# Local Model Troubleshooting

> Diagnose PyTorch CUDA mismatches and diffusion model issues.

- Skill: `wcpaka-lgtm/local-model-troubleshooting` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add wcpaka-lgtm/local-model-troubleshooting`
- Raw SKILL.md: https://api.skillmd.com/api/skills/wcpaka-lgtm/local-model-troubleshooting/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: wcpaka-lgtm (https://skillmd.com/u/wcpaka-lgtm)
- Updated: 2026-09-21
- Page: https://skillmd.com/skills/wcpaka-lgtm/local-model-troubleshooting

---


# 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
```bash
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
```bash
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:
```bash
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:
```bash
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).