MXFP4 and MXFP8 loading support

[yiliu30]

• Support loading the quantized MXFP4/MXFP8 via transformers.from_pretrained` API.
• Other small fixes

[Copilot]

Pull Request Overview

This PR adds support for MXFP4 and MXFP8 quantization formats, including new quantization modules, backend configurations, and test coverage. The implementation introduces specialized quantized linear layers for MXFP4/8 formats with proper loading and inference support.

Key Changes:

• Adds new MXFP4QuantLinear and MXFP8QuantLinear modules with specialized quantization/dequantization logic
• Updates backend infrastructure to support MXFP4/8 schemes with proper device compatibility checks
• Implements FP4 utility functions for packing/unpacking operations

Reviewed Changes

Copilot reviewed 10 out of 10 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
auto_round/experimental/qmodules/mx.py	New quantized linear modules for MXFP4/8 with weight initialization and dequantization
auto_round/experimental/qmodules/init.py	Exports the new MXFP quantization modules
auto_round/data_type/fp4_utils.py	FP4 unpacking utilities for converting uint8 to fp4 values
auto_round/inference/backend.py	Backend configuration and scheme checking for MXFP4/8 support
auto_round/inference/convert_model.py	Model conversion logic updates for MXFP format handling
auto_round/schemes.py	Adds act_sym parameter to MXFP4/8 quantization schemes
auto_round/export/export_to_autoround/qlinear_fp.py	Updates weight buffer initialization for MXFP quantization
auto_round/export/export_to_autoround/export.py	Renames FP8_STATIC format enum and updates references
auto_round/autoround.py	Updates format references for consistency
test/test_cuda/test_mx_quant.py	End-to-end test for MXFP4/8 quantization and inference

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}

[wenhuach21]

@n1ck-guo @WeiweiZhang1 please have a review