HologramHologram

Connect

Hugging Face tools

Point the tools you already have at the hub with one environment variable.

Markdown

Your code already downloads models through huggingface_hub. It reads one variable to decide where from. Set it and nothing else changes: same commands, same cache directory, same file layout.

export HF_ENDPOINT=https://gethologram.ai

What that reaches

Tool Setting Verified
hf CLI, huggingface_hub (hf_hub_download, snapshot_download, HfApi) HF_ENDPOINT 1.x and 0.36
transformers, diffusers, sentence-transformers (from_pretrained) HF_ENDPOINT metadata and download path
vLLM, SGLang, TGI, KServe hf:// HF_ENDPOINT same client path; a full serving run has not been recorded
llama.cpp -hf, llama-server -hf MODEL_ENDPOINT=https://gethologram.ai/ build 11028
transformers.js, @huggingface/hub in a browser the URL, cross-origin every read route sends Access-Control-Allow-Origin: *

docker model pull hf.co/…, LM Studio and Jan hardcode huggingface.co and cannot be pointed here.

Download a model

export HF_ENDPOINT=https://gethologram.ai
hf download hexgrad/Kokoro-82M

Or from Python:

import os
os.environ["HF_ENDPOINT"] = "https://gethologram.ai"

from huggingface_hub import HfApi, hf_hub_download, snapshot_download

api = HfApi()
print([m.id for m in api.list_models(search="kokoro", limit=3)])
# ['hexgrad/Kokoro-82M', 'oddadmix/Kokoro-7M-Distill']

info = api.model_info("hexgrad/Kokoro-82M")
print(info.sha, len(info.siblings))
# f3ff3571791e39611d31c381e3a41a3af07b4987 72

path = hf_hub_download("hexgrad/Kokoro-82M", "EVAL.md")
local = snapshot_download("hexgrad/Kokoro-82M")

main resolves to the indexed revision, so snapshot_download lands in snapshots/f3ff3571…/, the same directory a download from huggingface.co of that commit would use. The two caches are interchangeable.

Load a model

import os
os.environ["HF_ENDPOINT"] = "https://gethologram.ai"

from transformers import AutoConfig, AutoTokenizer
config = AutoConfig.from_pretrained("Qwen/Qwen3-0.6B")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-0.6B")
print(config.model_type, tokenizer("hub")["input_ids"])
# qwen3 [26682]

from_pretrained for weights takes the same path; it needs a framework such as PyTorch installed, which the hub does not change.

llama.cpp

MODEL_ENDPOINT=https://gethologram.ai/ llama-server -hf bartowski/MiniCPM5-2B-GGUF:IQ2_M

The GGUF lands in the shared Hugging Face cache layout and loads.

Verify afterwards

huggingface_hub checks the size of each file, not its hash. For the check that matters, ask it to hash what it downloaded against the tree:

hf download hexgrad/Kokoro-82M
hf cache verify hexgrad/Kokoro-82M

It checks a large file by the SHA-256 in lfs.oid and a small one by the git blob SHA-1 in oid, and exits non-zero on any mismatch.

Search from the command line

The list route is Hugging Face's, with the same filters and the same rows. It answers in a few hundred bytes, not a catalogue.

curl -s "https://gethologram.ai/api/models?search=qwen&pipeline_tag=text-generation&filter=gguf&sort=downloads&limit=5"
Parameter Meaning
search case-insensitive substring of the model id
author exact <owner>
pipeline_tag exact task, e.g. text-generation, text-to-speech
library exact library, e.g. transformers, gguf
filter a tag that must be present, e.g. gguf, apache-2.0; repeat or comma-separate for several
sort trendingScore (default), downloads, likes, createdAt
direction 1 ascending; anything else descending
limit 1 to 500, default 50

What is not here

The hub indexes a daily snapshot of trending open models, not all of Hugging Face. A model that is not in the index answers 404 with a sentence saying so, and the request is recorded for the next index run. Until then, huggingface.co still works for it: unset HF_ENDPOINT for that one call. See Limits and scope.