spark-control

Configuration needed. Run the Configure Sparks action in StartOS to set hostnames, then run Test Connection.

Spark hardware

Spark connectivity history

Most recent up/down transitions per Spark. Tracked since this dashboard was installed.

OpenAI-compatible endpoint

Base URL —

Model ID —

curl example

Swap in progress 0:00

Starting…

Show technical logs

Add WhisperX recommended

WhisperX is a single-container speech pipeline (faster-whisper for transcription + pyannote 3.1 for diarization) designed to handle long audio cleanly. Replaces the Parakeet + Sortformer combo we patched together, which crashed on a 90-min meeting. Pulled and built directly on Spark 2 (~10–15 min first time; you only do this once).

Requires a Hugging Face token at ~/.cache/huggingface/token on Spark 2 (already set up).

Installing WhisperX…

Starting…

0:00

Build log

Always-on services

Install a NVIDIA NIM container

Pick a curated container below or paste any image from the NGC NIM catalog. Spark Control will docker pull and docker run it on the target Spark.

Custom image Image (nvcr.io/...) Container name Port Kind Target Spark

Installing…

Starting…

0:00

Log

Speech model patches

Spark Control adds Sortformer speaker diarization to the third-party Parakeet ASR container via two Python overlays (diarizer.py + a patched main.py). Overlays survive container restart but not a fresh redeploy — if the parakeet container is ever rebuilt, click Reapply patches below to restore them.

Reapplying speech-model patches…

Copying overlays into the parakeet container, verifying syntax, restarting, waiting for both models to load. Takes ~60–120 s.

LLM swap

Add downloaded model to catalog

It will appear as a new card you can swap to. Knob values become its default launch flags — you can tweak later via the model's "Advanced" panel.

Key (URL-safe id) Display name Repo (read-only) Size (GB) Mode Description (optional)

Default launch knobs Max context (tokens) GPU memory % 0.85 Fast safetensors loading Prefix caching FP8 KV cache

Delete model weights from disk?

This is reversible — you can re-download from the catalog at any time. The catalog entry stays intact.

Advanced settings

Custom values are stored in the package volume and survive package updates. Empty fields fall back to defaults.

Max context (tokens) GPU memory % Fast safetensors loading (faster cold start) Prefix caching (speeds up repeated prefixes) FP8 KV cache (halves context memory)

HuggingFace repo ↗

Browse vLLM-compatible models · NVFP4-quantized models (e.g. RedHatAI/...) are best for Blackwell hardware

Where Spark 1 only Spark 2 only Both Sparks (for cluster models)

For solo models, download to wherever you'll run them. For cluster models (-tp 2), both Sparks need the weights — "Both" downloads to one Spark and rsyncs to the other in parallel.

Downloading… 0:00

Connecting…

Show technical logs

Updates to eugr/spark-vllm-docker — the upstream project that orchestrates vLLM on your Sparks (launch-cluster.sh, recipes, mods). These are not firmware, OS, or model updates.

Checking for updates…

Pending commits

Explained by the loaded LLM

Applying update…

0:00

Show technical logs

vLLM Parakeet Magpie