Xiafei Qiu

We may consult torch-mlir community. If compatibility is committed, our legacy converter should be maintained for quite a long time. 😢

Seems because we only handle `DynamicPadOp` through out the lowering pipeline. (for simplicity??) How about add a pass somewhere in early stage to convert `PadOp` to `DynamicPadOp`?

> Nit: Could be better to add some overflow tests This Op is just used to carry quant info and use `int32` for underlying representation, won't be overflow I think...

# To refine project structure of tensorflow_blade 1. Currently, tensorflow_blade is fully managed by Bazel, while in pytorch_blade Bazel only takes control of C++ code. After some discussion, we decide...

Just surprised that add/mul/sqrt are in the list..

One more thing: - [ ] Add TF-CPU CR tests

`dump.[date].mlir` is mhlo IR converted from input torch model, which will be fed to BladeDISC compiler. `mhlo_conpile.[date].log` is the output of BladeDISC compiler, it contains detailed output IR if every...

It seemed to fail on some check in `aten::conv2d`'s converter, roll the log up, there should some log. Otherwise, you can try setting env var `TORCH_DISC_USE_TORCH_MLIR` to `true` to use...

maybe your zk cluster is configured with acl.

It currently use `~/.zk-web-conf.clj` or `conf/zk-web-conf.clj`. What's the scenario for constom path ?