PulseAugur
中
实时 07:09:37

llama.cpp 项目发布预发布更新,包含性能优化

llama.cpp 项目发布了多个预发布更新,包括对 MoE expert 的 GPU 缓存管理改进、针对 Apple Silicon 的各种量化类型的优化,以及对 Metal fusion 操作的修复。这些更新还解决了 temperature-zero chains 的采样逻辑,并引入了新的词汇 token。这些发布涵盖了 macOS、iOS、Linux、Android 和 Windows 等广泛平台,并针对 Apple Silicon 和 openEuler 进行了特定优化。 AI

影响 这些更新提高了本地 LLM 在各种硬件和操作系统上推理的性能和兼容性。

排序理由 该集群包含 llama.cpp 项目的多个预发布更新,侧重于软件改进和优化,而不是新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 71 个来源。 我们如何撰写摘要 →

llama.cpp 项目发布预发布更新,包含性能优化

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
Tool
该集群包含 llama.cpp 项目的多个预发布更新,侧重于软件改进和优化,而不是新模型发布或重大行业事件。
Source corroboration
71 independent sources
Strong cross-source corroboration — multiple independent publishers covered this within the clustering window.
Topics
infra, product
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
48 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 [71]

  1. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11499

    <details open=""> <p>cuda : Use byte strides for roll to allow non-contiguous ROLL operations (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29547">#29547</a>)</p> </details> <p><strong>Website:</strong></p> <ul> <li><a href="https://llama.a…

  2. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11498

    <details open=""> <p>CUDA: fix norm family kernels when ne[2]/ne[3] exceed grid dim limits (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28175">#28175</a>)</p> </details> <p><strong>Website:</strong></p> <ul> <li><a href="https://llama.app"…

  3. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11497

    <details open=""> <p>sycl: fuse the delta-net alpha gate (add + unary + mul) (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29687">#29687</a>)</p> <ul> <li> <p>sycl: fuse the delta-net alpha gate (add + unary + mul)</p> </li> <li> <p>tests: …

  4. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11495

    <details open=""> <p>model : support classifier_activation for rerankers (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29692">#29692</a>)</p> <ul> <li>model : support classifier_activation for rerankers</li> </ul> <p>Assisted-by: Claude Opu…

  5. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11494

    <details open=""> <p>ggml-cuda: assign four GDN state columns per warp (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30087">#30087</a>)</p> <ul> <li> <p>ggml-cuda: assign two GDN state columns per warp</p> </li> <li> <p>ggml-cuda: use 4 GDN…

  6. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11493

    <details open=""> <p>sycl: add grouped MoE XMX GEMM (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29245">#29245</a>)</p> <p>Co-authored-by: cwriter <a href="mailto:cwriter@localhost">cwriter@localhost</a></p> </details> <p><strong>Website:<…

  7. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11491

    <details open=""> <p>sycl: remove duplicate block-size defines from op headers (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29507">#29507</a>)</p> <p>All were duplicates of the values in ggml/src/ggml-sycl/presets.hpp</p> </details> <p><st…

  8. llama.cpp — Releases TIER_1 (SO) · cwriter ·

    b11496

    <p>sycl: stage bulk uploads (model loading) through a pinned ring buffer…</p>

  9. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11490

    <details open=""> <p>hexagon: enable alloc_buffer_n (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30126">#30126</a>)</p> <ul> <li> <p>hex-bufs: add support for alloc_buffer_n</p> </li> <li> <p>hex-bufs: add support for splitting large tenso…

  10. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11489

    <details open=""> <p>hexagon: Q6_K weight dequant speedup (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30121">#30121</a>)</p> <ul> <li>hexagon: Q6_K weight dequant speedup</li> </ul> <p>Assisted-by: OpenCode</p> <ul> <li>unroll by another …

  11. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11488

    <details open=""> <p>model : add LiquidAI/d1-omni-600M decision model (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30114">#30114</a>)</p> <ul> <li>model : add LiquidAI/d1-omni-600M decision model</li> </ul> <p>Assisted-by: Claude Opus 5.5<…

  12. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11487

    <details open=""> <p>hexagon: support tiled Q4_K and Q6_K GET_ROWS (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30115">#30115</a>)</p> <ul> <li>hexagon: support tiled Q4_K GET_ROWS</li> </ul> <p>Assisted-by: OpenCode</p> <ul> <li>properly …

  13. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11486

    <details open=""> <p>hexagon: improved GELU accuracy (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30104">#30104</a>)</p> <p>Assisted-by: OpenCode</p> </details> <p><strong>Website:</strong></p> <ul> <li><a href="https://llama.app" rel="nof…

