llama.cpp-xrip

wmantly/llama.cpp-xrip

0 Fork 0

16a020ce1e ggml-cpu: add AVX1 Q4_K and IQ4_XS paths xrip 2026-09-06 16:24:17 +00:00
337be720fc Fix prefill speed degradation Denis Artyomov 2026-08-31 08:19:44 +05:00
007188b6bd Move the DFlash2 selector out of the graph to save VRAM (without reducing the speed of infence) Denis Artyomov 2026-08-31 00:30:44 +05:00
81cdd3ce48 limit DFlash2 draft ubatch to reduce compute buffer Denis Artyomov 2026-08-30 20:43:30 +05:00
d00e039632 fix DFlash2 with tensor split (assisted by DeepSeek v4 Flash) Denis Artyomov 2026-08-28 12:02:24 +05:00
e49563cdf8 ggml-cuda: use DP4A offset for all Q6_K MMVQ paths xrip 2026-09-03 13:39:06 +00:00
255a4421fd ggml-cuda: use DP4A offset for Q6_K on Turing xrip 2026-09-03 13:24:25 +00:00
00fd692403 Revert "ggml-cuda: reduce chunked AllReduce launch width" xrip 2026-09-03 13:11:46 +00:00
83629f2632 ggml-cuda: reduce chunked AllReduce launch width xrip 2026-09-03 12:17:46 +00:00
982f36c66c ggml-cuda: speed up one-column Q6_K MMVQ on Turing xrip 2026-09-03 11:53:44 +00:00
0338c487f3 ggml-cuda: pack IQ4_XS header loads on Turing xrip 2026-09-03 11:26:27 +00:00
35aa76ba18 ggml-cuda: use wide IQ4_XS loads on Turing xrip 2026-09-03 09:09:02 +00:00
b96d5c8267 ggml-cuda: add hybrid AllReduce mode xrip 2026-09-03 07:16:53 +00:00
7391a178e6 MoE expert cache: GPU-resident LRU cache for host-offloaded expert weights csantiago78andClaude Fable 5 2026-08-28 02:36:40 -04:00
302bd55b5f qwen4exp: direct reads for the lazy PLE table (--lazy-mode on-direct) Josh Leverette 2026-08-31 20:30:06 -05:00
5a459773e3 ggml: add AVX1 IQ4_XS panel kernels xrip 2026-09-01 12:12:26 +00:00
112667913f ggml-rpc: add NUMA split server support xrip 2026-09-01 12:17:44 +03:00
e09889fb79 ggml : add AVX1 Q4_K kernels and prefill xrip 2026-08-31 11:59:44 +03:00
fb34fc262c metal : fix mask bounds in flash attention block pre-pass (#29220) Foad Abo Dahood 2026-09-21 20:31:56 +03:00
c641dfa833 test-save-load-state : compare logits with NMSE and feed expected tokens (#29238) Georgi Gerganov 2026-09-21 20:19:11 +03:00
9655061365 llama-context : report graph inputs and input tensors during sched reserve (#26625) Georgi Gerganov 2026-09-21 19:13:04 +03:00
b1c2863e2c cuda: fix sm_70 tile compilation error (#29224) lingyezhixing 2026-09-22 00:11:29 +08:00
f4e276a206 ggml-cuda : convert contiguous tensors four elements at a time (#29155) Piotr Wilkin (ilintar) 2026-09-21 18:00:51 +02:00
e6cef8152f cuda : accelerate conv2d with implicit GEMM (#29135) leejet 2026-09-21 23:11:43 +08:00
c21284cdf5 ggml : fix dimension and stride truncation in ggml_permute (#29227) leejet 2026-09-21 22:25:44 +08:00
6f41ac59e0 vendor : update cpp-httplib to 0.57.0 (#29214) Adrien Gallouët 2026-09-21 13:44:43 +02:00
ec91ab5add docker : bump cuda to 13.4.1 (#29207) Sigbjørn Skjæret 2026-09-21 13:03:41 +02:00
