llama.cpp 发布带来 Vulkan、CUDA 等优化

作者 PulseAugur 编辑部 · [51 个来源] · 2026-09-04 17:57

llama.cpp 项目发布了多个更新，包括 b10883，该更新针对 Apple Silicon 和其他平台上的矩阵乘法进行了 Vulkan 优化。其他近期发布如 b10881 和 b10877 解决了特定的问题，例如 Vulkan 工作组分配和针对专家混合 (MoE) 模型的 CUDA 优化。这些更新还包括针对不同操作系统和硬件配置的各种错误修复和性能增强。 AI

影响 llama.cpp 的持续改进提高了本地 LLM 在各种硬件和操作系统上的部署性能和兼容性。

排序理由该集群包含 llama.cpp 项目的多个发布说明，详细介绍了增量更新和错误修复，而不是新模型发布或重大的研究突破。

在 llama.cpp — Releases 阅读 →

AI 生成摘要 · Google Gemini · 来自 51 个来源。我们如何撰写摘要 →

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; 该集群包含 llama.cpp 项目的多个发布说明，详细介绍了增量更新和错误修复，而不是新模型发布或重大的研究突破。
Source corroboration: 51 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: 5 days old; Aged out of breaking-news scoring windows; ranking reflects the durable signal from the full source set.

完整方法见我们的编辑标准。

报道来源 [51]

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-09 17:32

b10883

<details open=""> vulkan: use spec constant for matrix matrix multiplication A-type (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25773">#25773</a>) <ul> <li>vulkan: use spec constant for mul mat type_a</li> </ul> vulkan: use map …
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-09 15:34

b10881

<details open=""> vulkan: Convert FILL to distribute workgroups in 2D to avoid exceeding maxComputeWorkGroupCount (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28592">#28592</a>) <ul> <li>divide workload to 2D</li> </ul> This is t…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-09 13:25

b10878

<details open=""> llama : use int32_t for llama_sampler_chain_n return type (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28631">#28631</a>) Contributes to <a class="issue-link js-issue-link" href="https://github.com/ggml-org/llam…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-09 12:55

b10877

<details open=""> CUDA: size routed MoE MMQ N-tiles from typical expert width on RDNA3 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28552">#28552</a>) Recreated from <a class="issue-link js-issue-link" href="https://github.com/gg…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-09 11:47

b10876

<details open=""> CUDA: replace GGML_FA_ALL_QUANTS with GGML_FA_QUANTS, more control over what is compiled (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28079">#28079</a>) <ul> <li>CUDA: add configurable FA quant combinations</li> </…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-09 11:17

b10875

<details open=""> args: officially deprecate --mmap|mlock|dio (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28334">#28334</a>) Signed-off-by: Aaron Teo <a href="mailto:[email protected]">[email protected]</a> </details> <…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-09 09:56

b10874

<details open=""> model: fix granite3 moe unknown parameter count (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28632">#28632</a>) Signed-off-by: Aaron Teo <a href="mailto:[email protected]">[email protected]</a> </details> …
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-09 09:24

b10873

<details open=""> mtmd: propagate video ID to bitmap (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28601">#28601</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.app</…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-09 07:47

b10872

<details open=""> jinja: treat a null left operand of in as a plain lookup (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28620">#28620</a>) Templates that default an optional variable to none and then test its membership in …
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-09 07:20

b10871

<details open=""> vulkan: add dedicated iq4_xs mat-vec shader (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28426">#28426</a>) <ul> <li>vulkan: add dedicated iq4_xs mat-vec shader</li> </ul> Dedicated mul_mat_vec_iq4_xs for the dm…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-09 06:40

b10870

<details open=""> vulkan: add f16 B-type matmul pipelines and warp tile size tuning for Intel coopmat1 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27471">#27471</a>) <ul> <li> vulkan: add f16 B-type matmul pipelines and warp til…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-09 05:27

b10869

<details open=""> tests : use less threads for data initialization (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28325">#28325</a>) <ul> <li> tests : use 1 thread for data initialization </li> <li> cont : scale threads with…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-09 02:03

