Commit Graph

Select branches

Hide Pull Requests

master

Mono Color

• 8e2d31e0eb graph : reorder get_rows for embeddings (#30160) Georgi Gerganov 2026-10-09 20:43:05 +03:00
• baef3ed9a1 ui: Models Manager Follow-up Improvements (#30228) Aleksander GrygierandPascal 2026-10-09 19:28:54 +02:00
• f39148a953 llama-bench: respect -fitc if bigger than required benchmark size (#28331) Martin Emrich 2026-10-09 19:14:30 +02:00
• 50e3e3e480 CUDA: Remove redundant CUDA copies after SSM_SCAN (#29807) Anav Prasad 2026-10-09 16:23:28 +00:00
• 64df9183f5 opencl: fix kernel compilation for a6x GPUs (#30176) jingzhou 2026-10-09 09:19:20 -07:00
• 6184e92c57 model : use exact GELU for ModernBERT encoders (#30108) bosh 2026-10-09 22:26:56 +07:00
• 8b54361025 chat : refactor API (#30210) Aldehir Rojas 2026-10-09 10:06:22 -05:00
• a518119d30 llama: keep the backend sampling graph static across ubatches (#30223) Pascal 2026-10-09 16:20:25 +02:00
• 8ae386707b ggml: fix OOB write in ggml_acc with negative offset (#30135) Jasmine-tim 2026-10-09 21:57:33 +08:00
• e60eff95fd meta : handle host views (#30217) Georgi Gerganov 2026-10-09 16:01:04 +03:00
• ba6439a6b5 webgpu: use 2D workgroup dispatch for all the ops which use 1D dispatch (e.g., rms_norm) (#30219) Masashi Yoshimura 2026-10-09 22:00:37 +09:00
• 5e4878e978 vulkan: fix rms_norm workgroup count overflow (#30145) Ruben Ortlam 2026-10-09 15:00:14 +02:00
• 609290be6b convert : support compressed-tensor mixed-precision NVFP4 checkpoint (#28636) ynankani 2026-10-09 11:54:58 +00:00
• 86a2835320 musa: drop mp_21 from the default architectures (#30203) R0CKSTAR 2026-10-09 18:51:24 +08:00
• e94acad853 ci: fix Models Backend webgpu by supporting GGML_OP_DUP (#30216) Pascal 2026-10-09 12:27:32 +02:00
• 5e1d74043e ci : set run-name for publish release (#30214) Sigbjørn Skjæret 2026-10-09 11:12:24 +02:00
• b42b7e6d30 server: use port 9931 by default (#30159) Xuan-Son Nguyen 2026-10-09 10:40:03 +02:00
• b013e56a71 ui : add the models manager (#29583) Aleksander Grygier 2026-10-09 10:34:45 +02:00
• 2c564f43df ci: fix HIP quality check by ignoring the 320/256 FA kernel spill (#30212) Pascal 2026-10-09 10:29:22 +02:00
• 1f8fa52318 metal: add the 128/96 flash attention kernels, accept every compiled pair (#30209) Pascal 2026-10-09 10:15:32 +02:00
• 8a1a9b5126 CUDA: pass src1 precision to host MMQ config helpers (#30168) ynankani 2026-10-09 06:32:52 +00:00
• d4d82d67f4 hexagon: fix IM2COL patch-embed DMA ring overflow (#30189) Mark TodorovichandMax Krasnyansky 2026-10-08 23:04:09 -07:00
• 3d65c90d04 sycl : Q5_K reorder-layout MMVQ and fused GLU (#29375) Nicolas Mowen 2026-10-08 20:19:32 -06:00
• de7fa0a3c6 Musa FWHT fix (#30167) bri-prism 2026-10-08 12:56:54 -07:00
• 71ad0590f4 CUDA: improve top-k algorithm selection (#28713) 3 people 2026-10-08 22:04:27 +05:30
