English(EN) b11735: metal : fix moe_reduce fusion race when dst aliases its inputs (#30795)

llama.cpp 修复了 Metal 后端 moe_reduce 融合中的竞态条件

作者 PulseAugur 编辑部 · [1 个来源] · 2026-10-19 08:55

llama.cpp 项目发布了更新 b11735，解决了其 Metal 后端中与 `moe_reduce` 融合相关的竞态条件。此错误在处理 Kolibri 等某些模型时可能导致罕见的、非确定性的错误结果。该修复程序可在目标张量与输入重叠时阻止融合，并恢复到更安全的、未融合的乘加操作。 AI

影响确保使用 llama.cpp 上的 Metal 后端的模型的推理更稳定和正确。

排序理由这是针对推理基础设施的开源项目的错误修复，而不是新的模型发布或重大研究。

AI 生成摘要 · Google Gemini · 来自 1 个来源。我们如何撰写摘要 →

本文如何被排名

Signal score: 3 / 100; Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket: Tool; 这是针对推理基础设施的开源项目的错误修复，而不是新的模型发布或重大研究。
Source corroboration: Single-source cluster; Only one publisher covered this so far. Single-source stories can still rank when the publisher is high-authority, but they lack cross-source corroboration.
Topics: infra; Editorial topic classification. Feeds into how the story surfaces on /topic/<slug> hub pages and into the per-entity coverage mix.
AI-industry relevance: High; Clearly on-topic for AI-industry coverage.
Story freshness: Same-day; Cluster formed today. Ranking reflects the current source set at time of score.

完整方法见我们的编辑标准。

llama.cpp — Releases TIER_1 English(EN) · ggerganov · 2026-10-19 08:55

b11735: metal : 修复 moe_reduce fusion 竞争条件，当 dst 别名化其输入时 (#30795)

<p>The fused moe_reduce kernel computes dst = sum(experts[e] * weights[e]) in a<br /> single dispatch, reading experts and weights while writing dst. The graph<br /> allocator computes tensor liveness on the unfused graph, where the inputs are<br /> fully consumed by earlier node…