fd9bd632f346085920d523dd307a3e4c11cc0a05	e08e473cf292ba73789769b993bb35bcdfc31689	2026-06-01T14:15:35+08:00	wangbomeng	HEAD -> dev, origin/dev	llama-bench: fix device-info
e08e473cf292ba73789769b993bb35bcdfc31689	f86112d6758298241ab2fa84ad4f0f7b1ace35c6	2026-06-01T11:47:01+08:00	wangbomeng		llama-bench:add device-info
ebce9e1a548c2329aa97f53b8b2ad50d968088dc	e850ff4de25ad1221abc9bf3bb30df252ccd9c36	2026-05-27T09:38:00+08:00	Yunzhe Jia	origin/runtime_replace	Refactor cal-llm integration by removing runtime capacity loading and adjusting request handling
e850ff4de25ad1221abc9bf3bb30df252ccd9c36	9985688f7fa96f755882f96a2cd8a18114fdc42d	2026-05-26T10:59:45+08:00	Yunzhe Jia		Enhance cal-llm integration by adding runtime checks for CALRT_LIBRARY and updating link settings for Linux and Windows
38ca6895815671a87ceec86b4aa4eb976580ce76	ddd60c928224b52eb4e1e60e01e8a1ecbb80ee45	2026-05-20T16:37:13+08:00	Yunzhe Jia		Add support for cal-llm profile dumping to JSONL file
2716130010c58f1cdd9d94f0e2f8e047bc92dc66	b35bda124ccd4ed581dd6088d3ffa4931c2809b6	2026-05-18T08:40:54+08:00	Yunzhe Jia		fix n_ctx_slot error by adding runtime capacity loading for cal-llm
a43741244a646d3f8fa3ff01bb9b7d059223d0a5	b2003d6c90ef0fa26ab9e930ba3e22b355eb3d39	2026-05-15T14:39:53+08:00	wangbomeng		run llama-passkey successfully, fix llama-bench, copy 3 times input to infer qwen2.5vl
1bec0db5a675ddc60fb793be2a746e8f3c8855dc	465df20c3e1f3f58d0f5531c796e0941a49c32b0	2026-04-30T16:02:59+08:00	wangbomeng	origin/dev-ben	bench: support 2 calbins
eeeef3f6d6a5aad9296323ca15d9d929745b8932	431ad6f1c787f1b47e9e7be14b2e1a7d8565485c	2026-04-28T16:16:19+08:00	wangbomeng		add unknown time info: prefill and decode
a2d5bf362d6c9a546f9deadf4f8975605a915d33	bd3a57a11aa008e0b5dc2a30f82fcd3a02587208	2026-04-09T21:01:46+08:00	Bomeng Wang		server: comment fixed time
18814f9bf1a423ebe8f32d6043c9cbc5b21b0218	bbee06d6d851e3f84f75b578047967d7e0e4a8dc	2026-04-08T13:55:52+08:00	wangbomeng		server : reset inference time
bbee06d6d851e3f84f75b578047967d7e0e4a8dc	e126c21ed7b793b648d63533ea7ee30885f5a82a	2026-04-08T11:29:36+08:00	wangbomeng		calrt: add test time print
1e8a16cd0e3ad3c9a0f8746852a0a38c6e52322e	893e52bda0fee99bbda1521ea733a760bc4201f1	2026-04-01T10:14:22+08:00	wangbomeng		calrt: fix slice and add infer/copy time
98e8f9acfe87cd93646482cb2d942b8a132f0852	8c8152a04036bb0d4756a4e08e54f514dae4e64d	2025-12-10T00:20:57+08:00	Yunzhe Jia		add timestamp
5d2387931528fbe04aa8d66de935147efb65ca4d	e6285a748657add8d974b78ca920c5b41753ee12	2025-12-08T16:40:34+08:00	Yunzhe Jia		add timestamp for one decode process
7e994168b1ccc12337ba8de939c4fd466107c1fb	bcfa87622ae46be6345a8e3dfdbdc5ba5414042b	2025-11-03T03:35:33+02:00	shani-f		SYCL: optimized repeat_back kernel (3× fewer asm instructions, 2× faster)Feature/sycl repeat back opt (#16869)
7fd205a8e8832ead273f62c5fd81d2b8aa910535	2f68ce7cfd20e9e7098514bf730e5389b7bba908	2025-11-02T00:15:31+02:00	Georgi Gerganov		scripts : add script to bench models (#16894)
d261223d24e97f2df50220e4a5b7f0adb69bba81	dcca0d3ab840ebe9b2ccd4719033d408eeb758d7	2025-10-30T23:19:14+08:00	JJJYmmm		model: add support for qwen3vl series (#16780)
8b11deea4663f29d3e042ce1056ba643264cd5f1	b9ce94017729465895402cbcfffb51fa926c15e3	2025-10-30T04:34:15+01:00	Oliver Simons		Hide latency of bias and gate-loading (#16847)
b9ce94017729465895402cbcfffb51fa926c15e3	3464bdac37027c5e9661621fc75ffcef3c19c6ef	2025-10-29T15:13:10-05:00	Jeff Bolz		vulkan: Fuse rope+set_rows (#16769)
3eb2be1ca5f37480aeb16102970d9e65f43347fe	e41bcce8f0b53032a1fed275cd253e931c041cf6	2025-10-29T06:29:12-07:00	Max Krasnyansky		Hexagon Op queue & dispatch optimizations (#16820)
f549b0007dbdd683215820f7229ce180a12b191d	9a3ea685b937c0f0cbfda2e50004ea54bf187512	2025-10-29T03:53:04-05:00	Jeff Bolz		vulkan: Call ggml_vk_buffer_write_2d from ggml_vk_buffer_copy (#16793)
338074c383c81366320d176d83b94b0a567ee0c2	851553ea6b24cb39fd5fd188b437d777cb411de8	2025-10-29T08:14:39+02:00	YaelLogic		sycl: add RMS_NORM_BACK operation support (#16808)
a8ca18b4b815a2abdbecb958ee5f4c542d69aac7	8284efc35c909217ee1b9a06903245d808ac2283	2025-10-28T19:41:43+02:00	Georgi Gerganov		llama-bench : clarify benchmarked parts of the computation (#16823)
ad8d36beffd791db10c94eb9e964afb891e3ca55	c053e18a66dd95dc340aa61317877c2a41d4e3cf	2025-10-28T03:50:33+02:00	tamarPal		sycl: add SSM_CONV operation support (#16800)
8423d019318b446640bb620e4ce80066d8530f05	5cca2542ac3f3f86831d32bce744d08fc2b353b0	2025-10-25T00:04:12-05:00	Jeff Bolz		vulkan: Optimize SSM_SCAN (#16645)
63d2fc46e17a06be5b4b5823a5ada088317f1f0a	a2e0088d9242bd9e57f8b852b05a6e47843b5a45	2025-10-22T13:47:09-07:00	Max Krasnyansky		Add experimental ggml-hexagon backend for the Hexagon NPU (#16547)
9b9201f65a22c02cee8e300f58f480a588591227	19a5a3edfd306516cc419679d69d6435943b6816	2025-10-22T16:58:23+02:00	Pascal		webui: introduce OpenAI-compatible model selector in JSON payload (#16562)
6ea37f57391d27736c35cd3c20c1f990b7952b74	fb349848f387f355450c3187556e71e6d32c145f	2025-10-20T22:26:17-07:00	lhez		opencl: fix warnings and clean up profiling (#16688)
84bf3c677857279037adf67cdcfd89eaa4ca9281	c9c1972e2c2cc6a771fcc145bfa138700179f961	2025-10-20T21:38:20+02:00	Sigbjørn Skjæret		model : add BailingMoeV2 support (#16063)
9ad4f1931ee0f3b41d9355245ef744786aaae0aa	79967ec596c0dacfd2251b085a57e79df292b1cc	2025-10-17T02:33:58-04:00	Ilia Ilmer		metal : add `CONV_TRANSPOSE_2D` (#16542)
ceff6bb253dd306f5404d7ccb3f11fadafe71b52	1bb4f43380944e94c9a86e305789ba103f5e62bd	2025-10-17T05:36:40+03:00	GittyBurstein		SYCL SET operator optimized for F32 tensors (#16350)
adc9b60f190c1016a09f439862fa1cbb302262ac	ee50ee1eadff58777ae746827b04de7ba0befc55	2025-10-16T13:10:32+08:00	takuya kodama		ggml-cpu: replace putenv with setenv for const-correctness (#16573)
f4ce81c45e7bd910e36bf44c253fc5255c49b1e4	17304cbcc1dd24de7741cbe57925d58e90a98ac1	2025-10-15T23:05:56+09:00	Sam/Samuel		metal: optimise `GGML_OP_SUM` (#16559)
7ea15bb64c81e3813eb0babf9a57e1bc5697f569	9c7185dd28416cf67f5e3b268381f311b5e3da56	2025-10-14T07:51:36-05:00	Jeff Bolz		vulkan: Improve build time for MSVC (#16545)
56fc38b9655fbe1869d8bd6cfb269418196cea69	1fb9504eb744969a990bfe4cfcf1d3d7a479541c	2025-10-13T17:01:24+08:00	Chenguang Li		CANN: fix CPU memory leak in CANN backend (#16549)
1fb9504eb744969a990bfe4cfcf1d3d7a479541c	3f750f8d760ab5a61491e6a9409072dfeee4b4d7	2025-10-13T10:55:32+02:00	Pascal		fix: add remark plugin to render raw HTML as literal text (#16505)
f9bc66c3ebcfddb5f09e4b21253623caeb8e414a	a31cf36ad946a13b3a646bf0dadf2a481e89f944	2025-10-13T08:52:22+08:00	hipudding		CANN: Update several operators to support FP16 data format (#16251)
a2fba89a426ff8005d303c73f0436e7e67368b70	20cc625edc2264aae2779e71bef1593e6a4e8c43	2025-10-12T07:19:06+02:00	Daniel Bevenius		hparams : add check for layer index in is_recurrent (#16511)
31d0ff1869aa2ea31f4e96d5877d0343e9a2171b	97870e64975b26c5e06a3540a8dc0ff601351e86	2025-10-11T21:39:04+08:00	Yann Follet		server / ranking : add sorting and management of top_n (#16403)
1deee0f8d494981c32597dca8b5f8696d399b0f2	d00cbea63c671cd85a57adaa50abf60b3b87d86f	2025-10-09T22:11:15+03:00	duduta		cpu : optimize the ggml NORM operation (#15953)
74b8fc17f92ada295a648e3c5eb28f46bca7d892	aeaf8a36f06b5810f5ae4bbefe26edb33925cf5e	2025-10-07T13:48:56-07:00	Reese Levine		ggml webgpu: profiling, CI updates, reworking of command submission (#16452)
8ae32dc9ecd7aeebf3a5b43557e0552a0a04cd4f	3df2244df40c67dfd6ad548b40ccc507a066af2b	2025-10-07T08:21:40+03:00	Georgi Gerganov		metal : various optimizations + refactoring (#16446)
04e632a4aab8e6bfff0f8bc216b36ceb1e199ff9	a80ff183abe4e5a76316257ffa597da41b3b6fa0	2025-10-06T14:56:59+02:00	Daniel Bevenius		ci : remove missing reranker model files (#16444)
ca71fb9b368e3db96e028f80c4c9df6b6b370edd	35266573b968e1c947b367782fb4b3eddbb4f3c0	2025-10-05T06:57:47-06:00	Gabe Goodhart		model : Granite docling + Idefics3 preprocessing (SmolVLM) (#16206)
898acba6816ad23b6a9491347d30e7570bffadfd	e29acf74fea996014380d59d31aa504ae8964258	2025-10-04T12:49:16+03:00	Radoslav Gerganov		rpc : add support for multiple devices (#16276)
e29acf74fea996014380d59d31aa504ae8964258	128d522c04286e019666bd6ee4d18e3fbf8772e2	2025-10-04T11:42:56+02:00	Acly		vulkan : incremental shader builds (#16341)
ef07a4090672a3438d7f64f197795d7dc1c18957	34fcc5a4ace8c69476ef2ea3857f39a60334acc4	2025-10-02T11:00:31-07:00	Reese Levine		ggml webgpu: add support for soft_max, optimize rms_norm (#16357)
34fcc5a4ace8c69476ef2ea3857f39a60334acc4	91a2a5655658bb9ab77894716b82fae7ecb4b4d1	2025-10-02T19:43:22+02:00	Piotr Wilkin (ilintar)		model : Apertus model implementation (#15852)
364a7a6d4a786e98947c8a90430ea581213c0ba9	2df5bcf357dba0c49b4df1d684b2ffc9e88b7054	2025-09-30T16:39:44+02:00	Adrien Gallouët		common : remove common_has_curl() (#16351)
6a2c6145a0b91b40eb3c3dba7b20ccc4b270490f	3b53634fe35771e2e318227aa81585726bae7234	2025-09-28T09:34:44+03:00	Georgi Gerganov		metal : extend mat-mat multiplication support (#16225)
e6d65fb02d553bd79cad94e517cdca18b687788d	8656f5de688cddcaea1d6174535eb60ee23ef6a0	2025-09-27T16:43:39-04:00	Jeff Bolz		vulkan: support arbitrary KV dimension in flash attention (#16160)
1a189278944d030211f336e103e96b65f976c361	e0539eb6aed346d4b25a6ea019044e88771e7690	2025-09-26T18:35:42+02:00	Aleksander Grygier		Allow viewing conversations even when llama server is down (#16255)
9b26511857ac09ae69ab485168fe2d3ee5fb1d6e	00217cd41388328c93e9e9644921c4319bb03bcc	2025-09-26T18:27:25+08:00	Aaron Teo		ggml-cpu: implement MXFP4 SIMD for s390x (#16193)
077c94d0caf87fbd3cf3288dbb5c0fd9670294cf	aa3ee0eb0b80efca126cedf9bcb4fb5864b46ce3	2025-09-25T22:35:05+08:00	Aman Gupta		CUDA: add a fused top-K MoE kernel (#16130)
02a6a82ae7c7ddd1819ae26b0cc36675879aecee	c498fc82fe5b83fc8c6e1627286bdc1f93caddbf	2025-09-25T11:29:08+03:00	Georgi Gerganov		metal : restore im2col perf (#16219)
bee378e0988b44bbe93b0768208080951b312363	5fb557653b8756ac2b6c6a102b1e44abcadf552f	2025-09-25T05:06:06Z	Eve		ci: run the x64 and arm ci on the github machines instead (#16183)
f2a789e33490deb483a2694b066b37e45524bb79	3a599719673c850647e3bb911ed6d91109bb91d2	2025-09-24T16:17:49+02:00	Acly		ggml : split graph allocations according to backend max buffer size (#15815)
3a599719673c850647e3bb911ed6d91109bb91d2	63b54c81a620981be020184ab99e63a8e50e47cb	2025-09-24T13:42:26+02:00	Tarek Dakhran		model : add label for LiquidAI LFM2-2.6B model (#16204)
351f3da39c85f59d581fc184f09283da7f099a3b	3ecb2f671a2f49d56357f99d135a94e841759178	2025-09-23T01:57:46+08:00	Haiyue Wang		clang-tidy : disable warning about performance enum size (#16127)
4d0a7cbc617e384fc355077a304c883b5c7d4fb6	9073a73d82a916cea0809de225ef5175c3a86e91	2025-09-22T08:31:40+03:00	Georgi Gerganov		ci : adjust params for less runtime (#16167)
1eeb523c3e0c7ffbd59469f5463dcbdecba3535e	5bb4a3edec297e74b0f7bd4ed5d0fdd12e28d858	2025-09-21T08:31:55+02:00	Giuseppe Scrivano		vulkan: optimize UMA buffer operations and fix driver hangs (#16059)
803dac2e48ef3ba26a504eb27c4e77ec2d21f7d0	459c0c2c1a400f960d7b8e8d94d31a8426f80986	2025-09-20T10:42:56+02:00	Ruben Ortlam		vulkan: use vec dot for matrix matrix multiplications (#16056)
be79d9fdd95ab8955527c4aaa67b90e8b9516718	f432d8d83e7407073634c5e4fd81a3d23a10827f	2025-09-19T15:15:21-07:00	ssweens		llama-bench: add --devices and --list-devices support (#16039)
0dd58b6877f3dc106593d6bc68d98305f553c566	69ffd891631befa9e6b485fd646a16dab4f2c007	2025-09-19T11:31:56+07:00	Xuan-Son Nguyen		ggml : refactor forward_dup for cpu backend (#16062)
3edd87cd055a45d885fa914d879d36d33ecfc3e1	c0b45097c33e2667a94444f08cc9e36bec0a5e2e	2025-09-18T12:03:34-07:00	Shawn Gu		opencl: optimize mxfp4 kernels (#16037)
c0b45097c33e2667a94444f08cc9e36bec0a5e2e	38dbdf4c057515ccea9bec0ca2518f86d5e4d28e	2025-09-18T13:46:17-05:00	Jeff Bolz		rename optimize_graph to graph_optimize (#16082)
38dbdf4c057515ccea9bec0ca2518f86d5e4d28e	368560a1e3b9a3bc83af741b0b2bc9e46fb420d2	2025-09-18T11:26:03-07:00	Bowen Han		CUDA: Optimize PAD_REFLECT_1D (#15957)
0320ac5264279d74f8ee91bafa6c90e9ab9bbb91	a7a98e0fffed794396b3fbad4dcdbbc184963645	2025-09-17T20:38:12+03:00	Georgi Gerganov		metal : refactor + optimize v2 (#15995)
c959b676be29e93f8dbc3bd6056ceba812a9eb72	cd08fc3ecc0264b4414b68af3874a6c689ed60c1	2025-09-17T15:32:42+02:00	Johannes Gäßler		CUDA: fix FA occupancy, optimize tile kernel (#15982)
d5fabe3682de515fd09d6c981f7a0d1b75614455	8ff206097c2bf3ca1c7aa95f9d6db779fc7bdd68	2025-09-17T14:33:08+08:00	Chenguang Li		CANN: Optimize ggml_cann_set_device (#15935)
8ff206097c2bf3ca1c7aa95f9d6db779fc7bdd68	77475530b8bbea3bf578632507e1284cdfe2c8c0	2025-09-16T16:17:08+02:00	jacekpoplawski		llama-bench: add --n-cpu-moe support (#15952)
77475530b8bbea3bf578632507e1284cdfe2c8c0	3913f8730ec6d6245480affc30ae3049107956f4	2025-09-16T15:27:52+02:00	Daniel Bevenius		ci : use macos-latest for arm64 webgpu build (#16029)
3913f8730ec6d6245480affc30ae3049107956f4	76888d202ed2b835ae19ea9f9db6baf39e419297	2025-09-16T15:25:57+02:00	Daniel Bevenius		ggml : fix padding in timestep embedding kernels (#15932)
76888d202ed2b835ae19ea9f9db6baf39e419297	f1fbffb5c0b34b2a68febb7da3fd0f8333f1ed4c	2025-09-16T13:41:38+02:00	Daniel Bevenius		ci : upload xcframework artifact from ios-xcode-build job (#16010)
106220562aca42b6738b8f51acfce0db1b8a2fb6	a68f31edd71cc39141113f05f7133a3e9ece8c61	2025-09-15T17:35:11+08:00	Aman Gupta		CUDA: some micro-optimizations in mmf.cuh for mul_mat_id (#15926)
b9c9c9f789cd57fb0b28b25223305613cd90fa10	50f4281a6f5c3a5d68bdeb12f904fa01e0e2ba91	2025-09-13T16:23:30+01:00	Jeff Bolz		vulkan: initialize vulkan-hpp to allow using extension function pointers (#15705)
f161463a54d9f93d41246286aa4a9569a91d804d	84d7b2fca11d1be118ce776f6d72a486c4883b74	2025-09-13T13:54:28+03:00	Georgi Gerganov		metal : allow ops to run concurrently (#15929)
00681dfc16ba4cebb9c7fbd2cf2656e06a0692a4	4f658855fa8f2e42b7ed9a5b298fa39a2e39b096	2025-09-10T22:04:03+02:00	Oliver Simons		CUDA: Add `fastdiv` to `k_bin_bcast*`, giving 1-3% E2E performance (#15872)
9de447d94e1ae9d1a36e5a2e5bf47483352c0d9c	0f0a3c2851134d49955f3c85afbb0b1bb47c3e07	2025-09-10T17:31:40+02:00	Daniel Bevenius		ggml-cpu : fix padding in ggml_timestep_embedding (#15917)
e7b6d83b524bbc24a1343d862de6dba8e8eddbd6	2cfef4d117d67ab1dec002915b48a15d11ee1973	2025-09-10T14:17:09+02:00	Daniel Bevenius		tests : filter out no-ops from coverage report (#15900)
ff02caf9eed261423289d1531a56536fbf57bfc2	ae355f6f7108540297d8b7f7ae71d20fe610a0b7	2025-09-10T05:23:19+02:00	Daniel Bevenius		ci : cache ROCm installation in windows-latest-cmake-hip (#15887)
a972faebed5fdc4a3d2a844d92d476058c02e02d	550cf726e133fd0a069d991287fd3a2a3e3e1cbd	2025-09-09T14:38:02+08:00	Aman Gupta		CUDA: Add mul_mat_id support for the mmf kernel (#15767)
70cd37dbbebdb7a2f84f08207f18eabb0b291a55	acc1b008cfd95e63d5f99a370dbffb98e5a99d2c	2025-09-09T06:06:52+02:00	Daniel Bevenius		requirements : update transformers/torch for Embedding Gemma (#15828)
e68aa10d8f3d26fdad5b912540362d79de5460e3	0a16bf52e6874369cb4fd2bdc4863ef984b688e7	2025-09-08T13:10:07-05:00	Jeff Bolz		vulkan: sort graph to allow more parallel execution (#15850)
56920f56651908d5cce7a310dabf54ac4f6fbb7f	b0d52998b962bd2681c34bf52af993af79f178b8	2025-09-08T21:50:05+07:00	Xuan-Son Nguyen		server : bring back timings_per_token (#15879)
f28d4f4ac963f182ea9d0fe9e269f3f5f3782aaf	9fcb29f22f5c33c04c7f0daebb24057899d67a1a	2025-09-08T13:34:56+03:00	Georgi Gerganov		metal : refactor + optimize (#15857)
a885dcff11a7b73f9377812d6151f6b15d307de0	663027fd5490438ce9c27ea866a560e1e268d11f	2025-09-08T10:27:07+03:00	Georgi Gerganov		batched-bench : fix llama_synchronize usage during prompt processing (#15835)
3c3635d2f20424d557b5b0605a2a356214ffe048	61bdfd5298a78593be649a1035ee2a120b13c4f0	2025-09-06T19:45:24+07:00	Xuan-Son Nguyen		server : speed up tests (#15836)
61bdfd5298a78593be649a1035ee2a120b13c4f0	01806e77714ae8a78130d432945b959a0956c56f	2025-09-06T18:35:04+07:00	Xuan-Son Nguyen		server : implement prompt processing progress report in stream mode (#15827)
c1c354e44c06d259679bb5bb7a8fa9f0b28480e4	a68d9144262f1d0ef4f6ba7ad4a7e73e977ba78c	2025-09-04T20:20:14+08:00	Chenguang Li		CANN: Refactor ND to NZ workspace to be per-device (#15763)
