llama.cpp releases multiple updates with performance and build improvements

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-04 15:05

b10798

<details open=""> common : make build info output stream configurable (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28322">#28322</a>) Let llama_print_build_info write to a caller-provided FILE* instead of hardcoding stderr.…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-04 08:38

b10797

<details open=""> ggml-cpu(s390x) : fix q5_1 uninitialized v_acc (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28332">#28332</a>) Signed-off-by: Aaron Teo <a href="mailto:[email protected]">[email protected]</a> </details> <…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-04 05:33

b10796

<details open=""> src : add n_expert_used_max function (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28323">#28323</a>) <ul> <li>src : add n_expert_used_max function</li> </ul> With Commit <a class="commit-link" href="https://gith…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-04 04:32

b10795

<details open=""> sycl: fuse rms_norm+mul+add and add+add residual chains (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27610">#27610</a>) Fuse RMS_NORM+MUL+ADD and ADD+ADD under GGML_SYCL_ENABLE_FUSION. ADD+ADD uses the sa…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-04 03:22

b10794

<details open=""> SYCL: Refactor GGML_SYCL_ENABLE_MKL_FA to global var (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26863">#26863</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">h…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-03 22:19

b10793

<details open=""> llama: fix whole source code rebuilt on each new commit (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28278">#28278</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-03 21:12

b10792

<details open=""> common/json-schema : fix GBNF grammar generation for empty object schemas (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28279">#28279</a>) </details> Website: <ul> <li><a href="https://llama.…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-03 20:24

b10791

<details open=""> opencl: quant lm_head / decode GEMV and medium-batch GEMM optimizations (speculative decoding/MTP) (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26477">#26477</a>) <ul> <li> opencl: quant lm_head / decode GEMV an…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-03 19:54

b10790

<details open=""> tune MMVQ to MMQ crossover for SM87 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28285">#28285</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.app<…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-03 12:45

b10786

<details open=""> mtmd: propagate const to preproc class (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28310">#28310</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.a…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-03 12:21

b10785

<details open=""> metal : add sparse FA (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28098">#28098</a>) <ul> <li>metal : support n_kv_max sparse mask hint in flash attention vec kernel</li> </ul> <ul> <li>add kernel_flash_attn_ext_v…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-03 11:54

b10784

<details open=""> metal : fix glu dispatch with ne00 = 1 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28306">#28306</a>) <ul> <li> metal : fix glu dispatch with ne00 = 1 </li> <li> tests : disable ill-defined tests </l…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-03 11:18

b10783

<details open=""> mtmd : add <code>const</code> in various places (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28307">#28307</a>) <ul> <li>mtmd : mark context as const in more methods</li> </ul> Mark <code>mtmd_context</code> as …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-03 10:28

b10782

<details open=""> CUDA: Allow concurrent streams per split for multi-GPU (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28198">#28198</a>) <ul> <li>CUDA: Allow CUDA optimization per split for multi-GPU.</li> </ul> Previous guard ca…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-03 09:59

b10781

<details open=""> vulkan: fix FA dequant path engagement (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28190">#28190</a>) Skip the nb[3] check when ne[3] == 1, the shader never reads it for a single stream. Cache views carry…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-03 09:35

b10780

<details open=""> sycl : enhance the api to support peer-to-peer copy (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27550">#27550</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">ht…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-03 08:59

b10778

<details open=""> misc : prevent RAM peaking at model loading stage (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27483">#27483</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">http…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-03 08:29

b10777

<details open=""> sycl: reduce redundant work in Q4_K multi-column MMVQ (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27062">#27062</a>) <ul> <li> sycl: Q4_K Weight unpack optimization and reuse between destination Columns </l…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-03 07:59

b10776

<details open=""> model: add NVIDIA Nemotron-3-Puzzle-75B-A9B (NemotronHPuzzle) support (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25444">#25444</a>) <ul> <li>hparams: add per-layer n_ff_exp/n_expert_used arrays with scalar-or-arr…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-03 02:18

b10775

<details open=""> mtmd: fix idefics3 preproc (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28273">#28273</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] · 2026-09-03 01:54

b10774

<details open=""> finetune: fix no KV cache (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27199">#27199</a>) <ul> <li> training: fix no KV cache </li> <li> apply @ ggerganov suggestion </li> </ul> </details> <s…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-03 01:27

b10773

<details open=""> server : accept data: URLs for input_video and input_audio (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27735">#27735</a>) <ul> <li>server : accept data: URLs for input_video and input_audio</li> </ul> input_vid…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-03 00:50

b10772

<details open=""> ggml-hexagon: add F16 support for unary ops (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28228">#28228</a>) Extend the HTP backend's F16 unary op coverage to include ABS on top of the existing NORM/RMS_NOR…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-03 00:26

b10771

<details open=""> mtmd: add mtmd_tokenize_from_parts() (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28250">#28250</a>) <ul> <li> add mtmd_tokenize_from_parts </li> <li> use it in mtmd-cli </li> <li> move add_special…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-03 00:02

b10770

<details open=""> metal : add fa-vec tunings for M3 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28236">#28236</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] · 2026-09-02 23:39

b10769

<details open=""> metal : fix memory query under low-memory conditions (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27701">#27701</a>) <ul> <li> metal: Fix memory query under low-memory conditions </li> <li> Simply variabl…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-02 23:00

b10767

<details open=""> Update ROCm to 10.0.0 release (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27803">#27803</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.app</a></l…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-02 20:35

b10766

<details open=""> model: correctly support input vision for deepseek4 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28154">#28154</a>) <ul> <li> model: correctly support input vision for deepseek4 </li> <li> nits </li> …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-02 19:59

b10764

<details open=""> ggml-cuda : remove unused vars (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28235">#28235</a>) Signed-off-by: Adrien Gallouët <a href="mailto:[email protected]">[email protected]</a> </details> <stron…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-02 19:04

b10763

<details open=""> common, server : enable preserve_reasoning kwarg by default, log its effective state (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28174">#28174</a>) <ul> <li>common, server : enable preserve_reasoning kwarg by defa…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-02 18:39

b10762

<details open=""> mtmd: support DeepSeek-V4-Flash-Vision-Exp (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28133">#28133</a>) <ul> <li> mtmd: support DeepSeek-V4-Flash-Vision-Exp </li> <li> handle min/max token counts from …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-02 11:16

b10760

<details open=""> mtmd: Fix Qwen3-tts-0.6b (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28231">#28231</a>) <ul> <li>mtmd: load the qwen3-tts code predictor proj_in as optional</li> </ul> The talker and the code predictor share th…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-02 10:23

b10759

<details open=""> ggml: avoid KleidiAI buffer type init on dispatch (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27891">#27891</a>) Co-authored-by: Acmmi <a href="mailto:[email protected]">[email protected]…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-02 09:31

b10758

<details open=""> hexagon: MUL_MAT and MUL_MAT_ID fusion and fixes (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28202">#28202</a>) <ul> <li> hex-mm: fuse QKV and FFN matmuls that land on HMX </li> <li> hex-mm: remove hardc…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-02 08:44

b10757

<details open=""> vulkan: handle larger batch sizes (>4) efficiently for IQ3_S mat-vec (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27449">#27449</a>) <ul> <li>vulkan: handle larger batch sizes (>4) efficiently for IQ3_S mat-v…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-02 08:19

b10756

<details open=""> vulkan : only request VK_KHR_shader_bfloat16 extension if supported (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28155">#28155</a>) </details> Website: <ul> <li><a href="https://llama.app" r…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-02 06:28

b10754

<details open=""> opencl: fix out‐of‐bound reads in the Adreno image kernels (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27632">#27632</a>) <ul> <li> opencl: clamp the q4_K decode GEMV's fetch row on a padded x-grid </li> <l…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-02 06:03

b10753

<details open=""> hexagon: add missing FARF logs for cpy/get_rows/set_rows/gdn ops (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28217">#28217</a>) <ul> <li> hexagon: fix bug ne[2] printed in proc_op_req prep-src log </li> <li…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-02 00:15

b10752

<details open=""> metal : add metallib build support for xcframework (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28163">#28163</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">htt…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-01 20:16

b10751

<details open=""> cuda: fuse MoE weighted expert reduction (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25952">#25952</a>) <ul> <li>cuda : fuse MoE weighted reduction (mul + view + add)</li> </ul> The MoE combine tail currently w…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-01 19:23

b10750

<details open=""> kv-cells: look up the n-gram history in the sequence position index (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28040">#28040</a>) get_prev_tokens() rebuilt a (seq, pos) -> token hash map on every ubat…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-01 18:59

b10749

<details open=""> context : autoscale n_ctx_train when yarn scaling specified (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28030">#28030</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofo…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-01 18:20

b10743

<details open=""> metal : add fa-vec tuning for M2 Pro (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28122">#28122</a>) <ul> <li> metal: add fa-vec tuning for M2 Pro </li> <li> metal : update fa-vec tuning for M2 Pro with n…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-01 17:06

b10742

<details open=""> metal : add fa-vec tunings for A18 Pro (MacBook Neo) (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28152">#28152</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">h…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-01 16:37

