Efficient Workflow Serving for Diffusion Models with Many Adapters
summary
System decouples heavy ControlNets from the base diffusion model by deploying them on dedicated GPUs as ControlNet-as-a-Service, enabling caching, parallel execution, and multiplexed sharing; lightweight LoRAs are loaded with bounded asynchronous loading that overlaps up to K denoising steps to hide I/O without degrading image quality; base model steps run across GPUs using latent parallelism and kernel optimizations. Evaluation on SDXL with H800s reports up to 7.8× latency reduction and 1.7× throughput improvement while preserving image quality.
AI-written summary · about
chapters (5)
- 0:00CAS and diffusion models intro
- 2:47Serving adapters in diffusion workflows
- 5:15CAS key designs for adapters
- 9:56Performance evaluation of CAS
- 11:19Summary
related talks
19mYou Might Not Need 50 Diffusion Steps
29mIntegrating Stable Diffusion Generative AI in Artist Workflows
32mFinetuning Stable Diffusion with low-rank adaptation techniques
28mComputing Large Foundational Models
26mEfficiently serving LLMs at scale.
34mHow To Evolve Your LLM Self-Hosting Platform: A Practical Guide To Adopting…
01h 16mOptimizing workloads with Transformers and Diffusers
13mGNNAdvisor: An Adaptive and Efficient Runtime System for GNN Acceleration on GPUs