0fce7a1248b74148c1eb0d368b7e18e8bcb96809	8227695d7a3e5b357cb37fad263f00a8ca6db710	2025-09-03T13:33:15-05:00	Jeff Bolz		vulkan: don't use std::string in load_shaders, to improve compile time (#15724)
661ae31c9c68201577e70278285b349a5a662caf	407c23786dd0d3a503e9429eead96e611d3950c9	2025-09-03T19:59:16+02:00	Oliver Simons		CUDA: Optimize `rms_norm_f32` kernel and its fused variants, giving 1-6% perf E2E (#15715)
cdedb70a998cea7052560fe0b0615a839443564d	2c8dac72eb6acd4e20c0da251535dfc46d35178b	2025-09-03T18:16:26+03:00	Georgi Gerganov		sampling : optimize dist sampler (#15704)
2c8dac72eb6acd4e20c0da251535dfc46d35178b	40a751ea9a94364da73537b86502a808ebe1fc3a	2025-09-03T13:35:49+02:00	Daniel Bevenius		llama : fix incorrect model type for  Gemma 270M (#15764)
05c0380f2ae5c7193445e03ebc7b6de9ce49f1a6	8c3fdf44ecf08335942f2ba558955f55e88c7991	2025-09-03T16:16:21+08:00	xctan		ggml-cpu : optimize RVV kernels (#15720)
9961d244f2df6baf40af2f1ddc0927f8d91578c8	25f1045f07cf0daf667d63e35618842e3174a8c7	2025-09-02T17:12:37+08:00	hipudding		CANN: Resolve soft_max precision issue (#15730)
02c1813517412f3e00aa6ca7c0273fea64edb492	77dee9de97be75b7143a213bc48893e0c0b29af7	2025-09-01T16:19:07+02:00	Ruben Ortlam		Vulkan: Add Integer Dot Product mul_mat_vec shader for legacy quants (#14903)
b9382c3877c6067feccf182efe9449a2d1cb24c7	3dc7397a2799bdc07bccf637ab7ae5a1e786d1a4	2025-09-01T08:57:23+08:00	hipudding		CANN: Optimize MUL_MAT_ID (#15658)
3dc7397a2799bdc07bccf637ab7ae5a1e786d1a4	e92d53b29e393fc4c0f9f1f7c3fe651be8d36faa	2025-09-01T08:57:00+08:00	hipudding		CANN: fix RoPE cache issue on multi-device (#15629)
e92d53b29e393fc4c0f9f1f7c3fe651be8d36faa	0d161f021aa33ec0e90cce96f5d1a88925557327	2025-08-31T20:41:02+03:00	Georgi Gerganov		sampling : optimize samplers by reusing bucket sort (#15665)
c37052ab4d6d1ae73c0e90bc6e560cc6409e1311	5c16b9c87d840e4d5d55fa83c732c6b693346f40	2025-08-31T02:06:43-05:00	Jeff Bolz		vulkan: mul_mat_id coopmat2 optimizations (#15546)
e8d99dd0b67f2ecc1e45fca8074a3a18c3e036d2	a8bca68f727844e7dcf24a956003b3c2039ea563	2025-08-28T18:39:31-06:00	Gabe Goodhart		nvidia nemotron nano v2 (nemotronh) (#15507)
fbef0fad7a7c765939f6c9e322fa05cd52cf0c15	da54f9f1a2db07aaae390024ac466e7867685d94	2025-08-27T20:58:09+02:00	Johannes Gäßler		server: higher timeout for tests (#15621)
1e7489745a74996fc36e8fd05b73aa16bc184e0c	1cf123a343ab7ca5586aacb9e0a1d2de7fe33be4	2025-08-27T17:21:41+08:00	Chenguang Li		CANN: refactor mask handling and improve performance in FA (#15561)
8b696861364360770e9f61a3422d32941a477824	8ce3ff1d91245e158d98d8062cd64b0dd98dcfe3	2025-08-27T00:27:49+05:30	Akarshan Biswas		SYCL: fix rms_norm_mul_add for tensor dim not a multiple of sg_size (#15592)
44b1efa41acd4df3f56ee0e46f898135ecd1a054	a6a58d64785cb458ed9de52f391aa38142d38d64	2025-08-26T15:42:49Z	Eve		tests: add performance test for mul mat id (#15543)
a6a58d64785cb458ed9de52f391aa38142d38d64	0373486dbc0dccbdcb3b5fdd65759d88cec06196	2025-08-26T21:05:25+05:30	shalinib-ibm		llamafile: PowerPC Sgemm Optimization (#15558)
b3964c1e890ef8c947afb36a5124ce6fcb2136d4	79a546220c719e6a70627b243a478ab8d84dc9e1	2025-08-26T14:22:14+03:00	Georgi Gerganov		metal : optimize FA vec for large sequences and BS <= 8 (#15566)
1d8d83deaa48d4a5491820d58ff1c0d8cf9d196c	c4e9239064a564de7b94ee2b401ae907235a8fca	2025-08-26T12:46:15+03:00	Georgi Gerganov		metal : improve `MUL_MAT_ID` (#15541)
0115cfb7c1dec0753aecb34f3af4338bf9de8eeb	3a617cb18121a85eed17ac5585788810c5d9c1e0 0d5a470223fc90b6b6807921d68011ff06ae7f9e	2025-08-26T15:19:56+08:00	Yunzhe Jia		Merge branch 'master' into dev
74f52f77f28a5ad6d6075231afcb8d1ad763ca32	f7207b0415986dd7f48447149da7de3a82338276	2025-08-26T05:21:22+08:00	Qeeweew		CUDA: Accelerate MXFP4 table lookup using `__byte_perm` (#15451)
4d917cd4f64cc37744e76d084659475819fb0728	886b97a5d693550c2da470c091d9d27bf38398f8	2025-08-25T17:56:59+02:00	Ruben Ortlam		vulkan: fix min subgroup 16 condition for mmid subgroup optimization (#15565)
5eff6ec9b1220b599a43b594b1110487ab6aca08	dfd9b5f6c7586c88588f06a644c131bec071a0a1	2025-08-25T17:23:40+02:00	Johannes Gäßler		CUDA: MoE helper in device code, better tile sizes (#15525)
6b64f74b55628e4193f4fb00313f07dbd8556528	0d5a470223fc90b6b6807921d68011ff06ae7f9e	2025-08-25T13:56:43+03:00	Georgi Gerganov		batched-bench : fix unified KV cache handling + pp timing (#15562)
043fb27d3808766d8ea8195bbd12359727264402	b730706a49e576fb882dc34d9966345778b3ab0b	2025-08-24T19:36:36+02:00	Ruben Ortlam		vulkan: apply MUL_MAT_ID subgroup optimization to non-coopmat devices (#15524)
611f419cff11e4952228162a1c44cb35dff2274a	b1afcab804e3281867a5471fbd701e32eb32e512	2025-08-23T13:16:17-05:00	Jeff Bolz		vulkan: optimize rms_norm, and allow the work to spread across multiple SMs (#15281)
9ef536907de1b50c30e0369284898d30472a755a	21dc4ddaf21b8ed551d717e7606abd2cffbacdbf	2025-08-23T12:58:58+02:00	Johannes Gäßler		scripts: fix compare-llama-bench.py (#15521)
289bf4113ef5c02d8f5eb0cf2d86683d8b8bc4d9	b55f06e1aa67fb10e89f53e31bbccf37eb2678ea	2025-08-23T02:33:36-05:00	Jeff Bolz		vulkan: Rewrite synchronization to allow some overlap between nodes (#15489)
330c3d2d21b55bca5517db7d2eea2ea8f131df4a	e92734d51bcb82cc35f0a6b5a14928f0036b2c90	2025-08-23T01:31:54-05:00	Jeff Bolz		vulkan: optimize mul_mat_id loading row ids into shared memory (#15427)
ad5c975c2d0297124fad210776ef8eed6b90d578	4afb0a746f22abaa545b3ebdb76a400d7da3a713	2025-08-22T16:11:04+08:00	Aaron Teo		ggml-cpu: Support Q5_0 and Q5_1 on s390x (#15486)
a0f98dd604d34826eb5ea2560d1e23fe726921df	54a241f505d515d625767b993bfd573ecee306b9	2025-08-22T14:12:07+08:00	Chenguang Li		CANN: Optimize RMS_NORM using cache (#15419)
2758fa10dab0556e6c3f130e664750fd6773dc7c	b108e429043ee5c9fc8fa4957a0a52c3e490d5c9	2025-08-21T12:16:54+02:00	Daniel Bevenius		examples : add model conversion tool/example (#15455)
fec9519802ae1567048abb126cdd5ea160a22d0f	657b8a77bd01854f99d37a47318fa24f2e7e298f	2025-08-20T09:33:14-05:00	Jeff Bolz		vulkan: shorten pipeline name strings (#15431)
37f10f955f70e0158d50343d0b9a3f92d194daae	2f37014073f4c6ddc8f241c927db87337c71aa52	2025-08-20T12:31:16+02:00	Daniel Bevenius		make : remove make in favor of CMake (#15449)
a6d3cfe7fa6ea1fb0e1ba8243b731db22ddc0b49	67f09a3a27db443f9870aac87e163dba0d08131e	2025-08-19T21:28:22+08:00	SHUAI YANG		CANN: optimize rope operator (#15335)
6424594c56f4dbd0573455d89a0d89a0ac093d13	e9288e886970884f288533cd597b3798995b4099	2025-08-19T10:54:31+02:00	Marvin Gießing		ggml-cpu: add mxfp4 VSX intrinsics for Power9+ (ppc64le) hardware (#15385)
f0d3c7405c323784a60f14ddddfbac3f7404d417	f08c4c0d8d0cb6caaf8b7ad316039232b9fa059c	2025-08-19T08:45:12+03:00	Georgi Gerganov		batched-bench : use rand tokens (#15398)
de5627910df74298c998e6bb36ee3217375a5719	65349f26f2299e06477ec8e85e46243046801358	2025-08-17T03:41:45-05:00	Jeff Bolz		vulkan: Optimize argsort (#15354)
1fe00296f587dfca0957e006d146f5875b61e43d	de2192794f4e8e04f2e8167ef2424905145e88fc	2025-08-16T11:48:22-05:00	Jeff Bolz		vulkan: fuse adds (#15252)
863d341eeb81db104902b14b6f9413daa515e957	d32e03f4495d3efa1c5126f53b449f1d429c5664	2025-08-14T08:38:10-05:00	Jeff Bolz		vulkan: perf_logger improvements (#15246)
5cdb27e0917479d2d742cea7beee089574bb09fa	3ea913f1ce9567289aedd866a569dbab8fb8e419	2025-08-14T03:03:57-07:00	Jonathan Graehl		finetune: SGD optimizer, more CLI args (#13873)
6028bf74351d35a06bd98498624f8c2f029f7d1a	bc5182272c373267352bc689e5fca276934bea2d	2025-08-13T10:04:46+02:00	Oliver Simons		CUDA: Optimize `reduce_rows_f32` kernel, leading up to 25x perf improvement on kernel-level and 10% perf increase for Gemma3n (#15132)
bbd57b7eafb3b32e3f7f3a2175bce0b35abc7de8	25ff6f7659f6a5c47d6a73eada5813f0495331f0	2025-08-12T16:12:13+08:00	Chenguang Li		CANN: GGML_OP_CPY optimization (#15070)
4850b52aedceeb70bb4fe49f2d7cd1df6ee98682	cd6983d56d2cce94ecb86bb114ae8379a609073c	2025-08-08T23:04:36+02:00	Johannes Gäßler		server-bench: external OAI servers, sqlite (#15179)
9a96389544a08fd829fccda28142ce2066017fde	1d72c841888b9450916bdd5a9b3274da380f5b36	2025-08-07T13:45:41+02:00	Christian Kastner		ggml: Skip backend library linking code when GGML_BACKEND_DL=ON (#15094)
2241453252147bb7362a286977ee9f9a92130062	9515c6131aecaccc955fdedcfe16c3e030aaefcb	2025-08-06T14:12:42+08:00	Chenguang Li		CANN: add support for ACL Graph (#15065)
fd1234cb468935ea087d6929b2487926c3afff4b	f324a3b715d5c1081c110ce459f8a8486fb1ee89	2025-08-05T22:10:36+03:00	Georgi Gerganov		llama : add gpt-oss (#15091)
4cb208c93c1c938591a5b40354e2a6f9b94489bc	3025b621d12a6931ff5e9775d4f644719980ad91	2025-08-02T04:21:37-05:00	Jeff Bolz		vulkan: coopmat2 mul_mat optimizations (#14934)
3025b621d12a6931ff5e9775d4f644719980ad91	ec0b18802c91badd3ff1388ffd09ee163251bd72	2025-08-02T17:20:40+08:00	R0CKSTAR		llama-bench: rename DB table name from test to llama_bench (#15003)
a9f7541ec25c4c8547daf5ff48700ad2836e2b7d	9c35706b98ea271858acef4194f526a71b24cdc9	2025-08-02T02:57:04-05:00	Jeff Bolz		vulkan: optimizations for direct convolution (#14933)
484b2091ce5017901483b5204c07878f171d1441	daf2dd788066b8b239cb7f68210e090c2124c199	2025-08-01T08:47:27+08:00	R0CKSTAR		compare-commits.sh: support both llama-bench and test-backend-ops (#14392)
c7aa1364fd59b2ac06fd9e0a719253d968472dc3	1a67fcc30677e96dda76bb1b290788e7d8852b51	2025-07-29T17:43:43+02:00	uvos		HIP: Ignore unsupported unroll transformation in fattn-vec (#14931)
bbd0f917797e9d524680f1b30d34a46eb06d7651	0a5036bee9cfb946870689db4400e9e0d17844c9	2025-07-29T10:40:50+02:00	Johannes Gäßler		server-bench: make seed choice configurable (#14929)
bda62193b2a6bebbf515c3c389303094a44458c1	c556418b600ad5792440942079d93e393595688b	2025-07-28T09:04:27-07:00	Leonard Mosescu		test-backend-ops : extend test case filtering (#14865)
c556418b600ad5792440942079d93e393595688b	db16e2831c0f344f041af3d067db81c42b16eb22	2025-07-28T18:59:04+03:00	Radoslav Gerganov		llama-bench : use local GPUs along with RPC servers (#14917)
66906cd82a4a1fd10151707cee3f66cb61fc4055	11dd5a44eb180e1d69fac24d3852b5222d66fb7f	2025-07-26T18:28:14-04:00	deepsek		HIP: Enable Matrix cores for MMQ Kernels, Enable stream-K for CDNA 3 (#14624)
793c0d7f46384001738c337d7afa46b45ae32745	ce111d39d666f4a3b6a561ad020f6feb8cc67790	2025-07-25T10:47:39-06:00	Gabe Goodhart		metal: SSM_SCAN performance (#14743)
e4868d16d24dec55e61bcaadaca28feed8f98b13	820de57d4faa427a3d0bfb14e48057247fae036e	2025-07-24T16:31:48+03:00	Georgi Gerganov		context : perform output reorder lazily upon access after sync (#14853)
a86f52b2859dae4db5a7a0bbc0f1ad9de6b43ec6	b284197df426fb189cdcfe56a43c863a788ac756	2025-07-23T21:43:25+02:00	Johannes Gäßler		CUDA: fix overflow in FA, tune performance (#14840)
d1aa0cc5d13ec3fa553de5d298f9166679e58479	c8ade30036139e32108fee53d8b7164dbfda4bee	2025-07-22T13:33:37+01:00	Ed Addario		imatrix: add option to display importance score statistics for a given imatrix file (#12718)
a979ca22db0d737af1e548a73291193655c6be99	90083283ec254fa8d33897746dea229aee401b37	2025-07-19T21:59:08+02:00	Ervin Áron Tasnádi		ggml: adds CONV_2D op and direct GEMM Vulkan implementation (#14316)
90083283ec254fa8d33897746dea229aee401b37	d4b91ea7b2da253e1355b503f0fcb7b428ce005d	2025-07-19T12:51:22-04:00	compilade		imatrix : use GGUF to store importance matrices (#9400)
01612b74090df592663cfa01f661c9628f403b59	086cf81e88fb75287b71ff19c08a206b7bc2e02f	2025-07-17T19:08:33+03:00	Georgi Gerganov		llama : reuse compute graphs (#14482)
225e7a1438f4ea85eaa7b5ef3ab3b266ee4d9c06	ab140198211385b85eeeb0abd549a4bbe259e10d	2025-07-16T16:35:42+03:00	Georgi Gerganov		llama : add high-throughput mode (#14363)
ab140198211385b85eeeb0abd549a4bbe259e10d	64978340b0b4a0a6e2fb74270c1509383d2eff32	2025-07-16T20:03:51+08:00	Aman Gupta		Support diffusion models: Add Dream 7B (#14644)
5cae76654113160f691f581930b69fc5535e8159	4b91d6f71f14040979bcdb7b6729b3bca93ec1c1	2025-07-16T09:33:28+02:00	Johannes Gäßler		scripts: synthetic prompt mode for server-bench.py (#14695)
79e0b68c178656bb0632cb8602d2940b755077f8	c81f4192f91a1e209c1eec7a84fe5371ef9175da	2025-07-16T12:00:42+08:00	Min-Hua		llama: add LLAMA_API to deprecated llama_kv_self_seq_div (#14708)
494c5899cb76859f32ddd913534f2685fd684a3d	0f4c6ec0f1a9607ba67071f8a02c69b0afc2f91e	2025-07-14T13:14:30+02:00	Johannes Gäßler		scripts: benchmark for HTTP server throughput (#14668)
05fec5bd298d3c0243cbb9336e59b8b6aff75a81	dcf7f2ea3c4e3cf36dd9ab5a36785c00e6033267	2025-07-13T10:36:33+03:00	Georgi Gerganov		ggml : add build-time message to remind about ggml_set_rows (#14661)
b3ad3a0191994d6c47b2bd389d5c7431526ecd2c	98197e5c98388470030d908f355ec5937dcccaaa	2025-07-12T05:12:26-05:00	Jeff Bolz		vulkan: support SET_ROWS (#14587)
98197e5c98388470030d908f355ec5937dcccaaa	f5e96b368f1acc7f53c390001b936517c4d18999	2025-07-12T04:51:58-05:00	Jeff Bolz		vulkan: optimizations for deepseek prompt processing (#14555)
0aedae00e6fb48680324a5ac5da9cba0e35de6b5	6bdda13981d6c8189b7dc4f9fb8ecb91c21529f8	2025-07-10T18:20:13-06:00	Gabe Goodhart		model : Granite Four (#13550)
4a5686da22057867c23bd4a6be941ddc8c51e585	98bab638fb28cf95a5a66dd2d51b40d6c8f6d69a	2025-07-09T14:59:57-04:00	compilade		llama : support Jamba hybrid Transformer-Mamba models (#7531)
6efcd65945a98cf6883cdd9de4c8ccd8c79d219a	699f4392a33f57c3352cf8d60bdc53db7ca235e7	2025-07-08T13:11:42-05:00	Jeff Bolz		vulkan: optimize flash attention split_k_reduce (#14554)
53903ae6fa5f1caf187889c839cdd1ad25da4018	4d0dcd4a06080e796e6742a88f2ffa7fc41b28b8	2025-07-08T02:38:31-05:00	Jeff Bolz		vulkan: increase timeout for CI (#14574)
5d46babdc2d4675d96ebcf23cac098a02f0d30cc	e17991c466ac835b2c71bd813c1ca7ff8dd97b94	2025-07-02T13:10:24-04:00	compilade		llama : initial Mamba-2 support (#9126)
caf5681fcb47dfe9bafee94ef9aa8f669ac986c7	83790b0e7e09ab17238b16452a33053a71dbdfad	2025-06-29T20:02:53+02:00	matteo		server : support jinja extra template kwargs (Qwen3 enable_thinking feature), from command line and from client (#13196)
a0535ffa0d35fccfec3e1a0a3bfc9dbb6054d7c0	bd9c981d7226107f18deb8344c3301450311bb8b	2025-06-29T11:04:10+02:00	Sigbjørn Skjæret		ggml : implement REGLU/GEGLU/SWIGLU ops (#14158)
bd9c981d7226107f18deb8344c3301450311bb8b	27208bf657cfe7262791df473927225e48efe482	2025-06-29T02:43:36-05:00	Jeff Bolz		vulkan: Add fusion support for RMS_NORM+MUL (#14366)
8d94219a4a7f2da72ee542019ca01f36af93d1d6	f667f1e6244e1f420512fa66692b7096ff17f366	2025-06-27T16:41:40+03:00	Radoslav Gerganov		ggml : add ggml_set_rows (#14274)
8f52c6f8abb2ed9ad967cc0bfb90acb073011193	d2f325130c4ed77a3dcff2c23a926587fb08e2cc	2025-06-26T15:10:45+08:00	Yunzhe Jia		Squashed commit of the following:
60ef23d6c14d325d83eae5752e5de39ad268e9b0	b193d5306912a2adae0fde7481819f6ee0941bc6	2025-06-26T05:49:04+08:00	Aaron Teo		ggml-cpu: enable IBM NNPA Vector Intrinsics (#14317)
0142961a2e67909e33cdf410274b56c08c5dce7a	ce82bd0117bd3598300b3a089d13d401b90279c7	2025-06-24T01:12:56+02:00	uvos		CUDA/HIP: optimize mmv paths taken for HIP devices (#14324)
812939a9e90f99d1bd5bb1bc6b99d12600671d50	4c9fdfbe1580a66fd7d77c77418ce2c606a29fdd	2025-06-20T10:50:27+03:00	Georgi Gerganov		model : more uniform output id handling (#14275)
8f71d0f3e86ccbba059350058af8758cafed73e6	381174bbdaf10d6a80dc2099f284b20544d86962	2025-06-19T12:24:14-07:00	Diego Devesa		ggml-cpu : remove unnecesary arm feature detection (#14281)
fffcce535ebbdc25f81966a15b758658788e7466	5fc7856815920f828f9e90cb759ac82c7f0c1ea5	2025-06-19T13:24:12+03:00	bashayer hijji		llama-bench : add --no-warmup flag (#14224) (#14270)
8d947136546773f6410756f37fcc5d3e65b8135d	50d2227953ca9024f04255b4f116d06fcc0db74c	2025-06-19T01:10:26+08:00	Aaron Teo		docs: add s390x build documentation (#14264)
0dbcabde8c006d5cf781ca0fe071c41559572a72	ad590be98c83217fcf1a101d78d9ab389fd5dc0b	2025-06-16T10:32:13-03:00	bandoti		cmake: clean up external project logic for vulkan-shaders-gen (#14179)
b9912ac570de8945ae9383c9ca8291027bf287dd	00ba7726100d7e1941d9f5a06f56a7559945b33c	2025-06-15T09:18:37+03:00	Georgi Gerganov		batch : auto-gen positions + verify multi-sequence input (#14177)
2e42be42bd6bf1dcc643d6ac4e77419bfe5dd24f	fb85a288d72abbd5e5daa8de96e6f8bfa7b5ab46	2025-06-14T16:34:20+08:00	Aman Gupta		compare-llama-bench: add option to plot (#14169)
c61285e7396c8e526fe7794c19e8d4f1c99bfc51	09cf2c7c655c90e53e100f29b830a788bab0653d	2025-06-13T08:45:37+01:00	Ewan Crawford		SYCL: Bump oneMath commit (#14152)
