2024-07-10 –, Else (1.3)
Deep Learning models are often composed of a few building blocks operating on tensors with relatively little control flow compared to more traditional code. Coil proposes a mechanism to extract the tensor subprogram from the model call graph and lift its operations to the MLIR representation. By leveraging the IREE compiler stack, Coil is able to fuse and optimize operations across Julia function calls since the whole model can be observed.
Deep Learning models are often composed of a few building blocks operating on tensors with relatively little control flow compared to more traditional code. Coil proposes a mechanism to extract the tensor subprogram from the model call graph and lift its operations to the MLIR representation. By leveraging the IREE compiler stack, Coil is able to fuse and optimize operations across Julia function calls since the whole model can be observed.
PhD student, Julia enthusiast.