  14. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11485

    <details open=""> <p>chat: fix jinja parser for TranslateGemma (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30096">#30096</a>)</p> <ul> <li> <p>chat: fix jinja parser for translategemma</p> </li> <li> <p>log if missing soruce_lang_code or …

  15. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11484

    <details open=""> <p>chat : name tool and argument parser rules by index (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30088">#30088</a>)</p> <ul> <li> <p>bugfix: infinite recursion caused by a tool named 'call'(<a class="issue-link js-issu…

  16. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11483

    <details open=""> <p>model : add LiquidAI/d1-3B decision model (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30110">#30110</a>)</p> <ul> <li>model : add LiquidAI/d1-3b decision model</li> </ul> <p>mtmd : read LFM2 image resize algo from GGU…

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

    b11482

    <details open=""> <p>cuda: FWHT kernels for block widths above 512 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29100">#29100</a>)</p> <p>The CUDA FWHT covers widths 64 to 512. It runs one row per warp and keeps N/32<br /> values per lane,…

  18. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11481

    <details open=""> <p>mtmd: add cohere2 vision support (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30062">#30062</a>)</p> <ul> <li> <p>cohere2 vision model</p> </li> <li> <p>address comments</p> </li> <li> <p>remove redundant mapping</p> <…

  19. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11480

    <details open=""> <p>llama : add a GPU cache for MoE experts kept in host memory (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29887">#29887</a>)</p> <ul> <li>llama : add a GPU cache for MoE experts kept in host memory</li> </ul> <p>Assiste…

  20. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11476

    <details open=""> <p>metal : few-row MMA mat-mul for the remaining src0 types (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30065">#30065</a>)</p> <p>The generic few-row MMA kernel works for any type with a 16-weight<br /> dequantizer, so i…

  21. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11475

    <details open=""> <p>metal : fix MUL_MAT+ADD fusion when the residual is itself a MUL_MAT (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30100">#30100</a>)</p> <ul> <li>metal : fix MUL_MAT+ADD fusion when the residual is itself a MUL_MAT</li…

  22. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11472

    <details open=""> <p>sampling : use greedy selection for eligible temperature-zero chains (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29797">#29797</a>)</p> <ul> <li>sampling : use greedy selection for eligible temperature-zero chains</li…

  23. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11471

    <details open=""> <p>vocab : add plamo fim tokens (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30090">#30090</a>)</p> </details> <p><strong>Website:</strong></p> <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.app</a></li…

  24. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11469

    <details open=""> <p>musa: use the tile lightning indexer kernel (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30080">#30080</a>)</p> <p>Signed-off-by: Xiaodong Ye <a href="mailto:[email protected]">[email protected]</a></p> <…

  25. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11468

    <details open=""> <p>vendor : update cpp-httplib to 0.60.0 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30081">#30081</a>)</p> </details> <p><strong>Website:</strong></p> <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.ap…

  26. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11467

    <details open=""> <p>imatrix : include clocale for std::setlocale (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30079">#30079</a>)</p> <p>Co-authored-by: yarik <a href="mailto:[email protected]">[email protected]</a></p> </details> <p><st…

  27. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11466

    <details open=""> <p>ggml-webgpu: fix flash_attn supports_op check for overlapping KV (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28205">#28205</a>)</p> </details> <p><strong>Website:</strong></p> <ul> <li><a href="https://llama.app" rel=…

  28. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11465

    <details open=""> <p>tests : retain the anchor when testing recurrent rollback (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29923">#29923</a>)</p> </details> <p><strong>Website:</strong></p> <ul> <li><a href="https://llama.app" rel="nofoll…

  29. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11464

    <details open=""> <p>[SYCL] fix the issue in mixed different model GPUs in FA (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29071">#29071</a>)</p> <ul> <li> <p>fix mixed different model GPUs issue</p> </li> <li> <p>Update ggml/src/ggml-sycl…

  30. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11463

    <details open=""> <p>sycl: accelerate GLM MLA prefill with MKL flash attention (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29171">#29171</a>)</p> <ul> <li>sycl: accelerate GLM MLA prefill with MKL flash attention</li> </ul> <p>GLM-4.7 Fla…

  31. llama.cpp — Releases TIER_1 (SO) · Pratyush-gg ·

    b11462

    <p>sycl : fattn_kv_buffers cleanup (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27689">#27689</a>)</p>

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

    b11461

    <details open=""> <p>vulkan: fix amd iGPU slow checkpoint read (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30049">#30049</a>)</p> </details> <p><strong>Website:</strong></p> <ul> <li><a href="https://llama.app" rel="nofollow">https://llam…

  33. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11460

    <details open=""> <p>sycl: add IQ3_S multi-column MMVQ (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29500">#29500</a>)</p> </details> <p><strong>Website:</strong></p> <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.app</a…

