Release Release v3.1.0 · iree-org/iree

Notable changes

The iree-import-onnx tool now includes options that improve handling of large models: ab3c9bb. See the ONNX framework documentation at https://iree.dev/guides/ml-frameworks/onnx/.
The PJRT plugin code at integrations/pjrt/ has been refreshed and is now tested continuously. Thank you @PragmaTwice !

Compiler support for "tuning" continues to improve. See the new documentation at https://iree.dev/reference/tuning/ and the discussion on #19214.
Support for ukernels in GPU targets has evolved:
- AMDGPU ukernels were ported from HIP to self-contained C in d704051.
- AMDGPU ukernels for select targets are now bundled with the compiler and are selected, loaded, and linked in during compilation thanks to 8272490, cbb11f2, and dc29ee7. Externally-defined ukernels can also be provided: 82a89e3.
Compilation time for large models has improved substantially with 71d6de7.

There were breaking changes to some iree_hal_* APIs in eae7bfb:
- iree_hal_buffer_subspan now requires an iree_allocator_t host_allocator that was previously implicit.
- iree_hal_subspan_buffer_initialize was removed as it was not safe.
- iree_hal_deferred_buffer_t was removed as placement checks without an allocated buffer reference are not possible.
- iree_hal_heap_buffer_wrap and all other buffer creation now requires a placement.
- iree_hal_buffer_initialize used by HAL implementations now requires a placement.
The HIP runtime can now create a single logical device backed by multiple physical devices: 0e71e72, 1f19761.
Improved support for loading large programs in the HIP HAL: 47ccd93, 7ff83ea.

The iree.build package now has reference documentation at https://iree-python-api.readthedocs.io/en/latest/compiler/build.html, has more robust HTTP fetching (9789438), and supports out of process concurrency (ef4ecf3).
Development versions of Python packages are now easier to install using build_tools/pkgci/setup_venv.py and a new package bisect script is available at build_tools/pkgci/bisect.