b10868

<details open=""> Add IQ type handling for MoE (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28476">#28476</a>) Co-authored-by: cwriter <a href="mailto:cwriter@localhost">cwriter@localhost</a> </details> Website:</s…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-08 17:33

b10867

<details open=""> llama: disable lazy tensor loading by default on iGPUs (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28326">#28326</a>) <ul> <li> llama: add lazy mode auto, fix iGPU regression </li> <li> revert changes ex…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-08 15:14

b10865

<details open=""> Revert "ggml-cuda : restore prop.integrated on HIP builds (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24233">#24233</a>)" (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28604">#2…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-08 14:33

b10864

<details open=""> server : apply checkpoint min-step eviction only when the checkpoint list is full (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28302">#28302</a>) The spacing eviction in create_checkpoint() keeps the oldest chec…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-08 14:01

b10863

<details open=""> metal : fix idle threads in mul_mv_iq3_xxs for ne00 < 1024 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28086">#28086</a>) <ul> <li> metal : fix half-idle simdgroup in kernel_mul_mv_iq3_xxs_f32 for ne00 < …
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-08 12:36

b10859

<details open=""> llama : add missing headers (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28566">#28566</a>) <ul> <li>fix compile-error: add missing header</li> </ul> Bug: <a class="issue-link js-issue-link" href="https://github…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-08 12:08

b10858

<details open=""> vulkan : fuse UNARY(GELU|SIGMOID|SILU|SOFTPLUS) + MUL (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27220">#27220</a>) <ul> <li> vulkan : fuse UNARY(SIGMOID|SILU|SOFTPLUS) + MUL </li> <li> vulkan : fuse UN…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-08 11:38

b10857

<details open=""> Fix Vulkan-Hpp handle usage on 32-bit targets. (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/22892">#22892</a>) On 32-bit platforms, Vulkan non-dispatchable handles such as VkBuffer are represented as uint6…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-08 10:50

b10856

<details open=""> chat : split specialized parsers into common/parsers (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27764">#27764</a>) <ul> <li>chat : split specialized parsers into common/parsers</li> </ul> Move the 14 dedicated…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-08 10:06

b10855

<details open=""> opencl: properly handle non-contiguous inputs to conv2d (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28503">#28503</a>) <ul> <li> opencl: fix conv2d non-contiguous strides </li> <li> opencl: format </…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-08 04:00

b10853

<details open=""> model : support Kimi-K3 recurrent-state rollback (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28466">#28466</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-08 00:42

b10852

<details open=""> hexagon: add RELU and LEAKY_RELU ops (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28585">#28585</a>) <ul> <li> hexagon: add RELU op </li> <li> hexagon: add LEAKY_RELU op too </li> </ul> </details> …
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-07 20:05

b10850

<details open=""> tests : initialize the L2_NORM batch array (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28553">#28553</a>) <ul> <li>tests: bind the L2_NORM batch count to a local</li> </ul> GCC cannot prove the loop fills norms…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-07 19:30

b10844

<details open=""> vulkan: add DeepSeek-V4 hyper-connection fused ops (DSV4_HC_COMB/PRE/POST) (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26578">#26578</a>) <ul> <li>vulkan: add DeepSeek-V4 hyper-connection fused ops (DSV4_HC_COMB/P…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-07 18:16

b10842

<details open=""> ggml: add gfx90c HIP support (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26454">#26454</a>) <ul> <li> ggml: add gfx90c HIP support </li> <li> ggml: make gfx90c HIP support compliant with specifications</…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-07 14:33

b10840

<details open=""> CUDA: branchless Q4_K/Q5_K unpack to speed up mmvq, L2 prefetch on DGX Spark (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26705">#26705</a>) <ul> <li> Update Q4_K and Q5_K to use branchless computation, which st…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-07 11:16

b10839

<details open=""> vulkan: support type-aligned GET_ROWS (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28253">#28253</a>) <ul> <li>vulkan: fall back to CPU for GET_ROWS with misaligned offsets</li> </ul> The Vulkan GET_ROWS shader …
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-07 08:37

b10837

<details open=""> caps : recheck typed content if template checks for string (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28511">#28511</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofol…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-07 08:11

