The training and evaluation scripts assume CUDA is available. The default Qwen2.5-7B configuration uses 4 GPUs for training and tensor parallel size 2 for vLLM evaluation.
Some results have been hidden because they may be inaccessible to you
Show inaccessible results