Improve GPT OSS Conversion Script #42902
Description
Feature request
The GPT OSS Conversion Script exposes parameters that are not needed or used, has incorrect documentation, and crashes due to a tiktoken bug.
I improved the script and validated it works with GPT-OSS-20B.
Motivation
The original motivation is from huggingface/accelerate#3882 (comment) where I would like to use accelerate to load a GPT-OSS-20B model. This does not work out of the box, however this conversion script helps. Unfortunately the script was broken and had incorrect documentation and dangerous arguments.