llama.cpp releases multiple updates with cross-platform optimizations

By PulseAugur Editorial · [118 sources] · 2026-07-14 12:07

The llama.cpp project has released several updates, including versions b10106, b10105, b10108, b10099, b10098, b10094, b10093, b10092, b10091, and b10103. These releases introduce various improvements and fixes across different platforms and hardware accelerators. Notable updates include enhancements for CUDA, Metal, Hexagon, and OpenVINO, as well as fixes for specific model templates like DeepSeekv4 and improvements to argument parsing and memory management. AI

IMPACT Ongoing improvements to a popular open-source inference engine enhance its performance and compatibility across diverse hardware and operating systems.

RANK_REASON The cluster consists of multiple release notes for the llama.cpp project, detailing software updates and bug fixes.

Read on llama.cpp — Releases →

product

AI-generated summary · Google Gemini · from 118 sources. How we write summaries →

llama.cpp releases multiple updates with cross-platform optimizations

How we ranked this

Signal score: 0 / 100; Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket: Tool; The cluster consists of multiple release notes for the llama.cpp project, detailing software updates and bug fixes.
Source corroboration: 118 independent sources; Strong cross-source corroboration — multiple independent publishers covered this within the clustering window.
Topics: product; Editorial topic classification. Feeds into how the story surfaces on /topic/<slug> hub pages and into the per-entity coverage mix.
AI-industry relevance: High; Clearly on-topic for AI-industry coverage.
Story freshness: 58 days old; Aged out of breaking-news scoring windows; ranking reflects the durable signal from the full source set.
Coverage growth since scoring: +46 source(s) since last score; New sources have picked up this story since our last re-score. Score will update on the next scoring pass.

Full methodology in our editorial standards.

COVERAGE [118]

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-31 14:26

b10206

<details open=""> llama : enforce the same K and V cache types for DeepSeek V4; enable FA if V cache is quantized (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25871">#25871</a>) <ul> <li> llama : enforce the same K and V cache ty…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-30 20:34

b10199

<details open=""> server: support inp embd to generate next token (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26313">#26313</a>) <ul> <li> server: support embd for sampled token </li> <li> fix ~server_batch() </li> </…
llama.cpp — Releases TIER_1 (SO) · ngxson · 2026-07-30 19:59

b10200

mtmd: add lanczos resize method [no release] (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26341">#26341</a>)
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-30 19:39

b10198

<details open=""> vulkan: Support quantized concat (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25684">#25684</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] · 2026-07-30 19:07

b10197

<details open=""> Test support for alternative conv layout (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25617">#25617</a>) <ul> <li> add bool cwhn = true to conv_2d test cases </li> <li> add layout check at graph building …
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-30 18:30

b10196

<details open=""> llama-context : sync pending async copies before clearing embd_seq (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25676">#25676</a>) </details> Website: <ul> <li><a href="https://llama.app" re…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-30 17:25

b10195

<details open=""> tests : avoid building get-model.cpp many times (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26317">#26317</a>) <ul> <li> tests : remove get-model.cpp </li> <li> tests : fix quant type selection </li>…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-30 15:02

b10194

<details open=""> ggml-cuda: Allow transpose-free gemmv computation (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26171">#26171</a>) When matrix's weights are shaped 1xK is leverage a transpose-free computation to use mat_mu…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-30 13:46

b10192

<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/b1019…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-30 09:35

b10189

<details open=""> Remove custom cpu op from the M3 graph, express with stock ops (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26297">#26297</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="n…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-30 08:59

b10188

<details open=""> metal: fix memory unwire if model is freed without any GPU operations (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26082">#26082</a>) <ul> <li> metal: fix memory leak if model is freed without any GPU operations…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-30 07:37

b10186

<details open=""> ggml : Fix issue with kleidiai ci and stringop overflow warning (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26277">#26277</a>) Signed-off-by: Jonathan Clohessy <a href="mailto:[email protected]">Jonatha…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-30 04:42

b10184

<details open=""> mimo2: address MTP review feedback (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26228">#26228</a>) Co-authored-by: tnhnyc <a href="mailto:[email protected]">[email protected]…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-29 17:03

