gpt_server/model_worker/base/model_worker_base.py

@@ -4,8 +4,7 @@
 from abc import ABC, abstractmethod
 from fastapi import BackgroundTasks, Request, FastAPI
 from fastapi.responses import JSONResponse, StreamingResponse
-from fastchat.utils import get_context_length as get_context_length_
-
+from fastchat.utils import SEQUENCE_LENGTH_KEYS
 from loguru import logger
 import os
 from transformers import (
@@ -23,6 +22,24 @@
 app = FastAPI()
 
 
+def get_context_length_(config):
+    """Get the context length of a model from a huggingface model config."""
+    rope_scaling = getattr(config, "rope_scaling", None)
+    if rope_scaling:
+        try:
+            rope_scaling_factor = config.rope_scaling["factor"]
+        except KeyError:
+            rope_scaling_factor = 1
+    else:
+        rope_scaling_factor = 1
+
+    for key in SEQUENCE_LENGTH_KEYS:
+        val = getattr(config, key, None)
+        if val is not None:
+            return int(rope_scaling_factor * val)
+    return 2048
+
+
 class ModelWorkerBase(BaseModelWorker, ABC):
     def __init__(
         self,

install.sh

@@ -1,2 +1,2 @@
 pip install -r requirements.txt
-pip install --force-reinstall lmdeploy==0.6.0 --no-deps
+pip install --force-reinstall lmdeploy==0.6.2 --no-deps

pyproject.toml

@@ -22,6 +22,7 @@ dependencies = [
     "torchvision==0.19",
     "transformers==4.45.2",
     "vllm==0.6.3",
+    "qwen_vl_utils"
 ]
 
 

requirements.txt

@@ -5,11 +5,12 @@ fschat==0.2.36
 transformers==4.45.2
 accelerate
 vllm==0.6.3
-git+https://gitee.com/liuyu_1997/lmdeploy.git#egg=lmdeploy # lmdeploy==0.6.0
+git+https://gitee.com/liuyu_1997/lmdeploy.git # lmdeploy==0.6.2
 openai==1.44.0
 loguru
 fastapi==0.114.1
 infinity-emb[all]==0.0.53
 git+https://gitee.com/liuyu_1997/ffmpy.git # 解决setuptools最新版本导致的重大事故
 gradio==4.26.0
-streamlit==1.39.0
+streamlit==1.39.0
+qwen_vl_utils

tests/test_openai_vl_chat.py

@@ -1,5 +1,22 @@
+import base64
 from openai import OpenAI
 
+
+def image_to_base64(image_path):
+    """将图片转换为Base64字符串"""
+    base64_prefix = "data:image/png;base64,"
+
+    with open(image_path, "rb") as image_file:
+        base64_string = base64.b64encode(image_file.read()).decode("utf-8")
+    return base64_prefix + base64_string
+
+
+image_path = "../assets/logo.png"
+# 使用本地的图片
+url = image_to_base64(image_path)
+# 使用网络图片
+url = "https://opencompass.oss-cn-shanghai.aliyuncs.com/image/compass-hub/botchat_banner.png"
+
 # 新版本 opnai
 client = OpenAI(api_key="EMPTY", base_url="http://localhost:8082/v1")
 
@@ -17,7 +34,7 @@
                 {
                     "type": "image_url",
                     "image_url": {
-                        "url": "https://opencompass.oss-cn-shanghai.aliyuncs.com/image/compass-hub/botchat_banner.png",
+                        "url": url,
                     },
                 },
             ],