hrx: prefill attention keeps its rate at long key counts (+26 % prefill at 24k tokens on Qwen3-4B; engine #1989) - #289

hrx: prefill attention keeps its rate at long key counts (+26 % prefill at 24k tokens on Qwen3-4B; engine #1989)#289

bong-water-water-bong merged 1 commit into

1bit/hrx-vulkan-patched1bit-MONSTER/llama.cpp:1bit/hrx-vulkan-patchedfrom

1bit/1989-fa-value-transpose1bit-MONSTER/llama.cpp:1bit/1989-fa-value-transposeCopy head branch name to clipboard