Refactor FP8 dequantization and detection using registry pattern

[scopophobic]

Refactor FP8 dequantization and detection using registry pattern

Summary

Refactors FP8 dequantization and detection logic to use a registry pattern, making FP8 support more explicit, maintainable, and extensible. Pure refactoring - no behavior changes, fully backward compatible.

What Changed

• Added FP8_DEQUANT_REGISTRY with @register_fp8_layer() decorator
• Registered CompressedLinear and FP8Linear handlers
• Refactored is_fp8_linear() to use registry-first detection
• Refactored convert_fp8_layer_to_linear() to use registry dispatch
• Updated convert_fp8_model_to_16b_model() to support all registered types

Benefits

• Explicit: Only registered layer types are supported
• Consistent: Detection and dequantization stay in sync
• Extensible: Adding new FP8 types requires just a handler + registration
• Maintainable: Clear separation of concerns

Files Changed

• auto_round/utils/model.py (only file modified)

Type: Refactoring | Breaking: No | Migration: No

[yiliu30]

Hi @scopophobic, the CI is currently blocked due to the Transformers v5 upgrade. I’ll get back to you once it’s fixed.