llama.cpp releases multiple updates with performance and bug fixes

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-14 11:26

b9999

<details open=""> kleidiai : add SME2 f32 kernel (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24414">#24414</a>) <ul> <li> kleidiai : add SME2 f32 kernel </li> <li> enable dynamic scheduling for SME2 f32 kernel </li> <…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-14 10:43

b9996

<details open=""> arg: Flush log before exiting after usage() (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25504">#25504</a>) Under certain conditions, it's possible for messages emitted via LOG() to get lost before exit, a…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-14 09:48

b9995

<details open=""> sycl: set fattn_vec_nthreads to 256 for Battlemage (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25205">#25205</a>) Currently detects lunarlake + battlemage / xe2 and sets the value to 256. Keeps def…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-14 05:30

b9994

<details open=""> metal : add Q2_0 support (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25419">#25419</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b9994…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-13 23:10

b9993

<details open=""> model: add Hy3 (hy_v3) support with MTP speculative decoding (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25395">#25395</a>) <ul> <li>model: add Hy3 (hy_v3) architecture support</li> </ul> Adds Tencent Hunyuan 3…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-13 21:06

b9992

<details open=""> CUDA: refactor MMQ kernel configuration (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24127">#24127</a>) <ul> <li> CUDA: refactor MMQ kernel configuration </li> <li> fix Blackwell config </li> <li> …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-13 19:53

b9990

<details open=""> spec: add Minimax2 eagle3 support <ul> <li> Fix nullptr in minimax2 EAGLE3 </li> <li> minor : add newline </li> </ul> <hr /> Co-authored-by: Georgi Gerganov <a href="mailto:[email protected]">[email protected]</a> </details> <s…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-13 19:12

b9988

<details open=""> tests: Harmonize header use (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25616">#25616</a>) <ul> <li> tests: Harmonize the use of private ggml includes </li> <li> tests: In test-backend-ops, use quoted in…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-13 15:42

b9987

<details open=""> gguf : add tensor shape accessor (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24405">#24405</a>) <ul> <li> gguf : add tensor shape accessors </li> <li> gguf : return tensor shape as const int64_t * </…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-13 09:08

b9986

<details open=""> chat : fix reasoning leak with force-opened bare templates (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24674">#24674</a>) <ul> <li>chat : fix reasoning leak with force-opened bare templates</li> </ul> The reaso…

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

b9985

<details open=""> sycl: add fused top-k MoE (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25217">#25217</a>) <ul> <li> sycl: add fused top-k MoE </li> <li> sycl: address review: GGML_SYCL_ENABLE_FUSION env, move fusion disp…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-13 07:40

b9984

<details open=""> sycl: add Q2_K to DMMV reorder path (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25064">#25064</a>) Signed-off-by: Todd Malsbary <a href="mailto:[email protected]">[email protected]</a> </details…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-13 01:30

b9982

<details open=""> server: honour per-request reasoning_budget_tokens in chat completions (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/23116">#23116</a>) <ul> <li>server: honour per-request reasoning_budget_tokens in chat completions…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-13 00:55

b9981

<details open=""> vendor : update cpp-httplib to 0.50.1 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25576">#25576</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama.cpp/releases/d…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-13 00:08

b9980

<details open=""> server: Don't consider models with --no-mmproj-auto as multimodal (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25590">#25590</a>) If mmproj is explicitly disabled via the model preset or command-line param…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-12 23:38

b9979

<details open=""> mtmd: fix silent prompt truncation on embedded NUL (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25548">#25548</a>) <ul> <li>mtmd: fix silent prompt truncation on embedded NUL</li> </ul> mtmd_input_text carried t…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-12 21:34

b9978

<details open=""> server : evict checkpoints within min-step of each other (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25472">#25472</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/l…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-12 16:19

b9977

<details open=""> server : fix image blocks in tool_result being dropped during Anthropic OpenAI conversion (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/22536">#22536</a>) <ul> <li>server : fix image blocks in tool_result being drop…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-12 14:50

