A hf download that stalled or died does not need to start over. Re-run the exact same command:
finished files are skipped from the cache and the interrupted file continues from its .incomplete
partial. If a partial is corrupt, prune it and pull again.
1. Re-run the same command
hf download bartowski/Meta-Llama-3.1-8B-Instruct-GGUF Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf
Completed files come straight from the cache and the stalled transfer picks up where it stopped, so nothing is fetched twice.
2. Clear broken partials
hf cache ls
Shows cached repos and flags leftover incomplete downloads with a hint.
hf cache prune
Deletes unreferenced revisions and every leftover .incomplete partial, which clears a corrupt
half-file so the next download is clean (hf cache rm never touches partials).
3. Speed it up with hf_xet
pip install -U huggingface_hub
Installs hf_xet alongside the library, which enables chunked, resumable transfers automatically —
no flag needed.
The older hf_transfer package and its HF_HUB_ENABLE_HF_TRANSFER=1 switch are now deprecated; use
hf_xet instead.
4. Check disk and cache location
export HF_HOME=/mnt/big/hf
Moves the whole cache to another drive; the default is ~/.cache/huggingface/hub, which fills the
home partition fast. See where the Hugging Face cache lives
to confirm the path first.
Verify it worked
hf cache verify bartowski/Meta-Llama-3.1-8B-Instruct-GGUF
Checks the cached files against the Hub checksums and reports any mismatch. If it fails:
- Download keeps timing out — raise the timeout from its default of 10 seconds and re-run:
export HF_HUB_DOWNLOAD_TIMEOUT=30. - 401 or “gated repo” — accept the licence and log in first; see access gated models.
- No space left on device — run
hf cache prune, or pointHF_HOMEat a larger disk.
Next: the full hf download workflow. Cache internals: Hugging Face’s cache guide.