b10182

<details open=""> llama: move suppress_tokens handling to common/sampling (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26276">#26276</a>) <ul> <li> llama: move suppress_tokens handling to common/sampling </li> <li> address…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-29 15:11

b10181

<details open=""> ggml-cuda : disable MMQ on devices with less than 48 KiB shared memory (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26141">#26141</a>) ggml_cuda_should_use_mmq() selects MMQ purely from the quantization ty…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-29 14:35

b10180

<details open=""> sycl: contiguous fast path + 32-bit index math for unary elementwise ops (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25946">#25946</a>) <ul> <li> sycl: contiguous fast path + 32-bit index math for unary element…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-29 13:52

b10179

<details open=""> vendor: update BoringSSL to 0.20260728.0 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26241">#26241</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-29 13:03

b10178

<details open=""> server : add trace logging for slot similarity checking (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26271">#26271</a>) Adds trace logging in server-context.cpp for slot similarity checking during prompt c…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-29 08:52

b10176

<details open=""> RPC: add tensor_memset (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25912">#25912</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.app</a></li> </ul…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-29 08:07

b10175

<details open=""> add rdna3.5, and 3 to mmq configs so they can be tuned independently. (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26199">#26199</a>) </details> Website: <ul> <li><a href="https://llama.app"…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-29 07:17

b10174

<details open=""> model: add NextN/MTP speculative decoding support for GLM_DSA (GLM-5.2) (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25980">#25980</a>) <ul> <li>model: add NextN/MTP speculative decoding support for GLM_DSA (GLM-5.…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-28 20:06

b10173

<details open=""> model: Add Laguna-S-2.1 LLM_TYPE (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26233">#26233</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] · 2026-07-28 19:29

b10172

<details open=""> ggml-webgpu: Fix some binding alias issues to support all archs, fix recurrent-state-rollback test (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25931">#25931</a>) <ul> <li> Add overlap glu variant to support all…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-28 18:55

b10171

<details open=""> opencl: skip the Adreno KQ/KQV image kernels for multi-stream batches (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26189">#26189</a>) The Adreno KQ/KQV image1d kernels (ggml_cl_mul_mat_kq_kqv_adreno) ignore<br /…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-28 17:04

b10167

<details open=""> server: abstract llama_memory calls to common_memory (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26221">#26221</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">h…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-28 16:17

b10166

<details open=""> ggml : set output of view src (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25729">#25729</a>) <ul> <li> llama-graph: set_outputs to t->view_src </li> <li> change set_output to GGML_ASSERT about views n…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-28 15:31

b10165

<details open=""> vulkan: add iq4_nl support back to FA (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24585">#24585</a>) <ul> <li>vulkan: add iq4_nl support back to FA</li> </ul> I was originally concerned about wasting shared mem…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-28 14:30

b10164

<details open=""> ggml-cuda: add chunked SSD matmul for Mamba-2 prefill acceleration (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/22675">#22675</a>) <ul> <li> ggml-cuda: add chunked SSD matmul for Mamba-2 prefill acceleration…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-28 11:45

b10159

<details open=""> ggml-metal: FWHT kernel for metal backend (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25924">#25924</a>) <ul> <li> metal fwht wip </li> <li> shape guard and formatting </li> <li> formatting </…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-28 08:56

b10158

<details open=""> spec: add eagle3-v3 support for gpt-oss model (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25794">#25794</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-28 00:42

b10156

<details open=""> Disable -ffast-math on HIP (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25495">#25495</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] · 2026-07-27 22:01

b10155

<details open=""> mtmd: support MiMo-V2.5 audio input (RVQ-based model) (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26190">#26190</a>) <ul> <li> gguf converter for mimo audio </li> <li> fix conv </li> <li> cpp impl…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-27 17:01

b10154

<details open=""> common : add common_print_available_devices() (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26170">#26170</a>) Signed-off-by: Adrien Gallouët <a href="mailto:[email protected]">[email protected]</a> </det…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-27 15:58

b10153

<details open=""> model: Add support for Nanbeige4.2 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25994">#25994</a>) <ul> <li> support nanbeige4.2 model </li> <li> fix </li> <li> fix flake8 Lint check </li> <li>…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-27 14:15

