What problem does it solve? Evaluating Stable Diffusion text-to-image inference performance on NVIDIA GPUs requires coordinating Docker containers, TensorRT engine builds, model weights, and log parsing, which is error-prone when done manually. This Skill guides an executor through the full evaluation workflow so throughput and latency metrics are collected consistently. ## Core Features & Use Cases - Containerized Evaluation Workflow: Launches a prebuilt Docker image with TensorRT, diffusers, and transformers, mounting model, code, log, and temp directories. - Multi-Model Support: Benchmarks stable-diffusion-v1-5 and stable-diffusion-2-1 with configurable batch size (1-2) and resolutions (512/768/960) in FP16. - Metric Collection: Extracts Throughput, Pipeline Latency, per-module latency (CLIP, UNet, VAE-Dec), GPU memory usage, and model loading time from inference logs. - Use Case: A performance engineer needs to compare Stable Diffusion 1.5 throughput at 512x512 versus 768x768 on an NVIDIA GPU; the Skill drives container startup, runs demo_txt2img.py, and parses the logs into comparable metrics. ## Quick Start Ask the agent to run a Stable Diffusion text-to-image TensorRT inference benchmark on NVIDIA GPU and collect throughput and latency metrics.
