llama.cpp-xrip

wmantly/llama.cpp-xrip

0 Fork 0

6d78fb0727 llama : fix init in several tools/examples (#29632) Georgi Gerganov 2026-09-29 10:32:05 +03:00
18bbc46b46 metal: FWHT perf optimizations (#29602) bri-prism 2026-09-29 00:29:51 -07:00
0bc845d356 vulkan : reuse descriptor sets when bindings are constant (#29280) Andrew Lee 2026-09-28 23:54:55 -07:00
139997d8e7 chat : fix Muse Glimmer ignoring response_format json_schema with --jinja (#29615) vaibhavdedhiaandAlde Rojas 2026-09-29 11:48:58 +05:30
76a5bc86d1 common : use fs::path for cache dirs (#29595) Adrien Gallouët 2026-09-29 07:24:03 +02:00
46e17a6352 tests : skip pytest workers when PYTEST_WORKERS=1 (#29610) Georgi Gerganov 2026-09-29 08:20:54 +03:00
fc07d781e6 ci : update the oneAPI toolkit to 2026.1 (#29273) Asahi-PrvandAsahi-Prv 2026-09-29 11:19:51 +09:00
526c43b8f7 mtmd: fix GCC 15 stringop-overflow in decode_embd_batch (#29607) Pascal 2026-09-29 01:21:37 +02:00
1c4729414d hex-scripts: show trace events smaller than 100nsec in perfetto (#29614) Trivikram Reddy 2026-09-28 17:02:58 -05:00
680a036285 server : support typed content (vision/audio/video) input for /v1/embeddings endpoint (#29556)
Tim Wang timothywang21 Xuan Son Nguyen
2026-09-28 15:40:38 -04:00
66e665c427 vulkan: include functional header (#29597) Linux User 2026-09-28 19:31:42 +00:00
57b557cb95 models: pad on the left with ggml_pad_ext (#29567) Pascal 2026-09-28 20:56:15 +02:00
14ebbd5f2f ggml-openvino: mark unaligned batch-stride views unsupported (#29603) Ravi Panchumarthy 2026-09-29 00:24:01 +05:30
f1ea206218 batch: migrate speculative, mtmd and server to batch_ext (#29385) Xuan-Son Nguyen 2026-09-28 19:52:45 +02:00
6c7a87f7e5 common : fix HF cache paths on Windows (#29475) Adrien Gallouët 2026-09-28 16:25:24 +02:00
f00a64c147 webgpu: Handle unaligned writes in ggml_backend_webgpu_buffer_set_tensor (#29471) jbooth 2026-09-28 21:38:10 +08:00
d77dd0806d tests : refactor test-recurrent-state-rollback (#29426) Georgi Gerganov 2026-09-28 16:36:38 +03:00
6f767fe960 ggml-cpu: enable tiled flash attention for non-vector-multiple head dims on x86 (#29423) SXX 2026-09-28 21:23:31 +08:00
f916130d00 ci : ignore more vgpr spills in > 256 DQK fattn kernels (#29571) uvos 2026-09-28 14:52:07 +02:00
03a667aa30 vulkan: fuse qwen4exp's SCALE -> SIGMOID -> SCALE -> hc_post chain (#29520) François-Xavier Gsell 2026-09-28 20:20:14 +08:00
c2a9e16068 HIP: fix template skip for DKQ > 256 mfma kernels (#29559) uvos 2026-09-28 13:47:29 +02:00
4364bf7232 metal: support left and circular padding in GGML_OP_PAD (#29561) Pascal 2026-09-28 12:26:50 +02:00
ed7ac35e1e context : do not re-reserve the scheduler when toggling causal_attn (#28751) Sihan YuandGeorgi Gerganov 2026-09-28 11:58:51 +03:00
0c6a6a7ce5 Enables Windows ARM64 build with MSVC cl.exe (#28362) Sarah Wu 2026-09-28 01:07:27 -07:00