b9975

<details open=""> gguf : reject empty metadata keys (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24917">#24917</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama.cpp/releases/downl…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-12 14:19

b9974

<details open=""> cuda: Don't crash when querying memory on device with no free memory. (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25157">#25157</a>) If a Cuda device has no or limited available memory, the actual call to…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-12 13:41

b9973

<details open=""> DeepseekV4: clear cache only for seq rather than full (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25521">#25521</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llam…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-12 13:10

b9972

<details open=""> server: allow stream for exec_shell_command (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25526">#25526</a>) <ul> <li> init stream </li> <li> add stream for shell tool </li> <li> add test </li> …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-12 12:38

b9971

<details open=""> server: refactor server_stream (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25541">#25541</a>) <ul> <li> server: refactoring, remove spipe from server_http_res </li> <li> wip </li> <li> remove non-…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-12 12:05

b9970

<details open=""> ggml : add GGML_OP_LIGHTNING_INDEXER that implements DeepSeek V3.2/V4 lightning indexer (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24231">#24231</a>) <ul> <li> ggml : add GGML_OP_LIGHTNING_INDEXER that impleme…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-12 05:32

b9969

<details open=""> Vulkan: route large matmuls to medium tile on Adreno (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24877">#24877</a>) <ul> <li>[Vulkan] Fixes llama-cli breaking over longer promts sizes</li> </ul> The llama-cli w…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-12 01:14

b9968

<details open=""> opencl: add int8 dp4 dense and MoE prefill optimization for Adreno GPUs (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25537">#25537</a>) <ul> <li> opencl: add int8 dp4 dense and moe GEMM </li> <li> opencl:…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-11 18:56

b9966

<details open=""> llama : make tensor-split regex patterns static (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24710">#24710</a>) llama_meta_device_get_split_state() recompiled 29 std::regex on every call. In -sm tensor mod…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-11 18:12

b9965

<details open=""> hexagon: improve ARGSORT performance for small tensors (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25512">#25512</a>) <ul> <li> hex-sort: add efficient bitomic sort in hvx regs up to 1024 elements </li> <li…

llama.cpp — Releases TIER_1 (SO) · kdkd · 2026-07-11 18:03

b9976

Fix conditional to display 'LLAMA_SPLIT_MODE_TENSOR not implemented f…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-11 10:54

b9964

<details open=""> arg: prevent duplicate spec model downloads (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25527">#25527</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] · 2026-07-11 08:40

b9963

<details open=""> mtmd: deepseek-ocr v1 multi-tile (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24717">#24717</a>) <ul> <li> mtmd: deepseek-ocr v1 multi-tile dynamic resolution + unified image-preprocessors for both versions (ds-…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-11 08:03

b9960

<details open=""> server: remove loading.html (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25500">#25500</a>) <ul> <li> server: remove loading.html </li> <li> apply ui changes </li> </ul> </details> macOS/iO…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-11 06:46

b9959

<details open=""> sync : ggml </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b9959/llama-b9959-bin-macos-arm64.tar.gz">macOS Apple Silicon (arm64)</a></li> <li>macOS Apple Silicon (arm64, KleidiAI ena…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-10 20:19

b9957

<details open=""> server: improve tools, remove apply_diff (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25498">#25498</a>) <ul> <li> server: improve tools, remove apply_diff </li> <li> improve edit tool </li> <li> a…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-10 19:48

b9956

<details open=""> cli: fix crash on wrong server base url (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25497">#25497</a>) <ul> <li> llama-cli: fix crash on wrong server base url by catching exceptions and graceful exit </li> …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-10 17:20

b9952

<details open=""> llama : make all KQ masks f16 if FA is used, remove zero attention bias, remove raw_k repeats in DeepSeek V4 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25370">#25370</a>) <ul> <li> llama : make all KQ masks (e…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-10 14:17

