llama.cpp 发布增加 IQ4_XS 支持、WebGPU ARANGE 和 HVX 置换优化

作者 PulseAugur 编辑部 · [40 个来源] · 2026-10-31 02:24

llama.cpp 项目发布了几个预发布版本，包括 b12069，它为 OpenCL 增加了初始的 IQ4_XS 支持；b12064，它为 hexagon 设备引入了分块的 HVX 置换操作。其他版本侧重于各种优化和功能添加，例如 WebGPU ARANGE 支持、改进 PocketTTS 语音在音频生成中的处理，以及更高效的用于多序列处理的统一 KV 缓存。这些更新还包括特定平台的调整，例如在 Apple 平台上使用 vm_allocate 以及在 ggml-zdnn 中与主机共享缓冲区。 AI

影响对 llama.cpp 的这些更新增强了在各种硬件上运行 AI 模型的性能并增加了新功能。

排序理由对一个专注于软件优化和功能添加的开源项目的多个预发布更新。

在 llama.cpp — Releases 阅读 →

AI 生成摘要 · Google Gemini · 来自 40 个来源。我们如何撰写摘要 →

llama.cpp 发布增加 IQ4_XS 支持、WebGPU ARANGE 和 HVX 置换优化

本文如何被排名

Signal score: 0 / 100; Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket: Research; 对一个专注于软件优化和功能添加的开源项目的多个预发布更新。
Source corroboration: 40 independent sources; Strong cross-source corroboration — multiple independent publishers covered this within the clustering window.
Topics: model release, 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: 3 days old; Aged out of breaking-news scoring windows; ranking reflects the durable signal from the full source set.
Coverage growth since scoring: +8 source(s) since last score; New sources have picked up this story since our last re-score. Score will update on the next scoring pass.

完整方法见我们的编辑标准。

报道来源 [40]

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-11-03 18:28

b12078

<details open=""> tools : reject mismatched shards in GGUF split merge (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/31579">#31579</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">h…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-11-03 17:55

b12077

<details open=""> meta: support GLM-5.3 next sm tensor (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30499">#30499</a>) <ul> <li> meta: support GLM-5.3 next sm tensor </li> <li> glm5-next : cont q_absorbed for -sm tensor wi…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-11-03 17:26

b12076

<details open=""> metal : fix cpy kernel for dst non-contiguous in dim 0 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/31581">#31581</a>) kernel_cpy_t_t indexed the dst innermost dim with an implicit element-size stride inst…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-11-03 15:43

b12074

<details open=""> delta_net_base: refactor conv state (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30878">#30878</a>) <ul> <li> delta_net: access conv state via views </li> <li> fix sm tensor </li> <li> delta_net: f…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-11-03 15:21

b12073

<details open=""> vulkan: fix out-of-bounds NVFP4 UE4M3 scale lookup (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/31079">#31079</a>) <ul> <li>vulkan : fix out-of-bounds NVFP4 UE4M3 scale lookup</li> </ul> On the non-OCP path, ue4…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-11-03 15:03

b12072

<details open=""> llama : drop the redundant cnv from llm_attn_mask (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/31558">#31558</a>) The dspark block mask is now derived at input-fill time: llm_attn_mask::block_uniform fills…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-11-03 11:20

b12069

<details open=""> opencl: add initial iq4_xs support (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/31339">#31339</a>) <ul> <li>opencl: add iq4_xs AOS mv</li> </ul> Assisted-by: Codex <ul> <li>opencl: add iq4_xs SOA mv and tile…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-11-03 10:48

b12068

<details open=""> chat : allow constraint type after gpt-oss final channel (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/31556">#31556</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollo…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-11-03 10:04

b12066

<details open=""> webgpu: add ARANGE support (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/31455">#31455</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.app</a></li> …
llama.cpp — Releases TIER_1 (SO) · FamousMonsterr · 2026-11-03 07:53

b12075

vulkan: skip maxStorageBufferRange pre-check for non-compute tensors …
llama.cpp — Releases TIER_1 (SO) · am17an · 2026-11-03 06:08

b12071

dsv4.1: fix llama-context tests with partial offload (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/31292">#31292</a>)
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-11-03 05:45

b12064

<details open=""> hexagon: tiled hvx permute (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/31446">#31446</a>) <ul> <li>hexagon: add tiled HVX transpose path for permuted CONT/CPY</li> </ul> Permuted copies where dst dim 0 is strid…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-11-03 01:42

b12063

<details open=""> dflash: copy hidden features via device copies (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/31506">#31506</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https:/…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-11-02 23:21

b12061

<details open=""> mtmd: Pocket tts - support for voices without voice cloning (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/31429">#31429</a>) <ul> <li>feat: support PocketTTS voices with no voice cloning</li> </ul> The kyutai/poc…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-11-02 22:44

b12059

<details open=""> llama : flash_attn_ext_rows - more efficient unified kv for multi-seq (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29510">#29510</a>) <ul> <li>llama : use llm_attn_mask and kv_rows for unified attention</li> </ul> …
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-11-02 19:33

