[GPT-OSS] Expanded Support for Activation Quantization

[kylesayrs]

Weight-only quantization support for GPT-OSS has recently been merged via [CPU] Linearize gpt_oss model and add example to quantize it to w4a8 from @isharif168 🔥.

There are a couple follow ups that will make this work even better and more expansive

1. Refactor convert_model_for_quantization_gptoss to use the MoECalibrationModule interface
2. Add support for the calibrate_all_experts option in the forward definition
3. Test the forward function with modifiers such as GPTQ or AWQ
4. Add an example of how to run the quantized model in vLLM (with reference to any necessary patches)
5. Potential code cleanup/simplification, described in comments here

[saraswatmks]

Is someone already planned to work on this? I would be interested to contribute here.

[dsikka]

I don’t believe anybody is! If you’re interested, let me know and I’ll assign it to you!

[saraswatmks]

I don’t believe anybody is! If you’re interested, let me know and I’ll assign it to you!

Yes please, want to work on this.

[dsikka]

Awesome, thank you