b9950

<details open=""> llama-batch: add unit test (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25471">#25471</a>) <ul> <li> llama-batch: add unit test </li> <li> fix win32 builds </li> <li> add not implemented assertion …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-10 10:39

b9949

<details open=""> opencl: cluster-parallel decode FA for Adreno (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25473">#25473</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] · 2026-07-10 07:45

b9948

<details open=""> ggml : process data in smaller chunks in CUDA ggml_top_k() and ggml_argsort() to reduce temporary buffers memory usage (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24776">#24776</a>) <ul> <li> ggml : process dat…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-10 04:48

b9947

<details open=""> cli: add --output option (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25484">#25484</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b9947…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-09 22:22

b9946

<details open=""> hexagon: tiling, tracing and optimizations for unary ops (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25474">#25474</a>) <ul> <li> hexagon: tile wide rows in pointwise unary ops to avoid VTCM overflow </li> …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-09 21:48

b9945

<details open=""> server : move chat-template thinking probe inside the init try/catch (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24093">#24093</a>) A model whose chat template parses at init but fails parser generation a…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-09 20:19

b9941

<details open=""> Only index by compile times + always multiply/add (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25445">#25445</a>) The first one avoids relying on compile to optimize local memory away, and the second is ch…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-09 11:39

b9940

<details open=""> llama-bench : init params.offline (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25476">#25476</a>) Signed-off-by: Adrien Gallouët <a href="mailto:[email protected]">[email protected]</a> </details> <st…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-09 11:01

b9939

<details open=""> metal : add CONV_2D_DW (depthwise convolution) support (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/21565">#21565</a>) <ul> <li> metal : add CONV_2D_DW (depthwise 2D convolution) support </li> <li> test :…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-09 10:10

b9938

<details open=""> ggml-hip: enable -funsafe-math-optimizations (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24668">#24668</a>) CUDA is compiled with fast math and AMD/HIP is not — this flag lets AMD use fast math too. We c…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-09 08:31

b9937

<details open=""> cuda: align snake fusion matcher with the other backends (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25460">#25460</a>) <ul> <li>cuda: fix snake fusion type predicate, a and inv_b are F32</li> </ul> The matcher…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-09 07:11

b9936

<details open=""> server : respect min-step when splitting prompt batches (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25420">#25420</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/ll…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-09 05:33

b9935

<details open=""> hexagon: add VISION RoPE support (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25216">#25216</a>) <ul> <li> hexagon: add VISION RoPE support </li> <li> hexagon: support RoPE on strided half-dim views for a…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-09 00:20

b9934

<details open=""> ggml-webgpu: tune subgroup split (d_split) in flash_attn_vec (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25418">#25418</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-o…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-08 23:24

b9933

<details open=""> opencl: Q6_K GEMM/GEMV fix for ne01 of weights that are not multiples of 128. (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25464">#25464</a>) <ul> <li>opencl: fix garbled output for Q6_K weights with ne01 % 128 != …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-08 21:47

b9932

<details open=""> vulkan: disable FA mask_opt on GCN to improve performance (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24362">#24362</a>) <ul> <li> vulkan: disable FA mask_opt on GCN to improve performance </li> <li> ree…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-08 21:13

b9931

<details open=""> opencl: ragged-tile MoE prefill FP16 GEMM optimization (skip padded expert tiles) (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25433">#25433</a>) <ul> <li>opencl: ragged-tile MoE prefill GEMM (skip padded expert ti…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-08 20:38

b9930

<details open=""> llama-batch: fix allowed decreasing pos in a seq (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25449">#25449</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama.cpp…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-08 19:55

b9929

<details open=""> vulkan: for small AMD GPUs, reduce submission threshold based on CU count (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25240">#25240</a>) </details> macOS/iOS: <ul> <li><a href="https://gith…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-08 19:17

