Changelog

v0.3.3 Beta 1 (In progress)

Add support for tensorrt 9.1
Remove incorrect shape checking when the model has multiple inputs with different shapes.
TensorrtTensor does not support non-float input types; this version fixes this issue.

Fix the bug that compilation fails due to missing setMaxThreads when the TensorRT version is too low；Compiled successfully under the image nvcr.io/nvidia/pytorch:21.07-py3
Add ppl.cv compilation option (BUILD_PPLCV).

torchpipe.utils.test.throughput: API updated. Dependency on onnx and tensorrt python libraries removed.
timm: python -m torchpipe.utils.test.throughput --model=resnet18 --config instance_num:2 and test_throughput_from_timm

Unit tests in test_engine.py may fail due to the use of random weights when creating resnet18. This issue has been fixed
The dependency on the ONNX Python library introduced in previous version has been removed

filter: rename Continue to Run. Deprecate Continue.
Torch: Supports cross-GPU device data transfer
Added some helper functions:
- torchpipe.utils.models.onnx_export
- torchpipe.utils.test.throughput
- torchpipe.utils.models.register_model
- torchpipe.utils.models.list_models
- torchpipe.utils.models.create_model
Fixed a bug that caused infinite looping during initialization and crashed the program when using SyncTensor[Sequential[…,SyncTensor]]
fix the issue of check_dynamic_batchsize failing for some Unsqueeze layers during initialization(0.3.0b4-0.3.2b1).