|
576 | 576 | * Move some lowering passes to graph level logging ([0266f41](https://github.com/NVIDIA/TRTorch/commit/0266f41)) |
577 | 577 | * **//py:** Fix trtorch.Device alternate contructor options ([ac26841](https://github.com/NVIDIA/TRTorch/commit/ac26841)) |
578 | 578 |
|
| 579 | +# 1.0.0 (2021-11-09) |
| 580 | + |
| 581 | + |
| 582 | +### Bug Fixes |
| 583 | + |
| 584 | +* aten::gelu call was wrong in test ([40bc4e3](https://github.com/NVIDIA/TRTorch/commit/40bc4e3)) |
| 585 | +* Fix a core partitioning algo bug where non-tensor input segments are not updated correctly ([cc10876](https://github.com/NVIDIA/TRTorch/commit/cc10876)) |
| 586 | +* Fix modules_as_engines test case to use trt_mod instead of pyt_mod ([282e98a](https://github.com/NVIDIA/TRTorch/commit/282e98a)) |
| 587 | +* Fix plugin registration macro ([8afab22](https://github.com/NVIDIA/TRTorch/commit/8afab22)) |
| 588 | +* Fix python API tests for mobilenet v2 ([e5a38ff](https://github.com/NVIDIA/TRTorch/commit/e5a38ff)) |
| 589 | +* Partial compilation translation to internal settings was incorrect ([648bad3](https://github.com/NVIDIA/TRTorch/commit/648bad3)) |
| 590 | +* **//py:** Don't crash harshly on import when CUDA is not available ([07e16fd](https://github.com/NVIDIA/TRTorch/commit/07e16fd)) |
| 591 | +* Renable backtrace and make it less repetitive ([1435845](https://github.com/NVIDIA/TRTorch/commit/1435845)) |
| 592 | +* **//core/lowering:** Fixes module level fallback recursion ([f94ae8f](https://github.com/NVIDIA/TRTorch/commit/f94ae8f)) |
| 593 | +* **//core/partitioing:** Fixing support for paritally compiling ([748ecf3](https://github.com/NVIDIA/TRTorch/commit/748ecf3)) |
| 594 | +* **//docker:** Update docker container build script to use release path ([9982855](https://github.com/NVIDIA/TRTorch/commit/9982855)) |
| 595 | +* **//py:** Add new dirs to remove during clean ([d2cc1e9](https://github.com/NVIDIA/TRTorch/commit/d2cc1e9)) |
| 596 | +* **//py:** Fix some api import issues ([840ca89](https://github.com/NVIDIA/TRTorch/commit/840ca89)) |
| 597 | +* **//py:** Fix trtorch.Device alternate contructor options ([fa08311](https://github.com/NVIDIA/TRTorch/commit/fa08311)) |
| 598 | +* **//py:** Fix trtorch.Device alternate contructor options ([ac26841](https://github.com/NVIDIA/TRTorch/commit/ac26841)) |
| 599 | +* Update notebooks with new library name Torch-TensorRT ([8274fd9](https://github.com/NVIDIA/TRTorch/commit/8274fd9)) |
| 600 | +* **aten::conv1d:** Update namespace, fix typo in dest IR for conv1d ([d53f136](https://github.com/NVIDIA/TRTorch/commit/d53f136)) |
| 601 | +* **eval:** Rollback 1.11a0 change + namespace issues ([ba743f5](https://github.com/NVIDIA/TRTorch/commit/ba743f5)) |
| 602 | +* Use scripting instead of tracing for module fallback tests ([32e8b53](https://github.com/NVIDIA/TRTorch/commit/32e8b53)) |
| 603 | +* Workspace defaults for other apis and centralize cuda api use ([930321e](https://github.com/NVIDIA/TRTorch/commit/930321e)) |
| 604 | + |
| 605 | + |
| 606 | +### Features |
| 607 | + |
| 608 | +* Add functionality for tests to use precompiled libraries ([b5c324a](https://github.com/NVIDIA/TRTorch/commit/b5c324a)) |
| 609 | +* Add QAT patch which modifies scale factor dtype to INT32 ([4a10673](https://github.com/NVIDIA/TRTorch/commit/4a10673)) |
| 610 | +* Add TF32 override flag in bazelrc for CI-Testing ([7a0c9a5](https://github.com/NVIDIA/TRTorch/commit/7a0c9a5)) |
