cogito/calculet-npu-research-archive
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
98e6cadcb6c28c7800d3bb013eab848b74510b35
calculet-npu-research-archive/history/llama.cpp/llama.cpp-dev-vs-master-commits.tsv
T

140 KiB
Raw Blame History

1fd9bd632f346085920d523dd307a3e4c11cc0a05e08e473cf292ba73789769b993bb35bcdfc316892026-06-01T14:15:35+08:00wangbomengllama-bench: fix device-info
2e08e473cf292ba73789769b993bb35bcdfc31689f86112d6758298241ab2fa84ad4f0f7b1ace35c62026-06-01T11:47:01+08:00wangbomengllama-bench:add device-info
3f86112d6758298241ab2fa84ad4f0f7b1ace35c6d1bc743c10080fed1253686ce01749b42611ded42026-05-29T16:40:38+08:00wangbomengadd dump_pos,fix pos_tensor of pre_by_embeds
4d1bc743c10080fed1253686ce01749b42611ded42f201160b1960fd7be18b70d5770599d7082dd2f2026-05-28T14:23:06+08:00wangbomengset u_batch = max_seq_len
52f201160b1960fd7be18b70d5770599d7082dd2ff131bf1908b8c5dd1f49d7ae7f616506fc4716662026-05-27T09:58:31+08:00wangbomengdelete useless code
6f131bf1908b8c5dd1f49d7ae7f616506fc47166633d96dae28e17208ef61a71dba40cda3c04f135a2026-05-27T09:38:51+08:00wangbomengfix patch of qwen2.5vl,n_patch of minicpm = 64,fix n_ctx_slot,comment out M-ROPE
733d96dae28e17208ef61a71dba40cda3c04f135a37f068d87cc2d0a0b40ce978031fc9932fc890772026-05-22T14:33:45+08:00wangbomengadd llama-build.sh
837f068d87cc2d0a0b40ce978031fc9932fc8907713508e5e18bbb1202ae60c2ee5ecbdda1cca817b2026-05-21T14:07:00+08:00wangbomengrelease
913508e5e18bbb1202ae60c2ee5ecbdda1cca817bb6b7919468126d5ba30a10941825154789e9082f2026-05-21T11:43:59+08:00wangbomengreduce cmake log
10b6b7919468126d5ba30a10941825154789e9082f46847c24fc5c52dd57bb4fb17fc4d684149bb4fa2026-05-20T16:40:33+08:00wangbomengreduce cmake log
1146847c24fc5c52dd57bb4fb17fc4d684149bb4fad028722a12c5a463cc97207787cfd03d7a2590b02026-05-20T16:02:12+08:00wangbomengSimplify CMakeLists.txt
12d028722a12c5a463cc97207787cfd03d7a2590b063442fde8dc285817f725bd6c5fae80f478a38132026-05-20T14:41:05+08:00wangbomengreplace cparam.n_ubatch with n_ubatch()
1363442fde8dc285817f725bd6c5fae80f478a3813fc3f4a9fab88e98eff8eeca8cbc6617333edba2c2026-05-20T11:47:42+08:00wangbomengfix ubatch and cmakelist
14fc3f4a9fab88e98eff8eeca8cbc6617333edba2cc931ef3163940227c9b6291e04216245cce214292026-05-19T16:54:16+08:00wangbomengadd dump and rt case generation
15c931ef3163940227c9b6291e04216245cce21429a43741244a646d3f8fa3ff01bb9b7d059223d0a52026-05-18T17:34:37+08:00wangbomengcalrt: fix prefill_by_ids.fix create_buf
16a43741244a646d3f8fa3ff01bb9b7d059223d0a5b2003d6c90ef0fa26ab9e930ba3e22b355eb3d392026-05-15T14:39:53+08:00wangbomengrun llama-passkey successfully, fix llama-bench, copy 3 times input to infer qwen2.5vl
17b2003d6c90ef0fa26ab9e930ba3e22b355eb3d397800c772d44a7415640ec1669a691232939927cf2026-05-13T15:24:24+08:00wangbomengtry passkey and embedding
187800c772d44a7415640ec1669a691232939927cfdde8b8228081769c8d8f1e0751d47fa7875f84232026-05-13T15:23:54+08:00wangbomengtry passkey and embedding
19dde8b8228081769c8d8f1e0751d47fa7875f8423fabcc7dac4b34d83fe430980bb364b81087c7ee92026-05-12T10:46:08+08:00wangbomengclip:add minicpm patch;
20fabcc7dac4b34d83fe430980bb364b81087c7ee91e9787962f20a7e82c71589ac1dd0585454e4bfe2026-04-30T16:17:17+08:00wangbomengcalrt: fix encode dump
211e9787962f20a7e82c71589ac1dd0585454e4bfef584311c716dd078c4b737eebbfda97fe12b7274 1bec0db5a675ddc60fb793be2a746e8f3c8855dc2026-04-30T16:06:23+08:00wangbomengMerge remote-tracking branch 'origin/dev-ben' into dev
22f584311c716dd078c4b737eebbfda97fe12b7274a339954cc2a145a80c5ea349e7896211916cbed92026-04-30T16:05:15+08:00wangbomengcalrt: fix encode dump
231bec0db5a675ddc60fb793be2a746e8f3c8855dc465df20c3e1f3f58d0f5531c796e0941a49c32b02026-04-30T16:02:59+08:00wangbomengbench: support 2 calbins
24465df20c3e1f3f58d0f5531c796e0941a49c32b0a339954cc2a145a80c5ea349e7896211916cbed92026-04-30T09:08:12+08:00wangbomengsupport one calbin
25a339954cc2a145a80c5ea349e7896211916cbed98e76fc330ad624d1768b412d7894e272dbe696f4 eeeef3f6d6a5aad9296323ca15d9d929745b89322026-04-29T14:57:55+08:00wangbomengMerge branch 'dev' into dev-cli
268e76fc330ad624d1768b412d7894e272dbe696f4431ad6f1c787f1b47e9e7be14b2e1a7d8565485c2026-04-29T14:56:59+08:00wangbomengfinish dev of llama-cli
27eeeef3f6d6a5aad9296323ca15d9d929745b8932431ad6f1c787f1b47e9e7be14b2e1a7d8565485c2026-04-28T16:16:19+08:00wangbomengadd unknown time info: prefill and decode
28431ad6f1c787f1b47e9e7be14b2e1a7d8565485cefa1876bb8b2a449a55054a8c3745892f2c0f2622026-04-27T16:52:15+08:00wangbomengupdate calrt_infer
29efa1876bb8b2a449a55054a8c3745892f2c0f2624a2a3181f1cea170105c771e6ba69d851b82b9372026-04-27T11:35:58+08:00wangbomengupdate version print
304a2a3181f1cea170105c771e6ba69d851b82b937e3ea5541259f35871c3e995331770aaca297e9ec2026-04-26T10:56:02+08:00wangbomengadd calrt-version
31e3ea5541259f35871c3e995331770aaca297e9ec0b8e43943846a1cd05c26c9186b9ef778ebae6e22026-04-26T10:55:48+08:00wangbomengadd calrt-version
320b8e43943846a1cd05c26c9186b9ef778ebae6e25eb46fffbd2e260c540fb112314ad67bbb8f56b62026-04-26T08:43:43+08:00wangbomengto debug
335eb46fffbd2e260c540fb112314ad67bbb8f56b6005112c51e00d998b457337ca3dad53fc46e2d0b2026-04-24T16:24:51+08:00wangbomengclip: add image_to_patches; calrt: vision copy data from bf16 to fp32
34005112c51e00d998b457337ca3dad53fc46e2d0b509670642e8090a1713f5d73590b549be827aaef2026-04-24T16:24:31+08:00wangbomengclip: add image_to_patches; calrt: vision copy data from bf16 to fp32
35509670642e8090a1713f5d73590b549be827aaef5d219bfcef239c1e3ddd1baecc0514a9462ff6a82026-04-24T09:40:34+08:00wangbomengrm debug code
365d219bfcef239c1e3ddd1baecc0514a9462ff6a86787a53ab9b4897c5a7b51a4062d7b90e6697d352026-04-24T09:31:32+08:00wangbomengcalrt: recover base = batch_index * dict_len
376787a53ab9b4897c5a7b51a4062d7b90e6697d35e1a76abf66178ea85656479a5e19aea7785c09f42026-04-24T09:22:02+08:00wangbomengcalrt:input of multimodel from fp32 to bf16
38e1a76abf66178ea85656479a5e19aea7785c09f499dd308adb8488fa869bb669e3335e646a938eff2026-04-22T16:29:22+08:00wangbomengcalrt:add CalcoreRT version print
3999dd308adb8488fa869bb669e3335e646a938eff06c7852e99e34c71e24e3ef6001cb54c72970e4d2026-04-21T15:50:01+08:00wangbomengserver: fix max_seq_len
4006c7852e99e34c71e24e3ef6001cb54c72970e4d96b4e267e562f44e76ec6c965ee0fa361f0439bf ed62eb33447f993d578e156a3c7e38c91b420ece2026-04-21T15:06:01+08:00wangbomengMerge branch 'dev-ot' into dev
41ed62eb33447f993d578e156a3c7e38c91b420ece015e988a271573a7c11e2a4b45084e1219d69d502026-04-21T15:05:22+08:00wangbomengtry server-test:not success
42015e988a271573a7c11e2a4b45084e1219d69d5099962021f2dd1994dbbf90d5712f0ca2633f20f82026-04-17T16:56:08+08:00wangbomenglittle change
4396b4e267e562f44e76ec6c965ee0fa361f0439bf99962021f2dd1994dbbf90d5712f0ca2633f20f82026-04-17T15:33:01+08:00wangbomengmerge dev-ot
4499962021f2dd1994dbbf90d5712f0ca2633f20f89af6682ef101c6d006c743c703cf150ac25f466f2026-04-17T14:48:33+08:00wangbomengcalrt: add onetoken slice and untile_prefill
459af6682ef101c6d006c743c703cf150ac25f466fe24f775ee26ea5249d5ff6cd23b8a0a17fe1438d2026-04-16T10:09:44+08:00wangbomengargs: add -ca
46e24f775ee26ea5249d5ff6cd23b8a0a17fe1438da206460e610aef995e36e50b33e8bd41fe4775412026-04-15T10:01:09+08:00wangbomengcalrt: add copy_data_to_ibuf for vision model
47a206460e610aef995e36e50b33e8bd41fe477541d8b2aaa00049978cebac15041f960d75b552f97b2026-04-14T10:23:43+08:00wangbomengcalrt: support mixture of prefill_by_ids and prefill_by_embeds
48d8b2aaa00049978cebac15041f960d75b552f97bb6f29ec8140028619efaa50aa6e73146cecf631d2026-04-13T11:23:44+08:00wangbomengserver: fix context-shift
49b6f29ec8140028619efaa50aa6e73146cecf631d6d55eae7e76b768acfcec045b8e84cded8ae3b552026-04-13T11:23:31+08:00wangbomengserver: fix context-shift
506d55eae7e76b768acfcec045b8e84cded8ae3b55af8055f5f033a730e7e1fe80185b18b7e94739662026-04-10T14:58:56+08:00wangbomengadd annotations
51af8055f5f033a730e7e1fe80185b18b7e94739666f54682df62b8ec6bb4154b99c47b5becda3291f2026-04-10T14:58:19+08:00wangbomengadd annotations
526f54682df62b8ec6bb4154b99c47b5becda3291f42a181674565411efa5c8b318bc77d6fd084df8a a2d5bf362d6c9a546f9deadf4f8975605a915d332026-04-10T09:07:20+08:00Yunzhe JiaMerge branch 'dev' of http://192.168.10.26:1000/yunzhe/llama.cpp into dev
5342a181674565411efa5c8b318bc77d6fd084df8abd3a57a11aa008e0b5dc2a30f82fcd3a025872082026-04-10T09:06:55+08:00Yunzhe Jiafix prompt_cache issue
54a2d5bf362d6c9a546f9deadf4f8975605a915d33bd3a57a11aa008e0b5dc2a30f82fcd3a025872082026-04-09T21:01:46+08:00Bomeng Wangserver: comment fixed time
55bd3a57a11aa008e0b5dc2a30f82fcd3a02587208fcfcdf80b01664a08fb12df879a05df3cc0eba70 74d437ab1456831573def7ca385a74d0ca460c372026-04-09T19:42:57+08:00Yunzhe JiaMerge branch 'dev' of http://192.168.10.26:1000/yunzhe/llama.cpp into dev
56fcfcdf80b01664a08fb12df879a05df3cc0eba700cd44929b442860dc0e5f88976108be82350f9dc2026-04-09T19:41:52+08:00Yunzhe Jiatmp fix one run >4K problem
5774d437ab1456831573def7ca385a74d0ca460c375813c943005d1ede551e67ee98e35aefd210ff22 0cd44929b442860dc0e5f88976108be82350f9dc2026-04-09T17:44:26+08:00wangbomengMerge branch 'dev' of http://192.168.10.26:1000/yunzhe/llama.cpp into dev
585813c943005d1ede551e67ee98e35aefd210ff221f0d5bb0453f93ffa51faf68237c89251c2e0c402026-04-09T17:38:57+08:00wangbomengadd --device-info
591f0d5bb0453f93ffa51faf68237c89251c2e0c40b8153b6b8f244b223c9f4c2940ae1f212634519e2026-04-09T17:38:37+08:00wangbomengadd --device-info
600cd44929b442860dc0e5f88976108be82350f9dcb8153b6b8f244b223c9f4c2940ae1f212634519e2026-04-09T15:08:50+08:00Bomeng Wangrm calculet0.txt
61b8153b6b8f244b223c9f4c2940ae1f212634519e82842cfc3fa3d5c004c4540aeff8a81f38509bee2026-04-09T15:07:02+08:00wangbomengserver: fix limit tokens to max_seq_len
6282842cfc3fa3d5c004c4540aeff8a81f38509bee622a22991089c6debd5f4b57738f3df3a0bda8b52026-04-09T15:06:26+08:00wangbomengserver: fix limit tokens to max_seq_len
63622a22991089c6debd5f4b57738f3df3a0bda8b55f47a738ba56a37d0ff281a16212f41979568e352026-04-09T10:36:31+08:00wangbomengserver: comment circle print
645f47a738ba56a37d0ff281a16212f41979568e35ecf01a17651cd7b1e8b85fd6075e83831463ee23 207ca38c751f7e6c251833eb11e89a2a0bb0fd072026-04-09T09:48:20+08:00Yunzhe JiaMerge branch 'dev' of http://192.168.10.26:1000/yunzhe/llama.cpp into dev
65ecf01a17651cd7b1e8b85fd6075e83831463ee2399d2078e89305035d87d966cee7987c7d50b39252026-04-09T09:47:55+08:00Yunzhe Jiafix n_parallel problem
66207ca38c751f7e6c251833eb11e89a2a0bb0fd0799d2078e89305035d87d966cee7987c7d50b39252026-04-08T19:14:37+08:00wangbomengdelet extra printf
6799d2078e89305035d87d966cee7987c7d50b39259626239f5a9e568c9975d67dd6745fef90279a872026-04-08T19:01:16+08:00Yunzhe Jiafix buffer resize problem
689626239f5a9e568c9975d67dd6745fef90279a87da724f3e2a8af2c2537e3edcff41e9415ac66fef2026-04-08T17:47:04+08:00wangbomengcalrt: fix MapBuf
69da724f3e2a8af2c2537e3edcff41e9415ac66fef18814f9bf1a423ebe8f32d6043c9cbc5b21b02182026-04-08T16:58:32+08:00Bomeng Wangcalrt: add MapBuf
7018814f9bf1a423ebe8f32d6043c9cbc5b21b0218bbee06d6d851e3f84f75b578047967d7e0e4a8dc2026-04-08T13:55:52+08:00wangbomengserver : reset inference time
71bbee06d6d851e3f84f75b578047967d7e0e4a8dce126c21ed7b793b648d63533ea7ee30885f5a82a2026-04-08T11:29:36+08:00wangbomengcalrt: add test time print
72e126c21ed7b793b648d63533ea7ee30885f5a82a5603e050af07173589f3b8850121263d86da01fd2026-04-01T10:07:12+08:00wangbomengsever: fix commit id
735603e050af07173589f3b8850121263d86da01fd1e8a16cd0e3ad3c9a0f8746852a0a38c6e52322e2026-04-01T05:57:55+08:00wangbomengcalrt: add t_incopy and t_outcopy
741e8a16cd0e3ad3c9a0f8746852a0a38c6e52322e893e52bda0fee99bbda1521ea733a760bc4201f12026-04-01T10:14:22+08:00wangbomengcalrt: fix slice and add infer/copy time
75893e52bda0fee99bbda1521ea733a760bc4201f1398ecf71f5f574037ce33f84dd1576fa164909ad2026-03-30T11:35:40+08:00Bomeng Wangdelete unnecessary commit id
76398ecf71f5f574037ce33f84dd1576fa164909ad6d5e83902bdd3142a8ff0d3db19f02a8619e95a02026-03-27T02:01:23+08:00wangbomengrm llama-server_old.cpp
776d5e83902bdd3142a8ff0d3db19f02a8619e95a0e4eaab700333490e80ebefc8d023f6c1adfcc3122026-03-27T00:57:11+08:00wangbomengserver: printf calrt commit ID
78e4eaab700333490e80ebefc8d023f6c1adfcc312aeacc23a883400db0ebffb50e23355ada054da662026-03-25T09:03:55+08:00wangbomengcalrt: prefil to prefill
79aeacc23a883400db0ebffb50e23355ada054da66cbaa7492663630132adeddccd9fec5e44ffa33362026-03-25T08:49:22+08:00wangbomengcalrt: add slice only on decode
80cbaa7492663630132adeddccd9fec5e44ffa33363c08ebf0e442cd730ddffd959ec676f9caf31c822026-03-25T03:21:02+08:00wangbomengreduce warning
813c08ebf0e442cd730ddffd959ec676f9caf31c82613bc2b9a9d7af92efbe42d876ffcdb244f7a00e2026-03-24T02:43:03+08:00wangbomengcalrt: comment LOG_ERROR in untile_decode_cpy
82613bc2b9a9d7af92efbe42d876ffcdb244f7a00e4e194ecc6c4b8ebe4c72d93f6055e3605dbfcefa2026-03-24T02:30:12+08:00wangbomengcalrt: fix max_seq_len judgment
834e194ecc6c4b8ebe4c72d93f6055e3605dbfcefa16cbf81a731f5839a2f6b0eb9d8539826353748b2026-03-24T01:59:03+08:00wangbomengcalrt: fix ubatch.embd cpy
8416cbf81a731f5839a2f6b0eb9d8539826353748b8b4e17d990c23025148c4abadefe1010a0dd07342026-03-23T09:38:21+08:00wangbomengcalrt: fix model_name of untile_cpy
858b4e17d990c23025148c4abadefe1010a0dd07346dacdd2cf9574214ca272a86674e7bd7ba21c4f0 41e6636ea1fb3884e41fb2023f92b0e7262ef09b2026-03-23T14:50:27+08:00wangbomengmerge origin dev
866dacdd2cf9574214ca272a86674e7bd7ba21c4f03e6ad64467771224f072e5bad90e270951ba4cdf2026-03-23T14:42:28+08:00wangbomengcalrt: add multimodal
873e6ad64467771224f072e5bad90e270951ba4cdfb37b7d9756feb3dac2db526eeab38b572a095d472026-03-23T14:36:41+08:00wangbomengcalrt: add multimodal
8841e6636ea1fb3884e41fb2023f92b0e7262ef09bd0a1b2c758440720d42fa108b3444f54593daa732026-03-12T16:36:07+08:00Bomeng Wangserver: correct the max_seq_len judgment
89d0a1b2c758440720d42fa108b3444f54593daa73b37b7d9756feb3dac2db526eeab38b572a095d472026-03-12T16:10:28+08:00Bomeng Wangcalrt: comment out dump
90b37b7d9756feb3dac2db526eeab38b572a095d4707817cefc14fa359dcecad0630defd3610f8f5c8 7ddafbdcb9426ae3af016925b8b596f11e8742d02026-03-10T16:32:31+08:00Yunzhe JiaMerge branch 'dev' of http://192.168.10.26:1000/yunzhe/llama.cpp into dev
9107817cefc14fa359dcecad0630defd3610f8f5c8b2c45d8ddefb7ec09435231c8ec5b736fabf9c912026-03-10T16:32:07+08:00Yunzhe Jiafix kv issue
927ddafbdcb9426ae3af016925b8b596f11e8742d0059009eeaf20a569abfbac5eb37dad3bb93764282026-03-10T07:13:04+08:00wangbomengcalrt: commented out cal_ctx->encode(batch)
93059009eeaf20a569abfbac5eb37dad3bb93764281b073661d1311b5300173993b9f31ee7241d86062026-03-10T07:04:22+08:00wangbomengcalrt: add multimodel
941b073661d1311b5300173993b9f31ee7241d86066295273c5866b9249bacdedbfc3a31877e6e6a85 b2c45d8ddefb7ec09435231c8ec5b736fabf9c912026-03-10T06:58:35+08:00wangbomengMerge branch 'dev' of http://192.168.20.229:1000/yunzhe/llama.cpp into dev
956295273c5866b9249bacdedbfc3a31877e6e6a85c6ea3f0d2468e3b8eab5de80bacee7fe4c9dd3d72026-03-10T06:54:41+08:00wangbomengarg: update hf_repo url to https://download.calculet.tech:9443
96b2c45d8ddefb7ec09435231c8ec5b736fabf9c91c6ea3f0d2468e3b8eab5de80bacee7fe4c9dd3d72026-03-09T15:03:26+08:00Bomeng Wangclart: past_kv_len = n_tokens
97c6ea3f0d2468e3b8eab5de80bacee7fe4c9dd3d7f8fba66b657c51a1df1efc7267bfea9f6140cebf2026-03-09T03:26:47+08:00wangbomengCMakeLists: STATIC to SHARED
98f8fba66b657c51a1df1efc7267bfea9f6140cebfdb4a61d2234c2f457b42ecc3f7219a1e1a35508a2026-03-06T15:06:36+08:00Yunzhe Jiafix calrt_install include path
99db4a61d2234c2f457b42ecc3f7219a1e1a35508ae672e02ca1a582f9e3a9d011d9e1b4fea4942a252026-03-06T06:59:58+08:00Yunzhe Jiaadapt to host-rt install structure
100e672e02ca1a582f9e3a9d011d9e1b4fea4942a2527b82f64252779ad160a3fa76bd378c39421de36 7a3a9a0e76bb5f530beafcfe52e87e388e93117a2026-03-06T11:47:09+08:00Yunzhe JiaMerge branch 'dev' of http://192.168.10.26:1000/yunzhe/llama.cpp into dev
10127b82f64252779ad160a3fa76bd378c39421de360bd95719fed2f0247c8d0199dd897d24157dfae62026-03-06T11:42:45+08:00Yunzhe Jiaadapt to calrt_objs target
1027a3a9a0e76bb5f530beafcfe52e87e388e93117a583c387efaf7bc030574e554088de79d09a27fbd2026-03-05T07:24:59+08:00wangbomengcalrt_infer: fix past_kv_len
103583c387efaf7bc030574e554088de79d09a27fbd0edc14a60a2d5596f71d728c655e3523da924d892026-02-27T02:24:56+08:00wangbomengcaserver: lim generated tokens to n_ctx_per_seq
1040edc14a60a2d5596f71d728c655e3523da924d89aa0a17d719326a63e0d6fea60aa0ced44e7abc332026-02-27T01:57:39+08:00wangbomengcaserver: limit generated tokens to n_ctx_per_seq
105aa0a17d719326a63e0d6fea60aa0ced44e7abc33fa332e773da681f3e77d6ffe83a87edb557f758b2026-02-27T01:40:11+08:00wangbomengcalrt: add bf16_to_fp32
106fa332e773da681f3e77d6ffe83a87edb557f758baa54a7c637f2c92d924a5f35386975b1c27de8982026-02-27T01:31:42+08:00wangbomengAdd bf16_to_fp32
107aa54a7c637f2c92d924a5f35386975b1c27de89884f2c0b5f334a7ab77347c6779b6027a951d0ff02026-02-27T01:30:36+08:00wangbomengAdd bf16_to_fp32
10884f2c0b5f334a7ab77347c6779b6027a951d0ff00bd95719fed2f0247c8d0199dd897d24157dfae62026-02-09T01:46:36+08:00wangbomengadd bf16_to_fp32 in untile_decode/prefill_cpy
1090bd95719fed2f0247c8d0199dd897d24157dfae6365eb0b719e30b0f9e348d854f11c9c3f954a96e2026-02-07T10:25:06+08:00Yunzhe Jiafix get_model_by_name
