vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-10-23 14:50:00 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Commit graph

  • 6475b2f6b7 added support for taeqi tinier concedo_experimental Concedo 2026-10-23 18:20:17 +08:00
  • 99bfa84a7d fix for rocm search path Concedo 2026-10-23 00:20:45 +08:00
  • f7eac25dbc fix for https://github.com/LostRuins/koboldcpp/issues/2552 Concedo 2026-10-23 00:07:14 +08:00
  • dff7b54b77 Merge commit '1b29621bac' into concedo_experimental Concedo 2026-10-22 16:20:34 +08:00
  • c7fc509417
         opencl: add F32-to-I32 ARGMAX (#30897) upstream kurquhar 2026-10-21 18:02:50 -07:00
  • a37cc46104
         models : add Flint model (FlintForCausalLM) (#30928) alexey-rsi 2026-10-21 23:26:22 +02:00
  • 99fb997184
         jinja : fix namespace attributes shadowed by dict methods (#30941) Tarek Dakhran 2026-10-21 23:06:56 +02:00
  • 7775ffb7e4
         model: Add support for SenseNova U1 model (#28919) Maphist0 2026-10-22 03:54:16 +08:00
  • d454a204ef
         llama: create new fused op for QSA mask (#30640) Aman Gupta 2026-10-22 01:02:22 +08:00
  • 0756e83fc1
         sycl: q8 fast dequant (#30775) 92es 2026-10-21 10:01:03 -07:00
  • 9febe63d5a
         jinja: support hexadecimal string escapes (#30929) alexey-rsi 2026-10-21 18:04:36 +02:00
  • 9b56ccf00d
         ggml-opencl : add IM2COL_3D support (#30895) kurquhar 2026-10-21 07:59:55 -07:00
  • cbd7e272f2 allow reading vocab from gemma (ltx) and umt5 (wan) gguf, no longer bundle them Concedo 2026-10-21 22:08:03 +08:00
  • 0a1ef309cd
         mtmd: add state object for mtmd_gen (#30902) Xuan-Son Nguyen 2026-10-21 12:59:54 +02:00
  • 261700b896
         server : forward llama.cpp sampling params on /v1/messages (#30856) David Garcia 2026-10-21 06:58:52 -04:00
  • 01e3b5032c allow using tokenizers baked into the gguf llm as a last resort Concedo 2026-10-21 17:57:28 +08:00
  • f849095365
         vendor: update BoringSSL to 0.20261020.0 (#30889) Alessandro de Oliveira Faria (A.K.A.CABELO) 2026-10-21 06:15:54 -03:00
  • 824eaecf48
         llama : optimize vocabulary loading (#30679) Evgen Mykhailenko 2026-10-21 05:03:27 -04:00
  • eb182d85c0
         llama : place d2t by its set_rows use so it can stay on the GPU (#30912) Aman Gupta 2026-10-21 14:02:29 +05:30
  • ea9e3446b5
         metal : fix memory leaks and cleanup logic in GPU capture path (#30890) Niklas Wenzel 2026-10-21 09:54:33 +02:00
  • 0efe8e4c6b
         hexagon: neon-vectorized and threaded repack (tiling) (#30901) Max Krasnyansky 2026-10-21 00:48:39 -07:00
  • b531e4c749
         opencl: add dp4a GEMM variants to vector load activation (#30753) lhez 2026-10-21 00:48:13 -07:00
  • 74b115ca6e
         sycl : fix fa crash for Query head dimension mismatch (neq0 != HSK / D=2048/4096) (#30286) Neo Zhang 2026-10-21 15:46:43 +08:00
  • 96caa80059
         sycl: collapse duplicate need_check launch bodies in mmq (#30739) Titaniumtown 2026-10-21 00:46:00 -07:00
  • 479a52779f
         ggml-cpu : compute f32 gelu_quick and geglu_quick without the fp16 table (#30737) bosh 2026-10-21 14:34:37 +07:00
