Use llama serve in Jobs serving guide (llama.cpp unified CLI)

[davanstrien]

Follow-up to #2554 addressing @julien-c's review feedback: push the new unified llama serve command instead of the standalone llama-server.

What changed

The llama.cpp section now uses llama serve on the full-cuda image (both the -hf example and the volume-mount tip), with the surrounding prose and warning updated to match.

Why the image also changed

llama serve and llama-server are the same server — the unified llama CLI (ggml-org/llama.cpp#23875, "git commit vs git-commit"). But the unified llama binary only ships in the full-cuda image; server-cuda contains only /app/llama-server. So using llama serve requires full-cuda — a ~1 GB larger pull (3.6 GB vs 2.6 GB compressed). Startup is dominated by the model download/load, which is unchanged, so the "a few minutes" framing still holds.

Tested

Ran on live Jobs (full-cuda + a10g-small):

• llama serve -hf ggml-org/gemma-4-E4B-it-GGUF --host 0.0.0.0 --port 8080 -ngl 99 → server reached ready, returned a completion through the exposed-port proxy with a Bearer token.
• The volume-mounted --model /model/...gguf variant → ready in ~40s (nothing to download), completion returned.

🤖 Generated with Claude Code

---

Note

Low Risk
Documentation-only wording and example command updates with no runtime or security impact.

Overview
Updates the Serve GGUF models with llama.cpp section in jobs-serving.md to document llama.cpp’s unified llama serve entry point instead of the standalone llama-server.

Prose, the main -hf job example, the volume-mount tip, and the 0.0.0.0 warning now refer to llama serve and use /app/llama serve in the sample commands; server flags and the ghcr.io/ggml-org/llama.cpp:server-cuda image tag in the snippets are unchanged in this diff.

^{Reviewed by Cursor Bugbot for commit 3f9c7ac. 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.

[pcuenca]

Ah, I just verified it and the llama binary is indeed included in the full-cuda image (6.58 GB in my system) but not in the server-cuda one (4.35 GB). I wonder if we should include llama in server builds as well (but not all the other binaries/utilities that come with the full package).

[julien-c]

@angt makes sense to add it to server-cuda builds?

[angt]

let's go llama everywhere by default then ?

[davanstrien]

will hold off merging this for now to switch back to the lighter image

[pcuenca]

For this particular use case (optimized server image), I'd include the same stuff we do now, but also a llama binary if that's possible. So calling llama serve will work, it's ok that llama cli doesn't, imo.

[angt]

let's go llama everywhere by default then ?

Forgot it was already enabled by default 😅

[julien-c]

can you check if it's in the latest server-cuda images?

[davanstrien]

llama is in the latest server-cuda

Verified directly by running the latest image on a Job:

$ hf jobs run --flavor cpu-basic ghcr.io/ggml-org/llama.cpp:server-cuda \
    sh -c "ls -la /app/llama && /app/llama --version"

Switched back to the server image. IMO can merge now?