532802f938c6a18cc6a704057ab571f253fd77ed	d4e0d95cf581f50c9a21d06eaecae2dd580076bd	2025-06-11T19:07:44Z	Christian Kastner		Implement GGML_CPU_ALL_VARIANTS for ARM (#14080)
3a12db23b6918682ccb70ab15d897f11fe5fc320	ae92c1855b1a5b604fa1bfcced19a556ab3e78c5	2025-06-10T16:48:07+01:00	Juk Armstrong		Fixed spec timings to: accepted/tested instead of accepted/drafted (#14104)
228f34c9ceefa3ea4f4d6933edd858121e8106cb	0974ad7a7cd4bca846b15c484ff3be890135a52c	2025-06-07T18:58:20+05:30	Akarshan Biswas		SYCL: Implement few same quantized type copy kernels (#13739)
7e00e60ef86645a01fda738fef85b74afa016a34	ea1431b0fa3a8108aac1e0a94a13ccc4a749963e	2025-06-03T13:30:22-05:00	Jeff Bolz		vulkan: fix warnings in perf logger querypool code (#13937)
71e74a3ac929b8af91f16f73f3c2b9b2f796d207	bfb1e012a0b7658e8f00ed4333d059943ea9d648	2025-06-02T16:54:58-07:00	lhez		opencl: add `backend_synchronize` (#13939)
093e3f1feb16e25e58f7d61e01266c830dd424b8	663445b0deb21fb602176da030d4154197a4fca6	2025-06-02T17:48:36+05:30	shalinib-ibm		cmake : Handle mixed-case 'Power' strings in POWER CPU detection (#13966)
053b1539c02617eff744f89525ee57497c3c1fbe	b3a89c3d9e34c28c5be70d8b687a84775746d4a0	2025-05-31T15:39:19-07:00	Max Krasnyansky		threading: support for GGML_SCHED_PRIO_LOW, update thread info on Windows to avoid throttling (#12995)
b49a8ff96b769b8a4c36d89fb783ec0135be582b	53f925074de02c5304b00c14b4d6d8910c58667d	2025-05-30T19:40:57+05:30	Akarshan Biswas		SYCL: Add mrope kernel (#13755)
54a2c7a8cd8a32b44e3a98c2999b0f5c9114be5c	21fcc21ad5e5de2daa5da8d08dbbcc86b8d815d7	2025-05-29T19:39:20+08:00	Yibo Cai		arm64: optimize q4_k_q8_k kernel with i8mm (#13886)
bef817638780dcbd8c0c80c97b7a4f8e92c8fe74	34b7c0439ed0f98575cc4689dfecd98991dee8be	2025-05-27T11:39:07-05:00	Jeff Bolz		vulkan: use timestamp queries for GGML_VULKAN_PERF (#13817)
bc583e3c63c04a11d287c108ea9e6a515ead0423	72b090da2c50e540143fd312a2f9aa5f151e6136	2025-05-27T14:06:10+02:00	Xuan-Son Nguyen		mtmd : support Qwen 2.5 Omni (input audio+vision, no audio output) (#13784)
f5cd27b71da3ac375a04a41643d14fc779a8057b	a2d02d5793fd9af7a7224773456501691b95fd02	2025-05-25T01:48:08+01:00	Olivier Chafik		`server`: streaming of tool calls and thoughts when `--jinja` is on (#12379)
2bd1b30f6979235ec67b95c183b9b77baa7ab9ce	259469c4b57c1a32606353bcac52ba683424a990	2025-05-24T13:26:47-07:00	Diego Devesa		ggml-cpu : set openmp wait time if not set (#13758)
f7c9429c85748cde9599499601ba48d0057722e6	1dfbf2cf3a9f15193dd893396d07762bbd2c4785	2025-05-20T02:54:43+02:00	Nicolò Scipione		sycl : Overcoming workaround for mmap() allocation on Windows (#13482)
725f23f1f3f0d3adf49f95d8dfa6e7c74adff149	92ecdcc06a4c405a415bcaa0cb772bc560aa23b1	2025-05-19T14:38:20+01:00	Alberto Cabrera Pérez		sycl : backend documentation review (#13544)
6c8b91500e75df6664278d1e9af3e39e8a2fb0d0	3cc1f1f1d24472a6558c942b1c78989ff4b0e569	2025-05-15T06:46:55-07:00	Diego Devesa		llama-bench : fix -ot with dl backends (#13563)
b2838049ccf50859cea4c390e81405c2fb01e820	aa48e373f256df9608395ee6881e7010840d6202	2025-05-15T05:57:02+03:00	Georgi Gerganov		bench : handle decode errors (#13548)
5ab5d5fb256aa23f8ab4de8463515ea627f43006	3198405e98530c683d12fd123e3920f2bd2aafa5	2025-05-15T03:53:52+08:00	Yibo Cai		arm64: optimize q6_k_q8_k kernel with i8mm (#13519)
24e86cae7219b0f3ede1d5abdf5bf3ad515cccb8	bb1681fbd532eba26ae4c14cd8be884c8afeb31c	2025-05-14T18:55:26+09:00	Jeff Bolz		vulkan: KHR_coopmat flash attention (#13506)
be1d4a13db26750fac702ceb3af88ae4f39dc9f4	ab3971f2a0a526564bfde0e4cc8a5b90f9d33ad2	2025-05-14T08:41:01+02:00	Sigbjørn Skjæret		scripts : fix compare-llama-bench.py show parameter (#13514)
f0995d28ce3d15095b6845d94ce4465e46575873	c252e0c4097b34666e5a81db9d0450d71fa3098f	2025-05-13T18:04:39+03:00	Georgi Gerganov		metal : use FA-vec kernel up to batch size 20 (#13496)
c252e0c4097b34666e5a81db9d0450d71fa3098f	4f711afed5e7ef4304b567c8888ee1aa60e868eb	2025-05-13T18:04:00+03:00	Georgi Gerganov		metal : optimize multi-sequence FA vec kernel (#13493)
b89d605a91dee0a518ecd582f8991a07d523e2fa	b4726345aca49e2ad62d615e6e370b3dbad6434f	2025-05-13T18:01:53+03:00	Georgi Gerganov		batched-bench : fix pp batch contents (#13492)
bf7937112058f2815fc3825a9ff7b536ecafa3bb	d590cd4c244e5f260c42c290b83a358b9d86d763	2025-05-13T15:31:12+02:00	Sigbjørn Skjæret		scripts : support arbitrary input file formats in compare-llama-bench.py (#13455)
cf0a43bb6490bd49344775abb22ba26f8047cb54	f0d46ef15717cd609a7b69cf6190edde64d466c8	2025-05-12T15:31:37-07:00	Diego Devesa		llama-bench : add defrag-thold, check for invalid ranges (#13487)
22cdab343b63edd0906f1c132616746085f81983	a71a4075cdb8e81eaaa834e48ffda88b038286bc	2025-05-12T13:08:22+02:00	Diego Devesa		llama-bench : accept ranges for integer parameters (#13410)
09232370fc6426aa5dd9be01a8271b9c28f5af3a	7474e00b34629e9cd8b06bc87ad935584ea30f8e	2025-05-11T16:20:39+02:00	Sigbjørn Skjæret		scripts : exit compare-llama-bench.py gracefully when there's nothing to compare (#13451)
7f323a589f8684c0eb722e7309074cb5eac0c8b5	3eac209319a6726fd9687c6188fc6b916b65953d	2025-05-11T20:18:39+08:00	David Huang		Add `--no-op-offload` to improve `-ot` pp perf in MoE models like llama4 400B (#13386)
dc1d2adfc0f4de84da7923866d00781ec5c4e666	7c28a74e0783f4bb74a246fb9f19bf212139e365	2025-05-09T23:07:07-07:00	Jeff Bolz		vulkan: scalar flash attention implementation (#13324)
33eff4024084d1f0c8441b79f7208a52fad79858	17512a94d636c4b6c1332370acb3e5af3ca70918	2025-05-09T19:29:37+02:00	Xuan-Son Nguyen		 server : vision support via libmtmd (#12898)
611aa914ef4231fab5d1ad04773c42e119ae2d2e	0cf6725e9f9a164c39f7a87214d60342f7f946d8	2025-05-09T15:14:56+03:00	Georgi Gerganov		metal : optimize MoE for large batches (#13388)
1e333d5bba18e99bc328bb87ac1ee6a4e6260e0e	2f54e348ad2999c4e31b8777592247622b20420f	2025-05-06T20:27:06+05:30	Akarshan Biswas		SYCL: Disable reorder optimize by default and stop setting tensor extras when optimize is disabled (#13254)
b34c859146630dff136943abc9852ca173a7c9d6	9b61acf06041dcbaff6afa5f28940e93297f8520	2025-05-05T17:03:31+03:00	igardev		server : Webui - change setText command from parent window to also send the message. (#13309)
b34443923cad751483cc53af2e680d595daadce7	a75cb30dc9e63488c3614e2d5a9fe2306eaf47cd	2025-05-02T20:54:30+03:00	Georgi Gerganov		sync : ggml (#13268)
3f3769ba76061a511f02f2a48da2ad2d93fce511	2f567611c0234bbca0a4009762acb47b56866095	2025-05-02T22:23:12+05:30	shalinib-ibm		ggml : Enable MMA for BF16 in llamafile_sgemm (#13148)
19e899ce21a7c9ffcf8bb2b22269a75f6e078f8f	e2e1ddb93a01ce282e304431b37e60b3cddb6114	2025-04-29T23:32:04+02:00	Johannes Gäßler		scripts: n_depth for compare-llama-bench [no ci] (#13201)
5a6398011704c31178d7b774be67856ba57647c8	cdf76586b23c67abd3ca064ee2c084c57ae240bd	2025-04-29T16:24:36+01:00	Alberto Cabrera Pérez		llama-bench: fixed size of fields to correctly map to values (#13183)
1831f538f720d1d99fba146f24f0a8e970838cc4	4e87962e34a4b257ec374c4baf6b1568554b81a9	2025-04-28T20:20:39+05:30	Vishal Agarwal		llama-bench: add `-d` depth arg (#13096)
c0a97b762e5ec767dc414f0dc4979befd4c09a52	ced44be34290fab450f8344efa047d8a08e723b4	2025-04-27T14:48:26-07:00	4onen		llama-bench : Add `--override-tensors` arg (#12922)
2d451c80590b9ac250322769ac13d3b4870dbcf7	4753791e70acd4d4e02f2098f14a03df26c992bd	2025-04-26T22:58:12+02:00	Xuan-Son Nguyen		common : add common_remote_get_content (#13123)
77d5e9a76a7b4a8a7c5bf9cf6ebef91860123cba	d5fe4e81bd447124836ecfb47d794f8768665b9f	2025-04-26T22:05:31+08:00	SXX		ggml: move fp16/bf16 conversion optimizations to CPU backend + export conversion APIs (#13107)
553a5c3a9fdf771be2101bc3529937963f817457	13be08daf992c89d5169518229b3740041c0f419	2025-04-25T10:08:08+03:00	Radoslav Gerganov		rpc : do not wait for response when sending RPC_CMD_SET_TENSOR (#12943)
7a395f67a7a02bb361d944b816d6e933889e28e1	971f245b3b5f3f55991bb779cb541b00f82eea1d	2025-04-17T20:34:16+08:00	hipudding		CANN: Add support for async operator submission (#12864)
015022bb53387baa8b23817ac03743705c7d472b	b43d89e311c5e7fbf62e5ec3c0401eb536677267	2025-04-16T13:37:25-05:00	Jeff Bolz		vulkan: enable coopmat2 FA gqa and split_k optimizations more often (#12931)
84778e97703740d8ac5fb64e14d83b80eafa0f3c	510676475f885ec064ff147af9f20ee7a9b12a50	2025-04-15T17:20:38+08:00	David Huang		CUDA/HIP: Share the same unified memory allocation logic. (#12934)
daa422881a0ec7944771bcc8ff8de34d11f5bd3b	eccc7a1602f0752507de4aaad1008b9618a282c8	2025-04-15T07:49:57+01:00	Juk Armstrong		llama : DeepSeek V2/V3 MLA implementation (#12801)
0019279bb56f028e4eee19b59b19750736c719f7	b0c75ac9f93322b45e3766e149417334b4fd1ed9	2025-04-15T10:09:35+08:00	Chenguang Li		CANN: Opt ROPE optimization (#12865)
b0c75ac9f93322b45e3766e149417334b4fd1ed9	d6d2c2ab8c8865784ba9fef37f2b2de3f2134d33	2025-04-15T10:04:24+08:00	Xinpeng Dou		CANN: Optimize CANN buffer pool memory management (#12875)
d6d2c2ab8c8865784ba9fef37f2b2de3f2134d33	75afa0ae31f0a51aaadcc5ff146eb7a32a7f9088	2025-04-15T01:18:20+08:00	Russyyds		Add performance print for gemma3 in example (#12929)
8b9cc7cdd8a0dcf0176c60c755322c95b5965299	64eda5deb9859e87a020e56bab5d2f9ca956f1de	2025-04-10T22:57:16+02:00	Xuan-Son Nguyen		llava : introduce libmtmd (#12849)
0090950f679475c5ecaac2f7bca5049cca96492b	7ecd780b1a1d5214b8d04c25ebfc194d310816ed	2025-04-09T00:25:08-05:00	Jeff Bolz		vulkan: In coopmat2 mmq, load q4_k/q5_k scales through shared memory (#12833)
8ca6e1c3a4deb6bb27ee294bfd5706098d94ae88	656babd6c21a3b9b3622324cfcc80a2ab78da25b	2025-04-08T14:14:59+05:00	characharm		server : webui : Improve Chat Input with Auto-Sizing Textarea (#12785)
6bf28f0111ff9f21b3c1b1eace20c590281e7ba6	f1e3eb4249db68d97be352c0adf16eef7ae53795	2025-04-05T18:04:03+02:00	0cc4m		Vulkan: Tune Vulkan mmq int dot shader for performance (#12767)
74d4f5b041ad837153b0e90fc864b8290e01d8d5	35e592eb30832187412360912ab8f2f5b7984df1	2025-04-04T00:54:35-05:00	Jeff Bolz		vulkan: Hybrid waitForFences/getFenceStatus to reduce fence latency (#12630)
c262beddf29f3f3be5bbbf167b56029a19876956	5dd5d1ab00d074e3b7c02ca3ae12f6bf3e86336a	2025-04-03T21:50:29+05:30	Gaurav Garg		CUDA: Prefer vector flash decoding kernel for Gemma models (#12738)
be0a0f8cae039e2286f757612accebfb8f21b36e	92e3006bb69dfeb656ccf5c7c1c1efadb03c88c2	2025-04-02T12:40:32-05:00	Jeff Bolz		vulkan: Implement grouped query attention in the coopmat2 FA shader (#12559)
9bacd6b37461608385360fd64326c13247ccf18e	267c1399f15a278ec8c3cdcf9c90dc94151fbc38	2025-04-02T15:22:13+08:00	Chenguang Li		[CANN] get_rows and dup optimization (#12671)
e408d4351a4ad143656672479d8ae1479306ce31	3891e183c61c1e25c2c61afe68d7b95019ea3f89	2025-03-24T09:53:38+01:00	Daniel Bevenius		ggml : add logging for native build options/vars (whisper/2935)
3891e183c61c1e25c2c61afe68d7b95019ea3f89	af6ae1efb27a9a7c3f7f7f84639d2243f7303ac1	2025-03-20T07:02:18+01:00	Daniel Bevenius		examples : command.wasm updates (whisper/2904)
b4ae50810e4304d052e630784c14bde7e79e4132	b86f6007234da4bff51a3ebef2bdb952b52059c6	2025-03-28T20:21:59+02:00	Georgi Gerganov		metal : improve FA + improve MoE (#12612)
5d01670266859444366e4f333ade5e0e5e2ae63d	ef03229ff423dd1991f4f44ef1352f03334d86eb	2025-03-28T01:05:44-07:00	Benson Wong		server : include speculative decoding stats when timings_per_token is enabled (#12603)
13731766db91ec927c1b61bf502ac7a9be2b11b9	ab6ab8f809bd514d88e02a63869b4d619f13fa86	2025-03-28T13:13:22+05:30	amritahs-ibm		llamafile : ppc64le GEMV forwarding for FP32. (#12594)
c7b43ab60855f752ae79937fb93d561bc30b69a4	24feaec05792b972d5ff3e2b12d9237ebd50d1ac	2025-03-27T12:21:47+05:30	amritahs-ibm		llamafile : ppc64le MMA implementation for Q4_0. (#12489)
ef19c71769681a0b3dde6bc90911728376e5d236	053b3f9aae63151732eccf6b7408c6418ba8746e	2025-03-25T17:46:11Z	Eric Curtin		run: de-duplicate fmt and format functions and optimize (#11596)
e2f560175a195f63c3276972a3d1caec0bd13e05	36ee06dd2dcf8d3d1157ebe36366d7670770e75f	2025-03-25T16:10:18+05:30	Akarshan Biswas		SYCL: disable Q4_0 reorder optimization (#12560)
eddfb438502bd5d1014d63a812e9b6d03d326f8c	4375415b4abf94fb36a5fd15f233ac0ee23c0bd1	2025-03-22T03:40:11-05:00	Jeff Bolz		vulkan: Optimize mul_mat_vec p021 and nc shaders (#12505)
732b5fbf5e7f9cf069942f0c5850ee959ef321ba	568013d0cd3d5add37c376b3d5e959809b711fc7	2025-03-20T02:36:37-04:00	Bartowski		convert : avoid calls to tokenizer.added_tokens_decoder (#12473)
517b5ddbf002b91fd6d6daf5d8db8c88a0173039	a9b59288e222f39fc0311dc66944ed5a86c815fa	2025-03-20T01:22:06+05:30	Gaurav Garg		CUDA: Improve flash decoding kernel GPU occupancy for BS=1 case (#12183)
a9b59288e222f39fc0311dc66944ed5a86c815fa	0fd8487b142b2b92565bc95b39ddc440955a237c	2025-03-19T13:56:23-05:00	Jeff Bolz		vulkan: optimize iq1 coopmat2 dequant functions (#12427)
c446b2edd2a9fe2772a1a18923c3e54a6749c364	d84635b1b085d54d6a21924e6171688d6e3dfb46	2025-03-19T02:26:26-05:00	Jeff Bolz		vulkan: Submit once enough matmul work has been recorded (#12406)
d84635b1b085d54d6a21924e6171688d6e3dfb46	75422e8bc42646005be0754f7aa438b97a5e777e	2025-03-18T12:54:55-07:00	lhez		opencl: improve profiling (#12442)
484a8ab513bbd740cc49f30280c1acf52cb4e7e9	cf2270e4d3685ac46f4a166d8718997ba7cbc45a	2025-03-17T09:26:18-05:00	Jeff Bolz		vulkan: Add N/2 and N/4 optimized paths in coopmat2 shader (#12312)
891c63956dbfbdf7ed2ecd0b5882cff49dbfe90f	2f21123c1deb3ce1be3c0578c5f6980fe19ed077	2025-03-17T04:41:59-05:00	Jeff Bolz		vulkan: Pad N dimension of B matrix for coopmat2 perf, to avoid bounds checking (#12273)
92a391327e9201b9b5b32fdd3afb452026c22d4c	9f2250ba722738ec0e6ab684636268a79160c854	2025-03-15T09:31:08+08:00	Chenguang Li		[CANN]MUL_MAT optimization (#12382)
9f2250ba722738ec0e6ab684636268a79160c854	774973b8f3d5e375b0b74d58638eeb1817e950a8	2025-03-14T16:41:20Z	Eric Curtin		Add CLI arg to llama-run to adjust the number of threads used (#12370)
5e2d57b2b2e43eadbe6d66ba3e873a824b95e725	f1648e91cf6c52e9593810aa70857e412d474c09	2025-03-07T15:35:57+08:00	BB-fat		metal : simplify kernel arguments using a struct (#3229) (#12194)
e721c05c9336a72fbb59d5c75967360bc67036c6	57b6abf85acb1f697913a44e5e105e333d894b78	2025-03-06T08:20:52+01:00	uvos		HIP/CUDA: set the paramerter value in maintain_cuda_graph instead of replaceing it. (#12209)
ed4ce0dda24986c80d58e2ce112049af00f632f4	07d15723470a0a5b15d8ccad1aff5b20354ffbe1	2025-03-06T09:30:05+08:00	simon886212		opencl : fix profile-related errors (#12095)
669912d9a5bf927312c553332ff997f0a99da8fb	fa31c438e0e709242ab7334d26fc3be3dcda07a0	2025-03-05T13:05:13Z	Olivier Chafik		`tool-call`: fix Qwen 2.5 Coder support, add micro benchmarks, support trigger patterns for lazy grammars (#12034)
3ccbfe5a71c74ac574b00607067d0aa0a49df04c	06a92a193a07afe445929607be9d5e4d033956fb	2025-03-05T08:34:02+01:00	Daniel Bevenius		ci : remove xframework upload (#12190)
2679c3b55d1c9c3fed56dea00fea713622e42594	c43af9276b119dae7436b7878d59671d0f6b1a97	2025-03-03T16:17:36+01:00	Daniel Bevenius		ci : set GITHUB_ACTION env var for server tests (#12162)
7b69003af74924ac04d97313c2e57c45ba56b616	ece9745bb8079b9f4af45df29b67ad0c6e50584d	2025-03-03T11:42:45+01:00	Xuan-Son Nguyen		webui : add ?m=... and ?q=... params (#12148)
438a83926afcff3643ffef5543db67545ceffe39	9c42b1718ca8299f9afeabdc122badeab64c9690	2025-02-28T09:42:52+01:00	Rémy O		vulkan: add specific MMV kernels for IQ2 and IQ3 quants + optimizations (#11595)
581650b7cacec2872982fde381bd3bcda0f78699	b95c8af37ccf169b0a3216b7ed691af0534e5091	2025-02-28T06:52:51Z	Daniele		vulkan: improve im2col (#11826)
7a2c913e66353362d7f28d612fd3c9d51a831eda	08d5986290cc42d2c52739e046642b8252f97e4b	2025-02-24T09:09:51-07:00	Alex Brooks		llava : Add Granite Vision Support (#11794)
08d5986290cc42d2c52739e046642b8252f97e4b	651adf4b6675339465179b81b194d98cc14704d6	2025-02-24T22:33:23+08:00	Neo Zhang Jianyu		[SYCL] Optimize mul_mat for Q4_0 on Intel GPU (#12035)
af7747c95ae71a5db4184947f837179e82c70b77	a28e0d5eb18c18e6a4598286158f427269b1444e	2025-02-23T05:39:24+08:00	Aaron Teo		ggml-cpu: Support s390x SIMD Instruction Set (#12019)
36c258ee921dbb5c96bdc57c0872e4a9a129bef6	f3e64859edb0d55d4223ead78672597cd1a218df	2025-02-22T22:28:28+08:00	Ting Lou		llava: build clip image from pixels (#11999)
5fa07c2f93c73161bf09ef0b23b5d2686f9a073e	335eb04a91f481f37c0c9b302ee31b449b04c3e9	2025-02-22T12:20:17+01:00	Johannes Gäßler		CUDA: optimize FA for GQA + large batches (#12014)