  • 7a94ebf4c3 tool call improvemetns Concedo 2026-10-21 14:52:18 +08:00
  • 8baabd53b7
         llama : scale moe down projection per expert token (#30861) Georgi Gerganov 2026-10-21 09:01:00 +03:00
  • 493253fe5a
         Sycl iq3 crash fix (#30604) Anant Shrivastava 2026-10-21 10:34:13 +05:30
  • bb77c8a33d widen id range to reduce collision, better handling for named tool requests Concedo 2026-10-21 11:38:44 +08:00
  • 339f3242c7
         opencl: fix rope program build error with newer Clang (#30893) Ben Ashbaugh 2026-10-20 17:19:08 -07:00
  • 9a3b91c680
         server : fix usage.completion_tokens undercount with n > 1 (#29486) Unbug Lee 2026-10-21 07:19:54 +08:00
  • 5c0b49a23b
         server : add POST /tts endpoint for TTS models (#26603) Xuan-Son Nguyen 2026-10-21 01:19:42 +02:00
  • 1242be66f6
         server: add realtime transcription API (#30636) Xuan-Son Nguyen 2026-10-21 00:04:15 +02:00
  • 84e66d0321
         refactor : move LFM2 shortconv and FFN blocks into a shared LFM base (#30874) Tarek Dakhran 2026-10-20 23:17:13 +02:00
  • 52ffa02745
         model : support Nemotron-3(.5)-Embed (#30877) Takamichi Ohara 2026-10-21 05:48:45 +09:00
  • 839a572a4c
         qwen: reshape output to 2d instead of 3d (#30823) Aman Gupta 2026-10-21 00:39:56 +05:30
  • ead9f32058
         server: Support logprobs with speculative decoding (#27196) Nigel Bosch 2026-10-20 12:36:14 -05:00
  • b684b498b6
         hexagon: fix VTCM size query on images that need layout out-params (#30655) Tarek Dakhran 2026-10-20 19:22:08 +02:00
  • 4a45de4f2c
         mtmd : lfm2 tiles and layout match transformers (#30261) Yuri Khrustalev 2026-10-20 12:54:46 -04:00
  • 7e3eeee048
         model : add H2O-Lightning-4B decision model (#30361) PSEUDOTENSOR / Jonathan McKinney 2026-10-20 09:54:27 -07:00
  • 5ba3b9fd3d
         mtmd: fix gemma 4 ffn act (#30882) Xuan-Son Nguyen 2026-10-20 18:33:32 +02:00
  • 5a9a34d888
         server : avoid repeated mmproj memory margin addition on sleep/resume (#28233) TANGBUDU 2026-10-21 01:32:12 +09:00
  • 39da130fcd
         ggml-cuda : support ssm_conv kernel size 7 (#30873) Tarek Dakhran 2026-10-20 17:47:44 +02:00
  • 4e201388c3
         server: Router wasn't propagating the client disconnects fully to the child, the response wait could cause a delayed cancellation (#29707) Me42019 2026-10-20 10:43:45 -05:00
  • 5c13e63c66 paralleltotalcontext flag Concedo 2026-10-20 23:09:06 +08:00
  • 1ba568c5ad
         server: fix missing output_index in Responses API SSE events for compatibility (#30863) Renzis 2026-10-20 22:17:06 +08:00
  • 36140dbee2
         mtmd : read LFM2 tiling params from GGUF metadata (#30840) Tarek Dakhran 2026-10-20 16:05:30 +02:00
  • df11702136
         metal : optimize argmax and pool_1d (#30862) Georgi Gerganov 2026-10-20 16:46:00 +03:00
  • 3c28d676c0 compact some embedded json files Concedo 2026-10-20 21:21:24 +08:00
  • 9df0b73c70
         model : drop the redundant precision overrides in the Kolibri-1 router (#30865) Georgi Gerganov 2026-10-20 15:58:09 +03:00
