Torch2trt Versions Save

• Added converter for torch.nn.functional.group_norm using native TensorRT layers
• Added converter for torch.nn.ReflectionPad2d using plugin layer
• Added torch2trt_plugins library
• Added support for Deep Learning Accelerator (DLA)
• Added support for explicit batch
• Added support for TensorRT 8

This version includes the introduction of the Quantization Aware Training workflow in torch2trt.contrib (thanks to @SrivastavaKshitij). It also contains various converters added since the previous release. Please see the notes below.

Added

• Added converter for torch.nn.functional.adaptive_avg_pool3d
• Added converter for torch.nn.functional.adaptive_max_pool3d
• Added converter for torch.maxpool3d and torch.nn.functional.max_pool3d
• Added Quantization Aware Training (QAT) workflow to contrib
• Added converter for torch.roll
• Added converter for torch.nn.functional.layer_norm
• Added converter for torch.nn.functional.gelu
• Added converter for torch.nn.functional.linear
• Added converter for torch.nn.functional.silu

Added

• Added converter for torch.Tensor.expand
• Added support for custom converters for methods defined outside of torch module
• Added names for TensorRT layers
• Added GroupNorm plugin which internally uses PyTorch aten::group_norm
• Replaced Tensor.ndim references with len(tensor.shape) to support older pytorch versions
• Added reduced precision documentation page
• Added converters for floordiv, mod, ne, and torch.tensor operations
• Extended relu converter to support Tensor.relu operation
• Extended sigmoid converter to support Tensor.sigmoid operation

• torch2trt method
  • conversion hooks
  • conversion context
• TRTModule class
• converters to support most torchvision image classification models
• image classification example notebook