v4.3.0

• Rename num_batches parameter to batch_size for idiom.
• Redirect stdout to stderr so as to avoid corrupted pipe due to warning messages from TensorRT.
• Add auto_download parameter to allow only download the specified model at first run rather than forcibly download all models at once.
• Bump PyTorch and Torch-TensorRT to 2.6.0.

v4.2.0

• Add num_batches and trt_static_shape parameters.
• Remove trt_int8, trt_int8_sample_step and trt_int8_batch_size parameters.
• Improve performance by using separate streams for transfering tensors between CPU and GPU.

v4.1.0

• Lower default trt_min_shape.
• Mildly decrease TRT engine building time.
• Bump PyTorch to 2.4.0.dev.
• Remove vstools dependency.

v4.0.0

• Add support for TensorRT dynamic shapes.
• Add support for TensorRT INT8 mode using Post Training Quantization (PTQ), giving 2x performance increase over FP16 mode.
• Bump PyTorch to 2.3.
• Bump VapourSynth to R66.
• Bump TensorRT to 10.0.1.

v3.1.1

• Remove nvfuser and cuda_graphs parameters.
• Bump PyTorch to 2.0.1.
• Bump TensorRT to 8.6.1.
• Bump VapourSynth to R60.

v3.0.1

• Allow strength clip to be any GRAY format.
• Don't globally set default floating point tensor type when the input is of RGBH format.

v3.0.0

• Switch to PyTorch again for inference.
• Change function name to lowercase.

v2.3.0

• Fix strength clip not properly normalized.
• Allow GRAY8 format for strength clip.
• Add dual parameter to perform inference in two threads for better performance. Mostly useful for TensorRT, not so useful for CUDA, and not supported for DirectML.

v2.2.0

• Add trt_max_workspace_size parameter.
• Allow specifying GRAYS clip for strength parameter.

v2.1.0

• Add AMD MIGraphX provider.