10 releases 11 tags

RSS feed

• build-18-44948c7 44948c70c4

  Compare

  build-18-44948c7 Stable

  taxmalalas0001 released this 2026-10-18 15:20:42 +00:00 | 8 commits to master since this release

  Commit: 44948c70c4

  Build: release

  Message:

  Fix crash when asking the Explain chat a question

  • Guard the first GPU replies with the crash marker, not just init: a GPU
    that initialises but faults in prefill/decode crashed on every question
    and never fell back to CPU. A GPU generation error also disables GPU.
  • Declare libvndksupport.so next to libOpenCL.so (per LiteRT-LM docs), so
    Adreno OpenCL actually loads instead of LiteRT falling back to WebGPU.
  • Cap maxNumTokens at 4096 so the KV cache isn't sized for the full
    context window on the first prefill.
  • Stream via the callback API: cancelling now calls cancelProcess(), the
    native conversation is only freed after generation has ended (closing
    mid-reply was a use-after-free), and the chunk buffer is unlimited so
    tokens are no longer dropped under a busy main thread.

  Downloads

  • Source code (ZIP)
    0 downloads
  • Source code (TAR.GZ)
    0 downloads

  ---

  • app-release.apk
    9 downloads · 2026-10-18 15:20:43 +00:00 · 53 MiB