anilmartha opened a new pull request #6343: URL: https://github.com/apache/incubator-tvm/pull/6343
This PR implements the Vitis-AI codegen using the BYOC flow and enables us to offload subgraphs to FPGA DPU accelerators (cloud/edge). Below are the features added as part of this PR - Annotate the graph for the given Vitis-AI DPU (Deep Learning Processing Unit) target. - During codegen phase, convert the relay subgraph into PyXIR and save XGraph. - Vitis-AI runtime supports SaveToBinary and LoadFromBinary. We save the XGraph it in our own format and serialize it in the Module by keeping track of the path to the files. - Tests include a complete resnet18 model test partly offloaded to PyXIR for DPU acceleration. However, we don't have access to an FPGA instance in the CI docker environment and therefore the offloaded subgraph is just executed on CPU. This PR depends on following [Vitis-AI CI contribution PR](https://github.com/apache/incubator-tvm/pull/6342). The RFC for this PR can be found from [here](https://discuss.tvm.ai/t/rfc-byoc-vitis-ai-integration/7544). This work is co-authored by Jorn Tuyls <[email protected]> @jtuyls, Elliott Delaye <[email protected]> @edelaye and Sumit Nagpal <[email protected]> @sumitn-xilinx. ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected]