b10152

<details open=""> fit : count nextn (MTP) blocks in n_gpu_layers so front layers stay on GPU (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26177">#26177</a>) </details> Website: <ul> <li><a href="https://llama…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-27 13:26

b10151

<details open=""> sycl(build): parallelize ocloc invocations (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25903">#25903</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://lla…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-27 12:47

b10150

<details open=""> ggml : adjust logic for offloading ops to weight's backend (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25832">#25832</a>) <ul> <li> ggml : adjust logic for offloading ops to weight's backend </li> <li> l…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-27 11:47

b10149

<details open=""> tests : remove unnecessary sync in test-save-load-state (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26166">#26166</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-27 11:16

b10148

<details open=""> common: fix explicit -md precedence over draft sidecar resolution (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26165">#26165</a>) <ul> <li>common: fix explicit -md precedence over draft sidecar resolution</li> </ul…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-27 10:16

b10146

<details open=""> ggml-cpu: Enable BF16 tiled gemm optimization on PowerPC (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26068">#26068</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollo…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-27 09:30

b10145

<details open=""> args: add <code>-lm mlock</code> where it mlocks but doesnt mmap (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26135">#26135</a>) <ul> <li>arg: add <code>-lm mlock</code> where it mlocks but doesnt mmap</li> </ul> <…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-27 06:15

b10144

<details open=""> server + ui: fix stream routes for model names containing a slash (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26137">#26137</a>) <ul> <li>server + ui: refactor resumable stream routes to query string conv_id</li> …
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-27 00:22

b10142

<details open=""> mtmd: Add Vision Support for Minimax-M3 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25113">#25113</a>) <ul> <li>Add preliminary MiniMax-M3 support</li> </ul> Text-only port that re-uses existing components: Min…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-26 23:05

b10141

<details open=""> mtmd: fix android build (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26150">#26150</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.app</a></li> </u…
llama.cpp — Releases TIER_1 (SO) · ServeurpersoCom · 2026-07-26 04:51

b10133

ui: fix context gauge card regressions and land at the conversation e…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-24 14:02

b10107

<details open=""> hexagon: fix Windows crash when op_poll is enabled (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/26029">#26029</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">htt…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-24 11:36

b10106

<details open=""> CUDA: fix external compilation of q1_0 MMQ (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25778">#25778</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://lla…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-24 09:35

b10105

<details open=""> args: refactor mlock/mmap/directio into load-mode (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/20834">#20834</a>) <ul> <li>args: overhaul mmap/mlock/dio into single arg</li> </ul> Signed-off-by: Aaron Teo <a hre…
llama.cpp — Releases TIER_1 (SO) · max-krasnyansky · 2026-07-24 02:13

b10108

hexagon: further improved pipeline of the core bits (L2, DMA, MM, FA)…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-23 23:35

b10099

<details open=""> CUDA: Improve NVFP4 W4A4 activation quantization (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25730">#25730</a>) <ul> <li>Squash history before conflict-resolution during rebase on master</li> </ul> WIP commit</…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-23 20:13

b10098

<details open=""> hexagon: activation ops update (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25974">#25974</a>) <ul> <li> hex-geglu: optimized all-in-one geglu microkernel </li> <li> hex-geglu: enable non-contiguous src a…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-23 18:29

b10094

<details open=""> common: infer the speculative type from the draft repo sidecars (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25989">#25989</a>) With -hfd pointing to a repo that ships mtp-/dflash-/eagle3- sidecars and no …
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-23 17:21

b10093

<details open=""> Fix DeepSeek4 crafted template (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25414">#25414</a>) <ul> <li> chat: fix DS4 template to explicitly follow reference behavior </li> <li> Support DeepSeekv4 flag (…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-23 05:32

b10092

<details open=""> ggml: enable PowerPC backend variants on AIX (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25983">#25983</a>) <ul> <li>ggml: enable PowerPC backend variants on AIX</li> </ul> Allow the PowerPC CPU backend variant…
llama.cpp — Releases TIER_1 (SO) · iliailmer · 2026-07-23 03:45

b10103

metal : add f16 type support to leaky relu (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25981">#25981</a>)
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-22 21:13

b10091

<details open=""> ci : fix SYCL package shared library lookup (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25987">#25987</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://ll…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-22 20:25

