Completing the Torch2MHLO operator conversion

Open tanyokwok opened this issue 3 years ago • 1 comments

Background: We are working on building Torch2MHLO passes based on TorchMLIR. The goal for the most recent is to switch from the old converter framework to TorchMLIR:

[ ] Run through all UTs using the new conversion #371
[ ] Run through Bert & ASR model inference
[ ] Fix dynamic shape lowering: erase shape information from tensor type.

Jun 08 '22 07:06 tanyokwok

For the task "Run through all UTs using the new conversion", there are the following unit tests:

test_disc_activation.py test_disc_batchnorm.py test_disc_binary_ops.py test_disc_black_ops.py test_disc_block_ops.py test_disc_broadcast.py test_disc_const.py test_disc_conv_ops.py @qiuxiafei test_disc_engine.py test_disc_list.py test_disc_matmul.py test_disc_multi_streams.py test_disc_nn_ops.py test_disc_reduction.py test_disc_shapes.py test_disc_tensor.py test_disc_tools.py test_disc_unary_ops.py @fortianyou Also, we will add conversion passes unit tests based on FileCheck, https://github.com/alibaba/BladeDISC/issues/366

@Yancey1989 @wyzero @qiuxiafei

Jun 08 '22 07:06 tanyokwok