HMCLIP / BACKUP_RESTORE.md
SteveZeyuZhang's picture
Add source files and streaming backup restore instructions
cf7e640
|
Raw History Blame Contribute Delete
1.29 kB
# RadIR snapshot backup
This repository displays the source code and stores the large data, model, output,
cache, and log directories as a split `tar.zst` archive in Git LFS.
- Source commit: `5b926d2179afa417784ba57459e86d2336d50a52`
- Archived paths: `dataset, outputs, hf_models, hf_cache, runs, wandb, __pycache__`
- Each volume is at most 64 MiB and has a SHA-256 checksum in `backup-manifest.json`.
- This is a snapshot of committed files, not an export of all historical Git objects.
- Original source files and the original local Git history were not modified.
## Restore
Wait until `backup-manifest.json` says `complete: true`. Download the repository
with its LFS objects (not just pointer files). Install `zstd` and `tar`, then run:
```bash
python3 restore_backup.py
```
The restore tool verifies every volume, then streams decompression directly into
`tar`. It does not create a second concatenated archive. Extracted contents keep
their original relative paths. Source code is already available as regular files.
The uploader uses `hf-mirror.com` for Git and LFS metadata. LFS payloads use the
storage URLs returned by that service, as in a normal Git LFS transfer. It keeps
at most 16 archive volumes in memory (1 GiB total), plus small Git metadata and progress files.