110365eb0b719e30b0f9e348d854f11c9c3f954a96e886cd1fb3b217efcf51c46209fcb6940223467972026-02-07T09:05:17+08:00Yunzhe Jiacomment out pos buffer file
111886cd1fb3b217efcf51c46209fcb6940223467979328f21378275f2354a19c3af907332b216ab4922026-02-06T17:36:54+08:00Yunzhe Jiafix compile problem
1129328f21378275f2354a19c3af907332b216ab49249fea70a20f550bfdb59d0ffe041b54919c224472026-01-28T15:32:36+08:00Yunzhe Jiaadd prefill-only mode
11349fea70a20f550bfdb59d0ffe041b54919c2244774a118df31b6a4acc8876b2a2db90d2395cc6ec32026-01-28T09:11:16+08:00Yunzhe Jiaadapt to new calbin test_case
11474a118df31b6a4acc8876b2a2db90d2395cc6ec398e8f9acfe87cd93646482cb2d942b8a132f08522026-01-13T14:32:47+08:00Yunzhe Jiasync with calrt update
11598e8f9acfe87cd93646482cb2d942b8a132f08528c8152a04036bb0d4756a4e08e54f514dae4e64d2025-12-10T00:20:57+08:00Yunzhe Jiaadd timestamp
1168c8152a04036bb0d4756a4e08e54f514dae4e64d5d2387931528fbe04aa8d66de935147efb65ca4d2025-12-09T14:40:04+08:00Yunzhe Jiaadapt to calrt csr
1175d2387931528fbe04aa8d66de935147efb65ca4de6285a748657add8d974b78ca920c5b41753ee122025-12-08T16:40:34+08:00Yunzhe Jiaadd timestamp for one decode process
118e6285a748657add8d974b78ca920c5b41753ee126b1ecdbbe3cc72ed8888be38f9e934cfd79b9fbb2025-11-27T16:58:35+08:00Yunzhe Jiaadapt to calrt
1196b1ecdbbe3cc72ed8888be38f9e934cfd79b9fbb240d9bb75241c91896afe125f2fc74698a7395f32025-11-25T17:06:14+08:00Yunzhe Jiaadd kv_tensor for pld test
120240d9bb75241c91896afe125f2fc74698a7395f37eeb241aad429a68fb8cfa3ba85d000fb2348cd22025-11-24T12:03:00+08:00Yunzhe Jiaadapt to calrt modification
1217eeb241aad429a68fb8cfa3ba85d000fb2348cd2dfb6250031a25058dbc3757e8a7ec75d90dcb48a2025-11-21T09:20:01+08:00Yunzhe Jiafix byte_size error
122dfb6250031a25058dbc3757e8a7ec75d90dcb48ae31bce25e85dc7fdf0a0e69fd0e58ba4a2ffaf3e2025-11-18T10:28:00+08:00Yunzhe Jiaadd elf in CMakelist
123e31bce25e85dc7fdf0a0e69fd0e58ba4a2ffaf3e6444e227c36df16199c40d8cead2244fe014f3772025-11-07T11:52:17+08:00Yunzhe Jiadump src tensor
1246444e227c36df16199c40d8cead2244fe014f377d81d7b190ff5f21325167936496dfe5ab48b922e2025-11-07T09:19:11+08:00Yunzhe Jiatest golden case
125d81d7b190ff5f21325167936496dfe5ab48b922ee492f5dc8cec17529f53ef2bbdf7b502107f8148 1f5accb8d0056e6099cd5b772b1cb787dd590a132025-11-04T14:43:32+08:00Yunzhe JiaMerge branch 'master' into dev
126e492f5dc8cec17529f53ef2bbdf7b502107f81482e05afd22b3d77c7013b11eee88e88b17188f21a2025-11-04T14:09:59+08:00Yunzhe Jiaredo untile logic
1271f5accb8d0056e6099cd5b772b1cb787dd590a132759ccdb4adc8568add4316780d5e675519b07752025-11-04T05:04:59ZNoahFix garbled output with REPACK at high thread counts (#16956)
1282759ccdb4adc8568add4316780d5e675519b0775c5023daf607c578d6344c628eb7da18ac3d92d322025-11-04T10:53:48+08:00Aman GuptaCUDA: avoid mul + bias fusion when doing fusion (#16935)
129c5023daf607c578d6344c628eb7da18ac3d92d32e7da30b584dc1f2ee0414c4a1298ce64eef97e8d2025-11-03T11:47:57-08:00lhezopencl: support imrope (#16914)
130e7da30b584dc1f2ee0414c4a1298ce64eef97e8ded8aa63320393512bdcfe4b05b5ae01ba91888e12025-11-03T18:53:26+01:00Aleksander Grygierfix: Viewing multiple PDF attachments (#16974)
131ed8aa63320393512bdcfe4b05b5ae01ba91888e148bd26501b08a3f0bff1249db47f313641f7bebb2025-11-03T18:01:59+01:00Daniel Beveniusmodel-conversion : pass config to from_pretrained (#16963)
13248bd26501b08a3f0bff1249db47f313641f7bebb622cd010ff4a65bef67edbf2f9bf4707c01f98f72025-11-03T15:38:23+02:00Georgi Gerganovserver : add props.model_alias (#16943)
133622cd010ff4a65bef67edbf2f9bf4707c01f98f7070ff4d5356083d60b807bb34d36b31c3653a29e2025-11-03T14:29:11+01:00theo77186ggml: CUDA: add head size 72 for flash-attn (#16962)
134070ff4d5356083d60b807bb34d36b31c3653a29ebf7b0c9725ed0c406c5debaea022ec72584306342025-11-03T11:11:18+01:00Xuan-Son Nguyenmtmd: add --image-min/max-tokens (#16921)
135bf7b0c9725ed0c406c5debaea022ec7258430634fcfce040e816c542f374ad51461b3561d73c4bc92025-11-03T10:25:55+01:00Xuan-Son Nguyenmtmd: pad mask for qwen2.5vl (#16954)
136fcfce040e816c542f374ad51461b3561d73c4bc9ee3a5a10adf9e83722d1914dddc56a0623ececaf2025-11-03T14:40:02+08:00Jinyang Heggml : LoongArch fixes (#16958)
137ee3a5a10adf9e83722d1914dddc56a0623ececaf7e994168b1ccc12337ba8de939c4fd466107c1fb2025-11-03T05:33:56ZOlivier Chafiksync: minja (glm 4.6 & minmax m2 templates) (#16949)
1387e994168b1ccc12337ba8de939c4fd466107c1fbbcfa87622ae46be6345a8e3dfdbdc5ba5414042b2025-11-03T03:35:33+02:00shani-fSYCL: optimized repeat_back kernel (3× fewer asm instructions, 2× faster)Feature/sycl repeat back opt (#16869)
139bcfa87622ae46be6345a8e3dfdbdc5ba5414042ba2054e3a8ff0da3978a4acc18c349ff58554d3362025-11-03T00:41:08+01:00Sascha Rogmannfeat(webui): improve LaTeX rendering with currency detection (#16508)
140a2054e3a8ff0da3978a4acc18c349ff58554d336dd5286805004db1f9ac3176a1cbbfe373bdda0f82025-11-03T04:40:30+05:30Shagun Beratest-backend-ops : fix segfault in moe-expert-reduce test in support mode and coverage (#16936)
141dd5286805004db1f9ac3176a1cbbfe373bdda0f86b9a52422bac0f50dd8f1f8386744fa3ce9783bf2025-11-02T23:11:21+01:00Sigbjørn Skjæretci : disable failing riscv cross build (#16952)
1426b9a52422bac0f50dd8f1f8386744fa3ce9783bf2f966b8ed87514e74bb96592217226cb6a6974dd2025-11-02T13:08:04-08:00Zhiyong Wangmodel: add Janus Pro for image understanding (#16906)
1432f966b8ed87514e74bb96592217226cb6a6974ddcd5e3b57541ecc52421130742f4d89acbcf77cd42025-11-02T22:21:48+02:00Georgi Gerganovclip : use FA (#16837)
144cd5e3b57541ecc52421130742f4d89acbcf77cd487c9efc3b297b8a498716b1db3d061842e6fc85b2025-11-02T18:14:04+02:00Georgi Gerganovserver : support unified cache across slots (#16736)
14587c9efc3b297b8a498716b1db3d061842e6fc85b76af40aaaad78c42faecd8016a88362c788b84b02025-11-02T08:56:28-06:00Aldehir Rojascommon : move gpt-oss reasoning processing to init params (#16937)
14676af40aaaad78c42faecd8016a88362c788b84b07db35a7958a943be1693879f42d166f1526139792025-11-02T10:28:37+01:00Adrian Lundbergdocs: remove llama_sampler_accept reference in sampling sample usage (#16920)
1477db35a7958a943be1693879f42d166f152613979a864132ba546b6385922226480ee4e392aaa065c2025-11-02T08:42:57+05:30mnehete32CUDA: add FLOOR, CEIL, ROUND, TRUNC unary ops (#16917)
148a864132ba546b6385922226480ee4e392aaa065cd38d9f0877a5872daa3c5f06fb9a86376bf15d502025-11-02T08:48:46+08:00Aaron Teodevops: fix failing s390x docker build (#16918)
149d38d9f0877a5872daa3c5f06fb9a86376bf15d507fd205a8e8832ead273f62c5fd81d2b8aa9105352025-11-02T08:48:23+08:00Aaron Teoggml: add s390x cpu-feats (#16774)
1507fd205a8e8832ead273f62c5fd81d2b8aa9105352f68ce7cfd20e9e7098514bf730e5389b7bba9082025-11-02T00:15:31+02:00Georgi Gerganovscripts : add script to bench models (#16894)
1512f68ce7cfd20e9e7098514bf730e5389b7bba908e4a71599e5846110159955dec0008eb4aa24222b2025-11-01T19:49:51+01:00Pascalwebui: auto-refresh /props on inference start to resync model metadata (#16784)
152e4a71599e5846110159955dec0008eb4aa24222bdd5e8cab512c7752392b6e51a6f118f348fb3f162025-11-01T17:14:54+01:00Pascalwebui: add HTML/JS preview support to MarkdownContent with sandboxed iframe (#16757)
153dd5e8cab512c7752392b6e51a6f118f348fb3f16cf659bbb8ef9eb048e5153a27cf787fd83c055602025-11-01T16:52:17+01:00Adrien Gallouëtvendor : update cpp-httplib to 0.27.0 (#16846)
154cf659bbb8ef9eb048e5153a27cf787fd83c05560d8b860a219c2415faac8cc0e50b48b4aa11e3b642025-11-01T15:51:36+01:00Xuan-Son Nguyenmtmd: refactor preprocessing + support max/min pixels (#16878)
155d8b860a219c2415faac8cc0e50b48b4aa11e3b641ae74882f8f6755e44dff8f23f3abdc5b53ab7c12025-11-01T15:35:57+01:00Aleksander GrygierAdd a setting to display message generation statistics (#16901)
1561ae74882f8f6755e44dff8f23f3abdc5b53ab7c1961660b8c395afb8903f61ace69396a158ea0cb42025-11-01T15:02:57+01:00Jaromír Hradílekwebui: recognize AsciiDoc files as valid text files (#16850)
157961660b8c395afb8903f61ace69396a158ea0cb474fef4129fa58f6277c243777a4894371479dbad2025-11-01T11:01:42+01:00Sigbjørn Skjæretcommon : allow --system-prompt-file for diffusion-cli (#16903)
15874fef4129fa58f6277c243777a4894371479dbad5d8bb900bc7daa84bfa7bb1d25ab7e32394919f32025-11-01T08:55:25+01:00Sigbjørn Skjæretcodeowners : update after refactor (#16905)
1595d8bb900bc7daa84bfa7bb1d25ab7e32394919f32e76e013600cb0d51ccf158571ca1d0502952a072025-11-01T00:52:14-05:00Jeff Bolzvulkan: Fix multi_add invalid descriptor usage (#16899)
1602e76e013600cb0d51ccf158571ca1d0502952a07d3dc9dd898be805c23a408cc36daed5b3bf292212025-11-01T00:45:28-05:00Jeff Bolzvulkan: fuse mul_mat+add and mul_mat_id+add_id (#16868)
161d3dc9dd898be805c23a408cc36daed5b3bf29221bea04522ff1a0d8559ccfd353aa018dcfbb608cc2025-11-01T06:13:26+01:00Oliver SimonsCUDA: Remove unneded bias/gate dims in fused mmvq (#16858)
162bea04522ff1a0d8559ccfd353aa018dcfbb608cc0de0a01576772032008a689afc4d7c80685074c42025-10-31T23:40:23+01:00Piotr Wilkin (ilintar)refactor : llama-model.cpp (#16252)
1630de0a01576772032008a689afc4d7c80685074c4e58d585604bbbbbc24f8149effe444621b5191c62025-10-31T21:20:47+01:00Piotr Wilkin (ilintar)model : Minimax M2 (#16831)
164e58d585604bbbbbc24f8149effe444621b5191c631c511a968348281e11d590446bb815048a1e9122025-10-31T21:20:07+01:00Giuseppe Scrivanomodel : add Granite Hybrid nano types (#16896)
16531c511a968348281e11d590446bb815048a1e9126d39015a744b47bb7643ea73f412e6d64677f3052025-10-31T15:57:19+01:00Johannes GäßlerCUDA: Volta tensor core support for MMF (#16843)
1666d39015a744b47bb7643ea73f412e6d64677f3054146d6a1a6228711a487a1e3e9ddd120f8d027d72025-10-31T16:25:50+02:00Georgi Gerganovsync : ggml
1674146d6a1a6228711a487a1e3e9ddd120f8d027d78da3c0e200a586f768ada6f38745acb01380174c2025-10-31T20:05:07+08:00Aman GuptaCUDA: add expert reduce kernel (#16857)
1688da3c0e200a586f768ada6f38745acb01380174cc22473b580807929fd9e3a3344a48e8cfbe6c88f2025-10-31T13:50:33+02:00Georgi Gerganovbatch : fix consistency checks for the input positions (#16890)
169c22473b580807929fd9e3a3344a48e8cfbe6c88f0f715b4e759acceccb9f437cfd2a988fff85514a2025-10-31T10:54:19+02:00Georgi Gerganovserver : don't print user inputs to console (#16871)
1700f715b4e759acceccb9f437cfd2a988fff85514ad2d931f173b8a736b08999436e9259aafddec7182025-10-31T09:51:26+01:00Daniel Beveniusserver : fix typos in server.cpp comments [no ci] (#16883)
171d2d931f173b8a736b08999436e9259aafddec7182976b0374d36609b0429dd6ce48807e2ad39a7c22025-10-31T02:34:47-05:00Jeff Bolzvulkan: disable spirv-opt for rope shaders (#16872)
1722976b0374d36609b0429dd6ce48807e2ad39a7c2d2a2673dd1c86a01ab010e18b13b8bb959968c482025-10-31T16:18:59+09:00Masato Nakasakavulkan: Fix crash when FP16 mul_mat accumulation is not supported (#16796)
173d2a2673dd1c86a01ab010e18b13b8bb959968c4813002a08960e51a76c4d696165b5d7638d2f2b992025-10-31T08:14:49+01:00Ruben Ortlamvulkan: fix shmem overrun in mmq id shader (#16873)
1742e05afd22b3d77c7013b11eee88e88b17188f21ac20c0a5c0c44179f5267a262546624854b1203d12025-10-31T14:06:06+08:00Yunzhe Jiaadd pcie support
17513002a08960e51a76c4d696165b5d7638d2f2b996eb208d17ea29bb60295d9a2b5e7122dfb8f4b552025-10-31T12:46:31+08:00l3utterflyggml-hexagon: respect input size when getting/setting tensor data (#16836)
1766eb208d17ea29bb60295d9a2b5e7122dfb8f4b559984cbb61d12491d604484fb38b678fa150640612025-10-31T00:34:27+01:00Sigbjørn Skjæretci : enable free-disk-space on cuda docker build (#16877)
1779984cbb61d12491d604484fb38b678fa15064061ce18efeaf1234bd08f25bc08f88ef95cf5d9e51f2025-10-30T16:00:20-07:00lhezopencl: fix boundary handling for mul_mm (#16875)
178ce18efeaf1234bd08f25bc08f88ef95cf5d9e51f16724b5b6836a2d4b8936a5824d2ff27c52b45172025-10-30T23:15:03+01:00RodriMoraconvert : update transformers requirements (#16866)
17916724b5b6836a2d4b8936a5824d2ff27c52b4517b52edd25586fabb70f0c21b274473b307cf144992025-10-30T14:22:23-04:00chansikparkserver : bump request URI max length to 32768 (#16862)
180b52edd25586fabb70f0c21b274473b307cf14499517b7170e1a4d733583c4b07c5b7a49acc05911c2025-10-30T18:42:57+02:00Georgi Gerganovserver : remove n_past (#16818)
181517b7170e1a4d733583c4b07c5b7a49acc05911c835e918d8428f5119927d7150bf5a26176dedda02025-10-30T09:06:13-07:00Max Krasnyanskycpu: introduce chunking for repack matmuls and enable matmul-id chunking on ARM64 (#16833)
182835e918d8428f5119927d7150bf5a26176dedda0d261223d24e97f2df50220e4a5b7f0adb69bba812025-10-30T21:17:31+05:30Shagun Beracommon: fix typo in cli help text (#16864)
183d261223d24e97f2df50220e4a5b7f0adb69bba81dcca0d3ab840ebe9b2ccd4719033d408eeb758d72025-10-30T23:19:14+08:00JJJYmmmmodel: add support for qwen3vl series (#16780)
184dcca0d3ab840ebe9b2ccd4719033d408eeb758d7bacddc049a00786df44e682262f6e298742bfbc32025-10-30T05:26:05-07:00Max Krasnyanskycpu: introduce chunking for flash attention (#16829)
185bacddc049a00786df44e682262f6e298742bfbc3229bf686287d18f82c44e89888cc662145ecfdb42025-10-30T07:18:50-04:00Tianyue-Zhaomodel: Add support for CogVLM model (#15002)
186229bf686287d18f82c44e89888cc662145ecfdb4d7395115baf395b75a73a17b0b796e746e468da92025-10-30T08:56:28+01:00Sigbjørn Skjæretcuda : fix argsort with 64k+ rows (#16849)
187d7395115baf395b75a73a17b0b796e746e468da9052df28b0e3ca0398e5928c61c7de402543178942025-10-30T14:30:58+08:00Jan Boonllama : use std::abs instead of abs (#16853)
188052df28b0e3ca0398e5928c61c7de402543178948b11deea4663f29d3e042ce1056ba643264cd5f12025-10-30T01:27:41-05:00Jeff Bolzvulkan: Handle argsort with a large number of rows (#16851)
1898b11deea4663f29d3e042ce1056ba643264cd5f1b9ce94017729465895402cbcfffb51fa926c15e32025-10-30T04:34:15+01:00Oliver SimonsHide latency of bias and gate-loading (#16847)
190b9ce94017729465895402cbcfffb51fa926c15e33464bdac37027c5e9661621fc75ffcef3c19c6ef2025-10-29T15:13:10-05:00Jeff Bolzvulkan: Fuse rope+set_rows (#16769)
1913464bdac37027c5e9661621fc75ffcef3c19c6efe3af5563bd049141e036b50f843196db33d23e972025-10-29T20:11:39+01:00Xuan-Son Nguyenllama: fix ASAN error with M-RoPE (#16848)
192e3af5563bd049141e036b50f843196db33d23e9710fcc41290e233788f5a4215314156e8e023eb922025-10-29T18:09:18+01:00Xuan-Son Nguyenllama: store mrope data in KV cell (#16825)
19310fcc41290e233788f5a4215314156e8e023eb92bcf5bda6f5df559565d11d7c8e8295c1159a85ec2025-10-29T08:44:29-05:00Jeff Bolzvulkan: Update topk_moe fusion to handle gpt's late softmax (#16656)
194bcf5bda6f5df559565d11d7c8e8295c1159a85ec3eb2be1ca5f37480aeb16102970d9e65f43347fe2025-10-29T14:39:03+01:00Ruben OrtlamVulkan MMQ Integer Dot Refactor and K-Quant support (#16536)
1953eb2be1ca5f37480aeb16102970d9e65f43347fee41bcce8f0b53032a1fed275cd253e931c041cf62025-10-29T06:29:12-07:00Max KrasnyanskyHexagon Op queue & dispatch optimizations (#16820)
196e41bcce8f0b53032a1fed275cd253e931c041cf6144a4ce824b6bd0e48d62009d10cae1daf5308db2025-10-29T21:11:53+08:00Aman GuptaCUDA: use fastdiv in set-rows (#16834)
197144a4ce824b6bd0e48d62009d10cae1daf5308dbf549b0007dbdd683215820f7229ce180a12b191d2025-10-29T14:09:50+01:00Sigbjørn Skjæretvendor : sync minja (#16500)
198f549b0007dbdd683215820f7229ce180a12b191d9a3ea685b937c0f0cbfda2e50004ea54bf1875122025-10-29T03:53:04-05:00Jeff Bolzvulkan: Call ggml_vk_buffer_write_2d from ggml_vk_buffer_copy (#16793)
1999a3ea685b937c0f0cbfda2e50004ea54bf187512338074c383c81366320d176d83b94b0a567ee0c22025-10-29T15:55:06+08:00Aman GuptaCUDA: Fix bug in topk-moe for gpt-oss (#16821)
200338074c383c81366320d176d83b94b0a567ee0c2851553ea6b24cb39fd5fd188b437d777cb411de82025-10-29T08:14:39+02:00YaelLogicsycl: add RMS_NORM_BACK operation support (#16808)
201851553ea6b24cb39fd5fd188b437d777cb411de885a7d8677bf2200981e52f744a21d5267964ffcf2025-10-28T21:10:28+02:00YaelGitAccountcuda: add SET operation support (#16804)
20285a7d8677bf2200981e52f744a21d5267964ffcfa8ca18b4b815a2abdbecb958ee5f4c542d69aac72025-10-28T20:19:44+02:00Georgi Gerganovmemory : remove KV cache size padding (#16812)
203a8ca18b4b815a2abdbecb958ee5f4c542d69aac78284efc35c909217ee1b9a06903245d808ac22832025-10-28T19:41:43+02:00Georgi Gerganovllama-bench : clarify benchmarked parts of the computation (#16823)
2048284efc35c909217ee1b9a06903245d808ac22831c1409e13169d0ced4b1b4d39fb5b268b75250912025-10-28T23:16:20+08:00l3utterflyinitialise buffer.device in ggml_hexagon_session (#16816)
2051c1409e13169d0ced4b1b4d39fb5b268b75250917a0e900e3615fa46c074a7fdf900b47d3c0a1c7e2025-10-28T03:51:41-07:00Sam Malayekembedding: add raw option for --embd-output-format (#16541)
2067a0e900e3615fa46c074a7fdf900b47d3c0a1c7e280d97be9660e7a5feaa28a6e7a299bc73dd83fc2025-10-28T11:23:54+01:00Johannes Gäßlerllama: consistent ctx <-> buf order for KV cache (#16746)
207280d97be9660e7a5feaa28a6e7a299bc73dd83fc3479efd112b3910d2d008ad08fe4cb48956059032025-10-28T03:37:52-05:00Aldehir Rojasgrammar : support array references in json schema (#16792)
2083479efd112b3910d2d008ad08fe4cb4895605903463bbf20bfbe0da10ac58984d4d62bed5a49362a2025-10-28T10:54:53+08:00Chenguang LiCANN: Improve device ID handling and aclnnArange checks (#16752)
209463bbf20bfbe0da10ac58984d4d62bed5a49362aad8d36beffd791db10c94eb9e964afb891e3ca552025-10-28T10:31:21+08:00Aman GuptaCUDA: add unused vars to mmvf and mmvq (#16807)
210ad8d36beffd791db10c94eb9e964afb891e3ca55c053e18a66dd95dc340aa61317877c2a41d4e3cf2025-10-28T03:50:33+02:00tamarPalsycl: add SSM_CONV operation support (#16800)
