Native libraries for React Native ExecuTorch, built from software-mansion-labs/executorch @ @ms/separate-backends-1.4.1 (13b1d158).
Same content as v0.10.2-libs, with the XNNPACK backend rebuilt to fix a use-after-free of unpacked constants in the weights cache (upstream pytorch/executorch#22779). finalize_for_runtime freed every unpacked constant once the runtime was created, but XNNPACK does not pack all of them: PReLU reads its slope straight out of that memory for the life of the runtime, so any delegated model with PReLU read freed heap. It surfaced as NaN outputs from Face Mesh. The weights cache is enabled by both the Android and Apple presets, so every published XNNPACK library was affected.
Only xnnpack-android-arm64-v8a, xnnpack-android-x86_64 and xnnpack-ios change. The other ten artifacts are the v0.10.2-libs files byte for byte.
Provenance: rebuilding the parent commit 0e18ed635 reproduced the published v0.10.2-libs arm64 libxnnpack_executorch_backend.so md5 exactly (02e21da366d2646cf2e189b0cbcfc85a), so the only delta in these libraries is the fix. libXnnpackBackend.a keeps the same 1098 objects and the same undefined ExecuTorch symbol set apart from the one changed signature. Verified on device: Face Mesh renders its 468 landmarks from the published .pte, which returned all-NaN before.