• a11f57ba93 model : fix DFlash output head sharing (#30111) Hrishith Thadicherla 2026-10-08 21:12:28 +05:30
• fc9ce6b9d5 CUDA: fix MMQ out-of-bounds reads (#29953) Johannes Gäßler 2026-10-08 16:41:49 +02:00
• c35b66744f CUDA : looped PAD kernel for more than 65535 rows or slices (#30147) Prabhsimran Singh 2026-10-08 19:06:53 +05:30
• 1167d3f42c CUDA: fix CCCL version guard breaking on major version rollover (#29453) hey-gmandOliver Simons 2026-10-08 14:29:34 +01:00
• 4f92965a7b ui: apply ui_settings on first visit in router mode (#29668) Simon Sudarushkin 2026-10-08 18:08:22 +05:00
• c811cb8f0a llama: support MoE cache over multiple GPUs (#30112) Aman Gupta 2026-10-08 15:50:49 +05:30
• 033df86b69 server : preserve context checkpoints across slot save/restore (#26004) Leebr Data ConsultingandIgor Okulist 2026-10-08 12:18:20 +02:00
• ff5888f999 vulkan : fix TOP_K for +inf/NaN inputs and k = 1 on negative values (#30107) gianni-cor 2026-10-08 12:14:35 +02:00
• dac3087394 vulkan: extend sparse FA support to coopmat2 (#30003) Jeff Bolz 2026-10-08 04:36:20 -05:00
• 03aa006acb vendor : update cpp-httplib to 0.60.1 (#30134) Adrien Gallouët 2026-10-08 10:57:47 +02:00
• 291dc55c16 Merge branch 'ggml-org:master' into master Ilya Maslennikov 2026-10-08 11:16:31 +03:00
• 08246a28f6 cuda : support arbitrary striding for unary ops on f16, f32, and bf16 (#29781) Muhammad SaadandJohannes Gäßler 2026-10-08 03:21:42 -04:00
• 46baf1f1fe sycl: FWHT optimizations (#29605) bri-prism 2026-10-07 23:50:22 -07:00
• 097f5b5332 hex-mmadd: do not assume aligned read/write when bias-add is fused (#30133) Max Krasnyansky 2026-10-07 23:48:06 -07:00
• d888016041 cuda : Use byte strides for roll to allow non-contiguous ROLL operations (#29547) Bertay Eren 2026-10-08 09:46:38 +03:00
• fda1866613 CUDA: fix norm family kernels when ne[2]/ne[3] exceed grid dim limits (#28175) Eden 2026-10-08 14:46:21 +08:00
• ff30363a0e sycl: fuse the delta-net alpha gate (add + unary + mul) (#29687) Titaniumtown 2026-10-07 23:45:21 -07:00
• 000bee54a5 sycl: stage bulk uploads (model loading) through a pinned ring buffer (#29608) cwriterandcwriter 2026-10-08 08:30:00 +02:00
• 37ac634566 model : support classifier_activation for rerankers (#29692) bosh 2026-10-08 13:29:20 +07:00
• 24e41838e0 ggml-cuda: assign four GDN state columns per warp (#30087) SXX 2026-10-08 14:07:34 +08:00
• 847f447c31 sycl: add grouped MoE XMX GEMM (#29245) cwriterandcwriter 2026-10-08 07:41:13 +02:00
• 75118a3a59 convert : support Qwen3.5 embedding models (#27920) Sam Malayek 2026-10-08 01:40:01 -04:00
• 9b4ed0ca57 sycl: remove duplicate block-size defines from op headers (#29507) Titaniumtown 2026-10-07 22:39:25 -07:00
• dad523e050 Revert "MoE expert cache: GPU-resident LRU cache for host-offloaded expert weights" xrip 2026-10-08 08:07:34 +03:00
