nvidia-inference-multimodal - SKILL.md Agent Skill

name: nvidia-inference-multimodal description: "Go beyond text: understand images (vision) and generate images via the NVIDIA inference hub through the in-mac router. Use when a task involves looking at an image/screenshot or producing an image — no local GPU needed (routes to hosted models)." version: "0.1.0" license: Apache-2.0 platforms: [linux, macos, windows] tools:

Read
Shell metadata: author: mac fleet (jordanh) tags: [nvidia, inference, multimodal, vision, vlm, image-generation, sdxl, flux, router] hermes: tags: [vision, image, multimodal, image-gen, nvidia, router] related_skills: [omniverse-realtime-viewer]

NVIDIA inference: vision + image generation (via the in-mac router)

The fleet's chat already routes through the in-mac router (/v1/chat/completions) to the NVIDIA inference hub, and the hub hosts many model types (chat, vision, text-to-image, embeddings, …). This is a routing capability — you do not need a local GPU or a local Stable Diffusion install; you call the router and it forwards to NVIDIA's hosted models.

Call the router at the hub URL with the mac bearer:

URL base: ${MAC_HUB_URL:-http://127.0.0.1:8789} (spokes reach the hub via this)
Auth header: Authorization: Bearer $MAC_API_TOKEN

1) Understand an image (vision / VLM) — works today

The routed chat model is multimodal, so send the image as message content. Use a data: URL (base64) for local files, or an http(s) URL the model can fetch.

B64=$(base64 -w0 path/to/image.png)
curl -s -X POST "$MAC_HUB_URL/v1/chat/completions" \
  -H "Authorization: Bearer $MAC_API_TOKEN" -H 'Content-Type: application/json' \
  -d "{\"model\":\"*\",\"max_tokens\":300,\"messages\":[{\"role\":\"user\",\"content\":[
        {\"type\":\"text\",\"text\":\"Describe this image.\"},
        {\"type\":\"image_url\",\"image_url\":{\"url\":\"data:image/png;base64,$B64\"}}]}]}"

"model":"*" lets the router pick the configured chat model (it's vision-capable). Verified: the chat path returns 200 and analyzes the image.

2) Generate an image (text-to-image) — verified working via FLUX

POST to the router's image proxy POST /v1/genai/<model>; it forwards to NVIDIA's hosted image models. Use FLUX (verified on this fleet — HTTP 200 → base64 image):

# FLUX.1-schnell — fast (steps 4). Recommended default.
curl -s -X POST "$MAC_HUB_URL/v1/genai/black-forest-labs/flux.1-schnell" \
  -H "Authorization: Bearer $MAC_API_TOKEN" -H 'Accept: application/json' -H 'Content-Type: application/json' \
  -d '{"prompt":"a red cube on a white table","steps":4,"seed":0}' -o out.json

# FLUX.1-dev — higher quality (steps MUST be >= 5, else HTTP 422)
curl -s -X POST "$MAC_HUB_URL/v1/genai/black-forest-labs/flux.1-dev" \
  -H "Authorization: Bearer $MAC_API_TOKEN" -H 'Content-Type: application/json' \
  -d '{"prompt":"a red cube on a white table","steps":8,"seed":0}' -o out.json

The response carries the image as base64 in artifacts[0].base64; decode it:

python3 -c "import json,base64,sys; d=json.load(open('out.json'));
b=(d.get('artifacts') or [{}])[0].get('base64') or d.get('image') or d.get('b64_json');
open('out.png','wb').write(base64.b64decode(b)) if b else sys.exit('no image: '+str(d)[:200])"

Notes:

Older stabilityai/sdxl-turbo / stable-diffusion-xl return 404 on this account — prefer the FLUX models above.
HTTP 401 "Authentication failed" means the hub's nvidia-image vault key isn't a build.nvidia.com image key. It is set on this fleet; if it regresses, re-escrow a build.nvidia.com nvapi-… key as the nvidia-image secret.

3) Speech (ASR/TTS) and video — when configured

The hub also proxies POST /v1/audio/{path} and POST /v1/video/{path} to configurable upstreams, wired only when the operator set them at cluster init (separate URL + key per modality). If configured:

# Speech-to-text (ASR)
curl -s -X POST "$MAC_HUB_URL/v1/audio/transcriptions" \
  -H "Authorization: Bearer $MAC_API_TOKEN" -F file=@clip.wav -F model=<asr-model>
# Text-to-speech (TTS)
curl -s -X POST "$MAC_HUB_URL/v1/audio/synthesize" \
  -H "Authorization: Bearer $MAC_API_TOKEN" -H 'Content-Type: application/json' \
  -d '{"text":"hello","voice":"<voice>"}' -o speech.wav
# Video generation
curl -s -X POST "$MAC_HUB_URL/v1/video/<org>/<model>" \
  -H "Authorization: Bearer $MAC_API_TOKEN" -H 'Content-Type: application/json' \
  -d '{"prompt":"a drone shot over a forest","seed":0}' -o out.json

A 404 on these means the modality isn't configured on this hub yet — the operator sets router.audio / router.video (URL + key) at init.

When you'd run a model locally instead

Only for models not on the hub, data that can't leave (privacy), very high throughput, or custom/fine-tuned models — then a GPU node can host a NIM / vLLM / diffusion service. For the standard set above, the router + hub cover it.