sglang pipeline模式启动，加载模型出现:AutoImageProcessor.register() got multiple values for argument 'exist_ok'

[neoyxm]

System Info / 系統信息

linux: ubuntu 22.04
cuda: NVIDIA-SMI 550.144.03 Driver Version: 550.144.03 CUDA Version: 12.4
pypi:

sglang                    0.5.6.post3.dev1178+g4346db5fa
transformers              5.0.0.dev0
diffusers                 0.37.0.dev0

Who can help? / 谁可以帮助到您？

No response

Information / 问题信息

• The official example scripts / 官方的示例脚本
• My own modified scripts / 我自己修改的脚本和任务

Reproduction / 复现过程

1. 我安装了以下版本的pypi库:

sglang                    0.5.6.post3.dev1178+g4346db5fa
transformers              5.0.0.dev0
diffusers                 0.37.0.dev0

2. 然后 用命令: sglang serve --model-path /data/models/ZhipuAI/GLM-Image/

3. 出现:

Traceback (most recent call last):
  File "/data1/condaenvs/envs/glm-image/bin/sglang", line 7, in <module>
    sys.exit(main())
             ^^^^^^
  File "/data1/condaenvs/envs/glm-image/lib/python3.11/site-packages/sglang/cli/main.py", line 26, in main
    args.func(args, extra_argv)
  File "/data1/condaenvs/envs/glm-image/lib/python3.11/site-packages/sglang/cli/serve.py", line 52, in serve
    from sglang.multimodal_gen.runtime.entrypoints.cli.serve import (
  File "/data1/condaenvs/envs/glm-image/lib/python3.11/site-packages/sglang/multimodal_gen/__init__.py", line 4, in <module>
    from sglang.multimodal_gen.runtime.entrypoints.diffusion_generator import DiffGenerator
  File "/data1/condaenvs/envs/glm-image/lib/python3.11/site-packages/sglang/multimodal_gen/runtime/entrypoints/diffusion_generator.py", line 30, in <module>
    from sglang.multimodal_gen.runtime.launch_server import launch_server
  File "/data1/condaenvs/envs/glm-image/lib/python3.11/site-packages/sglang/multimodal_gen/runtime/launch_server.py", line 9, in <module>
    from sglang.multimodal_gen.runtime.entrypoints.http_server import create_app
  File "/data1/condaenvs/envs/glm-image/lib/python3.11/site-packages/sglang/multimodal_gen/runtime/entrypoints/http_server.py", line 21, in <module>
    from sglang.srt.managers.io_struct import VertexGenerateReqInput
  File "/data1/condaenvs/envs/glm-image/lib/python3.11/site-packages/sglang/srt/managers/io_struct.py", line 31, in <module>
    from sglang.srt.managers.schedule_batch import BaseFinishReason
  File "/data1/condaenvs/envs/glm-image/lib/python3.11/site-packages/sglang/srt/managers/schedule_batch.py", line 5, in <module>
    from sglang.srt.dllm.config import DllmConfig
  File "/data1/condaenvs/envs/glm-image/lib/python3.11/site-packages/sglang/srt/dllm/config.py", line 3, in <module>
    from sglang.srt.configs.model_config import ModelConfig
  File "/data1/condaenvs/envs/glm-image/lib/python3.11/site-packages/sglang/srt/configs/__init__.py", line 8, in <module>
    from sglang.srt.configs.janus_pro import MultiModalityConfig
  File "/data1/condaenvs/envs/glm-image/lib/python3.11/site-packages/sglang/srt/configs/janus_pro.py", line 634, in <module>
    register_image_processor(MultiModalityConfig, VLMImageProcessor)
  File "/data1/condaenvs/envs/glm-image/lib/python3.11/site-packages/sglang/srt/configs/utils.py", line 18, in register_image_processor
    AutoImageProcessor.register(config, None, image_processor, None, exist_ok=True)
TypeError: AutoImageProcessor.register() got multiple values for argument 'exist_ok'

Expected behavior / 期待表现

正常启动

[neoyxm]

应该是git里的transformer太新了，导致sglang和transformer不匹配， 能贴下你们用的transformer的版本不 ？

[maolixiaolin-ailab]

我的也报这个错误：
sglang 0.5.6.post3.dev1178+g4346db5fa
transformers 5.0.0.dev0
diffusers 0.37.0.dev0

[neoyxm]

我的也报这个错误： sglang 0.5.6.post3.dev1178+g4346db5fa transformers 5.0.0.dev0 diffusers 0.37.0.dev0

我看了下原因， 目前sglang不兼容 transformer git里的最新代码，但如果把transformer退回4.57.*, transformer又不支持glm-image. 死结了。

我不知道官方是怎么跑出来的。

[Dangzii]

直接修改报错文件/data1/condaenvs/envs/glm-image/lib/python3.11/site-packages/sglang/srt/configs/utils.py：
AutoImageProcessor.register(config, image_processor, exist_ok=True)

[neoyxm]

直接修改报错文件/data1/condaenvs/envs/glm-image/lib/python3.11/site-packages/sglang/srt/configs/utils.py： AutoImageProcessor.register(config, image_processor, exist_ok=True)

好吧，这也太粗暴了

[zRzRzRzRzRzRzR]

这个问题现在的sgl应该在修了，已经同步反馈

[neoyxm]

这个问题现在的sgl应该在修了，已经同步反馈

我很好奇，那开发团队当时是怎么跑过的？ 因为已经写到readme里了，说明应该是跑通过的。

[zRzRzRzRzRzRzR]

因为transformers的mian分支一直在更新，hf model card上有能过的commit