Serving transcription models on Jobs
Part of uv-scripts/transcription.
Two models can run as a server in a Job with an exposed port, so you pay the setup once and send requests from anywhere with an HF token.
Nemotron: serve, explore, scale
Run the server on its own to pay the ~8 min setup once, then send requests with an HF token as Bearer auth:
# Serve: the job logs https://<job_id>--8000.hf.jobs once the model is up
hf jobs uv run --detach --expose 8000 --timeout 2h \
https://huggingface.co/datasets/uv-scripts/transcription/raw/main/nemotron-transcribe.py \
--serve
# Scale: the same script as a driver, from a laptop
uv run https://huggingface.co/datasets/uv-scripts/transcription/raw/main/nemotron-transcribe.py \
./audio ./transcripts --server https://<job_id>--8000.hf.jobs
For exploring, any OpenAI-compatible client works: base_url="https://<job_id>--8000.hf.jobs/v1", api_key=<HF token>, with the audio as a data:audio/flac;base64,... audio_url part (see the model card). Measured from a laptop: a warm 2-minute clip came back in 2.4 s, and the driver did two half-hour episodes with --timestamps in 54 s. The first requests after start are slow (~1 min) while vLLM compiles two Triton kernels that its warmup misses. Cancel the job when done (hf jobs cancel <job_id>); exposed ports are billed on top of the hardware price, and each half-hour file is uploaded as ~30 MB, so for bulk work the in-job driver is still cheaper.
MOSS-Transcribe-Diarize as an OpenAI-compatible API
MOSS-Transcribe-Diarize can be served as an OpenAI-compatible transcription API on Jobs (see Serving on Jobs). Two verified paths; both expose /v1/audio/transcriptions at https://<job-id>--8000.hf.jobs (requests need an HF token with read access to your namespace).
vLLM (simplest; returns the raw [start][Sxx]text[end] transcript). Model support was merged into vLLM on 2026-07-08 — after the v0.24.0 release — so use a commit-pinned nightly image until the next release (then plain vllm/vllm-openai:latest works). The image needs the audio extras installed first:
hf jobs run --detach --expose 8000 --flavor l4x1 -s HF_TOKEN --timeout 2h \
vllm/vllm-openai:nightly-2c17d33f4291a55b447317640c81eb61077b1b00 -- \
bash -c "pip install librosa soundfile && vllm serve OpenMOSS-Team/MOSS-Transcribe-Diarize --trust-remote-code"
Parse the response text into segments with parse_transcript from the model's GitHub package.
sglang-omni (upstream's recommended production server; adds response_format=verbose_json with parsed speaker segments). Its prebuilt image predates this model, so install it at job start over a current sglang nightly image — the tag below is the tested pin; newer nightlies should also work. The clone + fresh-venv sequence is upstream's documented install and adds under a minute:
hf jobs run --detach --expose 8000 --flavor l4x1 -s HF_TOKEN --timeout 2h \
lmsysorg/sglang:nightly-dev-cu13-20260709-074bb928 -- \
bash -c "pip install -q uv 2>/dev/null; git clone --depth 1 https://github.com/sgl-project/sglang-omni.git && cd sglang-omni && uv venv .venv -p 3.12 && . .venv/bin/activate && uv pip install . && sgl-omni serve --model-path OpenMOSS-Team/MOSS-Transcribe-Diarize --host 0.0.0.0 --port 8000 --max-running-requests 16 --mem-fraction-static 0.80"
Query either server the same way (response_format=verbose_json on sglang-omni only; raise max_new_tokens, e.g. 65536, for long recordings). Caution — long audio on the serve path: sglang-omni silently truncates input past ~62 minutes (observed deterministically at ~62.5 min on the same tape that transcribes fully via the transformers recipe, regardless of max_new_tokens). Split longer recordings before posting — moss-transcribe-diarize-server.py does this automatically:
curl -X POST https://<job-id>--8000.hf.jobs/v1/audio/transcriptions \
-H "Authorization: Bearer $HF_TOKEN" \
-F model=OpenMOSS-Team/MOSS-Transcribe-Diarize \
-F file=@meeting.wav \
-F response_format=verbose_json \
-F max_new_tokens=65536
The job — and its billing — stops at --timeout or hf jobs cancel <job_id>.
sglang-omni install
the lmsysorg/sglang-omni:dev image predates this model — don't use it. Install from git over a current lmsysorg/sglang nightly image as shown above, and follow the clone + fresh-venv sequence exactly: installing outside the repo skips its [tool.uv] dependency overrides and fails on a protobuf conflict. Once they publish a fresh image, sgl-omni serve alone will do.