Commit Graph

Select branches

Hide Pull Requests

master

Mono Color

• 6b790a9c29 vulkan: handle misalignment in conv_2d and conv_3d (#29365) Ruben Ortlam 2026-09-24 14:08:26 +02:00
• af084752ab Merge branch 'ggml-org:master' into master Ilya Maslennikov 2026-09-24 15:00:13 +03:00
• 3423f940e8 vulkan: tune KHR cooperative matrix support for Adreno GPUs (#29328) Raman Shinde 2026-09-24 16:38:40 +05:30
• 53ed051ce5 cuda : add conv3d with implicit GEMM (#29137) leejet 2026-09-24 15:24:57 +08:00
• f830688e91 model : add Ling 3.0 VL support (#29151) Tobyandaetherbird 2026-09-24 02:57:31 -04:00
• 2b70583997 server,common : fix the GCC 12 stringop-overread false positive (again) (#29325) Adrien Gallouët 2026-09-24 08:40:06 +02:00
• 4c5957c277 test-save-load-state : print a per-model results table in --models mode (#29316) Georgi Gerganov 2026-09-24 09:17:04 +03:00
• 0831f472a9 Merge branch 'ggml-org:master' into master Ilya Maslennikov 2026-09-24 09:05:19 +03:00
• 9710a32175 hexagon: reject MUL_MAT_ID when src1 precision is F32 (#29348) Jhen-Jie Hong 2026-09-24 13:02:17 +08:00
• 013b31c03c scripts : make-release-desc - link previous release in changelog title (#29336) Georgi Gerganov 2026-09-24 07:46:38 +03:00
• bd4f514db1 convert : allow vision target for DFlash/Dspark (#29339) Tarek Dakhran 2026-09-24 01:16:43 +02:00
• b9ae43a5d4 server: allow preset to set log file (#29334) Xuan-Son Nguyen 2026-09-24 01:16:00 +02:00
• d2e54583c7 tests: add -b/--backend option to test-llama-archs for testing a specific backend (#27372) Masashi YoshimuraandJohannes Gäßler 2026-09-24 05:27:27 +09:00
• 6e60f35608 ci : use hf-jobs-cpu-xl runner in server sanitize workflow (#29297) Georgi Gerganov 2026-09-23 21:30:39 +03:00
• fee39dd926 opencl: add A8 Q6_K non-MoE dp4a binary kernel (#29057) shaofeiqi 2026-09-23 10:37:04 -07:00
• 7fe450e193 llama.cpp : bump version to 0.5.0 (#29333) Georgi Gerganov 2026-09-23 20:32:51 +03:00
• 177cd8cc70 sync : ggml Georgi Gerganov 2026-09-23 20:28:23 +03:00
• e4e2f62325 ggml : bump version to 0.25.1 (ggml/1637) Georgi Gerganov 2026-09-23 20:10:09 +03:00
• 66fba63af1 CUDA: add a reserve to avoid spurious warning on older GCC builds (#29317) Aman Gupta 2026-09-24 00:52:40 +08:00
• bddf8263c3 common : keep HF cache dir as path, expose UTF-8 only for logs (#29320) Adrien Gallouët 2026-09-23 18:24:23 +02:00
• 3dd3ac88a8 Merge branch 'ggml-org:master' into master Ilya Maslennikov 2026-09-23 19:13:59 +03:00
• 9575389609 metal: add the missing f32 x bf16 mul_mv variants (#28741) Pascal 2026-09-23 17:29:00 +02:00
• dc9879cf66 CUDA: enable sparse-fa for dsv4 prefill (again) (#29298) Aman GuptaandPascal 2026-09-23 23:20:40 +08:00
• 0361e7c46e Merge branch 'ggml-org:master' into master Ilya Maslennikov 2026-09-23 18:03:41 +03:00
• 42916d83f4 server: fix token counting API crash on sleep (#29309) Will 2026-09-23 21:28:49 +08:00
• 4e416ee730 jinja : parse unary +/- before variables (#29244) Si Chen 2026-09-23 19:29:45 +08:00
• ee3ecce05c metal : key the fa-vec tuned table by family instead of SKU (#29075) YiChen Lv 2026-09-23 19:23:15 +08:00
• 057494f93f server: accept OpenAI video_url content type and data: video URIs (#27921) calebrio02 2026-09-23 04:58:09 -06:00
• bcbc936a87 server: Dedup the draft HF model via dedup-cache-models (#27934) Xie Wenxiang 2026-09-23 03:57:53 -07:00
• 26758d38f9 ci : fix build-cmake runner target (#29299) Sigbjørn Skjæret 2026-09-23 12:57:14 +02:00
• 18f9f7bef9 model-conversion : add causal-compare-logits recipe (#29305) Daniel Bevenius 2026-09-23 12:51:35 +02:00
• 633733d0ae model : support Gemma4 DSpark draft backbone (#29226) Hrishith Thadicherla 2026-09-23 03:34:09 -07:00