6dde1782483d6b0a1d59f5a5fbcb3119b9d34c27	fc10c38ded84670955d4c44770f8acfb64617e15	2025-02-15T20:23:22+01:00	Johannes Gäßler		scripts: fix compare-llama-bench commit hash logic (#11891)
22885105a6b034abaf1c1471ad90fb2ae96146bb	c2cd24fbfdfeacc2fc6ad03878379de104264114	2025-02-15T19:39:20+01:00	Adrian Kretz		metal : optimize dequant q6_K kernel (#11892)
fc1b0d0936e4dfc52a81f38e7420c7d23f6caa88	89daa2564f6eab33be53c6a1b39273af536d6bb3	2025-02-15T09:01:40+01:00	Rémy O		vulkan: initial support for IQ1_S and IQ1_M quantizations (#11528)
38e32eb6a0cec32130b699ce9fefad6c74571953	a4f011e8d02179f032627130f961eb77ee30401c	2025-02-14T16:54:27+08:00	Jinyang He		ggml: optimize some vec dot functions for LoongArch ASX (#11842)
a7b8ce226071b2b0faaad0d36cc5ebd7fb074730	04045bb84288dc7e9e424d4e2bf7f40d259b4c6a	2025-02-14T09:13:43+08:00	theraininsky		llama-bench : fix unexpected global variable initialize sequence issue (#11832)
27e8a23300e30cd6ff6107ce262acf832ca60597	e4376270d971cff7992bdb6c5412a739195b1459	2025-02-13T00:45:57-06:00	Vinesh Janarthanan		sampling: add Top-nσ sampler (#11223)
be3bbd62153820ff6d358c817360927f429105c4	31afcbee0eebbc998ab311aa314e389d60aa2127	2025-02-13T00:33:45+01:00	Xuan-Son Nguyen		ggml : x2 speed for WASM by optimizing SIMD (#11453)
c2a67efe38e6782c0217e1de3edc68de6951612b	b044a0fe3ca0cbef9dd041edce3ebda8c501fae4	2025-02-10T07:17:21+01:00	Danny Milosavljevic		vulkan: Make Vulkan optional at runtime (#11493). (#11494)
55ac8c7791ff44aeb82bd7fe3ca2041844fc77f1	e6e658319952f7ad269dc11275b9edddc721fc6d	2025-02-08T21:54:50+01:00	Xuan-Son Nguyen		server : (webui) revamp Settings dialog, add Pyodide interpreter (#11759)
2d219b389e8c8c40bce547b08c8aa7add60fde1f	333820d7491cd31c707a340ff23b984a84e40154	2025-02-07T08:55:47-05:00	Christian Fillion		vocab : ignore invalid UTF-8 input in the BPE tokenizer (#11729)
7ee953a64a40c09438b2064539becdbc577cefd0	ec3bc8270bc67b58955748d40a3e558a05b2d8f2	2025-02-07T04:33:27-05:00	Christian Fillion		llama : add llama_sampler_init for safe usage of llama_sampler_free (#11727)
225bbbfa39930cda38a2e5d1f3e5b38226732009	855cd0734aca26c86cc23d94aefd34f934464ac9	2025-02-07T15:38:31+08:00	Jinyang He		ggml : optimize and build warning fix for LoongArch (#11709)
2fb3c32a1634488a5265d1304ab37628eeb5480d	9ab42dc722ad19a12af80f38a06474d498f96da3	2025-02-06T17:32:29+01:00	Xuan-Son Nguyen		server : (webui) migrate project to ReactJS with typescript (#11688)
2c6c8df56d8a3edd657b9a295e95d469a37f0044	8a7e3bf17aa5a8412854787746c92a28623a8925	2025-02-06T00:15:30-06:00	Jeff Bolz		vulkan: optimize coopmat2 iq2/iq3 callbacks (#11521)
84ec8a58f7b6aad6887bbfbd1321f3ff417341a5	bfcce4d693617ec843d0b2510f6ee16e6bc6720d	2025-02-02T16:14:48+01:00	Eric Curtin		Name colors (#11573)
ff227703d6d6e1888bdc7af6138514092ffcdb96	0cec062a638700495673f5494d200b74340538be	2025-02-01T23:55:32-08:00	Michał Moskal		sampling : support for llguidance grammars (#10224)
66ee4f297cff3c7ce98b31dbc0ce909d41b9e408	e51c47b401f8cb5f21630a05171e2529cde4d186	2025-01-29T18:29:39+01:00	Rémy Oudompheng		vulkan: implement initial support for IQ2 and IQ3 quantizations (#11360)
be5ef7963fcf14a9c77c963fdd3f7b606eacb498	cae9fb4361138b937464524eed907328731b81f6	2025-01-28T23:06:32+01:00	uvos		HIP: Supress transformation warning in softmax.cu
cae9fb4361138b937464524eed907328731b81f6	7fee2889e6565830631fbe76d47ef85cf8fd946a	2025-01-28T07:42:20-08:00	Nikita Sarychev		HIP: Only call rocblas_initialize on rocblas versions with the multiple instantation bug (#11080)
f643120bad8ab3a753daa64aaac8288ee5800e06	6e84b0ab8e10b8f6f99a32855f976ebcd35b0353	2025-01-28T11:42:32+01:00	Nuno		docker: add perplexity and bench commands to full image (#11438)
a5203b4465c5c87813936bde98170e25bb09024f	df984e014714cba4c99ef894b20b51cbcef31b16	2025-01-27T17:42:09+04:00	lexasub		llama : minor fixes for up llama load model speed (#11448)
4a75d19376f2f00dbae6c266eb9c4f3001872b52	26771a1491f3a4c3d5b99c4c267b81aca9a7dfa0	2025-01-25T15:29:57-06:00	Jeff Bolz		vulkan: compile shaders on-demand (#11406)
58456616408c828a1ce6d2481940fd1c97bce3b5	f211d1dc10332b7e89a4abd1041903fa63bfed27	2025-01-23T13:56:05+01:00	Xuan Son Nguyen		server : add more clean up when cancel_tasks is called (#11340)
92bc493917d43b83e592349e138b54c90b1c3ea7	b9daaffe02d6a77d85f0420bce5dfe0e00daeff6	2025-01-19T20:22:30+02:00	Georgi Gerganov		tests : increase timeout when sanitizers are enabled (#11300)
44e18ef93995f3040660750b527e5becf85899d0	3edfa7d3753c29e44b964c0ff424d2ea8d5fdee6	2025-01-18T02:26:50-06:00	Jeff Bolz		vulkan: fix coopmat2 flash attention for non-contiguous inputs (#11281)
466300fe1416de2802b710215817db28d4496f41	206bc53422521a67de5e2caba661e21d590d2bae	2025-01-16T15:23:49-06:00	Jeff Bolz		vulkan: optimize coopmat2 q4_k/q5_k dequant functions. (#11206)
206bc53422521a67de5e2caba661e21d590d2bae	4dbc8b9cb71876e005724f4e8f73a3544646bcf5	2025-01-16T15:16:39-06:00	Jeff Bolz		vulkan: optimize coopmat2 q2_k dequant function (#11130)
adc5dd92e8aea98f5e7ac84f6e1bc15de35130b5	f11cfdfd7fe29436fce512d934c2ff6b94bd89d2	2025-01-15T19:50:13Z	Eve		vulkan: scale caching for k quants + misc fixes (#11081)
39509fb082895d1eae2486f8ad2cbf0e905346c4	a29f0870d4846f52eda14ae28cea612ab66d903c	2025-01-13T16:45:53+01:00	Andreas Kieslinger		cuda : CUDA Graph Compute Function Refactor (precursor for performance improvements) (#11042)
afa8a9ec9b520137bbd1ca6838cda93ee39baf20	c05e8c9934f94fde49bc1bc9dc51eed282605150	2025-01-12T11:32:42+02:00	Georgi Gerganov		llama : add `llama_vocab`, functions -> methods, naming (#11110)
8cef75c743ba13ebbd6d380c531200c768a8b8aa	0d52a69e4bf0d6181beec7853307bdcdeec9905b	2025-01-08T16:24:19+05:30	amritahs-ibm		llamafile : ppc64le MMA INT8 implementation (#10912)
02f04301417e7fb44fa1025bc1b0aef866e2ca89	bec2183f2c8d37cf1278c11d1adb9311e9eaa242	2025-01-08T09:18:13+01:00	Mathieu Baudier		Disable GL_KHR_cooperative_matrix Vulkan extension if not available. (#11117)
2f0ee84b9b02d2a98742308026f060ebdc2423f1	0da5d860266c6928b8c9408efbd264ae59fedda6	2025-01-02T18:06:12+01:00	Pierrick Hymbert		server: bench: minor fixes (#10765)
716bd6dec3e044e5c325386b5b0483392b24cefe	c250ecb3157f3bae0a45f44c3c953b5414d4c2f7	2024-12-30T11:27:11-06:00	Jeff Bolz		vulkan: optimize mul_mat for small values of N (#10991)
a813badbbdf0d38705f249df7a0c99af5cdee678	fdd21889123bec62b1db3b2fc22b5a4abab32174	2024-12-29T03:16:34-06:00	Jeff Bolz		vulkan: im2col and matmul optimizations for stable diffusion (#10942)
2cd43f4900ba0e34124fdcbf02a7f9df25a10a3d	09fe2e76137dde850b13313f720e7ffa17efdefa	2024-12-24T18:54:49+01:00	Djip007		ggml : more perfo with llamafile tinyblas on x86_64 (#10714)
7024d59e6a572730626cb11896829d115043a1b1	7c0e28585843b366864b43b48f92425e2ea17df6	2024-12-22T16:20:11-08:00	yuri@FreeBSD		ggml : fix run-time on FreeBSD in get_executable_path() (#10948)
a91a41364b25705dbb81ae996bc35c3440c63b35	e34c5af43f941f0ddb92466776339897295aca11	2024-12-21T01:04:45-06:00	Jeff Bolz		vulkan: optimize coopmat2 dequant functions (#10855)
644fd71b44c4cdbfc6482fbf0353d289c3bc29e6	4ddd199f6f6b980e0a7ed9f9b44efeae2fbdf5c4	2024-12-16T12:31:14+02:00	Georgi Gerganov		sampling : refactor + optimize penalties sampler (#10803)
87cf323cef80f6aa530f047ab05b539ebc6b7e3c	5478bbcd173e7027af7689493c7421719f5c43df	2024-12-15T18:44:47+02:00	Georgi Gerganov		scripts : change build path to "build-bench" for compare-commits.sh (#10836)
e52aba537a34d51a65cddec6bc6dafc9031edc63	ba1cb19cdd0d92e012e0f6e009e0620f854b6afd	2024-12-14T18:17:36Z	Evgeny Kurnevsky		nix: allow to override rocm gpu targets (#10794)
64ae0655114f84f11a724bc6878c6f8f4a55560b	83ed24a97b500ccdb32b90b94e6f9621ad8db79e	2024-12-13T08:42:04Z	Eve		vulkan: small mul_mat_vec optimizations (#10665)
235f6e14bf0ed0211c51aeff14139038ae1000aa	1a31d0dc00ba946d448e16ecc915ce5e8355994e	2024-12-11T20:52:14+01:00	Xuan Son Nguyen		server : (UI) add tok/s, get rid of completion.js (#10786)
19d8762ab61df8286367588a80b9c7db4cb568db	c2a16c0bdbe2e51adf318918bad82f0c3e3d6f3b	2024-12-07T13:37:50+01:00	Djip007		ggml : refactor online repacking (#10446)
3df784b3050f657ea681f804187ce5bddb433e88	86a1934978ce5daffc7e5b4251f6540ee5e7b47b	2024-12-07T10:24:15+01:00	0cc4m		Vulkan: VK_KHR_cooperative_matrix support to speed up prompt processing (#10597)
40c6d79fb52f995f47507fedfeaae2ac05d9b35c	98036d5670f21e9b9a99d5e3dbb3bf7589f5c4e3	2024-12-04T02:29:20+01:00	Nicolò Scipione		SYCL : Move to compile time oneMKL interface backend selection for NVIDIA backend (#10584)
cc98896db858df7aa40d0e16a505883ef196a482	91c36c269bca75b2d08119c653512cd20b4ea2ba	2024-12-03T13:29:54-06:00	Jeff Bolz		vulkan: optimize and reenable split_k (#10637)
efb6ae963031709fc331e6e48cc4606ac8f9c3a7	667d70d1704dfa6977505f5d01d4638669b90dce	2024-12-02T19:27:24+01:00	PAB		feat: add `GGML_UNARY_OP_ARGMAX` Metal kernel (ggml/1019)
64ed2091b24b2f9747148fdf49a34ed5938762c3	991f8aabeec89d801300bb179e52013fb0eb0584	2024-12-02T21:45:54+08:00	haopeng		server: Add "tokens per second" information in the backend (#10548)
938f6087421889a3af7d0786c64406ced2be81b8	f095a649ec390e04dfab1b04e646ae8549dafaef	2024-11-29T14:46:55+08:00	Chenguang Li		CANN: RoPE operator optimization (#10563)
f095a649ec390e04dfab1b04e646ae8549dafaef	678d7994f4da0af3d29046be99950ac999ee9762	2024-11-29T00:18:02-06:00	Jeff Bolz		vulkan: get the first command buffer submitted sooner (#10499)
c202cef1686182a78f8f4e253ab8d0c0ffe2fcc8	2025fa67e94358deda4740a74fe9803916cb2f60	2024-11-28T20:52:03+08:00	Shupei Fan		ggml-cpu: support IQ4_NL_4_4 by runtime repack (#10541)
b7420131bf8ab3e067bc660439ab1ab18be7edbd	9f912511bc9414fa7a3c521378b6388cd932b58d	2024-11-28T14:24:46+08:00	Chenguang Li		CANN: ROPE operator optimization (#10540)
3ad5451f3b75809e3033e4e577b9f60bcaf6676a	46c69e0e752ff16206347bb12f96ed69f4a01abf	2024-11-27T17:10:08+01:00	uvos		Add some minimal optimizations for CDNA (#10498)
46c69e0e752ff16206347bb12f96ed69f4a01abf	9e2301f4a4ef1690bd99360c11de43fe830b1c8d	2024-11-27T11:03:25+01:00	Diego Devesa		ci : faster CUDA toolkit installation method and use ccache (#10537)
249a7902ec710c8d027b9cc0ed10219d2b4184f8	71a64989a5d2e25c13507efada145f12cf358914	2024-11-27T01:21:59-06:00	Jeff Bolz		vulkan: further optimize q5_k mul_mat_vec (#10479)
4a57d362e1948ada50af997a92c3cbff9711e78b	c9b00a70b080d5c0668608024afc3e0e2fed822f	2024-11-27T01:00:50-06:00	Jeff Bolz		vulkan: optimize Q2_K and Q3_K mul_mat_vec (#10459)
9a4b79bcfa4338b922fa8cf903bd5ac058aaf46f	7066b4cce2898993e943ad6af5d8f1de5840c8e9	2024-11-26T18:08:37+08:00	Shanshan Shen		CANN: Improve the Inferencing Performance for Ascend NPU Device (#10454)
7066b4cce2898993e943ad6af5d8f1de5840c8e9	0eb4e12beebabae46d37b78742f4c5d4dbe52dc1	2024-11-26T17:31:05+08:00	Chenguang Li		CANN: RoPE and CANCAT operator optimization (#10488)
6dfcfef0787e9902df29f510b63621f60a09a50b	599b3e0cd40432cd1975a8906f3db70bbe53b627	2024-11-22T17:44:08+08:00	蕭澧邦		ci: Update oneAPI runtime dll packaging (#10428)
a5e47592b6171ae21f3eaa1aba6fb2b707875063	1bb30bf28cb5a7adf111bc41c935bdaf128397e7	2024-11-21T18:18:50+01:00	Diego Devesa		cuda : optimize argmax (#10441)
1bacb9f62514b520bdf74ed6feb46c80508dad38	ad21c9e1f14d82b8c15ae369a8839019e3d498b4	2024-11-20T01:11:00-06:00	Jeff Bolz		vulkan: further optimize mul_mat_vec using larger loads (#10387)
2a1507c1629975d9d20a503d6a14f44eff292c25	b3e585988fc65d3a8083c6d94dfc0629f9ce226d	2024-11-19T09:02:23+01:00	Romain Biessy		sycl : Add option to set the SYCL architecture for all targets (#10266)
b3e585988fc65d3a8083c6d94dfc0629f9ce226d	557924f22237c76387a39c4db5abae154d57e754	2024-11-19T01:25:17-06:00	Jeff Bolz		vulkan: Optimize soft_max (#10301)
cf32a9b93ad859ea592c31785b6bd3b4b2121463	a43178299c2f74f14bcfc659bfd9fd32d931d1f4	2024-11-17T11:23:01+02:00	Georgi Gerganov		metal : refactor kernel args into structs (#10238)
8a43e940ab0daaff198809bf9277289994ec62f5	5c9a8b22b10132db620529435e3cfa49304b65cc	2024-11-16T22:17:59+02:00	Johannes Gäßler		ggml: new optimization interface (ggml/988)
f245cc28d4eb900efad0bc740145f58d713c6e4f	772703c8fffdd83d2e28f60119e83525f1189412	2024-11-16T10:32:50+02:00	Georgi Gerganov		scripts : fix missing key in compare-llama-bench.py (#10332)
772703c8fffdd83d2e28f60119e83525f1189412	dd3a6ce9f84d21ba05fe98af9f983bdea0398e6c	2024-11-16T00:26:57-06:00	Jeff Bolz		vulkan: Optimize some mat-vec mul quant shaders (#10296)
1e58ee1318429a3e97aa66f3034cdfd65ffc6c34	89e4caaaf081f4712af61a3e08cb67b406c02b80	2024-11-16T01:53:37+01:00	Dan Johansson		ggml : optimize Q4_0 into Q4_0_X_Y repack (#10324)
4047be74da398acb8717a4d21b77b929ad7ed4f7	883d206fbd2c5b2b9b589a9328503b9005e146c9	2024-11-15T21:19:03+01:00	Johannes Gäßler		scripts: update compare-llama-bench.py (#10319)
18429220bdb344da1bc7df9bc580c7b41b3cd57b	f0204a0ec70d50ca60e07bc0096ec1d6508ab0c7	2024-11-15T11:47:58Z	Eve		AVX BF16 and single scale quant optimizations (#10212)
5a54af4d4f588f109f31e456483fdf77096399d9	1607a5e5b08f4e55f118af3d7de325949d8f1835	2024-11-15T04:09:12+01:00	Romain Biessy		sycl: Use syclcompat::dp4a (#10267)
af148c9386da825a60c7038549c121c35ca56b50	66798e42fbe636f1cb6236e4bc30939d23ef7c25	2024-11-13T23:22:55-06:00	Jeff Bolz		vulkan: Optimize binary ops (#10270)
2e82ffa4af29f87e7d3d6dff8060a2a79613b72f	80dd7ff22fd050fed58b552cc8001aaf968b7ebf	2024-11-13T09:40:57Z	Alberto Cabrera Pérez		sycl : Fixes to broken builds and test-backend-ops (#10257)
80dd7ff22fd050fed58b552cc8001aaf968b7ebf	54ef9cfc726a799e6f454ac22c4815d037716eda	2024-11-13T00:58:57-06:00	Jeff Bolz		vulkan: Optimize contiguous copies (#10254)
e89213492d3e01705739789733f0f2d250b4c449	8fc393f246c550d2481e53323a47644a94e8d01f	2024-11-09T12:47:50+05:30	amritahs-ibm		ggml : optimize llamafile cpu matrix multiplication for ppc64le (#10156)
841f27abdbbcecc9daac14dc540ba6202e4ffe40	d05b3127bd30515955aa4ee2bacdb68ebafe88f4	2024-11-08T13:47:22+02:00	Georgi Gerganov		metal : optimize FA kernels (#10171)
3bcd40b3c593d14261fb2abfabad3c0fb5b9e318	5c333e014059122245c318e7ed4ec27d1085573c	2024-11-07T18:19:10+11:00	Zhiyuan Li		Optimize RWKV6 Operator Naming and Implement Multi-core CPU/ SYCL Acceleration (#10133)
340736477651095a98a3b10e19b038ec62593a1d	d5a409e57fe8bd24fef597ab8a31110d390a6392	2024-11-04T22:06:31Z	Eve		Q6_K AVX improvements (#10118)
fc83a9e58479e4dd70054daa7afe5184c1bbe545	c5b0f4b5d90297f3e729fca7f78ddb25fcab5ddc	2024-10-30T15:00:40+08:00	xctan		ggml : add Q4_0_8_8 RISC-V GEMV and GEMM kernels (#10029)
8f275a7c4593aa34147595a90282cf950a853690	8d8ff715367480b856ad86ac3888e9742b13a6fa	2024-10-29T17:52:56+09:00	Changyeon Kim		ggml: Add POOL2D OP for GPU acceleration to the Vulkan backend in the MobileVLM model. (#9763)
2d3aba9ee8da9c026d54e8a912a1d64f56809be3	80273a306d07ed95059d6130389deacb3b2d7196	2024-10-23T17:16:56+03:00	Georgi Gerganov		llama.vim : bump generation time limit to 3s [no ci]
4c9388fb96ac2415fbb1239b7ba8346616606e2e	873279b1592e433c4d9eb5065091cc98473c7bee	2024-10-23T19:33:45+09:00	Jun Hee Yoo		metal : add POOL2D and fix IM2COL (#9943)
cda0e4b648dde8fac162b3430b14a99597d3d74f	afd9909a6481402844aecefa8a8908afdd7f52f1	2024-10-18T23:18:01+02:00	Xuan Son Nguyen		llama : remove all_pos_0, all_pos_1, all_seq_id from llama_batch (#9745)
13dca2a54a394757d56fdd652b9f0df08f44ea22	d4c19c0f5cdb1e512573e8c86c79e8d0238c73c4	2024-10-14T01:49:08+01:00	agray3		Vectorize load instructions in dmmv f16 CUDA kernel (#9816)
edc265661cd707327297b6ec4d83423c43cb50a5	1bde94dd024b632f98428f4bf2ce483295130779	2024-10-12T16:14:27+03:00	Georgi Gerganov		server : add option to time limit the generation phase (#9865)
a39ab216aa624308fda7fa84439c6b61dc98b87a	f536f4c4391bec74c432a924625c04e8c484d3ee	2024-10-02T15:49:55+02:00	Xuan Son Nguyen		llama : reduce compile time and binary size (#9712)
76b37d1541a880b9645557c8715a343fd074cc5c	148844fe97fff4c1563a3111bf238ba4dd22ef56	2024-10-02T15:21:57+08:00	Zhenwei Jin		gguf-split : improve --split and --merge logic (#9619)
148844fe97fff4c1563a3111bf238ba4dd22ef56	3f1ae2e32cde00c39b96be6d01c2997c29bae555	2024-10-02T10:14:44+03:00	Georgi Gerganov		examples : remove benchmark (#9704)
cb00020504416606601f8cb35f55ee07b710b4b7	6c5322481a75f1be54e58a000c3f78484d07f948	2024-09-30T09:14:09+02:00	Salvatore Mesoraca		vulkan : mul_mat: fix UB with small warps (ggml/952)
1b2f992cd2cff0b69e5abe78bb8888d51ed19d67	739842703e32cd43443c45e0b4f6647cc4e6b3d6	2024-09-28T14:32:46+02:00	slaren		test-backend-ops : use flops for some performance tests (#9657)
