[GitHub] [tvm] masahi commented on pull request #7231: Adding aten::unsqueeze_ and aten::copy_ ops to PT Frontend

[GitBox <gi...@apache.org>]

masahi commented on pull request #7231:
URL: https://github.com/apache/tvm/pull/7231#issuecomment-756693366


   cc @t-vi @yongwww 
   
   This is probably not how we should support `copy_`. We discussed this issue before.
   
   Later I found a promising way to support inplace copy: We first convert `aten::copy_` to `aten::index_put` (see https://pytorch.org/docs/stable/tensors.html#torch.Tensor.index_put_) using PyTorch jit passes, and then convert `aten::index_put` to relay `scatter_nd`. This is also how PyTorch ONNX export supports inplace assignment, see https://github.com/pytorch/pytorch/pull/26941
   
   Relay `scatter_nd` op was added in https://github.com/apache/tvm/pull/6854. You can convert `aten::copy_` to `aten::inplace_put` by adding the following after https://github.com/apache/tvm/blob/d1399f378e37e9e4d9bfadc5cdae57fdc5bcaf7f/python/tvm/relay/frontend/pytorch.py#L2643
   
   ```
   torch._C._jit_pass_onnx_prepare_inplace_ops_for_onnx(graph)
   torch._C._jit_pass_dce_allow_deleting_nodes_with_side_effects(graph)
   ```
   
   @t-vi Does this sound reasonable a way to support inplace assignment?


----------------------------------------------------------------
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:
users@infra.apache.org