b9928

<details open=""> hexagon: new vtcm layouts and improved pipelines for MUL_MAT, MUL_MAT_ID and FLASH_ATTN_EXT (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25425">#25425</a>) <ul> <li> hex-fa: refactor kernel param compute to use …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-08 18:30

b9927

<details open=""> cli : move to HTTP-based implementation (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24948">#24948</a>) <ul> <li> cli: move to HTTP-based implementation </li> <li> wip </li> <li> working </li> …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-08 16:59

b9925

<details open=""> cuda : add support for f16->f16 GGML_OP_SET_ROWS (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25367">#25367</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama.…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-08 16:14

b9924

<details open=""> llama: refactor fused ops (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24646">#24646</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b992…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-08 15:28

b9923

<details open=""> server-stream: follow-up on SSE Replay Buffer (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/23226">#23226</a>) (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25047">#25047</a>)…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-08 14:53

b9922

<details open=""> llama-batch: add n_keep_tail in split_equal for recurrent models (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25278">#25278</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/gg…

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

b9918

<details open=""> metal : add set_rows with src0 f16 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25434">#25434</a>) Co-authored-by: Georgi Gerganov <a href="mailto:[email protected]">[email protected]</a> </details> <…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-08 12:22

b9916

<details open=""> ggml : fix A indexing in simd_gemm scalar tail-column path (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25390">#25390</a>) <code>simd_gemm()</code> has an incorrect A-matrix index in the scalar tail-column path …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-08 11:49

b9915

<details open=""> ggml : add support for CPU f16->f16 GGML_OP_SET_ROWS (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25344">#25344</a>) <ul> <li> ggml : add support for CPU f16->f16 GGML_OP_SET_ROWS </li> <li> ggml : …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-08 11:14

b9914

<details open=""> opencl: fix potential crash in aos reconstruct (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25383">#25383</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama.cpp/r…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-08 10:35

b9913

<details open=""> Add Q2_0 quantization: type definition and CPU backend (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24448">#24448</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/lla…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-08 07:56

b9912

<details open=""> spec : fix naming, spacing (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25410">#25410</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b99…

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

b9911

<details open=""> CUDA: Fuse MMVQ post-scale for NVFP4 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24481">#24481</a>) <ul> <li>CUDA: Fuse MMVQ for NVFP4 and BS 1</li> </ul> TODO: <ol> <li>Add tests to test-backend-ops (did v…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-08 04:26

b9910

<details open=""> server : fix draft model fit vs load inconsistency (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25056">#25056</a>) <ul> <li> fix: draft model fit vs load inconsistency </li> <li> refactor(server): unify d…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-08 03:34

b9909

<details open=""> server : add timings and progress to /responses API stream (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25348">#25348</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-08 02:55

b9908

<details open=""> server: enforce prompt cache RAM limit (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25070">#25070</a>) Before this commit, --cache-ram was not a hard limit: <ul> <li>The cache always kept at least one entry,…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-08 02:22

b9907

<details open=""> common : add missing include in common.h (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25220">#25220</a>) Signed-off-by: zhangrunda <a href="mailto:[email protected]">[email protected]</a> <…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-08 01:51

b9906

<details open=""> ggml-hip : add -fno-finite-math-only alongside -ffast-math (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25373">#25373</a>) -ffast-math implies -ffinite-math-only under ROCm/clang 22, which disables INFINIT…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-08 00:25

b9905

<details open=""> llama: fix quantized kv-cache for dsv4 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25202">#25202</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama.cpp/releases/…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-07 23:50

b9904

<details open=""> [SYCL] fix unsupported UT cases of CONT & CPY (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25231">#25231</a>) <ul> <li> fix unsupported UT cases of CONT & CPY </li> <li> update ops.md </li> <l…

