trunk/65a423d37d91acd10d8628d614731fd1728b12f4: [compile] Compile time diffusion benchmarks (#175294)
New benchmarks target diffusion models with notoriously high 120-second compile times, aiming to cut them down.
Deep Dive
PyTorch developers have added a new compile-time benchmark suite to the main codebase. The tool specifically measures the compilation overhead for complex diffusion models like 'auroflow', which currently take over 108 seconds. This move formalizes internal testing to help engineers identify and optimize slow compilation steps, directly speeding up the development and iteration cycle for AI researchers building image and video generation models.
Why It Matters
Faster compilation means researchers can test new AI model architectures more quickly, accelerating the pace of innovation.