whichllm Files

plan owner/repo can now fetch public Hugging Face repositories outside the cached catalog.

Model grouping keeps organizations, model versions, checkpoint dates, and instruction/chat variants distinct. JSON family_id values change to preserve those distinctions.

This release also fixes the ROCm detection case where a generic GPU name hides a recognized shared-memory SKU, and documents the limits of multi-GPU memory-placement estimates.

Update with uv tool upgrade whichllm or pip install --upgrade whichllm.

Full changelog