Reliable LLM weight downloads on restricted networks: dlmodel's multi-source fallback
When downloading tens of gigabytes of weights on a private or restricted network, the hard part is not the download itself but what happens after an interruption. dlmodel is designed so that every step has a fallback.
Two levels of fallback
- Between sources: the auto source tries ModelScope first and falls back to a Hugging Face mirror (hf-mirror.com by default), no proxy needed.
- Within a source: aria2 multi-connection (16 connections per file by default) → single-stream resumable HTTP Range → the native SDK as a last resort. When aria2 is unavailable it steps down automatically and the task does not fail.
Pausing is a first-class state
Pausing is not cancelling: it stops the network stream safely and keeps the task, progress, downloaded files, resume points and logs, then continues through the aria2 control file, HTTP Range or SDK temporary files on resume. Pause / resume / cancel can interleave with queuing and downloading at any moment, and only one of them takes effect.
Restarts lose nothing
On shutdown, queued and running tasks are parked and saved, and requeued automatically at the next start; errors that retries cannot fix, such as a full disk or permission denied, are not retried automatically: free up space and click "Retry" on the failed task to continue from the resume point.
Verification and storage
After downloading, LFS files are checked against the SHA-256 recorded in the repository and saved in the vLLM layout, ready for vLLM or Ollama.
Learn more on the dlmodel product page.