BladeDISC
BladeDISC copied to clipboard
alibaba
BladeDISC is an end-to-end DynamIc Shape Compiler project for machine learning workloads.
Currently, DISC lowers all pad-like op into `mhlo::DynamicPadOp`. However, the `DPadToPad` pattern in `mhlo::DynamicPadOp`'s CanonicalizationPatterns may convert `mhlo::DynamicPadOp` into `mhlo::PadOp` since the rebase https://github.com/alibaba/BladeDISC/pull/525. It will cause lowering exception of...
It requires external only results have the same num-elements. Non external only results can have different number of elements.
The dump toolkit dumps the transformer training graph based on LTC, next we should implement following op converters: - [ ] aten::_log_softmax - [ ] aten::_log_softmax_backward_data - [ ] aten::_softmax...
To utilize BladeDISC accelerating bert training, we need to cluster more op into disc, as the experiment, we found 244 `FusionGroups` in the first iteration with LTC, to reduce the...
https://github.com/alibaba/BladeDISC/blob/336736c89b4ccc2a5da28c1c43afd85be5e72893/pytorch_blade/src/BUILD#L15-L21 By depending on the above, bazel targets may have C++ source root as include root other than bazel workspace root. As a consequence, targets my reference any header file...
This issue lists all disc op unit tests that torch-mlir converter needed - [x] test_unary_ops.py @fortianyou - [x] test_activation.py @Yancey1989 #382 - [ ] test_batchnorm.py - [x] test_binary_ops.py @fortianyou -...
## Background When putting `pytorch_blade` and `tensorflow_blade` to open source, BladeDISC's project structure gets more complex. Currently we've already had the following essential directories: - `tao`: TF bridge of BladeDISC...
TorchMlir follows the latest PyTorch version, but BladeDISC CI runs some older versions, e.g. `1.7.1`, `1.8.1`, and `1.10`, the ATen symbol may be incompatiable between the latest Pytorch and older...