Windows RTX 5090 local generation crashes with access violation (0xC0000005) during model component load #50
Description
Summary
Local generation on Windows crashes the backend process with an access violation during model component loading on an RTX 5090.
The crash is not a handled Python exception. The backend exits with Windows code 3221225477 (0xC0000005).
Environment
- OS: Windows 11
- GPU: NVIDIA GeForce RTX 5090
- NVIDIA driver:
591.74 nvidia-smiCUDA version:13.1- App: LTX Desktop
1.0.1 - Bundled Python:
3.13.12 - Torch:
2.10.0+cu128 - cuDNN:
91002 ltx-core:1.0.0ltx-pipelines:1.0.0sageattention:1.0.6triton-windows:3.6.0.post25
What happens
Starting a local text-to-video generation causes the backend to crash and restart.
The failure happens even after disabling:
- SageAttention
- FP8 quantization
I also reproduced the failure outside the app process using the bundled Python/runtime, so this does not appear to be Electron/UI-specific.
Repro steps
- Launch LTX Desktop on Windows with the local backend.
- Use local generation with the downloaded fast model.
- Start a text-to-video generation, for example:
- resolution:
540p - fps:
24 - frames:
121
- resolution:
- Backend crashes mid-generation.
Observed logs
Session log excerpt:
[t2v] Generation started (model=fast, 960x544, 121 frames, 24 fps)
[t2v] Pipeline load: 0.00s
[t2v] Text encoding (local): 0.00s
[fast-native] Loading text encoder
...
Python backend exited with code 3221225477
In another run it progressed slightly further:
[fast-native] Loading text encoder
[fast-native] Loading video encoder and transformer
Python backend exited with code 3221225477
After forcing safer paths, startup confirmed:
SageAttention: disabled
FP8 quantization: disabled
but the crash still occurred.
Standalone repro outside the app
Using the bundled Python/runtime, I can reproduce the failure while loading model components directly.
This survives:
pipeline creation
text_encoder() load
and then crashes while loading other core model components such as:
video_encoder()
or, in later iterations, even earlier during local text-encoder load depending on patching path
Notes
This looks like a native runtime/kernel/access violation rather than a normal Python exception.
Given the stack, this may be related to:
Windows + Blackwell/RTX 5090 support maturity
PyTorch/cu128/driver interaction
LTX local runtime compatibility on this platform
Requested help
Please advise:
Whether RTX 5090 on Windows is currently supported for local LTX Desktop generation
Whether a different recommended runtime stack exists for Blackwell
Whether there is a known fix/workaround
Whether Linux/WSL2 is currently the expected path for local generation on RTX 5090