llama.cpp 发布引入 ZDNN 后端、Vulkan 调优和 Hexagon 支持

作者 PulseAugur 编辑部 · [47 个来源] · 2026-09-28 12:45

llama.cpp 项目发布了多项更新，包括添加 ZDNN 后端构建和各种 CI 改进的版本 b11269。其他近期版本侧重于特定优化和错误修复，例如改进 Adreno GPU 的 OpenCL 内核、为 Intel 硬件调优 Vulkan 着色器，以及增强 AVX512-FP16 的 FP32 点积累加。这些更新还包括对 PLaMo 模型词汇处理的调整，以及对 Hexagon 处理器上 FP32 GELU_ERF 和 GEGLU_ERF 的支持。 AI

影响 llama.cpp 的持续性能优化和后端支持，实现了更广泛的硬件兼容性和效率。

排序理由这是一系列开源项目的软件发布，而非前沿模型发布或重大行业事件。

在 llama.cpp — Releases 阅读 →

AI 生成摘要 · Google Gemini · 来自 47 个来源。我们如何撰写摘要 →

llama.cpp 发布引入 ZDNN 后端、Vulkan 调优和 Hexagon 支持

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: Tool; 这是一系列开源项目的软件发布，而非前沿模型发布或重大行业事件。
Source corroboration: 47 independent sources; Strong cross-source corroboration — multiple independent publishers covered this within the clustering window.
Topics: product, 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: +18 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 [47]

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-30 20:52

b11299

<details open=""> cli: exit on stdin EOF and drop the console wide Ctrl+C broadcast (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29722">#29722</a>) <ul> <li>cli: exit on stdin EOF and drop the console wide Ctrl+C broadcast</li> </ul…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-30 20:22

b11298

<details open=""> mimo : support dflash (convert + feature extraction) (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29650">#29650</a>) <ul> <li> convert : update to support dflash </li> <li> cont : fix </li> </ul> C…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-30 19:40

b11297

<details open=""> jinja : support coerced array attributes (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29574">#29574</a>) <ul> <li> support coerced array attributes </li> <li> add tests </li> </ul> </details> <stro…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-30 18:53

b11295

<details open=""> ci : fix Models Backend Check by shortening the hrm_text fixture (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29744">#29744</a>) The fixture recycles its two blocks over 8 cache slots, so the fp16 error bu…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-30 18:17

b11294

<details open=""> llama: llama_prefetch_rows (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29599">#29599</a>) <ul> <li> llama: llama_prefetch_rows </li> <li> llama: support row prefetch on Windows </li> </ul> Apply t…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-30 17:38

b11293

<details open=""> ggml : add BF16 unary, GLU, binary and scale ops (CPU, CUDA) (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29675">#29675</a>) <ul> <li> ggml : add BF16 unary, GLU, binary and scale ops (CPU, CUDA) </li> <li> …
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-30 16:21

b11292

<details open=""> cpu: accept BF16 in src1 of mul_mat (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28937">#28937</a>) <ul> <li>cpu: accept BF16 in src1 of mul_mat</li> </ul> ggml_conv_1d_dw builds its im2col in F32 when the kerne…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-30 12:39

b11284

<details open=""> openvino: serve GET_ROWS on a weight view from the base Constant (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28381">#28381</a>) <ul> <li>openvino: serve GET_ROWS on a weight view from the base Constant</li> </ul> …
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-30 11:59

b11282

<details open=""> musa : define CUDA_ARCH for device passes (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29508">#29508</a>) The MUSA vendor header never defined CUDA_ARCH, so every architecture<b…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-30 11:30

b11280

<details open=""> SYCL: reduce tensor allreduce sync with pinned host buffers (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29604">#29604</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofo…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-30 11:01

b11279

<details open=""> add GLM-5.3-Flash (GLM5-Next) support (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27773">#27773</a>) <ul> <li> Rebase GLM-Next support onto master, and migrate to llama-memory-hybrid-idx </li> <li> Add i…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-30 10:35