| 611 | +* Add VGG QAT sample notebook which demonstrates end-end workflow for QAT models ([8bf6dd6](https://github.com/NVIDIA/TRTorch/commit/8bf6dd6)) |
| 612 | +* Augment python package to include bin, lib, include directories ([ddc0685](https://github.com/NVIDIA/TRTorch/commit/ddc0685)) |
| 613 | +* handle scalar type of size [] in shape_analysis ([fca53ce](https://github.com/NVIDIA/TRTorch/commit/fca53ce)) |
| 614 | +* support aten::__and__.bool evaluator ([6d73e43](https://github.com/NVIDIA/TRTorch/commit/6d73e43)) |
| 615 | +* support aten::conv1d and aten::conv_transpose1d ([c8dc6e9](https://github.com/NVIDIA/TRTorch/commit/c8dc6e9)) |
| 616 | +* support aten::eq.str evaluator ([5643972](https://github.com/NVIDIA/TRTorch/commit/5643972)) |
| 617 | +* support setting input types of subgraph in fallback, handle Tensor type in evaluated_value_map branch in MarkOutputs ([4778b2b](https://github.com/NVIDIA/TRTorch/commit/4778b2b)) |
| 618 | +* support truncate_long_and_double in fallback subgraph input type ([0bc3c05](https://github.com/NVIDIA/TRTorch/commit/0bc3c05)) |
| 619 | +* Update documentation with new library name Torch-TensorRT ([e5f96d9](https://github.com/NVIDIA/TRTorch/commit/e5f96d9)) |
| 620 | +* Updating the pre_built to prebuilt ([51412c7](https://github.com/NVIDIA/TRTorch/commit/51412c7)) |
| 621 | +* **//:libtrtorch:** Ship a WORKSPACE file and BUILD file with the ([7ac6f1c](https://github.com/NVIDIA/TRTorch/commit/7ac6f1c)) |
| 622 | +* **//core/partitioning:** Improved logging and code org for the ([8927e77](https://github.com/NVIDIA/TRTorch/commit/8927e77)) |
| 623 | +* **//cpp:** Adding example tensors as a way to set input spec ([70a7bb3](https://github.com/NVIDIA/TRTorch/commit/70a7bb3)) |
| 624 | +* **//py:** Add the git revision to non release builds ([4a0a918](https://github.com/NVIDIA/TRTorch/commit/4a0a918)) |
| 625 | +* **//py:** Allow example tensors from torch to set shape ([01d525d](https://github.com/NVIDIA/TRTorch/commit/01d525d)) |
| 626 | + |
| 627 | + |
| 628 | +* feat!: Changing the default behavior for selecting the input type ([a234335](https://github.com/NVIDIA/TRTorch/commit/a234335)) |
| 629 | +* refactor!: Removing deprecated InputRange, op_precision and input_shapes ([621bc67](https://github.com/NVIDIA/TRTorch/commit/621bc67)) |
| 630 | +* feat(//py)!: Porting forward the API to use kwargs ([17e0e8a](https://github.com/NVIDIA/TRTorch/commit/17e0e8a)) |
| 631 | +* refactor(//py)!: Kwargs updates and support for shifting internal apis ([2a0d1c8](https://github.com/NVIDIA/TRTorch/commit/2a0d1c8)) |
| 632 | +* refactor!(//cpp): Inlining partial compilation settings since the ([19ecc64](https://github.com/NVIDIA/TRTorch/commit/19ecc64)) |
| 633 | +* refactor! : Update default workspace size based on platforms. ([391a4c0](https://github.com/NVIDIA/TRTorch/commit/391a4c0)) |
| 634 | +* feat!: Turning on partial compilation by default ([52e2f05](https://github.com/NVIDIA/TRTorch/commit/52e2f05)) |
| 635 | +* refactor!: API level rename ([483ef59](https://github.com/NVIDIA/TRTorch/commit/483ef59)) |
| 636 | +* refactor!: Changing the C++ api to be snake case ([f34e230](https://github.com/NVIDIA/TRTorch/commit/f34e230)) |
| 637 | +* refactor! : Update Pytorch version to 1.10 ([cc7d0b7](https://github.com/NVIDIA/TRTorch/commit/cc7d0b7)) |
