Pulls and Dependencies

A Pull's lifecycle

A Pull names an Upstream, a Repo, a Revision (the upstream default branch if you give none) and a file selection: everything, glob include and exclude patterns, explicit paths, GGUF quants, or for datasets Configs, Splits and Parquet-only.

1. Plan. Hangar resolves the Revision, fetches the upstream git history so Commits, trees and pointer files match exactly, lists the files with their sizes and SHA-256 values and applies the selection. Objects already anywhere in the Instance are marked present without a transfer. The Mirror's branches point at the upstream Commits as soon as planning finishes, so you can browse the history straight away.
2. Transfer. One pull.file Job per missing object streams it from the Upstream into the Blob store, hashing as it goes. Interrupted transfers resume with HTTP ranges, and an object that does not match its SHA-256 or size is rejected and retried, so nothing unverified becomes visible.
3. Finish. Hangar indexes and scans the new Commit, resolves Dependencies and fires the pull.completed webhook. If some files still failed after their retries, the Pull ends partial, with an error per file and a Retry failed action.

Small, non-LFS files come with the git history and need no transfer.

Network controls

Admins can cap the bandwidth Pulls use across the whole Instance, limit how many files transfer at once, and confine Pulls to time windows such as 01:00-07:00. See Pull windows and bandwidth.

Dependencies

After indexing a Commit, Hangar reads the files that name other Repos:

Source	Finds	Kind	Pulled automatically
auto_map in config.json, tokenizer_config.json and the other config files	Remote code in another Repo, as in jinaai/xlm-roberta-flash-implementation--modeling_lora.XLMRobertaLoRA	code	Yes, recursively (pull.dependencies.auto_pull_code)
adapter_config.json base_model_name_or_path	A PEFT adapter's base model	adapter_base	Only with pull.dependencies.auto_pull_base_models
Model card base_model	Where the model came from	card_base_model	No
Model card datasets	Training datasets	dataset	No
modules.json, model_index.json	sentence-transformers and diffusers components in other Repos	code	Yes

A Pull is not complete until its code Dependencies are, so from_pretrained(..., trust_remote_code=True) works offline right after it. Suggested Dependencies are listed on the Repo page and by hangar info, each with a one-click Pull.