hao-ai-lab-FastVideo/consumer-guide

FastVideo Config Generator

A standalone Next.js app that helps users generate FastVideo inference commands. Deployed under /config-generator on the FastVideo docs site.

Two pages:

• Quick Start (/config-generator/) — pick a task + hardware + output mode, get a ready-to-run command.
• Advanced Tuning (/config-generator/tuning/) — tune generation parameters and see the real benchmarked time.

Run locally

pnpm install
pnpm dev

Open http://localhost:3003/config-generator/ (the app is served under the /config-generator base path).

Build

pnpm build   # static export to ./out

Data files

All data lives in data/. The .tsx pages contain no hardcoded model or benchmark data — they read these JSON files:

• data/quickstart.json — drives the Quick Start page: curated product profiles, plus task / tier / hardware option lists and defaults.
• data/tuning.json — drives the Advanced Tuning page: every benchmark run is one real measured configuration, plus model / workload / gpu / preset option lists and defaults.

Both are the machine-readable record of the RTX 5090 / 4090 inference benchmarks (5090_inference.xlsx, 4090_inference.xlsx). A configuration is included only when the benchmark actually recorded a generation time — nothing is estimated.

Adding a configuration

The pages are data-driven — to add or change a configuration, edit the JSON, not the .tsx:

• Quick Start — add an entry to profiles[] in data/quickstart.json. If it introduces a new task / hardware / tier, also add it to the matching option list.
• Advanced Tuning — add an entry to runs[] in data/tuning.json. If it introduces a new model / workload / gpu / preset, also add it to the matching list.

The only thing not in JSON: a task's icon must be one of the lucide-react icons imported in basic_page.tsx (Film, Image, Sparkles, Gamepad2). A new icon needs a small .tsx edit.

Quick Start page logic

• Matches the user's selection (task + tier + gpu + vram) against quickstart.json profiles[]. One profile per combination.
• Tier rule: Quick Generate = fastest small/distilled model; High Quality = largest model. Prefer FastVideo-owned models; fall back to other models. Both must have a passing benchmark.
• If no profile matches, the combination shows "Not suitable" — no command is generated.
• All displayed numbers (time, VRAM) come straight from the benchmark.

Advanced Tuning page logic

• The Model dropdown is filtered by Workload Type + GPU: it only lists models with a passing benchmark for that combination. Models that ran out of memory on a GPU do not appear for it.
• Est. Time is a real measured generation time, looked up by model + GPU + attention backend — there is no formula. Selecting a model snaps resolution / frames / steps / attention to its benchmarked values; tuning them away flags the time with ⚠.
• Top metrics — Video Length (frames/fps), Total Pixels (h*w*frames), Latent Tokens (((frames-1)/4+1)*(h/8)*(w/8)) — are deterministic arithmetic on the chosen config.
• The generated Python command maps each control to a real FastVideo API field (VideoGenerator.from_pretrained kwargs, SamplingParam attributes).

Notes

This app is independent from the main FastVideo Python package. Run all commands from this consumer-guide/ directory.