b10090

<details open=""> webgpu : add CONV_2D_DW (depthwise conv2d) kernel (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25847">#25847</a>) <ul> <li>webgpu : add CONV_2D_DW (depthwise conv2d) kernel</li> </ul> Implement GGML_OP_CONV_2D_D…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-22 19:47

b10089

<details open=""> cuda: GET_ROWS quants (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25962">#25962</a>) <ul> <li>cuda: add k-quant support to GET_ROWS</li> </ul> Device-side embedding lookups require GET_ROWS to handle the k-quan…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-22 11:05

b10087

<details open=""> Add support for Laguna XS.2 & M.1 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25165">#25165</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.ap…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-22 09:54

b10085

<details open=""> mtmd : use align_corners for qwen3vl vision position embedding interpolation (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25781">#25781</a>) The Qwen3-VL learned position embedding is interpolated to the runtime…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-22 09:14

b10084

<details open=""> hexagon: check tensor type when reusing descriptors (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25968">#25968</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">ht…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-22 08:35

b10083

<details open=""> cuda: add sqrt_softplus in topk-moe for dsv4 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25896">#25896</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://l…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-22 06:29

b10082

<details open=""> kleidiai : warn once when a weight type has no KleidiAI kernel (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25701">#25701</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="n…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-22 05:44

b10081

<details open=""> common: resolve draft repo to its requested sidecar (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25955">#25955</a>) With -hfd pointing to a repo shipping speculative sidecars, the draft resolved to the mai…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-22 05:07

b10080

<details open=""> server: return 400 instead of 500 on validation error with X-Conversation-Id (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25760">#25760</a>) <ul> <li>server: return 400 instead of 500 on validation error with X-Con…
llama.cpp — Releases TIER_1 (SO) · helanfxz · 2026-07-22 02:55

b10088

llama-arch: fix DeepSeek4 APE tensor op (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25945">#25945</a>)
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-22 02:05

b10079

<details open=""> server : properly handle null llama_context (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25868">#25868</a>) Co-authored-by: Stanisław Szymczyk <a href="mailto:[email protected]">[email protected]</a> </det…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-21 22:40

b10078

<details open=""> vulkan: Refactor vk_queue to use per-instance mutexes and unique handles (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/23570">#23570</a>) <ul> <li> Refactor vk_queue to use per-instance mutexes and unique handles…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-21 21:55

b10077

<details open=""> ggml-openvino: Add GGML_BACKEND_DL_IMPL invocation for OpenVINO backend (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25795">#25795</a>) This adds the missing <code>GGML_BACKEND_DL_IMPL()</code> macro invocation,…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-21 15:54

b10076

<details open=""> CUDA: vectorize same-type get_rows with int4 copy (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25929">#25929</a>) k_get_rows_float did a scalar one-element-per-thread copy and recomputed the row-invariant …
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-20 06:31

b10069

<details open=""> opencl: Support broadcast for Adreno MUL_MAT and honor <code>view_offs</code> for Adreno Q8_0 MUL_MAT for llama-server multi-stream (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25910">#25910</a>) <ul> <li> openc…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-18 13:38

b10068

<details open=""> model: rotate injected K/V cache for DFlash (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25823">#25823</a>) <ul> <li> dflash: rotate injected K/V cache when using K/V quantization </li> <li> Update src/mo…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-18 12:14

b10067

<details open=""> llama-quant : exclude i32 ffn_gate_tid2eid routing table from quantization (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25787">#25787</a>) DeepSeek-V4's ffn_gate_tid2eid tensor is an i32 token-id -> expert-id…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-17 23:13

b10066

<details open=""> opencl: load and use <code>kernel_gemm_moe_q6_k_f32_ns</code> from bin kernel lib (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25797">#25797</a>) </details> Website: <ul> <li><a href="https:…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-17 15:42

b10064

<details open=""> opencl: transpose q4_K noshuffle scales for coalesced reads (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25805">#25805</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofo…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-17 15:00

b10063

<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/b1006…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-17 14:20

b10061

<details open=""> tests : initialize all tensors in test_dsv4_hc to avoid NaNs in sentinel tensors (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25822">#25822</a>) Co-authored-by: Stanisław Szymczyk <a href="mailto:sszymczy@gmail.…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-17 12:24