• d7c332228f Merge remote-tracking branch 'upstream/master' xrip 2026-10-08 08:06:53 +03:00
• 9c2e0e491a hexagon: enable alloc_buffer_n (#30126) Max KrasnyanskyandJhen-Jie Hong 2026-10-07 18:30:06 -07:00
• 06cad0b9e7 hexagon: Q6_K weight dequant speedup (#30121) kurquhar 2026-10-07 17:13:21 -07:00
• a657f7e981 model : add LiquidAI/d1-omni-600M decision model (#30114) Tarek Dakhran 2026-10-08 02:07:49 +02:00
• aa5e0092fd hexagon: support tiled Q4_K and Q6_K GET_ROWS (#30115) kurquhar 2026-10-07 17:04:45 -07:00
• 70815103c8 hexagon: improved GELU accuracy (#30104) kurquhar 2026-10-07 16:57:09 -07:00
• bd4eeaa047 chat: fix jinja parser for TranslateGemma (#30096) Nicolas Budyn 2026-10-07 22:48:06 +02:00
• 5de733437b chat : name tool and argument parser rules by index (#30088) Frost-54andAlde Rojas 2026-10-08 06:37:51 +10:00
• 88dcc460d6 model : add LiquidAI/d1-3B decision model (#30110) Tarek Dakhran 2026-10-07 22:04:39 +02:00
• b86d2f0754 cuda: FWHT kernels for block widths above 512 (#29100) bri-prism 2026-10-07 11:56:18 -07:00
• 50a6c5cf7c mtmd: add cohere2 vision support (#30062) Terrence Zhao 2026-10-07 14:07:50 -04:00
• d6cf9acb25 llama : add a GPU cache for MoE experts kept in host memory (#29887) Aman Gupta 2026-10-07 23:37:45 +05:30
• 42c787e8c1 cuda: update uncoalesced memory reads in pool2d (#29425) ShobhitandGeorgi Gerganov 2026-10-08 01:53:43 +08:00
• 18b5f8b186 server : accumulate generated text and tokens as parse input (#29876) Aldehir Rojas 2026-10-07 11:49:17 -05:00
• ac22f2b99c Merge branch 'ggml-org:master' into master Ilya Maslennikov 2026-10-07 19:00:18 +03:00
• 448147d42a llama: share the nextn tensor flags between models (#30097) Pascal 2026-10-07 16:30:00 +02:00
• 988190680d metal : few-row MMA mat-mul for the remaining src0 types (#30065) pratiknarola-t 2026-10-07 19:19:37 +05:30
• 7e8324f5fe metal : fix MUL_MAT+ADD fusion when the residual is itself a MUL_MAT (#30100) Niko MaroulisandGeorgi Gerganov 2026-10-07 09:41:50 -04:00
• bd5ea24694 Merge branch 'ggml-org:master' into master Ilya Maslennikov 2026-10-07 15:51:44 +03:00
• b9acf138a1 feat: add GLM5Next MTP, optimize (#29928) 3 people 2026-10-07 13:16:53 +02:00
• 48499d2e1c qwen3tts : guard speaker_encoder_config patch for CustomVoice variant (#29179) SIDDARTHA REDDY 2026-10-07 06:57:48 -04:00
• d0b490f25e sampling : use greedy selection for eligible temperature-zero chains (#29797) Hrishith ThadicherlaandGeorgi Gerganov 2026-10-07 03:34:18 -07:00
• 42b021b4dc vocab : add plamo fim tokens (#30090) Sigbjørn Skjæret 2026-10-07 11:40:35 +02:00
• 7481354a17 convert : fix token configuration for PLaMo-3 (#29843) Toki Nasin 2026-10-07 18:34:16 +09:00
• ad21565331 musa: use the tile lightning indexer kernel (#30080) R0CKSTAR 2026-10-07 17:22:56 +08:00
• b7dafa01e5 vendor : update cpp-httplib to 0.60.0 (#30081) Alessandro de Oliveira Faria (A.K.A.CABELO) 2026-10-07 06:18:33 -03:00