b11278

<details open=""> vendor: update BoringSSL to 0.20260929.0 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29669">#29669</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-30 09:01

b11276

<details open=""> Hexagon f16 activation ops (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29209">#29209</a>) <ul> <li>hexagon: add F16 support for activation ops (SILU/GELU/GELU_QUICK/GEGLU/SWIGLU)</li> </ul> Widens ggml_hexagon_…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-30 08:33

b11275

<details open=""> gguf : reject tensor size that wraps after padding (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26979">#26979</a>) GGML_PAD(nbytes, alignment) wraps to 0 when nbytes is within (alignment - 1) of SIZE_MAX, …
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-30 07:58

b11274

<details open=""> ggml : check row bounds in get_rows_back (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29575">#29575</a>) Signed-off-by: Adrien Gallouët <a href="mailto:[email protected]">[email protected]</a> </details>…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-30 07:25

b11273

<details open=""> model : support classifier_pooling for rerankers (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29627">#29627</a>) <ul> <li>model : support classifier_pooling for ModernBERT rerankers</li> </ul> Assisted-by: Claud…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-30 06:14

b11272

<details open=""> hexagon: optimize concat op (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29673">#29673</a>) <ul> <li>hex-concat: reduce pkts in gather/transpose hot loop</li> </ul> gather directly into dst buffer, use special i…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-30 03:56

b11270

<details open=""> ggml-cuda: HIP: optimize packed byte subtraction (<code>__vsubss4</code> -> <code>__vsub4</code>) (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29478">#29478</a>) <ul> <li> ggml-cuda: HIP: optimize non-saturat…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-30 01:21

b11269

<details open=""> ci: add zdnn backend build but not test (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29541">#29541</a>) <ul> <li>ci: add zdnn backend build but not test</li> </ul> Signed-off-by: Aaron Teo <a href="mailto:aaron.…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-30 00:49

b11268

<details open=""> opencl: fix get_tensor for q5_K adreno gemm_nonshuffle kernel (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29555">#29555</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="no…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-30 00:19

b11267

<details open=""> vulkan: Tune GDN kernel, fix Intel performance (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29476">#29476</a>) <ul> <li> vulkan: tune GDN shader </li> <li> tune for Intel </li> </ul> </details> <st…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-29 23:54

b11266

<details open=""> vulkan : Load F32 A matrix 2 at a time when its 2-aligned (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29254">#29254</a>) It turns out Intel doesn't particularly like loading F32s one at a time and we alre…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-29 23:27

b11265

<details open=""> vulkan: MOE aware mat_mul_id tile selection (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29182">#29182</a>) mut_mul_id selected its matmul tile with total token count. For MoE dispatch grid the true N per …
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-29 23:00

b11264

<details open=""> ggml : fix c++ odr by properly using GGML_COMMON_DECL_CPP (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29504">#29504</a>) <ul> <li> fix c++ odr by properly using GGML_COMMON_DECL_CPP </li> <li> using actu…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-29 22:35

b11263

<details open=""> vocab : keep NORMAL in PLaMo-2 and PLaMo-3 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29580">#29580</a>) <ul> <li>vocab : keep NORMAL in PLaMo-2 and PLaMo-3</li> </ul> The PLaMo-2 and PLaMo-3 vocabularies mark…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-29 21:53

b11262

<details open=""> ggml : accumulate f16 dot products in f32 on AVX512-FP16 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29545">#29545</a>) Supersedes <a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-29 21:22

b11261

<details open=""> ggml : require input tensors to be GGML_OP_NONE (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29647">#29647</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https:…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-29 19:59

b11260

<details open=""> hexagon: add FP32 GELU_ERF and GEGLU_ERF support (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29631">#29631</a>) <ul> <li> hexagon: add FP32 GELU_ERF and GEGLU_ERF support </li> <li> hex-erf: reduce regis…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-29 17:17

