Release tracker - 3.2.0

[ScottTodd]

Overview

Our previous stable release was cut on 2025-01-07: https://github.com/iree-org/iree/releases/tag/v3.1.0. We aim for roughly one stable release every 6 weeks, which will be around 2025-02-17 2025-02-10. This issue tracks the release process for the next release.

Release process docs:

• https://iree.dev/developers/general/release-management/
• https://github.com/iree-org/iree/blob/main/RELEASING.md

Release checklist:

• Watch for major/breaking changes and decide to either batch them with this release or defer them until the next release
• Choose a release candidate from nightly releases to promote. The candidate should contain no major regressions and should include all packages, even those marked experimental in .github/workflows/build_package.yml like macOS and Windows packages.
• Compile release notes
• Promote to stable (on GitHub) and push packages to PyPI
• Roll forward downstream packages like iree-turbine and push similar package updates to PyPI
• Update release process docs for next time

Release notes

Candidate: https://github.com/iree-org/iree/releases/tag/iree-3.2.0rc20250206

Commit range

• Previous release to main: v3.1.0...main
• Previous release to release 3.2.0rc20250206: v3.1.0...iree-3.2.0rc20250206

(Contributors can edit this section as they land patches)

---

Notable changes

Framework support

• The TOSA dialect, commonly used when importing from TensorFlow Lite (LiteRT), is in the process of incrementing to v1.0: https://discourse.llvm.org/t/rfc-tosa-dialect-increment-to-v1-0/83708. During this transition, we anticipate that files imported from .tflite to .mlirbc may not be compatible with the IREE compiler. See Bump to llvm/llvm-project@3f1486f #19683 for more details.
• Support for importing and compiling TensorFlow models is known to be unstable. We expect this to improve after migrating to new APIs. Follow [RFC] Deprecate the iree-import-tf tool and iree-tools-tf package #19917 for updates.

Compiler

• Remove legacy sync path #19714: The legacy_sync compilation mode has been removed since all in-tree compiler targets have been migrated off of it. The remaining asynchronous mode produces non-blocking operations that enable more multi-device parallelism.
• [Tuner][Codegen] Make default tuning be available per-SKU instead of per-arch #19720: The tuner now supports specializing for GPUs that share the same ISA but have different hardware capabilities.
• Compilation time improvements, particularly for large, sharded models
  • [Stream][NFC] Iterate over blocks to find return op #19791
  • [Stream][NFC] Move topological sort out of loop #19794
• [CPU] Add support for parsing AArch64 cpu features. #19881: The llvm-cpu target now supports parsing AArch64 cpu features.

Runtime

• Bump tracy to wolfpld/tracy@5479a42 and update local build system. #19842: IREE is now using a newer (currently unreleased) version of Tracy. To capture and view traces of the IREE runtime, a matching version of the Tracy tools should be used. See also https://iree.dev/developers/performance/profiling-with-tracy/.
• Supporting (and renaming) IREE_HAL_WHOLE_BUFFER in binding table resolve. #19640: The special IREE_WHOLE_BUFFER value was renamed to IREE_HAL_WHOLE_BUFFER.

Development tools

• Update GH actions with Dependabot #19663: More source dependencies are now managed by Dependabot.
• [Python] Enable building Python bindings as editable wheels, document it #19716: IREE's Python bindings can now be installed as editable wheels. See the documentation at https://iree.dev/building-from-source/getting-started/#python-bindings.
• Port compiler and iree-dialects to nanobind #19790: IREE now uses nanobind for its runtime and compiler Python bindings, instead of mixing nanobind and pybind11.

Full changelog

v3.1.0...v3.2.0

---

[ScottTodd]

New release date target: Feb 10, instead of Feb 17.

[ScottTodd]

We've updated the downstream shark-ai packages to use 3.2.0rc20250206 (the most recent, from last night) in nod-ai/shark-ai#928 and I'm not aware of any particularly risky or breaking changes recently, so I'll vote for that to be our release candidate this release cycle.

