[Feature] Add train/eval/inference configs for OpenAI gpt-oss models

[wizeng23]

Feature request

Plan to support both 20B and 120B.

Upgrades we need to make on our side to support all recommended optimizations in the HF blog:

• Job configs should use H100 GPUs at least, due to mxfp4 quant
• Need torch 2.7/2.8 (we have 2.6) -- 2.8 is recommended, and was just released today
• Need to upgrade our transformers version to 4.55
• Also need accelerate, kernels, triton 3.4, and triton_kernels
• Make sure we can support adjustable reasoning levels (seems to be set as the system prompt via the chat template)
• If not using remote/vllm/HF inference, need to support Harmony response format.
• Support Flash Attention 3 w/ attention sinks from vLLM. It seems the attn_implementation field was generalized to now support pulling in arbitrary kernels from HF Hub; we should add support for this
• Ensure it works on vLLM. I get an error right now following their instructions here: TypeError: flash_attn_varlen_func() got an unexpected keyword argument 's_aux'

Motivation / references

https://huggingface.co/blog/welcome-openai-gpt-oss
#1661
https://openai.com/open-models/
https://cookbook.openai.com/articles/gpt-oss/fine-tune-transfomers
https://github.com/huggingface/gpt-oss-recipes/blob/main/sft.py

Your contribution

PR