DinoML

DinoML compiles machine learning models into optimized standalone modules.

System and platform support

System	Support
Windows	✅
Linux	✅

Platform	Support
CUDA	✅
ROCm	Partial

Optimizations

• Custom kernels for standard operators
• Custom fused kernels that combine several operations into a single kernel
• Graph transformations deduplicate and reduce the number of needed operations
• Graph transformations fuse common patterns of operations
• Profiling to find the best kernel for a problem size
• Intermediate workspace is reused to reduce memory usage

and more!

Comparison

	Dynamic shape	Portability	Dynamic memory usage
DinoML	✅	✅	✅
torch.compile	Limited	Limited	❌
TensorRT	Limited	Limited	❌

Documentation

• Getting started
• Modules
• Modeling