b10741

<details open=""> model : load hparams.n_layer_nextn before n_layer() calls (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28159">#28159</a>) <ul> <li> load hparams.n_layer_nextn before n_layer() calls </li> <li> remove dupl…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-01 15:43

b10740

<details open=""> metal : fix more leaks due to missing autoreleasepools (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27883">#27883</a>) <ul> <li> metal : fix more leaks due to missing autoreleasepools </li> <li> metal : r…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-01 15:17

b10739

<details open=""> metal : add fa-vec tuning for M2 Max (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28015">#28015</a>) Rows for M2 Max (30 GPU cores) collected with 'ggml-metal-tuning fa-vec --dtype f16,q8_0', pasted into f…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-01 14:41

b10738

<details open=""> sycl : support limit max alloc memory within 2GB for host-pinned memory (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27559">#27559</a>) </details> Website: <ul> <li><a href="https://llama.ap…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-01 14:05

b10737

<details open=""> qwen4exp: fix seq_cp, block position keying, mtmd input, cuda abort, add tests (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27941">#27941</a>) <ul> <li> qwen4exp: follow up fixes </li> <li> -kvu NaN colla…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-01 13:32

b10736

<details open=""> tests : fix log verbosity for test-llama-archs (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28147">#28147</a>) <ul> <li> tests : fix log verbosity for test-llama-archs </li> <li> cont : naming </li> <…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-01 11:59

b10734

<details open=""> metal: enable Metal 4.0 tensor API on M5+/A19+ (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27461">#27461</a>) <ul> <li> metal : request Metal 4.0 language version for the tensor API </li> <li> metal : lo…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-01 10:02

b10733

<details open=""> ci: Bump ggml-org/ccache-action to v1.2.24 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28083">#28083</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://lla…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-01 05:36

b10731

<details open=""> qwen4exp: support recurrent state rollback (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28123">#28123</a>) MTP speculative decoding needs the target state to move back by the number of rejected draft token…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-09-01 05:10

b10730

<details open=""> qwen4exp: sum the indexer heads by slices (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28023">#28023</a>) <ul> <li>qwen4exp: sum the indexer heads by slices</li> </ul> The head reduction went through a transpose…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-31 22:24

b10729

<details open=""> metal : add fa-vec tunings for M1 Ultra (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28088">#28088</a>) <ul> <li> metal : add fa-vec tunings for M1 Ultra </li> <li> metal : move M1 Ultra tunings after M1 …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-31 21:48

b10728

<details open=""> CUDA: XOR swizzle flash attn K,V smem fp16 tiles (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25635">#25635</a>) <ul> <li>CUDA: XOR swizzle flash attn K,V smem fp16 tiles</li> </ul> Signed-off-by: ynankani <a hr…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-31 20:56

b10727

<details open=""> metal : add concat support for quantized types (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28116">#28116</a>) Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-0731 </details> Website: <ul…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-31 19:43

b10726

<details open=""> AVX2: Speed up large batch size prompt processing of IQ models (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27402">#27402</a>) <ul> <li>Batched gemm for grid IQ quants</li> </ul> Style updates and a bit more per…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-31 17:29

b10724

<details open=""> kv-cache : optimize restoring non-contiguous cells (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27991">#27991</a>) <ul> <li>kv cache : batch state restore scatter reads per contiguous run</li> </ul> When restori…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-31 16:33

b10723

<details open=""> opencl: tune the quant paths for Intel Xe-LP GPUs to improve its TG and PP performance (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26438">#26438</a>) <ul> <li> opencl: Q4_K/Q5_K mul_mv N_DST 4->8 on Intel fo…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-31 15:56

b10721

<details open=""> webgpu : avoid crash when offset is not multiple of 4 in WebGPU ggml_backend_tensor_get() implementation (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28045">#28045</a>) <ul> <li> webgpu : avoid crash when offset…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-31 15:30

b10720

<details open=""> ROCm: add radix TOP_K for long rows (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27466">#27466</a>) <ul> <li>ROCm: add radix TOP_K for long rows</li> </ul> </details> Website: <ul> <li><a hr…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-31 15:06

b10719

<details open=""> metal : add fa-vec tunings for M1 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28078">#28078</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] · 2026-08-31 14:24

b10718

<details open=""> CUDA: extend MOE fusion to specdec, earlier MOE glu fusion and topk-router fusion were restricted to 1 token (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27621">#27621</a>) <ul> <li>CUDA: extend MOE fusion to specd…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-31 11:21

b10717

<details open=""> sycl : Enhance to get the free memory of Intel GPU (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27968">#27968</a>) <ul> <li> enhance get mem info by l0 an SYCL API </li> <li> remove debug code, format the…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-31 10:13

b10715

<details open=""> spec : fuse the DFlash encoder into the KV cache injection (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27310">#27310</a>) <ul> <li>dflash : fuse the encoder into the KV injection decode</li> </ul> The encoder i…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-31 09:41

b10714

<details open=""> vulkan: tune mat-vec rows for batched inference on Strix Halo (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27909">#27909</a>) <ul> <li>vulkan: RDNA3 static mat-vec rows above four columns</li> </ul> On RDNA3 abo…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-31 08:51

b10713

<details open=""> ggml : add MUL_MAT to the list of ops that may need additional memory (for WebGPU) (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28071">#28071</a>) Co-authored-by: Stanisław Szymczyk <a href="mailto:sszymczy@gmai…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-31 05:41

b10712

<details open=""> vulkan: top_k radix select for k >= 1024 for Qwen 3.8 Flash Next (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28032">#28032</a>) <ul> <li> vulkan: add top-k radix sort shader for k >= 1024 </li> <li> <…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-31 04:47

b10711

<details open=""> hexagon: fix CPY fence bug (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28033">#28033</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] · 2026-08-31 04:20

b10710

<details open=""> metal : add remaining Q4_1/Q5_0/Q5_1 fa-vec tunings for M2 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28017">#28017</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofol…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-31 03:50

b10709

<details open=""> rpc: avoid serializing buffers from other servers (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26500">#26500</a>) <ul> <li>rpc: avoid serializing buffers from other servers</li> </ul> Only include remote buffer …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-31 03:27

b10708

<details open=""> ggml : fix ggml_backend_buft_get_alloc_size() guard (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28038">#28038</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">ht…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-31 03:04

b10707

<details open=""> kv-cells: stop the sequence scan once all sequences are seen (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28011">#28011</a>) for_each_token_in tested all LLAMA_MAX_SEQ sequences for every used cell, while …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-31 02:40

b10706

<details open=""> ggml: add SWIGLU_CLAMP (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27930">#27930</a>) <ul> <li> ggml: add SWIGLU_CLAMP </li> <li> add vulkan shader </li> </ul> </details> Website:…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-30 23:17

b10705

<details open=""> llama: improve TENSOR_READ_LAZY handling (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27837">#27837</a>) <ul> <li> force lazy tensor on cpu if lazy is on </li> <li> llama: improve TENSOR_READ_LAZY handlin…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-30 22:55

b10704

<details open=""> CUDA: use the fast mm_ids_helper path for any n_expert_used (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27978">#27978</a>) The optimized path grouped warp lanes by token and required warp_size % n_expert_…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-30 22:33

b10703

<details open=""> hip: tune rdna 3 mmq config (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26284">#26284</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] · 2026-08-30 21:21

b10702

<details open=""> hip : optimize Q2_0 dot-product path for gfx1201 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26753">#26753</a>) <ul> <li> hip/gfx1201: optimize q2_0 vec_dot_q2_0_q8_1 with native amdgcn perm </li> <li> B…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-30 18:58

b10701

<details open=""> dflash: pass missing NVFP4 scales to attention operations (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/28000">#28000</a>) <ul> <li>DFlash2 NVFP4 draft models produced almost no accepted speculative tokens bec…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-30 18:35

b10700

<details open=""> common: rename --tensor-read-lazy to --lazy-mode, add -lzm shorthand (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27969">#27969</a>) Rename the --tensor-read-lazy CLI argument to --lazy-mode, to match the …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-30 18:09

b10699

<details open=""> ggml : add ggml_backend_op_alloc_size_may_expand, use it in RPC (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27960">#27960</a>) some backends (Metal, SYCL, WebGPU) require additional memory for fleeting da…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-30 14:49

b10698

<details open=""> rpc: fix apple rdma error spew on teardown (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27908">#27908</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://lla…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-30 14:26

b10697

<details open=""> metal: add fa-vec tunings for M3 Ultra (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27999">#27999</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.a…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-30 14:01

b10696

<details open=""> metal : Add fa-vec tuning for M3 Pro (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27963">#27963</a>) Related issue: #27668 </details> Website: <ul> <li><a href="https://llama.app" rel…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-30 13:32

b10695

<details open=""> vendor : update cpp-httplib to 0.54.0 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27919">#27919</a>) <ul> <li> vendor : update cpp-httplib to 0.54.0 </li> <li> vendor : update cpp-httplib to 0.54.0 and 0…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-30 13:08

