llama.cpp excludes Adreno MoE kernels to fix garbage output
Local inference on Adreno devices broke for MoE models until this exclusion landed. Users on affected hardware must pull the tagged build to restore correct output. The source does not state whether a corrected kernel will return in a later release. llama.cpp tagged b10034 on 2026-07-15, excluding some MoE kernels on Adreno A7x and A6x after compilers miscom
Introduction
llama.cpp tagged b10034 on 2026-07-15, excluding some MoE kernels on Adreno A7x and A6x after compilers miscompiled repack kernels and corrupted weights, causing MoE models to generate garbage output. The release also lists build matrices for macOS, Linux, Android, and Windows. The fix is an upstream GitHub release record. Official upstream release record from llama.cpp; verify the linked page before approval.
What changed
llama.cpp tagged b10034 on 2026-07-15, excluding some MoE kernels on Adreno A7x and A6x after compilers miscompiled repack kernels and corrupted weights, causing MoE models to generate garbage output. The release also lists build matrices for macOS, Linux, Android, and Windows. The fix is an upstream GitHub release record. Official upstream release record from llama.cpp; verify the linked page before approval. [S10]
Why it matters now
Local inference on Adreno devices broke for MoE models until this exclusion landed. Users on affected hardware must pull the tagged build to restore correct output. The source does not state whether a corrected kernel will return in a later release. llama.cpp tagged b10034 on 2026-07-15, excluding some MoE kernels on Adreno A7x and A6x after compilers miscompiled repack kernels and corrupted weights, causing MoE models to generate garbage output. The release also lists build matrices for macOS, Linux, Android, and Windows. The fix is an upstream GitHub release record. [S10]
The local and hardware angle
Adreno A6x/A7x local MoE inference produced corrupted output; b10034 disables the broken kernels. Portability on Android ARM is restored by avoidance rather than fix. Local inference on Adreno devices broke for MoE models until this exclusion landed. Users on affected hardware must pull the tagged build to restore correct output. The source does not state whether a corrected kernel will return in a later release. llama.cpp tagged b10034 on 2026-07-15, excluding some MoE kernels on Adreno A7x and A6x after compilers miscompiled repack kernels and corrupted weights, causing MoE models to generate garbage output. The release also lists build matrices for macOS, Linux, Android, and Windows. The fix is an upstream GitHub release record. [S10]
Local inference on Adreno devices broke for MoE models until this exclusion landed. Users on affected hardware must pull the tagged build to restore correct output. The source does not state whether a corrected kernel will return in a later release. Adreno A6x/A7x local MoE inference produced corrupted output; b10034 disables the broken kernels. Portability on Android ARM is restored by avoidance rather than fix. llama.cpp tagged b10034 on 2026-07-15, excluding some MoE kernels on Adreno A7x and A6x after compilers miscompiled repack kernels and corrupted weights, causing MoE models to generate garbage output. The release also lists build matrices for macOS, Linux, Android, and Windows. The fix is an upstream GitHub release record.
Primary sources
- [S10]b10034
llama.cpp · July 15, 2026
Official upstream release record from llama.cpp; verify the linked page before approval.
Sources, review, and corrections
This article was reviewed against the cited source records. Evidence limitations remain visible, and material corrections are reflected in the updated date.
Editorial note: Expanded from story 7 of week-2026-07-13. The article opens with the documented change, then separates context, consequence, and any supported local-workflow angle.
Read our evidence and corrections policy →