211c053e18a66dd95dc340aa61317877c2a41d4e3cfe1ab0848037c9f9bfe68b3e1cee9ee375e1018a32025-10-27T18:54:01-04:00Yuri Khrustalevchat: Add LFM2 tool handling (#16763)
212e1ab0848037c9f9bfe68b3e1cee9ee375e1018a35a4ff43e7dd049e35942bc3d12361dab2f1555442025-10-27T23:12:16+01:00Xuan-Son Nguyenmtmd : fix idefics3 preprocessing (#16806)
2135a4ff43e7dd049e35942bc3d12361dab2f15554410640e31aab0819f31c1e1f2d008b019ee7372322025-10-27T13:51:28-07:00Diego Devesallama : disable pipeline parallelism if compute buffer allocation fails (#16748)
21410640e31aab0819f31c1e1f2d008b019ee73723280d28f104c0c3e61c11d6af073642b246a9fc19c2025-10-27T21:50:22+01:00Aclyggml : fix interpolate with align-corners and ne=1 (#16700)
21580d28f104c0c3e61c11d6af073642b246a9fc19cc55d53acec864f64afa1ba92972203dce1bf88f52025-10-27T21:39:49+01:00Johannes GäßlerHIP: fix AMDGPU_TARGETS, update documentation (#16803)
216c55d53acec864f64afa1ba92972203dce1bf88f5945501f5ea4b8ca56b181ecb035e9ee3fb31f4322025-10-27T16:02:58+01:00Xuan-Son Nguyenmodel : add LightOnOCR-1B model (#16764)
217945501f5ea4b8ca56b181ecb035e9ee3fb31f43275cbdd3fce38ea12d50cd19e73a069aa5dbbd5fa2025-10-27T09:17:31+01:00Johannes Gäßlerllama: fix leaked buffers for mmap + split files (#16765)
21875cbdd3fce38ea12d50cd19e73a069aa5dbbd5fa2b9bd9bf4e759c05db629ec1c391dc8aeaa718872025-10-27T09:25:10+08:00Aman Guptatest-backend-ops: print failed tests at the end (#16785)
2192b9bd9bf4e759c05db629ec1c391dc8aeaa7188759fc1ec8e83b14354c1a3a8acf8c5c2cbf9af42f2025-10-27T03:20:24+02:00tamarPalsycl: add ROLL operation support (#16665)
22059fc1ec8e83b14354c1a3a8acf8c5c2cbf9af42f75d33b9302f84a5b89f82205d2bcd8def5a64e0a2025-10-27T03:19:50+02:00shani-fsycl: add REPEAT_BACK operation support (#16734)
22175d33b9302f84a5b89f82205d2bcd8def5a64e0a3470a5c891dcc94363e492a3760af92b6b07241c2025-10-27T09:06:16+08:00Aman GuptaCUDA: support for weight clamp in top-k norm (#16702)
2223470a5c891dcc94363e492a3760af92b6b07241cbd562fe4f7bd55625511d5f9d639c4fb1db1d4402025-10-26T23:19:03+01:00Aclyggml-alloc : make gallocr prefer chunks that allow memory reuse (#16788)
223bd562fe4f7bd55625511d5f9d639c4fb1db1d440bbac6a26b2bd7f7c1f0831cb1e7b52734c66673b2025-10-26T21:31:41+01:00Sigbjørn Skjæretcuda : use fast copy when src and dst are of different type and contiguous (#16789)
224bbac6a26b2bd7f7c1f0831cb1e7b52734c66673b73a48c9790d320476b3e5ef75bda09f2f8269e6e2025-10-27T02:13:31+08:00leejetggml: fix cuda kernel launch configuration for k_compute_batched_ptrs to support large batch (#16744)
22573a48c9790d320476b3e5ef75bda09f2f8269e6ef696428ce8e4d16c17acbffeaa7feac3b0fb90612025-10-26T17:21:23+01:00Sigbjørn Skjæretconvert : enable expert group selection for all models with it (#16691)
226f696428ce8e4d16c17acbffeaa7feac3b0fb90617cce4f8158f0c4c88d8dadd4c23d33938127b8972025-10-26T17:20:32+01:00Sigbjørn Skjæretgraph : add clamping to ffn_moe_weights_sum to avoid div-by-zero (#16655)
2277cce4f8158f0c4c88d8dadd4c23d33938127b8978d8862829cd770fc9c0c7a726ed162de824ff5ea2025-10-26T16:08:52+01:00Sigbjørn Skjæretmodel : set res->t_embd in SmallThinker models (#16782)
2288d8862829cd770fc9c0c7a726ed162de824ff5eaf77c13b91f4d25754b6a0b857f98a6bc922a0aa72025-10-26T14:01:20+02:00amirai21docs : add Jamba to Text-only models list (#16778)
229f77c13b91f4d25754b6a0b857f98a6bc922a0aa73cfa9c3f125763305b4226bc032f1954f08990dc2025-10-26T19:28:04+08:00Aman GuptaCUDA: General GEMV fusion (#16715)
2303cfa9c3f125763305b4226bc032f1954f08990dc5d195f17bc60eacc15cfb929f9403cf29ccdf4192025-10-26T06:37:38+02:00Gilad S.vulkan: deduplicate Microsoft Direct3D12 devices (#16689)
2315d195f17bc60eacc15cfb929f9403cf29ccdf419226f295f4dd92ad714533adc5497afed5fa88bb82025-10-25T20:41:36+02:00Galunidconvert : handle mmproj filename/path properly (#16760)
232226f295f4dd92ad714533adc5497afed5fa88bb8f90b4a8efe4466215c51155a5c22c1ae6207da232025-10-25T19:26:27+09:00Shunta Saitomodel : set res->t_embd in PLaMo2 models (#16766)
233f90b4a8efe4466215c51155a5c22c1ae6207da238423d019318b446640bb620e4ce80066d8530f052025-10-25T10:59:54+02:00Giuseppe Scrivanovulkan: delete dead code (#16732)
2348423d019318b446640bb620e4ce80066d8530f055cca2542ac3f3f86831d32bce744d08fc2b353b02025-10-25T00:04:12-05:00Jeff Bolzvulkan: Optimize SSM_SCAN (#16645)
2355cca2542ac3f3f86831d32bce744d08fc2b353b055945d2ef51b93821d4b6f4a9b994393344a90db2025-10-24T20:52:00-04:00compiladeconvert : avoid dequantizing mxfp4 for GPT-OSS (#16756)
23655945d2ef51b93821d4b6f4a9b994393344a90db0bcb40b48c6fc6f17ba9672625e526ab2574344b2025-10-25T03:39:37+08:00leejetggml: fix CUDA grid launch condition for large block_nums.y in binbcast (#16742)
2370bcb40b48c6fc6f17ba9672625e526ab2574344b69e9ff010309a1155d704cf9320bdb3aaf4160ca2025-10-24T20:46:19+08:00Aman GuptaCUDA: use CUB for arbitary size argsort (#16754)
23869e9ff010309a1155d704cf9320bdb3aaf4160ca5a91109a5d7dab5d7adc40bedb397ede99a705b12025-10-24T14:10:29+02:00Florian Badiewebui: support q URL parameter (#16728)
2395a91109a5d7dab5d7adc40bedb397ede99a705b1f8f071faddf32ea09f4234edb6e809b380a9ee262025-10-24T12:02:02+02:00Daniel Beveniusmodel-conversion : add trust_remote_code for orig model run [no ci] (#16751)
240f8f071faddf32ea09f4234edb6e809b380a9ee260bf47a1dbba4d36f2aff4e8c34b06210ba34e6882025-10-23T16:31:41-04:00compiladeconvert : handle pre-quantized models (#14810)
2410bf47a1dbba4d36f2aff4e8c34b06210ba34e688dd62dcfab97e420949519fd0eac9fca7bf97e6352025-10-23T21:30:17+02:00Johannes Gäßlerserver: add memory breakdown print (#16740)
242dd62dcfab97e420949519fd0eac9fca7bf97e635d0660f237a5c31771a3d6d1030ebe3e0c409ba922025-10-23T15:54:46+02:00Julien Denizeconvert : Make mistral-common dependency optional (#16738)
243d0660f237a5c31771a3d6d1030ebe3e0c409ba92fe6a9882acf5c02f96624ed8f80144100d7006cb2025-10-23T15:00:49+02:00Xuan-Son Nguyenmtmd-cli : allow using --jinja (#16718)
244fe6a9882acf5c02f96624ed8f80144100d7006cb061f0eff02dc9a82f7bd850db3bd70b8a0b5e87a2025-10-23T17:07:31+05:30Prajwal B MehendarkarManually link -lbsd to resolve flock symbol on AIX (#16610)
245061f0eff02dc9a82f7bd850db3bd70b8a0b5e87a8cf6b42d467d05fa7d9776d2bcc69974ecce69002025-10-23T19:14:06+08:00Aman Guptaggml-cuda: use passed ops instead of hardcoded ops (#16712)
2468cf6b42d467d05fa7d9776d2bcc69974ecce69009de9672adb0f4ca4e39483ac3ffed52b3f70a55d2025-10-23T11:32:24+02:00matteoserver : send partial stop string when <EOG> is reached (#15007)
2479de9672adb0f4ca4e39483ac3ffed52b3f70a55d63d2fc46e17a06be5b4b5823a5ada088317f1f0a2025-10-22T20:05:15-05:00Matthew Michelsycl: use async memory allocation to fix crashes during graph recording (#16644)
24863d2fc46e17a06be5b4b5823a5ada088317f1f0aa2e0088d9242bd9e57f8b852b05a6e47843b5a452025-10-22T13:47:09-07:00Max KrasnyanskyAdd experimental ggml-hexagon backend for the Hexagon NPU (#16547)
2499b9201f65a22c02cee8e300f58f480a58859122719a5a3edfd306516cc419679d69d6435943b68162025-10-22T16:58:23+02:00Pascalwebui: introduce OpenAI-compatible model selector in JSON payload (#16562)
25019a5a3edfd306516cc419679d69d6435943b6816d8eaa26e4d9228df3aa46a930db60c8eaab67c1b2025-10-22T05:14:14-05:00sirus20x6ggml : Leverage the existing GGML_F32_VEC helpers to vectorize ggml_vec_set_f32 for faster fills (#16522)
251d8eaa26e4d9228df3aa46a930db60c8eaab67c1b9285325ce0174631c3cd6121d56084adc4ef2d8f2025-10-22T12:01:22+02:00Aclytests : fix test-thread-safety when compiling with multiple backends (#16699)
2529285325ce0174631c3cd6121d56084adc4ef2d8f03792ad93609fc67e41041c6347d9aa14e5e0d742025-10-22T12:33:08+08:00Aman GuptaCUDA: fix bug in topk-moe softmax (#16711)
25303792ad93609fc67e41041c6347d9aa14e5e0d7451d1a8c997bd2629ef211a30208058ea87a309822025-10-21T22:40:38+08:00Aman GuptaCUDA: topk-moe: add optional parameter for gpt-oss (#16649)
25451d1a8c997bd2629ef211a30208058ea87a309824926419c4d74a1cf724e7163d937eb72f36e7b262025-10-21T15:27:53+02:00Johannes GäßlerCUDA: better error for FA kernel with 0 occupancy (#16643)
2554926419c4d74a1cf724e7163d937eb72f36e7b266ea37f57391d27736c35cd3c20c1f990b7952b742025-10-21T16:43:14+08:00Aman Guptaggml: add ggml_can_fuse_subgraph (#16662)
2566ea37f57391d27736c35cd3c20c1f990b7952b74fb349848f387f355450c3187556e71e6d32c145f2025-10-20T22:26:17-07:00lhezopencl: fix warnings and clean up profiling (#16688)
257fb349848f387f355450c3187556e71e6d32c145f6de8ed75196c7cd98c1f34bbf3a7452451ba8ac22025-10-20T22:16:08-05:00Jeff Bolzvulkan: Handle FA with all -inf mask values (#16447)
2586de8ed75196c7cd98c1f34bbf3a7452451ba8ac284bf3c677857279037adf67cdcfd89eaa4ca92812025-10-21T01:21:12+03:00YehuditEsycl : add PAD_REFLECT_D1 operator support (#16145)
25984bf3c677857279037adf67cdcfd89eaa4ca9281c9c1972e2c2cc6a771fcc145bfa138700179f9612025-10-20T21:38:20+02:00Sigbjørn Skjæretmodel : add BailingMoeV2 support (#16063)
260c9c1972e2c2cc6a771fcc145bfa138700179f961b617cfd2896edd592a36ebbc041817eb030a10052025-10-20T19:49:02+02:00Aleksander GrygierHandle legacy 'context' attachments (#16687)
261b617cfd2896edd592a36ebbc041817eb030a100579068501fac9a74cca7129a8e5a8281b410a853e2025-10-20T05:53:50-07:00Diego Devesaggml-alloc : fix leak when reusing a tensor with a larger size (#16679)
26279068501fac9a74cca7129a8e5a8281b410a853e0e4a0cf2fae667d3efcf52f2f52398779d986b1d2025-10-20T14:21:12+02:00Aleksander GrygierPrevent premature submission on IME input (#16673)
2630e4a0cf2fae667d3efcf52f2f52398779d986b1d13f2cfad4170c096c51a02c24a6a158cb47f14802025-10-20T13:29:14+02:00Aleksander GrygierImport/Export UX improvements (#16619)
26413f2cfad4170c096c51a02c24a6a158cb47f148006332e28672356b964d6dfc2ba4657e20581cd432025-10-20T12:41:13+02:00Aleksander GrygierEnable per-conversation loading states to allow having parallel conversations (#16327)
26506332e28672356b964d6dfc2ba4657e20581cd4372d53e6c4decee8b339e49aed8cc0e234b9639dc2025-10-20T16:27:09+08:00takuya kodamallama-batch: fix build fails with `-Werror=missing-braces` (#16614)
26672d53e6c4decee8b339e49aed8cc0e234b9639dc2330de7b847ca84eac766df372c604c26db727472025-10-20T10:20:04+02:00Ron Evansreadme: update bindings (#16651)
2672330de7b847ca84eac766df372c604c26db727477062dd8460685d6700ed7621e50a22c6f3400ca32025-10-20T11:08:32+03:00safranowithSYCL: Add support for FLOOR,CEIL,ROUND and TRUNC unary operators (#16613)
2687062dd8460685d6700ed7621e50a22c6f3400ca30398752dd450dfabdd1b9e289f6364c2600f6ab52025-10-20T15:44:21+08:00takuya kodamallama-context: only warn on pooling_type when user specified (#16674)
269c20c0a5c0c44179f5267a262546624854b1203d162425b3a7e8b169a3bbe6ce8c7324513a7fcee8c2025-10-20T14:41:08+08:00Yunzhe Jiakv: turn on common-prefix mtmd: fix output size bug
2700398752dd450dfabdd1b9e289f6364c2600f6ab54f73d0a95120687e2c527739f771330a5271259a2025-10-19T23:54:31+02:00Giuseppe Scrivanomodel : add Granite Hybrid types (#16635)
2714f73d0a95120687e2c527739f771330a5271259acec5edbcaec69bbf6d5851cabce4ac148be417012025-10-20T05:06:39+08:00Aaron Teoci : fix binaries release failure for s390x (binaries may not work yet) (#16664)
272cec5edbcaec69bbf6d5851cabce4ac148be41701fcb235b46618921cbd826acd49b553b5302233aa2025-10-19T14:03:25+02:00Sigbjørn Skjæretci : avoid manual updates of docs/ops.md (#16663)
273fcb235b46618921cbd826acd49b553b5302233aa55754bebd5d570960cde9c0ba991a0b2991f6b1e2025-10-19T18:37:47+08:00Aaron Teoci: include s390x release binaries (#16648)
27455754bebd5d570960cde9c0ba991a0b2991f6b1eee09828cb057460b369576410601a3a09279e23c2025-10-19T15:37:12+08:00Aman GuptaCODEOWNERS: update for ggml-cuda/mmf (#16660)
275ee09828cb057460b369576410601a3a09279e23ce56abd2098dd2e2b0804691b93c13b48ae4216272025-10-18T14:47:32+02:00Johannes GäßlerHIP: fix GPU_TARGETS (#16642)
276e56abd2098dd2e2b0804691b93c13b48ae42162738355c6c8e43204e11a22daa7483082c0ff01e712025-10-18T05:22:57-05:00Jeff Bolzvulkan: Implement topk_moe fused shader, ported from CUDA (#16641)
27738355c6c8e43204e11a22daa7483082c0ff01e7181387858f1fbcc1acedbd308486e1016618ca8f82025-10-18T17:52:53+08:00Aman GuptaCUDA: use registers instead of smem in topk-moe (#16647)
27881387858f1fbcc1acedbd308486e1016618ca8f866b0dbcb2d462e7b70ba5a69ee8c3899ac2efb1c2025-10-17T17:55:32-07:00Shawn Guopencl: transposed gemm/gemv moe kernel with mxfp4,f32 (#16602)
27966b0dbcb2d462e7b70ba5a69ee8c3899ac2efb1c41386cf365d894134ee0813d15e2f5d76f6a4d8e2025-10-17T17:41:09+02:00Johannes Gäßlerllama-model: fix insonsistent ctxs <-> bufs order (#16581)
28041386cf365d894134ee0813d15e2f5d76f6a4d8e3d4e86bbeb15f487d6da6174ba6191b7c212cc252025-10-17T18:02:52+03:00Radoslav Gerganovrpc : report actual free memory (#16616)
2813d4e86bbeb15f487d6da6174ba6191b7c212cc25342c728d031d50673feded797520a44127d733792025-10-17T14:23:47+02:00Giuseppe Scrivanovulkan: Add State Space Model (SSM) Operations Support (#16463)
282342c728d031d50673feded797520a44127d73379ababae7e1ec3e9cfdab0322ee55ea3389e82a4d52025-10-17T18:01:23+08:00muggle-stackggml : fix SpaceMit IME array out-of-bounds in task assignment (#16629)
283ababae7e1ec3e9cfdab0322ee55ea3389e82a4d5b19491599d4d42c606601d75e95c1d1de3291f8e2025-10-17T10:35:03+02:00Pascalwebui: reorganize settings layout (#16607)
284b19491599d4d42c606601d75e95c1d1de3291f8e9ad4f1931ee0f3b41d9355245ef744786aaae0aa2025-10-17T02:31:04-05:00Jeff Bolzvulkan: fix debug build (add_rms_len/data not found) (#16624)
2859ad4f1931ee0f3b41d9355245ef744786aaae0aa79967ec596c0dacfd2251b085a57e79df292b1cc2025-10-17T02:33:58-04:00Ilia Ilmermetal : add `CONV_TRANSPOSE_2D` (#16542)
28679967ec596c0dacfd2251b085a57e79df292b1ccceff6bb253dd306f5404d7ccb3f11fadafe71b522025-10-17T06:59:31+01:00Olivier Chafikgrammar : use int64_t to avoid int overflows in int schema to grammar conversion logic (#16626)
287ceff6bb253dd306f5404d7ccb3f11fadafe71b521bb4f43380944e94c9a86e305789ba103f5e62bd2025-10-17T05:36:40+03:00GittyBursteinSYCL SET operator optimized for F32 tensors (#16350)
2881bb4f43380944e94c9a86e305789ba103f5e62bd683fa6ba4ed3a23b939d3e11e6dc860bd47a0ccf2025-10-16T19:00:31+02:00Xuan-Son Nguyenmtmd : support home-cooked Mistral Small Omni (#14928)
289683fa6ba4ed3a23b939d3e11e6dc860bd47a0ccfb22572e97dc51757d3ebe917a5a283385010ec682025-10-16T16:28:41+02:00Pascalfix: added a normalization step for MathJax-style \[\] and \(\) delimiters (#16599)
290b22572e97dc51757d3ebe917a5a283385010ec687a50cf388a530127cbbdd2a507ef81a451c9d8192025-10-16T16:26:21+03:00GittyBursteinsycl : add ARANGE operator (#16362)
29162425b3a7e8b169a3bbe6ce8c7324513a7fcee8c76a63a2998286c1649de2496bccb7d8a465498952025-10-16T17:34:39+08:00Yunzhe Jiakv: let seq_rm free whole seq since current hw-op cannot support common-prefix
2927a50cf388a530127cbbdd2a507ef81a451c9d8196f5d924637a15abedb111cbbffd7da5f31c818552025-10-16T16:41:11+08:00Chenguang LiCANN: format code using .clang-format (#15863)
29376a63a2998286c1649de2496bccb7d8a46549895b67217078aa748b06e1b2269d88dd18c0aca2c262025-10-16T16:31:15+08:00Yunzhe Jiakv: fix common-prefix bug by implementing llama_memory_seq_rm
2946f5d924637a15abedb111cbbffd7da5f31c81855adc9b60f190c1016a09f439862fa1cbb302262ac2025-10-16T01:11:33-04:00takasurazeemcommon : Update the docs on -t --threads (#16236)
295adc9b60f190c1016a09f439862fa1cbb302262acee50ee1eadff58777ae746827b04de7ba0befc552025-10-16T13:10:32+08:00takuya kodamaggml-cpu: replace putenv with setenv for const-correctness (#16573)
296ee50ee1eadff58777ae746827b04de7ba0befc557adc79c03234de9a20661fd6dbf2d02c32ca7acb2025-10-16T07:21:28+03:00yael-worksSYCL: Add GGML_OP_MEAN operator support (#16009)
2977adc79c03234de9a20661fd6dbf2d02c32ca7acb466c1911ab736f0b7366127edee99f8ee56874172025-10-15T22:43:08+02:00Aleksei Nikiforovgguf-py : add support for endian conversion of BF16 data (#16594)
298466c1911ab736f0b7366127edee99f8ee56874170cb7a0683b0529172472d74d21f05470a607f2972025-10-15T22:24:51+03:00safranowithcpu : add FLOOR, CEIL, ROUND and TRUNC unary operators (#16083)
2990cb7a0683b0529172472d74d21f05470a607f297d93f8439b08c4f35e13a41a7366901fdbe770fc82025-10-15T10:51:04-07:00lhezopencl: add q8_0 mm support (#16469)
300d93f8439b08c4f35e13a41a7366901fdbe770fc8f9fb33f2630b4b4ba9081ce9c0c921f8cd8ba4eb2025-10-15T10:48:28-07:00lhezopencl: fix FA for f32 (#16584)
301f9fb33f2630b4b4ba9081ce9c0c921f8cd8ba4ebf4ce81c45e7bd910e36bf44c253fc5255c49b1e42025-10-15T16:22:20+02:00Aleksander GrygierAdd server-driven parameter defaults and syncing (#16515)
302f4ce81c45e7bd910e36bf44c253fc5255c49b1e417304cbcc1dd24de7741cbe57925d58e90a98ac12025-10-15T23:05:56+09:00Sam/Samuelmetal: optimise `GGML_OP_SUM` (#16559)
30317304cbcc1dd24de7741cbe57925d58e90a98ac13e3cb19f6449f9168a128eaeae01f8f41b049acc2025-10-15T16:53:12+03:00Georgi Gerganovserver : fix img token logs (#16595)
3043e3cb19f6449f9168a128eaeae01f8f41b049acc5acd455460f457942d8dd02e3dd9b1eebfce99fe2025-10-15T14:48:08+02:00Xuan-Son Nguyenllama-quant: add support for mmproj (#16592)
3055acd455460f457942d8dd02e3dd9b1eebfce99fe554fd578a5ed78a10f371f5850c6a69aa83df15a2025-10-15T13:54:15+02:00Julius TischbeinCUDA: Changing the CUDA scheduling strategy to spin (#16585)
306554fd578a5ed78a10f371f5850c6a69aa83df15afa882fd2b1bcb663de23af06fdc391489d05b0072025-10-15T12:51:27+03:00Georgi Gerganovserver : fix mtmd checkpoints (#16591)
307fa882fd2b1bcb663de23af06fdc391489d05b007ffa059034c1e41f3e58363ef3c46d2fcf854bc4d2025-10-14T20:33:05+03:00Georgi Gerganovmetal : avoid using Metal's gpuAddress property (#16576)