| 638 | +* refactor!: Updating bazel version for py build container ([06533fe](https://github.com/NVIDIA/TRTorch/commit/06533fe)) |
| 639 | + |
| 640 | + |
| 641 | +### BREAKING CHANGES |
| 642 | + |
| 643 | +* This removes the InputRange Class and op_precision and |
| 644 | +input shape fields which were deprecated in TRTorch v0.4.0 |
| 645 | + |
| 646 | +Signed-off-by: Naren Dasan <[email protected]> |
| 647 | +Signed-off-by: Naren Dasan <[email protected]> |
| 648 | +* This change updates the bazel version |
| 649 | +to build Torch-TensorRT to 4.2.1. |
| 650 | + |
| 651 | +This was done since the only version of bazel available |
| 652 | +in our build container for python apis is 4.2.1 |
| 653 | + |
| 654 | +Signed-off-by: Naren Dasan <[email protected]> |
| 655 | +Signed-off-by: Naren Dasan <[email protected]> |
| 656 | +* This changes the API for compile settings |
| 657 | +from a dictionary of settings to a set of kwargs for the various |
| 658 | +compilation functions. This will break existing code. However |
| 659 | +there is simple guidance to port forward your code: |
| 660 | + |
| 661 | +Given a dict of valid TRTorch CompileSpec settings |
| 662 | + |
| 663 | +```py |
| 664 | +spec = { |
| 665 | + "inputs": ... |
| 666 | + ... |
| 667 | +} |
| 668 | +``` |
| 669 | + |
| 670 | +You can use this same dict with the new APIs by changing your code from: |
| 671 | + |
| 672 | +```py |
| 673 | +trtorch.compile(mod, spec) |
| 674 | +``` |
| 675 | + |
| 676 | +to: |
| 677 | + |
| 678 | +```py |
| 679 | +trtorch.compile(mod, **spec) |
| 680 | +``` |
| 681 | +which will unpack the dictionary as arguments to the function |
| 682 | + |
| 683 | +Signed-off-by: Naren Dasan <[email protected]> |
| 684 | +Signed-off-by: Naren Dasan <[email protected]> |
| 685 | +* This commit changes the APIs from a dictionary of |
| 686 | +arguements to a set of kwargs. You can port forward using |
| 687 | + |
| 688 | +```py |
| 689 | +trtorch.compile(mod, **spec) |
| 690 | +``` |
| 691 | + |
| 692 | +Also in preparation for partial compilation to be enabled by default |
| 693 | +settings related to torch fallback have been moved to the top level |
| 694 | + |
| 695 | +instead of |
| 696 | + |
| 697 | +```py |
| 698 | +"torch_fallback": { |
| 699 | + "enabled": True, |
| 700 | + "min_block_size" " 3, |
| 701 | + "forced_fallback_ops" : ["aten::add"], |
| 702 | + "forced_fallback_mods" : ["MySubModule"] |
| 703 | +} |
| 704 | +``` |
| 705 | + |
| 706 | +now there are new settings |
| 707 | + |
| 708 | +```py |
| 709 | +require_full_compilation=False, |
| 710 | +min_block_size=3, |
| 711 | +torch_executed_ops=["aten::add"], |
| 712 | +torch_executed_modules=["MySubModule"] |
| 713 | +``` |
| 714 | + |
| 715 | +Signed-off-by: Naren Dasan <[email protected]> |
| 716 | +Signed-off-by: Naren Dasan <[email protected]> |
| 717 | +* This commit changes the API for automatic fallback |
| 718 | +to inline settings regarding partial compilation in preparation |
| 719 | +for it to be turned on by default |
| 720 | + |
| 721 | +Now in the compile spec instead of a `torch_fallback` field with its |
| 722 | +associated struct, there are four new fields in the compile spec |
| 723 | + |
| 724 | +```c++ |
| 725 | +bool require_full_compilation = true; |
| 726 | +uint64_t min_block_size = 3; |
| 727 | +std::vector<std::string> torch_executed_ops = {}; |