b10694

<details open=""> rpc : fix pre-rdma macOS versions (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27815">#27815</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] · 2026-08-30 12:45

b10693

<details open=""> hexagon: support for device discovery and create sessions on demand (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27785">#27785</a>) <ul> <li>hex-devices: add support for lazy session allocation and cleanup dev inte…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-30 12:23

b10692

<details open=""> sycl: split long rows in TOP_K instead of one work-group per row (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27847">#27847</a>) </details> Website: <ul> <li><a href="https://llama.app" rel=…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-30 12:00

b10691

<details open=""> metal : fix null-pipeline crash for F16 src1 mul_mat/mul_mat_id (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25648">#25648</a>) <ul> <li> metal : fail closed on mul_mat shapes with missing F16 kernels </li> …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-30 07:46

b10690

<details open=""> memory : copy Hadamard matrix to k_rot tensor only if it has buffer assigned to prevent crashes during context shift of unquantized K cache (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27967">#27967</a>) Co-auth…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-30 04:04

b10689

<details open=""> ggml: allow passing alloc dependencies in graph_optimize (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27301">#27301</a>) <ul> <li> ggml: allow passing alloc dependencies in graph_optimize </li> <li> add a…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-30 00:13

b10688

<details open=""> metal : add fa-vec tunings for M2 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27940">#27940</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] · 2026-08-29 18:24

b10687

<details open=""> opencl: use a better matmul path on two Adreno GPU generations (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27640">#27640</a>) <ul> <li>opencl: default the Adreno xmem F16xF32 GEMM on for X2E</li> </ul> kernel_m…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-29 16:47

b10686

<details open=""> metal : assert shared memory padding (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27951">#27951</a>) <ul> <li> metal : assert shared memory padding </li> <li> cont : add ref </li> </ul> </details> …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-29 16:25

b10685

<details open=""> metal : add remaining fa-vec tunings for M4 Pro (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27915">#27915</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https:…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-29 15:59

b10684

<details open=""> sycl: make --fit respect --fit-target better (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27629">#27629</a>) improve the --fit algorithm to take into account the actual peak required VRAM for a given conte…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-29 15:26

b10683

<details open=""> vulkan: combine duplicated fastdiv functions, rename the one optimizing small divs (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27526">#27526</a>) <ul> <li> vulkan: combine duplicated fastdiv functions, rename t…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-29 15:02

b10682

<details open=""> metal : add fa-vec tunings for M1 Max (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27932">#27932</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.ap…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-29 14:36

b10681

<details open=""> vulkan: Change mul_mat_id to pad K rather than N (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27925">#27925</a>) The N padding is needed for mul_mat, but not mul_mat_id. For mul_mat_id, we indirect the row…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-28 19:24

b10679

<details open=""> bench: add --tensor-read-lazy (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27881">#27881</a>) <ul> <li> bench: add --tensor-read-lazy </li> <li> rm the alias </li> <li> rename to LLAMA_LAZY_MODE_*<…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-28 18:53

b10678

<details open=""> model: qwen4exp: reduce number of graph splits (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27880">#27880</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https:/…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-28 18:29

b10677

<details open=""> vulkan: fix missing view-alias dependencies in ggml_vk_graph_optimize (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27812">#27812</a>) <ul> <li>vulkan: fix missing view-alias dependencies in ggml_vk_graph_optimize</…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-28 18:03

b10676

<details open=""> ggml : fix conv_transpose_2d for multiple batches (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26132">#26132</a>) <ul> <li>ggml : fix conv_transpose_2d for multiple batches</li> </ul> ggml_compute_forward_conv_t…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-28 15:50

b10675

<details open=""> Vulkan: add hoisting support for row IDs and expert count in shaders (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26686">#26686</a>) <ul> <li> vulkan: add hoisting support for row IDs and expert count in shaders…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-28 14:32

b10673

<details open=""> metal : add fa-vec tunings for M4 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27875">#27875</a>) This adds fa_vec_tuned_table records for Apple M4 to ggml-metal-tuning.cpp. Includes F16, Q4_0, Q4_1, Q5_0…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-28 13:53

b10672

<details open=""> OpenVINO: Update OV to 2026.3.1, whisper.cpp support, Qwen3.5 on NPU, and new ops (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27843">#27843</a>) <ul> <li> OpenVINO Backend: Fuse IM2COL + MatMul convolution into…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-28 13:14

b10670

<details open=""> sycl: use TILE for quantized KV decode on BMG (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26689">#26689</a>) Route quantized KV decode to TILE on Xe2 (BMG) only, keep VEC on other archs until validated there.</…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-28 12:45

b10669

<details open=""> sycl: bind the f16 KV cache in place for the oneDNN SDPA path (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27468">#27468</a>) Measured at a live KV length of 34816 (32768 depth plus one 2048 ubatch), on Qw…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-28 12:19

b10668

<details open=""> metal : add fa-vec tunings for M3 Max, M5 and M5 Pro (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27863">#27863</a>) <ul> <li>metal : add fa-vec tunings for M5</li> </ul> This is a followup contribution to <a cl…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-28 11:55

b10667

<details open=""> metal : add fa-vec tunings for M4 Pro (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27824">#27824</a>) This is a followup contribution to <a class="commit-link" href="https://github.com/ggml-org/llama.cpp/commit/…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-28 00:19

b10665

<details open=""> model: add DSpark support for Nemotron3.5 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27804">#27804</a>) <ul> <li> model: add DSpark support for Nemotron3.5 </li> <li> Update src/models/dflash.cpp </…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-27 22:49

b10664

<details open=""> ggml-hexagon: add HTP unary ops for ABS and LOG (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27786">#27786</a>) Add HVX-accelerated implementations for GGML_OP_LOG and GGML_UNARY_OP_ABS on the HTP backend.…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-27 22:25

b10663

<details open=""> hex-unary: fix RMS_NORM_MUL weight-offset bugs for grouped/broadcast norms (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27798">#27798</a>) </details> Website: <ul> <li><a href="https://llama…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-27 21:09

b10662

<details open=""> server: add ctx-per-slot (--kv-unified-per-slot) (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24124">#24124</a>) <ul> <li> Add ctx-per-slot argument for unifid KV cache </li> <li> Swap out ctx fractions f…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-27 20:39

b10661

<details open=""> ci : build only the ggml-hip backend for windows-rocm release (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27753">#27753</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="no…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-27 20:05

b10660

<details open=""> model: add Qwen3.8-Flash-Next (qwen4exp) (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27742">#27742</a>) <ul> <li>gguf: add qwen4exp (Qwen3.8-Flash-Next) arch and converter</li> </ul> Adds the GGUF-side plumbing…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-27 19:09

b10659

<details open=""> ci : bundle HIP runtime DLLs with Windows ROCm release (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26973">#26973</a>) Copy amdhip64_7, amd_comgr and rocm_kpack next to the binaries so the correct HIP runt…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-27 18:34

b10658

<details open=""> spec : add DFlash2 support (local convolution + candidate selector) (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27342">#27342</a>) (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-27 18:08

b10657

<details open=""> opencl: add bin kernels <code>kernel_gemm_moe_q4_0_q8_1_dp4a_bin</code>, <code>kernel_gemm_moe_mxfp4_q8_1_dp4a_bin</code> (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27768">#27768</a>) </details> Websit…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-27 17:42

b10656

<details open=""> quantize: cap working memory size to avoid loading big tensors onto RAM (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27795">#27795</a>) </details> Website: <ul> <li><a href="https://llama.ap…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-27 14:20

b10655

<details open=""> Feature: Added LIGHTNING_INDEXER support for Deepseek V4 ops on Vulkan Backend (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27453">#27453</a>) <ul> <li> vulkan: add LIGHTNING_INDEXER op </li> <li> vulkan:…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-27 13:50

b10653

<details open=""> llama: model_loader: add TENSOR_READ_LAZY (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27794">#27794</a>) <ul> <li> llama: model_loader: add TENSOR_GET_ROW_LAZY </li> <li> add --tensor-read-lazy </li>…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-27 12:30

b10649

<details open=""> spec: Add benchmark-only synthetic speculative acceptance options (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27711">#27711</a>) <ul> <li> Add benchmark-only synthetic speculative acceptance to llama-server and…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-27 12:01

b10648

<details open=""> model : simplify MiniMax-01 graph (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27790">#27790</a>) Co-authored-by: Stanisław Szymczyk <a href="mailto:[email protected]">[email protected]</a> </details> <…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-27 11:27

b10647

<details open=""> args: add --video-* CLI arguments (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24318">#24318</a>) <ul> <li> args: add --video-* CLI arguments </li> <li> gen docs </li> <li> nits </li> <li> a…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-27 10:48

b10646

<details open=""> metal : fix memory leaks due to missing autoreleasepools (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27758">#27758</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollo…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-27 10:04

b10645

<details open=""> llama : add --n-cpu-ffn option (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26622">#26622</a>) <ul> <li> common : dedupe --n-cpu-moe / --spec-draft-n-cpu-moe override loops </li> <li> common : add --n-cpu…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-27 05:24