b10835

<details open=""> ggml-cuda: fix divergent barrier in f16 flash attention (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27870">#27870</a>) <ul> <li> ggml-cuda: fix divergent barrier in f16 flash attention </li> <li> ggml-cu…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-07 07:20

b10834

<details open=""> ggml: allow backend inputs to not create another split (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28387">#28387</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow"…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-07 06:51

b10833

<details open=""> vulkan: rms_norm fusion opportunities (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28024">#28024</a>) Support RMS_NORM + MUL + ADD (+ MUL) and RMS_NORM + VIEW + SET_ROWS. Extend ROPE + VIEW + SET_ROWS to s…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-07 05:21

b10831

<details open=""> vulkan: add TQ1_0 support (mm, mat-vec, mat-vec-id, dequant, get_rows) (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27765">#27765</a>) <ul> <li> vulkan: add TQ1_0 support (mm, mat-vec, dequant, get_rows) </l…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-06 23:30

b10830

<details open=""> convert : add <code>--fuse-qkv</code> flag to fuse Q/K/V into QKV during HF-to-GGUF conversion (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/22780">#22780</a>) </details> Website: <ul> <li><a…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-06 23:06

b10829

<details open=""> models : fix GDN normalization from <code>max</code> to <code>rsqrt</code> (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28068">#28068</a>) <ul> <li>models: use flash-linear-attention's l2norm for gated delta net q/…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-06 22:11

b10828

<details open=""> [Model] Support for Spark2_5ForCausalLM implementation (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27868">#27868</a>) <ul> <li>Add Spark3 Model</li> <li>rename spark3 -> spark2_5</li> </ul> Co-authored-by: S…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-06 20:52

b10827

<details open=""> opencl: properly choose weights pack for q4_K, q5_K mul_mat (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28402">#28402</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofo…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-06 16:42

b10826

<details open=""> cuda: fixes races in mmid and mmf (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28475">#28475</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.app</a…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-06 13:24

b10825

<details open=""> grammar : fix max repetition threshold (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28469">#28469</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.a…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-06 11:59

b10823

<details open=""> common: add --log-jsonl (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28437">#28437</a>) <ul> <li> common: add --log-jsonl </li> <li> rename unknown to none </li> </ul> </details> Website:</…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-06 11:32

b10822

<details open=""> ui : embed assets directly with CMake (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28445">#28445</a>) Remove the build-time C++ helper and external gzip dependency, simplifying cross-compilation. Keep the …
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-06 10:01

b10821

<details open=""> metal : add remaining fa-vec tunings for M2 Max (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28458">#28458</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https:…
llama.cpp — Releases TIER_1 (SO) · JohannesGaessler · 2026-09-05 20:42

b10820

Github: limit blank issues to maintainers (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28435">#28435</a>)
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-05 10:38

b10819

<details open=""> metal : fix memory leak in early return (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28399">#28399</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-05 09:28

b10817

<details open=""> sycl: attribute device allocations by site (GGML_SYCL_MEMTRACE) (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27631">#27631</a>) define two new environment variables to better understand how much memory is …
llama.cpp — Releases TIER_1 (SO) · philip-jingxin · 2026-09-05 02:37

b10818

sycl : fix test-backend-ops CI break && restore Kronecker product FWH…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-04 20:00

b10816

<details open=""> metal : add remaining fa-vec tunings for M3 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28396">#28396</a>) <ul> <li> addition of m3 in fa_vec_tuned_table </li> <li> adding q4_0,q4_1,q5_0,q5_1 in ggml-met…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-04 19:32

b10814

<details open=""> opencl: extend the elementwise and data‐movement op coverage (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27633">#27633</a>) <ul> <li>opencl: add extended elementwise unary ops (sgn, step, elu, hardswish, hardsigmo…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-04 19:05

b10813

<details open=""> opencl: add Adreno xmem SDPA path (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26331">#26331</a>) <ul> <li>opencl: add Adreno xmem SDPA path</li> </ul> Assisted-by: Codex <ul> <li> Removed the Adreno-spec…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-04 17:57

b10809

<details open=""> llama.cpp : bump version to 0.4.0 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28386">#28386</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.app</a…

本文如何被排名

报道来源 [51]

相关实体

相关话题