• 86b2daa730 ci : run python (jinja) test (#29302) Sigbjørn Skjæret 2026-09-23 11:49:43 +02:00
• 183d2a04c2 make-release : update summary prompt Georgi Gerganov 2026-09-23 11:36:50 +03:00
• 45062d4056 sync : ggml Georgi Gerganov 2026-09-23 11:34:47 +03:00
• 503549c5f4 ggml : bump version to 0.25.0 (ggml/1635) Georgi Gerganov 2026-09-23 11:25:40 +03:00
• e97545d916 sycl : fix compile warnings Georgi Gerganov 2026-09-23 10:44:07 +03:00
• b1ff4ca236 vulkan: add IQ4_XS MMQ/MMV matmul kernels (#28415) Piotr Wilkin (ilintar) 2026-09-23 09:00:06 +02:00
• 16c566a2f4 Merge branch 'ggml-org:master' into master Ilya Maslennikov 2026-09-23 09:54:03 +03:00
• 94256114c2 ggml-meta: resolve multi buffer views (#29266) Ruben Ortlam 2026-09-23 07:35:24 +02:00
• 1a679828f3 cuda: top-k MoE should always fire (#28432) Aman Gupta 2026-09-23 13:26:05 +08:00
• 384a534ce3 sycl : support new UT case for mul_mat_hadamard fp16 (#29218) Neo Zhang 2026-09-23 13:22:44 +08:00
• 5e48b31000 sycl: extend MMVQ GLU fusion, add rms_norm+scale and ssm_conv+silu fusions (#28931) Anant Shrivastava 2026-09-23 10:48:17 +05:30
• 4d7d7703fe sycl : support op get_rows_back, only support fp32/fp16 (#25266) Neo Zhang 2026-09-23 13:15:27 +08:00
• 08b1d2aea5 vulkan: hide internal symbols to prevent duplicate-dlopen state destruction (#29139) Erik Winter 2026-09-23 07:13:01 +02:00
• 441df11f65 sampler: reduce the size of the probe (#29285) Max Krasnyansky 2026-09-22 22:12:04 -07:00
• e6ab7c1a41 hex-dma: introduce direct-mapped DMA cache that is better suited for HVX FA mask handling (#29282) Max Krasnyansky 2026-09-22 15:20:08 -07:00
• f46bc30cb6 HIP : optimize IQ2/IQ3 (__vsub4 __vcmpne4) using SWAR (#27962) Felix Ye 2026-09-23 04:31:12 +08:00
• 709fe755df jinja : fix dangling reference warning in for_statement (#29279) Georgi Gerganov 2026-09-22 22:45:33 +03:00
• d5f66492e6 opencl: add bin kernel kernel_gemm_noshuffle_q4_k_q8_1_dp4a_ila_a8_bin (#29056) shaofeiqi 2026-09-22 12:39:11 -07:00
• 9919911185 server: fix router eviction races with the existing queue (#29217) Pascal 2026-09-22 21:38:54 +02:00
• bbf99b1b33 server: do not pass log file to children (#29212) David Friehs 2026-09-22 21:30:40 +02:00
• 4098fdc922 server: support input_image in function_call_output (#20663) (#22575) Empressia 2026-09-23 04:04:58 +09:00
• 6d6c5f1e91 Merge branch 'ggml-org:master' into master Ilya Maslennikov 2026-09-22 19:43:49 +03:00
• 4ceb171910 vulkan: add Intel Xe flash attention optimization kernels (2/3, Xe-LPG Plus/Xe2/Xe3) (#24406) Jiang, FishandLiu, Russell 2026-09-23 00:05:37 +08:00
• 73c941b111 mtmd: add various sanity checks (#29276) Xuan-Son Nguyen 2026-09-22 17:53:05 +02:00
• 8a60a0fa90 Merge branch 'ggml-org:master' into master Ilya Maslennikov 2026-09-22 18:49:28 +03:00
• 0f8a414b75 metal : gate mul_mm_id src1 rescale behind ggml_prec (#29029) Michael de Gansandyomaytk 2026-09-22 17:32:28 +02:00
• f95b0d9539 ggml : IQ1_M build prefix sums once per block (#28706) Bartowski 2026-09-22 15:54:45 +02:00
• c350a40bbd Performance tune for gemma4-26b-a4b flash attention shape. (#28450) David M. Rogers 2026-09-22 09:43:29 -04:00
• 9b421fa946 ui : Accept WEBM video files (#28622) Eric Rodrigues Pires 2026-09-22 10:40:24 -03:00
• 348f853b7a jinja: use const for statement::execute and ::visit (#29271) Xuan-Son Nguyen 2026-09-22 15:27:59 +02:00
• 217f81c266 server: Add support for binding to multiple addresses (#28690) Emanuil RusevandXuan Son Nguyen 2026-09-22 16:16:40 +03:00
• 828fdf282e spec : support DFlash for HunyuanOCR (#28890) wendadawen 2026-09-22 21:04:52 +08:00
• bfd73a876e convert: add MiMo-V2.6 support (#29257) 3 people 2026-09-22 05:38:09 -07:00
