masahi opened a new issue #6594: URL: https://github.com/apache/incubator-tvm/issues/6594
Let's summarize steps needed to update our PyTorch support for the latest version, 1.6 (the CI is at 1.4). - [ ] Quantization support: It is completely broken due to a bug introduced in 1.6 (see https://github.com/pytorch/pytorch/issues/42497). I found a simple workaround for that problem (finally!) and the fix is WIP. The representation of quantized weights and APIs of quantized ops have changed. To support both 1.6 and older versions, we need to handle both cases (with version check) for now. - [ ] torchvision detection model support: Need to add converters for `aten::tensor`, `aten::empty`, `aten::numel` and `aten::dim`. I found that other missing ops I mentioned in https://github.com/apache/incubator-tvm/pull/6449#issuecomment-693023530 can be removed using `torch._C._jit_pass_onnx_function_substitution(...)` pass. - [ ] Upgrade CI docker image cc @t-vi @siju-samuel @kevinthesun @yongwww @tqchen ---------------------------------------------------------------- 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]
