Profile Video Generation with AIPerf

First follow Build from Source and Build a Bundle. This example assumes the existing Wan2.1 full profile has been built as wan21-t2v-1.3b.bundle and passed its family's correctness checks: 832×480, 17 frames, FP16 with the profile's existing FP32 layer selection. Run from the repository root in your prepared GPU environment. Install FFmpeg with libx264rgb and put ffmpeg, ffprobe, and trtmc-server on PATH.

1. Install AIPerf

python -m pip install -e '.[serve]' -C py-only=true
python -m venv .venv-aiperf
.venv-aiperf/bin/pip install aiperf==0.13.0

2. Start the built model (terminal 1)

trtmc-server ./wan21-t2v-1.3b.bundle \
  --model-name wan21-t2v-1.3b \
  --startup-timeout 1800 --request-timeout 14400 \
  --records artifacts/aiperf-video-server.jsonl

Wait for Uvicorn running on http://127.0.0.1:8000. Add --runtime-root when native libraries are in a separate directory.

3. Benchmark it (terminal 2)

printf '%s\n' '{"text":"A cat sitting on a beach watching the sunset"}' > /tmp/video-prompts.jsonl
.venv-aiperf/bin/aiperf profile \
  --url http://127.0.0.1:8000 --model wan21-t2v-1.3b \
  --endpoint-type video_generation --download-video-content \
  --tokenizer builtin --input-file /tmp/video-prompts.jsonl \
  --custom-dataset-type single_turn \
  --extra-inputs size:832x480 num_frames:17 num_inference_steps:30 seed:42 fps:24 \
  --request-timeout-seconds 14400 \
  --concurrency 1 --warmup-request-count 1 --request-count 2 \
  --export-level raw --artifact-dir artifacts/aiperf-videos

AIPerf submits multipart POST /v1/videos, polls GET /v1/videos/{id}, then downloads GET /v1/videos/{id}/content. Keep --download-video-content to include the complete MP4 transfer. The server encodes every native RGB frame losslessly; fps controls playback. Frame count is fixed when the bundle is built. Video serving requires explicit size and frame count in the bundle's core build provenance; rebuild older video bundles that lack this metadata. Existing still-image bundles retain their image endpoint, including bundles with missing provenance or family-default dimensions. A mismatched size/frame count, seconds, image conditioning, streaming, or batching is rejected.

Read client latency in profile_export_aiperf.json. It includes submission, polling, encoding, and download. The separate server video_completion record reports model_call_ms (public_task_call_wall), the original request ID, job ID, complete file hash and ordered decoded-frame hash. Task timing excludes build, loading, warmup and external media encoding. Two requests demonstrate the workflow; they do not establish serving capacity.

AIPerf 0.13 exports job JSON and intentionally omits downloaded video bytes. To save a clip, take its job ID from the raw export and run:

curl --fail http://127.0.0.1:8000/v1/videos/VIDEO_ID/content -o clip.mp4

Completed clips expire after ten minutes; retained jobs and reserved media storage are bounded. Use --max-video-jobs, --max-video-bytes, and --video-ttl-seconds to set those limits. Shutdown waits for active native calls and encoding before removing temporary files.

Use another built video model

Replace the bundle, API name, prompt and settings with those of its existing full text-to-video profile. The current full profiles are:

Profile	Size	Frames	Steps	Seed	Explicit guidance
wan21-t2v-1.3b	832×480	17	30	42	Omit to retain the public Task default
wan22-ti2v-5b	1280×704	121	50	42	guidance_scale:5
minimax-h3-768p	1344×768	124	50	0	Omit to retain the public Task default

MiniMax-H3's existing prompt is the prompt string in families/minimax_h3/tests/prompts/t2va-example-1.json; pass that whole string unchanged. This endpoint returns its video frames. Audio output and image-to-video are separate workloads.

Cosmos3-Nano and LTX-Video currently have only partial/L0 profiles and are not full-model qualifications. Inventory all variants, including these gaps, with python -m trtmc_aiperf_qual endpoint-models --repo . --group videos after installing the qualification tools.