b10644

<details open=""> models : support nanbeige4.2-3B (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27730">#27730</a>) Co-authored-by: admin <a href="mailto:[email protected]">[email protected]</a> </details> <stron…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-27 02:16

b10643

<details open=""> hexagon: support for multi-NPU devices (IQ9, IQ10) and fully asynchronous backend (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26501">#26501</a>) <ul> <li> hexagon: use non-host bufs by default and make the back…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-26 22:02

b10642

<details open=""> llama: add token ID tracking to KV cell (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27762">#27762</a>) <ul> <li> kv: track token id </li> <li> rm get_prev_tokens, move it to the main pr </li> <li> <p…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-26 16:53

b10639

<details open=""> vulkan: warptiles currently assume warp sizes <= 64, clamp to work around larger warps (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27726">#27726</a>) </details> Website: <ul> <li><a href…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-26 16:11

b10638

<details open=""> Implemented vulkan cross_entropy_loss and cross_entropy_loss_back (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27216">#27216</a>) </details> Website: <ul> <li><a href="https://llama.app" rel…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-26 12:46

b10636

<details open=""> ci: Clean up UI builds from releases (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27706">#27706</a>) <ul> <li> ci : inline UI version resolution into ui-build.yml </li> <li> ci : build UI once and reuse t…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-26 11:34

b10635

<details open=""> cuda: unblock mmq for MoE on sm_60 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26264">#26264</a>) <ul> <li> cuda: unblock mmq for MoE on sm_60 </li> <li> cuda: duplicate mmq-config-pascal for dp4a and ol…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-26 09:31

b10632

<details open=""> ggml-metal: add chunked SSD MMA for Mamba-2 prefill optimization (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26647">#26647</a>) <ul> <li> metal: WIP chunked SSD SSM_SCAN kernels for multi-token prefill </li…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-26 06:06

b10631

<details open=""> ggml-meta: propagate buffer usage and call init on the new tensors (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27586">#27586</a>) </details> Website: <ul> <li><a href="https://llama.app" re…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-25 21:38

b10630

<details open=""> kleidiai: Rework KleidiAI Build System/Integration (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26077">#26077</a>) <ul> <li>Rework KleidiAI Build System/Integration</li> </ul> Signed-off-by: Jonathan Clohessy <a…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-25 20:26

b10629

<details open=""> ci : update OS used for ROCM to Ubuntu 24.04 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27681">#27681</a>) This matches what other build targets use and also what AMD advertises wheels as supporting.…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-25 18:08

b10628

<details open=""> rpc: support apple RDMA as an RPC transport (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26421">#26421</a>) <ul> <li> rpc: support apple RDMA as an RPC transport </li> <li> remove set_tensor micro optimiz…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-25 17:28

b10625

<details open=""> chat : scope qwen3-coder workarounds (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27679">#27679</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.app…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-25 12:19

b10622

<details open=""> metal : null-check buffer alloc to fix OOM crash (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25371">#25371</a>) <ul> <li>metal : null-check ggml_metal_buffer_init result to avoid OOM crash</li> </ul> ggml_backe…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-25 10:18

b10621

<details open=""> llama.cpp : bump version to 0.3.0 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27696">#27696</a>) <ul> <li> llama.cpp : bump version to 0.3.0 </li> <li> ci : update release default desc </li> <li> …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-25 09:37

b10620

<details open=""> sync : ggml </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.app</a></li> </ul> Attestations: <ul> <li><a href="https://github.com/ggml-org/llama.cpp/attestations/4281028…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-25 07:09

b10618

<details open=""> grammar : parse - in char classes as literal hyphen (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27591">#27591</a>) <ul> <li>grammar : accept "-" escape in character classes</li> </ul> gbnf_escape_char_class() e…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-25 06:47

b10617

<details open=""> sycl : mark tq2_0 as not supported (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27660">#27660</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.app</…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-24 21:54

b10615

<details open=""> metal : per-device tuned (Q, NE) for flash-attn vec (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26570">#26570</a>) <ul> <li> metal : per-device tuned (Q, NE) for flash-attn vec (<a class="issue-link js-issue-li…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-24 21:25

b10614

<details open=""> metal: per-op source split + parallel compile (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26561">#26561</a>) <ul> <li> metal : per-op source split + parallel compile (<a class="issue-link js-issue-link" href="h…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-24 13:43

b10612

<details open=""> tests : disable DOTS3NOTE arch test for WebGPU (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27654">#27654</a>) Co-authored-by: Stanisław Szymczyk <a href="mailto:[email protected]">[email protected]</a> </…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-24 13:18

b10610

<details open=""> ggml : shorten virtual device naming in CUDA and Metal (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27608">#27608</a>) <ul> <li>ggml : shorten virtual device naming in CUDA and Metal</li> </ul> Assisted-by: llam…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-24 12:56

b10608

<details open=""> mtmd: video: fix moov atom at the end of file (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27596">#27596</a>) <ul> <li>mtmd: video: fix moov at the end of file</li> </ul> Co-authored-by: rkfg <a href="mailto:rkf…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-24 12:33

b10606

<details open=""> ggml : fix ggml_clamp (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27644">#27644</a>) <ul> <li> ggml : fix ggml_clamp </li> <li> cont : update ggml-alloc </li> </ul> </details> Website:</st…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-24 08:59

b10605

<details open=""> mamba2 : Flatten in/out projections to dispatch GEMM instead of GEMV (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27513">#27513</a>) <ul> <li> mamba2 : flatten mamba2 in/out projections to dispatch gemm instead …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-24 07:18

b10604

<details open=""> Deepseek 4: <code>-sm tensor</code> (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26490">#26490</a>) <ul> <li> DSV4: sm tensor </li> <li> set coarser granularity for head splits </li> <li> fix dspar…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-23 18:47

b10603

<details open=""> model : support MTP in GLM-4.5-Air (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26534">#26534</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.app</…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-23 17:41

b10599

<details open=""> test: move tools/parser to tests (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27548">#27548</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] · 2026-08-23 17:18

b10598

<details open=""> mtmd: use pillow-accurate algo, correct resize_algo for all models (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27594">#27594</a>) <ul> <li> mtmd: use pillow-accurate resize algo, correct resize_algo for all mod…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-23 13:45

b10595

<details open=""> server : add LLAMA_SERVER_SLOTS_N_DIFF (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27600">#27600</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.a…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-23 13:23

b10594

<details open=""> common : skip device_info loop if it's not going to be printed (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26692">#26692</a>) The device_info loop iterates over the discovered devices and gets the availab…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-23 11:41

b10593

<details open=""> DeepseekV4: fix rollback with multi-seq (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26756">#26756</a>) <ul> <li> DeepseekV4: fix rollback with multi-seq </li> <li> fix model loading </li> <li> mak…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-23 08:22

b10590

<details open=""> vendor : update subprocess.h (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27409">#27409</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] · 2026-08-23 08:00

b10589

<details open=""> cuda : add POOL_1D support (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27573">#27573</a>) <ul> <li> cuda : add POOL_1D support </li> <li> fix: add missing trailing newline for editorconfig compliance…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-22 23:42

b10588

<details open=""> common: json.h: fix clang lto (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27575">#27575</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.app</a></l…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-22 20:13

b10587

<details open=""> vulkan : added the PAD_REFLECT_1D operation (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26586">#26586</a>) <ul> <li>vulkan : added PAD_REFLECT_1D operation</li> </ul> Implemented the GGML_OP_PAD_REFLECT_1D oper…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-22 16:05

b10586

<details open=""> mtmd: use ggml_rope_set_offset (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27521">#27521</a>) <ul> <li> mtmd: use ggml_rope_set_offset </li> <li> add comment </li> </ul> </details> Website…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-22 15:42

b10585

<details open=""> common: add json.h abstraction (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27511">#27511</a>) <ul> <li> add common/json </li> <li> migrate common </li> <li> adapt jinja </li> <li> migrate s…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-22 15:07

b10584

<details open=""> fit: also take into account n_streams (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27496">#27496</a>) <ul> <li> fit: also take into account n_streams </li> <li> server: make the draft context follow the t…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-22 12:59

b10582

<details open=""> ci : Restore ROCm job for Ubuntu (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27399">#27399</a>) <ul> <li>Revert "ci : disable ubuntu-rocm (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llam…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-22 10:14

b10581

<details open=""> model : support DSpark for bailingmoe3 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27508">#27508</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.a…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-22 09:35

b10580

<details open=""> mtmd: support dots3-note vision+audio (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27524">#27524</a>) <ul> <li> text: conversion </li> <li> init impl </li> <li> mtmd: conversion </li> <li> i…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-22 09:12

b10578

<details open=""> ggml: optimize concat op by replacing per-element memcpy with row-level memcpy (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24575">#24575</a>) <ul> <li> ggml: optimize concat op by replacing per-element memcpy w…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-22 08:15

b10577

<details open=""> common : fix draft-mtp with embeddings (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/issues/26352">#26352</a>, <a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/issues/27299">#27299</a>) (<a cla…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-22 07:50