6a0f7794847244fb3b99a983e03137d7e832b585	89f9944981010d195e411a9fbfbb19959412f710	2024-09-28T14:06:16+02:00	Dan Johansson		ggml : add run-time detection of neon, i8mm and sve (#9331)
1e436302188a704ac9ea044af03193648806f19c	afbbfaa537a96f562c34df4542930fa951b40d9e	2024-09-25T15:12:20+02:00	Charles Xu		ggml : remove assert for AArch64 GEMV and GEMM Q4 kernels (#9217)
904837e0cb2f5f01bf5d5901b7aa57a026860ae4	70392f1f81470607ba3afef04aa56c9f65587664	2024-09-25T11:30:38+08:00	Dou Xinpeng		cann: fix crash when llama-bench is running on multiple cann devices (#9627)
f0c7b5edf82aa200656fd88c11ae3a805d7130bf	1d48e98e4f3316bd2f6b187d288c7b6cb88d5cb3	2024-09-23T11:42:43-07:00	Max Krasnyansky		threads: improve ggml_barrier scaling with large number of threads (#9598)
424c5d00a9b97dd5559635872db9b57f87c23b02	a6809c6a2e8163cba296d4cc0c5cd4bbc8073638	2024-09-20T19:04:44+03:00	Johannes Gäßler		ggml/examples: add backend support for numerical optimization (ggml/949)
7be099fa817e9c53ffb4c3ed7d063e1cffcd675a	8b836ae731bbb2c5640bc47df5b0a78ffcb129cb	2024-09-17T22:41:38+02:00	Michael Podvitskiy		llama-bench: correct argument parsing error message (#9524)
acb2c32c336ce60d765bb189563cc216e57e9fc2	a6a3a5c531c73aef85750a847d21e7d4671e723d	2024-09-16T13:07:13+02:00	Daniel Bevenius		llama : rename n_embed to n_embd in rwkv6_time_mix (#9504)
5c3d0f1824714e9a97fc9b06e046eefcb6ecc721	0aadac10c7dd704f8285ddf5a63d6f764cb340aa	2024-09-16T06:48:24Z	Eve		ggml : IQ4_NL sgemm + Q4_0 AVX optimization (#9422)
0abc6a2c25272d5cf01384dda8ee8bfec4ba8745	bd35cb0ae357185c173345f10dc89a4ff925fc25	2024-09-13T09:53:38+03:00	Georgi Gerganov		llama : llama_perf + option to disable timings during decode (#9355)
d2b496bff4f353a6429f8e833448f071bd237ba7	b34e02348064c2f0cef1f89b44d9bee4eb15b9e7	2024-09-11T10:03:54+03:00	Georgi Gerganov		batched-bench : remove unused code (#9305)
0b4ac75772b744bb0a0d674927587621d1057884	fb3f2498156b3140e2050ec9c7bf61372f63ff56	2024-09-10T15:02:30+08:00	Molly Sophia		RWKV v6: Add time_mix_decay_w1/w2 in quant exclusion list (#9387)
b2e89a327457179a34eae4d7de0d412ed945679c	daa9623ab051a8162ae750b150b9522571b55f21	2024-09-09T09:02:45+02:00	Dan Johansson		Arm AArch64: Documentation updates (#9321)
1b9ae5189cd279c6b45e36d43e4f9ccae628d02f	e32d0816edfd247784f6780b0bb9ae0bceef5e47	2024-09-07T20:43:51+02:00	Xuan Son Nguyen		common : refactor arg parser (#9308)
df270ef74596da8f1178f08991f4c51f18c9ee82	947538acb8617756a092042ff7e58db18dde05ec	2024-09-07T15:16:19+03:00	Georgi Gerganov		llama : refactor sampling v2 (#9294)
134bc38ecf3e2c5460581badce289a1ffa680453	815b1fb20a53e439882171757825bacb1350de04	2024-09-07T00:03:01+03:00	Aarni Koskela		llama-bench : log benchmark progress (#9287)
815b1fb20a53e439882171757825bacb1350de04	409dc4f8bb5185786087f52259ee4626be93f54d	2024-09-06T18:59:58+03:00	Aarni Koskela		batched-bench : add `--output-format jsonl` option (#9293)
9bc6db28d011d47a5f318dc4aebbe7927fac4629	32b2ec88bc44b086f3807c739daf28a1613abde1	2024-09-05T21:48:47-04:00	compilade		ggml-quants : ternary packing for TriLMs and BitNet b1.58 (#8151)
bdf314f38a2c90e18285f7d7067e8d736a14000a	581c305186a0ff93f360346c57e21fe16e967bb7	2024-09-05T02:19:39+02:00	slaren		llama-bench : fix NUL terminators in CPU name (#9313)
8962422b1c6f9b8b15f5aeaea42600bcc2d44177	b69a480af4db8ffd6cbb2efe7ed8132bc7d39073	2024-09-03T20:58:54+03:00	Aarni Koskela		llama-bench : add JSONL (NDJSON) output mode (#9288)
48baa61eccdca9205daf8d620ba28055c2347b64	f1485161e58d562099dd050f8ac3a9ea9f4cd765	2024-09-02T22:08:38+02:00	Xuan Son Nguyen		server : test script : add timeout for all requests (#9282)
8f1d81a0b6f50b9bad72db0b6fcd299ad9ecd48c	a47667cff41f5a198eb791974e0afcc1cddd3229	2024-09-01T22:38:17+08:00	Molly Sophia		llama : support RWKV v6 models (#8980)
ea5d7478b1edcfc83f8ea8f9f0934585cc0b92e3	49271efbaf3f7a4ae52c4ca299b6d4e82598a97d	2024-08-31T13:50:35+05:30	Srihari-mcw		sgemm : improved Q4_0 and Q8_0 performance via 4xN and Mx4 gemm (#8908)
42c76d1358021ccdbe8ba89109c143dd7ae166df	9f7d4bcf5c27d37b0c7da82eeaf9c1499510554b	2024-08-29T19:20:53-04:00	Faisal Zaghloul		Threadpool: take 2 (#8672)
f12ceaca0c7b59def30b7a832a6904df7ed3f4f7	436787f170329b3f549e6c2c46593d2af8482e7c	2024-08-26T11:03:30+02:00	slaren		ggml-ci : try to improve build time (#9160)
436787f170329b3f549e6c2c46593d2af8482e7c	93bc3839f980ff14be86efe408b4cd7e89b26835	2024-08-25T23:09:53-07:00	Justine Tunney		llama : fix time complexity of string replacement (#9163)
2f3c1466ff46a2413b0e363a5005c46538186ee6	50addec9a532a6518146ab837a85504850627316	2024-08-21T04:00:00+09:00	Changyeon Kim		llava: Add ACC OP for GPU acceleration to the Vulkan backend in the LLAVA CLIP model. (#8984)
d5492f0525fa533817a67e93a4bde9d71d81cf58	234b30676a97ce227b604c38beb9dcaca406dea9	2024-08-15T10:11:11+03:00	Georgi Gerganov		ci : disable bench workflow (#9010)
5fd89a70ead34d1a17015ddecad05aaa2490ca46	98a532d474c73d3494a5353024cb6a4fbbabbb35	2024-08-14T18:32:53+02:00	0cc4m		Vulkan Optimizations and Fixes (#8959)
7c5bfd57f83fd3630934cfa70892aa4022d3faf7	6e02327e8b7837358e0406bf90a4632e18e27846	2024-08-11T10:09:09+02:00	Markus Tavenrath		Optimize Vulkan backend for better CPU performance and less GPU synchronization overhead. (#8943)
506122d854c5d05b4a3d45a294f14bd4c02d9868	725e3d94379d5b619c027347308bccf2e0ead89f	2024-08-07T09:01:06+08:00	Zhenwei Jin		llama-bench : add support for getting cpu info on Windows (#8824)
bc0f887e159c0d78c28121e2c8b5c58094170875	b42978e7e4d56eaaa93588414e804d9fbbc3cae2	2024-08-05T21:10:37+08:00	wangshuai09		cann: fix buffer_num and runtime speed slowly error (#8865)
a3738b2fa7c60ef2c4592435d1aa7fb8f1f69c3e	655858ace0cf2720e56eb01f84ad05e0c94ada3c	2024-08-04T17:28:08+02:00	0cc4m		vulkan : implement Stable Diffusion operators (ggml/904)
0d6fb52be0c1b7e77eb855f3adc4952771c8ce4c	978ba3d83d17b10fdf9807006048432b5b3769fc	2024-08-04T14:17:16-04:00	Brandon Squizzato		Install curl in runtime layer (#8693)
ecf6b7f23e664afd7ff856ec39034240ce438daa	01aae2b4975b57a265ce8194928fd87f2d71027e	2024-08-04T03:55:03-07:00	Brian Cunnie		batched-bench : handle empty `-npl` (#8839)
76614f352e94d25659306d9e97321f204e5de0d3	b72c20b85c1029d135022d39e9a20d4807c11893	2024-08-04T01:34:41+09:00	jdomke		ggml : reading the runtime sve config of the cpu (#8709)
ed9d2854c9de4ae1f448334294e61167b04bec2a	398ede5efeb07b9adf9fbda7ea63f630d476a792	2024-07-31T15:51:06-04:00	Clint Herron		Build: Fix potential race condition (#8781)
268c5660062270a2c19a36fc655168aa287aaec2	7e72aa74fd676a093eb9970e761085ec22734c71	2024-07-30T23:35:30+03:00	Someone		nix: cuda: rely on propagatedBuildInputs (#8772)
c887d8b01726b11ea03dbcaa9d44fa74422d0076	75af08c475e285888f66556d0f459c533b7deb95	2024-07-30T14:56:51+08:00	zhentaoyu		[SYCL] Add `TIMESTEP_EMBEDDING` OP (#8707)
a05ca9369716a8319014cd1fc365980d43f8aae9	9f77d899b7b0d56496f679e54b797da6199fed8e	2024-07-25T00:54:08-07:00	Mahesh Madhav		ggml : loop tiling optimizations for scalar path (ggml/898)
b5e95468b1676e1e5c9d80d1eeeb26f542a38f42	92090eca212650727e38b335c1d4accfbcc9b79c	2024-07-27T05:03:45-07:00	Jeffrey Morgan		llama : add support for llama 3.1 rope scaling factors (#8676)
46e47417aa4f18c08738afd4d9a3e838e97ca03f	e7e6487ba06634edf58dfdf9673bad9df41b445a	2024-07-23T10:50:40+02:00	Jeroen Mostert		Allow all RDNA2 archs to use sdot4 intrinsic (#8629)
566daa5a5b38018b2727950bbd280239adb981b6	6f11a83e4e7700fdf353ed4a29599cb662c792f6	2024-07-22T15:44:53+02:00	Jiří Podivín		*.py: Stylistic adjustments for python (#8233)
22f281aa16f44d8f6ec2c180a0685ff27e04e714	328884f4219c0228673cf1870ac63987fb4f9fd0	2024-07-20T22:09:17-04:00	M-A		examples : Rewrite pydantic_models_to_grammar_examples.py (#8493)
7acfd4e8d55082c1b597dfc3ffe04fb5d530c6dc	97bdd26eee11fe109dec00de75690ceef61c03f2	2024-07-15T23:13:10-04:00	compilade		convert_hf : faster lazy safetensors (#8482)
808aba39161e5d7ca2ff24110b5aa14d2e536988	a977c115448e40856fb9cbe3ceb6d8ce802553b0	2024-07-11T16:47:47+02:00	Johannes Gäßler		CUDA: optimize and refactor MMQ (#8416)
dd07a123b79f9bd9e8a4ba0447427b3083e9347a	f4444d992c16b6b9442f4770c7c3a10b19a08343	2024-07-10T12:35:18-04:00	Clint Herron		Name Migration: Build the deprecation-warning 'main' binary every time (#8404)
0f1a39f3439825acf7e3a1663566d410be152170	83321c6958acf20747d288031174cc1bf8816e33	2024-07-10T07:14:51-05:00	Dibakar Gope		ggml : add AArch64 optimized GEMV and GEMM Q4 kernels (#5780)
fd560fe680c72fd0a0af2bc8881add20ad919071	e500d6135ac42c4a23baf6a9a1a934dc023e5c7d	2024-07-09T11:58:44-07:00	Andy Salerno		Update README.md to fix broken link to docs (#8399)
e500d6135ac42c4a23baf6a9a1a934dc023e5c7d	a03e8dd99d3954e7547137936c5a2a3b348bdb7f	2024-07-09T11:54:43-04:00	Clint Herron		Deprecation warning to assist with migration to new binary names (#8283)
470939d483d1c89b7292f78bac1fd27c42c171ce	6f0dbf6ab087bcd286fb78560099ca0458316735	2024-07-08T03:26:53-04:00	Kevin Wang		common : preallocate sampling token data vector (#8363)
3fd62a6b1c9ca7b7c0093e984cc9c133c6f2726d	a8db2a9ce64cd4417f6a312ab61858f17f0f8584	2024-07-07T15:04:39-04:00	compilade		py : type-check all Python scripts with Pyright (#8341)
905942abdba5ba0b28a1b0805e51e4f818c54bc9	b5040086d436e7345e4fa33a5b9558060c75603f	2024-07-07T20:52:10+08:00	toyer		llama : support glm3 and glm4 (#8031)
0a423800ffe4e5da3d83527ef3473da88cd78146	d12f781074b92589a72a36ffabb583933f7b9dc0	2024-07-05T07:06:09Z	Daniele		CUDA: revert part of the RDNA1 optimizations (#8309)
f09b7cb609d80b8031803f89255991dc8b35db69	a38b884c6c4b0c256583acfaaabdf556c62fabea	2024-07-05T10:32:29+08:00	Neo Zhang Jianyu		rm get_work_group_size() by local cache for performance (#8286)
d23287f122c34ebef368742116d53a0ccb2041ee	5f2d4e60e202aabee10051e6615bb821e51787be	2024-07-03T23:02:58Z	Daniele		Define and optimize  RDNA1 (#8085)
cb5fad4c6c2cbef92e9b8b63449e1cb7664e4846	dae57a1ebc1c9bd5693ab999e19d77c5506ae559	2024-07-01T20:39:06+02:00	Johannes Gäßler		CUDA: refactor and optimize IQ MMVQ (#8215)
9a590c82262dd518137f85406e65e452fdf2aca3	52fc8705a0617452df08333e1161838726c322b4	2024-06-24T12:41:23+02:00	Johannes Gäßler		CUDA: optimize MMQ int8 tensor core performance (#8062)
95f57bb5d5b18ef0beb2702a0d6c06e46804075c	e112b610a1a75cb7fa8351e1a933e2e7a755a5ce	2024-06-24T03:07:59+02:00	slaren		ggml : remove ggml_task_type and GGML_PERF (#8017)
a927b0f3dd9a86ee042cd2bdcc8c9da4a855926b	80ea089d771f0c2d97afa8bead80ded412f600d7	2024-06-21T08:51:28+03:00	Georgi Gerganov		llama : optimize long word tokenization with WPM (#8034)
61665277afde2add00c0d387acb94ed5feb95917	b96f9afb0d58b003ac8d1d0c94cd99393a3bc437	2024-06-18T14:00:14+02:00	Ulrich Drepper		Allow compiling with CUDA without CUDA runtime installed (#7989)
6a2f0b3474d479bda4ac2ee7cfd5dcdcf0be1f79	21be9cab94e0b5b53cb6edeeebf8c8c799baad03	2024-06-17T16:10:15+02:00	Markus Tavenrath		Implement non-mapped async IO for CUDA on Windows.  (#7896)
b5fcf8ef5c29df53cfff60e180b4992a3b2332a6	398105ff4373eea385ea8e8625cb417b2ae51134	2024-06-16T16:53:11+08:00	Hong Bo PENG		ggml : fix and optimize ppc64le (ggml/849)
e65bbf606c61f49dc06c7ac060cd5ba7ae446025	6fcd1331efbfbb89c8c96eba2321bb7b4d0c40e4	2024-06-14T16:47:41+03:00	Radoslav Gerganov		llama-bench : fix RPC indication (#7936)
f578b86b2123d0f92afbaa98a031df4d4464e582	1c641e6aac5c18b964e7b32d9dbbb4bf5301d0d7	2024-06-13T03:11:35+02:00	slaren		move BLAS to a separate backend (#6210)
148995e5e57b313cce2672f75610db58c6327a51	4bfe50f741479c1df1c377260c3ff5702586719e	2024-06-11T14:45:40+02:00	Johannes Gäßler		llama-bench: more compact markdown tables (#7879)
5795b941827fdec6c1662986de962badff456718	ed9f2521185706481501a5e6d5315397b11802ff	2024-06-08T22:47:25-04:00	compilade		convert-hf : match model part name prefix and suffix (#7687)
55b2d0849d3ec9e45e4a4d9e480f5fa7977872a6	f5d7b268ec4bf8628aa6ccc9f6631d0230dde76f	2024-06-06T10:07:06+01:00	Olivier Chafik		grammars: x{min,max} repetition operator (#6640)
1442677f92e45a475be7b4d056e3633d1d6f813b	554c247caffed64465f372661f2826640cb10430	2024-06-04T21:23:39+03:00	Georgi Gerganov		common : refactor cli arg parsing (#7675)
adc9ff384121f4d550d28638a646b336d051bf42	987d743d6bc4cee4bde6820733ea33a2abc0afac	2024-06-04T14:32:42+02:00	slaren		llama-bench : allow using a different printer for stderr with -oe (#7722)
2ac95c9d5678d05e253691fb1f26471675bff5ad	750f60c03e4d3f53fa51910551ce87a3d508d2d7	2024-06-01T21:50:18+05:30	HanishKVC		SimpleChat: Simple histogram/repeatMatching driven garbageTrimming, Settings UI, Streaming mode, OpenAi Compat (Model, Authorization Bearer), Save/Restore session, Auto Settings UI (#7548)
c8047d538f3addab40e3112be60bb92e70ce1a50	30e238b246f8002cc6eb7cb79afe242243f1f66d	2024-05-31T16:26:21+02:00	Johannes Gäßler		scripts: update compare_llama_bench.py [no ci] (#7673)
210d99173dc82aafb48f6e39d787c387951fe3a9	87bdf2a199acd62e19814d7a4d0500a04a7f09f3	2024-05-29T14:45:44+03:00	Radoslav Gerganov		llama-bench : add support for the RPC backend (#7435)
b9adcbbf92fc7096bee23fe61496d25652ebf765	9588f196b1d7b21bdff013fcf958c249576b2619	2024-05-26T06:26:34+05:30	HanishKVC		SimpleChat Completion Mode flexibility and cleanup, Settings gMe, Optional sliding window (#7480)
1e374365d170b7f692fd7753c145e21bc14486c8	197ff91462dd05bb9a3be03578114abf0c355536	2024-05-22T23:23:21+05:30	HanishKVC		SimpleChat: a simple and dumb web front end for testing /chat/completions and /completions end points and try chat (#7350)
b18532a4efeca8796fea8e36195c81cbfd596a4a	fcda1128bc5f8eb7e1811708fe9d9867b9aec815	2024-05-22T16:10:46+02:00	slaren		phi3 : duplicate rope factors in each layer (#7447)
65c58207ece92ad213f4bfd0f91dcb2dfb664f5b	1cc0155d04918cb3017afa472acea51b77483c4a	2024-05-20T15:19:21+08:00	junchao-loongson		ggml : add loongarch lsx and lasx support (#6454)
1cc0155d04918cb3017afa472acea51b77483c4a	e932094d58f513d5996c3efc9f6fed8238894c57	2024-05-20T10:16:41+03:00	Georgi Gerganov		server : tuning tests (#7388)
d273c1402b25086fd91aef2467ac13f2e49fa0ea	27b040691cbe45314147c2745e891a38e9c048d4	2024-05-17T15:11:45+03:00	Aarni Koskela		py : convert-hf-to-gguf-update improvements (#7340)
934266c0e0b2aa9781fdba2deb112c161ff038a9	9c4fdcbec8c7fcc428e723b0d8a1cf1f351ba642	2024-05-17T02:58:52-04:00	Justine Tunney		ggml : rewrite silu and softmax for cpu (#7154)
24ecb58168dce81646c2ed425690a106591c8c6d	9afdffe70ebf3166d429b4434783bb0b7f97bdeb	2024-05-16T20:43:45+02:00	Pierrick Hymbert		Revert "server bench: fix bench not waiting for model load (#7284)" (#7334)
13ad16af1231ab2d245d35df3295bcfa23de1305	8f7080bf48828b538bc9387c3d150bbd4fb4cf2d	2024-05-15T19:47:36-07:00	Max Krasnyansky		Add support for properly optimized Windows ARM64 builds with LLVM and MSVC (#7191)
583fd6b000ec9ad1b465b5c98524f4a0ae388077	9f773486ab78d65f5cca3f7e31c862b7043bf721	2024-05-15T08:44:16+02:00	Johannes Gäßler		server bench: fix bench not waiting for model load (#7284)
0d26d8ccd8caebab75af697c0275f599075fdacf	4f0263633b40e94e8b69fd6e7e4395cfedfd5c12	2024-05-12T17:17:18+08:00	Hong Bo PENG		ggml : optimize for ppc64le using VSX intrinsics (ggml/784)
cbf75894d256f1861f6409565db599365de3d4b8	0d5cef78aeafae4d4e6d56e2d4bcda771af58cc9	2024-05-13T08:04:29+08:00	Neo Zhang		[SYCL] Add oneapi runtime dll files to win release package (#7241)
e849648888a11de13aaaa4cb2eda3f5a9c7b444d	18e437665ce626dddbd79119aa7498493e7cb13b	2024-05-10T18:03:54+02:00	slaren		llama-bench : add pp+tg test type (#7199)
fd9f92b154850014146f61717cd292a59a5cee5a	22842164bcae3251b81ad9e497a16ef66833cb9e	2024-05-09T13:03:29+02:00	Daniel Bevenius		llama : update llama_timings.n_p_eval setting (#7160)
bc4bba364fb96d908f2698e908648df5e6f55e02	c12452c7aec8a02264afc00196a13caa591a13ac	2024-05-08T21:55:49+01:00	agray3		Introduction of CUDA Graphs to LLama.cpp (#6766)
acdce3cdef6fc2f0b7b5623231fd7762c0884d1c	3855416027cb25d9a708ffa5581cf503a87856a6	2024-05-08T18:54:39+10:00	Brian		compare-llama-bench.py: add missing basicConfig (#7138)
3855416027cb25d9a708ffa5581cf503a87856a6	c0e6fbf8c380718102bd25fcb8d2e55f8f9480d1	2024-05-08T02:30:09-04:00	Justine Tunney		ggml : introduce bfloat16 support (#6412)
858f6b73f6e57a62523d16a955d565254be889b4	b3a995b416e13ae3123a117a743e11d0ede0ca4c	2024-05-06T11:12:14-07:00	William Tambellini		Add an option to build without CUDA VMM (#7067)
