Fix equality-vs-assignment bug in GptqHfQuantizer.update_device_map

[gapsong]

Fixes a typo in GptqHfQuantizer.update_device_map where an equality comparison (==) was used instead of an assignment (=). Due to this, device_map was never updated to CUDA when using auto-gptq (which does not support CPU), leaving the model on CPU contrary to the intent expressed in the comment.

• File: src/transformers/quantizers/quantizer_gptq.py
• Function: GptqHfQuantizer.update_device_map

Before:

if not is_gptqmodel_available() and device_map in ("cpu", {"": torch.device("cpu")}):
    device_map == {"": 0}

After:

if not is_gptqmodel_available() and device_map in ("cpu", {"": torch.device("cpu")}):
    device_map = {"": 0}

This change makes the code match the intended behavior: when gptqmodel is not available and the model would otherwise be placed on CPU, move to CUDA device map ({"": 0}).

Who can review?

• quantization (bitsandbytes, autogpt, gptqmodel): @SunMarc @MekkCyber

Anyone in the community is also welcome to review.

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: gptq

[MekkCyber]

Thanks for the fix 🤗 ! Sorry a PR was already merged to fix this issue

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.