b10576

<details open=""> sycl : add Q2_K reordered MMVQ and ESIMD kernels (again) (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27490">#27490</a>) <ul> <li>Revert "Revert "sycl : add Q2_K reordered MMVQ and ESIMD kernels (<a class="issue-li…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-22 05:40

b10573

<details open=""> mtmd: support webp via ffmpeg (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27520">#27520</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.app</a></l…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-22 01:19

b10569

<details open=""> model: add dots3-note (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27060">#27060</a>) <ul> <li> text: conversion </li> <li> init impl </li> <li> address review comments </li> <li> fix rope</…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-21 22:58

b10568

<details open=""> model: use ggml_rope_set_offset() (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27382">#27382</a>) <ul> <li> model: use ggml_rope_set_offset() </li> <li> partially apply to deepseek2 </li> </ul> </deta…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-21 19:00

b10567

<details open=""> ci : run ccache-clear as the last step of release jobs (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27503">#27503</a>) <ul> <li>ci : run ccache-clear as the last step of release jobs</li> </ul> Assisted-by: pi:l…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-21 18:27

b10566

<details open=""> llama.cpp : bump version to 0.2.0 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27498">#27498</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] · 2026-08-21 09:24

b10549

<details open=""> TP: enable tensor split for LFM2/LFM2MOE (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26993">#26993</a>) Assisted-by: deepseek-v4-flash </details> Website: <ul> <li><a href="https://l…

llama.cpp — Releases TIER_1 (SO) · itsvedantkumar · 2026-08-21 04:36

b10548

docs: fix typos in ET.md (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27457">#27457</a>)

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-21 03:50

b10545

<details open=""> metal : clamp K extent in tensor API mat-mat kernel for K not a multiple of 32 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27450">#27450</a>) The Tensor API mat-mat path of kernel_mul_mm (GGML_METAL_HAS_TENSOR)…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-21 03:29

b10541

<details open=""> mtmd: add --mmproj-device argument (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/23255">#23255</a>) <ul> <li> feat: add --mmproj-device arg & backwards compatible MTMD_BACKEND_DEVICE env var </li> <li> <p…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-21 03:05

b10539

<details open=""> vulkan: FA MMQ should use fp32 for Q quantization calculations (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27413">#27413</a>) Codex found that qd could be a denorm and 1/qd would overflow. </details> <st…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-21 02:44

b10538

<details open=""> metal : dequant kv cache only for large batches (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27438">#27438</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https:…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-21 02:23

b10537

<details open=""> CI: Use LLVM's OpenMP over MSVC_DEBUG_non_redist on Windows (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26678">#26678</a>) <ul> <li>CI: Use LLVM's OpenMP over MSFT_DEBUG_non_redist on Windows</li> </ul> Current…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-21 02:05

b10536

<details open=""> server: (router) lazy-load startup_models after main setup (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27424">#27424</a>) <ul> <li> server: (router) lazy-load startup_models after main setup </li> <li> o…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-21 01:43

b10534

<details open=""> CUDA: adding switch points per HW and quant type to tune the mvq->MMQ decode crossover (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26079">#26079</a>) <ul> <li>CUDA: runtime GGML_CUDA_MMVQ_MAX to tune the mvq-&g…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-21 01:08

b10533

<details open=""> common : gracefully fallback on unsupported regex patterns in JSON schema (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26939">#26939</a>) </details> Website: <ul> <li><a href="https://llama.…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-21 00:44

b10532

<details open=""> metal : dequantize quantized KV to F16 before flash attention (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27390">#27390</a>) <ul> <li>metal: dequantize q8_0 KV to f16 before flash attention</li> </ul> Add a pre…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-21 00:21

b10531

<details open=""> Revert "tensor-split meta backend fixes (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26502">#26502</a>)" (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27433">#27433</a>) T…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-20 21:05

b10520

<details open=""> ggml-cpu: gate __fp16 on __ARM_FP16_FORMAT_IEEE (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26860">#26860</a>) <ul> <li>ggml-cpu: gate __fp16 on __ARM_FP16_FORMAT_IEEE</li> </ul> __ARM_NEON only signals NEON av…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-20 20:33

b10519

<details open=""> server: refactor sleep handling, allow access /metrics during sleep (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27376">#27376</a>) <ul> <li> add cached responses </li> <li> refactor on_sleeping_state…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-20 17:45

b10517

<details open=""> vulkan : dequant q8_0 KV once in coopmat1 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25494">#25494</a>) <ul> <li>vulkan : dequant q8_0 KV once in coopmat1</li> </ul> Assisted-by: Claude (Opus 4.8) <ul> <li…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-20 13:22

b10516

<details open=""> vulkan: add null checks in ggml_vk_queue_command_pools_cleanup (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27353">#27353</a>) <ul> <li>Guard against null queue pointers.</li> </ul> </details> Website:</…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-20 11:19

b10514

<details open=""> model : GraniteSWAForCausalLM / GraniteMoeSWAForCausalLM (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25505">#25505</a>) <ul> <li>feat(convert): Add conversion for GraniteSWAForCausalLM</li> </ul> Branch: Granit…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-20 06:54

b10509

<details open=""> ggml: add ggml_rope_set_offset (+ metal support) (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27120">#27120</a>) <ul> <li> add params </li> <li> cpu kernel </li> <li> metal kernel </li> <li> <p…

llama.cpp — Releases TIER_1 (SO) · wanghqc · 2026-08-20 03:40

b10524

opencl: make the MoE expert scatter deterministic (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26464">#26464</a>)

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-20 03:23

b10507

<details open=""> mtmd: add mtmd_bitmap_set_mergeable (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27348">#27348</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.app<…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-20 02:59

b10506

<details open=""> metal : dequantize q8_0 using packed types (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27370">#27370</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://lla…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-20 02:38

b10505

<details open=""> server: add dedup-cache-models preset option (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27346">#27346</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://l…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-19 14:22

b10502

<details open=""> ci : add attestation for signed release artifacts (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25933">#25933</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">http…

llama.cpp — Releases TIER_1 (SO) · ngxson · 2026-08-18 23:45

b10499

server: (cosmetic) do not print cmd_child_to_router messages [no rele…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-18 11:07

b10488

<details open=""> ci : Update OpenVINO to 2026.3, skip nemotron-h rollback test (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27292">#27292</a>) <ul> <li> update to ov-2026.3, update device drivers </li> <li> ci: skip nemot…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-18 10:45

b10486

<details open=""> mtmd: fix LFM2 image tiling threshold (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27057">#27057</a>) <ul> <li> mtmd: fix LFM2 image tiling threshold </li> <li> refactor testing </li> <li> fix …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-18 10:10

b10485

<details open=""> sync : ggml </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.app</a></li> </ul> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b1048…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-18 09:26

b10483

<details open=""> build : fix xcframework + cmake clean-up (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27304">#27304</a>) <ul> <li> xcframework : fix build </li> <li> mtmd : remove unused include path </li> <li> ve…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-17 17:18

b10472

<details open=""> cuda : skip UMA override for HIP builds (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27083">#27083</a>) AMD APUs report accurate memory via hipMemGetInfo. Using MemAvailable over-promises on small-carveout…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-17 14:01

b10470

<details open=""> ci : push release tag explicitly in release.yml (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27261">#27261</a>) Add a "Create and push git tag" step to the release job, right before the "Create release" st…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-17 06:31

b10456

<details open=""> sycl: fix thread/block count in quantized cpy kernel launches (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27160">#27160</a>) Adjusts the thread/block count to be proportional to the size of the quant, red…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-17 05:49

b10455

<details open=""> [SYCL] support OP OPT_STEP_ADAMW, OPT_STEP_SGD (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25268">#25268</a>) <ul> <li> fix conflict </li> <li> fix conflict of ops.md </li> <li> fix conflict of op…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-16 12:55

b10453

<details open=""> model : remove some ggml_concat (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27176">#27176</a>) Co-authored-by: Xuan Son Nguyen <a href="mailto:[email protected]">[email protected]</a> </details> <stron…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-16 11:23

b10452

<details open=""> chat: refactor handling supports_string_content / supports_typed_content (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27130">#27130</a>) <ul> <li> better supports_string_content cap detect </li> <li> test…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-16 07:26

b10451

<details open=""> llama : check LoRA tensor data is within file bounds (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27056">#27056</a>) <ul> <li> llama : check LoRA tensor data is within file bounds </li> <li> Update src/ll…

llama.cpp — Releases TIER_1 (SO) · crowmoed · 2026-08-15 20:52

b10450

ui: mask API Key field in settings and error splash to stop browser a…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-15 20:50

b10448

<details open=""> model: add Kimi-K3 text model (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26185">#26185</a>) <ul> <li>model: add Kimi-K3 text model</li> </ul> Hybrid KDA (linear) + MLA (full) attention as in Kimi-Linear-48B, p…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-15 20:15

b10447

<details open=""> server: re-design yield_to_queue thread model (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27133">#27133</a>) <ul> <li> run common_speculative_process in worker </li> <li> swap worker <--> main thre…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-15 19:45

b10446

<details open=""> vendor: update BoringSSL to 0.20260813.0 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27099">#27099</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-15 18:43