bb3c853c30 sycl : support gated DSV4_HC_PRE and optional HC_POST comb matrix (#29132) cwriterandcwriter 2026-09-21 12:59:38 +02:00
af911149c5 sycl : pinned memory use right device context instead of 0 (#28895) Łukasz Ślusarczyk 2026-09-21 12:58:59 +02:00
1884824fda CUDA: Follow up of #25635, refactoring FA shared smem swizzle (#28536) ynankani 2026-09-21 10:58:28 +00:00
161755f29e test-llama-archs : make tensor data stdev configurable and improve help (#29133) Georgi Gerganov 2026-09-21 13:57:36 +03:00
1d72b05d38 tests/test-backend-ops : allow regex entries in the -o filter (#29204) Georgi Gerganov 2026-09-21 13:57:19 +03:00
542e9202d7 ci : refactor build-self-hosted into backend-specific workflows (#28991) Sigbjørn Skjæret 2026-09-21 12:51:53 +02:00
e0dff58475 args: add env vars for temperature, top-p, min-p and penalties (#27380) Mikolaj Kucharski 2026-09-21 10:47:38 +00:00
982a3329af server : do not forward --api-key-file to router-spawned child instances (#28938) Nandan Vallamdasu 2026-09-21 16:15:58 +05:30
711f60beeb tests : remove stale comment (#29140) Mostafa 2026-09-21 13:38:48 +03:00
335b21fcbd ggml-metal : simplify fusion pattern op list declaration (#29206) Georgi Gerganov 2026-09-21 12:37:24 +03:00
26394b4e67 json: Fixed json enum handling (#28518) Silverside 2026-09-21 03:32:07 -05:00
1aa2954bde sycl : coalesce MKL-FA softmax loads instead of one work-item per row (#28918) Anant Shrivastava 2026-09-21 13:37:04 +05:30
8034c1d1f1 ggml-cpu: ARM Repack kernels for Q1_0 (#23492) pl752andCopilot Autofix powered by AI 2026-09-21 13:04:51 +05:00
6ad1af5603 ci : Upgrade CUDA to 13.4 for Ubuntu CUDA Release Builds (#29202) Samriddha Sinha 2026-09-21 13:32:40 +05:30
0c3626ec06 hexagon: overhaul of buffer and DMA handling to support 64bit mappings + improvements (#29197) Max Krasnyansky 2026-09-21 01:00:28 -07:00
68d9053afd cuda : tune MMVQ to MMQ crossover for SM70 (Volta) (#28912) Yangyu ChenandJohannes Gäßler 2026-09-21 15:45:31 +08:00
8aa161b54a metal : fix deprecation warnings from macOS 27 SDK (#29136) Niklas Wenzel 2026-09-21 09:44:40 +02:00
932a68e068 webgpu : add fused gdn + cpy (#28976) Masashi Yoshimura 2026-09-21 16:39:30 +09:00
62668d6b26 convert: enable --fuse-qkv for muse-glimmer (#29203) David Friehs 2026-09-21 09:38:28 +02:00
ce8caa6e60 CUDA: tune FA for Gemma 4 on Ampere or newer (#29152) Johannes Gäßler 2026-09-20 22:20:12 +02:00
a894dae939 metal : support arbitrary hc in dsv4_hc_pre (#29169) Georgi Gerganov 2026-09-20 17:52:20 +03:00
3d82ef62d4 common/peg : handle invalid utf-8 sequences in the AST (#29161) Aldehir Rojas 2026-09-20 06:51:40 -05:00
3cf03257f2 CUDA: enable sparse fa for qwen4 (#28770) Aman Gupta 2026-09-20 16:08:11 +08:00
b23efaa2ef ui: Fix mobile breakpoint + content overflow issues (#29108) Aleksander GrygierandPascal 2026-09-20 07:59:43 +02:00
