Skip to main content

Profile Image Generation with AIPerf

First follow Build from Source and Build a Bundle, using an existing image-generation profile. This example assumes the Stable Diffusion v1.5 profile has already been built as stable-diffusion-v1-5.bundle and passed its family's correctness checks. Its workload is one 256×256 image, 10 steps, guidance 7.5.

Run both terminals in your prepared GPU environment (the same container when using Docker). trtmc-server must be 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

The client comes from PyPI; an AIPerf source checkout is unnecessary.

2. Start the built model (terminal 1)​

trtmc-server ./stable-diffusion-v1-5.bundle \
  --model-name stable-diffusion-v1-5 \
  --startup-timeout 1800 \
  --records artifacts/aiperf-images-server.jsonl

Large image bundles can take several minutes to load. Wait for Uvicorn running on http://127.0.0.1:8000. Add --runtime-root if the native libraries are in a separate directory.

3. Benchmark it (terminal 2)​

echo '{"text":"a photograph of an astronaut riding a horse"}' > /tmp/image-prompts.jsonl
.venv-aiperf/bin/aiperf profile \
  --url http://127.0.0.1:8000 --model stable-diffusion-v1-5 \
  --endpoint-type image_generation --tokenizer builtin \
  --input-file /tmp/image-prompts.jsonl --custom-dataset-type single_turn \
  --extra-inputs size:256x256 num_inference_steps:10 guidance_scale:7.5 cfg_scale:7.5 seed:0 \
  --concurrency 1 --warmup-request-count 1 --request-count 5 \
  --export-level raw --artifact-dir artifacts/aiperf-images

Each response contains one complete PNG in data[0].b64_json. It remains usable after the server removes its temporary files. Raw AIPerf exports retain the response; server records identify the same PNG by SHA-256 and include its actual dimensions, count, generation controls and request ID. Server records hash both positive and negative prompt text rather than storing it.

Read client request latency and throughput in profile_export_aiperf.json. With one image per request, requests/sec is images/sec. Server model_call_ms measures public_task_call_wall, including family preprocessing and generation. PNG encoding, HTTP transfer, build, loading, startup and warmup are outside that steady-state Task measurement. Keep client latency and Task time separate.

Run another built image model​

Replace the bundle, API name, prompt and explicit generation controls with those of an existing full-model profile. The bundle must declare image_generation or SDK text_to_image and produce one image. Video and image editing are separate workloads. Dimensions must match the built profile. The server accepts seed, num_inference_steps, guidance_scale, cfg_scale, negative_prompt and size; it rejects batching, streaming, URL responses and unsupported options such as quality rather than ignoring them.

For matched-noise validation, initial_latents_b64 accepts the owning family's existing raw float32 latent layout. A shared seed alone does not guarantee the same random noise between native C++ and a Python reference. Increase --max-body-bytes when replay noise needs a larger request (up to the worker's 16 MiB auxiliary-file limit). Model-specific latent creation and quality checks stay in the owning family.

For automated request/metric joins, PNG identity checks and saved images, see image-profile.