  34. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11459

    <details open=""> <p>ggml-webgpu: no dawn native features on wasi (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27069">#27069</a>)</p> </details> <p><strong>Website:</strong></p> <ul> <li><a href="https://llama.app" rel="nofollow">https://l…

  35. llama.cpp — Releases TIER_1 (SO) · max-krasnyansky ·

    b11458

    <p>hexagon: CPY/CONCAT/CONT/DUP overhaul to use DMA/HVX for all cases (#…</p>

  36. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11457

    <details open=""> <p>cuda : add BF16 support for XIELU (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29955">#29955</a>)</p> <p>The XIELU CUDA kernel template is already generic over the element<br /> type; only the F32/F16 type assertion an…

  37. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11456

    <details open=""> <p>ggml-cuda: use per-thread stream for buffer-init padding memset (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28782">#28782</a>)</p> <ul> <li> <p>ggml-cuda: use per-thread stream for buffer-init padding memset</p> </li>…

  38. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11455

    <details open=""> <p>vocab : implement PLaMo-3 tokenizer pre-segmentation (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30045">#30045</a>)</p> <ul> <li>vocab : implement PLaMo-3 tokenizer pre-segmentation</li> </ul> <p>The PLaMo-3 tokenizer…

  39. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11451

    <details open=""> <p>opencl: fix OOB read in adreno xmem GEMM (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30041">#30041</a>)</p> </details> <p><strong>Website:</strong></p> <ul> <li><a href="https://llama.app" rel="nofollow">https://llama…

  40. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11450

    <details open=""> <p>RPC: add <code>-sm tensor</code> (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26610">#26610</a>)</p> <ul> <li> <p>rpc: allow -sm tensor</p> </li> <li> <p>fix flush for apple rdma</p> </li> <li> <p>move graph_uids to rp…

  41. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11449

    <details open=""> <p>ggml: fix CLAMP on non-contiguous views (CPU, CUDA) (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29517">#29517</a>)</p> <ul> <li>ggml: fix CLAMP on non-contiguous views (CPU, CUDA)</li> </ul> <p>CUDA clamped ggml_nelem…

  42. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11448

    <details open=""> <p>cuda: BF16/FP16 conversion to f32 chunking (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29442">#29442</a>)</p> <ul> <li> <p>ggml-cuda: chunk large BF16/FP16 to F32 conversions</p> </li> <li> <p>Update ggml/src/ggml-cud…

  43. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11447

    <details open=""> <p>models: support pplx-decider (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30044">#30044</a>)</p> </details> <p><strong>Website:</strong></p> <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.app</a></li…

  44. llama.cpp — Releases TIER_1 (SO) · ngxson ·

    b11452

    <p>model: support embeddinggemma2 (text+vision+audio) (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30054">#30054</a>)</p>

  45. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11445

    <details open=""> <p>vulkan : check for null vkEnumerateInstanceVersion (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29872">#29872</a>)</p> <p>A 1.0 loader (e.g. Android 8.1) has no vkEnumerateInstanceVersion, so<br /> backend init called …

  46. llama.cpp — Releases TIER_1 (SO) · masterFoad ·

    b11446

    <p>metal : fix excess threadgroup memory in quantized flash attention (#…</p>

  47. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11443

    <details open=""> <p>models : consolidate nextn row cropping into shared helpers (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30017">#30017</a>)</p> <ul> <li>mimo2 : always emit h_nextn</li> </ul> <p>the other nextn-capable models set it u…

  48. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11440

    <details open=""> <p>llama : re-reserve the sched when the nextn extraction flags change (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30020">#30020</a>)</p> <p>The speculative MTP init enables NextN extraction on the target and draft<br />…

  49. llama.cpp — Releases TIER_1 (SO) · JohannesGaessler ·

    b11444

    <p>HIP: use -O0 for host code in debug builds (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29795">#29795</a>)</p>

  50. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11438

    <details open=""> <p>test-llama-archs : initialize backends before generating models (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30034">#30034</a>)</p> <p>With GGML_BACKEND_DL=ON, backends must be loaded explicitly before<br /> creating m…

  51. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11435

    <details open=""> <p>llama: fix k-pool scatter data race on shared sequences (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29994">#29994</a>)</p> <ul> <li>llama: re-pool each shared k-pool rep once</li> </ul> <p>With shared cells every pool…

  52. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11433

    <details open=""> <p>hexagon: add pool op support (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29995">#29995</a>)</p> <ul> <li> <p>hexagon: add pool_2d support</p> </li> <li> <p>hexagon: add pool_1d support</p> </li> <li> <p>hex-pool: dma …

