Pei Zhang

- In CPP I run the following test program: ``` torch::Tensor input = torch::rand({5, 18, 4, 4}, torch::TensorOptions(torch::kFloat)); torch::Tensor xla_input = CopyToDevice(input, device); torch::Tensor xla_output = torch::pixel_shuffle(xla_input, upscale_factor); std::cout (f32[5,2,12,12])...

This is something to do with `at::functionalization::functionalize_aten_op::call(self, upscale_factor)` called [here](https://github.com/pytorch/xla/blob/5a113aff98ce42420891c724843ccb30691dc24a/torch_xla/csrc/aten_xla_type.cpp#L3639-L3645). `pixel_shuffle` has it's own decomposition declared in torch [here](https://github.com/pytorch/pytorch/blob/58047205ed098c04ec045e66fc39dcc70b60600b/torch/_refs/nn/functional/__init__.py#L1169). If we remove the `pixel_shuffle` op from `xla_native_functions.yaml` and let torch...

> any updates on this issue? No, let me follow up with @bdhirsh .

Hi @bdhirsh , @wonjoolee95 , thanks for the following up. I checked the decomposition trace, it turns out that if we move the tensor to the XLA device, it will...

The `v2.3.0-rc` is not out yet for pytorch, does this cause any concern if we make the change ahead of the tag creation?

> hmm it is in https://github.com/pytorch/xla/blob/master/torch_xla/csrc/xla_graph_executor.cpp#L621-L624 > > One possible reason is that this model compiles way too many times so LRU cache kick out one of the graphs. You...

`opacus_cifar10` shows the same issue in the latest run.

> * [hf_T5_generate](https://github.com/pytorch/pytorch/blob/main/benchmarks/dynamo/torchbench.py#L101) Those two got skipped because `No install.py is found` for those two models and we didn't install the model.

Sorry, I mean when I try `python install.py --continue_on_fail` in https://github.com/pytorch/benchmark, it shows model `cm3leon_generate` and `hf_T5_generate` are skipped due to no install.py under those two models.