Built for the model, not around it.

Two things make a long render dependable: it survives interruption (shown below, with the hashes), and — in resident mode — it yields to urgent work without losing progress. Both rest on the same per-step checkpoint.

Each resume picked up at the next step, so all 40 steps were computed exactly once — none repeated, none skipped.

Why teams and GPU farms care

• One GPU, many uses. A single consumer card runs a long batch render and still answers an urgent request in between — no second machine, no waiting for the batch to finish.
• No compute thrown away. Preemption lands on a step boundary and checkpoints there, so slotting in urgent work never discards the long render's hours of progress.
• The model stays resident. Switching between jobs costs a checkpoint write, not a cold model load, so the urgent answer comes back fast.
• Exact on resume. The resumed render is bit-for-bit what an uninterrupted one would have been — the same guarantee this page already shows for crash-resume, now under deliberate preemption.

How it is used

Run one resident server and send renders to it. When an urgent render arrives, the render already running is the one that pauses: it writes a .vzckpt checkpoint, yields the GPU to the urgent render, and then resumes itself from that checkpoint once the urgent one finishes — from the exact step it stopped at. You never pause or restart the long render by hand; the engine does the checkpoint and the resume.

# 1. start a resident server — the model loads once and stays loaded
$ veizik serve wan22_t2v_a14b --port 8000
# 2. submit a LONG render — it starts running on the GPU
$ curl -s localhost:8000/v1/videos/generations \
$ -d '{"prompt":"a 5-second cinematic drone shot over a coastline at dawn","seed":42}'
# 3. while render #2 is still going, send an URGENT one (priority: urgent):
#    render #2 auto-pauses to a .vzckpt, the urgent render runs, then render #2
#    auto-resumes from its checkpoint. you never stop or restart #2 by hand.
$ curl -s localhost:8000/v1/videos/generations \
$ -d '{"prompt":"a 2-second product spin on white","seed":7,"priority":"urgent"}'

Available now. Resident serving with priority preempt and .vzckpt resume ships in the current NVIDIA release — the Wan engine and the veizik command together, on Releases. Install it, start a resident server, and send an urgent render to see the pause-and-resume above. Crash-resume ships in the same build.

• 170.1 s per step, mean of all 40 steps on one NVIDIA GeForce RTX 3090 at 420 W — the fair figure for a render that ran in pieces.
• The pieces add up to 113 min of stepping. That is a sum, not a run: the wall clock between first and last attempt includes the crashes, so it is not published as a render time.
• For clean run times, see the modes rendered in one go above.
• Our own figures only. Another job was using the machine's other GPU during attempts 3–5.

Prompt sha256 d224fa1df3bf…

Cinematic photoreal close-up of a single capybara relaxing peacefully in a steaming natural hot spring at dusk, only its lower body in the water and its head held calmly well above the surface, dry soft fluffy fur on the head and face with no dripping, eyes gently half closed in bliss and slowly blinking, soft warm steam drifting upward into the air above the pool, gentle ripples spreading slowly across the calm glassy water around its body, a yuzu citrus fruit floating softly beside it, warm amber and soft rose light over a dark forest background, slow smooth cinematic push-in, shallow depth of field, detailed dry fluffy fur, calm still water surface, serene cozy and tranquil, 35mm film look, no text, no people.

Negative prompt sha256 dd682ecfdfa1…

text, watermark, logo, people, human, hands, fingers, extra capybara, morphing, warping, flicker, deformed, extra limbs, low detail, blur, cartoon, cyan, blue, purple, neon, oversaturated, water dripping on head, droplets on head, water on head, water falling on head, dripping from chin, dripping from jaw, dripping from face, soaked fur, spiky wet matted fur, drenched, shaking off water, splashing, water splashes, spray, wet face, rain, raindrops, waterfall, mist on the face.

Command

# produced with a development build; the NVIDIA build is on /releases
$ veizik run wan22_t2v_a14b --prompt-file prompt.txt --neg-file neg.txt --width 1280 --height 720 --frames 81 --steps 40 --seed 42 --mode quality --checkpoint-every 1
# after any interruption: the same command, plus
$ veizik run wan22_t2v_a14b --prompt-file prompt.txt --neg-file neg.txt --width 1280 --height 720 --frames 81 --steps 40 --seed 42 --mode quality --checkpoint-every 1 --resume

• The same answer on every machine. Greedy decoding is the default and is deterministic — the same prompt gives the same tokens. For the committed benchmark prompt the output hash was identical on the RTX 3090, RTX 4090 and RTX 5090.
• A local OpenAI-compatible server. veizik serve answers POST /v1/chat/completions, /v1/completions and /v1/models on 127.0.0.1 — point your existing OpenAI client at it and nothing leaves the machine.
• No Python, no framework. One native engine reads the published Hugging Face checkpoint and runs on the NVIDIA driver alone: no CUDA toolkit, no framework stack, no cloud at inference time.