Stuck on either loading safetensors or fetching the files.

[tedcar]

I'm experiencing inconsistent behavior while using an NVIDIA L4 GPU. The main issues are:

1. SafeTensors Loading Problem:

   • Loading SafeTensors gets stuck and can take hours instead of the expected 2 minutes
   • Sometimes it freezes at the file fetching stage

2. Installation Method Issues:

   • Direct GitHub clone method doesn't work reliably
   • HuggingFace space clone worked perfectly the first time
   • Subsequent attempts using the same HuggingFace space method resulted in the same errors
   • It seems very hit-or-miss with no clear pattern

3. Environment Details:

   • I'm running this on Lightning AI's cloud service (paid subscription)
   • Have access to high-end GPUs

4. VRAM Usage Concerns:

   • When it does work, it only uses 15GB of VRAM
   • I have access to GPUs with 25GB+ (sometimes even 100GB+) of VRAM
   • Image editing takes 1-1.5 minutes, which seems slow given the available resources
   • There should be an option to allocate more VRAM to improve processing speed

As someone who isn't particularly technical, I'd appreciate if there was a built-in feature to control VRAM allocation for better performance. Could you please investigate these issues, particularly the inconsistent loading behavior and VRAM utilization?

[staoxiao]

Hi, @tedcar, thanks for your suggestions! We will further analyze these issues to improve the user experience. For loading behavior, the script will download files from HF the first time you run it, so it might take some time. On subsequent runs, it will read the files directly from the local storage if they are available, so it won't need to download them again.

[tedcar]

Appreciate that. This is still cutting-edge technology and I appreciate you guys making this open source