# Windows Ml Troubleshooting

> Debug CUDA crashes and Python venv contamination on Windows.

- Skill: `wcpaka-lgtm/windows-ml-troubleshooting` (Agent Skill)
- Install (CLI): `npx skillmds@latest add wcpaka-lgtm/windows-ml-troubleshooting`
- Raw SKILL.md: https://api.skillmd.com/api/skills/wcpaka-lgtm/windows-ml-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/windows-ml-troubleshooting

---


# Windows ML Troubleshooting

Diagnose and fix Python ML/AI tool crashes on Windows: CUDA version mismatches,
native DLL failures, and environment contamination when calling subprocess Python.

## When to use

- Python ML tool crashes with `0xC0000005` (access violation) or `0xC0000135` (DLL not found)
- `torch.cuda.is_available()` returns `False` but `nvidia-smi` shows a working GPU
- `cudaGetDeviceCount()` returns `cudaErrorNotSupported`
- Running a different venv's Python but it loads packages from the wrong environment

---

## 1. CUDA / PyTorch Version Mismatch → 0xC0000005 Crash

**Symptom:** ComfyUI (or any PyTorch app) crashes on startup with:
```
Windows fatal exception: access violation
Process exited with code 3221225477 / 0xC0000005
Stack (most recent call first):
  File "...torch\cuda\__init__.py", line 491 in _lazy_init
```
And/or:
```
cudaGetDeviceCount() returned cudaErrorNotSupported, likely using older driver
```

**Root cause:** PyTorch was installed with a CUDA build (e.g., `cu130`) that is
**newer** than the driver's supported CUDA version. The NVIDIA driver reports
its max CUDA version via `nvidia-smi`; PyTorch MUST be built with a CUDA version
≤ that number.

**Diagnosis:**
```bash
# Check driver's max CUDA version
nvidia-smi | grep "CUDA Version"
# → "CUDA Version: 12.9"   ← this is the MAX allowed

# Check what PyTorch was built with
python -c "import torch; print('CUDA built:', torch.version.cuda)"
# → "CUDA built: 13.0"     ← HIGHER than driver → CRASH
```

**Fix:** Uninstall and reinstall PyTorch with a compatible CUDA build.
`cu124` (CUDA 12.4) is the safest choice for most drivers (CUDA 12.0–12.9):
```bash
pip uninstall torch torchvision torchaudio -y
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124
```

**Verify:**
```bash
python -c "import torch; print(torch.cuda.is_available(), torch.cuda.get_device_name(0))"
# → "True NVIDIA GeForce GTX 1080 Ti"
```

### CUDA version compatibility table

| Driver CUDA | Safe PyTorch build | Notes |
|---|---|---|
| ≤ 11.8 | `cu118` | Legacy cards |
| 12.0–12.4 | `cu121` or `cu124` | `cu124` is most stable |
| 12.5–12.9 | `cu124` | `cu124` is forward-compatible |
| ≥ 13.0 | `cu130` | Very new drivers only |

---

## 2. PYTHONPATH Contamination (Wrong venv Packages Loaded)

**Symptom:** You call a different venv's Python (e.g., ComfyUI's `.venv/Scripts/python.exe`),
but it loads packages from the Hermes venv instead.

**Root cause:** Hermes sets `PYTHONPATH`, `PYTHONHOME`, and `VIRTUAL_ENV` environment
variables that can bleed into subprocess calls and override the target venv's isolation.

**Fix:** When running Python from a different venv via `subprocess.run()` or `terminal()`,
always clear these environment variables first:

```python
import subprocess, os

env = os.environ.copy()
env.pop('PYTHONPATH', None)
env.pop('PYTHONHOME', None)
env.pop('VIRTUAL_ENV', None)

result = subprocess.run(
    [r"E:\target\.venv\Scripts\python.exe", "-c", "import torch; print(torch.__file__)"],
    capture_output=True, text=True,
    env=env
)
```

**Verification:** The loaded `torch.__file__` should point to the target venv's
`site-packages`, not the Hermes venv.

---

## 3. General DLL Load Failures

**Symptom:** `OSError: [WinError 126] 지정된 모듈을 찾을 수 없습니다` when importing
torch or other native libraries.

**Common causes:**
- Missing Visual C++ Redistributable (install from https://aka.ms/vs/17/release/vc_redist.x64.exe)
- Missing CUDA toolkit DLLs (check `CUDA_PATH` environment variable)
- cuDNN not in PATH (should be in `CUDA_PATH\bin`)

---

## Pitfalls

- **Never assume `cu130` is safe** — many driver versions don't support CUDA 13.0 yet.
  Always check `nvidia-smi` first, then pick the matching PyTorch build.
- **`nvidia-smi` working ≠ PyTorch CUDA working** — the driver can be healthy but
  the PyTorch build can still be incompatible. Always verify with
  `torch.cuda.is_available()`.
- **Meta Virtual Monitor (Oculus/Meta Quest)** can interfere with CUDA device
  enumeration. If `nvidia-smi` shows the GPU but PyTorch can't see it, the virtual
  display adapter may be confusing CUDA initialization.