81ef10ea58 tests : fix ggml init (#29554) Georgi Gerganov 2026-09-28 10:24:21 +03:00
5262471615 vulkan: fix wrong results when a mul_mat reads a slice of a larger cache (#28956) Pascal 2026-09-28 08:36:38 +02:00
4da6337767 server : allow RANK pooling batch splitting for causal LLM rerankers (ie. Qwen3 and Qwen3-VL) (#28876) Tim Wangandtimothywang21 2026-09-27 17:28:10 -04:00
a97cce86a8 common : avoid side effects around params parsing (#29537) Georgi Gerganov 2026-09-27 20:18:56 +03:00
136887b665 common : make string_split<T> throw on invalid values (#29518) Adrien Gallouët 2026-09-27 18:04:11 +02:00
9adc7f420c convert : export YaRN scaling parameters for PLaMo-3 (#29528) Toki Nasin 2026-09-27 20:45:47 +09:00
6fd50a4094 ci : bump ty to 0.0.84 (#29529) Sigbjørn Skjæret 2026-09-27 13:43:08 +02:00
33c923db1b jinja : add support for dict builtin (#29477) Sigbjørn Skjæret 2026-09-27 13:41:50 +02:00
c9064dded7 opencl: refine bin kernel loading condition (#29503) lhez 2026-09-27 03:08:39 -07:00
c829670992 sycl: FWHT kernels for block widths above 512 (#29243) bri-prism 2026-09-27 03:08:19 -07:00
36d7b08340 CUDA: tune fp16 tile FlashAttention configs for head sizes 40-112 (#26289) Animesh 2026-09-27 15:37:37 +05:30
2ebd9ae621 HIP: Enable fattn-mma kernel on cdna for dkq > 256 for large batch sizes (#28907) uvos 2026-09-27 12:06:56 +02:00
cea74625fa vulkan: fix argsort kernel selection for Adreno (#29469) Ruben Ortlam 2026-09-27 12:06:06 +02:00
da6c28eb13 common : throw instead of abort on grammar without llguidance (#29516) Adrien Gallouët 2026-09-27 12:05:50 +02:00
d7fb90e8e2 RPC: use RDMA completion channel to not spin (#29440) Aman Gupta 2026-09-27 17:28:32 +08:00
7fb2b082ce ci : enable GGML_SCHED_DEBUG_REALLOC=1 for ctest workflows (#29514) Georgi Gerganov 2026-09-27 12:16:47 +03:00
187664b537 llama-bench : fix OOB access of hf_file (#29515) Adrien Gallouët 2026-09-27 10:24:54 +02:00
85ca3b52c3 hrm : fix layer placement of z_l_init weight (#29512) Georgi Gerganov 2026-09-27 10:16:23 +03:00
7ac59a6e3a hexagon: support tiled Q4_0 and Q8_0 GET_ROWS (#29511) kurquharandMax Krasnyansky 2026-09-26 23:02:37 -07:00
2b129ccfa0 hexagon: support for backend sampler (#29502) Max Krasnyansky 2026-09-26 20:29:36 -07:00
95887577ab cuda: support Nemotron 3 Puzzle state size 96 for ssm scan (#28717) Anav Prasad 2026-09-26 13:05:52 -07:00
694ec23548 musa: build the docker images from the PH1 MUSA SDK image (#29481) R0CKSTAR 2026-09-27 03:50:27 +08:00
6f856c7099 cuda: add F16 input to the FWHT (#29096) bri-prism 2026-09-26 12:36:09 -07:00
fcb3074f2b server : fix wake_fd warning on Windows (#29479) Adrien Gallouët 2026-09-26 21:25:25 +02:00
2145525a40 Revert "Change max context length for auto-fitting with unified KV (#28849)" (#29437) Gaurav Garg 2026-09-26 07:59:15 -07:00
81bc6b83f8 jinja : implement sameas test (#29448) Sigbjørn Skjæret 2026-09-26 10:14:58 +02:00
