NVIDIA RTX 5090 Kernel problem

[MorkBorlog]

Hi, I have an RTX 5090. Unfortunately, I'm getting errors as a kernel. Can you include this map? Because it's not that new anymore :)

[smimaks]

@MorkBorlog Hi! I have the same problm on 5070 and this how i fixed:

Conceptual fix for Forge on RTX 50xx / Blackwell GPUs

RTX 50-series GPUs use the Blackwell architecture (sm_120). Older PyTorch builds bundled with many Forge releases are built for CUDA 12.1/12.4 and do not support this architecture.

What needs to be done:

1. Use a recent NVIDIA driver with CUDA 12.8+ support.

2. Replace the PyTorch packages inside Forge’s Python environment with CUDA 12.8-compatible builds:

   • torch
   • torchvision
   • torchaudio

3. Make sure the packages are installed into Forge’s own Python/venv, not the system Python.

4. If dependency conflicts appear, pin older Forge-compatible versions of common packages such as:

   • numpy
   • Pillow
   • MarkupSafe
   • protobuf

5. Verify that PyTorch reports:

   • CUDA build: cu128
   • GPU capability: (12, 0)
   • CUDA available: True

In short:
Forge needs a PyTorch build with CUDA 12.8 / Blackwell (sm_120) support.