  • b38960f520
         ci : remove superfluous webgpu jobs (#30868) Sigbjørn Skjæret 2026-10-20 14:27:55 +02:00
  • a01ce9ffc6
         opencl: add several gemm variants for small batches and some fixes (#30850) lhez 2026-10-20 04:37:28 -07:00
  • 8115b4c31f
         sycl : replace memset due to not always ordered (#30812) Neo Zhang 2026-10-20 19:35:48 +08:00
  • 119f68d7d4
         ggml-cpu : select the NEON fp32->fp16 conversion (#30505) Bernard Ladenthin 2026-10-20 13:35:06 +02:00
  • b684498579
         HIP/CUDA: Fix reduction in mmf getting tile widths not devisiable by 64 (#30836) uvos 2026-10-20 13:32:38 +02:00
  • 522f2b6eed
         ggml-zdnn: introduce flash attention op (#30512) Aaron Teo 2026-10-20 19:31:35 +08:00
  • 83c2d012a2
         spec : save the dynamic ngram cache file (#22055) Peter Sideris 2026-10-20 14:31:10 +03:00
  • 13786fbb3a
         CUDA: Clear CUDA graphs after legacy memory pool flush (#26771) Bálint Bokros 2026-10-20 13:30:50 +02:00
  • d80619397c
         server : pass id_slot through the Anthropic /v1/messages conversion (#28554) Peter van Velzen 2026-10-20 13:23:41 +02:00
  • 704366d76f
         metal : fix missing MoE kernel variant for 9 experts per token (#30610) (#30699) Caladrel 2026-10-20 06:02:05 -05:00
  • 707616feb5
         ci: Fix --strict mode for API check (#30623) Christian Kastner 2026-10-20 12:38:45 +02:00
  • 5a7f5c500e cmake fix Concedo 2026-10-20 17:46:02 +08:00
  • aa034e1004
         sd: sync with master-975-95b3090 (#2542) Wagner Bruna 2026-10-20 06:35:33 -03:00
  • d8d92973c8
         vocab : add \p{M} to lookaheads in Tekken and GPT-4o pre-tokenizer regexes (#30845) tkran2 2026-10-20 03:37:46 -05:00
  • 7879491cf4
         tests : make MoE router selection independent of the backend noise (#30818) Georgi Gerganov 2026-10-20 10:40:15 +03:00
  • 8466395353
         hexagon: update to snapdragon-toolchain v0.8 and enable hexagon v85 (#30854) Max Krasnyansky 2026-10-19 23:49:59 -07:00
  • dd002b91f5
         sycl: fix compiler warnings - missing prototypes, dead code, extra semicolons (#30271) Michael Taylor 2026-10-20 01:47:25 -03:00
  • 1d88cb05f1
         ui: hide the native clear button in SearchInput (#30719) aiqubits 2026-10-20 11:58:38 +08:00
  • 3bb63cbcbf
         sycl : fix get_rows on reordered quantized weights (#30430) Sam Murdoch 2026-10-20 10:47:21 +11:00
  • f711c8f303
         mtmd : match the transformers Qwen-VL resize (#30736) bosh 2026-10-20 06:14:40 +07:00
  • 921257f500
         ci, docs: add WASI CI and docs (#30754) Mendy Berger 2026-10-19 19:09:55 -04:00
  • 3ffb62935b
         ci : re-enable type/lint-check for skills (#30713) Sigbjørn Skjæret 2026-10-20 01:03:03 +02:00
  • 7adf160c41
         server: add --connect, allow p2p access to server outside LAN (#28277) Xuan-Son Nguyen 2026-10-20 00:23:14 +02:00
  • ab7ed0d4b5
         ui : fix attachment text preview clipped when a line is longer than the container (#30806) A-Raphie 2026-10-19 23:22:14 +01:00