628b299106d1e9476fdecb3cbe546bf5c60f1b89	8f8acc8683a00ce40fa5a81161a079d2167126e6	2024-05-05T07:17:47-05:00	kunnis		Adding support for the --numa argument for llama-bench. (#7080)
8f8acc8683a00ce40fa5a81161a079d2167126e6	ca3632602091e959ed2ad4c09c67a7c790b10d31	2024-05-05T13:38:55+02:00	Sigbjørn Skjæret		Disable benchmark on forked repo (#7034)
a2ac89d6efb41b535778bfeaecaae8fe295b6ed3	433def286e98751bf17db75dce53847d075c0be5	2024-05-04T05:36:41+10:00	Brian		convert.py : add python logging instead of print() (#6511)
9c67c2773d4b706cf71d70ecf4aa180b62501960	952d03dbead16e4dbdd1d3458486340673cc2465	2024-04-30T12:16:08+03:00	Georgi Gerganov		ggml : add Flash Attention (#5021)
b8a7a5a90fd3187175d84227dad705ade395ba46	d2c898f746a527f09effb061829e68b2e1812a28	2024-04-29T17:02:45+01:00	Olivier Chafik		build(cmake): simplify instructions (`cmake -B build && cmake --build build ...`) (#6964)
5790c8dac1a4f0aa80b4efee3b962d8c04c829e8	46e12c4692a37bdd31a0432fc5153d7d22bc7f72	2024-04-26T09:26:16+02:00	Pierrick Hymbert		bench: server add stop word for PHI-2 (#6916)
5cf5e7d490dfdd2e70bface2d35dfd14aa44b4fb	40f74e4d739e9250431cf339ae7588b28d8d0663	2024-04-21T18:48:53+01:00	Olivier Chafik		`build`: generate hex dump of server assets during build (#6661)
9958c81b798a5872087b30b360e4674871f2479e	8b1b1f4982d3e9b994308d05a1c8b9e45c23edb5	2024-04-19T09:35:54Z	nopperl		Implement the OLMo architecture (#6741)
8cc91dc63c0df397d644a581b2cbeea74eb51ae0	dbceec87c0221ec952e69448df6a71f1372a7487	2024-04-16T14:55:30-04:00	Justine Tunney		ggml : add llamafile sgemm (#6414)
de17e3f7455dc7fd298cc61d86798533b9ca7a29	b5e7285baffb0da8a6619567b52d8e67de41291d	2024-04-14T10:42:29+08:00	Neo Zhang Jianyu		fix memcpy() crash, add missed cmd in guide, fix softmax (#6622)
ab9a3240a9da941fdef5cd4a25f2b97c2f5a67aa	fbbc030ba93561fac842af994c5c6c4c1147f13b	2024-04-12T19:43:38+01:00	Olivier Chafik		JSON schema conversion: ⚡️ faster repetitions, min/maxLength for strings, cap number length (#6555)
04a5ac211ef40936295980b7cdf0ba6e97093146	f7001ccc5aa359fcf41bba19d1c99c3d25c9bcc7	2024-04-11T21:44:50-04:00	Clint Herron		Optimization: eliminate addition of redundant stacks when advancing grammar. (#6616)
57dd02c44b2a0eb79e28f6c5eb8242a5d2d3174d	75cd4c77292034ecec587ecb401366f57338f7c0	2024-04-06T10:31:33-04:00	Clint Herron		Tests: Added integration tests for GBNF parser  (#6472)
75cd4c77292034ecec587ecb401366f57338f7c0	a8bd14d55717754a1f48313a846a2b16fa998ad2	2024-04-06T05:40:47+02:00	Pierrick Hymbert		ci: bench: support sse and fix prompt processing time / server: add tokens usage in stream OAI response (#6495)
87e21bbacd830437ab653cf03b6f26d45c15395d	1b496a745c315022df2d919374052e6004ced8d3	2024-04-06T01:34:53+07:00	Ting Sun		bench : make n_batch and n_ubatch configurable in Batched bench (#6500)
7dda1b727ef1730783a6077136d28b83e70dd397	c666ba26c39d5c07e07b4e1e411332f408e309ad	2024-04-05T01:30:53+09:00	Minsoo Cheong		ci: exempt master branch workflows from getting cancelled (#6486)
8120efee1d9931b514aeb5a047209d576f23286c	a74401f0e5ebb15fa4d8b6619d1baa6ea9179123	2024-04-04T16:59:04+02:00	Pierrick Hymbert		ci: bench fix concurrency for workflow trigger dispatch with sha1 (#6478)
7a2c92637ae265654a68f62e6a7610b358255d3f	4bcd6b959ca3991084ad1d8464caf2a734e29b1d	2024-04-04T11:57:58+02:00	Pierrick Hymbert		ci: bench: add more ftype, fix triggers and bot comment (#6466)
5fb1574c8112c757fc202fdae279da883f34e610	60cdf40cc32f0ad4cb11e0ca8fd38f3b93d8d640	2024-04-03T13:22:57-07:00	Fattire		A few small fixes to server's README docs (#6428)
33a52448061cfd2ea44da9e6cb30b2ec22e2f6d0	226e819371eec0f298d9075198394a07b23ecfa9	2024-04-01T13:30:43+02:00	Johannes Gäßler		compare-llama-bench.py: fix long hexsha args (#6424)
37e7854c104301c5b5323ccc40e07699f3a62c3e	c342d070c64a1ffe35d22c1b16b672e684a30297	2024-03-30T11:36:07+01:00	Pierrick Hymbert		ci: bench: fix Resource not accessible by integration on PR event (#6393)
057400a3fd457f4f214684eeb171444663b47a23	b75c38166cf0c66793a32d5c3d6cb69929dd083d	2024-03-29T08:23:22+01:00	Daniel Bevenius		llama : remove redundant reshape in build_kv_store (#6369)
28cb9a09c4d10a489be1238abe7a858dcd4d65f2	cfc4d75df6399b36153ef739f2c1abee4c114bb8	2024-03-28T11:27:56+01:00	Pierrick Hymbert		ci: bench: fix master not schedule, fix commit status failed on external repo (#6365)
a016026a3ac16d8c9b993a3573f19b9556d67de4	53c7ec53d5eca26b2c0c648605543a5fa6c12817	2024-03-27T20:26:49+01:00	Pierrick Hymbert		server: continuous performance monitoring and PR comment (#6283)
557410b8f06380560155ac7fcb8316d71ddc9837	55c1b2a3bbd470e9e2a3a0618b92cf64a885f806	2024-03-26T10:46:41-04:00	compilade		llama : greatly reduce output buffer memory usage (#6122)
55c1b2a3bbd470e9e2a3a0618b92cf64a885f806	e097633f63fdd26d492844f7eff056e4083fd9eb	2024-03-26T15:21:27+01:00	Kawrakow		IQ1_M: 1.75 bpw quantization (#6302)
64e7b47c6986221f2ff5c57c89dfc018bb0e9e6d	7733f0c76081b2a69b5f8b192db2db7c43629d58	2024-03-25T16:38:22+09:00	Minsoo Cheong		examples : add "retrieval" (#6193)
cfd3be76e37dab92c846d75a2421178f20db4a11	5b7b0ac8dfdd800c0fd0dc69b69991e8cb19fb46	2024-03-21T13:59:38+01:00	Kawrakow		ggml : same IQ4_NL quantization for CPU/CUDA/Metal (#6196)
5b7b0ac8dfdd800c0fd0dc69b69991e8cb19fb46	1943c0198125a0da1a200390e82cf461f9080d99	2024-03-21T11:50:43Z	Olivier Chafik		json-schema-to-grammar improvements (+ added to server) (#5978)
76aa30a26353f597e4fbe3cf776772ae812af89a	c5b8595e3f4f4ed319ef71c9c9d868d1b7a27626	2024-03-21T08:27:57+01:00	Kawrakow		Add ability to use Q5_0, Q5_1, and IQ4_NL for quantized K cache (#6183)
bd60d82d0cc8b6852ec535495a5042dbdf05de24	6c0b287748327741b113d7d6018b68c63039b1c5	2024-03-20T01:33:49-04:00	Jared Van Bortel		server tests : more pythonic process management; fix bare `except:` (#6146)
a56d09a4407f29c21e149b44fd5308f83aa1cb09	d84c48505f60bcd358b82a751d40418c4d235643	2024-03-16T13:20:53+01:00	Pierrick Hymbert		ci : close inactive issue with workflow (#6053)
b0bc9f4a9da7c19f4779106ea83b23feca747566	4755afd1cbd40d93c017e5b98c39796f52345314	2024-03-15T09:22:24+01:00	slaren		llama-bench : use random tokens to improve accuracy with mixtral (#6069)
6e0438da3cc95b89cdbf55f45fa4e324d9076792	727107707a73b3dc8a497cf9fc9405722c16dd2b	2024-03-14T14:29:32-04:00	Steve Grubb		gguf : fix resource leaks (#6061)
43241adf22e8231ffaf3827d2c9310cc0ffd5ac5	a44bc969e4cd62ca9f4332e17fe3c51f2093e7c6	2024-03-14T12:15:39+01:00	Pierrick Hymbert		server: disable debug release type sanitizer, simplify trigger (#6047)
2c4fb69246834503db7b78bcbedcef506bbc60c4	3ca23481dd309bd51cc31c73a4cc34f922cc372f	2024-03-14T11:56:48+01:00	Michael Podvitskiy		llama : optimize defrag moves + fix fragmentation calculation (#6037)
f30ea47a87ed4446ad55adb265755dc9102956a2	d8fd0ccf6ac8b07791ffd1575eed436930854ae3	2024-03-13T18:54:21+01:00	slaren		llama : add pipeline parallelism support (#6017)
44ca159faf4fbe1a7ace13a962845ba7cdfd95ec	05b06210c954491cf0f12034b0a62bd4d69ce78b	2024-03-11T16:53:15+01:00	Kawrakow		1.5 bit: we can do even better (#5999)
be858f620508385ad12d0e5e862010e666ca729c	ef3ced26a3817d92890b97b83acaeb018ade02d0	2024-03-11T07:51:49+01:00	Kawrakow		Better 1.5 bit quantization  (#5971)
621e86b331f8b0e71f79fd82a4ae1cd54c3e4396	77d1ac7e00bf049b9f2bba1b5a310a78318c49c4	2024-03-09T23:41:49+01:00	Pierrick Hymbert		server: benchmark: chat/completions scenario and other llm servers comparison (#5941)
c2101a2e909ac7c08976d414e64e96c90ee5fa9e	515f7d0d4fce41c752fc253acf30707c3be2531e	2024-03-08T17:31:00-05:00	compilade		llama : support Mamba Selective State Space Models (#5328)
76e868821a94072fbc87cb1fcca291694319eae8	e457fb3540e0aaec47cfde0abf784c213f9216ee	2024-03-08T12:25:04+01:00	Pierrick Hymbert		server: metrics: add llamacpp:prompt_seconds_total and llamacpp:tokens_predicted_seconds_total, reset bucket only on /metrics. Fix values cast to int. Add Process-Start-Time-Unix header. (#5937)
6cdabe652695167263c8b447520987b11856f7ca	89fb735fcfd21781a8194b211cf32824beb3f71f	2024-03-07T16:32:38+02:00	Georgi Gerganov		llama-bench : add embeddings option (#5924)
2002bc96bf2cbf5ab981a17d7e994d817c9801f5	ceca1aef0738b57951cd12c603c3477e75312dec	2024-03-07T11:41:53+02:00	Georgi Gerganov		server : refactor (#5882)
652ca2bded3c818320d92c70d2b67f64bdbff5e5	bd836944f826f07e19b7edcf994a78728da49c1c	2024-03-05T22:27:29+01:00	slaren		compare-llama-bench.py : remove mul_mat_q (#5892)
61d1c88e155515dd03940913a5707ea84a8b119b	21b08674331e1ea1b599f17c5ca91f0ed173be31	2024-03-05T13:33:42+01:00	0cc4m		Vulkan Improvements (#5835)
21b08674331e1ea1b599f17c5ca91f0ed173be31	6a87ac3a52668e117d97bcea07b529c93188b303	2024-03-05T16:08:35+08:00	Neo Zhang Jianyu		[SYCL] fix mul_mat fault in CI/unit-test (#5862)
7d43c585dc174bb586775c22c15e5db9242b5b4b	82f3e668adafba647de703f835991e91a96b5ac4	2024-03-03T20:23:52+08:00	leejet		add some new ops, fix some operators and add batch operations to certain operators. (ggml/747)
9731134296af3a6839cd682e51d9c2109a871de5	4a6e2d6142ab815c964924896891e9ab3e050632	2024-03-02T22:00:14+01:00	Pierrick Hymbert		server: tests: passkey challenge /  self-extend with context shift demo (#5832)
3ab8b3a92ede46df88bc5a2dfca3777de4a2b2b6	9600d59e010c18f5872580a21734ea1bf1968d04	2024-03-01T12:39:06+01:00	Pierrick Hymbert		llama : cleanup unused mmq flags (#5772)
9600d59e010c18f5872580a21734ea1bf1968d04	5cb02b4a012bb16c6c699c0c62c05ffa653eee0f	2024-03-01T03:15:36-06:00	Douglas Hanley		unicode : switch to multimap based nfd_map (#5799)
87c91c07663b707e831c59ec373b5e665ff9d64a	317709b2a81dbaf87850202686ec5bb2602a504e	2024-02-28T21:44:21+02:00	Georgi Gerganov		ci : reduce 3b ppl chunks to 1 to avoid timeout (#5771)
0becb22ac05b6542bd9d5f2235691aa1d3d4d307	c24a2a6e6005e5d424301525a42ba45a4a362d30	2024-02-27T16:34:24+02:00	Kawrakow		IQ4_XS: a 4.25 bpw quantization (#5747)
e3965cf35aac00d4e24998c8a3d0093ae1d98bd3	8b350356b28f782deab63d8b0e9ae103ceb25fcd	2024-02-25T22:48:33+01:00	Pierrick Hymbert		server: tests - slow inference causes timeout on the CI (#5715)
4c4cb30736582cacb1a164a9d4bc8e17b1014be7	525213d2f5da1eaf4b922b6b792cb52b2c613368	2024-02-24T16:23:52+02:00	Kawrakow		IQ3_S: a much better alternative to Q3_K (#5676)
70847553963c85e86051d06df848236829f5f951	4480542b2271ba1438f0daff8e5f3a74b1dc8609	2024-02-19T09:31:59+01:00	Daniel Bevenius		llava : avoid changing the original BakLLaVA model (#5577)
fc0c8d286a533363a9a663510b62af85ffad58b3	bd2d4e393b2b7d2a1b2e201058e26017c9728ead	2024-02-18T17:19:23+01:00	Daniel Bevenius		llava : update surgery script to not remove tensors (#5536)
8f1be0d42f23016cb6819dbae01126699c4bd9bc	6e4e973b2615f8d390b1c4f4a7e05a119078bb0f	2024-02-17T23:04:16+02:00	Georgi Gerganov		ggml : add ALiBi support for ggml_soft_max_ext (#5488)
d2819d5577b35507be83d0c3f4d2d3c0ab1488ca	4cb072769804c77ab466bc8351c76ede9d5ba49d	2024-02-16T15:14:40+02:00	Georgi Gerganov		scripts : add helpers script for bench comparing commits (#5521)
9060a1e9dfca6038906e819be5fa42217f49028c	9350a1cf21b1492c69b20175b73a419b897d6a3a	2024-02-15T16:49:01+01:00	slaren		cuda : print message when initialization fails (#5512)
0d4177126b0556e202efb85bf3f768be81076400	7930a8a6e89a04c77c51e3ae5dc1cd8e845b6b8f	2024-02-15T09:01:57+01:00	Elbios		llava : fix memory management bug (#5491)
704359e29985a06a389337a2617b7f3fa8eff908	594fca3fefe27b8e95cfb1656eb0e160ad15a793	2024-02-15T17:11:15+11:00	Neuman Vong		vulkan: Find optimal memory type but with fallback (#5381)
49cc1f7d67de2da99f3ac185f9ff1319b7bf35f8	99b8b43d7b185a6483f28cf798a2d968b2e16ca7	2024-02-13T13:01:29+02:00	Georgi Gerganov		bert : add tests + fix quantization (#5475)
85910c5b30f6e268321be8df044f5528a6efac52	139b62a839825ef20084ed75ed624db7a5ad554a	2024-02-11T15:35:50+02:00	Georgi Gerganov		main : ctrl+C print timing in non-interactive mode (#3873)
a07d0fee1f05c5c1dc49948ae1a3293db017275f	e4640d8fdf56f14a6db3d092bcd3d2d315cb5d04	2024-02-11T07:22:33-06:00	snadampal		ggml : add mmla kernels for quantized GEMM (#4966)
e4640d8fdf56f14a6db3d092bcd3d2d315cb5d04	907e08c1109f498b01036367804cff3082c44524	2024-02-11T12:44:51+01:00	Johannes Gäßler		lookup: add print for drafting performance (#5450)
4b7b38bef5addbd31f453871d79647fbae6bec8a	e00d2a62dd1441e3b089570ec06d05c18800d368	2024-02-10T05:30:19+11:00	Neuman Vong		vulkan: Set limit for task concurrency (#5427)
6fdfa2ecc684000a25a4ad91823bc82a6652b645	a2d60c9158435ae9a6f14632f07f1acf7a3becef	2024-02-05T10:46:06+02:00	Kawrakow		iq2_xxs: tune quantization (#5320)
3cc5ed353c07201d8d5b98b0a4713ab633da6d04	60ecf099eddfe70fec797ef6790572e452054add	2024-02-03T20:14:59+01:00	Johannes Gäßler		make: fix nvcc optimization flags for host code (#5309)
e920ed393d989ed35625ddaf182ebb52cda07fcd	52bb63c7082c859c3f1dfc527227e6a95b299c7c	2024-02-03T18:15:00+01:00	0cc4m		Vulkan Intel Fixes, Optimizations and Debugging Flags (#5301)
af3ba5d94627d337e32a95129e31a3064c459f6b	e1e721094d8169636d55f68efe37f222cd3f0677	2024-02-02T15:53:27+08:00	Neo Zhang Jianyu		[SYCL] update guide of SYCL backend (#5254)
128dcbd3c9c4b12f42b560a4430427d7b2828628	4d0924a8902010d31bd737b6f1f594943d120d0f	2024-02-02T03:48:53+08:00	Neo Zhang Jianyu		add --no-mmap in llama-bench (#5257)
e8dc55d0065d076d4c20f3c4bfca562701b4edfe	e0085fdf7c758f0bc2746fc106fb29dd9df959de	2024-01-30T19:04:37-05:00	Jared Van Bortel		kompute : llama-bench support and ggml_cpu_has_kompute() (#5226)
8e14e3ddb3744566aef7bc0fa734180e47ae6bdf	f4d7e5497485ce6ce0e322533930b7da4657dd2d	2024-01-30T15:15:07+02:00	Kawrakow		Faster AVX2 dot product for IQ2_XS (#5187)
f4d7e5497485ce6ce0e322533930b7da4657dd2d	2256f36b79a932a478d4dcdf02c1e5a60056e5f3	2024-01-30T15:14:12+02:00	Kawrakow		SOTA 3-bit quants  (#5196)
2307523d322af762ae06648b29ec5a9eb1c73032	0f648573dde61c510560f68244f70ece7e60d8c1	2024-01-28T18:03:59+01:00	0cc4m		ggml : add Vulkan backend (#2059)
0f648573dde61c510560f68244f70ece7e60d8c1	b764b8f1d079ba44d912801ce6d29bd0d94d51cf	2024-01-28T21:26:23+05:30	Abhilash Majumder		ggml : add unified SYCL backend for Intel GPUs (#2690)
7032f4f6349c17a8352f9f93f7d2122f45469e59	5f1925a8cef81eb9b372faaae34b0dd76d5361d4	2024-01-26T11:17:59-06:00	snadampal		ggml : update softmax n_task calculation (#5126)
5eaf9964fc797d4585c214db32a463d557f3ed33	d292f4f2047963f558dd516f1baaa71793e9acf2	2024-01-26T05:06:22+09:00	l3utterfly		llama : dynamic temperature sampling (#4972)
3ce7e8f8e7ccfce07e5947ac5f1f3f4628cf68ea	b2d80e105a59b54822edf7ce7f3ed5f317e96e21	2024-01-22T21:09:35+08:00	XiaotaoChen		llava : MobileVLM support (#4954)
726c0fa9a2da976e9c5d5c51e185d9dd453fc9e5	942c0107a7301434c0a5e7da46bc4cf2393aa556	2024-01-21T08:01:20+02:00	Kawrakow		Slightly faster imatrix (#5050)
77bc1bbd05f0c31cb45773eb5eb59b9ff2b07e1b	48e2b1337257bb77573bf76cfffcff3a5efa8704	2024-01-20T08:11:31Z	Herman Semenov		cmake : add support for ccache (#5002)
381ee195721d8e747ee31a60c0751822b3072f02	a5cacb22b2114fd9adf61c00cbb237384d86bced	2024-01-19T13:20:50-05:00	Uzo Nweke		finetune : fix ggml_allocr lifetimes (tmp workaround) (#5033)
7051aacfac0057fa5fac9ea46c55bffc3892d810	2b3b999cacc7ad1207c32fbdf3479a19c06e1a34	2024-01-19T11:39:11+02:00	Kawrakow		winogrande: evaluate log-probs in parallel (#5036)
3e945cc1e9c06d2001031360e4e303e9548fb02c	ad19812cda4062c9f154ef16315df41fbe6a770a	2024-01-18T19:18:21+02:00	Kawrakow		HellaSwag: speed up by parallelizing log-prob evaluation (#5020)
ba69bbc84ced580fe4fdb0713ca2d95634325b7a	44a1a4a41a4c0b03afaa7d9e06bcbc7cf95aa1e6	2024-01-17T18:46:30+02:00	Georgi Gerganov		imatrix : offload to GPU support (#4957)
44a1a4a41a4c0b03afaa7d9e06bcbc7cf95aa1e6	c918fe8dca8fa1c4602427e0a4b88e20046f6c34	2024-01-17T18:39:41+02:00	Georgi Gerganov		backend : add eval callback (#4935)
158f8c9e21302114bac3c646f80ea85b52ffa0bd	862f5e41ab1fdf12d6f59455aad3f5dd8258f805	2024-01-16T17:05:19Z	Paul Tsochantaris		metal : localized logic in `ggml_metal_graph_compute` (#4924)
862f5e41ab1fdf12d6f59455aad3f5dd8258f805	3a48d558a69c88ac17efcaa5900cd9eb19596ac4	2024-01-17T00:47:34+11:00	Neuman Vong		android : introduce starter project example (#4926)
a0b3ac8c48b66206b9c5921ce57bd5c0ea6557c3	d75c232e1da56f19ac4d2530dadbe0ab3a11fde5	2024-01-16T03:16:33-08:00	Justine Tunney		ggml : introduce GGML_CALL function annotation (#4850)
7dc78764e2ff86512e6e31cb0fcb8087df4b4708	356327feb3f66980ab687040495d722696d98970	2024-01-13T15:52:53+01:00	Johannes Gäßler		compare-llama-bench: tweak output format (#4910)