86a24a182b jinja : fix compile error (#29468) Sigbjørn Skjæret 2026-09-26 09:43:47 +02:00
08618ff8e7 llama : fix K/V and recurrent state cleanup after failed restores (#27530) Chipmunk 2026-09-26 16:23:03 +09:00
a1de614ba3 jinja : support noncall test statements with arg (#29443) Sigbjørn Skjæret 2026-09-26 08:56:48 +02:00
965f89794f polished Readme and llama-bench (#28968) Vladislavandplotnikov.v10 2026-09-26 09:21:21 +03:00
d834d44e64 ggml-cpu: tiled mul_mat for k-quants (#27851) jboothandGeorgi Gerganov 2026-09-26 13:38:57 +08:00
9f70b2cecd opencl: add A8 Q8_0 non-MoE dp4a binary kernel (#29439) shaofeiqi 2026-09-25 20:41:47 -07:00
4e7481175c hexagon: find software divide calls using binary inspection tool (#29449) Trivikram Reddy 2026-09-25 21:43:38 -05:00
171e8846b4 vendor : update cpp-httplib to 0.58.0 (#29407) Alessandro de Oliveira Faria (A.K.A.CABELO) 2026-09-25 20:55:18 -03:00
4b1a27fa0e common,rpc : simplify fs_create_directory_with_parents() (#29432) Adrien Gallouët 2026-09-25 20:33:37 +02:00
fcc891545b mtmd: fix mel preprocessor in LFM2 audio (#29403) Yuri Khrustalev 2026-09-25 12:52:44 -04:00
a25c9865fe opencl: add bin kernel kernel_gemm_noshuffle_q5_k_f32_32b_trans_ila_a8_bin, kernel_gemm_noshuffle_q5_k_q8_1_dp4a_ila_a8_bin (#29401) shaofeiqiandLi He 2026-09-25 07:35:18 -07:00
e85e15cf6d Fixing the vulkan build issue of legacy GLSLC version that has no cooperativeMatrix API support (https://github.com/ggml-org/llama.cpp/issues/29373) (#29409) sliu39 2026-09-25 20:33:49 +08:00
b248f4a3c1 gguf-py : ByteLevel processing defaults bos/eos to False (#29422) Sigbjørn Skjæret 2026-09-25 13:59:47 +02:00
d81aef1994 gguf-py : TemplateProcessing has final word on add_special_token (#29417) Sigbjørn Skjæret 2026-09-25 11:55:38 +02:00
27b20ba8b1 common : extract shared unicode path/string helpers (#29415) Adrien Gallouët 2026-09-25 11:40:41 +02:00
e351231c4f metal: FWHT kernels for block widths above 512 (#29095) bri-prismandGeorgi Gerganov 2026-09-25 02:15:33 -07:00
5a75f14c0f metal : split fa kernels into per-dtype libraries (#29329) Georgi Gerganov 2026-09-25 12:11:00 +03:00
e9f824d8c0 llama : add llama_prec_policy + model-driven W4A4 path (#24364)
ynankani Georgi Gerganov Johannes Gäßler
2026-09-25 08:36:35 +00:00
74facb828e Merge upstream/master into master xripandClaude Opus 5.5 2026-09-25 11:33:07 +03:00
d028c697b5 HIP: bump HIP_VERSION requried for fp8 to avoid missing __hip_fp8_e4m3 support in 6.2 (#29231) uvos 2026-09-25 09:36:36 +02:00
66963a8bc7 rpc: include nb in the get_alloc_size cache key and floor the result at ggml_nbytes (#29283) Jess SullivanandGeorgi Gerganov 2026-09-25 03:35:09 -04:00
cd74ef6274 [SYCL] support sparse FA (#28796) Neo Zhang 2026-09-25 15:17:51 +08:00
f9af9be219 musa: fix PH1 (MTT S5000) operator failures and build issues (#29193) R0CKSTAR 2026-09-25 15:16:29 +08:00