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

b9902

<details open=""> [SYCL] support OP cross_entropy_loss, cross_entropy_loss_back (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25236">#25236</a>) <ul> <li> support OP cross_entropy_loss, cross_entropy_loss_back </li> <li> co…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-07 16:43

b9901

<details open=""> sycl : set K_QUANTS_PER_ITERATION to 1 on DMMV path (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25063">#25063</a>) <ul> <li>sycl: add supported types to ggml_sycl_supports_reorder_dmmv</li> </ul> The reordered …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-07 15:20

b9899

<details open=""> sycl : enhance argsort to support all UT cases (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25125">#25125</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama.cpp/r…

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

b9898

<details open=""> sycl : use sycl func to fix AOT double type issue (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25081">#25081</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama.cp…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-07 13:23

b9897

<details open=""> sycl : rename the env vars from "disable" to "enable" (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25042">#25042</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llam…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-07 09:38

b9895

<details open=""> speculative : fix out-of-bounds read in ngram-map on prompt shrink (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/23936">#23936</a>) <ul> <li> speculative : fix out-of-bounds read in ngram-map on prompt shrink…

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

b9894

<details open=""> vulkan : check src0 type in GGML_OP_SET_ROWS to avoid failures due to unimplemented f16 support (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25351">#25351</a>) <ul> <li> vulkan : check src0 type in GGML_OP_SET_R…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-07 03:39

b9893

<details open=""> opencl: general flash attention decode performance optimizations (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25366">#25366</a>) <ul> <li> opencl: vec flash-attention decode kernels for f16/q8_0/q4_0 KV </li…

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

b9892

<details open=""> common: Set optimal default thread count for ppc ( linux as well as AIX) (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25237">#25237</a>) </details> macOS/iOS: <ul> <li><a href="https://githu…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-06 21:24

b9891

<details open=""> metal: add col2im_1d op (f32/f16/bf16) (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25176">#25176</a>) <ul> <li>metal: add col2im_1d op (f32/f16/bf16)</li> </ul> Gather kernel mirroring the CPU/CUDA path: each o…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-06 20:47

b9890

<details open=""> CUDA: remove -sm row, refactor cuBLAS (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24216">#24216</a>) <ul> <li> CUDA: remove -sm row, refactor cuBLAS </li> <li> fix CDNA + BF16 logic </li> <li> fix…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-06 14:59

b9888

<details open=""> CUDA: extend K-type validation to V-types for flash attention (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24403">#24403</a>) <ul> <li> CUDA: extend K-type validation to V-types for flash attention </li> <li…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-06 11:48

b9886

<details open=""> ggml-cpu: use UE4M3 LUT in ARM NVFP4 dot product (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25331">#25331</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama.cpp…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-06 11:06

b9885

<details open=""> ggml-cpu: Enable tiled matmul on AIX (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25199">#25199</a>) The matmul_tiled path uses large local stack buffers for A_pack and B_pack. On AIX this can trigger a segmenta…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-06 09:37

b9884

<details open=""> vulkan: fix 32-bit integer overflow in CEIL_DIV (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25245">#25245</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama.cpp/…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-06 09:03

b9882

<details open=""> scripts : use HF_TOKEN when downloading UI assets (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25280">#25280</a>) Signed-off-by: Adrien Gallouët <a href="mailto:[email protected]">[email protected]</a> <…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-06 08:24

b9881

<details open=""> ggml-hip: enable -ffast-math for HIP builds (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/23862">#23862</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama.cpp/rele…

llama.cpp — Releases TIER_1 (SO) · adavyas · 2026-07-06 03:49

b9879

ggml-cuda: optimize conv_transpose_1d indexing (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25310">#25310</a>)

