add mxfp8 qat

[lkk12014402]

Type of Change

add mxfp8 QAT code

Description

1. support mxfp8 forward and bf16 backward for QAT
2. export quantized hf model to llm-compressor, which can be deployed using vllm with mxfp8
3. add a simple qat example for testing currently
4. leverage some functions from AutoRound

[yiliu30]

Please add UTs to https://github.com/intel/neural-compressor/tree/master/test/3x/torch

[yiliu30]

Where is this function used?

[lkk12014402]

please check the usages in the https://github.com/intel/neural-compressor/pull/2299/files#diff-c587158e3584c62ab2d6cf97658da96ec6c88a1621aef0cd330ee16cfaa2da1aR190

[Copilot]

Pull Request Overview

This PR adds MxFP8 Quantization-Aware Training (QAT) support to Neural Compressor, enabling training with mxfp8 forward pass and bf16 backward pass, along with model export capabilities for VLLM deployment.

Key changes include:

• Implementation of MxFP8 QAT infrastructure with tensor quantizers and quantized linear layers
• Export functionality to convert quantized models to llm-compressor format for VLLM deployment
• Integration with AutoRound library for quantization functions

Reviewed Changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 7 comments.

Show a summary per file

File	Description
neural_compressor/torch/export/export_hf.py	Adds HuggingFace model export functionality for quantized models
neural_compressor/torch/algorithms/qat/tensor_quantizer.py	Implements core tensor quantization module with MxFP8 support
neural_compressor/torch/algorithms/qat/quant_utils.py	Provides utility functions for quantization configuration and model conversion
neural_compressor/torch/algorithms/qat/quant_linear.py	Implements quantized linear layer with weight/input/output quantization
neural_compressor/torch/algorithms/qat/init.py	Package initialization file for QAT module

---

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

[yiliu30]

Please add a simple end-to-end unit test to demonstrate the usage. Other submodules look good to me.