308ffa059034c1e41f3e58363ef3c46d2fcf854bc4d120bf7046d85a893f064c12abe58bfeebd735f842025-10-14T19:18:05+02:00SavicStefanvulkan: Add ACC_TYPE_VEC2 implementation (#16203)
309120bf7046d85a893f064c12abe58bfeebd735f844258e0cfe72c72ad2787be1e9f93253e892194552025-10-14T22:48:08+08:00Aman GuptaCUDA + openCL: fix bug in accessing rms_norm->src while doing fusion (#16577)
3104258e0cfe72c72ad2787be1e9f93253e892194557ea15bb64c81e3813eb0babf9a57e1bc5697f5692025-10-14T08:53:37-05:00Jeff Bolzvulkan: Support FA with K/V in F32 (#16543)
3117ea15bb64c81e3813eb0babf9a57e1bc5697f5699c7185dd28416cf67f5e3b268381f311b5e3da562025-10-14T07:51:36-05:00Jeff Bolzvulkan: Improve build time for MSVC (#16545)
3129c7185dd28416cf67f5e3b268381f311b5e3da561ee9d0b415cdf5240418c110a18b419f4002b1542025-10-14T14:22:47+02:00Johannes GäßlerCUDA: enable FA for FP32 KV cache (#16546)
3131ee9d0b415cdf5240418c110a18b419f4002b15448e2fa9fb7c2de1e53808fdb65ec33f916020fc42025-10-14T19:16:21+08:00Aman GuptaCUDA: use fastdiv + ggml_cuda_mad for mmvf (#16557)
31448e2fa9fb7c2de1e53808fdb65ec33f916020fc45b6913c47b6bc71a6f927805a45387d5657d8b892025-10-14T19:15:15+08:00Aman GuptaCUDA: add fp kernel for larger batch size MoE (#16512)
3155b6913c47b6bc71a6f927805a45387d5657d8b89bc07349a7f87ba6eb31ed4b0ea9d9a73521852132025-10-14T09:53:49ZAnav Prasadcuda : remove legacy copy-op pointer indirection code (#16485)
316bc07349a7f87ba6eb31ed4b0ea9d9a7352185213e60f241eacec42d3bd7c9edd37d236ebf35132a82025-10-14T08:48:50+03:00Georgi Gerganovserver : dynamic token limit for prompt cache (#16560)
317e60f241eacec42d3bd7c9edd37d236ebf35132a8e38b7c6e9e4453e3b3e96d76e38bc2ccb6bce4582025-10-13T23:07:57+03:00Georgi Gerganovmetal : FA support F32 K and V and head size = 32 (#16531)
318e38b7c6e9e4453e3b3e96d76e38bc2ccb6bce4585016b7286240d29f8f640039989b84ea3a8543442025-10-13T22:42:37+03:00Georgi Gerganovgraph : support cacheless embeddings with FA and iSWA (#16528)
3195016b7286240d29f8f640039989b84ea3a8543447049736b2dd9011bf819e298b844ebbc4b5afdc92025-10-13T11:50:37-07:00lhezopencl: fix build targeting CL 2 (#16554)
3207049736b2dd9011bf819e298b844ebbc4b5afdc901d2bdc2bc61b676706830305b286b08b9885a412025-10-13T16:29:45+02:00Johannes GäßlerCUDA: fix numerical issues in tile FA kernel (#16540)
32101d2bdc2bc61b676706830305b286b08b9885a4156fc38b9655fbe1869d8bd6cfb269418196cea692025-10-13T20:48:47+08:00Jie Fu (傅杰)ggml : fix build broken with -march=armv9-a on MacOS (#16520)
32256fc38b9655fbe1869d8bd6cfb269418196cea691fb9504eb744969a990bfe4cfcf1d3d7a479541c2025-10-13T17:01:24+08:00Chenguang LiCANN: fix CPU memory leak in CANN backend (#16549)
3231fb9504eb744969a990bfe4cfcf1d3d7a479541c3f750f8d760ab5a61491e6a9409072dfeee4b4d72025-10-13T10:55:32+02:00Pascalfix: add remark plugin to render raw HTML as literal text (#16505)
324b67217078aa748b06e1b2269d88dd18c0aca2c2698c8269abefdfa8fde47dfb2769ceebc8b704e862025-10-13T11:48:14+08:00Yunzhe Jiaadd different path for unified and seperate compilation
3253f750f8d760ab5a61491e6a9409072dfeee4b4d7c515fc577166042234241c6bd0da9b08dcbe2bb92025-10-13T16:25:02+08:00Sam/Samuelmetal: add support for opt_step_sgd (#16539)
326c515fc577166042234241c6bd0da9b08dcbe2bb9f9bc66c3ebcfddb5f09e4b21253623caeb8e414a2025-10-13T11:22:27+03:00Georgi Gerganovggml : fix scalar path for computing norm (#16558)
327f9bc66c3ebcfddb5f09e4b21253623caeb8e414aa31cf36ad946a13b3a646bf0dadf2a481e89f9442025-10-13T08:52:22+08:00hipuddingCANN: Update several operators to support FP16 data format (#16251)
328a31cf36ad946a13b3a646bf0dadf2a481e89f94481d54bbfd599811b354c39f04550888168be77802025-10-13T02:43:14+08:00Sam/Samuelmetal : add opt_step_adamw and op_sum (#16529)
32981d54bbfd599811b354c39f04550888168be7780c7be9febcbafa9af7d1b9443f86475c59c9c5f872025-10-12T18:06:41+02:00Pascalwebui: remove client-side context pre-check and rely on backend for limits (#16506)
330c7be9febcbafa9af7d1b9443f86475c59c9c5f878415f61e23d04427cd0d912fbb9d33b85f8494562025-10-12T21:53:35+08:00Neo Zhang Jianyu[SYCL] fix UT fault cases: count-equal, argsort, pad OPs (#16521)
3318415f61e23d04427cd0d912fbb9d33b85f8494562c301e91abb92d03c1a682b4b540ba835562a74b2025-10-12T15:48:03+02:00Mathieu Baudierci : add Vulkan on Ubuntu with default packages build (#16532)
3322c301e91abb92d03c1a682b4b540ba835562a74b4b2dae383df708e2afc49c4859a81cd074f5ac102025-10-12T08:18:47-05:00Aldehir Rojascommon : handle unicode during partial json parsing (#16526)
3334b2dae383df708e2afc49c4859a81cd074f5ac1041aac5c69b5fb281bc1f486afb053f78101bb39e2025-10-12T09:29:13+03:00Georgi Gerganovcommon : update presets (#16504)
33441aac5c69b5fb281bc1f486afb053f78101bb39ea2fba89a426ff8005d303c73f0436e7e67368b702025-10-12T00:25:37-05:00sirus20x6ggml : Fix FP16 ELU positive branch (#16519)
335a2fba89a426ff8005d303c73f0436e7e67368b7020cc625edc2264aae2779e71bef1593e6a4e8c432025-10-12T07:19:06+02:00Daniel Beveniushparams : add check for layer index in is_recurrent (#16511)
33620cc625edc2264aae2779e71bef1593e6a4e8c4311f0af5504252e453d57406a935480c909e3ff372025-10-12T00:15:00-05:00sirus20x6ggml: Correct SVE implementation in ggml_vec_dot_f16_unroll (#16518)
33711f0af5504252e453d57406a935480c909e3ff37a3cb04744fb5c591985f53b749fef5407d07a1452025-10-11T20:54:32+02:00Johannes GäßlerCUDA: faster tile FA, add oob checks, more HSs (#16492)
338a3cb04744fb5c591985f53b749fef5407d07a1454a8fbe0a5eb75f339782ebcf29c17848122184d32025-10-11T16:54:10+03:00Georgi Gerganovmetal : fix mul-mm condition + fix mul-mv permuted kernels (#16494)
3394a8fbe0a5eb75f339782ebcf29c17848122184d331d0ff1869aa2ea31f4e96d5877d0343e9a2171b2025-10-11T15:50:49+02:00Pascalfeat: render user content as markdown option (#16358)
34031d0ff1869aa2ea31f4e96d5877d0343e9a2171b97870e64975b26c5e06a3540a8dc0ff601351e862025-10-11T21:39:04+08:00Yann Folletserver / ranking : add sorting and management of top_n (#16403)
34197870e64975b26c5e06a3540a8dc0ff601351e86477a66b03501cf3bd067f8968b77ca4d053ff1bd2025-10-11T04:02:26-07:00Diego Devesacuda : avoid initializing unused devices (#16510)
342477a66b03501cf3bd067f8968b77ca4d053ff1bde60f01d941bc5b7fae62dd57fee4cec76ec0ea6e2025-10-11T11:33:41+03:00amirai21convert : correctly handle LLaMA tokenizer for Jamba (#16470)
343e60f01d941bc5b7fae62dd57fee4cec76ec0ea6e81086cd6a3ca1252f0dc0f938171648399179c532025-10-10T22:15:05+03:00Georgi Gerganovserver : fix division by zero when reporting stats (#16501)
34481086cd6a3ca1252f0dc0f938171648399179c5368ee98ae181a5c83a5cc6261daeee69a1f588c152025-10-10T17:17:31+03:00Georgi Gerganovvocab : mark EOT token for Granite models (#16499)
34568ee98ae181a5c83a5cc6261daeee69a1f588c15cdb6da468cc33323955a523738d2e1675aeb5e9a2025-10-10T17:11:07+03:00Radoslav Gerganovserver : return HTTP 400 if prompt exceeds context length (#16486)
346cdb6da468cc33323955a523738d2e1675aeb5e9a6d69ab3f262eca3f0c6ad0ce075d2c80b9924d0e2025-10-10T13:22:27+03:00Radoslav Gerganovserver : log requests to /v1/completions (#16495)
3476d69ab3f262eca3f0c6ad0ce075d2c80b9924d0e1faa13a1187051af66b0fd9f0d6effe4c77f0b3e2025-10-10T13:45:46+05:30Prajwal B Mehendarkarcmake : Dont define XOPENSOURCE on AIX (#16481)
3481faa13a1187051af66b0fd9f0d6effe4c77f0b3e1deee0f8d494981c32597dca8b5f8696d399b0f22025-10-09T22:54:57+02:00Pascalwebui: updated the chat service to only include max_tokens in the req… (#16489)
3491deee0f8d494981c32597dca8b5f8696d399b0f2d00cbea63c671cd85a57adaa50abf60b3b87d86f2025-10-09T22:11:15+03:00dudutacpu : optimize the ggml NORM operation (#15953)
350d00cbea63c671cd85a57adaa50abf60b3b87d86f8328fd4bae76fc027f8ca0e9a05acd3788dabe3f2025-10-09T18:54:51+03:00Georgi Gerganovserver : host-memory prompt caching (#16391)
3518328fd4bae76fc027f8ca0e9a05acd3788dabe3f56b4795842d852152222ca7a2d4304008facf1b92025-10-09T17:36:29+02:00PascalNo markdown in cot (#16483)
35256b4795842d852152222ca7a2d4304008facf1b92c0d875ae6c6043fbbafd2831e160955e9ca6af12025-10-09T14:35:22+02:00Daniel Beveniusmodel-conversion : add support for SentenceTransformers (#16387)
3532c0d875ae6c6043fbbafd2831e160955e9ca6af1aa4711d369b0d4adb6802b98ad6ea362f838710e2025-10-09T09:13:18+01:00sudhiarmci: add ARM64 Kleidiai build and test support (#16462)
354aa4711d369b0d4adb6802b98ad6ea362f838710ed80d6d2400b8faa80654c723cb5bdf9fc8f4db062025-10-09T15:50:25+08:00Chenguang LiCANN: Improve ACL graph matching (#16166)
355d80d6d2400b8faa80654c723cb5bdf9fc8f4db06b2602137557b2b28a39e03612717d85ead9a6f5a2025-10-09T09:29:17+02:00Charles Xukleidiai: kernel interface refactoring (#16460)
356b2602137557b2b28a39e03612717d85ead9a6f5ae08db4259521de493b7aeb49dadf29ebd1ee966a2025-10-09T15:25:11+08:00Neo Zhang Jianyu[SYCL] refactor soft_max, add soft_max_back (#16472)
357e08db4259521de493b7aeb49dadf29ebd1ee966a12bbc3fa50b6df03318a4451c9a2210200a0b28d2025-10-09T08:39:18+02:00Saba Fallahmodel: EmbeddingGemma Adding Support for SentenceTransformers Dense Modules (#16367)
35812bbc3fa50b6df03318a4451c9a2210200a0b28d9d0882840e6c3fb62965d03af0e22880ea90e0122025-10-08T22:18:41+02:00Pascalrefactor: centralize CoT parsing in backend for streaming mode (#16394)
3599d0882840e6c3fb62965d03af0e22880ea90e012d2ee056e1df0fdf646270fb5621e9f92084b59a72025-10-08T20:21:46+02:00ai-fonsiDisable CUDA host buffers on integrated GPUs (#16308)
360d2ee056e1df0fdf646270fb5621e9f92084b59a7b2c08c9ec4cd59ed88c1ebdd94f109af5ccc978e2025-10-08T17:20:18+09:00issixxserver : fix cancel pending task (#16467)
361b2c08c9ec4cd59ed88c1ebdd94f109af5ccc978e7fdd16b432d247121fe5fe7b21f8805f85266c852025-10-08T10:57:53+03:00Georgi Gerganovmetal : mark FA blocks (#16372)
3627fdd16b432d247121fe5fe7b21f8805f85266c8574b8fc17f92ada295a648e3c5eb28f46bca7d8922025-10-08T10:57:29+03:00Georgi Gerganovserver : improve context checkpoint logic (#16440)
36374b8fc17f92ada295a648e3c5eb28f46bca7d892aeaf8a36f06b5810f5ae4bbefe26edb33925cf5e2025-10-07T13:48:56-07:00Reese Levineggml webgpu: profiling, CI updates, reworking of command submission (#16452)
364aeaf8a36f06b5810f5ae4bbefe26edb33925cf5edf1b612e29ba97a2e67db339b1e8c7465702b7e82025-10-07T20:03:35+02:00Tarek Dakhranllama : support LiquidAI LFM2-MoE hybrid model (#16464)
365df1b612e29ba97a2e67db339b1e8c7465702b7e84e0388aa8a1f4ef1065701dc9c3947aea0a1b9a52025-10-07T15:57:14+03:00Georgi Gerganovserver : add `/v1/health` endpoint (#16461)
3664e0388aa8a1f4ef1065701dc9c3947aea0a1b9a5ef4c5b87ea2556ff8ca99cca3abdf48bdbca22f22025-10-07T11:11:08+02:00Sascha Rogmannwebui : added download action (#13552) (#16282)
367ef4c5b87ea2556ff8ca99cca3abdf48bdbca22f2c61ae20d05bd4fdd8551311325a28453364494262025-10-07T10:32:32+03:00Georgi Gerganovpresets : fix pooling param for embedding models (#16455)
368c61ae20d05bd4fdd8551311325a28453364494260123ff38f53d34752f29239a29d0e40a6dc4110f2025-10-07T09:59:13+03:00Radoslav Gerganovrpc : update documentation (#16441)
3690123ff38f53d34752f29239a29d0e40a6dc4110f0a319bb75ed29d968e2a9b544011b09ccb9329152025-10-07T08:24:17+03:00Georgi Gerganovmemory : use sequential equal splits for recurrent modules (#16442)
3700a319bb75ed29d968e2a9b544011b09ccb9329151d6092fc72f4d10f4486ac95edfd414bc08b62b82025-10-07T08:23:30+03:00Georgi Gerganovmetal : add support for non-padded FA KV (#16148)
3711d6092fc72f4d10f4486ac95edfd414bc08b62b88ae32dc9ecd7aeebf3a5b43557e0552a0a04cd4f2025-10-07T08:22:35+03:00Georgi Gerganovtests : add -INF blocks to the KQ mask in the FA tests (#16380)
3728ae32dc9ecd7aeebf3a5b43557e0552a0a04cd4f3df2244df40c67dfd6ad548b40ccc507a066af2b2025-10-07T08:21:40+03:00Georgi Gerganovmetal : various optimizations + refactoring (#16446)
3733df2244df40c67dfd6ad548b40ccc507a066af2bc08002a1988348403a5fd59b1fa3de3a10a6f92f2025-10-06T12:55:53-05:00Gadflyiillama : add --no-host to disable host buffers (#16310)
374c08002a1988348403a5fd59b1fa3de3a10a6f92f3a002afafa8e06555f11aed88b02d055d8a166f32025-10-06T10:59:40-06:00Gabe Goodhartchat : Granite Docling stopping (#16438)
3753a002afafa8e06555f11aed88b02d055d8a166f3a23b9bdbd3b64ce172f9962249f432d01aea74372025-10-06T17:40:21+02:00Sigbjørn Skjæretci : refactor sdk caching to minimize storage (#16414)
376a23b9bdbd3b64ce172f9962249f432d01aea743704e632a4aab8e6bfff0f8bc216b36ceb1e199ff92025-10-06T16:05:27+03:00Georgi Gerganovggml : fix unaligned access in AMX code (#16315)
37704e632a4aab8e6bfff0f8bc216b36ceb1e199ff9a80ff183abe4e5a76316257ffa597da41b3b6fa02025-10-06T14:56:59+02:00Daniel Beveniusci : remove missing reranker model files (#16444)
378a80ff183abe4e5a76316257ffa597da41b3b6fa01d49ca37594fb49db6aa9518ba7c512e5ccd01082025-10-06T14:17:12+02:00Daniel Beveniusggml-cpu : fix leftover handling in ggml_vec_scale_f32 for SVE (#16443)
3791d49ca37594fb49db6aa9518ba7c512e5ccd0108c5fef0fcea3b978a2318b1af170209ecec7c37b42025-10-06T09:29:56ZYuannannix : removed metal for nix (#16118)
380c5fef0fcea3b978a2318b1af170209ecec7c37b4ca71fb9b368e3db96e028f80c4c9df6b6b370edd2025-10-06T03:53:31-04:00Oleksandr Kuvshynovserver: update readme to mention n_past_max metric (#16436)
381ca71fb9b368e3db96e028f80c4c9df6b6b370edd35266573b968e1c947b367782fb4b3eddbb4f3c02025-10-05T06:57:47-06:00Gabe Goodhartmodel : Granite docling + Idefics3 preprocessing (SmolVLM) (#16206)
38235266573b968e1c947b367782fb4b3eddbb4f3c086df2c9ae4f2f1ee63d2558a9dc797b98524639b2025-10-04T20:59:31-07:00Reese Levineggml webgpu: actually add softmax, fix rms_norm offset (#16400)
38386df2c9ae4f2f1ee63d2558a9dc797b98524639bf39283960b58a92ecc0c72567711318b20e22b552025-10-04T20:04:27ZEvevulkan: use a more appropriate amount of threads when generating shaders (#16418)
384f39283960b58a92ecc0c72567711318b20e22b55898acba6816ad23b6a9491347d30e7570bffadfd2025-10-04T16:22:45+03:00Radoslav Gerganovrpc : check src buffer when copying tensor (#16421)
385898acba6816ad23b6a9491347d30e7570bffadfde29acf74fea996014380d59d31aa504ae89642582025-10-04T12:49:16+03:00Radoslav Gerganovrpc : add support for multiple devices (#16276)
386e29acf74fea996014380d59d31aa504ae8964258128d522c04286e019666bd6ee4d18e3fbf8772e22025-10-04T11:42:56+02:00Aclyvulkan : incremental shader builds (#16341)
387128d522c04286e019666bd6ee4d18e3fbf8772e2f6dcda390004b627ef30af378d0c01ad2519289e2025-10-03T20:51:48+02:00Pascalchat : support Magistral thinking (#16413)
388f6dcda390004b627ef30af378d0c01ad2519289e606a73f53175077429484b23dcf799f69a31d0bd2025-10-03T13:34:51-05:00ddh0server : context checkpointing for hybrid and recurrent models (#16382)
389606a73f53175077429484b23dcf799f69a31d0bd946f71ed9ade07e319859b5ce656144140e066fb2025-10-03T19:18:56+03:00Georgi Gerganovmetal : fix loop bound in ggml_mem_ranges (#16412)
390946f71ed9ade07e319859b5ce656144140e066fb638d330246954e88dffc22ce01fec15e6894e5442025-10-03T14:40:25+02:00Sigbjørn Skjæretllama : fix shapes for bert/mpt q/k norm (#16409)
391638d330246954e88dffc22ce01fec15e6894e54484c8e305e8010a1a3d43bdd0a25f737ac67809a42025-10-03T13:49:08+02:00Aclyggml : fix graph reallocation with multiple chunks (#16396)
39284c8e305e8010a1a3d43bdd0a25f737ac67809a42aaf0a2a2056d75d0dd53ab8a181473760e6ab222025-10-03T12:51:40+02:00Aleksander GrygierFix missing messages on sibling navigation (#16408)
3932aaf0a2a2056d75d0dd53ab8a181473760e6ab220e1f83855609d73beaf05d818640b6cfd39d287b2025-10-03T05:50:46-05:00Jeff Bolzvulkan: Replace uses of maxMemoryAllocationSize and VK_WHOLE_SIZE (#16354)
3940e1f83855609d73beaf05d818640b6cfd39d287bad126479c25cf983a0f994a08ba0911cf49ed62b2025-10-03T04:52:46-05:00Jeff Bolzvulkan: Fix FA coopmat1 invalid array indexing (#16365)
395ad126479c25cf983a0f994a08ba0911cf49ed62b77233277c912d495d1069543ca540c21a2f9e5b42025-10-03T11:45:16+02:00Daniel Beveniusci : change macos-13 to macos-15-intel (#16401)
39677233277c912d495d1069543ca540c21a2f9e5b4e308efda8e54e3f07578937a45a5d83624eb79702025-10-03T11:30:39+02:00Aleksander GrygierCapture model name only after first token (streaming) or completed request (#16405)
397e308efda8e54e3f07578937a45a5d83624eb7970136bda78c5679b266362b39c58338fff46fd25922025-10-03T03:33:08-05:00Jeff Bolzvulkan: in flash attention, bounds check against nem1 (don't rely on GGML_KQ_MASK_PAD) (#16316)
398136bda78c5679b266362b39c58338fff46fd25925113efd34ceda709292de26c72716c49e024fb322025-10-03T09:11:34+02:00Aleksander Grygierwebui : Fix messages payload sent to chat completions (#16402)
3995113efd34ceda709292de26c72716c49e024fb32d64c8104f090b27b1f99e8da5995ffcfa6b726e22025-10-03T08:01:31+02:00Pascalfix: track viewportHeight via window.innerHeight to avoid unwanted scrolling (#16356)
400d64c8104f090b27b1f99e8da5995ffcfa6b726e2ef07a4090672a3438d7f64f197795d7dc1c189572025-10-02T20:10:12+02:00Sigbjørn Skjærettest-barrier : do not use more threads than physically available (#16389)
401ef07a4090672a3438d7f64f197795d7dc1c1895734fcc5a4ace8c69476ef2ea3857f39a60334acc42025-10-02T11:00:31-07:00Reese Levineggml webgpu: add support for soft_max, optimize rms_norm (#16357)
40234fcc5a4ace8c69476ef2ea3857f39a60334acc491a2a5655658bb9ab77894716b82fae7ecb4b4d12025-10-02T19:43:22+02:00Piotr Wilkin (ilintar)model : Apertus model implementation (#15852)
40391a2a5655658bb9ab77894716b82fae7ecb4b4d172ee736c447be4ededb51ab97904b4d33c90c2612025-10-02T21:29:56+08:00R0CKSTARmusa: update compile flags (#16265)
