Pull requests / #778

#778 iq_kernels: portable aligned shared-memory declaration for gfx1030

closed · @jollyroger1480 · 0 コメント · GitHub で見る

AMD / HIP

本文

HIP clang for gfx1030 (ROCm 7.1) rejects the alignas(16) placement in the two shared grid buffers (iq_kernels.cu:1039/1107): 'alignas attribute cannot be applied to types'. Main no longer compiles for gfx1030 (RX 6800/6900/6950 XT) with CMAKE_HIP_ARCHITECTURES=gfx1030. Moves alignment to a declarator attribute (same layout, compiles everywhere). With this + the split-shard arch-guard (#648), latest main runs the UD-Q4_K_XL abliterated re-split end to end on an RX 6950 XT — and 3/3 clean verify-window runs (the #649-class stall did not reproduce on this build).

関連リンク

インストール・モデル・リリースへの站内リンク。