SHARK is a high performance codegen compiler and runtime built on MLIR, IREE and custom RL based tuning infrastructure. Here are some results of using SHARK for the same model across PyTorch, Onnx, TF/XLA and SHARK.
We have an Triton Inference Server integration of SHARK that runs on CPU and CUDA devices here: https://github.com/nod-ai/SHARK/tree/main/inference and we would like to upstream it as an available Triton backend that anyone can build, test and deploy.
SHARK is a high performance codegen compiler and runtime built on MLIR, IREE and custom RL based tuning infrastructure. Here are some results of using SHARK for the same model across PyTorch, Onnx, TF/XLA and SHARK.
We have an Triton Inference Server integration of SHARK that runs on CPU and CUDA devices here: https://github.com/nod-ai/SHARK/tree/main/inference and we would like to upstream it as an available Triton backend that anyone can build, test and deploy.