#Unsloth doesn't detect rtx5060 on Windows

4 messages · Page 1 of 1 (latest)

vast salmonBOT
#

Your issue is that Unsloth (and PyTorch) do not detect your RTX 5060 GPU because PyTorch does not yet support CUDA 12.9. The latest supported CUDA versions for PyTorch and Unsloth are 12.8 and below. You need to install CUDA 12.8 and a matching PyTorch build (e.g., torch 2.7.0+cu128) for GPU support to work on Windows with RTX 50 series cards. CUDA 12.9 is not supported by Unsloth or PyTorch as of now, so torch.cuda.is_available() will always return False with CUDA 12.9 installed, regardless of your GPU or driver status.

See the official Unsloth Blackwell/RTX 50 series install guide, which specifies CUDA 12.8 as the required version and provides step-by-step instructions for both pip and conda environments. You must uninstall your current CUDA, install CUDA 12.8, and then install the correct PyTorch and Unsloth versions for compatibility. More details and troubleshooting steps are available in the Unsloth documentation and related GitHub issues for RTX 50 series support. Would you like a step-by-step guide for this process?

Sources:

spring wraith
#

solved

#

@vast salmon FastLanguageModel.from_pretrained

where models are saved?