40472ee736c447be4ededb51ab97904b4d33c90c261f09aefaa84d7f4d5df3f400f67944b94fef5b7952025-10-02T13:51:36+02:00Sigbjørn Skjæretci : fix ubuntu-latest-cmake-rpc (disable ccache) (#16388)
405f09aefaa84d7f4d5df3f400f67944b94fef5b795bbd32bc0384fbfcf07369617de58856b1e0e95a32025-10-02T08:10:07ZEveci: update vulkan ci (#16294)
406bbd32bc0384fbfcf07369617de58856b1e0e95a32be72c2b121ee99f33927149265ce6073ade9e592025-10-02T10:35:43+03:00Georgi Gerganovci : fix clean-up of old logs (#16381)
4072be72c2b121ee99f33927149265ce6073ade9e5995ce0985449c52fb9d6849b95563f7cf933ea0e32025-10-02T15:16:25+08:00Neo Zhang JianyuSYCL: Update to oneAPI 2025.2 (#16371)
40895ce0985449c52fb9d6849b95563f7cf933ea0e3c8dedc9999eccf7821a9fe5b29f10e8d075e22172025-10-02T05:52:59+02:00uvosHIP: add IMbackK to codeowner (#16375)
409c8dedc9999eccf7821a9fe5b29f10e8d075e2217e95fec640f43623911a2cd5bda8b19b1898c530c2025-10-01T23:32:39+02:00uvosCI: reenable cdna in rocm docker builds (#16376)
410e95fec640f43623911a2cd5bda8b19b1898c530cded67b94446ef4f7fd988dbde7a12deef9870c132025-10-01T23:09:25+02:00uvosHIP: Disable ROCWMMA fattn on CDNA when compiled against ROCWMMA 2.0.0 (#16221)
411ded67b94446ef4f7fd988dbde7a12deef9870c131fe4e38cc20af058ed320bd46cac9349911900562025-10-02T06:08:15+09:00Shunta Saitollama : parameter conversion and loading fixes for PLaMo2 variants (#16075)
4121fe4e38cc20af058ed320bd46cac9349911900564201deae9c2ae4732db8957b6ce0808d02ec597c2025-10-01T20:18:03+02:00uvosci: Properly install rocwmma for hip builds (#16305)
4134201deae9c2ae4732db8957b6ce0808d02ec597c764799279f801c696503bacca490b4358587d94c2025-10-01T19:22:18+02:00Adrien Gallouëtcommon: introduce http.h for httplib-based client (#16373)
414764799279f801c696503bacca490b4358587d94c2a9b63383a448ec18c754dfdc6e95cb853940a522025-10-01T18:18:10+02:00Aleksander GrygierConversation action dialogs as singletons from Chat Sidebar + apply conditional rendering for Actions Dropdown for Chat Conversation Items (#16369)
4152a9b63383a448ec18c754dfdc6e95cb853940a521104ca1a1c926b832274694a2773a17066982ad02025-10-01T15:54:42+02:00Aleksander GrygierImprove code block color theming (#16325)
4161104ca1a1c926b832274694a2773a17066982ad04f1575921cac9b489fe8f8bbf20aff985e7da45e2025-10-01T14:09:52+02:00Sigbjørn Skjæretci : use registry cache for docker builds (#16366)
417132d673554e65282e92505f4f77401ab971528bbaa9538a63aef35f0224b2502f13b19d650a8ce042025-10-01T07:56:36ZEvevulkan: make ggml_vk_default_dispatcher support older vulkan headers (#16345)
418aa9538a63aef35f0224b2502f13b19d650a8ce04e74c92e84236b2bab3f3c77bee4ead94928be3602025-10-01T07:40:26+02:00Aleksander Grygierwebui: Remove running `llama-server` within WebUI `dev.sh` script (#16363)
419e74c92e84236b2bab3f3c77bee4ead94928be360b2ba81dbe07b6dbea9c96b13346c66973dede32c2025-09-30T16:24:36-04:00Bartowskimodel : support GLM 4.6 (make a few NextN/MTP tensors not required) (#16359)
420b2ba81dbe07b6dbea9c96b13346c66973dede32cbf6f3b3a1965d70e07ca94aab7b01268fe483e962025-09-30T21:41:42+02:00Sigbjørn Skjæretci : fix ccache key for ubuntu-cpu-cmake (#16355)
421bf6f3b3a1965d70e07ca94aab7b01268fe483e967c156df4148eb524b22f7f363bfd2df00f264e0b2025-09-30T19:52:41+02:00Adrien Gallouëtcommon : disable progress bar without a tty (#16352)
4227c156df4148eb524b22f7f363bfd2df00f264e0b16b0ca0d2e63fc1a0a43795f8552f9cb61d9f7a52025-09-30T10:45:45-07:00lhezopencl: support pad_ext (#15888)
42316b0ca0d2e63fc1a0a43795f8552f9cb61d9f7a58d78cd2613ccdeb3cc86f59bc8f9ddd31cfbd3ed2025-09-30T19:18:54+02:00PascalChatapi ignore empty sampling (#16330)
4248d78cd2613ccdeb3cc86f59bc8f9ddd31cfbd3edd1c84a662daa91be975863913a59975b114581412025-09-30T09:57:51-07:00Reese Levineggml webgpu: support for rope,div,sub,glu,scale,cont operators (#16187)
425d1c84a662daa91be975863913a59975b11458141364a7a6d4a786e98947c8a90430ea581213c0ba92025-09-30T09:55:13-07:00lhezopencl: support ne3 in get_rows (#15866)
426364a7a6d4a786e98947c8a90430ea581213c0ba92df5bcf357dba0c49b4df1d684b2ffc9e88b70542025-09-30T16:39:44+02:00Adrien Gallouëtcommon : remove common_has_curl() (#16351)
4272df5bcf357dba0c49b4df1d684b2ffc9e88b7054075c01567bb7e93965ae60b7e119182c3e68f3e92025-09-30T15:38:01+02:00Sigbjørn Skjæretci : disable ccache for android (#16348)
428075c01567bb7e93965ae60b7e119182c3e68f3e9a014310374a16f9204f2bcc1b458fc1eda67e4692025-09-30T13:42:39+03:00Georgi Gerganovggml : bump version to 0.9.4 (ggml/1363)
42998c8269abefdfa8fde47dfb2769ceebc8b704e862790ecc304e384d9075c55d0628d8bf440025dc42025-09-30T17:42:57+08:00Yunzhe Jiaserver: adapt to calrt
4302790ecc304e384d9075c55d0628d8bf440025dc4f528e97d894b4d93934a0711a00253eea980e7992025-09-30T17:37:58+08:00Yunzhe Jiaadapt to calrt
431f528e97d894b4d93934a0711a00253eea980e799dcbe998a6d8f13b6c9cb36bbe7b355ced2ac38422025-09-11T14:13:35+08:00Yunzhe Jiaadd calrt mtmd support
432a014310374a16f9204f2bcc1b458fc1eda67e46935fb82497ec6c5904b0adb7e1c881a76c1c692db2025-09-30T08:13:22Zanavp-nvidiacuda : Enable CUDA Graph usage for Nemotron Nano v2 (NemotronH) (#16328)
43335fb82497ec6c5904b0adb7e1c881a76c1c692db3c62aed89fbe3e15f6f82035e805a0a1add263062025-09-30T11:03:23+03:00Georgi Gerganovmetal : dynamic simdgroups for MV kernels (#16340)
4343c62aed89fbe3e15f6f82035e805a0a1add26306f1eb1cb1eba042b2d583234e80f65e2e225d89952025-09-30T09:36:33+02:00Adrien Gallouëtcommon : simplify etag tracking by removing json (#16342)
435f1eb1cb1eba042b2d583234e80f65e2e225d8995de41f2b7bffab7582944b213ce12b605f8cf6e0f2025-09-30T09:07:20+02:00Charles Xukleidiai : fix work size and threads sync for fp16 (#16246)
436de41f2b7bffab7582944b213ce12b605f8cf6e0fa74a0d69f34f52fa10d4f0a7ce749fb3490d07742025-09-29T22:30:16-07:00lhezcodeowners: add codeowners for opencl backend (#16344)
437a74a0d69f34f52fa10d4f0a7ce749fb3490d07745f7e166cbf7b9ca928c7fad990098ef32358ac752025-09-29T19:26:34-05:00Jeff Bolztests: override test_set_rows::max_nmse_err to allow for occasional rounding differences (#16295)
4385f7e166cbf7b9ca928c7fad990098ef32358ac75d72f5f7ba260b546190338b0b76f2f152581424f2025-09-29T18:49:47+02:00PascalFix thinking blocks with quotes + add handling `[THINK]...[/THINK]` blocks (#16326)
439d72f5f7ba260b546190338b0b76f2f152581424fb77e6c18e1a6fac5705ed95f03af5436d67484c12025-09-29T17:51:48+03:00Georgi Gerganovci : add AMD runners and workflows (#16249)
440b77e6c18e1a6fac5705ed95f03af5436d67484c12ddd3f2356c313385fafc19a9f0ce678c8fe03ee2025-09-29T22:50:44+08:00alex-spacemitggml: riscv: add riscv spacemit backend (#15288)
4412ddd3f2356c313385fafc19a9f0ce678c8fe03ee4d3d455d3c8719eb8f206de328d1b9ba191efd7c2025-09-29T16:50:52+03:00Georgi Gerganovsync : ggml
4424d3d455d3c8719eb8f206de328d1b9ba191efd7cc9b1c06467aca8d30fafcab51292bcb285df5b0d2025-09-29T16:49:11+03:00Georgi Gerganovsync : whisper.cpp (ggml/1359)
443c9b1c06467aca8d30fafcab51292bcb285df5b0db6ae75afb49b23db3dfbc2b01e8aabfb047e68802025-09-26T17:34:42+02:00Daniel Beveniusggml : remove -dev suffix from release version (ggml/1355)
444b6ae75afb49b23db3dfbc2b01e8aabfb047e6880b6dff20e2fe352093039e2359370c6bb2b505e0b2025-09-25T14:39:05+02:00Daniel Beveniusggml : bump version to 0.9.3 (ggml/1353)
445b6dff20e2fe352093039e2359370c6bb2b505e0b2db78c75e4ef7497313fda6dbbb3fcaf9ca0752d2025-09-20T16:44:23+03:00Georgi Gerganovggml : prepare for development of 0.9.2-dev
4462db78c75e4ef7497313fda6dbbb3fcaf9ca0752d02463ab27b1379ff5e3d936ce8b3bfd356872ea62025-09-20T16:44:23+03:00Georgi Gerganovggml : bump version to 0.9.1
44702463ab27b1379ff5e3d936ce8b3bfd356872ea6adc76347d73b3d915e946efa5de8d0ad9f3904c22025-09-29T13:17:09+02:00Rafal Lewczukggml-backend : add root cause in error message if loading backend library fails (#16172)
448adc76347d73b3d915e946efa5de8d0ad9f3904c23a2bdcda0b31e51db954551e0cd49424133a84f32025-09-29T11:09:00+02:00Sigbjørn Skjæretggml : check cuda and metal argsort limits and add test (#16323)
4493a2bdcda0b31e51db954551e0cd49424133a84f366bb7985c3fcefda7a74aae9f8321f70eb1646c42025-09-29T10:37:20+02:00Aleksander GrygierImprove Mobile UI for dialogs and action dropdowns (#16222)
45066bb7985c3fcefda7a74aae9f8321f70eb1646c42f61c0f5bf8a620ca4c3872408803ab38cfb96132025-09-29T09:08:41+02:00Pascalfix: preserved zero values in chat settings inputs and textareas by switching to nullish coalescing for field values and default placeholders (#16312)
4512f61c0f5bf8a620ca4c3872408803ab38cfb96133ffd0fae473c954bb3e67526b31262048fb508d42025-09-29T12:33:12+05:30Vinkalllama-cli: prevent spurious assistant token (#16202)
4523ffd0fae473c954bb3e67526b31262048fb508d4a4a0aa5ea2a88e4858996199fefd5439f17b481c2025-09-29T01:30:45-05:00ddh0perplexity : show more kl-divergence data (#16321)
453a4a0aa5ea2a88e4858996199fefd5439f17b481c92cd103f627015a95f2107f1c27dc218c7b8ec642025-09-29T08:41:28+03:00Georgi Gerganovggml : fix dependencies for ggml_set_rows (#16318)
45492cd103f627015a95f2107f1c27dc218c7b8ec64b887d2f3413ac231e3cb5925260c39902af4a70c2025-09-28T23:50:37-05:00Jeff Bolzvulkan: Fix validation failure in quantized flash attention (#16292)
455b887d2f3413ac231e3cb5925260c39902af4a70cbd0af02fc96c2057726f33c0f0daf7bb8f3e462a2025-09-28T23:15:03+02:00Sigbjørn Skjæretggml : fix GGML_F32_VEC_FMA argument order in ggml_vec_mad1_f32 (#16307)
456bd0af02fc96c2057726f33c0f0daf7bb8f3e462ad9e0e7c8194dfd7d23bf3a86608c9ece68d77c932025-09-28T14:13:50-04:00crat0zcommon : fix reasoning before forced tool call via tool_choice = required (#16264)
457d9e0e7c8194dfd7d23bf3a86608c9ece68d77c930124ac989f7e7bf08803788f66dbe4106bdcdd582025-09-28T22:38:15+08:00R0CKSTARci : fix musa docker build (#16306)
4580124ac989f7e7bf08803788f66dbe4106bdcdd582811c65286ae954bec87049f75b86dc022006dcc2025-09-28T19:25:58+08:00Aaron Teodevops: switch to using ubuntu-22.04-s390x image (#16302)
4592811c65286ae954bec87049f75b86dc022006dccd8359f5fde480da030bf75c7711573c7c4d993ba2025-09-28T12:04:46+01:00Imad SaddikFixed a few typos in the README of the LLaMA.cpp HTTP Server [no ci] (#16297)
460d8359f5fde480da030bf75c7711573c7c4d993ba6a2c6145a0b91b40eb3c3dba7b20ccc4b270490f2025-09-28T01:38:37-05:00Jeff Bolzvulkan: 64-bit im2col (#16135)
4616a2c6145a0b91b40eb3c3dba7b20ccc4b270490f3b53634fe35771e2e318227aa81585726bae72342025-09-28T09:34:44+03:00Georgi Gerganovmetal : extend mat-mat multiplication support (#16225)
4623b53634fe35771e2e318227aa81585726bae72341384abf8b8d5894d32fada453ccf4d196ffba7de2025-09-28T09:34:05+03:00Georgi Gerganovmetal : fuse non-sequential nodes (#16102)
4631384abf8b8d5894d32fada453ccf4d196ffba7dee6d65fb02d553bd79cad94e517cdca18b687788d2025-09-27T20:36:34-05:00Jeff Bolzvulkan: handle mat_mul with A matrix > 4GB (#16176)
464e6d65fb02d553bd79cad94e517cdca18b687788d8656f5de688cddcaea1d6174535eb60ee23ef6a02025-09-27T16:43:39-04:00Jeff Bolzvulkan: support arbitrary KV dimension in flash attention (#16160)
4658656f5de688cddcaea1d6174535eb60ee23ef6a04807e8f96a61b2adccebd5e57444c94d18de72642025-09-27T22:41:03+02:00Aclyvulkan : make the vulkan.hpp dynamic dispatcher instance private (#16224)
4664807e8f96a61b2adccebd5e57444c94d18de7264c0bfc57af421f8fd63c946c13b7666aed82560e22025-09-27T19:56:40+02:00Aleksander GrygierShow message actions by default (#16289)
467c0bfc57af421f8fd63c946c13b7666aed82560e275a3a6c2cd0002ba40e2dcc92007bc9fdbc69f1a2025-09-28T00:49:32+08:00Aman GuptaCUDA: mul_mat_id for mmf for bs <= 64 for f16 and bs <= 32 for f32 (#16277)
46875a3a6c2cd0002ba40e2dcc92007bc9fdbc69f1a0499b29c6f64c705faaf5860dc4600fca23671f42025-09-27T18:45:07+02:00Johannes GäßlerCUDA: refactor and deduplicate vector FA kernels (#16208)
4690499b29c6f64c705faaf5860dc4600fca23671f4234e2ff8ed09716fb553437596779399bee31b112025-09-27T19:26:46+03:00Dmytro Minochkinvulkan: throw system error instead of SIGABRT during init on older devices (#16156)
470234e2ff8ed09716fb553437596779399bee31b113f81b4e91c1d5f098148af117e3f13cf4b077f522025-09-27T18:17:08+02:00Adrien Gallouëtserver : remove old LLAMA_SERVER_SSL (#16290)
4713f81b4e91c1d5f098148af117e3f13cf4b077f52ace6a54565444b6377bee8e7ac693238e77662792025-09-27T06:36:11-04:00Jeff Bolzvulkan: support GET_ROWS for k-quants (#16235)
472ace6a54565444b6377bee8e7ac693238e776627972b24d96c6888c609d562779a23787304ae4609c2025-09-27T11:12:46+02:00Adrien Gallouëtbuild : add LLAMA_OPENSSL option (#16287)
47372b24d96c6888c609d562779a23787304ae4609c624207e676ab5eb3ce7af631902bb45fb73a83592025-09-27T02:58:29+05:30Vinkalmodel : make minicpm embedding_scale, residual_scale and logit_scale optional with legacy defaults (#16273)
474624207e676ab5eb3ce7af631902bb45fb73a8359807e8c6d310952f2f5656afc63dab9d7083dcb5c2025-09-27T02:03:33+08:00Aaron Teodevops: add s390x & ppc64le CI (#15925)
475807e8c6d310952f2f5656afc63dab9d7083dcb5c1a189278944d030211f336e103e96b65f976c3612025-09-26T19:25:29+02:00Aleksander GrygierEnhance text file detection logic for file attachments (#16199)
4761a189278944d030211f336e103e96b65f976c361e0539eb6aed346d4b25a6ea019044e88771e76902025-09-26T18:35:42+02:00Aleksander GrygierAllow viewing conversations even when llama server is down (#16255)
477e0539eb6aed346d4b25a6ea019044e88771e76905d0a40f390732cbf85d8a3b7b0fc3cbebffe780a2025-09-26T11:36:48-04:00Isaac McFadyenwebui: switch to hash-based routing (alternative of #16079) (#16157)
4785d0a40f390732cbf85d8a3b7b0fc3cbebffe780ad12a9836597b1ae4440d64d5a6ed727d27ac07022025-09-26T15:59:07+02:00Aleksander GrygierAlways show message actions for mobile UI + improvements for user message sizing (#16076)
479d12a9836597b1ae4440d64d5a6ed727d27ac0702cc1cfa277b3aae1f6cc9180472072336597a78d42025-09-26T16:09:34+03:00Radoslav Gerganovcodeowners : add rgerganov as owner of RPC [no ci] (#16279)
480cc1cfa277b3aae1f6cc9180472072336597a78d454dbc37053f7d75ea5b0631d5ee88f19391e43142025-09-26T15:00:44+02:00Aleksei Nikiforovmtmd : fix uninitialized variable in bicubic_resize (#16275)
48154dbc37053f7d75ea5b0631d5ee88f19391e4314b995a10760cb93d23d617d76ecb82a5f95b5e0d32025-09-26T14:14:28+03:00Georgi Gerganovmetal : report OOM errors (#16274)
482b995a10760cb93d23d617d76ecb82a5f95b5e0d34710dd31bbcef79d04f85a3a6a8c7d9439c5c79a2025-09-26T13:12:19+02:00Adrien Gallouëtcommon : use cpp-httplib as a cURL alternative for downloads (#16185)
4834710dd31bbcef79d04f85a3a6a8c7d9439c5c79a9b26511857ac09ae69ab485168fe2d3ee5fb1d6e2025-09-26T12:39:35+02:00Adrien Gallouëtbuild : fix build-ios-device (#16257)
4849b26511857ac09ae69ab485168fe2d3ee5fb1d6e00217cd41388328c93e9e9644921c4319bb03bcc2025-09-26T18:27:25+08:00Aaron Teoggml-cpu: implement MXFP4 SIMD for s390x (#16193)
48500217cd41388328c93e9e9644921c4319bb03bcc3b337b01a1a85c2f5b49376e8c0bbdb3f521528e2025-09-26T13:19:23+03:00Radoslav Gerganovci : create git tags for released docker images (#16008)
4863b337b01a1a85c2f5b49376e8c0bbdb3f521528ea86a580a6692edd1da076d5422f944c344b4fe5e2025-09-26T07:53:36+02:00Daniel Beveniuscodeowners : add danbev as owner of build-xcframework.sh [no ci] (#16268)
487a86a580a6692edd1da076d5422f944c344b4fe5e0f7c69689f4e681948a6005adea9bee9c08ff9032025-09-26T08:56:38+08:00R0CKSTARmusa: upgrade musa sdk to 4.3.0 (#16240)
4880f7c69689f4e681948a6005adea9bee9c08ff903835b2b915c52bcabcd688d025eacff9a07b65f522025-09-26T08:56:10+08:00R0CKSTARmusa: fix build warnings (#15611)
489835b2b915c52bcabcd688d025eacff9a07b65f52b05a9d650f4da1e70e7e2cbe8fe44e61b39656db2025-09-25T19:50:28+02:00Sigbjørn Skjæretmodel : add GroveMoE support (#15510)
490b05a9d650f4da1e70e7e2cbe8fe44e61b39656db27052978e487957b3d39e3bd8fc8ee4aa304bff92025-09-25T23:38:10+08:00Aaron Teovendors: update miniaudio version (#16212)
49127052978e487957b3d39e3bd8fc8ee4aa304bff9077c94d0caf87fbd3cf3288dbb5c0fd9670294cf2025-09-25T18:20:34+03:00rtaluyevreadme : update bindings (#16144)
492077c94d0caf87fbd3cf3288dbb5c0fd9670294cfaa3ee0eb0b80efca126cedf9bcb4fb5864b46ce32025-09-25T22:35:05+08:00Aman GuptaCUDA: add a fused top-K MoE kernel (#16130)
493aa3ee0eb0b80efca126cedf9bcb4fb5864b46ce3d0991da39d3c39b3980c24cdfccb9a4c95a468702025-09-25T12:02:36+02:00Daniel Beveniusmodel-conversion : add embedding prompt file support (#15871)
494d0991da39d3c39b3980c24cdfccb9a4c95a46870aa719c2f886c445b78113bf1e5849f1fce4124a72025-09-25T11:36:47+02:00Daniel Beveniusserver : add support for external server for tests (#16243)
495aa719c2f886c445b78113bf1e5849f1fce4124a74cdd0bb4537f9617e9efcfef6b9454fcefe2ff082025-09-25T17:22:55+08:00junchao-zhaoggml : fix loongarch lsx compilation error (#15864)
4964cdd0bb4537f9617e9efcfef6b9454fcefe2ff08b5bd037832bcb8ed3086dfe26ce9090bea989af12025-09-25T11:12:27+02:00Johannes Gäßlerdocs: fix typo [no ci] (#16244)
497b5bd037832bcb8ed3086dfe26ce9090bea989af1dfcd53f7ecb9bb897a9d752d09c59d10be47237a2025-09-25T03:53:09-05:00Douglas Hanleyllama : add support for qwen3 reranker (#15824)
498dfcd53f7ecb9bb897a9d752d09c59d10be47237a4ea00794b8c995b6deaf4bac159c1778dc27419a2025-09-25T11:30:16+03:00Georgi Gerganovmetal : fuse NORM + MUL + ADD, support non-multiples of 4 (#16220)
4994ea00794b8c995b6deaf4bac159c1778dc27419a02a6a82ae7c7ddd1819ae26b0cc36675879aecee2025-09-25T11:29:42+03:00Georgi Gerganovmetal : relax reorder conditions (#16216)
50002a6a82ae7c7ddd1819ae26b0cc36675879aeceec498fc82fe5b83fc8c6e1627286bdc1f93caddbf2025-09-25T11:29:08+03:00Georgi Gerganovmetal : restore im2col perf (#16219)