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-05 19:44

b9878

<details open=""> Fix stale tensor-split params for draft models (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24814">#24814</a>) <ul> <li> meta: fix tensor split metadata for GQA attention </li> <li> Tidied the code a bit …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-05 19:14

b9877

<details open=""> abort if we see a multi buffer (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25276">#25276</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-05 18:30

b9876

<details open=""> ggml : fix tensor-parallel + -ncmoe crash on MoE models (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25028">#25028</a>) Tensor parallelism (-sm tensor) combined with -ncmoe (CPU-offloaded MoE experts) abor…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-05 15:55

b9874

<details open=""> cuda : concat implementation for quantized types (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25303">#25303</a>) <ul> <li> cuda : concat implementation for quantized types </li> <li> chore : apply am17an …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-04 21:11

b9873

<details open=""> llama : add guard for K/V rotation input when buffer is unallocated (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25215">#25215</a>) llm_graph_input_attn_kv::set_input and llm_graph_input_attn_kv_iswa::set_input<…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-04 12:17

b9871

<details open=""> ggml : fix broken CPU concat implementation for quantized types (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25247">#25247</a>) <ul> <li> ggml : fix broken CPU concat implementation for quantized types </li>…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-03 21:40

b9870

<details open=""> chat: trim messages sent to StepFun parser (fixes long reasoning loops) (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25238">#25238</a>) <ul> <li> chat: trim messages sent to StepFun parser (fixes long reasoning …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-03 14:57

b9867

<details open=""> spec: support spec-draft-p-min in DFlash (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25246">#25246</a>) <ul> <li> spec: support spec-draft-p-min in DFlash </li> <li> dflash: add n_min guard </li> <li…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-03 14:22

b9866

<details open=""> cuda: enable topk-moe fusion for 288 experts (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25267">#25267</a>) <ul> <li>cuda: enable topk-moe fusion for 288 experts</li> </ul> The topk-moe fusion only accepted pow…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-03 11:30

b9864

<details open=""> server + ui: ping silent SSE streams every 1s and kick only after 3s so slow prefill never drops healthy connections (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25241">#25241</a>) <ul> <li> server + ui: ping si…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-03 09:42

b9862

<details open=""> Remove redundant CUDA copies after gated_delta_net. (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/23940">#23940</a>) <ul> <li>Remove redundant CUDA copies after gated_delta_net.</li> </ul> Currently, GDN writes r…

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

b9861

<details open=""> vendor : update cpp-httplib to 0.49.0 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25218">#25218</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama.cpp/releases/d…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-02 16:29

b9860

<details open=""> llama : add llama_model_ftype_name() (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25134">#25134</a>) <ul> <li>llama : add llama_model_ftype_name()</li> </ul> Expose the model file type (quantization) name, e.g. …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-01 18:11

b9859

<details open=""> opencl: allow loading precompiled binary kernels from library (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/23042">#23042</a>) <ul> <li> opencl: allow loading binary kernel </li> <li> opencl: add libdl.h</…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-01 17:11

b9858

<details open=""> common : use hf primary split as model path (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25194">#25194</a>) Fixes <a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/issues/25181">#25…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-01 14:42

b9857

<details open=""> hexagon: flash attention rework (optimizations, accuracy improvements, etc) (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25085">#25085</a>) <ul> <li> hex-mm: fold mm quant tasks into the main matmul threads …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-01 11:11

b9856

<details open=""> CUDA: consistent use of restrict + PDL for FA (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25185">#25185</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-01 10:17

b9855

<details open=""> ggml-cpu: add AVX2 optimization for nvfp4 dot product and use UE4M3 LUT (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/23961">#23961</a>) </details> macOS/iOS: <ul> <li><a href="https://github…

llama.cpp — Releases TIER_1 (SO) · allozaur · 2026-07-01 05:32

b9853

ui: Remove PWA navigate fallback to prevent caching API endpoint requ…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-07-01 05:13

b9852

<details open=""> opencl: initial q1_0 support (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25160">#25160</a>) <ul> <li> opencl: general q1_0 support </li> <li> opencl: add Adreno GEMM/GEMV for q1_0 </li> </ul> </detai…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-06-30 19:37

b9851

<details open=""> cuda : prevent integer truncation and overflow errors when using KQ mask strides in flash_attn_mask_to_KV_max kernel (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24945">#24945</a>) Co-authored-by: Stanisław Szym…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-06-30 16:55

