If you want to use the NVIDIA GeForce RTX 5090 GPU with PyTorch ... So, I came to conclusion, VLLM chose kernel with native FP8 because first GPU (4090) has FP8 support, and then crashed because not ...

Did you update? pip install --upgrade unsloth unsloth_zoo Colab or Kaggle or local / cloud Number GPUs used, use nvidia-smi Which notebook? ... One GPU,4090 jupyter Qwen3(4B) GRPO SFTTrainer and ...