e7e4df031b9e29d4b55a4e0b0295187f6b213db1	584d674be622fbf1578694ada6e62eebedbfd377	2024-01-12T20:07:38+01:00	slaren		llama : ggml-backend integration (#4766)
e739de790921e6abbc8c70398303cacd74913f61	c910e3c28a1caee8cb1398143d582dd9ab697e68	2024-01-10T21:13:42+08:00	leejet		ggml : change GGML_MAX_NAME at compile time (ggml/682)
d34633d8db6c2e400355de4862cd699154ecc73f	4f56458d34cb13dcbf69aca650e9bf77d5497e6f	2024-01-10T14:37:09+01:00	John		clip : support more quantization types (#4846)
4f56458d34cb13dcbf69aca650e9bf77d5497e6f	6efb8eb30e7025b168f3fda3ff83b9b386428ad6	2024-01-10T01:04:33+01:00	Johannes Gäßler		Python script to compare commits with llama-bench (#4844)
6efb8eb30e7025b168f3fda3ff83b9b386428ad6	36e5a08b203542dca53cca4eaf172c5dc4bbc991	2024-01-09T13:46:46-05:00	Austin		convert.py : fix vanilla LLaMA model conversion (#4818)
dd5ae06405c5565b99889bdb3f168f4351252cfb	668b31fc7d86245435ad6574e0e1126e734049e2	2024-01-08T16:02:32+01:00	Kawrakow		SOTA 2-bit quants (#4773)
226460cc0d5b185bc6685fb76f418fd9418d7add	d5a410e8556191672465f7ff58682ea2474038b0	2024-01-07T17:59:01+01:00	slaren		llama-bench : add no-kv-offload parameter (#4812)
f3f62f0d835d559e80714bbeb05d03125574e3dd	0ef3ca2ac62016c0c545de1c89dc2e3e130f4a99	2024-01-02T21:07:47+02:00	Georgi Gerganov		metal : optimize ggml_mul_mat_id (faster Mixtral PP) (#4725)
26f3071d714f0b27ad7f021a46a66a1085480258	775ac8712a7b42cfead2585f42cec0dfd56644ab	2024-01-02T16:23:38+07:00	Nam D. Tran		py : re-enable mmap in convert hf (#4732)
68eccbdc5b56f2a2450f9a8463f9934388cafabf	97bbca6e8522d18041fcde6c3d0907a52ce36446	2023-12-29T06:42:26-08:00	Philip Taron		flake.nix : rewrite (#4605)
db49ff8ed7f0bb201176703441cc02911b08ef2a	60f55e888c29cbd87c4238dd19e85d0eef87245d	2023-12-29T06:24:12-08:00	Justine Tunney		server : replace sleep with condition variables (#4673)
afd997ab6011dfefe9e917425b04ef4d83614841	c8255f8a6b2a3b3ebc6cb340cc2487f39fc95ffc	2023-12-29T05:58:56-08:00	Peter Sugihara		llama.swiftui : fix infinite loop, ouput timings, buff UI (#4674)
65e5f6dadbba4b496bba27f573e473c66b446496	ea5497df5d138c83b2b0ca70aefdc4b1175c1001	2023-12-28T11:20:00-08:00	Justine Tunney		Fix OpenAI server sampling w.r.t. temp and seed (#4668)
f6793491b5af6da75edad34d6f503ef86d31b09f	879b690a9e1eb1ab0a29b58236fc76978fb4d902	2023-12-27T22:39:45+07:00	Nam D. Tran		llama : add AWQ for llama, llama2, mpt, and mistral models (#4593)
48b7ff193e64c97ab174280ba0eb8d14b47c49ba	48b24b170e3b4f9dc28200306840cb07d1c123df	2023-12-22T12:12:53+01:00	slaren		llama : fix platforms without mmap (#4578)
d232aca5a73b290e218a2e48b91023d5e994203f	31f27758faf4a4bd08101a57c7ec3a473f771f86	2023-12-21T21:07:46+01:00	slaren		llama : initial ggml-backend integration (#4520)
800a489e4a8be199122259a995b1ee9dd7fae320	f7f468a97dceec2f8fe8b1ed7a2091083446ebc7	2023-12-17T19:38:41+02:00	Georgi Gerganov		llama.swiftui : add bench functionality (#4483)
799a1cb13b0b1b560ab0ceff485caed68faa8f1f	fecac45658a99eddc4d6e36ba0310ca8f87a77f0	2023-12-13T13:04:25+01:00	slaren		llama : add Mixtral support (#4406)
bcc0eb4591bec5ec02fad3f2bdcb1b265052ea56	81bc9214a389362010f7a57f4cbc30e5f83a2d28	2023-12-07T13:03:17+02:00	Georgi Gerganov		llama : per-layer KV cache + quantum K cache (#4309)
81bc9214a389362010f7a57f4cbc30e5f83a2d28	05cd6e5036d72d0930de4d8f6be7bce09e8dda24	2023-12-07T02:25:22-08:00	Hongyu Ouyang		train : fix #4227 (double free in examples/train-text-from-scratch/train-text-from-scratch.cpp) (#4351)
5aa365d88fdb8fdd430ef3fc141c7a5fd37c3502	52c8bc3cf312e1caf02d37bfb9d9d865cbe33594	2023-12-05T10:19:18-07:00	Kerfuffle		llama : allow overriding GGUF metadata when loading model (#4092)
ef47ec18da469423c276b683dd9b5741cee7023e	1d144112c0fbbb4ecc07dbcf4f05a380148bd6de	2023-12-01T10:51:24+02:00	Georgi Gerganov		ggml : add ggml_soft_max_ext (#4256)
8e672efe632bb6a7333964a255c4b96f018b9a65	0b871f1a04ef60e114bbe43004fd9c21114e802d	2023-11-21T16:22:30+01:00	Galunid		stablelm : simplify + speedup generation (#4153)
f23c0359a32871947169a044eb1dc4dbffd0f405	40a34fe8d034bd484efd79ccbb95059ca6308dcb	2023-11-20T11:35:47+01:00	Galunid		ci : add flake8 to github actions (python linting) (#4129)
f7d5e975424ff0eea55ca5a9181ac8e15553c1fc	ba4cf5c0bf37a729d29e899dadf14541cddd23d4	2023-11-17T16:20:53+01:00	Jiří Podivín		py : remove superfluous import statements (#4076)
3e916a07ac093045d88ef0c4fa78647ae0efc010	947f64f1630bb8b0b363a3bb5e29e11425312d57	2023-11-17T14:48:19Z	gwjr		finetune : speed-up ggml_compute_forward_out_prod_f32 via BLAS (#4079)
91f6499393d2d999331fbfdba47a7f8b9f913f0d	8da46278e1a57107591653275f8e03a281de94f0	2023-11-16T19:14:37-07:00	Kerfuffle		Respect tokenizer.ggml.add_bos_token value when tokenizing (#4040)
ca190bca8e844d171020d6147687e71472d71734	71e3718abdb2771b50c9606d3a7569623a0b0afe	2023-11-01T09:28:28Z	Adrian Hesketh		server : re-enable completion and embedded at the same time (#3876)
71e3718abdb2771b50c9606d3a7569623a0b0afe	238657db2364cfb728c694470a4a81702afea760	2023-11-01T08:04:02+02:00	Georgi Gerganov		llama : refactor graph build code (#3837)
2f9ec7e271220a78fe27c9e6ccbcc0dda31cda0f	34b2a5e1ee4fe6295fb4420eb91131d743694c65	2023-10-27T17:01:23+03:00	Georgi Gerganov		cuda : improve text-generation and batched decoding performance (#3776)
6961c4bd0b5176e10ab03b35394f1e9eab761792	cc448774866e6479c750bd7c135cd8f92cedee67	2023-10-25T10:26:27+03:00	Georgi Gerganov		batched-bench : print params at start
2b4ea35e56792064598e922e46d081e02bc96b94	daab3d7f45832e10773c99f3484b0d5b14d86c0c	2023-10-24T16:48:37+03:00	Georgi Gerganov		cuda : add batched cuBLAS GEMM for faster attention (#3749)
438c2ca83045a00ef244093d27e9ed41a8cb4ea9	9e70cc03229df19ca2d28ce23cc817198f897278	2023-10-22T22:53:08+03:00	Georgi Gerganov		server : parallel decoding and multimodal (#3677)
a5e7dbd6141128bfa3c40a19c2945a181df625d3	d3956aea53369455008159cc405ed4c496976692	2023-10-22T12:14:56-06:00	Kerfuffle		llama : validate special token ids are in range when loading GGUF model (#3635)
f439e506e8ae8b01df2ae2156380f8156d7553e3	e78f3ef24af4ca74e77e725644b41ae8ca3b10a5	2023-10-20T10:02:12Z	Herman Semenov		ggml : fix rope + llama minor optimizations (#3560)
370359e5baf619f3a8d461023143d1494b1e8fde	9e24cc6e2e589d405bd1720c400f5b0b9d0ca3ee	2023-10-12T18:23:18+03:00	M. Yusuf Sarıgöz		examples: support LLaVA v1.5 (multimodal model) (#3436)
8c70a5ff25964f0a81e20d142a2f5ac5baff22fc	24ba3d829e31a6eda3fa1723f692608c2fa3adda	2023-10-11T21:25:33+03:00	Georgi Gerganov		batched : add bench tool (#3545)
b0ec5218c3d24755786b80ecce9cf4ffc07583f8	63d3b06a4318329f92b078e8aa0be7ab6e9f871f	2023-10-08T10:01:53+03:00	Georgi Gerganov		metal : support MTLGPUFamily < Apple7, formatting, style (#3524)
f93af02488179b9c52d0d391b08ae4c4d891b8d3	f72f8f22c9cb60465b2e79df2767e4ba9604e576	2023-10-04T15:29:58+03:00	Georgi Gerganov		sync : ggml (conv 1d + 2d updates, UB fixes) (#3468)
79f34abddb72ac5ddbf118f3d87520b611a10a7d	8186242b6d67cf87ae179fb1a62f52fdf0e5c5eb	2023-10-03T23:38:19+05:00	Tameem		ggml : add RISC-V Vector Support for K-Quants and improved the existing intrinsics (#3453)
f5ef5cfb18148131fcf45bdd2331f0db5ab7c3d0	40e07a60f9ce06e79f3ccd4c903eba300fb31b5e	2023-09-30T18:12:57+02:00	slaren		ggml-cuda : perform cublas mat mul of quantized types as f16 (#3412)
16bc66d9479edd5ee12ec734973554d4493c5dfa	0512d66670de3f650c579519833c085014b0f200	2023-09-28T21:42:38+02:00	slaren		llama.cpp : split llama_context_params into model and context params (#3301)
0e76a8992c8200237bbc6471a53fb8796b3872f7	2db94d98eda56982d80238840b0652b4137a2a84	2023-09-28T20:40:11+02:00	xaedes		train : finetune LORA (#2632)
ec893798b7a2a803466cc8f063051499ec3d96f7	45855b3f1c7bdd0320aa632334d0b3e8965c26c4	2023-09-28T19:04:36+03:00	Georgi Gerganov		llama : custom attention mask + parallel decoding + no context swaps (#3228)
da0400344be12074e67dcabc565140289cf7efaa	e519621010cac02c6fec0f8f3b16cda0591042c0	2023-09-28T12:08:28+02:00	slaren		ggml-cuda : perform cublas fp16 matrix multiplication as fp16 (#3370)
c091cdfb24621710c617ea85c92fcd347d0bf340	51a7cf5c6e490b2f51c82daa76c4ca4f8d845826	2023-09-23T21:48:24+02:00	slaren		llama-bench : add README (#3317)
65c2c1c5ab7c5089dbc6d10bc49b9c58f0164317	80834daecf4b9021770361a6d5e1b9c7a60e6854	2023-09-20T12:06:08-04:00	Cebtenzzre		benchmark-matmult : do not use integer abs() on a float (#3277)
d119c04c159d015a93567df7e73e0e45a22d0f1d	8781013ef654270cbead3e0011e33a6d690fb168	2023-09-20T10:02:39+03:00	Georgi Gerganov		examples : fix benchmark-matmult (#1554)
8c00b7a6ff38e27fa1e471452b8a480913772c2a	7e50d34be68aae2cc766203703dd188e910e033a	2023-09-15T19:06:03+03:00	Georgi Gerganov		sync : ggml (Metal F32 support + reduce ggml-alloc size) (#3192)
83a53b753a9499a2a3535c93975b430cb2c828a9	5c872dbca2c7979b1f6dafc97db0774b8bbf9372	2023-09-14T20:21:25+03:00	Alon		CI: add FreeBSD & simplify CUDA windows (#3053)
d54a4027a6ebda98ab0fef7fa0c2247d0bef132a	1b0d09259e37898c519edb6c52d58f4d096f10bd	2023-09-11T19:55:51+02:00	Johannes Gäßler		CUDA: lower GPU latency + fix Windows performance (#3110)
f31b6f4e2d6def3c0bd7c75f75c0c1e8698e0589	6eeb4d90839bac1e6085e5544654ab5c319ad09a	2023-09-11T09:30:11+02:00	Kawrakow		metal : PP speedup (#3084)
ba7ffbb2517ff8cf4c689f94a9ad866f3ee71225	e64f5b55783e910d8287363895d652b4bea6527a	2023-09-08T18:01:04+02:00	Kawrakow		metal : Q3_K speedup (#2995)
15b67a66c2f2d6032415b28a699b5131962318f1	be8c9c245bd129ebabb80e0a7a8dd7daeb4d30af	2023-09-07T15:52:34+02:00	slaren		llama-bench : use two tokens in the warmup run for prompt evals (#3059)
31035681445181fb414e0def7ec3f84462b3bd97	5b8530d88c489f9d0c0ef3d0886b369f655b792e	2023-09-04T06:40:18-04:00	Cebtenzzre		llama-bench : make cpp file non-executable (#2999)
47068e517004d90f13c16352bb3b4cafd53a00cd	8f429fa5111901f9646cf998643ac5310846d487	2023-09-03T15:12:08+03:00	Georgi Gerganov		speculative : PoC for speeding-up inference via speculative sampling (#2926)
ca82cf7bac0c91d03e3d320b3a865dd006f854ac	6a31a3bd9806c85ed08266f6ab65181da0f30d03	2023-09-03T11:06:22+03:00	Kawrakow		metal : more optimizations (#2959)
13268c533177a4dc76bce0b465645d74f0d51d55	4dcd47d71df8ca4edcc31302744bd93f0c31298e	2023-09-01T13:42:41+03:00	Georgi Gerganov		metal : slight speed-up for add and mul kernels (#2917)
95b6e5212f5e4e1419de1d833d7f8d788f9f2227	44c117f41ee01c5ac8fb86bba041f08d8b87b46d	2023-08-28T23:33:27-07:00	Marcus Dunn		added `struct` to llama_dump_timing_info_yaml's `llama_context` (#2857)
44c117f41ee01c5ac8fb86bba041f08d8b87b46d	43033b7bb4858da4f591715b3babdf906c9b7cbc	2023-08-28T21:51:47+02:00	xaedes		train : mem usage and other improvements  (#2439)
43033b7bb4858da4f591715b3babdf906c9b7cbc	6b73ef120114beb5664ea94aab48d07ed248ee52	2023-08-28T19:19:18+02:00	slaren		llama-bench : set locale to utf8 (#2832)
463173a6c0ff353055eb90665794884c888c790f	eaa13a48ff4136f01c1cdb79cacd61b67ec53095	2023-08-27T16:50:33+03:00	Kawrakow		llama : speedup tokenization (#2831)
789c8c945a2814e1487e18e68823d9926e3b1454	c1ac54b77aaba10d029084d152be786102010eb2	2023-08-27T09:03:27+02:00	slaren		ci : add LoRA test to CI (#2650)
730d9c681e339b76407659344e5a2cd50af7d7d5	c7d92e6dfec3f54849f3a0ba373054d29f321ea2	2023-08-26T14:13:36-06:00	Kerfuffle		convert.py : advanced option (#2753)
154725c5436808e5c519685d0279e850596dbe62	12e2e33a977af73e75885eeee91c5575a77f4e5f	2023-08-25T15:16:19+02:00	slaren		llama-bench : add model sizes (#2771)
6bbc598a632560cb45dd2c51ad403bda8723b629	3f460a2b723c8b936ac29ecfd02f244b3adeba55	2023-08-25T12:09:42+03:00	Henri Vasserman		ROCm Port (#1087)
38b16dfca6e5032e6cfb90c1653bf1ba4cf647b4	8f8c28e89cb9531211783da697d6e7c445e2af1d	2023-08-24T12:27:25-04:00	Shouzheng Liu		metal : bug-fix when enable ggml-alloc (#2757)
6e91a1b0706c2e0e52b9d9be7ee82d3c1e7a33c1	44d5462b5cddc1c5cbcd7647646f7b55b175b01f	2023-08-24T00:07:13-04:00	Evan Jones		llama : fix grammar sometimes generating null char (#2756)
cf658adc832badaaa2ca119fe86070e5a830f8f6	a192860cfec89a38d59a943623bf595b1fe4495b	2023-08-23T23:08:04+03:00	Georgi Gerganov		llm : add Falcon support (#2717)
7f7ddd5002040804e33fcdbde44aa22f8635f57d	b8ad1b66b23f9b2e6e4531e9a62753323036a556	2023-08-23T06:31:09-03:00	IgnacioFDM		Fix ggml to gguf conversion on Windows (#2733)
8e4364f2af9cd5d57240f23e83c0e29bc068bc02	1e3bc523d8053a77df3ac7126a84d0297ee97ef6	2023-08-22T09:56:03+02:00	slaren		llama-bench : minor fixes (#2695)
097e121e2f17ed3541cf02c55ff7e9febc091b19	eaf98c2649d7da705de255712f0038ac7e47c610	2023-08-18T12:44:58+02:00	slaren		llama : add benchmark example (#2626)
a872a2b28eaefc8d464eaa535c94deeb501666f9	0919a0f73d95cfb93a1646a1d1741a0615fe2c5e	2023-08-17T03:35:53-04:00	Shouzheng Liu		ggml-alloc : fix discrepency between measure&eval (#2639)
bf83bff6742c0f1795b4c18695a13a34ac7adf62	b5ffb2849d23afe73647f68eec7b68187af09be6	2023-08-16T16:07:04-04:00	Shouzheng Liu		metal : matrix-matrix multiplication kernel (#2615)
5517d6e69214cdead000a76983b9fe175c3f8329	f31b5397143009d682db90fd2a6cde83f1ef00eb	2023-08-14T15:16:54+08:00	Jhen-Jie Hong		server : implement json-schema-to-grammar.mjs & add grammar param in the UI (#2588)
f64d44a9b9581cd58f7ec40f4fa1c3ca5ca18e1e	b19edd54d51cef5e3616c18b1d0d8626895b2cba	2023-08-13T00:24:45+02:00	Johannes Gäßler		CUDA: Fixed OpenLLaMA 3b mmq, reduced compile time (#2590)
53dc399472d5bd35ee739b865e843b1996bd3814	9ca4abed893685692f90413e4d43153af12342d9	2023-08-12T06:35:14+08:00	Equim		server: fixed wrong variable name in timing json (#2579)
1a941869cbef8e9cc351a6c6987e4ae3b0f021f7	b5472ea0ada081a6e1c06998ebbc9a24aa2cd4a4	2023-07-27T11:00:54+03:00	Georgi Gerganov		metal : disable graph concurrency optimization due to bug (#2413)
da1889834a036a63ead2b0ca5c9ed8967712568c	82552b7f5403ca13957ac9a2cdc1732470057b62	2023-07-25T14:32:20+02:00	slaren		ggml : improve graph build time via hash table lookup (#2329)
1aa18ef994a6a2b531434eb13251ef48e56d345b	9a08eaf3c4010962d0126e9e5bfbe9af64b2ac90	2023-07-25T08:00:19-04:00	Shouzheng Liu		metal : concurrently dispatch commands (#2358)
9a08eaf3c4010962d0126e9e5bfbe9af64b2ac90	129d844c87d90e74aafc23dcc84c980fd408def4	2023-07-25T13:48:29+03:00	Kawrakow		Another speed gain for Q4_0 and Q4_1 on Metal (#2375)
2f9cf974a066ac0e03fbb235d834b01b0164d743	4f06592cc6b83979e4b442e8cb97b3948c857188	2023-07-24T00:19:47+03:00	Kawrakow		Some more Q4_K and Q5_K speedup on CUDA (#2346)
d2a43664f93ba30a84e42713bb69f936cbdacf2a	b9b7d94fc10a8039befd1bc3af4f4b09c620c351	2023-07-23T08:49:20+03:00	Kawrakow		Speed up Q4_K (#2322)
b47b8a9cfeb439d271bf997fb985fd6d82b3af5e	b5fe67f8c69113bd9354bc1adcfe2df6be323740	2023-07-22T21:17:57+03:00	Georgi Gerganov		llama : optimize memory buffers (#2325)
5d500e8ccf5eee3de3ae66685cc3be75e43e08b9	7d5f18468ceabd7a38f414f9f21b26b0c137f994	2023-07-22T11:48:22+03:00	Georgi Gerganov		ci : add 7B CUDA tests (#2319)
4d76a5f49b9b5382dba5d13d92edb9159536c225	0db14fef06836caaa13cc123c0a24dc598bdb9f0	2023-07-21T17:05:30+03:00	Kawrakow		Faster Q3_K implementation on Metal (#2307)
e782c9e735f93ab4767ffc37462c523b73a17ddc	785829dfe8baf0213f2ff66963d28c62f92d7930	2023-07-20T18:19:45+03:00	Kawrakow		Faster Q5_K and Q6_K on Metal (#2294)
417a85a0010519224cf154eb85d383ffeafeeead	294f424554c1599784ac9962462fc39ace92d8a5	2023-07-20T06:32:22-04:00	Shouzheng Liu		metal: minor q4 optimization and reduce code size (#2248)
294f424554c1599784ac9962462fc39ace92d8a5	45a1b07e9b20c33d71d8c849ff27d693a75a0269	2023-07-19T15:06:40+08:00	Rinne		llama : extend API to get max devices at runtime (#2253)
672dda10e4d8ac79df5d5970da7fb69d242ca9a7	27ab66e437797aedbb23b3599385756b6c26ac39	2023-07-17T03:57:28+08:00	Qingyou Meng		ggml : fixed runtime bugs and compile errors related to GGML_PERF and GGML_DEBUG (#2219)
32c54116318929c90fd7ae814cf9b5232cd44c36	ff5d58faecf1f02b05bd015bdfc6a394cf2bc9ba	2023-07-13T21:58:25+08:00	Howard Su		Revert "Support using mmap when applying LoRA (#2095)" (#2206)
c9c74b4e3f9dcfab8b0032749ff8a579ab4e4d8d	3ec7e596b2ba3f43c22f441254ca2bcfa91102ba	2023-07-12T00:18:43+08:00	Bach Le		llama : add classifier-free guidance (#2135)
18780e0a5e17348236230bbe891901b9b5718709	3bbc1a11f04a9adc0d0e08c2940ba4d2978755ab	2023-07-09T05:20:43-03:00	JackJollimore		readme : update Termux instructions (#2147)
1d656d6360359cfdaaf5d64ed9690047b600dbcb	72421402834141df6cbdcf595fe46dbd11874dce	2023-07-08T00:24:01+08:00	Qingyou Meng		ggml : change ggml_graph_compute() API to not require context (#1999)