b10059

<details open=""> ggml-blas: default hadamard mul_mat to cpu routine (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25710">#25710</a>) Signed-off-by: Aaron Teo <a href="mailto:[email protected]">[email protected]</a> </detail…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-17 11:45

b10058

<details open=""> vulkan: Support Q2_0 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25430">#25430</a>) <ul> <li>vulkan: Support Q2_0</li> </ul> The backend perf tests for mat-vec-mul weren't very good at first (worse than q…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-17 10:24

b10057

<details open=""> sycl: fix row calculation when K_QUANTS_PER_ITERATION is 1 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25690">#25690</a>) <ul> <li>sycl: fix incorrect row calculation when K_QUANTS_PER_ITERATION=1</li> </ul> Si…
llama.cpp — Releases TIER_1 (SO) · Gezahegne · 2026-07-17 05:13

b10056

opencl: add ABS op (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25115">#25115</a>)
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-17 00:30

b10054

<details open=""> docs: added a note about using OpenCl with Adreno 810 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25786">#25786</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-16 21:25

b10052

<details open=""> hexagon: L2 cache handling rework (dirty bit tracking with lazy flushing) and more MUL_MAT updates (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25762">#25762</a>) <ul> <li> hex-mm: fix artificial limit in the so…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-16 20:51

b10051

<details open=""> kleidiai: Add SME vs SME2 distinction in kernel dispatch (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25478">#25478</a>) The current integration treats SME as a single capability (CPU_FEATURE_SME) with no …
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-16 20:10

b10050

<details open=""> vulkan: when using transfer queue for async copies, sync on event_wait to avoid race (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25229">#25229</a>) </details> Website: <ul> <li><a href="htt…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-16 19:24

b10048

<details open=""> TP: fix Phi3, Bert, Plamo2/3, ChatGLM (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25536">#25536</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.ap…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-16 18:46

b10047

<details open=""> vendor: update BoringSSL to 0.20260713.0 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25624">#25624</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-16 17:08

b10046

<details open=""> tests: actually exercise <code>test-recurrent-state-rollback</code> (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25758">#25758</a>) </details> Website: <ul> <li><a href="https://llama.app" r…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-16 16:11

b10045

<details open=""> server : allow text-only slot save/restore with mtmd (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25076">#25076</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">h…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-16 14:04

b10043

<details open=""> CUDA: Support CUDA Virtual Devices (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25228">#25228</a>) <ul> <li> support cuda virtual devices </li> <li> disable NCCL path when virtual devices are used </l…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-16 11:24

b10042

<details open=""> Enable CUDA graphs on volta+turing (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25749">#25749</a>) </details> Website: <ul> <li><a href="https://llama.app" rel="nofollow">https://llama.app</…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-16 10:38

b10041

<details open=""> server: Ignore empty / non-existing <code>Origin</code> headers (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25756">#25756</a>) Otherwise this gives lots of unnecessary warnings: W srv operator(): (CORS) …
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-16 09:57

b10040

<details open=""> ggml-cuda : restore prop.integrated on HIP builds (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24233">#24233</a>) PR <a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/16308">#1…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-16 06:59

b10038

<details open=""> ci : add official website link to release notes (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25728">#25728</a>) Assisted-by: pi:llama.cpp/Qwen3.6-27B </details> Website: <ul> <li><a h…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-16 06:17

b10037

<details open=""> quant : allow using manual tensor types with --pure (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25716">#25716</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama.…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-16 05:32

b10036

<details open=""> opencl: disable FA and MoE weights repack to work around compiler issues for Adreno 850 GPU (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25745">#25745</a>) <ul> <li> opencl: workaround for A850 compiler compat</…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-16 05:00

b10035

<details open=""> cuda: extract Q1_0 elements via __byte_perm (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25628">#25628</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama.cpp/rele…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-15 20:37

b10034

<details open=""> opencl: exclude some moe kernels on Adreno a7x (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25698">#25698</a>) <ul> <li>opencl: exclude Adreno A7x from using Adreno MoE kernels</li> </ul> Some compilers for A7x …
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-15 19:53

