✨[Feature] Some operators need to be supported in Torch-TensorRT

[ruoqianguo]

Is your feature request related to a problem? Please describe.

There are some operators need to be supported in Torch-TensorRT. It is from a Video-Swin-Transformer model.

• aten::index.Tensor(Tensor self, Tensor?[] indices) -> (Tensor)
• aten::adaptive_avg_pool3d(Tensor self, int[3] output_size) -> (Tensor)
• aten::Int.Tensor(Tensor a) -> (int)
• aten::remainder.Tensor(Tensor self, Tensor other) -> (Tensor)
• aten::div.Tensor_mode(Tensor self, Tensor other, *, str? rounding_mode) -> (Tensor)
• aten::remainder.Scalar(Tensor self, Scalar other) -> (Tensor)
• aten::roll(Tensor self, int[1] shifts, int[1] dims=[]) -> (Tensor)
• aten::to.dtype_layout(Tensor(a) self, *, int? dtype=None, int? layout=None, Device? device=None, bool? pin_memory=None, bool non_blocking=False, bool copy=False, int? memory_format=None) -> (Tensor(a))

Describe the solution you'd like

Describe alternatives you've considered

Additional context

[ncomly-nvidia]

@guoruoqian does Swin-Transformer work with fallback enabled? Can you post the log if not?

[ruoqianguo]

@guoruoqian does Swin-Transformer work with fallback enabled? Can you post the log if not?

Hi @ncomly-nvidia, Video-Swin-Transformer doesn't work with fallback enabled.
We set compile settings as the following.

compile_settings.enabled_precisions = {torch_tensorrt::DataType::kFloat};
compile_settings.require_full_compilation = false;
compile_settings.truncate_long_and_double = true;

And I add a LOG_DEBUG("Runing analysis in progress " << seg_block) in there.
The output is log_with_truncate.txt.

We set compile settings as the following.

compile_settings.enabled_precisions = {torch_tensorrt::DataType::kFloat};
compile_settings.require_full_compilation = false;
compile_settings.truncate_long_and_double = false;

The output is log_without_truncate.txt

[ncomly-nvidia]

@guoruoqian there are still outstanding ops here, right? I think we should leave this open for that + the fallback issues.

[ncomly-nvidia]

Still ops remaining