| 728 | +std::vector<std::string> torch_executed_modules = {}; |
| 729 | +``` |
| 730 | + |
| 731 | +Signed-off-by: Naren Dasan <[email protected]> |
| 732 | +Signed-off-by: Naren Dasan <[email protected]> |
| 733 | +* This commit sets the default workspace size to 1GB for GPU platforms and 256MB for Jetson Nano/TX1 platforms whose compute capability is < 6. |
| 734 | + |
| 735 | +Signed-off-by: Dheeraj Peri < [email protected]> |
| 736 | + |
| 737 | +Signed-off-by: Dheeraj Peri < [email protected]> |
| 738 | + |
| 739 | +Signed-off-by: Dheeraj Peri < [email protected]> |
| 740 | + |
| 741 | +Signed-off-by: Dheeraj Peri < [email protected]> |
| 742 | + |
| 743 | +Signed-off-by: Dheeraj Peri < [email protected]> |
| 744 | +* This commit turns on partial compilation |
| 745 | +by default. Unsupported modules will attempt to be |
| 746 | +run partially in PyTorch and partially in TensorRT |
| 747 | + |
| 748 | +Signed-off-by: Naren Dasan <[email protected]> |
| 749 | +Signed-off-by: Naren Dasan <[email protected]> |
| 750 | +* This commit renames the namespaces of all |
| 751 | +TRTorch/Torch-TensorRT APIs. Now torchscript specific functions |
| 752 | +are segregated into their own torch_tensorrt::torchscript / |
| 753 | +torch_tensorrt.ts namespaces. Generic utils will remain in the |
| 754 | +torch_tensorrt namespace. Guidance on how to port forward will follow in |
| 755 | +the next commits |
| 756 | +* This changes the C++ API ::ts |
| 757 | +APIs to be snake case and for CompileModules to |
| 758 | +become just compile |
| 759 | + |
| 760 | +Signed-off-by: Naren Dasan <[email protected]> |
| 761 | +Signed-off-by: Naren Dasan <[email protected]> |
| 762 | +* This commit updates the pytorch version to 1.10. To use python API of torch_tensorrt, please upgrade your local pytorch to 1.10 to avoid ABI incompatibility errors. WORKSPACE and requirements files are updated accordingly |
| 763 | + |
| 764 | +Signed-off-by: Dheeraj Peri < [email protected]> |
| 765 | + |
| 766 | +Signed-off-by: Dheeraj Peri < [email protected]> |
| 767 | +* This commit changes the default behavior of |
| 768 | +the compiler where if the user does not specify an input data |
| 769 | +type explicity instead of using the enabled precision, now |
| 770 | +the compiler will inspect the model provided to infer the |
| 771 | +data type for the input that will not cause an error if |
| 772 | +the model was run in torch. In practice this means |
| 773 | + |
| 774 | +- If the weights are in FP32 for the first tensor calculation |
| 775 | + then default input type is FP32 |
| 776 | +- If the weights are in FP16 for the first tensor calculation |
| 777 | + then default input type is FP16 |
| 778 | +- etc. |
| 779 | + |
| 780 | +If the data type cannot be determined the compiler will |
| 781 | +default to FP32. |
| 782 | + |
| 783 | +This calculation is done per input tensor so if one input |
| 784 | +is inferred to use FP32 and another INT32 then the expected |
| 785 | +types will be the same (FP32, INT32) |
| 786 | + |
| 787 | +As was the same before if the user defines the data type |
| 788 | +explicitly or provides an example tensor the data type |
| 789 | +specified there will be respected |
| 790 | + |
| 791 | +Signed-off-by: Naren Dasan <[email protected]> |
| 792 | +Signed-off-by: Naren Dasan <[email protected]> |
| 793 | + |
| 794 | + |
579 | 795 |
|
0 commit comments