b12058

<details open=""> imatrix : fix --nextn crash (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/31522">#31522</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.app</a></li>…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-11-02 18:47

b12054

<details open=""> ggml-zdnn: share buffers with host (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/31481">#31481</a>) Signed-off-by: Aaron Teo <a href="mailto:[email protected]">[email protected]</a> </details> We…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-11-02 18:27

b12053

<details open=""> ggml : use <code>vm_allocate</code> instead of <code>posix_memalign</code> on all Apple platforms (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/31437">#31437</a>) <ul> <li> ggml : use <code>vm_allocate</code> ins…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-11-02 17:41

b12052

<details open=""> opencl: add topk_mask (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/31441">#31441</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.app</a></li> </ul>…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-11-02 16:57

b12050

<details open=""> vulkan: improve small-n split-K selection on RDNA3 and RDNA4 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30146">#30146</a>) <ul> <li> vulkan: improve small-n split-K selection on RDNA3 and RDNA4 </li> <li> …
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-11-02 16:14

b12048

<details open=""> sycl : fast ggml_sycl_comm_exchange() on Intel Arc Pro B70s (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30956">#30956</a>) <ul> <li>Update ggml-sycl.cpp</li> </ul> Increases -sm tensor prompt eval t/s by +19% w…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-11-02 15:51

b12047

<details open=""> sycl: dmmv code cleanup (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/31325">#31325</a>) <ul> <li> sycl: fold q8_0 reorder dmmv into the generic template </li> <li> sycl: template the dmmv non-K launchers<…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-11-02 14:09

b12040

<details open=""> tests : merge identical if branches (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/31523">#31523</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.app<…
llama.cpp — Releases TIER_1 (SO) · nikwen · 2026-11-02 12:13

b12057

context : simplify setters (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29730">#29730</a>)
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-11-02 11:35

b12038

<details open=""> ggml : fix clang-cl build with AVX512 BF16 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/31276">#31276</a>) <ul> <li> Correctly gate immintrin.h inclusion when building with clang-cl and avx512bf16 </li> <li>…
llama.cpp — Releases TIER_1 (SO) · o7si · 2026-11-02 08:34

b12035

cmake : add explicit thread linkage (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/31414">#31414</a>)
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-11-02 08:06

b12034

<details open=""> webgpu : fix CI by keeping scratch alloc sizes monotonic in batch length (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/31422">#31422</a>) <ul> <li>webgpu: keep scratch alloc sizes monotonic in the batch length</li> …
llama.cpp — Releases TIER_1 (SO) · CISC · 2026-11-01 17:23

b12033

requirements : fix numpy dependencies (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/31508">#31508</a>)
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-11-01 09:04

b12032

<details open=""> sycl: Try PkgConfig for LevelZero includes / libs (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/31076">#31076</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">http…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-11-01 07:45

b12031

<details open=""> opencl: add initial iq3 support (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/31338">#31338</a>) <ul> <li>opencl: add iq3_s AOS mv</li> </ul> Assisted-by: Codex <ul> <li>opencl: add iq3_s SOA mv and tiled mm<…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-11-01 07:00

b12029

<details open=""> mtmd : fix InternVL CLS token order (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30886">#30886</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.app<…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-11-01 06:45

b12028

<details open=""> tests : enable Gemma 4 in <code>test-llama-archs</code> (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/31401">#31401</a>) <ul> <li> tests : enable Gemma 4 in <code>test-llama-archs</code> </li> <li> tests :…
llama.cpp — Releases TIER_1 (SO) · nikwen · 2026-10-31 22:30

b12025

llama : fix rotation of attention tensors in layers with shared KV (#…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-10-31 21:38

b12024

<details open=""> reshaping so mm on 2d instead of 3d tensor for lfm sc (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30831">#30831</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-10-31 19:22

b12023

<details open=""> server: cap shared decision prefixes at the first output (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30710">#30710</a>) Limit the shared prefix to the first required output position of the parent and each child…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-10-31 19:02

b12022

<details open=""> server : fix /v1/responses endpoint to accept extra_params (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30959">#30959</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofol…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-10-31 18:46

b12021

<details open=""> mtmd : resize Gemma 4 images without letterbox padding (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30922">#30922</a>) <ul> <li> mtmd : resize Gemma 4 images without letterbox padding </li> <li> Apply sug…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-10-31 08:11

b12020

<details open=""> opencl: add fwht (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/31434">#31434</a>) Co-authored-by: Li He <a href="mailto:[email protected]">[email protected]</a> </details> Website:</…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-10-31 05:23

b12019

<details open=""> chat : strip the think block in the qwen3-coder with reasoning_format none (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/31380">#31380</a>) <ul> <li>chat : accept the think block in the qwen3-coder response format w…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-10-31 02:24

b12018

<details open=""> sycl: fix f32 to q2_0/q4_0/q5_0 rounding in CPY (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/31120">#31120</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https:…

本文如何被排名

报道来源 [40]

相关实体

相关话题