b9850

<details open=""> model : register t_layer_inp for qwen3next (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25141">#25141</a>) <ul> <li>Fix input assignment in layer processing loop</li> </ul> Fix DFLASH for qwen-coder-next <ul…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-06-30 14:57

b9849

<details open=""> common,server: handle bracketed IPv6 literals in URL authority (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25140">#25140</a>) <ul> <li>common,server: handle bracketed IPv6 literals in URL authority</li> </ul> P…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-06-30 14:11

b9848

<details open=""> CUDA: fix get_rows_back for tables with more than 65535 rows (grid-y clamp + stride) (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25103">#25103</a>) </details> macOS/iOS: <ul> <li><a href="h…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-06-30 13:18

b9847

<details open=""> CUDA: fix Gemma E4B MTP FlashAttention (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25148">#25148</a>) <ul> <li> CUDA: fix Gemma E4B MTP FlashAttention </li> <li> remove unused template declaration </…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-06-30 11:04

b9846

<details open=""> vulkan: roll bk loop in matmul for asahi linux (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24663">#24663</a>) <ul> <li> vulkan: roll bk loop in matmul for asahi linux </li> <li> vulkan: fix inline commen…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-06-30 08:58

b9844

<details open=""> ggml-webgpu: add support for NVFP4 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25143">#25143</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama.cpp/releases/down…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-06-30 01:19

b9843

<details open=""> Revert "sched : reintroduce less synchronizations during split compute (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/20793">#20793</a>)" (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/p…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-06-29 16:20

b9842

<details open=""> common : dedup preset and cached model entries in /v1/models (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25131">#25131</a>) Signed-off-by: Adrien Gallouët <a href="mailto:[email protected]">[email protected]…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-06-29 10:27

b9840

<details open=""> DeepSeek V4 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24162">#24162</a>) <ul> <li> convert: add dsv4 conversion </li> <li> add basic setup </li> <li> add llm_graph_input_dsv4 </li> <li> a…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-06-29 09:47

b9839

<details open=""> tools/ui: restore Tailwind scanning in ignored worktrees (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24879">#24879</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/l…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-06-29 07:36

b9838

<details open=""> common : remove unused regex-partial (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25118">#25118</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] · 2026-06-29 00:07

b9837

<details open=""> jinja, chat: add --reasoning-preserve flag (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25105">#25105</a>) <ul> <li> jinja, chat: add --reasoning-preserve flag </li> <li> correct help message </li> </…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-06-28 19:38

b9835

<details open=""> ui: fix stop and reasoning skip in single-model mode (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25084">#25084</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-06-28 15:34

b9833

<details open=""> chat : implement minicpm5 parser (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24889">#24889</a>) <ul> <li> Add minicpm5 tool call parser </li> <li> Refactor MiniCPM5 PEG parser per review feedback </l…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-06-28 14:20

b9832

<details open=""> jinja: add --dump-prog for debugging (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25086">#25086</a>) <ul> <li> jinja: add --dump-prog for debugging </li> <li> Update common/jinja/runtime.cpp </li> </u…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-06-28 13:40

b9831

<details open=""> spec : add DFlash support (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/22105">#22105</a>) <ul> <li> spec: add DFlash v2 support </li> <li> dflash: support sliding window attention per layer_types </li…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-06-28 11:05

b9830

<details open=""> common : allow --offline in llama download (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25091">#25091</a>) Expose the existing --offline flag to <code>llama download</code> so a script can run it to check …