[ScottTodd]

Release candidate install instructions, including the downstream shark-ai packages:

pip install \
  --find-links https://github.com/nod-ai/shark-ai/releases/expanded_assets/dev-wheels \
  shortfin[apps]==3.2.0rc20250207 \
  sharktank==3.2.0rc20250207

pip install \
  --find-links https://iree.dev/pip-release-links.html \
  iree-base-compiler==3.2.0rc20250206 \
  iree-base-runtime==3.2.0rc20250206 \
  iree-turbine==3.2.0rc20250205

Then check what was installed:

pip freeze | grep -E 'iree|shark|shortfin'

iree-base-compiler==3.2.0rc20250206
iree-base-runtime==3.2.0rc20250206
iree-turbine==3.2.0rc20250205
sharktank==3.2.0rc20250207
shortfin==3.2.0rc20250207

Note: we could also use iree-turbine==3.2.0rc20250207 to pick up a few more changes there.

[ScottTodd]

For future releases, I hope to reduce those instructions to a single pip install shark-ai -f [nightly asset urls] command, see nod-ai/shark-ai#929

[aviator19941]

Seems like to use shortfin with the above commands, more dependencies need to be installed:

Traceback (most recent call last):
  File "/home/nod/avsharma/SHARK-Platform/3.11.venv/lib/python3.11/site-packages/shortfin_apps/llm/_deps.py", line 17, in <module>
    __import__(dep)
ModuleNotFoundError: No module named 'dataclasses_json'

The above exception was the direct cause of the following exception:

Traceback (most recent call last):
  File "<frozen runpy>", line 189, in _run_module_as_main
  File "<frozen runpy>", line 112, in _get_module_details
  File "/home/nod/avsharma/SHARK-Platform/3.11.venv/lib/python3.11/site-packages/shortfin_apps/llm/__init__.py", line 7, in <module>
    from . import _deps
  File "/home/nod/avsharma/SHARK-Platform/3.11.venv/lib/python3.11/site-packages/shortfin_apps/llm/_deps.py", line 27, in <module>
    raise ShortfinDepNotFoundError(__name__, dep) from e
shortfin.support.deps.ShortfinDepNotFoundError: Shortfin is missing a dependency to use shortfin_apps.llm. This is typically available via `pip install dataclasses_json`

[ScottTodd]

Seems like to use shortfin with the above commands, more dependencies need to be installed:

Ah, let me amend the instructions then: install shortfin[apps]==3.2.0rc20250207. That should get dataclasses-json from https://github.com/nod-ai/shark-ai/blob/2c61420ff5b54de8990a7f2f8941f987023cd690/shortfin/pyproject.toml#L37-L45.

[ScottTodd]

Possible release blocking regression reported downstream: nod-ai/shark-ai#934. Waiting for more information to determine the impact. We could release IREE + iree-turbine 3.2.0 and then follow with a 3.2.1 patch release for downstream shark-ai to use in its own 3.2.0.

Should know more by Monday.

(Then for future releases we can also establish more of a process for branching, vetting release candidates, and incorporating feedback from downstream users/projects)

[ScottTodd]

New plan this morning is to not block the release on that issue (matching IREE issue: #19948). Might add to a list of known issues in one or both projects. I'll start the release rollout.

[ScottTodd]

https://github.com/iree-org/iree/releases/tag/v3.2.0 is now published and pushed to pypi.

Next up:

• iree-turbine
• shark-ai
• Update release docs

[ScottTodd]

All packages are pushed to pypi and all releases are published:

• https://github.com/iree-org/iree/releases/tag/v3.2.0
• https://github.com/iree-org/iree-turbine/releases/tag/v3.2.0
• https://github.com/nod-ai/shark-ai/releases/tag/v3.2.0

Working on process documentation updates next.