llama.cpp fixes quantized weight views in its OpenVINO backend

This edition covers the wider gap since the 09:00 UTC publication. llama.cpp b11284 fixes an OpenVINO inference failure involving GET_ROWS over quantized weight views, restoring those cases and passing the backend's full test suite in the contributor's validation.

AI developments in this edition

Copy Markdown
  1. llama.cpp b11284 restores OpenVINO GET_ROWS for quantized weight views

    llama.cpp b11284 changes its OpenVINO backend so a view over a quantized model weight is served from the underlying Constant, with the view's row offset folded into gather indices instead of treating the view as a dynamic Parameter. The bug caused GET_ROWS failures for quantized weight views; the merged fix reports 51/51 GET_ROWS cases and 4,344/4,344 full backend tests passing with OpenVINO 2026.4 on CPU, versus eight GET_ROWS failures before the change. For builders using llama.cpp through OpenVINO, this removes a concrete inference compatibility failure affecting quantized model weights.

Last checked

RSS