4260903678 fix(mamba) : make time-step projection input contiguous (#28832) Andrei 2026-09-20 00:58:14 -04:00
9a9f939b80 metal: add F16 input to the FWHT (#29094) bri-prism 2026-09-19 21:57:30 -07:00
f072b10371 chat : fix gemma4 required tool grammar (#29115) Aldehir Rojas 2026-09-19 18:59:43 -05:00
59657a613a chat : add dedicated Ling 3.0 (Bailing V3) parser (#28682)
Toby aetherbird Alde Rojas
2026-09-19 19:35:44 -04:00
e613ef2c81 hexagon: enable I32 GET_ROWS (#29116) Aparna M P 2026-09-19 22:18:31 +05:30
851cb34f21 hexagon: add support for GEGLU_QUICK (#29114) Aparna M P 2026-09-19 22:18:07 +05:30
7d4b92bb9b hexagon: enable support for TOP_K op (#29113) Aparna M P 2026-09-19 21:46:03 +05:30
1af554f8fc server : improve startup log messages (#29125) Georgi Gerganov 2026-09-19 15:37:13 +03:00
eb1e1f495f json-schema : accept escaped hyphen in regex patterns (#29127) chiheb ben cheikh 2026-09-19 13:11:13 +01:00
5b59b83f4e metal : add MoE and SSM_CONV fusion optimizations (#28948) Georgi Gerganov 2026-09-19 13:14:44 +03:00
60b06ab9a9 metal : fix FA support checks (#29122) Georgi Gerganov 2026-09-19 11:33:03 +03:00
efa28e950e test-llama-archs : generate dummy test vocab (#29084) Georgi Gerganov 2026-09-19 11:27:46 +03:00
59fc5a1ca3 metal : support qwen4exp hc ops (#29000) Georgi Gerganov 2026-09-19 11:27:30 +03:00
b23701f77d cuda : fix CUB argsort corruption caused by in-place keys (#28389)
TheArchitectit Claude Opus 4.6 Oliver Simons
2026-09-19 00:32:52 -05:00
60081bb2b5 opencl: add support for bin kernel flash_attn_f32_f16_bin (#29046) dsproule 2026-09-18 16:32:31 -07:00
2b1847030c hexagon: add ROLL op support (#29105) Todor Boinovski 2026-09-18 15:05:10 -07:00
50631b3d2c hexagon: im2col update (#29103) Todor Boinovski 2026-09-18 14:20:48 -07:00
18a04f09c2 hexagon: HMX flash-attention head_dim padding (support DK=DV=72) (#26539) Todor Boinovski 2026-09-18 13:15:08 -07:00
ec92815050 opencl: add bin kernel kernel_gemm_noshuffle_q6_k_f32_32b_trans_ila_a8_bin (#28678) shaofeiqi 2026-09-18 10:50:15 -07:00
4fea119de3 ggml-cpu: add F16 input to the FWHT (#27779) bri-prism 2026-09-18 07:17:38 -07:00
5b335f413e ggml : check for allocation failures to prevent crashes (#28149) Alexey Kopytko 2026-09-18 22:58:25 +09:00
542348a35c Model-Saver: Write the SWA pattern, 15 more architectures roundtrip (#29042) Pascal 2026-09-18 15:20:03 +02:00
d663dd3f3a ci: change ubuntu-latest to ubuntu-24.04 (#29079) Aaron Teo 2026-09-18 21:17:19 +08:00
44be98f057 ggml-webgpu: fix supports_op condition for GET_ROWS (#28978) Masashi Yoshimura 2026-09-18 20:47:07 +09:00
911f6cdc8a ggml : handle graph buffer reservation failure (#26070) z 2026-09-18 17:31:19 +08:00
bbd488c42a vulkan: add IQ3_S MMQ matmul kernels (#28822) Daniel Varga 2026-09-18 10:46:02 +02:00
