Add GPT-OSS quant support

[yiliu30]

.

[Copilot]

Pull Request Overview

This PR adds GPT-OSS quantization support to the auto_round library. The implementation includes a new MoE (Mixture of Experts) converter for GPT-OSS models along with comprehensive test coverage.

• Creates specialized handling for GPT-OSS models by converting fused expert operations to individual expert modules for quantization
• Refactors the MoE converter architecture to support multiple model types through a dispatch table
• Adds comprehensive test coverage for GPT-OSS quantization with MXFP4 and MXFP8 schemes

Reviewed Changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
test/test_cpu/test_gpt_oss.py	New test file with fixtures and parametrized tests for GPT-OSS quantization
auto_round/utils.py	Added debug logging for non-quantized layers
auto_round/special_model_handler.py	Refactored MoE converter to use dispatch table and added GPT-OSS support
auto_round/modelling/llama4.py	Extracted Llama4 MoE converter to dedicated module
auto_round/modelling/gpt_oss.py	New GPT-OSS MoE converter with specialized expert handling
auto_round/modelling/init.py	New package initialization file

---

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

[Copilot]

The magic numbers ::2 and 1::2 for tensor slicing should be replaced with named constants like GATE_STRIDE = 2 and GATE_OFFSET = 0, UP_OFFSET = 1 to improve code readability and maintainability.

[wenhuach21]

It would be better not to categorize it into too many detailed types. A single flag like model_need_to_convert, or a similar name, should be sufficient, since some models may require conversion even if they don’t have expert layers. We provide a converter function for each model if needed, regardless of which parts need to be converted.

[yiliu30]

Yeah, I agree, the replacement code could be organized better. Once we support more model replacements, we can refactor that part as needed. For now, how about leaving it as is, since we have some higher-priority tasks to focus on?

[wenhuach21]

I don’t think it will take much effort to change. You could also finish the higher-priority tasks first.

[yiliu30]

Opened an issue to track #899