501c498fc82fe5b83fc8c6e1627286bdc1f93caddbfe7a5130a20cf45a3358308356c249734ca9821432025-09-25T10:20:02+03:00Radoslav Gerganovrpc : use ggml logging facilities
502e7a5130a20cf45a3358308356c249734ca982143bee378e0988b44bbe93b0768208080951b3123632025-09-25T13:06:30+08:00Aaron Teocodeowners: add ownership of zdnn backend [no ci] (#16232)
503bee378e0988b44bbe93b0768208080951b3123635fb557653b8756ac2b6c6a102b1e44abcadf552f2025-09-25T05:06:06ZEveci: run the x64 and arm ci on the github machines instead (#16183)
5045fb557653b8756ac2b6c6a102b1e44abcadf552f4ae88d07d026e66b41e85afece74e88af54f4e662025-09-25T11:36:30+08:00Aaron Teodevops: fix s390x docker release failure (#16231)
5054ae88d07d026e66b41e85afece74e88af54f4e66e789095502b337690c7616db32d7c679a5bd25332025-09-25T00:25:04+08:00Aaron Teocodeowners: add ownership of zdnn backend [no ci] (#16229)
506e789095502b337690c7616db32d7c679a5bd2533f2a789e33490deb483a2694b066b37e45524bb792025-09-24T16:53:48+02:00Johannes Gäßlerllama: print memory breakdown on exit (#15860)
507f2a789e33490deb483a2694b066b37e45524bb793a599719673c850647e3bb911ed6d91109bb91d22025-09-24T16:17:49+02:00Aclyggml : split graph allocations according to backend max buffer size (#15815)
5083a599719673c850647e3bb911ed6d91109bb91d263b54c81a620981be020184ab99e63a8e50e47cb2025-09-24T13:42:26+02:00Tarek Dakhranmodel : add label for LiquidAI LFM2-2.6B model (#16204)
509152729f8848b692e1ae53c197ccca92ef10a523bc0c59c1157b5aeded285a3b25d2463c866f583d32025-09-24T08:53:47+02:00Uilian Riescommon : add missing chrono header for common.cpp (#16211)
510c0c59c1157b5aeded285a3b25d2463c866f583d37735706b939847df2c6c00b44c36f95a20137c612025-09-24T08:53:20+02:00Sigbjørn Skjæretcodeowners : match all requirements files (#16214)
5117735706b939847df2c6c00b44c36f95a20137c614d9ea03d170cf504a40bf3a85788af84cccaee802025-09-24T14:46:52+08:00Jie Fu (傅杰)model-conversion : run-org-model.py fails to run on mac m1 (#16213)
5124d9ea03d170cf504a40bf3a85788af84cccaee808ba548dae251f8f2e3834ed5226b76b6e4f4a4852025-09-24T08:10:09+02:00Daniel Beveniuscodeowners : use slash prefix for root files [no ci] (#16210)
5138ba548dae251f8f2e3834ed5226b76b6e4f4a485f505bd83ca7a43c4585ff3d59135e77eae9c793b2025-09-24T12:19:23+08:00Jie Fu (傅杰)model-conversion : fix the make targets in the README.md (#16209)
514f505bd83ca7a43c4585ff3d59135e77eae9c793b0889589dbe8c67dd518c58a57afbb02dde2dccbe2025-09-23T20:41:40+03:00Georgi Gerganovci : disable AMD workflows + update NVIDIA workflows (#16200)
5150889589dbe8c67dd518c58a57afbb02dde2dccbe4e29084ba4104c4ea529fd3163bb6e76f64383df2025-09-23T13:44:25+03:00Georgi Gerganovci : enable Vulkan workflow on Mac (#16194)
5164e29084ba4104c4ea529fd3163bb6e76f64383dff6b4af3d04763b1e0130f5b5fce19c4bc6f83f1c2025-09-23T01:58:12-07:00Xiangyan Sunggml-cpu: Respect cpumask settings (#16164)
517f6b4af3d04763b1e0130f5b5fce19c4bc6f83f1c264f1b51872c125e23fa0ac1da5e2a1170de9a082025-09-23T10:25:20+02:00Sigbjørn Skjæretggml : fix uninitialized is_on_grid in quantize_row_iq3_xxs_impl (#15928)
518264f1b51872c125e23fa0ac1da5e2a1170de9a080bc7cc715472fe28822e57f036f0746592ed2c042025-09-23T14:53:05+08:00Aaron Teozdnn: refactor codebase + add docs (#16178)
5190bc7cc715472fe28822e57f036f0746592ed2c044b9f4cb0f89a88de4bdf97727d0457b0c648804c2025-09-23T08:13:22+02:00Daniel Beveniuscodeowners : add @danbev to model-conversion example [no ci] (#16190)
5204b9f4cb0f89a88de4bdf97727d0457b0c648804c85e72271ba1ce78adf34fd8997803c991e617ca62025-09-23T13:59:34+08:00Aaron Teodevops: add s390x containers (#15915)
52185e72271ba1ce78adf34fd8997803c991e617ca61d0125bcf1cbd7195ad0faf826a20bc7cec7d3f42025-09-23T05:59:03+02:00Daniel Beveniusggml-cpu : fix typo in gemm comments [no ci] (#16189)
5221d0125bcf1cbd7195ad0faf826a20bc7cec7d3f4351f3da39c85f59d581fc184f09283da7f099a3b2025-09-22T12:40:10-06:00Gabe Goodhartfeat: Add conversion support in GraniteHybrid for non-hybrid (all attn) (#16177)
523351f3da39c85f59d581fc184f09283da7f099a3b3ecb2f671a2f49d56357f99d135a94e8417591782025-09-23T01:57:46+08:00Haiyue Wangclang-tidy : disable warning about performance enum size (#16127)
5243ecb2f671a2f49d56357f99d135a94e841759178432cf4304c5379bfbd1f3feed65ec205955b2f4b2025-09-22T19:13:00+02:00Sigbjørn Skjæretggml : implement set_rows with i32 index (#16159)
525432cf4304c5379bfbd1f3feed65ec205955b2f4b37a23c17bdc9c99c9c6ad41168e4ced3724b72cd2025-09-22T18:20:21+03:00Georgi Gerganovcodeowners : update + cleanup (#16174)
52637a23c17bdc9c99c9c6ad41168e4ced3724b72cd138c87ce8bd558b2cc134ada7316a3dad8eb67ac2025-09-22T14:13:51+02:00Adrien Gallouëtcommon : enable `--offline` mode without curl support (#16137)
527138c87ce8bd558b2cc134ada7316a3dad8eb67acc6db9a10278f40b4b8d3f6931728f2cce2356daa2025-09-22T10:53:13+02:00Quentin Bramaswebui : fix handling incomplete chunks (#16107)
528c6db9a10278f40b4b8d3f6931728f2cce2356daad05affbab7d1364fa7ac06c03cd33f03235ae8402025-09-22T10:49:58+02:00GideonSerfembedding : fix typos in README (#16171)
529d05affbab7d1364fa7ac06c03cd33f03235ae8404f324a556caa5be0be2261604ef4aab0faf36eb82025-09-22T16:48:42+08:00Haiyue Wangcommon : remove unused local variables (#16140)
5304f324a556caa5be0be2261604ef4aab0faf36eb8a71ae3ba7a857394103f12e30b387c48242c84f22025-09-22T11:12:37+03:00Georgi Gerganovggml : extend ggml_can_fuse to work with non-sequential nodes (#16123)
531a71ae3ba7a857394103f12e30b387c48242c84f205a2458121f9cdcc98ea6ffeeab6f230231362662025-09-22T11:12:09+03:00Georgi Gerganovggml : add ggml_op_is_empty (#16122)
53205a2458121f9cdcc98ea6ffeeab6f2302313626696fdca043be8fa733bbd59868a75517f926333762025-09-22T15:10:58+07:00Xuan-Son Nguyencodeowners : update ownership for @ngxson and @allozuar (#16128)
53396fdca043be8fa733bbd59868a75517f92633376b2d980fce063fa3591c59ad50b946c8da66c294f2025-09-22T17:04:01+09:00Shin-myoung-serpVulkan: add conv_transpose_2d operation (#16022)
534b2d980fce063fa3591c59ad50b946c8da66c294f5c6106a696af2b00ce01d14ba525979d5a32b1992025-09-22T09:59:05+02:00Sigbjørn Skjæretcodeowners : claim responsibility for ci, models, gguf-py and convert (#16124)
5355c6106a696af2b00ce01d14ba525979d5a32b199ec65fb52f0cc890e5085a6c5995ab08a156265fb2025-09-22T10:58:02+03:00Georgi Gerganovcontrib : update roles (#16113)
536ec65fb52f0cc890e5085a6c5995ab08a156265fb1d660d2fae42ea2e1d3569638e722bf7a37b6b192025-09-22T10:16:05+03:00Georgi Gerganovci : remove vulkaninfo calls (#16169)
5371d660d2fae42ea2e1d3569638e722bf7a37b6b19a20d810d79adfbf82e0eb703af6f27a0e2d1a5392025-09-22T09:11:39+03:00Georgi Gerganovci : use smaller model (#16168)
538a20d810d79adfbf82e0eb703af6f27a0e2d1a5394d0a7cbc617e384fc355077a304c883b5c7d4fb62025-09-22T00:37:17-05:00Jeff Bolzvulkan: add RTE variants of exp shader (#16165)
5394d0a7cbc617e384fc355077a304c883b5c7d4fb69073a73d82a916cea0809de225ef5175c3a86e912025-09-22T08:31:40+03:00Georgi Gerganovci : adjust params for less runtime (#16167)
5409073a73d82a916cea0809de225ef5175c3a86e9151f5a45fbe575dcd54bdd2a339ef8e8424d1c12a2025-09-22T07:22:43+02:00Ruben Ortlamvulkan: vec dot matrix multiplication fix (#16151)
54151f5a45fbe575dcd54bdd2a339ef8e8424d1c12ac4510dc9374e17dcb8726902ab5216067a92b3d32025-09-21T16:42:10-07:00lhezopencl: fix concat crash on win arm64 with Adreno (#15944)
542c4510dc9374e17dcb8726902ab5216067a92b3d3da30ab5f8696cabb2d4620cdc0aa41a298c54fd62025-09-21T14:48:44-07:00lhezopencl: initial `q8_0` mv support (#15732)
543da30ab5f8696cabb2d4620cdc0aa41a298c54fd628baac9c9f491c872e2c37762d3bd90446b005e92025-09-21T19:00:27+03:00Georgi Gerganovci : add label for the RISC-V runner (#16150)
54428baac9c9f491c872e2c37762d3bd90446b005e91eeb523c3e0c7ffbd59469f5463dcbdecba3535e2025-09-21T16:50:45+03:00Georgi Gerganovci : migrate ggml ci to self-hosted runners (#16116)
5451eeb523c3e0c7ffbd59469f5463dcbdecba3535e5bb4a3edec297e74b0f7bd4ed5d0fdd12e28d8582025-09-21T08:31:55+02:00Giuseppe Scrivanovulkan: optimize UMA buffer operations and fix driver hangs (#16059)
5465bb4a3edec297e74b0f7bd4ed5d0fdd12e28d8587f766929ca8e8e01dcceb1c526ee584f7e5e14082025-09-21T01:23:37-05:00Jeff Bolzvulkan: fix validation error about VK_PIPELINE_CREATE_CAPTURE_STATISTICS_BIT_KHR (#16086)
5477f766929ca8e8e01dcceb1c526ee584f7e5e1408405921dcefdd4e90ed948a4bf179007c2fa92b2d2025-09-20T12:55:47+03:00Georgi Gerganovsync : ggml
548405921dcefdd4e90ed948a4bf179007c2fa92b2dfa6383ca7e7ccb8ca3bdfeb37e348ddc4113aa262025-09-16T06:16:52+02:00Daniel Beveniusggml : introduce semantic versioning (ggml/1336)
549fa6383ca7e7ccb8ca3bdfeb37e348ddc4113aa26803dac2e48ef3ba26a504eb27c4e77ec2d21f7d02025-09-10T17:21:11+02:00Gregor JasnyCUDA : conditionally add cuda architectures (ggml/1341)
550803dac2e48ef3ba26a504eb27c4e77ec2d21f7d0459c0c2c1a400f960d7b8e8d94d31a8426f809862025-09-20T10:42:56+02:00Ruben Ortlamvulkan: use vec dot for matrix matrix multiplications (#16056)
551459c0c2c1a400f960d7b8e8d94d31a8426f80986be79d9fdd95ab8955527c4aaa67b90e8b95167182025-09-20T07:56:30+02:00Benniserver: fix SSE and OpenAI compatibility for error messages when streaming (#16109)
552be79d9fdd95ab8955527c4aaa67b90e8b9516718f432d8d83e7407073634c5e4fd81a3d23a10827f2025-09-19T15:15:21-07:00ssweensllama-bench: add --devices and --list-devices support (#16039)
553f432d8d83e7407073634c5e4fd81a3d23a10827f4067f07fc5aa5722b87db303b561597004696f6c2025-09-20T00:57:30+09:00shun095chat: Fix streaming parser for granite models (#15682)
5544067f07fc5aa5722b87db303b561597004696f6c4b8560ab56fdd9819358b47c338bbc8ec357c57e2025-09-19T09:52:27+02:00Aleksander Grygierfeat: Improve mobile UI for Settings Dialog (#16084)
5554b8560ab56fdd9819358b47c338bbc8ec357c57e0dd58b6877f3dc106593d6bc68d98305f553c5662025-09-19T13:02:51+07:00Xuan-Son Nguyenchat : fix build on arm64 (#16101)
5560dd58b6877f3dc106593d6bc68d98305f553c56669ffd891631befa9e6b485fd646a16dab4f2c0072025-09-19T11:31:56+07:00Xuan-Son Nguyenggml : refactor forward_dup for cpu backend (#16062)
55769ffd891631befa9e6b485fd646a16dab4f2c007246c0d9c795fb25da8268625d597580155a3672f2025-09-18T23:07:26+02:00Adrien Gallouëtggml-amx : fix ggml_amx_init() on generic Linux (#16049)
558246c0d9c795fb25da8268625d597580155a3672f3edd87cd055a45d885fa914d879d36d33ecfc3e12025-09-18T23:07:18+02:00Adrien Gallouëtcmake : fix static linking for OpenMP on Unix-like systems (#16031)
5593edd87cd055a45d885fa914d879d36d33ecfc3e1c0b45097c33e2667a94444f08cc9e36bec0a5e2e2025-09-18T12:03:34-07:00Shawn Guopencl: optimize mxfp4 kernels (#16037)
560c0b45097c33e2667a94444f08cc9e36bec0a5e2e38dbdf4c057515ccea9bec0ca2518f86d5e4d28e2025-09-18T13:46:17-05:00Jeff Bolzrename optimize_graph to graph_optimize (#16082)
56138dbdf4c057515ccea9bec0ca2518f86d5e4d28e368560a1e3b9a3bc83af741b0b2bc9e46fb420d22025-09-18T11:26:03-07:00Bowen HanCUDA: Optimize PAD_REFLECT_1D (#15957)
562368560a1e3b9a3bc83af741b0b2bc9e46fb420d24ca088b036313c2d8e682f4cfeb7c29edd85d0b92025-09-18T19:28:32+02:00Johannes GäßlerCUDA: fix compilation on CC 6.0 (#16091)
5634ca088b036313c2d8e682f4cfeb7c29edd85d0b9703f9e32c4eb3166f8d63007c26e31a1466c21af2025-09-18T16:22:50+01:00Eric CurtinAdd resumable downloads for llama-server model loading (#15963)
564703f9e32c4eb3166f8d63007c26e31a1466c21afad6bd9083bc346afd4ecd2fbb83e0306c0141c992025-09-18T16:28:41+03:00Georgi Gerganovmetal : use function constants for mul_mv_ext kernels (#16074)
565ad6bd9083bc346afd4ecd2fbb83e0306c0141c992b6b55a59f590a1e1eb2dcd09d5b8b6feb9cc7482025-09-18T13:28:22+02:00Sigbjørn Skjæretcuda : add missing F32<->I32 entries in ggml_cuda_cpy_fn (#16060)
5662b6b55a59f590a1e1eb2dcd09d5b8b6feb9cc748e58174cecbc45bf79bf653cd2c984395940c6ef42025-09-18T13:36:57+03:00Radoslav Gerganovserver : include usage statistics only when user request them (#16052)
567e58174cecbc45bf79bf653cd2c984395940c6ef4b213fce89bee8cb56b587b91e15a4278f8ed01802025-09-18T12:47:56+03:00Georgi Gerganovllama : bump max seq limit from 64 to 256 (#15916)
568b213fce89bee8cb56b587b91e15a4278f8ed0180e00f3fd8fff2cf5a8c8c9f475034bd089c8bcce42025-09-18T12:33:45+03:00Georgi Gerganovmetal : improve F32, F16 and BF16 mat-vec multiplication (#16057)
569e00f3fd8fff2cf5a8c8c9f475034bd089c8bcce4f2f28380ea68939c0481df3e4216b698824a59df2025-09-18T15:06:48+08:00Jhen-Jie Hongmetal : avoid call free for non-owned buffer (#16067)
570f2f28380ea68939c0481df3e4216b698824a59df62c3b645c56ad5288aa401901f043b050edac5a22025-09-18T10:03:24+03:00Georgi Gerganovmetal : handle nil cv during pipeline creation (#16065)
57162c3b645c56ad5288aa401901f043b050edac5a2d304f459d8619399eb232b1e3689379306f439d32025-09-18T09:26:33+08:00Chenguang LiCANN: Remove print (#16044)
572d304f459d8619399eb232b1e3689379306f439d30320ac5264279d74f8ee91bafa6c90e9ab9bbb912025-09-17T13:09:40-07:00Reese LevineGGML WebGPU: Support for ADD, MUL, RMS_NORM, GET_ROWS operators (#16018)
5730320ac5264279d74f8ee91bafa6c90e9ab9bbb91a7a98e0fffed794396b3fbad4dcdbbc1849636452025-09-17T20:38:12+03:00Georgi Gerganovmetal : refactor + optimize v2 (#15995)
574a7a98e0fffed794396b3fbad4dcdbbc1849636458f8f2274ee3601fecf6e2d57b52f701c81bede212025-09-17T19:29:13+02:00Aleksander GrygierSvelteKit-based WebUI (#14839)
5758f8f2274ee3601fecf6e2d57b52f701c81bede21c959b676be29e93f8dbc3bd6056ceba812a9eb722025-09-18T00:18:21+07:00Xuan-Son Nguyenconvert : add Llama4ForCausalLM (#16042)
576c959b676be29e93f8dbc3bd6056ceba812a9eb72cd08fc3ecc0264b4414b68af3874a6c689ed60c12025-09-17T15:32:42+02:00Johannes GäßlerCUDA: fix FA occupancy, optimize tile kernel (#15982)
577cd08fc3ecc0264b4414b68af3874a6c689ed60c1cb5bb6cc05119c24e7711ca2956cd0e6d409d3962025-09-17T01:08:02-07:00David Ribeiro Alvescommon : Fix corrupted memory error on json grammar initialization (#16038)
578cb5bb6cc05119c24e7711ca2956cd0e6d409d396a91d035b901e8a9edf810f63d130ee49adf27be22025-09-17T07:35:37ZEvevulkan: automatically remove unsupported devices (#15976)
579a91d035b901e8a9edf810f63d130ee49adf27be2745cbcf2fe1eb88f8db615ac622f0b944d924ad62025-09-17T09:34:09+02:00Daniel Beveniusci : revert back to macos-13 for macOS-latest-cmake-x64 (#16040)
580745cbcf2fe1eb88f8db615ac622f0b944d924ad61cbd80f8cf80a817715b1ccc5680fe2a3c5172c82025-09-17T15:30:55+08:00Jie Fu (傅杰)llama-quant : fix the verification of attention layers for encoder-decoder models (#16023)
5811cbd80f8cf80a817715b1ccc5680fe2a3c5172c885286f354813056f6c835046c0acfa3bf6ba94322025-09-17T15:29:00+08:00Jie Fu (傅杰)examples : support encoder-decoder models in the simple example (#16002)
58285286f354813056f6c835046c0acfa3bf6ba9432d5fabe3682de515fd09d6c981f7a0d1b756144552025-09-17T00:01:58-07:00Shane Amodel : add OLMo3 support (#16015)
583d5fabe3682de515fd09d6c981f7a0d1b756144558ff206097c2bf3ca1c7aa95f9d6db779fc7bdd682025-09-17T14:33:08+08:00Chenguang LiCANN: Optimize ggml_cann_set_device (#15935)
5848ff206097c2bf3ca1c7aa95f9d6db779fc7bdd6877475530b8bbea3bf578632507e1284cdfe2c8c02025-09-16T16:17:08+02:00jacekpoplawskillama-bench: add --n-cpu-moe support (#15952)
58577475530b8bbea3bf578632507e1284cdfe2c8c03913f8730ec6d6245480affc30ae3049107956f42025-09-16T15:27:52+02:00Daniel Beveniusci : use macos-latest for arm64 webgpu build (#16029)
5863913f8730ec6d6245480affc30ae3049107956f476888d202ed2b835ae19ea9f9db6baf39e4192972025-09-16T15:25:57+02:00Daniel Beveniusggml : fix padding in timestep embedding kernels (#15932)
58776888d202ed2b835ae19ea9f9db6baf39e419297f1fbffb5c0b34b2a68febb7da3fd0f8333f1ed4c2025-09-16T13:41:38+02:00Daniel Beveniusci : upload xcframework artifact from ios-xcode-build job (#16010)
588f1fbffb5c0b34b2a68febb7da3fd0f8333f1ed4c51abc96bdc52ba8cd6ad78dcf12ed9a041d7b4422025-09-15T23:59:19-07:00Bowen Hanfix: apply clang-format to CUDA macros (#16017)
58951abc96bdc52ba8cd6ad78dcf12ed9a041d7b44207808ebb07e2b1aa19032705e332679ddf9676142025-09-16T05:57:16+02:00Daniel Beveniusci : update macos-latest* jobs to use macos-latest (#15938)
59007808ebb07e2b1aa19032705e332679ddf9676146d758839ff741d4966ca92b7f801b7a8b5b963642025-09-15T22:54:44-04:00Yuri Khrustalevcmake : Do not install tools on iOS targets (#15903)
5916d758839ff741d4966ca92b7f801b7a8b5b963643d4053f77f0f78ee2b791088c02af653ebee42dd2025-09-16T10:38:28+08:00Aman GuptaAdd LLaDA-7b-MoE diffusion model (#16003)
5923d4053f77f0f78ee2b791088c02af653ebee42dddc381aa9a6dc45f00673471d34b8bddd30e775702025-09-15T16:28:31-06:00Jake KarnesCUDA: fix im2col_3d to respect non-contiguous inputs (views) (#15956)
593dc381aa9a6dc45f00673471d34b8bddd30e7757010d197409bd9537ff302ad09966fe406882fef9d2025-09-15T14:38:52-07:00Diego Devesadocker : enable rocWMMA in ROCm images, add gfx1151 (#15997)
59410d197409bd9537ff302ad09966fe406882fef9db907255f4bd169b0dc7dca9553b4c54af51708652025-09-15T14:38:42-07:00Diego Devesareleases : switch to rocWMMA develop branch, add gfx1151 (#15992)
595b907255f4bd169b0dc7dca9553b4c54af517086528c39da7c645185ade5436767929d7ec330060332025-09-15T19:51:35+03:00yael-worksSYCL: Add COUNT_EQUAL operator support (#15991)
59628c39da7c645185ade5436767929d7ec33006033106220562aca42b6738b8f51acfce0db1b8a2fb62025-09-15T13:08:30+03:00Nikolay Popovllama-run: Fix model download on Windows (#15988)
597106220562aca42b6738b8f51acfce0db1b8a2fb6a68f31edd71cc39141113f05f7133a3e9ece8c612025-09-15T17:35:11+08:00Aman GuptaCUDA: some micro-optimizations in mmf.cuh for mul_mat_id (#15926)