dc85f89c7e vocab : add ufakzeka pre-tokenizer (#29033) Sait Furkan Teke 2026-09-18 11:45:11 +03:00
8ed1a55efc cmake : fix build when GGML_CPU=OFF and GGML_CUDA=ON (#29026) Nikita Gordeev 2026-09-18 15:44:03 +07:00
bb11ebb682 gguf-py: fix Q8_1 block size in GGML_QUANT_SIZES (2+2+32) (#29036) yanghongandSigbjørn Skjæret 2026-09-18 16:43:06 +08:00
f03cf3e9b8 ci : disable GHA cache for copilot (#29068) Sigbjørn Skjæret 2026-09-18 10:20:53 +02:00
bdcbaaf6e7 ci : bump android-actions/setup-android to 4.0.4 (#29065) Sigbjørn Skjæret 2026-09-18 09:04:14 +02:00
5c53396b89 vulkan: raise the hoisted row-id limit for mul_mat_id from 256 to 512 experts (#28501) drluoto 2026-09-18 09:00:15 +02:00
972d2313bc ci : add missing evict-old-files (#29041) Sigbjørn Skjæret 2026-09-17 19:05:40 +02:00
c77ae695c9 rpc : skip ACCEL devices (#29020) Pedro Cuenca 2026-09-18 00:36:53 +09:00
b49650adb3 model : skip gate_up_exps if TENSOR_SKIP is set (#29014) David Friehs 2026-09-17 13:53:54 +02:00
7076180486 model : extend Nemotron MTP support (#29018) Kartik Gulia 2026-09-17 17:22:36 +05:30
ebbb185227 openvino : Update OpenVINO to 2026.4;fix clangd,MSVC warnings; (#29009) Ravi PanchumarthyandMostafa Faheem 2026-09-17 03:46:14 -07:00
4ff829ec2e ui: fix removed reasoning menu in single model mode on desktop (#27985)
Benjamin Babik Ben Babik Pascal
2026-09-17 10:18:13 +01:00
f172be756a vulkan: split buffers and debug code into separate files, add shared headers (#28732) Ruben Ortlam 2026-09-17 10:16:35 +02:00
87f9c82f2d ci : add API/ABI check to make-release workflow [no ci] (#28947) Daniel Bevenius 2026-09-17 09:55:16 +02:00
7f6f0c2a9d chat : add message delimiters to the DeepSeek V3.2/V4 parser (#29008) midagedevandSigbjørn Skjæret 2026-09-17 16:41:08 +09:00
81aeaeb74b gguf : align the data section relative to the GGUF start, not the file (#28993) Yuri KhrustalevandJohannes Gäßler 2026-09-17 03:19:44 -04:00
c9a5eeeb34 sycl : fix the B70 mem allocate error when >19.3GB (#28953) Neo Zhang 2026-09-17 14:56:03 +08:00
7490357f22 vulkan: skip unneeded MoE work in mul_mm coopmat1 path (#25483) Jiang, Fish 2026-09-17 14:53:07 +08:00
817e5f83eb sycl: ssm_conv: fuse the SiLU epilogue into the ssm_conv kernel (#28929) Titaniumtown 2026-09-16 23:51:50 -07:00
c57da6fd81 opencl: fix various warnings (#28984) lhez 2026-09-16 23:48:05 -07:00
79bfc1d43a docs: remove JG as CODEOWNER for test-llama-archs (#29003) Johannes Gäßler 2026-09-17 08:14:58 +02:00
05f2dcfdba vulkan: fix buffer_reference alignment in im2col shaders (#28996) Abir Deol 2026-09-16 22:14:33 -07:00
35822afe58 vulkan: support qwen4exp hc ops (#28988) Ruben Ortlam 2026-09-17 06:34:23 +02:00

... 9 10 11 12 13 ...

Commit Graph Select branches Hide Pull Requests master Mono Color

Commit Graph

Select branches

Hide Pull Requests

master