  53. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11425

    <details open=""> <p>CUDA: make the alloc_deps check batch independent (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29986">#29986</a>)</p> <p>Fixes <a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/issues/2998…

  54. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11424

    <details open=""> <p>vulkan: fix Flash Attention shmem write out of bounds (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29988">#29988</a>)</p> </details> <p><strong>Website:</strong></p> <ul> <li><a href="https://llama.app" rel="nofollow">…

  55. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11418

    <details open=""> <p>server: support vision input for Clef (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29969">#29969</a>)</p> <ul> <li> <p>server: support vision input for Clef</p> </li> <li> <p>move input_attn_causal to private</p> </li>…

  56. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11417

    <details open=""> <p>CUDA: Optimize accumulation in mmq for NVFP4 type (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29857">#29857</a>)</p> <ul> <li> <p>ggml_cuda: optimize accumulation in mmq_vec_dot_fp4_fp4_mma for better performance</p> …

  57. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11415

    <details open=""> <p>server: reject partial media truncation (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24076">#24076</a>)</p> <ul> <li> <p>server: reject partial media truncation</p> </li> <li> <p>server: keep only the keep_first fix</p…

  58. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11413

    <details open=""> <p>vulkan: sparse flash attention for quantized K/V (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29639">#29639</a>)</p> <ul> <li>vulkan: sparse flash attention for quantized K/V</li> </ul> <p>Assisted-by: Claude</p> <ul> …

  59. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11412

    <details open=""> <p>llama : fix unexpected graph reallocation in the k-pool models (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29958">#29958</a>)</p> <ul> <li>llama : fix unexpected graph reallocation in the k-pool models</li> </ul> <p>B…

  60. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11411

    <details open=""> <p>kv-cache: fix restoring mismatched KV cache rotation by saving exact rotation metadata (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28498">#28498</a>)</p> <ul> <li> <p>kv-cache: save exact KV rotation metadata, reject …

  61. llama.cpp — Releases TIER_1 (SO) · fxgsell ·

    b11414

    <p>vulkan: fix stale prealloc_y reuse across flash attention and soft_ma…</p>

  62. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11408

    <details open=""> <p>ggml-cpu : add Q8_0 IME1 matrix kernel for SpacemiT X60 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28479">#28479</a>)</p> <ul> <li>ggml-cpu : add Q8_0 IME1 matrix kernel for SpacemiT X60</li> </ul> <p>On the SpacemiT…

  63. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11407

    <details open=""> <p>vulkan : Fix undeclared identifiers when -DGGML_VULKAN_RUN_TESTS=ON (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29912">#29912</a>)</p> </details> <p><strong>Website:</strong></p> <ul> <li><a href="https://llama.app" r…

  64. llama.cpp — Releases TIER_1 (SO) · CISC ·

    b11410

    <p>ci : winget urls must be separate strings (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29978">#29978</a>)</p>

  65. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11405

    <details open=""> <p>cuda: tile the lightning indexer over keys and tokens for 4 heads (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29901">#29901</a>)</p> <ul> <li>cuda: tile the lightning indexer over keys and tokens for 4 heads</li> </ul…

  66. llama.cpp — Releases TIER_1 (SO) · yomaytk ·

    b11406

    <p>webgpu: add MMVQ support for Q1_0/Q5_0/Q5_1/Q3_K/Q5_K/Q6_K/MXFP4 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/issues/29">#29</a>…</p>

  67. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11404

    <details open=""> <p>metal : few-row MMA mat-mul (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29869">#29869</a>)</p> <ul> <li>metal : few-row MMA mat-mul and batched copies for speculative decoding</li> </ul> <p>Speculative decoding verifi…

  68. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11403

    <details open=""> <p>CUDA: use MMVF for thin f16/bf16 mul_mat at small batch size (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29633">#29633</a>)</p> <ul> <li>CUDA: use MMVF for thin f16/bf16 mul_mat at small batch size</li> </ul> <p>Signe…

  69. llama.cpp — Releases TIER_1 (SO) · anujj ·

    b11402

    <p>CUDA: prefer whole-tile FlashAttention scheduling for efficient two-s…</p>

  70. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11401

    <details open=""> <p>log, server: self contained colors, split child commands from logs in router mode (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29895">#29895</a>)</p> <ul> <li>log, server: make router child lines carry their own colors…

  71. llama.cpp — Releases TIER_1 (SO) · github-actions[bot] ·

    b11400

    <details open=""> <p>llama: support both embd + raw tokens in batch (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29622">#29622</a>)</p> <ul> <li> <p>llama: support both embd + raw tokens in batch</p> </li> <li> <p>add to test-llama-archs</…