b10444

<details open=""> common: support --models-dir loading MTP assistant models (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24431">#24431</a>) <ul> <li> common: support --models-dir loading MTP assistant models </li> <li> com…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-15 18:05

b10443

<details open=""> fix: check gguf array type before reading (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27075">#27075</a>) <ul> <li> fix: check gguf array type before reading </li> <li> update skill </li> </ul> </deta…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-15 14:59

b10442

<details open=""> vulkan: add SHMEM_STRIDE_PAD/APPLY_SLM_A_RESHAPE for coopmat1 on Intel Xe (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25380">#25380</a>) <ul> <li> vulkan: add SHMEM_STRIDE_PAD/APPLY_SLM_A_RESHAPE for coopmat mu…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-15 11:26

b10441

<details open=""> common: migrate the deprecated --mmap/--no-mmap to --load-mode (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26934">#26934</a>) Replace the deprecated --mmap, --no-mmap, --mlock, and --direct-io flags with …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-15 05:25

b10437

<details open=""> model : add support for MiniMaxText01ForCausalLM and MiniMaxM1ForCausalLM (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27018">#27018</a>) <ul> <li> llama : support for MiniMax-Text-01 model </li> <li> cho…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-15 04:14

b10436

<details open=""> mtmd, common: various fixes (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27071">#27071</a>) <ul> <li> apply fixes </li> <li> cont </li> <li> revert gguf fix </li> </ul> </details> We…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-14 22:29

b10435

<details open=""> jinja : fix quadratic cost in gather_string_parts (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27034">#27034</a>) <ul> <li> jinja : fix quadratic cost in gather_string_parts </li> <li> fix some comments</…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-14 19:24

b10434

<details open=""> chat : pass reasoning_effort to template <ul> <li>chat: add reasoning_effort to common_chat_templates_inputs</li> </ul> Store OpenAI Chat Completions reasoning_effort and make it available to jinja templates (with model specific translations<br /…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-14 18:21

b10433

<details open=""> sync : ggml </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.app</a></li> </ul> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b1043…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-14 17:44

b10431

<details open=""> ggml : recurrent state rollback for ggml_ssm_scan (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26623">#26623</a>) <ul> <li> Initial changes for Recurrent state rollback for nemotron for cpu and cuda </li> <l…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-14 12:59

b10430

<details open=""> llama : allow virtual igpu devices (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26953">#26953</a>) <ul> <li> llama : allow virtual igpu devices </li> <li> cont : better comment </li> </ul> </details> …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-14 12:15

b10429

<details open=""> server: allow accessing /metrics and /slots during llama_decode() (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27041">#27041</a>) <ul> <li> server_queue::worker </li> <li> call llama_decode inside yield_t…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-14 09:08

b10428

<details open=""> tests : replace personal home directory paths with generic placeholders (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27043">#27043</a>) Scrub developer-specific /home// paths from example docs and test fix…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-14 08:15

b10427

<details open=""> sycl: fuse mul_mat(gate) + mul_mat(up) + GLU for q4_K dense FFN (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26779">#26779</a>) Measured on Arc Pro B70 (Battlemage, Level Zero), llama-bench -r 20, two inte…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-14 07:30

b10426

<details open=""> ggml: force single thread on wasi (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25686">#25686</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] · 2026-08-14 06:52

b10425

<details open=""> sycl: fuse the gated-delta-net state writeback cpy (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26643">#26643</a>) Port of <a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/239…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-14 05:46

b10424

<details open=""> dflash : clarify output logging of target_layer_ids (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27013">#27013</a>) This commit tries to make the logging of target_layer_ids a bit clearer and easier to rea…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-13 22:45

b10423

<details open=""> common: apply CPU parameters across tools (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27026">#27026</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llam…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-13 22:13

b10419

<details open=""> OpenVINO: Qwen3.5, memory optimization, and test-recurrent-state-rollback (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26952">#26952</a>) <ul> <li> OpenVINO backend: 1) enable gpt-oss moe on OV bk; 2) enable mxf…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-13 21:38

b10418

<details open=""> [SYCL] Support host pinned mem to improve SYCL Host-to-Device Memory Access (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26789">#26789</a>) <ul> <li> support host pinned mem, ggml_backend_sycl_host_buffer_type_g…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-13 20:51

b10417

<details open=""> chat : fix LFM2 tool call arg name prefix ambiguity (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26960">#26960</a>) Assisted-by: Claude Opus 5 </details> Website: <ul> <li><a href="ht…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-13 20:08

b10416

<details open=""> server : serve index.html with no-cache (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27006">#27006</a>) index.html was served with <code>max-age=31536000, immutable</code> like the hashed assets, but its name is…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-13 19:30

b10415

<details open=""> spec : auto-detect mtp draft model type (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27005">#27005</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-13 18:42

b10414

<details open=""> metal : add TQ2_0 support (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26980">#26980</a>) <ul> <li>metal: add TQ2_0 support</li> </ul> Add support for the GGML_TYPE_TQ2_0 (ternary, 2 bits per element) type in<br…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-13 18:00

b10413

<details open=""> common : auto-detect spec type from draft GGUF metadata (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26814">#26814</a>) <ul> <li>common : auto-detect spec type from draft GGUF metadata</li> </ul> When -md loads …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-13 17:18

b10412

<details open=""> spec: enable backend sampling for both dflash & dspark (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26958">#26958</a>) <ul> <li> dflash: enable backend sampling for both dflash & dspark </li> <li> <p…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-13 16:37

b10411

<details open=""> ggml-cpu/ops: vectorize flash-attention V-cache F16 to F32 conversion (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26947">#26947</a>) Co-authored-by: jinzihao <a href="mailto:[email protected]">jinzih…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-13 15:54

b10410

<details open=""> sycl: remove separate fp32 type promotion in gemm non-oneDNN path (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26372">#26372</a>) <ul> <li> sycl: use automatic fp16 promotion in gemm </li> <li> sycl: remo…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-13 14:31

b10408

<details open=""> sycl : Add DMMV ESIMD Q3_K kernel (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26251">#26251</a>) <ul> <li>Add DMMV Q4_K and Q6_K ESIMD kernels</li> </ul> Configure cmake build with -DGGML_SYCL_ESIMD=ON to enabl…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-13 07:34

b10405

<details open=""> ggml-hip : remove -funsafe-math-optimizations (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26696">#26696</a>) It enables -fassociative-math, which reassociates FP reductions and can flip greedy argmax on R…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-13 06:05

b10400

<details open=""> ggml : fix arm builds, unused var (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26991">#26991</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] · 2026-08-12 12:20

b10375

<details open=""> chat : tighten bare function parsing for Qwen models (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26793">#26793</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">h…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-12 09:24

b10373

<details open=""> imatrix.cpp: Move finite check and only check touched experts (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26861">#26861</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="no…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-12 04:54

b10369

<details open=""> mtmd: support pocket-tts (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26871">#26871</a>) <ul> <li> adapt the api </li> <li> text model ok </li> <li> working impl, need verify and clean up </li>…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-11 17:14

b10361

<details open=""> model : fix SWA not being enabled for EXAONE 4.5 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26848">#26848</a>) <ul> <li>model : fix SWA not being enabled for EXAONE 4.5</li> </ul> load_arch_hparams tests <code…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-11 13:19

b10360

<details open=""> common/peg : suppress incomplete escape sequences (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26780">#26780</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">http…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-11 10:51

b10359

<details open=""> ggml-webgpu: fix CI errors from <a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25025">#25025</a> and <a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25262">#25262</a> (<a class="issue-…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-11 09:06

b10358

<details open=""> Address review comment of PR 25532 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26852">#26852</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.app</…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-11 05:19

b10357

<details open=""> opencl: transpose the K tile in local memory for FA prefill kernels (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26428">#26428</a>) </details> Website: <ul> <li><a href="https://llama.app" r…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-11 02:47

b10356

<details open=""> ci : target ROCm 7.14 for build and release (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25775">#25775</a>) <ul> <li>Switch ROCm from 7.2.1 to 7.14</li> </ul> ROCm 7.14 is the first production release using TheR…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-10 23:17

b10355

<details open=""> llama : support multi-output backend sampling (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25532">#25532</a>) <ul> <li> Enable backend sampling with token speculation </li> <li> Clamp the mask sum before …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-10 22:40

b10354

<details open=""> ggml-cpu : fix CPU affinity mask being ignored on Android (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26838">#26838</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofoll…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-10 22:02

b10353

<details open=""> ggml : require contiguous src for ROLL on CUDA and Metal (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25928">#25928</a>) ggml_roll only asserts nb[0] == ggml_type_size, so a permuted src is a valid input, …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-10 16:25

b10344

<details open=""> model: add MTP support for Nemotron model (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26725">#26725</a>) <ul> <li> model: add MTP support for Nemotron Nano model </li> <li> model: add mtp_flags for nemot…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-10 15:25

b10343

<details open=""> vendor : update cpp-httplib to 0.53.0 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26821">#26821</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.ap…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-10 12:49

