From 92757a9a27bb4bca64bc68f48e99ac845a68f085 Mon Sep 17 00:00:00 2001 From: bong-water-water-bong Date: Thu, 12 Nov 2026 11:31:55 -0400 Subject: [PATCH] Bump llama.cpp (hrx) 962507f8e -> 9dd1a1b72: hip_top_k_f32 gives the same indices every run (engine #2146) llama.cpp #328: the three-level radix select reset its shared bucket before every lane had read it, and the emit placed elements with atomicAdd, so the order (and, with ties, the set) changed between runs. The Flash-Next indexer's TOP_K feeds the sparse attention, which is why identical 18.8k prompts gave different replies. Also carries #326 (a DISABLE_DISPATCH=fused log line). Co-Authored-By: Claude Opus 5.5 --- registry/architectures.json | 2 +- third_party/llama.cpp | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/registry/architectures.json b/registry/architectures.json index d95183a71..ee338d5b8 100644 --- a/registry/architectures.json +++ b/registry/architectures.json @@ -1,7 +1,7 @@ { "about": "HF architecture -> GGUF architecture and the backends whose code accepts it. Generated by tools/registry_build.py from the pinned sources; do not edit.", "sources": { - "llama.cpp (hrx)": "962507f8e8ba2cdbecd6ba061bb49a6900e4881b" + "llama.cpp (hrx)": "9dd1a1b72384fbda84c100911296d2a0e96674cb" }, "counts": { "hrx": 298, diff --git a/third_party/llama.cpp b/third_party/llama.cpp index 962507f8e..9dd1a1b72 160000 --- a/third_party/llama.cpp +++ b/third_party/llama.cpp @@ -1 +1 @@ -Subproject commit 962507f8e8ba2cdbecd6ba061bb49a6900e4881b +Subproject commit 9dd1a1b72384fbda84c100911296d2a0e96674cb