huggingface config_error ai_generated true

ValueError: Cannot use device_map='auto' with pipeline when device is explicitly set to 0. Please set device_map=None or remove device argument.

ID: huggingface/pipeline-device-map-conflict

Also available as: JSON · Markdown · 中文
92%Fix Rate
90%Confidence
1Evidence
2024-03-22First Seen

Version Compatibility

VersionStatusIntroducedDeprecatedNotes
transformers>=4.30.0 active — — —
accelerate>=0.20.0 active — — —

Root Cause

The pipeline API conflicts when both `device` and `device_map` are specified; `device_map='auto'` requires the model to be loaded with Accelerate's device mapping, which is incompatible with a fixed device index.

generic

中文

当同时指定 `device` 和 `device_map` 时，pipeline API 发生冲突；`device_map='auto'` 要求模型使用 Accelerate 的设备映射加载，这与固定的设备索引不兼容。

Official Documentation

https://huggingface.co/docs/transformers/v4.30.0/en/main_classes/pipelines#pipeline-device-map

Workarounds

  1. 95% success Remove the `device` argument and use only `device_map='auto'`: pipe = pipeline('text-generation', model='model-name', device_map='auto'). This lets Accelerate handle device placement. 
    Remove the `device` argument and use only `device_map='auto'`: pipe = pipeline('text-generation', model='model-name', device_map='auto'). This lets Accelerate handle device placement.
  2. 90% success Remove `device_map` and use `device=0` explicitly: pipe = pipeline('text-generation', model='model-name', device=0). This forces the model onto GPU 0. 
    Remove `device_map` and use `device=0` explicitly: pipe = pipeline('text-generation', model='model-name', device=0). This forces the model onto GPU 0.

中文步骤

  1. 移除 `device` 参数，仅使用 `device_map='auto'`：pipe = pipeline('text-generation', model='model-name', device_map='auto')。这使 Accelerate 处理设备放置。
  2. 移除 `device_map` 并显式使用 `device=0`：pipe = pipeline('text-generation', model='model-name', device=0)。这强制模型使用 GPU 0。

Dead Ends

Common approaches that don't work:

  1. Set both `device=0` and `device_map='auto'` and expect the pipeline to resolve the conflict 100% fail

    The pipeline raises a ValueError immediately because the two arguments are mutually exclusive.

  2. Use `device_map='sequential'` instead of 'auto' 90% fail

    The same conflict applies; any non-None device_map with an explicit device argument will raise an error.

  3. Set `device=-1` to use CPU and keep device_map='auto' 80% fail

    device_map='auto' requires at least one GPU to be available; setting device=-1 will cause a separate error about no GPU found.