Expand description
hearth pull — get the bytes, prove they are the right bytes, and write
down where they came from.
This is the half of the aggregator that was missing. hearth_resolve is
deliberately network-free: it turns a reference into blob URLs, a chosen
quantization and a byte count, with serde and serde_json as its entire
dependency list. That is what lets the VRAM planner refuse a model before
forty gigabytes move. This crate is what moves them.
Three properties, and the third is the one nobody else offers:
- Digest-verified. Every blob is hashed and compared against the digest the registry published. A download that does not match is deleted, not kept — a corrupt GGUF fails much later as an unexplainable tokenizer error, three layers away from the truth.
- Resumable. A 40 GiB transfer that dies at 39 GiB resumes. Restarting is not a recovery strategy, it is a way to never finish on a bad link.
- Recorded.
PullStartedandPullCompletedland in the spine with the resolved source, so months later “where did this file come from” is a query —hearth why <model>— instead of an archaeology project. Those two transitions already existed inhearth_store; the history was ready before the downloader was.
Re-exports§
Modules§
- curl
- HTTP by driving
curl, not by linking one. - registry
- Turning a reference into a list of blobs to fetch.
- runtime
hearth runtime— fetch a prebuilt llama-server instead of hazing the operator with a compiler.
Structs§
- Pull
Config - Where blobs live, and how loud to be about it.
- Pulled
- What a pull did.
Functions§
- blob_
path - Content-addressed path for a digest. The filename IS the digest, which is
what makes “do we already have this” a
statrather than a re-hash. - fetch_
blob - Fetch one blob to its content-addressed home, verifying the digest.
- human_
bytes - Bytes as something a person can compare to a download page.
- pull
- Pull a model reference, recording the whole thing in the spine.