b10342

<details open=""> model : Granite-Switch Architecture (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25107">#25107</a>) <ul> <li>granite-switch: add llama.cpp backend (POC, CPU)</li> </ul> New "granite-switch" architecture: a dense…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-08 16:01

b10329

<details open=""> server, ui: only offer a working directory when a tool reads it (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26762">#26762</a>) The working directory chip showed up as soon as the server exposed any builti…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-08 15:23

b10328

<details open=""> server: add initial tool isolation support (via docker) (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26507">#26507</a>) <ul> <li> server: add initial tool isolation support (via docker) </li> <li> add doc…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-08 06:06

b10327

<details open=""> CUDA: fix thread/block count in quantized cpy kernel launches (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26731">#26731</a>) <ul> <li> CUDA: fix thread/block count in quantized cpy kernel launches </li> <li…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-07 21:25

b10326

<details open=""> tts: account for the vocoder pass in the timings line (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26733">#26733</a>) get_output runs the waveform work the pipeline defers to it, from a single trailing win…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-07 19:53

b10322

<details open=""> sycl: coalesce the ssm_conv window loads (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26612">#26612</a>) test-backend-ops perf -o SSM_CONV on an Arc Pro B70, interleaved A/B against master, 6 reps, us/run:…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-07 19:09

b10321

<details open=""> metal : fix NORM/RMS_NORM for row lengths that leave a partial simdgroup (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26708">#26708</a>) ggml_metal_op_norm sized the threadgroup with <code>nth = std::min(n…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-07 18:09

b10319

<details open=""> mtmd: fix longest_edge ignoring min/max pixels (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26638">#26638</a>) <ul> <li> mtmd: fix longest_edge ignoring min/max pixels </li> <li> nits </li> </ul> </de…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-07 17:37

b10318

<details open=""> sync : ggml </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.app</a></li> </ul> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b1031…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-07 17:00

b10313

<details open=""> server: (router) add LRU scheduler (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26572">#26572</a>) <ul> <li> add lru_sched </li> <li> handle coalescing (req leaves waiting queue) </li> <li> add tes…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-07 16:25

b10312

<details open=""> server: (router) do not evict busy models (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26567">#26567</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llam…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-07 15:46

b10311

<details open=""> mtmd: stop feeding the text stream again during Qwen3-TTS generation (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26706">#26706</a>) The reference implementation has two mutually exclusive prompt layouts. …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-07 15:00

b10310

<details open=""> ggml : add aarch64 HWCAP fallbacks and fix fp16 variant detection (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25554">#25554</a>) <ul> <li> ggml : add fallback definitions for missing aarch64 HWCAP bits </li…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-07 07:19

b10299

<details open=""> metal : avoid <code>threadgroup</code> matrix array instantiation in kernel_lightning_indexer (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26646">#26646</a>) <ul> <li> In MSL, declaring an array of matrix types …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-06 22:12

b10298

<details open=""> mtmd: add chunk save/load function (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26645">#26645</a>) <ul> <li> mtmd: add chunk save/load function </li> <li> nits </li> <li> add tests </li> <li> <…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-06 13:42

b10297

<details open=""> server: fix empty response for /cors-proxy (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26656">#26656</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://lla…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-06 12:58

b10295

<details open=""> model-loader : fix quantized reshaped tensor strides (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26672">#26672</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">h…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-06 09:14

b10293

<details open=""> ci : onboard AMD ROCm CI with gfx1151 fixes (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26544">#26544</a>) <ul> <li>ci: prepare for amd rocm ci</li> </ul> Signed-off-by: Aaron Teo <a href="mailto:aaron.teo1@ibm…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-06 08:03

b10291

<details open=""> vulkan: fix submission batching size, add debug tools for diagnosing causes of DeviceLost drivers errors (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26371">#26371</a>) <ul> <li> vulkan: add debug tooling to get…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-06 01:35

b10290

<details open=""> mtmd/ggml: add ggml_build_forward_order (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26649">#26649</a>) <ul> <li>ggml: add ggml_build_forward_order</li> </ul> ggml_build_forward_expand marks the tensor and all i…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-05 20:09

b10289

<details open=""> server: harden the file_glob_search directory walk (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26626">#26626</a>) <ul> <li>server: don't walk Windows junctions in file_glob_search</li> </ul> std::filesystem rep…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-05 17:42

b10288

<details open=""> tests: re-enable MiniMax M3 in <code>test-llama-archs</code> (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26633">#26633</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nof…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-05 17:06

b10287

<details open=""> mtmd: Unlimited-OCR fix max_tiles, setting in converter (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25614">#25614</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-05 16:24

b10286

<details open=""> grammar : degrade max repetition >= 2000 to unbounded (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26613">#26613</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollo…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-05 15:22

b10285

<details open=""> mtmd: support multi-row batching for deepseek-ocr (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26154">#26154</a>) <ul> <li> mtmd: support multi-row batching for deepseek-ocr </li> <li> mtmd: weave deepsee…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-05 14:39

b10284

<details open=""> fit: Fix memory allocation for MTP layers (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26605">#26605</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llam…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-05 13:51

b10282

<details open=""> server: Adding spec-decode counters to /metrics endpoint (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26389">#26389</a>) <ul> <li> <ul> <li>server: add spec-decode counters to /metrics endpoint</li> </ul> </li> <li…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-05 11:36

b10280

<details open=""> vendor : apply patches for subprocess.h (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26606">#26606</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-05 10:18

b10278

<details open=""> build : remove GGML_METAL_USE_BF16 from all build scripts (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26604">#26604</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofoll…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-05 00:23

b10276

<details open=""> Prefer npm ci over install for security (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26601">#26601</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-04 23:50

b10275

<details open=""> server: decode Windows OEM output to UTF-8 in built-in tools (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26597">#26597</a>) a child process writes in the OEM code page, which is not UTF-8 on a western Win…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-04 23:14

b10274

<details open=""> mtmd: correcting duplicate empty audio chunks for short inputs (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26536">#26536</a>) <ul> <li> correcting duplicate empty audio chunks for short inputs </li> <li> <p…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-04 20:33

b10273

<details open=""> sampler : remove "full-context windows" from history-based samplers (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26524">#26524</a>) <ul> <li>Resolve -1 to 1024 instead of ctx-len for samplers</li> </ul> Because …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-04 18:56

b10271

<details open=""> ui: CWD for agent (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26518">#26518</a>) <ul> <li> server : extend file_glob_search for UI pickers </li> <li> ui : add per-conversation working directory with pick…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-04 18:06

b10270

<details open=""> mtmd: support Qwen3-TTS (note: breaking change to llama-tts binary) (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26254">#26254</a>) <ul> <li> convert text model </li> <li> main model load ok </li> <li…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-04 17:24

b10269

<details open=""> models : fix dflash wo_a reshape on load (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26577">#26577</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-04 16:40

b10268

<details open=""> ci: fix pre-built binaries no longer working on macOS 15 and below (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26375">#26375</a>) <ul> <li> ci: fix pre-built binaries no longer working on macOS 15 and below…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-04 15:56

b10267

<details open=""> speculative : refactor enabled configs common_speculative_init (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26510">#26510</a>) This commit contains a suggestion to reduce some code duplication in common_sp…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-04 15:17

b10265

<details open=""> sync : ggml </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.app</a></li> </ul> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b1026…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-04 13:56

b10262

<details open=""> vulkan backend ops: implemented GATED_LINEAR_ATTN (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25601">#25601</a>) <ul> <li> vulkan : add GATED_LINEAR_ATTN op </li> <li> docs : update Vulkan ops </li> …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-04 13:12

b10261

<details open=""> vocab : validate plamo2 byte tokens (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26511">#26511</a>) <ul> <li> validate plamo2 byte tokens </li> <li> --typo </li> </ul> </details> Website:</…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-04 12:19

b10259

<details open=""> model : allow reshape of tensors during load (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26531">#26531</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://l…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-04 09:24

b10258

<details open=""> llama : move n_vocab from llama_sampler_data to penalty_sampler (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26520">#26520</a>) This matches how it is done for logit_bias and mirostat samplers, see <a clas…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-04 06:41

b10256

<details open=""> sycl: parallelize the non-contiguous concat kernel (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25852">#25852</a>) <ul> <li>sycl: parallelize the non-contiguous concat kernel</li> </ul> Launch geometry only: the…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-04 05:40

b10255

<details open=""> Extended SYCL oneDNN SDPA to non-FP16 KV caches (Q4_0–Q8_0 and FP32) (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25874">#25874</a>) <ul> <li>sycl: extend oneDNN SDPA to Q4_0-Q8_0 and F32 KV caches</li> </ul> Ex…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-04 04:55

b10254

<details open=""> chat : add new template for DeepSeek V4 Flash 0731 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26398">#26398</a>) <ul> <li>common/chat: update DeepSeek V4 templates</li> </ul> Align the DeepSeek V4 templates wi…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-04 04:25

b10253

<details open=""> vendor : update cpp-httplib to 0.52.0 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26485">#26485</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.ap…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-04 03:49

b10252

<details open=""> vendor : update BoringSSL to 0.20260803.0 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26523">#26523</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llam…

llama.cpp — Releases TIER_1 (SO) · netrunnereve · 2026-08-04 03:28

b10257

ci: fix vulkan llvmpipe runs (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26533">#26533</a>)