1ab7e5ad2d CUDA: fuse RMS_NORM + SCALE into one kernel (#29393) InflexCZE 2026-09-25 07:22:43 +02:00
f805c57a2d llama : fix tensor split for fused qkv with uneven K/V head sizes (#29294) Aman Gupta 2026-09-25 11:15:33 +08:00
4de0926596 hexagon: add q5_k quant type support (#29123) Jhen-Jie Hong 2026-09-25 10:03:59 +08:00
ed319febb1 hexagon: use DMA for contiguous dim1 CONCAT (#29404) kurquhar 2026-09-24 19:03:41 -07:00
84e76d8a23 metal : fix graph capture and handle empty graphs (#29390) Georgi Gerganov 2026-09-24 22:44:53 +03:00
cdc06426e7 metal : optimize sparse FA + clean-up (#29377) Georgi Gerganov 2026-09-24 22:44:36 +03:00
bced4595b8 sync : ggml (#29396) Georgi Gerganov 2026-09-24 22:44:05 +03:00
a02c7f58c1 hexagon: handle multi-sequence in concat_2d (#29344) Jhen-Jie Hong 2026-09-25 03:16:55 +08:00
5cf3a35287 llama-grammar: fix numeric truncation for token_id parsing (#29382) Daniel Kuts 2026-09-24 21:43:26 +03:00
07fc586e38 hexagon: dynamic quantizer improvements (#29395)
kurquhar Aparna M P Sigbjørn Skjæret Max Krasnyansky
2026-09-24 11:39:28 -07:00
97a418bdf4 hexagon: support I32 CPY and CONT (#29379) kurquhar 2026-09-24 10:48:57 -07:00
a72e04abe0 cuda : add F16 kernel support for CONV_2D_DW (#29064) yomi 2026-09-25 00:38:46 +08:00
8212c78024 test: flush status (#28352) kurquhar 2026-09-24 08:29:41 -07:00
945064fcea ui : fix missing svg use and animation elements in preview and download (#28962) Nandan Vallamdasu 2026-09-24 20:56:01 +05:30
fc343a84bb llama: add llama_batch_ext (#24669) Xuan-Son Nguyen 2026-09-24 16:25:07 +02:00
80034be9a8 rpc : hide unsupported NUMA split option xrip 2026-09-24 17:20:23 +03:00
3c1618659a rpc : guard optional CPU NUMA split support xrip 2026-09-24 17:18:54 +03:00
5aa2606c70 build : include RPC server in best build xrip 2026-09-24 17:12:38 +03:00
5031d1c935 docs : clarify SM75 decode result xrip 2026-09-24 17:10:14 +03:00
ee90e619dd docs : update SM75 prefill result xrip 2026-09-24 17:07:51 +03:00
230393b41e build : add best SM75 prefill recipe xrip 2026-09-24 17:04:20 +03:00
308883b335 server : change default pytest workers to 4 (#29376) Daniel Bevenius 2026-09-24 15:59:13 +02:00
70596c4dcb ci : use hf-jobs-cpu-performance, disable pytest workers (#29369) Georgi Gerganov 2026-09-24 16:30:05 +03:00
44ac39ff87 Merge branch 'ggml-org:master' into master Ilya Maslennikov 2026-09-24 16:19:09 +03:00
70c4e1582e vulkan: int8 coopmat1 matmul implementation for AMD RDNA3 and RDNA4 (#27952)
Ruben Ortlam Claude Opus 5 Piotr Wilkin
2026-09-24 15:18:47 +02:00
7a21ea7408 ggml : add SM75 prefill experiment controls xrip 2026-09-24 16:16:48 +03:00
f0945bccdc Merge branch 'ggml-org:master' into master Ilya Maslennikov 2026-09-24 15:54:45 +03:00

... 7 8 9 10 11 ...

Commit Graph Select branches Hide Pull Requests master Mono Color

Commit Graph

Select branches

Hide Pull Requests

master