Skip to content

Release v0.11.1 - #1637

Merged
IgorSwat merged 2 commits into
mainfrom
@is/binaries-v0.11.2
Nov 16, 2026
Merged

IgorSwat merged 2 commits into
mainfrom
@is/binaries-v0.11.2

Conversation

@IgorSwat

@IgorSwat IgorSwat commented Nov 16, 2026 •

Copy link
Copy Markdown
Contributor

Description

Points nativeLibsVersion at v0.11.1-libs.
That release was re-published with new Android core and vulkan libraries from
software-mansion-labs/executorch#32 (7764e753e). That branch contains 0a087190, the commit the original
v0.11.1-libs was built from, so nothing from it is lost. No released version pinned v0.11.1-libs (v0.11.0 and the
nightlies pin v0.11.0-libs), so replacing its assets affects no existing install.

What changed

Compared with the original v0.11.1-libs, only the Android core and Vulkan artifacts were replaced:

Artifact Change
vulkan-android-{arm64-v8a,x86_64} Vulkan backend from #32
core-android-{arm64-v8a,x86_64} libexecutorch.so rebuilt from the same commit; executorch.jar unchanged
all other 24 artifacts (headers, iOS, XNNPACK, OpenCV, QNN) unchanged since v0.11.1-libs was first published

Vulkan backend changes:

  • New kernels for Reverb TTS: an int4 linear (tiled matmul through shared memory), a tiled conv1d and an attention
    QK kernel.
  • Correct results on Adreno (Snapdragon): the Adreno compiler miscompiled a bounds-checked load pattern in the
    tiled kernels, and every Reverb model returned garbage there. Fixed by clamping the index and loading unconditionally.
  • fp16 LayerNorm fix: on GPUs that flush fp16 subnormals (Adreno does), LayerNorm returned NaN for all-zero rows,
    because its epsilon is subnormal in fp16. The normalisation factor is now computed in fp32. fp32 LayerNorm is
    unchanged.
  • Shared memory planning fix: before this change, the memory consumption was growing linearly with each method of the program sharing model's weights. This resulted in an enormous memory usage (~3.5 GB For Reverb-TTS backbone).

Introduces a breaking change?

  • Yes
  • No

Type of change

  • Bug fix (change which fixes an issue)
  • New feature (change which adds functionality)
  • Documentation update (improves or adds clarity to existing documentation)
  • Other (chores, tests, code style improvements etc.)

Tested on

  • iOS
  • Android

Testing instructions

The benchmark app on a OnePlus 12 (Adreno 750): yarn bench --platform android --backends vulkan --suite full, with the
v0.11.0 libs and with these:

  • All 75 Vulkan models run, 0 failures with both lib sets.
  • No regressions past the comparator's tolerance. The two Whisper forward flags in the first run were noise:
    a 3-repeat recheck came back within noise.
  • Style transfer is about 6.6% slower (~20 ms), which is under tolerance. It is likely the Mali style-transfer fix
    that the original v0.11.1-libs already shipped, not this change.
  • The Reverb models were also checked for numerical correctness against references on the Galaxy S24 FE (Xclipse) and
    the OnePlus.

Screenshots

Related issues

PR to the ExecuTorch codebase

Checklist

  • I have performed a self-review of my code
  • I have commented my code, particularly in hard-to-understand areas
  • I have updated the documentation accordingly
  • My changes generate no new warnings

Additional notes

@IgorSwat
IgorSwat requested a review from barhanc November 16, 2026 08:48
@msluszniak

Copy link
Copy Markdown
Member

Are these changes upstreamed on ExecuTorch repository or not? Or maybe they have been already applied in upstream ET?

@IgorSwat

Copy link
Copy Markdown
Contributor Author

Are these changes upstreamed on ExecuTorch repository or not? Or maybe they have been already applied in upstream ET?

There is a PR with them, I believe not merged yet.

@barhanc barhanc left a comment

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why did you skip to 0.11.2, I'm pretty sure the binaries under 0.11.1 prerelease can be updated / substituted for the new ones.

@msluszniak

Copy link
Copy Markdown
Member

Are these changes upstreamed on ExecuTorch repository or not? Or maybe they have been already applied in upstream ET?

There is a PR with them, I believe not merged yet.

Sure thing, could you just pin this PR here then?

@barhanc barhanc added the chore PRs that are chores label Nov 16, 2026
@IgorSwat IgorSwat mentioned this pull request Nov 16, 2026
7 of 12 tasks
@IgorSwat IgorSwat changed the title Release v0.11.2 Release v0.11.1 Nov 16, 2026
@IgorSwat
IgorSwat force-pushed the @is/binaries-v0.11.2 branch from 36b3a66 to 457beb2 Compare November 16, 2026 15:51
@IgorSwat
IgorSwat merged commit 339614a into main Nov 16, 2026
6 checks passed
@IgorSwat
IgorSwat deleted the @is/binaries-v0.11.2 branch November 16, 2026 15:56
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

chore PRs that are chores

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants