> ## Documentation Index
> Fetch the complete documentation index at: https://docs.veri.studio/llms.txt
> Use this file to discover all available pages before exploring further.

# Hugging Face (your account)

> Deploy to a dedicated HF Inference Endpoint under your own Hugging Face account — HF bills the compute, and it costs no Veri credits.

`provider="huggingface"` is different from every other provider: the deployment does not run on GPUs Veri rents. Instead, Veri creates and manages a dedicated [HF Inference Endpoint](https://huggingface.co/inference-endpoints) **under your own Hugging Face account** — your namespace, your billing, your data path.

It is currently feature-gated: a `provider="huggingface"` create returns a "coming soon" 400 until the gate lifts.

## Why use it

* **Compute on your own account.** Hugging Face bills you directly, by the minute, at their published instance rates. The deployment costs **zero Veri credits** — credits only meter training and Veri-hosted serving.
* **Burn existing spend.** HF endpoint charges land on your HF organization's bill, alongside whatever you already pay them.
* **Your data path.** Your apps can call the endpoint URL directly with your HF token; requests then never transit Veri. (The Veri chat proxy and playground still work if you want one pane of glass.)

## Prerequisites

1. **A connected Hugging Face token** with the *Manage Inference Endpoints* permission (and *Make calls to Inference Endpoints* if you'll invoke through it). Connect it in **Settings → Integrations**, or `PUT /v1/settings/integrations/huggingface`.
2. **A model that lives on the HF hub**, one of:
   * `source="huggingface"` — any repo your token can access,
   * `source="training_job"` — a job trained with `hf_push` (its checkpoint is already in your HF account),
   * `source="custom_model"` — a saved model you've pushed with `POST /v1/models/{id}/push-to-hf`.
3. **A payment method on your HF account** (HF requires one for Inference Endpoints).

## Planned serving shapes

The `gpu_type`/`gpu_count` you request maps onto HF's instance catalog:

| `gpu_type`  | GPUs per endpoint replica | HF instance                   | Status      |
| ----------- | ------------------------- | ----------------------------- | ----------- |
| `A10G-24GB` | 1, 4                      | `nvidia-a10g` (AWS us-east-1) | Coming soon |
| `L4-24GB`   | 1, 4                      | `nvidia-l4` (AWS us-east-1)   | Coming soon |
| `L40S-48GB` | 1, 4, 8                   | `nvidia-l40s` (AWS us-east-1) | Coming soon |
| `A100-80GB` | 1, 2, 4, 8                | `nvidia-a100` (AWS us-east-1) | Coming soon |
| `H100-80GB` | 1, 2, 4                   | `nvidia-h100` (GCP us-east4)  | Coming soon |

## Creating one

```python theme={null}
dep = client.deployments.create(
    model="your-namespace/your-model",
    source="huggingface",
    name="my-byoc-endpoint",
    gpu={"gpu_type": "A10G-24GB", "gpu_count": 1},
    provider="huggingface",
)
```

Or from the CLI:

```bash theme={null}
veri deploy your-namespace/your-model --provider huggingface --gpu-type A10G-24GB
```

## How it behaves

* **One endpoint, HF-native scaling.** `num_replicas` / `min_replicas` / `max_replicas` map onto the endpoint's own `minReplica`/`maxReplica`. `min_replicas=0` uses HF's native scale-to-zero (the endpoint wakes itself on the next request; the deployment stays `serving` in Veri).
* **Status is polled.** Veri polls the HF API and mirrors the endpoint's state onto the deployment (`provisioning` → `serving`, failures surface HF's error message). The endpoint also appears in your HF console, named `veri-<id>`.
* **Deleting is symmetric.** Stopping or deleting the deployment in Veri deletes the HF endpoint. Pausing or deleting it in the HF console settles the Veri deployment to `stopped`.
* **Calling it.** `POST /v1/deployments/{id}/chat/completions` proxies to the endpoint, or call the endpoint URL directly: it is an `authenticated`-type endpoint, so pass your own HF token as the bearer.

```bash theme={null}
curl https://<endpoint-url>/v1/chat/completions \
  -H "Authorization: Bearer $HF_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"model": "your-namespace/your-model", "messages": [{"role": "user", "content": "hi"}]}'
```

<Note>
  Not supported on BYOC endpoints: the `max` engine, `vllm_extra_args`, model caching, and multi-adapter serving. Requests that include them return a 400 at create.
</Note>

## Where to go next

<CardGroup cols={2}>
  <Card title="All serving providers" icon="server" href="/deployments/providers">
    Compare shapes, status, and the live CLI catalog.
  </Card>

  <Card title="Push a checkpoint to HF" icon="upload" href="/training">
    Train with hf\_push so your checkpoint is deployable on your account.
  </Card>
</CardGroup>