a17a2683d8fdb899ba497d0c28ccafb28c62efb6	31cfbb1013a482e89c72146e2063ac4362becae7	2023-07-06T09:17:50-07:00	tslmy		alpaca.sh : update model file name (#2074)
31cfbb1013a482e89c72146e2063ac4362becae7	983b555e9ddb36703cee4d22642afe958de093b7	2023-07-05T16:51:13-04:00	Tobias Lütke		Expose generation timings from server & update completions.js (#2116)
051c70dcd55709c9cbbfa849af035951fe720433	9e4475f5cf639315f61ed7b8da6258bb0c7c5ca9	2023-07-05T18:31:23+08:00	Howard Su		llama: Don't double count the sampling time (#2107)
b2132270678c473f7cd9ba871b03d694126bc33a	2f8cd979ecd1fa582852e7136e92ff8990b98fd8	2023-07-01T20:31:44+02:00	Daniel Drake		cmake : don't force -mcpu=native on aarch64 (#2063)
b8c8dda75fdf5fdea49c80af36818e7c30fe0ddf	96a712ca1b7f427e3bd7ffc0c70b2105cfc7fbf1	2023-06-29T21:15:15+08:00	Howard Su		Use unsigned for random seed (#2006)
96a712ca1b7f427e3bd7ffc0c70b2105cfc7fbf1	d3494bb86bf7ad5b0b60aae0220ea576f273b5c0	2023-06-29T11:56:43+08:00	LostRuins		Porting the improved K-Quant CUDA kernels to OpenCL (#1966)
a84ab1da8dc6a59a5b67420ae1322f09503ffc72	5743ca80928d8410754ec64a5673d5c2dd6cfbb7	2023-06-27T01:47:02+09:00	katsu560		tests : fix quantize perf (#1990)
6769e944c727c63612dcafbef52009d21ae00fff	cbebf61ca7584e9709265395f0127ae7fc0f1882	2023-06-26T19:43:07+03:00	Kawrakow		k-quants : support for super-block size of 64 (#2001)
18b35625c3c19c64b7818a12460ba5ddb006dfdc	ba4e85a8339b9dd7cdffad31838235f2fe45a8ea	2023-06-19T20:43:30+03:00	Georgi Gerganov		ggml : fix bug in LBFGS optimizer (found by ggml tests)
794db3e7b982fee37e3995db9c3a216a57ff65e3	5ddf7ea1fb42bac21026de2f77e0f9c069b92234	2023-06-17T07:53:04-04:00	Randall Fitzgerald		Server Example Refactor and Improvements (#1570)
3d0112261042b356621e93db3fa4c6798a5d098f	602c748863e15270d80d74aa2c3bf86ab8139e07	2023-06-16T20:08:44+03:00	Kawrakow		CUDA : faster k-quant dot kernels (#1862)
a09f9195be39afb4b023b646c0a6ec8a86915174	bed92756172d4514b23aaf9744cf8e2dc892fc7b	2023-06-15T21:49:08+02:00	Johannes Gäßler		Fixed CUDA runtime version check (#1879)
254a7a7a5ff4c874ff8488f1f5cbdd7e9c89d682	92549202659fc23ba9fec5e688227d0da9b06b40	2023-06-14T19:47:19+02:00	Johannes Gäßler		CUDA full GPU acceleration, KV cache in VRAM (#1827)
e32089b2c20b1b87b22912f4a8b93fe01647d5b9	2347e45e7bdb09c9a7d74b2c0bc86c2b65f0c343	2023-06-13T21:04:40+02:00	xaedes		train : improved training-from-scratch example (#1652)
2347e45e7bdb09c9a7d74b2c0bc86c2b65f0c343	74d4cfa3438cb58bd177eed30014e6588694aaa8	2023-06-13T20:20:07+03:00	Georgi Gerganov		llama : do a warm-up eval at start for better timings (#1824)
74a6d922f12ccfe16b0c265f43be8978c6f25e98	e4caa8da59c1c97dc23fa336f4d726984a20560f	2023-06-12T22:39:21+03:00	Kawrakow		Metal implementation for all k_quants (#1807)
e9b66ee9829039d4ab54550d6222e42a0b31e52a	4f0154b0bad775ac4651bf73b5c216eb43c45cdc	2023-06-10T11:28:11+03:00	Kawrakow		metal : add Q4_1 implementation (#1785)
4f0154b0bad775ac4651bf73b5c216eb43c45cdc	ef3171d16241c18581d4d08374f0b9e396ade6b7	2023-06-10T01:59:17-06:00	Kerfuffle		llama : support requantizing models instead of only allowing quantization from 16/32bit (#1691)
72ff5282bf0388c60821f504c4c8cc2b1f491aa6	0bf7cf1b296fc9fca05411b37afdf08a531487d2	2023-06-08T22:28:21+03:00	Kawrakow		metal : add Q2_K implementation (#1762)
0f291e1f65c1d68201e71ce99c89562a36686b6d	8fc8179919a11738910db07a800f2b176f8adf09	2023-06-08T19:46:22+03:00	Kawrakow		metal : Q6_K implementation (#1752)
4161bdc04debb70bf5f275492b4d89fd9330087c	0035858273ebe0694926bf4414d279f3e1cd109d	2023-06-08T10:08:23+03:00	Kawrakow		metal : add Q4_K implementation (#1733)
5c64a0952ee58b2d742ee84e8e3d43cce5d366db	5b57a5b72676540b6a45a3f527126299969ad241	2023-06-07T10:59:52+03:00	Georgi Gerganov		k-quants : allow to optionally disable at compile time (#1734)
35a84916fb029905c44746127026079268216e7a	2d7bf110edd8c49209401a16132052cba706ffd0	2023-06-07T04:10:17+02:00	Willy Tarreau		main: add the possibility to open the prompt cache read-only (#1640)
f4c55d3bd7e124b101bc974cbbf0e0dbbc32d5a3	f1465624c2cbc8ee65b566e3d87f2af27796d4c4	2023-06-05T23:32:36+03:00	Yuval Peled		docs : add performance troubleshoot + example benchmark documentation (#1674)
efe05076323f5c6bafece109e21cce046f5e4b07	e7fe66e670537990ccc075cce9286df88bba052a	2023-06-05T22:11:49+02:00	grahameth		ggml : fix internal overflow in ggml_time_us on Windows (#1702)
99009e72f8072fa552eb02efee436be596c71cdd	5220a991a5e92bddad9542267ab445a2c033681c	2023-06-05T22:56:18+03:00	Kawrakow		ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684)
5220a991a5e92bddad9542267ab445a2c033681c	d1f563a743a83dabc11e125d4a7d64189c16498c	2023-06-05T13:43:08+03:00	Henri Vasserman		Increase 3B scratch buffers. (#1698)
ecb217db4fcfa3880300ad08531a5fb6bb142d45	dcb2ed48268e421baf25adc00d602dad0f415564	2023-06-04T23:34:30+03:00	Georgi Gerganov		llama : Metal inference (#1642)
1fcdcc28b119a6608774d52de905931bd5f8a43d	ac7876ac20124a15a44fd6317721ff1aa2538806	2023-05-25T23:07:29+02:00	Johannes Gäßler		cuda : performance optimizations (#1530)
affc76edfdefa7b326f526e463cc65ff13fcfb92	ea600071cb005267e9e8f2629c1e406dd5fde083	2023-05-20T14:19:28+02:00	Johannes Gäßler		cuda : loading models directly into VRAM, norm calculation on GPU, broadcasting for ggml_mul (#1483)
2d5db48371052087a83974abda3767d1aedec598	6986c7835adc13ba3f9d933b95671bb1f3984dc6	2023-05-19T22:17:18+03:00	Georgi Gerganov		ggml : use F16 instead of F32 in Q4_0, Q4_1, Q8_0 (#1508)
c238b5873a1ea496db03ffcfe124c9d0d83afbc6	2b2646931bd2a2eb3e21c6f3733cc0e090b2e24b	2023-05-17T22:47:58+08:00	rankaiyx		benchmark-matmul: Print the average of the test results (#1490)
b5c9295eef2b56e307393b35b3a923e3518d226e	eb363627fda5f47de8ab5e9be8abd426049d00df	2023-05-14T22:46:00+02:00	slaren		benchmark-matmul: fix clang-tidy issues, report results in GFLOPS (#1458)
bda4d7c215aa16b2a78e522521dfc0e1c2e8b194	5a5aeb1e91009c72bf816400b758bb8a305616d7	2023-05-13T17:25:09+03:00	Georgi Gerganov		make : fix PERF build with cuBLAS
f954edda935a70a14cf0cc45ecc7fe7d60cf3e4b	f048af0230ad5381bb20b9e3c1467ec6fc7debdf	2023-05-13T14:56:40+02:00	xaedes		ggml : implement backward pass for llama + small training-llama-from-scratch example (#1360)
b9fd7eee57df101d4a3e3eabc9fd6c2cb13c9ca1	b608b55a3ea8e4760c617418538465449175bdb8	2023-05-12T00:23:08+03:00	Georgi Gerganov		ggml : remove bit shuffling (#1405)
1b0fd454650ef4d68a980e3225488b79e6e9af25	3924088512d9e12e90ed6dbf28a6c5712481d33e	2023-05-07T10:03:23+07:00	swittk		ggml : Allow usage of CLBlast alongside Accelerate.framework (#1336)
58b367c2d757c0ea12aec672382462b42204c724	ea3a0ad6b6b5ca4693b94acd4cb32e2803f66fae	2023-05-01T18:11:07+02:00	slaren		cuBLAS: refactor and optimize f16 mat mul performance (#1259)
f0d70f147d969e41fa410b8af2965a27aa901eb9	3e5aa8a1c44051153d6d7b3eeca2f4b4e5fb310c	2023-04-30T12:32:37Z	Stephan Walter		Various fixes to mat_mul benchmark (#1253)
7296c961d9303010a2b98379f738da2a8a55aa1b	78ec543733d10a1629f984fd0302fdaa4e87fe66	2023-04-28T16:57:16+02:00	0cc4m		ggml : add CLBlast support (#1164)
574406dc7e350ddbffaeca33bf0392b7bfeb1436	87a6f846d3e929632c45916dd08f1e2a9c72d2a3	2023-04-26T23:14:13+03:00	Georgi Gerganov		ggml : add Q5_0 and Q5_1 quantization (#1187)
87a6f846d3e929632c45916dd08f1e2a9c72d2a3	ea3ad7eb60cfb44526a58122e8019850f437cd1b	2023-04-26T20:08:43Z	Ásgeir Bjarni Ingvarsson		Allow setting the rng seed after initialization. (#1184)
ec9cdb6752dd96b3cc74d90ad1adeba5b4fa2b0e	e4422e299c10c7e84c8e987770ef40d31905a76b	2023-04-23T18:32:52+03:00	Georgi Gerganov		ggml : do not print perf ops that have not been used at all
e4422e299c10c7e84c8e987770ef40d31905a76b	53c8434398b3cba7ac6298cdd44abd40f0e640b1	2023-04-23T18:15:39+03:00	Georgi Gerganov		ggml : better PERF prints + support "LLAMA_PERF=1 make"
5f939498d517b4dddbe904f202e895a3ecfb9dc4	36b4f7e06406eed8a605cc9f2921d9244ef6a8e5	2023-04-22T11:10:39+02:00	unbounded		ggml : unit test for quantization functions (#953)
36b4f7e06406eed8a605cc9f2921d9244ef6a8e5	10f19c1121068ce3dab9bece03a8b9caaea2db36	2023-04-22T16:56:35+08:00	wbpxre150		llama : print timings on ctrl+c exit (#1021)
c5aa5e577741d0359ad26ec50b9e21a74c65d911	e9a9cb0c54461ffbda75b7b2f99f3ea5562291c2	2023-04-22T07:37:05Z	Stephan Walter		ggml : AVX2 optimization for vec_dot_q4_3_q8_0 and refactoring (#1099)
50cb666b8a2e35a49b08c0f6bc81138c8f6f2ac1	25d7abbd1f73582b7e0fdc422a936e8541c0780b	2023-04-21T21:59:17+02:00	slaren		Improve cuBLAS performance by using a memory pool (#1094)
1bfc153e2f35ddd9d64b084e8d1a5e6fa57ad1c9	3d59769c3bb7e72c915646ddb1e239b1face19f5	2023-04-21T17:18:26+02:00	Kawrakow		ggml : a faster version for Q4_1 x Q8_0 dot products (#1083)
c8c2c524827be8fd681a63f0e5a697b0bf4c587b	02d6988121510c067e06d498a273a351a888f5b9	2023-04-20T06:45:41Z	Stephan Walter		AVX2 optimization for vec_dot_q4_2_q8_0 (#1068)
02d6988121510c067e06d498a273a351a888f5b9	834695fe3a3ed2a962e774c9615e3f7b41d360a8	2023-04-20T03:14:14+02:00	slaren		Improve cuBLAS performance by dequantizing on the GPU (#1065)
f7d05095b404b5500b4a702ea16f67fc22446e49	884e7d7a2bfd7325b107442d6758983f5886ed3d	2023-04-19T20:20:14+02:00	Kawrakow		Q4_2 quantization with rmse-optimized scale and quants (#1062)
884e7d7a2bfd7325b107442d6758983f5886ed3d	7cd5c4a3e9106151d48f328bb3c94c298a211f18	2023-04-19T20:10:08+03:00	Georgi Gerganov		ggml : use 8-bit precision for Q4_1 intermediate results (#1047)
77a73403ca8eaced2590559d0f9cebd2b3649d32	50a8a2af97cb92e53e7a3195aa201c3d87da5415	2023-04-18T23:54:57+03:00	Georgi Gerganov		ggml : add new Q4_2 quantization (ARM only) (#1046)
50a8a2af97cb92e53e7a3195aa201c3d87da5415	4caebf6d408b91c2d29d0abc7b1e867b5de64db5	2023-04-18T23:11:23+03:00	Georgi Gerganov		ggml : scratch that - vmlaq_n_f32 is always better
dcdd65e2969bc03c91a1ebd1160162d5054e6923	5ecff35151156118c2df74899637ad34ee384b9b	2023-04-18T22:59:17+03:00	Georgi Gerganov		ggml : optimize ggml_vec_dot_q4_0_q8_0() using vectorized accumulators
5ecff35151156118c2df74899637ad34ee384b9b	7faa7460f03bdd88becf1e659cf359f274055404	2023-04-18T21:00:14+02:00	Kawrakow		Adding a simple program to measure speed of dot products (#1041)
f266259ad9a2bce5a34d919592310147af23f3dc	47f61aaa5f76d04286792e2fbd0c95b659ab2af0	2023-04-17T15:10:57+02:00	Ivan Komarov		Speedup the AVX-512 implementation of ggml_vec_dot_q4_0() (#933)
489537e6cf6c93b74a029a11533dbcaa89791dcc	2d3481c72125cd388258864c7ad8d7d36777bad7	2023-04-16T12:13:00+02:00	Pavol Rusnak		examples: add missing <ctime> include for time() (#1011)
e95b6554b493e71a0275764342e09bd5784a7026	aa485cee334e84437e21681c14b6f80b65876d8b	2023-04-15T17:53:22+03:00	Georgi Gerganov		ggml : add Q8_0 quantization for intermediate results (#951)
c12b14b77fced0ce9a0e2d81f670c3a746dec251	106faaf2971d6c89d6010279a9a95737772470ef	2023-04-15T07:51:54+02:00	Ivan Komarov		benchmark : fix result validation in benchmark-q4_0-matmult (#987)
723dac55fa2ba7adc6e3fc8609781d1ad0378906	0f07cacb05f49704d35a39aa27cfd4b419eb6f8d	2023-04-14T00:03:03-07:00	comex		py : new conversion script (#545)
c5d70f5c9ea5a8f0f6b0d6aa741455978a1dabfd	be87b6ed20a5f7528bf491a83e759a9fc6a24fea	2023-04-14T14:24:52+08:00	Howard Su		ggml : optimize rope function to avoid call powf in the tight loop (#807)
d990e3fffc5b0f5448e90a16c79a4f2675100af0	9190e8eac8bdc108c40d2d7505e9b45fa773251f	2023-04-13T18:32:36+03:00	Georgi Gerganov		ggml : speed-up ggml_vec_dot_q4_1() ARM_NEON + 32-bit ARM support (#900)
c85980acd04631a7c43d13676276f76ec72f5dfe	6232f2d7fd7a22d5eeb62182b2f21fcf01359754	2023-04-13T18:01:22+03:00	Georgi Gerganov		gitignore : benchmark
6232f2d7fd7a22d5eeb62182b2f21fcf01359754	6c248707f51c8a50f7792e7f7787ec481881db88	2023-04-13T14:59:50Z	Stephan Walter		ggml : optimize non-SIMD Q4_0 vector dot product (#703)
95ea26f6e92d620a5437f576b80868aee7f808d6	82d146df9b43cf677e0dbce20b03cf864958a0cc	2023-04-13T14:46:23+02:00	SebastianApel		benchmark : add tool for timing q4_0 matrix multiplication (#653)
c3ac702e5ee3533457e0489df4906ee112fe88e7	9d634ef452d0fc24fcd49592952d13d0ab0f41b7	2023-04-10T22:40:28+03:00	Georgi Gerganov		ggml : add ggml_cont() + optimize ggml_cpy() for contiguous dst
f963b63afa0e057cfb9eba4d88407c6a0850a0d8	aaf3b23debc1fe1a06733c8c6468fb84233cc44f	2023-04-08T12:24:37-07:00	comex		Rewrite loading code to try to satisfy everyone:
eeaa7b0492fc79baab8bb1fe195d6c87159f2bd3	986b6ce9f99503c51ec5afd8a10baa32359434c6	2023-04-05T22:11:03+03:00	Georgi Gerganov		ggml : multi-thread ggml_rope() (~3-4 times faster on M1) (#781)
986b6ce9f99503c51ec5afd8a10baa32359434c6	34162989297fdfe3ab7305451ce55bc87e3f4c9c	2023-04-05T22:07:33+03:00	Georgi Gerganov		ggml, llama : avoid heavy V transpose + improvements (#775)
0c44427df10ee024b4e7ef7bfec56e993daff1db	594cc95fabab0b662dabba3ea619ca5dca18bf6b	2023-04-05T17:38:37+03:00	Ivan Stepanov		make : missing host optimizations in CXXFLAGS (#763)
437e77855a54e69c86fe03bc501f63d9a3fddb0e	cd7fa956904cb8e321b72b3499f4a3a82e43c266	2023-04-03T09:52:28+02:00	SebastianApel		10+% performance improvement of ggml_vec_dot_q4_0 on AVX2 (#654)
1d08882afa647c44195f4f6495a68ea455650cae	02c5b27e91a6d18cf1043d3a2d8dbc59610ac257	2023-03-31T17:55:52+02:00	slaren		Optimize AVX2 ggml_vec_dot_q4_0 (#642)
78ca9838ee36660a776e97e3391b6fb5dcaacf7f	a017390358cdb23fffb30988dc84bb190d0403ca	2023-03-29T13:51:37-07:00	Justine Tunney		Make loading weights 10-100x faster
b51c717d5cf9181c33afcb84554e47f6d539c891	0ba76c1e73ae21038b80bfb5a746157376c88173	2023-03-29T22:15:34+03:00	Georgi Gerganov		ggml : init time on first ggml_init() call
53635c081c49321d523567112f9fddfbba6b787b	41318d708ed196ff727dce14d263a64b23c7333d	2023-03-29T19:29:26+03:00	Georgi Gerganov		py : add GPT4All conversion script
436e56193199a1625f8c561069f702e8840a9e08	20e1e84884376b3fb44ffbfd48d478b2934b0b5e	2023-03-28T16:48:20Z	Stephan Walter		all : be more strict about converting float to double (#458)
29b7baab670ae8b76ac0da21c2ded69ff18971ee	4a7129acd2e939b92d70dd568c746f2fa078232c	2023-03-25T15:34:23+01:00	slaren		Add timings for the prompt evaluation (#478)
481044d50cfe8eaa6cd0c1a1b445680e4b0b3ebc	563cdc391dde140f1084d1012234e8e6f57f881f	2023-03-24T08:19:26-07:00	Cameron Kaiser		additional optimizations for POWER9 (#454)
8d4a855c241ecb0f3ddc03447fe56002ebf27a37	b6b268d4415fd3b3e53f22b6619b724d4928f713	2023-03-24T08:05:13-07:00	Luciano		Add embedding mode with arg flag. Currently working (#282)
305ba6f0e6daa3796aad9dd18053a1945dd4cc58	4122dffff958cd137175b58f1f27c0913528d7ba	2023-03-22T18:16:35+01:00	tjohnman		Don't force immediate interactive without `-i` (#354)
f5a77a629bd0f37ae1696747633ab42a5530ec15	da0e9fe90ccf6e73597eb19dd0cfc0a28363fb3b	2023-03-22T07:32:36+02:00	Georgi Gerganov		Introduce C-style API (#370)
353ec251a42491f5192c48561da4b444ef67f23c	89d5d90f3b6d25f134da7a8e252c3432bffcf674	2023-03-21T14:21:50-03:00	Fabio R. Sluzala		We could use std::unordered_map over std::map (#305)
16ffc013c62f22bdaa3cdc022d7a13fd952d73fc	486ae645fd3eda8b9d7413d5ff34fb65a3e337fb	2023-03-21T09:42:25-07:00	comex		Importer for GPTQ quantized LLaMA models (#301)
486ae645fd3eda8b9d7413d5ff34fb65a3e337fb	3ab3e6582f7320c2b6568c892fdfc8215caf7e6c	2023-03-21T09:27:42-07:00	Gary Linscott		Compute perplexity over prompt (#270)
2e664f1ff413995506c9a54f3a8d5b8c64e37a91	8cf9f34eddc124d4ab28f4d2fe8e99d574510bde	2023-03-21T07:35:42-07:00	Casey Primozic		Add initial AVX512 support for dot product on Linux (#320)
2af23d30434a677c6416812eea52ccc0af65119c	904d2a8d6acd667c9633138d45a361d40fbf76d0	2023-03-17T10:47:06+01:00	Bernat Vadell		🚀 Dockerize llamacpp (#132)
904d2a8d6acd667c9633138d45a361d40fbf76d0	721311070e31464ac12bef9a4444093eb3eaebf7	2023-03-17T05:48:39+01:00	Matvey Soloviev		Q4_1 quantization (#193)
84d9015c4a91ab586ba65d5bd31a8482baf46ba1	63fd76fbb06f9b723ca11505352387a3148b1814	2023-03-13T18:36:44+02:00	Georgi Gerganov		Use vdotq_s32 to improve performance (#67)
63fd76fbb06f9b723ca11505352387a3148b1814	2a20f48efad692a8c2744f10c673bbdbe0c751b7	2023-03-14T01:33:43+09:00	uint256_t		Reduce model loading time (#43)
c80e2a8f2adeda202cbffe76ef800f134e51f03f	54a0e66ea0ed3248e6c95a070a2da0bf5c6d4817	2023-03-13T01:28:08+02:00	Georgi Gerganov		Revert "10% performance boost on ARM"
113a9e83ebc0f788f861394437087bf3ca0e019b	404fac0d623c9eea74ad7a9347da69e33f10984e	2023-03-13T00:56:10+02:00	Georgi Gerganov		10% performance boost on ARM