• 26908739bc imatrix : include clocale for std::setlocale (#30079) yarikandyarik 2026-10-07 08:58:22 +00:00
• 36a73916ee ggml-webgpu: fix flash_attn supports_op check for overlapping KV (#28205) Masashi Yoshimura 2026-10-07 16:56:53 +09:00
• 26d2cfd35b Merge branch 'ggml-org:master' into master Ilya Maslennikov 2026-10-07 10:51:47 +03:00
• fa3c2fab36 tests : retain the anchor when testing recurrent rollback (#29923) Evan Huus 2026-10-07 03:44:10 -04:00
• 005a1e127a [SYCL] fix the issue in mixed different model GPUs in FA (#29071) Neo ZhangandGeorgi Gerganov 2026-10-07 15:24:43 +08:00
• d2a79e6046 sycl: accelerate GLM MLA prefill with MKL flash attention (#29171) Anant Shrivastava 2026-10-07 12:54:07 +05:30
• 5e5b628eb5 sycl : fattn_kv_buffers cleanup (#27689) Pratyush Kumar 2026-10-07 12:53:28 +05:30
• 315e0b5922 Merge branch 'ggml-org:master' into master Ilya Maslennikov 2026-10-07 10:00:17 +03:00
• 4d756bc72b vulkan: fix amd iGPU slow checkpoint read (#30049) Ruben Ortlam 2026-10-07 08:20:04 +02:00
• 78651c410d sycl: add IQ3_S multi-column MMVQ (#29500) Clemens Wasser 2026-10-07 07:28:23 +02:00
• f498f864fb ggml-webgpu: no dawn native features on wasi (#27069) Mendy Berger 2026-10-07 00:29:30 -04:00
• c479922ac5 hexagon: CPY/CONCAT/CONT/DUP overhaul to use DMA/HVX for all cases (#30067) Max Krasnyansky 2026-10-06 15:24:17 -07:00
• 5ad1c5da0a cuda : add BF16 support for XIELU (#29955) qiao_px 2026-10-07 04:36:21 +08:00
• 51ce9c11a6 ggml-cuda: use per-thread stream for buffer-init padding memset (#28782) Harkirat Gill 2026-10-06 15:04:05 -04:00
• abeada335e vocab : implement PLaMo-3 tokenizer pre-segmentation (#30045) Toki NasinandSigbjørn Skjæret 2026-10-07 02:07:54 +09:00
• 4625240437 model : add K2 Horizon dense and MoVA support (#29535) 7 people 2026-10-06 20:42:24 +04:00
• 3109914090 llama: remove the gather path of the glm5-next sparse attention (#30042) Pascal 2026-10-06 18:31:13 +02:00
• 4fbc76dec5 model: support embeddinggemma2 (text+vision+audio) (#30054) Xuan-Son Nguyen 2026-10-06 18:20:31 +02:00
• 2207c8e57c opencl: fix OOB read in adreno xmem GEMM (#30041) lhez 2026-10-06 09:15:25 -07:00
• 50238faa0d Merge branch 'ggml-org:master' into master Ilya Maslennikov 2026-10-06 18:55:09 +03:00
• a46709b683 RPC: add -sm tensor (#26610) Aman GuptaandGeorgi Gerganov 2026-10-06 20:55:39 +05:30
• 65840ed53c ggml: fix CLAMP on non-contiguous views (CPU, CUDA) (#29517) Pascal 2026-10-06 16:18:34 +02:00
• ab09ea4c14 cuda: BF16/FP16 conversion to f32 chunking (#29442) thelittlefiremanandJohannes Gäßler 2026-10-06 16:16:49 +02:00
• da263e7275 models: support pplx-decider (#30044) Xuan-Son Nguyen 2026-10-06 16:07:46 +02:00
• a043d38a62 metal : fix excess threadgroup memory in quantized flash attention (#29340) Foad Abo Dahood 2026-10-06 16:40:16 +03:00