[BUG] Cuda OOM with gpt-oss-20b #959
Description
Describe the bug
A clear and concise description of what the bug is.
lighteval accelerate "model_name=openai/gpt-oss-20b,max_length=4096,skip_special_tokens=False,batch_size=1" "leaderboard|mmlu|5" --save-details --output-dir "openai_scores"
goes OOM (lighteval installed from main, kernels 0.10.1, accelerate 1.6.0, transformers 4.56.1)
using a RTX 5090
can this be fixed with some other settings?