llama.cpp 发布带来模型支持、性能和安全修复 · 跟踪 8 个来源

作者 PulseAugur 编辑部 · [8 个来源] · 2026-09-16 13:10

llama.cpp 项目发布了多个更新，包括 b11009 版本，为 Gemma4 和 Qwen35 模型添加了对融合 QKV 的支持。早期版本如 b11007 改进了 MTP 的 CUDA 图使用，而 b11006 为 Hexagon 添加了对 K-Quants Q4_K 和 Q6_K 的支持。b11005 版本修复了探针权重放置问题，b11003 版本添加了对 HrmTextForCausalLM 模型（有大量 AI 生成代码贡献）的支持。b11002 版本改进了 CUDA/HIP im2col 中的访问模式，b11001 版本修复了 spacemit 中 int16 数据的转置函数错误。此外，b11000 版本解决了与缓冲区释放和图重新计算相关的关键 RPC 漏洞，该漏洞可能导致远程代码执行。 AI

影响 llama.cpp 项目的持续开发和优化提高了在本地运行大型语言模型的效率和可访问性。

排序理由该集群包含 llama.cpp 项目的多个次要版本说明，详细介绍了错误修复和次要功能添加，而不是重大的新模型发布或研究突破。

在 llama.cpp — Releases 阅读 →

AI 生成摘要 · Google Gemini · 来自 8 个来源。我们如何撰写摘要 →

本文如何被排名

Signal score: 5 / 100; Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket: Tool; 该集群包含 llama.cpp 项目的多个次要版本说明，详细介绍了错误修复和次要功能添加，而不是重大的新模型发布或研究突破。
Source corroboration: 8 independent sources; Strong cross-source corroboration — multiple independent publishers covered this within the clustering window.
Topics: model release, product, safety; 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。

报道来源 [8]

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-16 19:41

b11009

<details open=""> TP: fix split state and granularity for fused QKV gemma4, qwen35 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28965">#28965</a>) <ul> <li>model: calculate split states for attn_qkv from n_head * n_embd_head_k</li> …
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-16 17:49

b11007

<details open=""> Enable CUDA graph for MTP draft (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28549">#28549</a>) <ul> <li> Improve CUDA graph usage for MTP </li> <li> Rename field </li> <li> Address review feedback…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-16 17:22

b11006

<details open=""> hexagon: Support for K-Quants Q4_K and Q6_K (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28994">#28994</a>) implement q6k/q4k kernels Squashed from: feat: implement q6k kernel hex-q6k: improve…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-16 16:53

b11005

<details open=""> hexagon: accept the zeroed rope probe in supports_op (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28995">#28995</a>) llama probes weight placement with a rope where all params are 0, so rejecting n_dims ==…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-16 15:31

b11003

<details open=""> model : add support for HrmTextForCausalLM (DFM Mimir 1B) (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27625">#27625</a>) <ul> <li>model : add support for HrmTextForCausalLM (DFM Mimir 1B)</li> </ul> HRM-Text ru…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-16 14:41

b11002

<details open=""> CUDA/HIP: improve access patterns in im2col (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28013">#28013</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://ll…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-16 14:05

b11001

<details open=""> spacemit : fix wrong transpose function for int16 data (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25161">#25161</a>) The <code>sizeof(int16_t)</code> branch in <code>permute_transpose_impl</code> calls <…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-16 13:10

b11000

<details open=""> rpc : invalidate cached compute graph when a referenced buffer is freed (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24292">#24292</a>) The server caches the most recent compute graph per device so that GR…

本文如何被排名

报道来源 [8]

相关实体

相关话题