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-06-28 06:48

b9829

<details open=""> logs : reduce v2 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25078">#25078</a>) <ul> <li> server : reduce logs </li> <li> cont : common </li> <li> cont : spec </li> <li> cont : CMN_ -> C…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-06-27 23:21

b9828

<details open=""> opencl: flash attention improvement (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25069">#25069</a>) <ul> <li> opencl: rework FA kernel for f16 and f32 </li> <li> opencl: flash-attention prefill prepass ke…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-06-27 12:50

b9827

<details open=""> [CUDA] Added a cudaMemcpy2DAsync fast path to ggml_cuda_cpy (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25057">#25057</a>) <ul> <li>[CUDA] Added a cudaMemcpy2DAsync fast path to ggml_cuda_cpy</li> </ul> Add a C…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-06-27 11:02

b9826

<details open=""> sycl : fix failed ut cases of norm (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25044">#25044</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama.cpp/releases/down…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-06-27 10:33

b9825

<details open=""> vulkan: fix step operator for 0 input (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25036">#25036</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama.cpp/releases/d…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-06-27 10:05

b9824

<details open=""> binaries : Improve rpc-server and export-graph-ops names. (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25045">#25045</a>) Tests are generally prefixed with -test, so rename export-graph-ops accordingly.</p…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-06-27 09:14

b9823

<details open=""> ci : add windows-openvino to check-release (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25022">#25022</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama.cpp/relea…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-06-27 08:39

b9822

<details open=""> tests : fix test-chat-template --no-common option (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25075">#25075</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama.cp…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-06-26 21:57

b9821

<details open=""> app : allow --version, --licenses & --help (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25054">#25054</a>) Signed-off-by: Adrien Gallouët <a href="mailto:[email protected]">[email protected]</a> </de…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-06-26 18:37

b9820

<details open=""> sched : reintroduce less synchronizations during split compute (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/20793">#20793</a>) <ul> <li> CUDA: Improve performance via less synchronizations between token (<a clas…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-06-26 17:29

b9817

<details open=""> openvino: Update to OV 2026.2.1, self-contained release packages, operator improvements (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24974">#24974</a>) <ul> <li> Update to OV 2026.2.1, Make OV release packages s…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-06-26 16:50

b9816

<details open=""> sync : ggml </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b9816/llama-b9816-bin-macos-arm64.tar.gz">macOS Apple Silicon (arm64)</a></li> <li>macOS Apple Silicon (arm64, KleidiAI ena…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-06-26 16:16

b9814

<details open=""> vulkan: opt mul_mat_vecq for mi50 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/22933">#22933</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama.cpp/releases/downl…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-06-26 15:40

b9813

<details open=""> vulkan: add INTEL_XE1 arch enum and enable coopmat1 on Intel Xe-LPG Plus (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24404">#24404</a>) <ul> <li>vulkan: add INTEL_PRE_XE2 arch enum and enable coopmat1 on Intel Xe-…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-06-26 13:06

b9811

<details open=""> vulkan: Workaround compiler bug in conv2d coopmat2 path (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/24924">#24924</a>) <ul> <li> vulkan: Workaround compiler bug in conv2d coopmat2 path </li> <li> apply s…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-06-26 11:46

b9810

<details open=""> CUDA: add cublasSgemmBatched mapping for HIP/MUSA vendor headers (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25033">#25033</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/gg…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-06-26 07:17

b9804

<details open=""> mamba2: remove hardcoded 2x expansion factor and invalid d_inner % d_state check (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/23082">#23082</a>) <ul> <li> mamba2: remove hardcoded 2x expansion factor, support an…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-06-26 02:28

b9803

<details open=""> opencl: flush profiling batch at shutdown for incomplete batches (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/25016">#25016</a>) </details> macOS/iOS: <ul> <li><a href="https://github.com/gg…

llama.cpp — Releases TIER_1 (SO) · github-actions[bot] · 2026-06-25 22:50

b9802

<details open=""> </details> macOS/iOS: <ul> <li><a href="https://github.com/ggml-org/llama.cpp/releases/download/b9802/llama-b9802-bin-macos-arm64.tar.gz">macOS Apple Silicon (arm64)</a></li> <li>macOS Apple Silicon (arm64, KleidiAI enabled) <a href="http…

COVERAGE [150]