Releases · iree-org/iree

Announcements:

We plan on switching from our current version scheme (e.g. 20241104.1068) to a new version scheme (e.g. 3.0.1) with the next release. See #18938 for full details

Python 3.13 is now supported.
Support for ONNX operations and models continues to grow.
The CMake IREE_TARGET_BACKEND_CUDA, IREE_HAL_DRIVER_CUDA, IREE_TARGET_BACKEND_ROCM, and IREE_HAL_DRIVER_HIP options now default to OFF: f2bf602
CUDA support is now included in published aarch64 packages: 3dda3cd
Indirect command buffers are now enabled by default: a730349
- --iree-hal-memoization=true
- --iree-hal-indirect-command-buffers=true
The --iree-codegen-*-transform-dialect-jit flags have been off by default for a while and were removed across all backends: 012f8a6
A warning is now issued when the --iree-llvmcpu-target-cpu= or --iree-llvmcpu-target-cpu-features flags are unset, since the default of "generic" is slow. These flags will be required in the future: 0c6a151
The ROCm and CUDA compiler targets now support executable linking: 49ffdac
Support for more data types, including 8-bit floats is improving: 2291b38
The compiler is now better able to optimize sequences of arithmetic operations common in ML frameworks like PyTorch: 0f28d44
GitHub Actions workflows are using a new cluster of self-hosted runners: c3cbfbd, cc891ba
The legacy ireec console script has been removed in favor of iree-compile: 84d0789

Support for data-tiling on GPU targets is in progress.
A new iree.build package is being developed to help construct model compilation pipelines: 0077358