  • 4c3c5810f8
         hybrid dna tokenizer fixes (#30825) Kashif Rasul 2026-10-19 23:01:40 +02:00
  • 1b29621bac
         opencl: add cooperative-k dp4a gemm for q4_0, q4_k and q6_k (#30741) lhez 2026-10-19 12:30:39 -07:00
  • 6fa4f6cd3e
         opencl: add get_rows for i32, cpy for i32 and f32 (#30756) lhez 2026-10-19 12:22:34 -07:00
  • 0ce8cd1141
         jinja : implement unique filter (#30819) Aldehir Rojas 2026-10-19 12:18:24 -05:00
  • 7f5372aed5 improve kv allocation when using parallelrequests, allocates more kv but in a diminishing way Concedo 2026-10-20 00:20:25 +08:00
  • 7c97f0148e Merge branch 'upstream' into concedo_experimental Concedo 2026-10-19 22:48:10 +08:00
  • f174553e16
         server : fix swa checkpoint invalidation logic (#30666) Tarek Dakhran 2026-10-19 15:58:14 +02:00
  • 312cb1eabc partial revert of https://github.com/ggml-org/llama.cpp/pull/30449 - prevent sdcpp large vae regressions Concedo 2026-10-19 21:40:02 +08:00
  • 3bc4cc5946 updated lite Concedo 2026-10-19 21:35:04 +08:00
  • 21b12cc8db
         sd: detect file type by content instead of extension (#2548) Wagner Bruna 2026-10-19 10:00:29 -03:00
  • 7951f601e8
         vendor : update cpp-httplib to 0.62.0 (#30785) Alessandro de Oliveira Faria (A.K.A.CABELO) 2026-10-19 09:23:23 -03:00
  • 33d6135cbe
         moe-cache : replace MoE LRU cache with a 50/50 SLRU cache (#30637) Evan Huus 2026-10-19 07:28:50 -04:00
  • 08fb3a6d87
         build: use -Ofast only for legacy ggml (#2546) Wagner Bruna 2026-10-19 06:55:16 -03:00
  • 098ce526b5
         metal : fix moe_reduce fusion race when dst aliases its inputs (#30795) Georgi Gerganov 2026-10-19 11:55:02 +03:00
  • f3fb2fcdbc
         cann : refactor #ifdef blocks to avoid unreachable code after return (#25273) Raul Torres 2026-10-19 09:17:54 +01:00
  • fda00f18cd
         tests: add big-endian conversion tests (#30705) Aaron Teo 2026-10-19 16:14:58 +08:00
  • 4866db350d
         CUDA: fix warning for nonexistent extra semicolon (#30774) Johannes Gäßler 2026-10-19 10:08:21 +02:00
  • 42b70032b1
         ggml-cpu: vectorize ggml_v_silu and ggml_v_expf (#30791) Aaron Teo 2026-10-19 15:59:49 +08:00
  • 4c1004c492
         cuda: cap MMQ Stream-K blocks at available work (#30553) Joel 2026-10-19 03:41:14 -04:00
  • d3c1b40be1 use 2.0 swa warmup multiplier Concedo 2026-10-19 15:25:16 +08:00
  • 8a4a15c854 improvement to swa warmup Concedo 2026-10-19 15:25:01 +08:00
  • f8e027021f remove forced 512 swa padding Concedo 2026-10-19 14:33:51 +08:00
  • 412a363c65
         qwen3 : bidirectional attention and dense_2 output projection (#30607) fzowl 2026-10-19 08:29:02 +02:00
  • 2b57148b08
         ui: Fix markdown tables and lists being cut off on phones (#30687) Aleksander Grygier 2026-10-19 08:28:16 +02:00
  • 8f97d82da3
         tools : reject truncated GGUF tensor data in split and export tools (#30497) ppAlexa 2026-10-19 11:57:46 +05:30
  • 27af58168a
         tests : update Metal fusion baseline for beam (#30764) Georgi Gerganov 2026-10-19 09:11:38 +03:00