HolyWu

How to use realesr-animevideov3 for X1, X2, X3?

2

In https://github.com/xinntao/Real-ESRGAN/blob/master/docs/anime_video_model.md it says `This model can also be used for X1, X2, X3`. But simply changing the upscale argument of SRVGGNetCompact from 4 to 1, 2 or 3 will...

🐛 [Bug] `torch.export.load` fails to load file saved with `torch_tensorrt.save` in PyTorch 2.4.0.dev

2

## Bug Description ``` DEBUG:torch_tensorrt.dynamo._compiler:Input graph: graph(): %linear_weight : [num_users=1] = get_attr[target=linear.weight] %linear_bias : [num_users=1] = get_attr[target=linear.bias] %x : [num_users=1] = placeholder[target=x] %view : [num_users=1] = call_function[target=torch.ops.aten.view.default](args = (%x, [2,...

# Description - Add missing resize modes. - Fix `align_corners` handling. - Remove ops from `torch_enabled_decompositions`. Unfortunately the converter still won't be in use due to decomposition in PyTorch. See...

## Bug Description It works fine with static-shaped model, but fails to load with dynamic-shaped one. ``` DEBUG:torch_tensorrt.dynamo._compiler:Input graph: graph(): %arg0_1 : [num_users=1] = placeholder[target=arg0_1] %relu : [num_users=1] = call_function[target=torch.ops.aten.relu.default](args...

Remove ops from decompositions where converters exist

4

# Description The last removed `aten.var_mean` doesn't have a converter, but it's a duplicate in the list, so I remove it as well. Partially #2665 ## Type of change -...

1. The [Resize](https://docs.nvidia.com/deeplearning/tensorrt/api/python_api/infer/Graph/Layers.html#iresizelayer) layer supports `NEAREST, LINEAR, CUBIC` resize modes, but only `NEAREST` and `LINEAR` are implemented. https://github.com/pytorch/TensorRT/blob/ad74a735056667726692c49a175a790647ef889e/py/torch_tensorrt/dynamo/conversion/impl/upsample.py#L36-L47 2. The `LINEAR` resize mode supports `linear (1D), bilinear (2D), trilinear (3D)`...

❓ [Question] operator being decomposed rather than being converted when a corresponding converter exists?

2

## ❓ Question From the debug log below, it seems that the `aten.grid_sampler_2d` operator gets decomposed into several lower-level operators. But isn't there a corresponding [converter](https://github.com/pytorch/TensorRT/blob/9a100b6414bee175040bcaa275ecb71df54836e4/py/torch_tensorrt/dynamo/conversion/aten_ops_converters.py#L333-L358) which should be used?...

mv.Mask shall be updated to support 16 bits as well

6

As have been mentioned on doom9.

feat: C++ runtime on Windows

7

# Description Fix and enable C++ runtime on Windows. Fixes #2247 Fixes #2371 Fixes #2484 ## Type of change - New feature (non-breaking change which adds functionality) # Checklist: -...