b10032

<details open=""> cuda : CUDA GGML_OP_LIGHTNING_INDEXER implementation (generic vector kernel + wmma kernel) (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25545">#25545</a>) <ul> <li> cuda : CUDA GGML_OP_LIGHTNING_INDEXER implemen…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-15 19:17

b10031

<details open=""> tokenize : drop --stdin mutual-exclusion check (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25672">#25672</a>) match cli and completion, which don't enforce it </details> macOS/iOS: <…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-15 16:12

b10025

<details open=""> cuda : relax tensor contiguity requirements for quantized concat (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25678">#25678</a>) <ul> <li> cuda : relax tensor contiguity requirements for quantized concat </l…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-15 14:22

b10021

<details open=""> DeepseekV4: reduce graph splits (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25702">#25702</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama.cpp/releases/downloa…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-15 13:39

b10020

<details open=""> sycl : fix get_rows Q2_K, Q4_K, Q5_K (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25656">#25656</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama.cpp/releases/do…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-15 13:03

b10019

<details open=""> sycl : support kernel type fp16 for conv2d_dw (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25653">#25653</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama.cpp/re…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-15 12:23

b10018

<details open=""> sycl : implement xielu op (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25550">#25550</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b100…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-15 11:39

b10017

<details open=""> sycl: Increase minimum buffer size for USM system allocations (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25525">#25525</a>) Raise the threshold for minimum buffer size from 1 GiB to 4 GiB, based on real-…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-15 09:14

b10016

<details open=""> [SYCL] Flash Attention with XMX engine via oneDNN (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25222">#25222</a>) <ul> <li> [SYCL] F16 (default) Flash Attention with XMX engine via oneDNN graph API; Qwen3.6-27b-…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-15 03:26

b10015

<details open=""> opencl: do not use <code>clCreateBufferWithProperties</code> when targeting CL 2.x (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25673">#25673</a>) </details> macOS/iOS: <ul> <li><a href="htt…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-14 21:36

b10012

<details open=""> hexagon: fix hmx-queue signal enum-narrowing problem (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25677">#25677</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-14 20:57

b10011

<details open=""> server : refactor prompt cache state ownership (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25649">#25649</a>) <ul> <li> server : clear checkpoints upon prompt clear </li> <li> server : move the prompt st…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-14 20:27

b10010

<details open=""> server: add --cors-* options (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25655">#25655</a>) <ul> <li> server: add --cors-* options </li> <li> add special "localhost" value </li> <li> add tests…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-14 19:44

b10007

<details open=""> opencl: fix a dp4a bug for devices where cl_khr_integer_dot_product is unavailable (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25639">#25639</a>) <ul> <li> opencl: do not fail backend init on devices without cl…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-14 18:59

b10005

<details open=""> DeepseekV4: fix seq_rm (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25588">#25588</a>) <ul> <li> DeepseekV4: fix seq_rm </li> <li> implement proper seq_cp </li> <li> create actual update context</p…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-14 18:18

b10004

<details open=""> vulkan/cpu: Support f16 as SET_ROWS src. (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25432">#25432</a>) <ul> <li>vulkan/cpu: Support f16 as SET_ROWS src.</li> </ul> This adds full support for f16 SET_ROWS (equi…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-14 17:44

b10003

<details open=""> tokenize : align usage by using common args (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25516">#25516</a>) Migrate the tokenize tool to common_params_parse, replacing its hand-rolled argv parsing, Windows…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-14 15:21

b10002

<details open=""> ggml : add a set of functions for checking contiguity of inner tensor dimensions (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25650">#25650</a>) Co-authored-by: Stanisław Szymczyk <a href="mailto:sszymczy@gmail.…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-14 12:47

b10001

<details open=""> tests: export-graph-ops: exit gracefully when called w/o arguments (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25619">#25619</a>) Fixes a segfault when <code>test-export-graph-ops</code> is called without any<b…
llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-14 12:07

b10000

<details open=""> ggml: uniformize im2col dst_type for all conv ops (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/23660">#23660</a>) <ul> <li> ggml: uniformize im2col dst_type for all conv ops </li> <li> Update ggml/src/ggm…

How we ranked this

COVERAGE [118]

RELATED ENTITIES

RELATED TOPICS