Add Serve Models guide for Jobs (vLLM + llama.cpp via exposed ports)

[davanstrien]

Important

Depends on #2551 — this guide builds on the Expose Ports documentation added there (it links to jobs-configuration#expose-ports). Best merged after it.

Adds a new Serve Models guide page (docs/hub/jobs-serving.md + toctree entry) showing how to use the new --expose feature (huggingface_hub 1.19.0) to run a Job as a temporary inference server.

What's covered

• One-command vLLM server — vllm serve LiquidAI/LFM2.5-8B-A1B on a GPU flavor with --expose 8000, following startup with hf jobs logs -f
• Connecting clients — OpenAI Python client and curl, with the HF token as the API key and the /v1 URL pattern; notes the token travels in the Authorization header (works from scripts/notebooks/agents, not directly in a browser)
• llama.cpp — llama-server pulling the Gemma 4 12B GGUF straight from the Hub with -hf, using Gemma's recommended sampling settings
• Positioning — when to use this vs Inference Endpoints (ephemeral vs managed)
• Gotchas found while testing: Jobs runs the command directly (no docker run-style entrypoint args), servers must bind 0.0.0.0, -s HF_TOKEN for authenticated model downloads, -- separator when the server's flags would collide with the CLI's

Testing

Every command and client snippet was run verbatim against live Jobs — servers reached ready and returned completions through the exposed-port proxy.

A task-titled guide page ("Serve Models") rather than a section under Examples, so people — and agents — searching for "serve a model on Jobs" land directly on it. Happy to fold it elsewhere if you'd prefer.

🤖 Generated with Claude Code

---

Note

Low Risk
Documentation-only changes with no runtime or API impact.

Overview
Adds a dedicated Hub docs page Serve Models under Jobs, wired into _toctree.yml between Popular Images and Examples.

The guide explains using --expose so a Job runs as a short-lived OpenAI-compatible inference endpoint: vLLM on a GPU flavor with -s HF_TOKEN, client setup (Python OpenAI + curl with Bearer token and /v1 base URL), and llama.cpp llama-server for Hub GGUF models (including -- for CLI flag separation, optional hf:// volume mount, and 0.0.0.0 binding). It contrasts ephemeral Jobs serving with Inference Endpoints, covers readiness via logs, cancellation/timeouts, and notes huggingface_hub ≥ 1.19.0 plus exposed-port billing.

^{Reviewed by Cursor Bugbot for commit 54d8156. Bugbot is set up for automated code reviews on this repo. Configure here.}

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[Wauplin]

Looks good, considering the Jobs commands (vllm and llama server) have been tested

[julien-c]

Suggested change

	... /app/llama-server -hf ggml-org/gemma-4-E4B-it-GGUF \
	... /app/llama serve -hf ggml-org/gemma-4-E4B-it-GGUF \

[julien-c]

Suggested change

	> Your server must listen on `0.0.0.0`. `llama-server` binds to `127.0.0.1` by default, which the jobs proxy can't reach — pass `--host 0.0.0.0` explicitly.
	> Your server must listen on `0.0.0.0`. `llama serve` binds to `127.0.0.1` by default, which the jobs proxy can't reach — pass `--host 0.0.0.0` explicitly.

let's try to push llama serve now that it's been out for a few weeks

[julien-c]

(nit)

[davanstrien]

#2593 fixes :)