598a68f31edd71cc39141113f05f7133a3e9ece8c61b8e09f08b9a91c0401bc67d17a17c907564203462025-09-15T02:54:57-05:00ddh0fix KLD percentile output (#15999)
599b8e09f08b9a91c0401bc67d17a17c907564203466c019cb04e86e2dacfe62ce7666c64e9717dde1f2025-09-14T23:00:59+02:00Sigbjørn Skjæretmodel : add grok-2 support (#15539)
6006c019cb04e86e2dacfe62ce7666c64e9717dde1f9dcd200d57bc6f05a59a6a8df361d5d183af41242025-09-14T21:17:04+02:00Sigbjørn Skjæretserver : only attempt to enable thinking if using jinja (#15967)
6019dcd200d57bc6f05a59a6a8df361d5d183af41240fa154e3502e940df914f03b41475a2b80b985b02025-09-14T22:02:32+03:00Georgi Gerganovmetal : remove memory pools (#15966)
6020fa154e3502e940df914f03b41475a2b80b985b0261e6a20ffdb79c4875e674b4f6b514bc73cff8f2025-09-15T04:43:54+10:00Adamrocm.Dockerfile: added gfx1200,gfx1201 architectures to support AMD Radeon RX 9000 series (#15994)
603261e6a20ffdb79c4875e674b4f6b514bc73cff8fa0e13dcbe5bae7025660349ef3e4ead060e507f22025-09-14T16:56:28+02:00Ruben OrtlamVulkan: Clean up mul_mm shader (#15987)
604a0e13dcbe5bae7025660349ef3e4ead060e507f2a14bd350141fb42b8bf2dd2342cebc27bfdce3992025-09-14T22:20:35+08:00lcybuild: fix the build failures of Windows HIP release job (#15984)
605a14bd350141fb42b8bf2dd2342cebc27bfdce399918b26f197f55d5d562446dfc876d0e637929d072025-09-14T15:33:22+03:00Georgi Gerganovmetal : fix kernel requirements (#15983)
606918b26f197f55d5d562446dfc876d0e637929d079ecb88434644c865232bb665d2f6f05049fc64562025-09-14T12:28:18+03:00Radoslav Gerganovrpc : fix regression when --device is used (#15981)
6079ecb88434644c865232bb665d2f6f05049fc6456d1c6f11f47bae063a68a6be8e4830c060a11b7bd2025-09-14T02:21:59-07:00Diego Devesareleases : update ROCM, add gfx1200, gfx1201, gfx1151 (#15972)
608d1c6f11f47bae063a68a6be8e4830c060a11b7bd6380d6a3e709cb02a8695afdd96b40e6744773322025-09-14T12:10:07+03:00Radoslav Gerganovdoc : update documentation for --tensor-split (#15980)
6096380d6a3e709cb02a8695afdd96b40e674477332aa0c461efe3603639af1a1defed2438d9c16ca0f2025-09-14T13:37:03+08:00Aaron Teoggml-zdnn: rm user mapped buffers (#15965)
610aa0c461efe3603639af1a1defed2438d9c16ca0fb9c9c9f789cd57fb0b28b25223305613cd90fa102025-09-13T16:29:43+01:00Jeff Bolzvulkan: fix failing dequant shaders (#15862)
611b9c9c9f789cd57fb0b28b25223305613cd90fa1050f4281a6f5c3a5d68bdeb12f904fa01e0e2ba912025-09-13T16:23:30+01:00Jeff Bolzvulkan: initialize vulkan-hpp to allow using extension function pointers (#15705)
61250f4281a6f5c3a5d68bdeb12f904fa01e0e2ba9155758b00cae1451a0d789c50a5eb8c9bee42b9a02025-09-13T07:49:49-07:00Diego Devesallama : allow using iGPUs with --device (#15951)
61355758b00cae1451a0d789c50a5eb8c9bee42b9a0f161463a54d9f93d41246286aa4a9569a91d804d2025-09-13T16:24:22+03:00Georgi Gerganovmetal : refactor kernel loading (#15964)
614f161463a54d9f93d41246286aa4a9569a91d804d84d7b2fca11d1be118ce776f6d72a486c4883b742025-09-13T13:54:28+03:00Georgi Gerganovmetal : allow ops to run concurrently (#15929)
61584d7b2fca11d1be118ce776f6d72a486c4883b7440be51152d4dc2d47444a4ed378285139859895b2025-09-13T12:45:04+03:00Georgi Gerganovmetal : fix memory leaks (#15962)
61640be51152d4dc2d47444a4ed378285139859895b4bf5549269d99bac936a65892da2312cc71b24212025-09-13T02:39:52+08:00Aaron Teoggml-zdnn: fix #15414, activate FP16 and BF16 acceleration and incorrect zTensor free (#15839)
6174bf5549269d99bac936a65892da2312cc71b2421f4e664f838cc65d89fa845c48c372e43852112e42025-09-12T16:31:50+01:00Eric CurtinAdd docker protocol support for llama-server model loading (#15790)
618f4e664f838cc65d89fa845c48c372e43852112e4f088b6a84f6aed0abb619dbb9d375e726fc888a62025-09-12T23:16:32+08:00Haiyue Wangcontext : remove redundant explicit casting to the same type (#15948)
619f088b6a84f6aed0abb619dbb9d375e726fc888a6304ac5693d1e2124f83e0584bc5eea6311d3d3b42025-09-12T17:02:55+03:00Georgi Gerganovserver : adjust prompt similarity thold + add logs (#15913)
620304ac5693d1e2124f83e0584bc5eea6311d3d3b46c88ad8fa741a2182a2dcf8c5353ae4b5c66e6562025-09-12T13:24:21+02:00Ruben OrtlamVulkan iGPU device selection overhaul and PCI ID API support (#15947)
6216c88ad8fa741a2182a2dcf8c5353ae4b5c66e656704d90c987cdf00751567b2088c4e54742aa2d3f2025-09-12T09:06:20+02:00Mathieu Baudiervulkan: Make device memory check more portable (#15939)
622360d6533db39e11577afe9b0aece20c6b5ddaf1f0e6ff0046f4a2983b2c77950aa75960fe4b4f0e22025-09-11T13:47:38-07:00Diego Devesaggml-backend : add GGML_BACKEND_DEVICE_TYPE_IGPU device type (#15797)
6230e6ff0046f4a2983b2c77950aa75960fe4b4f0e2df082f56309073ecf885eceaa21b86e8a487e61b2025-09-11T21:19:58+02:00Johannes GäßlerCUDA: larger SRAM reads for tile FA, AMD FP16 dot (#15927)
624df082f56309073ecf885eceaa21b86e8a487e61b24a6734daf6932ff29ba8c1ff0245c51d76f783e2025-09-11T12:12:34-05:00ddh0nitpick : correct MB to MiB (#15934)
62524a6734daf6932ff29ba8c1ff0245c51d76f783e2b3efea9a4d91216850856fbb77075db26f6a6eb2025-09-11T15:39:12+02:00Daniel Beveniusggml-cpu : add check for ARM MATMUL_INT8/i8mm support (#15922)
6262b3efea9a4d91216850856fbb77075db26f6a6ebc0389dba43d50695f9d3f57dd1f1a14cbefc100c2025-09-11T12:45:40+02:00Charles Xukleidiai: fix GGML_ASSERT(*cur_backend_id != -1) failed (#15614)
627c0389dba43d50695f9d3f57dd1f1a14cbefc100c00681dfc16ba4cebb9c7fbd2cf2656e06a0692a42025-09-11T15:59:37+08:00hipuddingCANN: Disable acl_graph for prefill stage (#15933)
62800681dfc16ba4cebb9c7fbd2cf2656e06a0692a44f658855fa8f2e42b7ed9a5b298fa39a2e39b0962025-09-10T22:04:03+02:00Oliver SimonsCUDA: Add `fastdiv` to `k_bin_bcast*`, giving 1-3% E2E performance (#15872)
6294f658855fa8f2e42b7ed9a5b298fa39a2e39b0966ab397e12ba8e9f776341cdae68f7ffb2f8d2cde2025-09-11T02:51:51+08:00Jie Fu (傅杰)llama : support T5 models with unequal number of encoder-decoder layers (#15909)
6306ab397e12ba8e9f776341cdae68f7ffb2f8d2cde9de447d94e1ae9d1a36e5a2e5bf47483352c0d9c2025-09-10T19:08:59+02:00Sigbjørn Skjæretgraph : support non-contiguous Q in build_attn_mha (#15908)
6319de447d94e1ae9d1a36e5a2e5bf47483352c0d9c0f0a3c2851134d49955f3c85afbb0b1bb47c3e072025-09-10T17:31:40+02:00Daniel Beveniusggml-cpu : fix padding in ggml_timestep_embedding (#15917)
6320f0a3c2851134d49955f3c85afbb0b1bb47c3e0733daece86b65607451d0d4378d2d04ba6a20ad552025-09-10T17:52:35+03:00Georgi Gerganovmetal : make the backend async (#15906)
63333daece86b65607451d0d4378d2d04ba6a20ad55e7b6d83b524bbc24a1343d862de6dba8e8eddbd62025-09-10T15:39:57+02:00Daniel Beveniusci : add caching for ROCm installation in release workflow (#15924)
634e7b6d83b524bbc24a1343d862de6dba8e8eddbd62cfef4d117d67ab1dec002915b48a15d11ee19732025-09-10T14:17:09+02:00Daniel Beveniustests : filter out no-ops from coverage report (#15900)
6352cfef4d117d67ab1dec002915b48a15d11ee197309e72a037c77b6823fde9e1e4cf28e815b9c41ab2025-09-10T12:51:28+01:00j-kmedia : add transparent icon svg and png [no ci] (#15891)
63609e72a037c77b6823fde9e1e4cf28e815b9c41ab10d8b2b6b0ac2cae252a80b4daea5da55ab63c2f2025-09-10T07:28:47-04:00Jessegitignore : Ignore vim swap files in tests (#15901)
63710d8b2b6b0ac2cae252a80b4daea5da55ab63c2f28b5f190ef1dbea5edf82dbc8b4407b721fadd132025-09-10T18:42:00+08:00Chenguang LiCANN: Add ROPE sin/cos cache for reuse (#15912)
63828b5f190ef1dbea5edf82dbc8b4407b721fadd1386587da03bd78df8f4e7d8b111a0c1d2494d6ed02025-09-10T15:29:12+08:00Chenguang LiCANN: implement LRU cache for ACL graphs (#15814)
63986587da03bd78df8f4e7d8b111a0c1d2494d6ed0ff02caf9eed261423289d1531a56536fbf57bfc22025-09-10T05:33:58+02:00Daniel Beveniusllama : check returned fn ptrs from ggml_backend_reg_get_proc_address (#15893)
640ff02caf9eed261423289d1531a56536fbf57bfc2ae355f6f7108540297d8b7f7ae71d20fe610a0b72025-09-10T05:23:19+02:00Daniel Beveniusci : cache ROCm installation in windows-latest-cmake-hip (#15887)
641ae355f6f7108540297d8b7f7ae71d20fe610a0b74f63cd705c7b6f457f36c63fdc053e07f6f3cc6b2025-09-09T22:26:03+02:00Ruben Ortlamvulkan: throw the oom error instead of no memory type found (#15905)
6424f63cd705c7b6f457f36c63fdc053e07f6f3cc6b17bc5a815f0bebc1844c99796760cd7df5e9b8d92025-09-09T07:41:15-05:00Jeff Bolzvulkan: Fix OOB accesses in soft_max_back (#15861)
64317bc5a815f0bebc1844c99796760cd7df5e9b8d9ed54e32558ffe0f2c3b1d31f3227211fae05bd492025-09-09T14:04:43+02:00Johannes GäßlerHIP: use v_dot2_f32_f16 instruction for FA (#15884)
644ed54e32558ffe0f2c3b1d31f3227211fae05bd49a972faebed5fdc4a3d2a844d92d476058c02e02d2025-09-09T15:01:15+03:00lksj92hsWorkaround for subgroup arithmetic failing on MoltenVK with AMD GPUs (issue 15846) (#15886)
645a972faebed5fdc4a3d2a844d92d476058c02e02d550cf726e133fd0a069d991287fd3a2a3e3e1cbd2025-09-09T14:38:02+08:00Aman GuptaCUDA: Add mul_mat_id support for the mmf kernel (#15767)
646550cf726e133fd0a069d991287fd3a2a3e3e1cbdc252ce67c4b99f056eeb18d42a289e28fee034752025-09-09T08:11:01+02:00Johannes GäßlerCUDA: fix GET_ROWS for large tensors (#15882)
647c252ce67c4b99f056eeb18d42a289e28fee0347570cd37dbbebdb7a2f84f08207f18eabb0b291a552025-09-09T08:42:10+03:00Georgi Gerganovcontrib : add notes about merging PRs (#15881)
64870cd37dbbebdb7a2f84f08207f18eabb0b291a55acc1b008cfd95e63d5f99a370dbffb98e5a99d2c2025-09-09T06:06:52+02:00Daniel Beveniusrequirements : update transformers/torch for Embedding Gemma (#15828)
649acc1b008cfd95e63d5f99a370dbffb98e5a99d2c7057faf64b514e991e2f70147f82bb13d544b1c02025-09-09T06:05:55+02:00Piotr Wilkin (ilintar)model-conversion : add extra debugging support for model conversion (#15877)
6507057faf64b514e991e2f70147f82bb13d544b1c0fe1c92cd7bb491e9c5767c9de413f2b7dc1e832b2025-09-08T16:14:32-05:00Aldehir Rojasjson : support `enum` values within `allOf` (#15830)
651fe1c92cd7bb491e9c5767c9de413f2b7dc1e832be68aa10d8f3d26fdad5b912540362d79de5460e32025-09-08T19:57:01+01:00j-kmedia : add llama1 icon (#15878)
652e68aa10d8f3d26fdad5b912540362d79de5460e30a16bf52e6874369cb4fd2bdc4863ef984b688e72025-09-08T13:10:07-05:00Jeff Bolzvulkan: sort graph to allow more parallel execution (#15850)
6530a16bf52e6874369cb4fd2bdc4863ef984b688e788021565f08e0b7c4e07ac089a15ec16fae9166c2025-09-09T01:23:46+08:00Aman GuptaCUDA: generate_cu_files.py - add missing mxfp4 (#15880)
65488021565f08e0b7c4e07ac089a15ec16fae9166c56920f56651908d5cce7a310dabf54ac4f6fbb7f2025-09-08T10:59:48-04:00Jessechat : Deepseek V3.1 reasoning and tool calling support (OpenAI Style) (#15533)
65556920f56651908d5cce7a310dabf54ac4f6fbb7fb0d52998b962bd2681c34bf52af993af79f178b82025-09-08T21:50:05+07:00Xuan-Son Nguyenserver : bring back timings_per_token (#15879)
656b0d52998b962bd2681c34bf52af993af79f178b8f28d4f4ac963f182ea9d0fe9e269f3f5f3782aaf2025-09-08T13:56:51+03:00Georgi Gerganovcuda : fix supports_op condition for get_rows when number of blocks is too large (#15868)
657f28d4f4ac963f182ea9d0fe9e269f3f5f3782aaf9fcb29f22f5c33c04c7f0daebb24057899d67a1a2025-09-08T13:34:56+03:00Georgi Gerganovmetal : refactor + optimize (#15857)
6589fcb29f22f5c33c04c7f0daebb24057899d67a1a5ef22d281de9c5eaaf616874bc490b89241128cb2025-09-08T17:33:01+07:00Xuan-Son Nguyenggml: allow casting between f32 and i32 (#15783)
6595ef22d281de9c5eaaf616874bc490b89241128cb233d773d02c37982badddf3994e9953a175f34da2025-09-08T11:55:44+02:00Sigbjørn SkjæretCUDA: non-contiguous src0 not supported for PAD (#15869)
660233d773d02c37982badddf3994e9953a175f34daa885dcff11a7b73f9377812d6151f6b15d307de02025-09-08T09:44:34+02:00Daniel Beveniusconvert : force setting sliding_window from original config (#15867)
661a885dcff11a7b73f9377812d6151f6b15d307de0663027fd5490438ce9c27ea866a560e1e268d11f2025-09-08T10:27:07+03:00Georgi Gerganovbatched-bench : fix llama_synchronize usage during prompt processing (#15835)
662663027fd5490438ce9c27ea866a560e1e268d11fcf0e3ba1500bd23635b444f64ba23cbdb56c92ef2025-09-08T10:26:36+03:00Georgi Gerganovcontext : fix n_outputs during reserve (#15858)
663cf0e3ba1500bd23635b444f64ba23cbdb56c92efd413dca00360a7e4cb71441dacecfa32556fcc312025-09-08T10:25:33+03:00Georgi Gerganovmodel : avoid ggml_cont_3d for fused QKV weights (#15662)
664d413dca00360a7e4cb71441dacecfa32556fcc3185ca66a74676e6d5df4433016488e039a4b464ae2025-09-07T23:23:41-05:00Jeff Bolztests: large sizes for get_rows (#15687)
66585ca66a74676e6d5df4433016488e039a4b464ae3976dfbe00f02a62c0deca32c46138e4f0ca81d82025-09-08T10:03:29+08:00Chenguang LiCANN: Stream sync between devices for acl_graph (#15809)
6663976dfbe00f02a62c0deca32c46138e4f0ca81d8d36e61c580bf7fc7879c443c542312a42b718e112025-09-07T13:50:26-05:00Jeff Bolzvulkan: support im2col_3d (#15795)
667d36e61c580bf7fc7879c443c542312a42b718e11c97b5e5854b47b18a248d77edb693c63018a08652025-09-08T02:18:28+08:00Aaron Teoggml-cpu: clean up s390x SIMD (#15855)
668c97b5e5854b47b18a248d77edb693c63018a0865267e99867f09bec8bcc2e424ad9bcddd6cccf9d02025-09-07T12:00:49-05:00Jeff Bolzvulkan: Support pad_ext (#15794)
669267e99867f09bec8bcc2e424ad9bcddd6cccf9d03b15924d71237a43bb5ad71f5b885ee66a8213422025-09-07T11:53:07-05:00Jeff Bolzvulkan: Use larger loads in scalar/coopmat1 matmul (#15729)
6703b15924d71237a43bb5ad71f5b885ee66a82134279bc429262268ad2ac8a364cfe6c2d6b9c5f008a2025-09-07T10:19:45+02:00Daniel Beveniusggml WebGPU: remove userdata from request adapter callback (#15527)
67179bc429262268ad2ac8a364cfe6c2d6b9c5f008ac4df49a42d396bdf7344501813e7de53bc9e7bb32025-09-07T00:26:28+02:00Johannes GäßlerCUDA: faster tile FA (Pascal/AMD), headsize 256 (#15769)
672c4df49a42d396bdf7344501813e7de53bc9e7bb33c3635d2f20424d557b5b0605a2a356214ffe0482025-09-06T16:08:43+02:00Charles Xukleidiai: generalize compute_forward_kv_cache to compute_forward_fp16 (#15817)
6733c3635d2f20424d557b5b0605a2a356214ffe04861bdfd5298a78593be649a1035ee2a120b13c4f02025-09-06T19:45:24+07:00Xuan-Son Nguyenserver : speed up tests (#15836)
67461bdfd5298a78593be649a1035ee2a120b13c4f001806e77714ae8a78130d432945b959a0956c56f2025-09-06T18:35:04+07:00Xuan-Son Nguyenserver : implement prompt processing progress report in stream mode (#15827)
675186415d59552bd5c90549cd8e9be3cc287621fd9fd621880f3fd908424e675a41715a2dc760247a22025-09-06T11:27:28+08:00Aaron Teoggml-cpu: drop support for nnpa intrinsics (#15821)
676fd621880f3fd908424e675a41715a2dc760247a24281c7b315f8a904549fe527039b976e08098d1a2025-09-05T17:32:39-06:00Gabe GoodhartaLoRA Support (#15327)
6774281c7b315f8a904549fe527039b976e08098d1a5fac79cbc77b6d12c9feb5f34fc63586b35fd5612025-09-06T01:21:15+02:00Sigbjørn Skjæretci : exempt correct research label (#15825)
6785fac79cbc77b6d12c9feb5f34fc63586b35fd561408ff524b40baf4f51a81d42a9828200dd4fcb6b2025-09-05T14:31:24-06:00Gabe GoodhartThinking model disabled assistant prefill (#15404)
679408ff524b40baf4f51a81d42a9828200dd4fcb6b5143fa895e7725c5bd2135daf7d8f793d98fa91c2025-09-05T19:43:59+01:00Eric CurtinImplement --log-colors with always/never/auto (#15792)
6805143fa895e7725c5bd2135daf7d8f793d98fa91c3a550b5ca4565c9e28f63880d47840feb27d0ff62025-09-05T16:07:02+02:00Johannes GäßlerCUDA: fastdiv, launch bounds for mmvq + q8_1 quant (#15802)
6813a550b5ca4565c9e28f63880d47840feb27d0ff6a81283820a466f2ace06ce4d4bc9512761f9365f2025-09-05T14:49:21+02:00Daniel Beveniustests : add --list-ops and --show-coverage options (#15745)
682a81283820a466f2ace06ce4d4bc9512761f9365fc610b6c11b1ef7d678671dcf15acd7187a7ad8f32025-09-05T11:34:28+02:00Erik Scholzgguf: gguf_writer refactor (#15691)
683c610b6c11b1ef7d678671dcf15acd7187a7ad8f35d6688de08e73acc2532d668380801ed79d704eb2025-09-05T10:39:22+03:00Georgi Gerganovkv-cache : fix SWA checks + disable cacheless iSWA (#15811)
684dcbe998a6d8f13b6c9cb36bbe7b355ced2ac384287932ec016f0ec81e98a13ce5212851f8c12458a2025-09-05T14:34:07+08:00Yunzhe Jiaadd calrt API used in server side
6855d6688de08e73acc2532d668380801ed79d704eb4fd1242bef6cb2325b4ff1c1a80f3b54b64508a62025-09-05T04:36:23+02:00Daniel Beveniusmodel-conversion : add --embeddings flag to modelcard.template [no ci] (#15801)
6864fd1242bef6cb2325b4ff1c1a80f3b54b64508a6b2426e469e2fdb6c44216d56baa4cfff4f39ae002025-09-05T01:24:08+02:00ExtReMLapinchat : fixed crash when Hermes 2 <tool_call> had a newline before it (#15639)
687b2426e469e2fdb6c44216d56baa4cfff4f39ae009e2b1e83c68a38ea0c64f726dd979439bd02189b2025-09-05T01:22:22+02:00Piotr Wilkin (ilintar)chat : nemotron thinking & toolcalling support (#15676)
6889e2b1e83c68a38ea0c64f726dd979439bd02189bfb15d649ed14ab447eeab911e0c9d21e35fb243e2025-09-05T01:05:12+02:00Piotr Wilkin (ilintar)scripts : add Jinja tester PySide6 simple app (#15756)
689fb15d649ed14ab447eeab911e0c9d21e35fb243e856ed0947f27b4ec3ad269fceda0402fbab263d32025-09-04T18:10:29+02:00Daniel Beveniusllama : add support for EmbeddingGemma 300m (#15798)
690856ed0947f27b4ec3ad269fceda0402fbab263d3d1e2adba65208d6de3d7f6f23b00c562ff5bb7772025-09-04T09:53:22-06:00Gabe Goodhartmetal : Add template specialization for mul_mm_id w/ ne20 == 10 (#15799)
691d1e2adba65208d6de3d7f6f23b00c562ff5bb777c1c354e44c06d259679bb5bb7a8fa9f0b28480e42025-09-04T15:40:44+02:00Daniel Beveniusllama : set n_outputs to 1 to avoid 0 outputs mean-pooling (#15791)
692c1c354e44c06d259679bb5bb7a8fa9f0b28480e4a68d9144262f1d0ef4f6ba7ad4a7e73e977ba78c2025-09-04T20:20:14+08:00Chenguang LiCANN: Refactor ND to NZ workspace to be per-device (#15763)
693a68d9144262f1d0ef4f6ba7ad4a7e73e977ba78cbadb80cadbc40e047b30c43611aba575fc8d68452025-09-04T11:50:23+02:00Xuan-Son Nguyenserver: add exceed_context_size_error type (#15780)
