wcpaka-lgtm Local AI Model Troubleshooting
Diagnose and fix common issues when running video/image diffusion models locally on consumer GPUs.
PyTorch CUDA Version Mismatch (0xC0000005 / access violation)
Symptom
Windows fatal exception: access violation
File "torch\cuda\__init__.py", line 491 in _lazy_init
And/or: cudaGetDeviceCount() returned cudaErrorNotSupported
Diagnosis
nvidia-smi # Driver CUDA version
python -c "import torch; print(torch.version.cuda)" # PyTorch CUDA version
If PyTorch's CUDA version > driver's CUDA version → mismatch.
Fix
pip uninstall torch torchvision torchaudio -y
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124
Known working: GTX 1080 Ti + driver 576.28 (CUDA 12.9) → torch 2.6.0+cu124
Large Downloads with Rate Limiting
Prevent disconnects on unstable connections:
curl -L --limit-rate 5M -o output.safetensors "https://huggingface.co/.../resolve/main/..."
Speed Optimization for Diffusion Models
Reduce scheduler steps: 20 → 10 nearly halves generation time with minimal quality loss. Below 5 causes noticeable degradation.
Model File Verification
After downloading large models, verify sizes match expected:
ls -lh /path/to/models/
Partial downloads are common — re-download if significantly smaller than expected.
References
references/minimax-h3.md— MiniMax H3 local setup (ComfyUI on GTX 1080 Ti, model files, resolution/time table), WanGP low-VRAM alternative (GGUF quant, verified VRAM claims, Pinokio launchers), Apple Silicon/Mac status (no MPS backend as of Aug 2026), H3 architecture notes (33B / 13B AdaLN-cacheable → ~20B effective; 768p local ceiling).