Hi, recently, my colleagues and I have use torch-mlir as front-end tool to convert Huggingface’s Bert model into TensorRT model.
We added some graph-opt passes(base on Torch Dialect) by using graph optimization tool of MLIR, and then used them together with TensorRT api and some customized TensorRT plugins, which achieved better performance than “torch-mlir + iree” and ONNX-TensorRT.
Here, I have two questions:
- Will the Mlir community have any plans to support TensorRT Dialect in the future?
- If we design a TensorRT Dialect, Is it necessary? Or will it be valuable? Are there any requirement if we want to contribute it to torch-mlir in the future?