Troubleshooting

Troubleshooting

Error / symptom	Likely cause	Remedy
not a git repository	Extracted ZIP has no .git	Git-clone a target code repo; use the Wiki publishing launcher for a separate Wiki clone
.ps1 not digitally signed	Restricted PowerShell policy	Review and run the provided .cmd launcher for a process-local override
publish-wiki.ps1 not recognized	Wrong working folder	Run from package root containing both scripts and wiki
Stale .wiki-publish-* checkout	Older publishing script used fixed path	New script creates a unique temp folder and cleans only its own clone
ModuleNotFoundError	Missing dependencies in selected Python	python -m pip install -r requirements.txt
CUDA unavailable	CPU-only PyTorch, driver/device mismatch	Check torch CUDA build, GPU driver, CUDA wheel
CUDA out of memory	Too-large model replica / batch	Decrease batch size, input tokens, output tokens; ensure enough VRAM per GPU
Jupyter spawn issues	Worker function not importable	Run standalone Python program via subprocess
predicted_label: null	Non-one-word LLM completion	Inspect llm_output; adjust prompt/normalization carefully
Worker timeout	Cold model download or hang	Check cache/network, model memory; adjust --worker-timeout

The project does not use DDP or tensor parallelism, so model weights are not shared across GPU memories.