Hugging Face Repositories

When using Nexus Repository as a proxy, files are downloaded to the repository before they are returned to the requesting to the client.

Large models take time to download, especially during the initial caching process. Clients may timeout when using the default timeout configuration. To prevent timeout errors, configure the following environment variable used by the client:

export HF_HUB_DOWNLOAD_TIMEOUT=120
export HF_HUB_ETAG_TIMEOUT=1800

Hugging Face may be slow and cause Nexus Repository to trigger timeout errors. As mentioned in the proxy configuration we recommend setting the proxy HTTP request timeout to 120 seconds.

As an alternative to environment variables, you may configure the client timeout settings directly in the snapshot_download request:

from huggingface_hub import snapshot_download
snapshot_download(repo_id="microsoft/OmniParser", repo_type="model", local_dir="/Users/Documents/hugging", etag_timeout=900)

Set timeout for transformers and diffusers:

from diffusers import StableDiffusionPipeline
from huggingface_hub import configure_http_backend
from requests.adapters import HTTPAdapter
from requests import Session
from urllib3.util.retry import Retry

timeout = 900
session = Session()
retries = Retry(total=5, backoff_factor=0.5, status_forcelist=[500, 502, 503, 504])

class TimeoutHTTPAdapter(HTTPAdapter):
    def __init__(self, *args, **kwargs):
        self.timeout = kwargs.pop("timeout", timeout)
        super().__init__(*args, **kwargs)

    def send(self, request, **kwargs):
        kwargs["timeout"] = self.timeout
        return super().send(request, **kwargs)

adapter = TimeoutHTTPAdapter(max_retries=retries, timeout=timeout)
session.mount("https://", adapter)
session.mount("http://", adapter)

def get_custom_session():
    return session

configure_http_backend(get_custom_session)
pipeline = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4")

Hugging Face uses Git commit hashes for versioning. Nexus Repository reflects these commit hashes when storing different model versions.

Example of Hugging Face using GIT commit hash

from huggingface_hub import snapshot_download
snapshot_download(repo_id="keras-io/deeplabv3p-resnet50", revision="9de024eda886161906e18b30cf84d8bcd8a0664c")

On Nexus Repository you will see the same Hugging Face commit hashes when downloading more than one version of the same model.

A two-node HA cluster (r6i.xlarge instances) with an RDS cluster (db.t3.medium with one writer and one read replica).

Due to the prolonged download times, configuring adequate HTTP connection timeout settings for Hugging Face proxies is recommended. Nexus Repository outperformed connecting directly to the Hugging Face Hub for large assets.