• a60f9aead0 cmake : allow repeated find_package calls for llama (#29228) miyan 2026-09-22 19:58:42 +08:00
• 7ab4ee7baa chat : Fix Muse Glimmer tool-call first parser error (#29242) Nicolas Mowen 2026-09-22 01:38:52 -06:00
• 0ee9435b8f ci : publish snapdragon builds in release workflow (#29007) Yuri Khrustalev 2026-09-22 03:35:03 -04:00
• 8cfc315a8a Add close button to UI toasts (#28246) Agustín Mista 2026-09-22 09:23:41 +02:00
• ec5a12b85a opencl: add A8 Q4_0 non-MoE dp4a binary kernel (#29055) shaofeiqi 2026-09-21 23:14:00 -07:00
• c550d2f60b ci : update Level Zero SDK to v1.33.1 and enable the L0/oneDNN CMake flags in the SYCL job (#29230) Asahi-Prv 2026-09-22 13:03:14 +09:00
• 58367713a6 hexagon: new HMX-optimized GATED_DELTA_NET (#29199) Max Krasnyansky 2026-09-21 14:49:52 -07:00
• ff0dbb975e vendor : update cpp-httplib to 0.57.1 (#29239) Adrien Gallouët 2026-09-21 22:50:20 +02:00
• ee422754be Merge branch 'ggml-org:master' into master Ilya Maslennikov 2026-09-21 22:31:16 +03:00
• c3530049be test-save-load-state : compare logits with NMSE and feed expected tokens (#29238) Georgi Gerganov 2026-09-21 20:19:11 +03:00
• 704bb2f222 build: add SM75 dense and MoE scripts xrip 2026-09-09 12:04:40 +03:00
• e14d7024f7 sm75: record tests and memory results xrip 2026-09-09 11:37:58 +03:00
• 6380ef7f3b llama: share SM75 MTP compute buffers on request xrip 2026-09-09 11:36:37 +03:00
• 3a90a98e7a llama: add opt-in SM75 recurrent state replay xrip 2026-09-09 11:36:37 +03:00
• 5fe543b020 cuda: add opt-in SM75 SwiGLU input fusion xrip 2026-09-09 11:36:08 +03:00
• e73342ebc1 bench: record the shared-codebook result (+5.98% tg on 27B IQ2_XXS) xrip 2026-09-08 19:55:57 +03:00
• 68769090cb cuda: gate IQ1 shared codebook by ncols_dst instead of excluding it xrip 2026-09-08 19:35:07 +03:00
• c25b32c556 cuda: exclude IQ1_S/IQ1_M from the SM75 shared-codebook path xrip 2026-09-08 19:11:08 +03:00
• 1536af3b94 cuda: shared-memory codebooks for all grid quants on SM75 xrip 2026-09-08 18:31:14 +03:00
• ab178e31d0 cuda: stage the IQ2_XXS codebook in shared memory on SM75 (experiment) xrip 2026-09-08 18:11:10 +03:00
• 61b1b6e98b Revert "wip: probe Turing nwarps for low-bpw types (experiment, do not merge)" xrip 2026-09-08 18:10:08 +03:00
• dbe732c0a6 wip: probe Turing nwarps for low-bpw types (experiment, do not merge) xrip 2026-09-08 17:17:17 +03:00
• 0b853410d8 bench: record IQ2_XXS decode-reuse negative result xrip 2026-09-08 14:54:19 +03:00
• a3336c9495 bench: journal of SM75/Ivy Bridge measurements xrip 2026-09-08 10:46:58 +03:00
• 622576f2ec top-k: single GGML_CUDA_USE_TOP_K_RADIX macro (review) rhonstin 2026-09-06 08:42:16 +00:00
• f56cf8ff73 CUDA: use radix TOP_K when CUB DeviceTopK is unavailable Rhonstin 2026-09-04 07:50:38 +00:00
• 23cbecae91 cuda: do not apply the Turing mmvq thresholds under GGML_CUDA_FORCE_CUBLAS xrip 2026-09-07 20:56:35 +03:00
• 85ff4657cb cuda: tune the mmvq/mmq crossover for Turing xrip 2026-09-07 20:24:30 +03:00
• 6f65493dde cuda: drop SM75 Q4_K decode reuse, superseded by upstream branchless unpack xrip 2026-09-07 19:21:09 +03:00
• 12877b6224 cuda : select SM75 prefill cuBLAS path; document optional 25% build gain xrip 2026-09-07 09:25:56 +00:00
• 70c9f4fd46 ggml-cuda: use two warps for SM75 IQ4_XS N=1 xrip 2026-09-07 09:02:02 +03:00
• 7b19064aae ggml-cuda: reuse Q4_K decode across SM75 token columns xrip 2026-09-07 09:02:01 +03:00
• 53404dd5a6 ggml-cuda: fuse and reuse IQ4_XS weights on SM75 xrip 2026-09-07 09:02:00 +03:00
• 0166cf95b9 ggml-cpu: add opt-in AVX1 IQ3_XXS panel kernels xrip 2026-09-07 09:02:00 +03:00
• ec192a9ca6 ggml-cpu: vectorize AVX1 Q8_Kx4 packing xrip 2026-09-07 09:01:59 +03:00