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-04 03:08

b10251

<details open=""> model : support MTP in GLM-4.7-Flash (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24868">#24868</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.app…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-04 02:07

b10250

<details open=""> tests: add model resolution test on synthetic repo listings (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26172">#26172</a>) <ul> <li>tests: add model resolution test on synthetic repo listings</li> </ul> Include…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-04 01:38

b10249

<details open=""> server: add get_info tool (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26522">#26522</a>) <ul> <li> server: add get_info tool </li> <li> fix --rpc in docs </li> <li> server: harden get_info probe r…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-04 01:03

b10248

<details open=""> vocab : validate default special token ids (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26506">#26506</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://lla…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-04 00:28

b10247

<details open=""> ggml: use dynamic allocation for split graph inputs (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/22789">#22789</a>) <ul> <li>ggml: use dynamic allocation for split graph inputs</li> </ul> Replace fixed-size GGML…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-03 23:40

b10246

<details open=""> opencl: route large q6_K lm_head to the flat GEMV (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26427">#26427</a>) <ul> <li>add a direct size condition for <code>large</code> weights; the original dimension co…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-03 23:00

b10245

<details open=""> graph : fix unused input tensors in minimax m3 graph (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26519">#26519</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">h…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-03 22:15

b10244

<details open=""> model: M3: Move MSA into a new memory implementation (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26338">#26338</a>) <ul> <li> Move MSA logic from llama-kv-cache into llama-kv-cache-msa </li> <li> cont : …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-03 20:12

b10243

<details open=""> llama : allocate indexer cache only in "full" indexer layers (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26474">#26474</a>) Co-authored-by: Stanisław Szymczyk <a href="mailto:[email protected]">sszymczy@gmail.…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-03 17:24

b10242

<details open=""> CUDA: Add backend sampler for penalties sampler (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25262">#25262</a>) <ul> <li>sampling: enhance penalty handling in common_sampler_init</li> </ul> <ul> <li>Set default val…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-03 15:00

b10241

<details open=""> CUDA: Fix data-races when reusing SMEM in block_reduce (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26385">#26385</a>) <ul> <li>CUDA: Fix data-races when reusing block_reduce</li> </ul> block_reduce currently do…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-03 11:26

b10240

<details open=""> server: add notice for upcoming default port change 8080 --> 9931 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26508">#26508</a>) <ul> <li> server: add notice for upcoming default port change 8080 --> 6631…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-03 08:59

b10238

<details open=""> model: MTP support for Qwen3-Next (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25589">#25589</a>) <ul> <li> mtp for qwen3nex </li> <li> fix for python type-check </li> <li> Fix to compute num_mtp f…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-03 07:51

b10237

<details open=""> llama : MTP support for DeepSeek V3.2 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26457">#26457</a>) <ul> <li> llama : MTP support for DeepSeek V3.2 </li> <li> model : no need to include MTP layers durin…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-03 05:13

b10236

<details open=""> metal: implement DSv4 Lightning Indexer (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25893">#25893</a>) <ul> <li>metal: implement F16 Lightning Indexer</li> </ul> <ul> <li>Implement GGML_OP_LIGHTNING_INDEXER for 12…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-02 21:04

b10235

<details open=""> metal : add SILU_BACK (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25982">#25982</a>) <ul> <li> feat(silu_back): implemented silu_back op for f32 </li> <li> fix(silu_back): removed redundant asserts in gg…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-02 20:19

b10234

<details open=""> metal : add F16 support for bin ops (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26465">#26465</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.app<…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-02 19:40

b10233

<details open=""> opencl: limit local workgroup size for GLU operation (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26383">#26383</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">h…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-02 18:59

b10232

<details open=""> metal: implement DeepSeek V4 hyper-connections (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26459">#26459</a>) <ul> <li>Implement GGML_OP_DSV4_HC_COMB, GGML_OP_DSV4_HC_PRE, and GGML_OP_DSV4_HC_POST with SIMDg…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-02 18:15

b10231

<details open=""> common: support the DSpark sidecar resolution (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26458">#26458</a>) The dspark- files resolve like the other speculative sidecars: the -hfd tag applies to them, a …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-02 14:33

b10229

<details open=""> opencl: bugfix increment ref_count in ggml_backend_opencl_init() (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26162">#26162</a>) Incrementing <code>ref_count</code> at the beginning is important later in t…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-02 13:30

b10228

<details open=""> DeepseekV4 MTP + DSpark (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25784">#25784</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.app</a></li> </u…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-02 09:45

b10227

<details open=""> chat : add qwen3 specialized parser (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26252">#26252</a>) <ul> <li> Add tagged thinking tool parser </li> <li> chat : refactor and add permute helper </li> <l…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-02 08:32

b10225

<details open=""> model : load MiMo V2 MTP tensors only if used (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26412">#26412</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-02 08:16

b10226

<details open=""> sycl: fix classification of iGPUs (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26105">#26105</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] · 2026-08-02 07:06

b10224

<details open=""> ggml-webgpu: add support for f16 repeat (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26307">#26307</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-01 22:50

b10223

<details open=""> test: fix some CI errors (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26415">#26415</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] · 2026-08-01 19:31

b10221

<details open=""> vendor : update BoringSSL to 0.20260730.0 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26353">#26353</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llam…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-01 16:48

b10219

<details open=""> cli : persist reasoning_content in chat history (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26362">#26362</a>) <ul> <li>cli : persist reasoning_content in chat history</li> </ul> llama-cli collected reasoning f…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-01 12:47

b10218

<details open=""> mtmd: add minicpmv46 downsample (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25993">#25993</a>) <ul> <li>add minicpmv46 downsample</li> </ul> Signed-off-by: tc-mb <a href="mailto:[email protected]">tianchi_…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-08-01 06:52

b10217

<details open=""> chat : enable tool call in thinking for DS4 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26269">#26269</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://ll…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-31 22:06

b10216

<details open=""> vulkan: add POOL_1D op (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25431">#25431</a>) <ul> <li>vulkan : add pool1d push constants and pipeline field</li> </ul> Declared data structures needed for POOL1D OP, whi…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-31 21:33

b10215

<details open=""> vulkan: Introduce driver version check for Windows Intel GPU to mitigate crashing (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25192">#25192</a>) <ul> <li>Removed crash guard for Intel</li> </ul> Crash fixed fro…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-31 20:47

b10214

<details open=""> mtmd: add n_embd_head (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26342">#26342</a>) Co-authored-by: Daniel Han <a href="mailto:[email protected]">[email protected]</a> </details> Website:</s…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-31 20:12

b10213

<details open=""> Support rotated kv cache quant (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26180">#26180</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] · 2026-07-31 19:27

b10212

<details open=""> llama : load MTP tensors only if they are really used (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26296">#26296</a>) <ul> <li> llama : load MTP tensors only if they are really used </li> <li> llama : ski…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-31 18:53

b10211

<details open=""> vulkan: update vulkan sdk to 1.4.357.0 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26303">#26303</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.a…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-31 18:11

b10210

<details open=""> server: correct accepted tokens when need draft token replay (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26320">#26320</a>) <ul> <li> spec: correct accepted tokens when need draft token replay </li> <li> <p…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-31 17:30

b10209

<details open=""> cuda: extract Q2_0 elements via __byte_perm (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25603">#25603</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://ll…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-31 16:05

b10208

<details open=""> SYCL: add oneMKL GEMM flash attention for XMX-accelerated prompt proc… (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25025">#25025</a>) <ul> <li> SYCL: add oneMKL GEMM flash attention for XMX-accelerated prompt p…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-31 15:14

b10207

<details open=""> [SYCL] support the missed types in cpy (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26005">#26005</a>) <ul> <li> support the missed types in cpy </li> <li> use correct funct </li> <li> rm unused co…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-31 13:28

b10205

<details open=""> ggml-zendnn : group matmul direct API for mul_mat_id (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25918">#25918</a>) <ul> <li> ggml-zendnn : group matmul API for mul_mat_id </li> <li> ggml-zendnn : scale …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-31 12:46

b10204

<details open=""> sycl : support dev2dev memcpy by DEV2DEV_MEMCPY_FORWARD (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26234">#26234</a>) Co-authored-by: Neo Zhang Jianyu <a href="mailto:[email protected]">jianyu.zhang@intel…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-31 12:16

b10203

<details open=""> [SYCL] Support q2 mul_mat (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26231">#26231</a>) <ul> <li> support q2_0 in mul_mat </li> <li> support more q2_0 case </li> </ul> </details> Website:…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-31 11:43

b10202

<details open=""> sycl: fuse RMS_NORM + MUL (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26015">#26015</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] · 2026-07-31 09:47

b10201

<details open=""> ggml-webgpu: improve flash_attn_vec for quantized KV at long contexts (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25956">#25956</a>) <ul> <li> improve fa of quantized kv cache </li> <li> Fix some bugs an…

How we ranked this

COVERAGE [357]