suryatmodulus
/

GPC-1 / docs /DOWNLOADS.md
suryatmodulus's picture harshatheg's picture
Duplicate from harshatheg/GPC-1
96a4100
|
Raw
History Blame Contribute Delete
2.31 kB
# Download GPC-1
Download the complete GPC-1 release: the GPC-1 backbone in `model/`, its matching adapter in `adapter/`, and the serving package. Keep both weight components and release metadata together; the server loads them automatically.
## Command line
```bash
python -m pip install "huggingface-hub>=1.7,<2"
hf download harshatheg/GPC-1 --local-dir gpc-1
python gpc-1/download.py verify gpc-1
```
For a reproducible deployment, pin a commit SHA and use the same revision in every environment:
```bash
hf download harshatheg/GPC-1 --revision YOUR_COMMIT_SHA --local-dir gpc-1
python gpc-1/download.py verify gpc-1
```
Private or gated access uses `HF_TOKEN` in the environment. Keep the token in your secret manager; do not put it in command arguments or logs.
## Python
If you already have `download.py` from the release package, run Python from its directory:
```python
from pathlib import Path
from download import fetch
package = fetch(
"harshatheg/GPC-1",
revision="YOUR_COMMIT_SHA",
local_dir=Path("gpc-1"),
cache_dir=None,
offline=False,
)
print(package["model"])
```
`fetch` uses `huggingface_hub.snapshot_download` for the complete repository. Verification checks that the root and packaged base configs match, all inventoried base files and indexed weight shards are present, and the runtime manifest, base metadata, and adapter hashes match. It does not hash every base weight shard; the server checks those hashes when it loads the model. No tracking endpoint is called.
## Cache and offline use
After an online download, verify a copied package without contacting the Hub:
```bash
python gpc-1/download.py verify gpc-1
```
For offline deployment, download the complete package with `--local-dir` while online, then copy that directory to the offline machine. Run the verification command above on the copied directory. Verification makes no Hub request. Use the materialized package directory, not a cache snapshot containing symlinks.
## CI
Pin the revision in your pipeline and keep the token in a CI secret. A complete package can then be mounted read-only by the serving process:
```bash
hf download harshatheg/GPC-1 --revision "$GPC1_REVISION" --local-dir gpc-1
python gpc-1/download.py verify gpc-1
export GPC1_MODEL_PATH="$PWD/gpc-1/model"
```