b11259

<details open=""> common : stop accepting draft tokens at EOG (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29638">#29638</a>) <ul> <li> common : stop accepting draft tokens at EOG </li> <li> cont : remove the test </li…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-29 16:33

b11258

<details open=""> server : remove the built-in UI's service worker when the UI is not served (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29565">#29565</a>) With --path or --no-ui, /sw.js returned 404, and a 404 does not remove a…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-29 16:02

b11257

<details open=""> common : use fs::path for config dir (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29649">#29649</a>) Signed-off-by: Adrien Gallouët <a href="mailto:[email protected]">[email protected]</a> </details> …
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-29 13:10

b11256

<details open=""> tests : adjust server string regex to also match m2 utlra results (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29648">#29648</a>) </details> Website: <ul> <li><a href="https://llama.app" rel…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-29 12:31

b11255

<details open=""> common : add fs_write_atomic() (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29642">#29642</a>) <ul> <li>Check for buffered write errors when closing downloaded files.</li> <li>Use UTF-8 paths when writing ETag file…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-29 11:22

b11254

<details open=""> ggml : collect all input tensors into graph_inputs (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29634">#29634</a>) graph_inputs was populated while splitting the graph, so it only contained the inputs that…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-29 09:40

b11249

<details open=""> llama : fix init in several tools/examples (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29632">#29632</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://lla…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-29 07:31

b11247

<details open=""> vulkan : reuse descriptor sets when bindings are constant (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29280">#29280</a>) <ul> <li> vulkan : reuse descriptor sets when bindings are constant </li> <li> vul…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-29 06:49

b11246

<details open=""> chat : fix Muse Glimmer ignoring response_format json_schema with --jinja (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29615">#29615</a>) <ul> <li>chat : fix Muse Glimmer ignoring response_format json_schema with -…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-29 06:16

b11245

<details open=""> common : use fs::path for cache dirs (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29595">#29595</a>) <ul> <li>Avoid useless string conversions on Windows.</li> <li>No need for BSD or emscripten special cases.</li> …
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-28 22:30

b11240

<details open=""> server : support typed content (vision/audio/video) input for /v1/embeddings endpoint (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29556">#29556</a>) <ul> <li>server : support multimodal input for /v1/embeddings (Q…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-28 21:31

b11238

<details open=""> models: pad on the left with ggml_pad_ext (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29567">#29567</a>) <ul> <li>models: pad on the left with ggml_pad_ext</li> </ul> The Parakeet, LFM2-Audio, Granite Speech an…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-28 20:55

b11237

<details open=""> ggml-openvino: mark unaligned batch-stride views unsupported (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29603">#29603</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nof…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-28 18:35

b11236

<details open=""> batch: migrate speculative, mtmd and server to batch_ext (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29385">#29385</a>) <ul> <li> adapt common </li> <li> add common_batch </li> <li> wip </li> …
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-28 16:49

b11235

<details open=""> common : fix HF cache paths on Windows (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29475">#29475</a>) Supersedes <a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29158">#2915…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-28 16:16

b11234

<details open=""> webgpu: Handle unaligned writes in ggml_backend_webgpu_buffer_set_tensor (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29471">#29471</a>) <ul> <li> Fix: Handle unaligned writes in ggml_backend_webgpu_buffer_set_t…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-28 15:37

b11233

<details open=""> tests : refactor test-recurrent-state-rollback (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29426">#29426</a>) <ul> <li>tests : use llama_context_ptr in test-recurrent-state-rollback</li> </ul> Replace raw llama…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-28 14:25

b11232

<details open=""> ggml-cpu: enable tiled flash attention for non-vector-multiple head dims on x86 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29423">#29423</a>) <ul> <li> ggml-cpu: enable tiled flash attention for non-vector-mul…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-28 12:45

b11229

<details open=""> HIP: fix template skip for DKQ > 256 mfma kernels (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29559">#29559</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">h…

本文如何被排名

报道来源 [47]

相关实体

相关话题