Skip to content
hal0

Hugging Face pulls

When you assign a model to a slot that isn’t already in the registry, hal0 pulls it from Hugging Face. The pull surfaces as a slot-level state transition (offline → pulling) and a byte-level SSE stream so the dashboard and CLI can render a live progress bar.

Pulling a model

Section titled “Pulling a model”

Three ways:

  • Dashboard. The Models view has a Pull button. Paste a Hugging Face repo ref (e.g. bartowski/Qwen2.5-Coder-7B-Instruct-GGUF) and pick the quant file.
  • CLI.
    Terminal window
    hal0 model pull bartowski/Qwen2.5-Coder-7B-Instruct-GGUF \
      --file Qwen2.5-Coder-7B-Instruct-Q4_K_M.gguf
  • Slot swap.
    Terminal window
    hal0 slot swap primary --model qwen2.5-coder-7b-instruct-q4_k_m
    If the registry doesn’t have it, the slot transitions through pulling before warming.

Where the bytes land

Section titled “Where the bytes land”

Models are written to /var/lib/hal0/models/<safe-ref>/<file> with a checksum sidecar. On a successful pull the registry entry is created atomically; a failed pull leaves no partial entry.

/var/lib/hal0/ survives hal0 update (only /usr/lib/hal0/current/ gets swapped), so pulled models persist across version upgrades. If you’re mounting the model store from an NFS export on your NAS, the path stays stable through upgrades too.

Progress streaming

Section titled “Progress streaming”

The dashboard and CLI subscribe to an SSE stream that emits one event per progress tick:

  • Total bytes
  • Bytes received
  • Throughput (bytes / second)
  • Elapsed time
  • ETA

The slot itself stays in pulling until the file is fully verified; only then does it transition to warming.

Status today

Section titled “Status today”

POST /api/models/{id}/pull is live. It queues a background pull job, streams the file from Hugging Face, verifies sha256, and stages atomically into the registry. Poll GET /api/models/{id}/pull/status for progress, or POST .../pull/cancel to abort. The CLI (hal0 model pull) and the FirstRun wizard both drive this same code path.

FLM tags (the Ollama-style family:size ids served by the FLM toolbox) also go through POST /api/models/{id}/pull. The route detects them via is_flm_tag() and shells flm pull <tag> inside the toolbox image instead of hitting Hugging Face.

Coming soon — outline

Section titled “Coming soon — outline”
  • Repo authentication for gated models (HF_TOKEN plumbing).
  • Multi-file pulls (sharded GGUFs).
  • Resume on interrupt.
  • Disk-space pre-flight warning.
  • Mirror configuration for self-hosted HF caches.