694badb80cadbc40e047b30c43611aba575fc8d68450a1b3982cd0bd18730d50a693053b88c13fd04a62025-09-04T10:49:44+01:00Eric CurtinDocument the new max GPU layers default in help (#15771)
6950a1b3982cd0bd18730d50a693053b88c13fd04a65421f63ab08ca1e1a093662a5ccd0117e461185f2025-09-04T16:38:49+08:00leejetggml: add ops for WAN video model (cuda && cpu) (#15669)
6965421f63ab08ca1e1a093662a5ccd0117e461185f820bc9853100708011036e10f40b923968dd9b662025-09-04T15:12:30+08:00hipuddingCANN: Fix precision issue on 310I DUO multi-devices (#15784)
697820bc9853100708011036e10f40b923968dd9b66239b60e8986bbcb944f57311a02ac994431bf6522025-09-04T08:30:28+02:00rmatifopencl: add hs=40 to FA (#15758)
698239b60e8986bbcb944f57311a02ac994431bf652dff7551bfdbdd6e57c13e523d7dcca317640e9072025-09-04T11:03:02+08:00Chenguang LiCANN: fix acl_rstd allocation size in ggml_cann_rms_norm (#15760)
699dff7551bfdbdd6e57c13e523d7dcca317640e9070fce7a1248b74148c1eb0d368b7e18e8bcb968092025-09-03T22:55:10+02:00Ruben Ortlamvulkan: fix mmv subgroup16 selection (#15775)
7000fce7a1248b74148c1eb0d368b7e18e8bcb968098227695d7a3e5b357cb37fad263f00a8ca6db7102025-09-03T13:33:15-05:00Jeff Bolzvulkan: don't use std::string in load_shaders, to improve compile time (#15724)
7018227695d7a3e5b357cb37fad263f00a8ca6db7100014fb4add3a7d000c14b763538045c6170f57d92025-09-03T20:24:50+02:00Daniel Beveniusvulkan : update ggml_vk_instance_validation_ext_available (#15666)
7020014fb4add3a7d000c14b763538045c6170f57d9661ae31c9c68201577e70278285b349a5a662caf2025-09-04T03:22:55+09:00Shin-myoung-serpggml vulkan: add hardsigmoid and hardswish operations (#15762)
703661ae31c9c68201577e70278285b349a5a662caf407c23786dd0d3a503e9429eead96e611d3950c92025-09-03T19:59:16+02:00Oliver SimonsCUDA: Optimize `rms_norm_f32` kernel and its fused variants, giving 1-6% perf E2E (#15715)
704407c23786dd0d3a503e9429eead96e611d3950c9cdedb70a998cea7052560fe0b0615a839443564d2025-09-03T18:28:36+02:00Daniel Beveniusmodel-conversion : fix pyright errors (#15770)
705cdedb70a998cea7052560fe0b0615a839443564d2c8dac72eb6acd4e20c0da251535dfc46d35178b2025-09-03T18:16:26+03:00Georgi Gerganovsampling : optimize dist sampler (#15704)
7062c8dac72eb6acd4e20c0da251535dfc46d35178b40a751ea9a94364da73537b86502a808ebe1fc3a2025-09-03T13:35:49+02:00Daniel Beveniusllama : fix incorrect model type for Gemma 270M (#15764)
70740a751ea9a94364da73537b86502a808ebe1fc3a5eae9348835037046f33fafd852df7c952c952092025-09-03T12:50:47+02:00Daniel Beveniusmodel-conversion : remove hardcoded /bin/bash shebangs [no ci] (#15765)
7085eae9348835037046f33fafd852df7c952c9520905c0380f2ae5c7193445e03ebc7b6de9ce49f1a62025-09-03T16:46:01+08:00hipuddingCANN: Add RoPE contiguous check for 310I DUP device (#15735)
70905c0380f2ae5c7193445e03ebc7b6de9ce49f1a68c3fdf44ecf08335942f2ba558955f55e88c79912025-09-03T16:16:21+08:00xctanggml-cpu : optimize RVV kernels (#15720)
7108c3fdf44ecf08335942f2ba558955f55e88c7991f6da8cb86a28f0319b40d9d2a957a26a7d875f8c2025-09-03T09:48:35+02:00Daniel Beveniusmodel-conversion : add missing curl script [no ci] (#15761)
711f6da8cb86a28f0319b40d9d2a957a26a7d875f8c8a2234ea0c89f212190c176d741b7742f00825822025-09-03T14:08:22+08:00hipuddingCANN: Mask unsupported TRANSPOSE_1D operator (#15733)
7128a2234ea0c89f212190c176d741b7742f00825823de008208b9b8a33f49f979097a99b4d59e6e5212025-09-03T10:43:53+08:00Chenguang LiCANN: Fix type float_t to float (#15736)
7133de008208b9b8a33f49f979097a99b4d59e6e52169db8a52e6dd0db81ec05c581150cc1e43b8ac462025-09-03T03:27:30+08:00SnA1lGofix: resolve unsigned int initialization warning for n_dims/size in gguf.cpp (#15754)
71469db8a52e6dd0db81ec05c581150cc1e43b8ac46c466abe1587bde458c806e2134e37750f2edf8fb2025-09-02T19:40:37+02:00Oliver Simonschore: Update `.clang-format` to use `BinPackArguments=true` (#15744)
715c466abe1587bde458c806e2134e37750f2edf8fb0a2a3841e8ebc570da343e8cf58a21c1010b41e72025-09-02T18:17:26+02:00Johannes Gäßlerllama: -fa 1/0/-1 aliases for -fa on/off/auto (#15746)
7160a2a3841e8ebc570da343e8cf58a21c1010b41e79961d244f2df6baf40af2f1ddc0927f8d91578c82025-09-02T16:02:26+02:00Ruben Ortlamvulkan: fix shaders gen when no integer dot is available (#15740)
7179961d244f2df6baf40af2f1ddc0927f8d91578c825f1045f07cf0daf667d63e35618842e3174a8c72025-09-02T17:12:37+08:00hipuddingCANN: Resolve soft_max precision issue (#15730)
71825f1045f07cf0daf667d63e35618842e3174a8c797669e40735d08db65ef094a8faae8e8411a97db2025-09-02T01:37:01-05:00Jeff Bolzvulkan: Fix macro parameter order for f32 matmul shaders (#15716)
71997669e40735d08db65ef094a8faae8e8411a97db2f853687b3bce15e143a22f678d1715060fd606c2025-09-02T08:26:53+02:00rmatifopencl: add attn sinks support for FA kernels (#15706)
7202f853687b3bce15e143a22f678d1715060fd606cef2af57ddf04ab367f6ba6e8ed100fc56785e4b72025-09-02T14:07:48+08:00Chenguang LiCANN: Support eager execution mode under ACL graph compilation (#15712)
721ef2af57ddf04ab367f6ba6e8ed100fc56785e4b75d804a4938d896f9089687a8b99911cdb43b0b942025-09-02T14:05:23+08:00hipuddingCANN: Support ext_factor in rope (#15710)
7225d804a4938d896f9089687a8b99911cdb43b0b94d4d8dbe383e8b9600cbe8b42016e3a4529b512192025-09-02T01:14:55+02:00Johannes Gäßlerggml-backend: raise GGML_MAX_SPLIT_INPUTS (#15722)
723d4d8dbe383e8b9600cbe8b42016e3a4529b5121935a42edac84690990a75726898d975cd08d220c02025-09-01T22:17:42+03:00Gilad S.vulkan: use memory budget extension to read memory usage (#15545)
72435a42edac84690990a75726898d975cd08d220c0fec7911f8febb22c27ce1acb12441800da24cc062025-09-01T14:01:10-05:00Jeff Bolzvulkan: add missing clamps in new mul_mat_id paths (#15702)
725fec7911f8febb22c27ce1acb12441800da24cc06078ce23ea77988f2fe1a42afb18d58bc084d55fa2025-09-01T20:58:35+02:00Ruben Ortlamvulkan: disable large mmv subgroups on older Nvidia GPUs (#15717)
726078ce23ea77988f2fe1a42afb18d58bc084d55faa0c2b207c596d1092a08615de61ab56a7d63515f2025-09-02T03:13:49+09:00s-goto-11ggml: SVE support for exponential functions (#15145)
727a0c2b207c596d1092a08615de61ab56a7d63515f4b20d8b7e31718d3fe8b4f12d220a9d19e4f19972025-09-01T23:43:16+05:30Prashant Vithuleggml: aarch64: Implement SVE F16 kernels for vector functions (#15115)
7284b20d8b7e31718d3fe8b4f12d220a9d19e4f199702c1813517412f3e00aa6ca7c0273fea64edb4922025-09-01T23:53:31+08:00Jie Fu (傅杰)convert : remove redundant code (#15708)
72902c1813517412f3e00aa6ca7c0273fea64edb49277dee9de97be75b7143a213bc48893e0c0b29af72025-09-01T16:19:07+02:00Ruben OrtlamVulkan: Add Integer Dot Product mul_mat_vec shader for legacy quants (#14903)
73077dee9de97be75b7143a213bc48893e0c0b29af74795c91c32fec7165a1364763d4d4f0c93abf9332025-09-01T14:28:49+02:00Daniel Beveniusggml : WebGPU add TRANSPOSE and RESHAPE to supported ops (#15695)
7314795c91c32fec7165a1364763d4d4f0c93abf933b66df9d9c942254d03209186ef24ed7c994a576e2025-09-01T15:34:59+08:00Jie Fu (傅杰)docs : add Hunyuan to models section (#15707)
732b66df9d9c942254d03209186ef24ed7c994a576eb9382c3877c6067feccf182efe9449a2d1cb24c72025-09-01T06:55:06+05:30Akarshan BiswasCUDA: fix build error from ambiguous __half conversions in conv2d (#15690)
733b9382c3877c6067feccf182efe9449a2d1cb24c73dc7397a2799bdc07bccf637ab7ae5a1e786d1a42025-09-01T08:57:23+08:00hipuddingCANN: Optimize MUL_MAT_ID (#15658)
7343dc7397a2799bdc07bccf637ab7ae5a1e786d1a4e92d53b29e393fc4c0f9f1f7c3fe651be8d36faa2025-09-01T08:57:00+08:00hipuddingCANN: fix RoPE cache issue on multi-device (#15629)
735e92d53b29e393fc4c0f9f1f7c3fe651be8d36faa0d161f021aa33ec0e90cce96f5d1a889255573272025-08-31T20:41:02+03:00Georgi Gerganovsampling : optimize samplers by reusing bucket sort (#15665)
7360d161f021aa33ec0e90cce96f5d1a889255573274efd5a83163ff383285b3a4c2106feabf5c695572025-08-31T20:11:58+03:00Georgi Gerganovserver : enable /slots by default and make it secure (#15630)
7374efd5a83163ff383285b3a4c2106feabf5c69557274966226f87f301ac132da898280ca3142b60e52025-08-31T19:43:30+03:00Georgi Gerganovmetal : fix checks for available FA kernels (#15700)
738274966226f87f301ac132da898280ca3142b60e59777032dccd67bdc7785aeab7497014a8be8dacc2025-08-31T08:47:05-07:00Diego Devesallama : fix fattn reserve call n_seqs parameter (#15699)
7399777032dccd67bdc7785aeab7497014a8be8dacc7d3c9f2b217acf0ce5db81ae83d3f375f49ab2c72025-08-31T06:49:03-07:00Diego Devesallama : separate compute buffer reserve from fattn check (#15696)
7407d3c9f2b217acf0ce5db81ae83d3f375f49ab2c7bbbf5ecccb35286521f735239d499eec4279a8402025-08-31T15:30:20+02:00Sigbjørn Skjæretci : explicitly set fa off or on (#15692)
741bbbf5ecccb35286521f735239d499eec4279a840c37052ab4d6d1ae73c0e90bc6e560cc6409e13112025-08-31T03:13:27-05:00Jeff Bolzvulkan: handle large sizes for get_rows (#15686)
742c37052ab4d6d1ae73c0e90bc6e560cc6409e13115c16b9c87d840e4d5d55fa83c732c6b693346f402025-08-31T02:06:43-05:00Jeff Bolzvulkan: mul_mat_id coopmat2 optimizations (#15546)
7435c16b9c87d840e4d5d55fa83c732c6b693346f40b97c9edc59d4a1b4069991aa670411190f4f3a3e2025-08-31T08:46:42+02:00Daniel Beveniusvulkan : remove unused portability_enumeration_ext variable (#15679)
744b97c9edc59d4a1b4069991aa670411190f4f3a3e94e82c7eadeb8fff0db4bfd1ab6d8cf65fa6f2e02025-08-31T01:30:54-05:00Jeff Bolzvulkan: Allow fallback to sysmem memory when vidmem is full (#15649)
74594e82c7eadeb8fff0db4bfd1ab6d8cf65fa6f2e04d74393bcc956ccd7df68a6a06d1a0575cfa712c2025-08-31T01:27:57-05:00Jeff Bolzvulkan: clamp matmul and FA results to the max finite value (#15652)
7464d74393bcc956ccd7df68a6a06d1a0575cfa712cdd892555b0681b7f56d38780f6fdfe00a195160f2025-08-30T18:03:42+02:00Charles Xuggml: update kleidiai to v1.13.0 (#15663)
747dd892555b0681b7f56d38780f6fdfe00a195160fe81b8e4b7f5ab870836fad26d154a7507b341b362025-08-30T08:51:28-07:00Diego DevesaUpdate build.md to remove MSVC arm64 notes (#15684)
748e81b8e4b7f5ab870836fad26d154a7507b341b3638ad381f9f5d4dd368a96d844fb19cf501ed9d222025-08-30T16:32:10+02:00Johannes Gäßlerllama: use FA + max. GPU layers by default (#15434)
74938ad381f9f5d4dd368a96d844fb19cf501ed9d22696fccf354e9dbdfbce135bc40b44c9dcc64dda92025-08-30T16:20:32+02:00Johannes GäßlerCUDA: use FP32 arithmetic for conv2d (#15683)
750696fccf354e9dbdfbce135bc40b44c9dcc64dda9ef476916bba4b44f44be0c98babc1cb025968e752025-08-30T04:11:22-05:00Jeff Bolzvulkan: Skip syncing for prealloc_y when it is reused (#15544)
751ef476916bba4b44f44be0c98babc1cb025968e75d82f6aa34a216f5df1945cdfe121ba5e6cd80be02025-08-30T10:18:35+08:00Chenguang LiCANN: FIx compiler warnings (#15661)
752d82f6aa34a216f5df1945cdfe121ba5e6cd80be03d16b29c3bb1ec816ac0e782f20d1690970639192025-08-30T00:12:53+02:00Sergey Alirzaevserver : removed obsolete doc (#15670)
753792b44f2ed9668cce7f267ff0ae4950ed9b4a5de81017865ee444cf49ce0136f2be1e41a0270ff912025-08-29T19:25:40+02:00ExtReMLapinserver : add documentation for `parallel_tool_calls` param (#15647)
75481017865ee444cf49ce0136f2be1e41a0270ff9160e5eee31f1af9bb579ac45380e3857d610020b92025-08-29T21:30:06+08:00Aman GuptaCUDA: fix bug in rms_norm fusion (#15660)
75560e5eee31f1af9bb579ac45380e3857d610020b9009b709d6efd24820ac67765ed339a72dc7978142025-08-29T14:53:41+02:00Piotr Wilkin (ilintar)chat : Seed OSS thinking + tool call support (#15552)
756009b709d6efd24820ac67765ed339a72dc797814e8d99dd0b67f2ecc1e45fca8074a3a18c3e036d22025-08-29T11:35:58+08:00Aman GuptaCUDA: fuse adds, fuse add with rms norm (#15631)
757e8d99dd0b67f2ecc1e45fca8074a3a18c3e036d2a8bca68f727844e7dcf24a956003b3c2039ea5632025-08-28T18:39:31-06:00Gabe Goodhartnvidia nemotron nano v2 (nemotronh) (#15507)
758a8bca68f727844e7dcf24a956003b3c2039ea563c97dc093912ad014f6d22743ede0d4d7fd82365a2025-08-28T15:27:36-05:00Gabe Goodhartfix: Compute the full sum in llama-eval-callback, not just the sum of printed values (#15637)
759c97dc093912ad014f6d22743ede0d4d7fd82365a6c442f42ff25564a0cd6b1435d9abc1b0178eac52025-08-29T00:03:03+05:30mnehete32CUDA: add conv2d (#15635)
7606c442f42ff25564a0cd6b1435d9abc1b0178eac573804145ab6c06888e314046abf08f66a01336792025-08-28T22:39:27+08:00Aaron Teoggml-cpu: fix invalid hsum build in debug s390x (#15634)
76173804145ab6c06888e314046abf08f66a0133679c8d0d14e77c3c45df5cbbddde9c2b4c377ec4e7a2025-08-28T10:11:36-04:00compiladeggml : fix SSM_SCAN for n_groups > 1 (#15625)
762c8d0d14e77c3c45df5cbbddde9c2b4c377ec4e7a84ab83cc0b4b7e769451ee48e4c7d1acef91ef252025-08-28T17:09:05+03:00Georgi Gerganovkv-cache : fix find_slot to not search for continuous slot (#15638)
76384ab83cc0b4b7e769451ee48e4c7d1acef91ef2555042b3692cb1467c9ee15c62c4a9fbf180f89e32025-08-28T15:49:50+02:00Sigbjørn Skjæretmodel : jina-embeddings-v3 support (#13693)
76455042b3692cb1467c9ee15c62c4a9fbf180f89e38a4280ce431da6b33e5a95ae1fd61472c8c3f8cc2025-08-28T19:23:22+08:00Aman Guptascripts: add sqlite3 check for compare-commits.sh (#15633)
7658a4280ce431da6b33e5a95ae1fd61472c8c3f8cc64387f6e95434b393ac3df285864692b7fd9c4d22025-08-28T12:27:02+03:00Georgi Gerganovkv-cache : remove LLAMA_SET_ROWS checks (#15505)
76664387f6e95434b393ac3df285864692b7fd9c4d2d35a1e8c41f747548775225973a99507896a8c612025-08-28T10:56:41+02:00Aleksei Nikiforovgguf-py: byteswapping improvements (#12851)
767d35a1e8c41f747548775225973a99507896a8c6146d9caa27a0281150e8cf082308c0f9e7576ebe52025-08-28T01:48:20-06:00Joshua Cogliaticli : change log to warning to explain reason for stopping (#15604)
76846d9caa27a0281150e8cf082308c0f9e7576ebe55a0e3ef6f00c658fbae53797f02d5a360ebf8fec2025-08-28T09:26:48+02:00Daniel Beveniusmodel-conversion : add mmproj conversion target (#15628)
7695a0e3ef6f00c658fbae53797f02d5a360ebf8fecfbef0fad7a7c765939f6c9e322fa05cd52cf0c152025-08-27T17:32:36-07:00matiaslincuda: Add cublasLt_static linking when GGML_STATIC is enabled (#15622)
770fbef0fad7a7c765939f6c9e322fa05cd52cf0c15da54f9f1a2db07aaae390024ac466e7867685d942025-08-27T20:58:09+02:00Johannes Gäßlerserver: higher timeout for tests (#15621)
771da54f9f1a2db07aaae390024ac466e7867685d9447373271f971aa5a0a6462b286f4a7b5bd4ba6442025-08-27T15:48:07+03:00Georgi Gerganovpresets : add qwen3-30B-a3b FIM (#15616)
77247373271f971aa5a0a6462b286f4a7b5bd4ba6441bded5a3b3376b2aa3ba2f11ade5910c550f354b2025-08-27T13:58:54+02:00uvosHIP: Enable support for ggml_backend_cuda_register_host_buffer (#15615)
7731bded5a3b3376b2aa3ba2f11ade5910c550f354b1e7489745a74996fc36e8fd05b73aa16bc184e0c2025-08-27T13:55:12+03:00Georgi Gerganovkv-cache : better estimate of n_kv for multi-sequence batches (#15610)
7741e7489745a74996fc36e8fd05b73aa16bc184e0c1cf123a343ab7ca5586aacb9e0a1d2de7fe33be42025-08-27T17:21:41+08:00Chenguang LiCANN: refactor mask handling and improve performance in FA (#15561)
7751cf123a343ab7ca5586aacb9e0a1d2de7fe33be4fcca2182a18c786c57d02d1d1927204b133f1fdc2025-08-27T16:44:22+08:00xctanggml-cpu : add basic RVV support for vector f32 ops (#15057)
776fcca2182a18c786c57d02d1d1927204b133f1fdc86076f92de1a547ef87c304facca3b4b9fae6c212025-08-27T10:28:53+02:00Daniel Beveniuscommon : add -m to bash completion for --model [no ci] (#15591)
77786076f92de1a547ef87c304facca3b4b9fae6c21bcbddcd54f0d5c22eab180831fdea6484107112f2025-08-27T08:36:05+02:00rmatifOpenCL: add fused group_norm/norm, mul, add (#15314)
77887932ec016f0ec81e98a13ce5212851f8c12458a0115cfb7c1dec0753aecb34f3af4338bf9de8eeb2025-08-27T14:29:18+08:00Yunzhe Jiafix merge problem
779bcbddcd54f0d5c22eab180831fdea6484107112f8b696861364360770e9f61a3422d32941a4778242025-08-26T13:14:38-07:00Diego Devesatests : fix test-opt with GGML_BACKEND_DL (#15599)
7808b696861364360770e9f61a3422d32941a4778248ce3ff1d91245e158d98d8062cd64b0dd98dcfe32025-08-27T00:27:49+05:30Akarshan BiswasSYCL: fix rms_norm_mul_add for tensor dim not a multiple of sg_size (#15592)
7818ce3ff1d91245e158d98d8062cd64b0dd98dcfe344b1efa41acd4df3f56ee0e46f898135ecd1a0542025-08-26T20:05:50+02:00fidorielmtmd : fix mtmd ios build (#15579)
78244b1efa41acd4df3f56ee0e46f898135ecd1a054a6a58d64785cb458ed9de52f391aa38142d38d642025-08-26T15:42:49ZEvetests: add performance test for mul mat id (#15543)
783a6a58d64785cb458ed9de52f391aa38142d38d640373486dbc0dccbdcb3b5fdd65759d88cec061962025-08-26T21:05:25+05:30shalinib-ibmllamafile: PowerPC Sgemm Optimization (#15558)
7840373486dbc0dccbdcb3b5fdd65759d88cec0619662cef26ac5b6b7acb635d3dc963813b43952dc2b2025-08-26T17:45:17+03:00Georgi Gerganovgraph : fix assert in memory-less build_attn (#15590)
78562cef26ac5b6b7acb635d3dc963813b43952dc2b8f5afa94c4f929da71f560db7c9f38ef6a783d952025-08-26T16:12:29+02:00Daniel Beveniusmodel-conversion : add qat-q4 quantization targets (#15588)
7868f5afa94c4f929da71f560db7c9f38ef6a783d95b3964c1e890ef8c947afb36a5124ce6fcb2136d42025-08-26T16:01:20+02:00Johannes GäßlerCUDA: return -1 for nonexistent compiled arch (#15587)
787b3964c1e890ef8c947afb36a5124ce6fcb2136d479a546220c719e6a70627b243a478ab8d84dc9e12025-08-26T14:22:14+03:00Georgi Gerganovmetal : optimize FA vec for large sequences and BS <= 8 (#15566)
78879a546220c719e6a70627b243a478ab8d84dc9e185cc1ae998e4898d9fa992cb9b8620338cee97bf2025-08-26T12:54:19+02:00Xuan-Son Nguyenmtmd : support Kimi VL model (#15458)
78985cc1ae998e4898d9fa992cb9b8620338cee97bf1d8d83deaa48d4a5491820d58ff1c0d8cf9d196c2025-08-26T12:47:00+03:00Georgi Gerganovcontext : print graph stats for memory-less contexts (#15586)
7901d8d83deaa48d4a5491820d58ff1c0d8cf9d196cc4e9239064a564de7b94ee2b401ae907235a8fca2025-08-26T12:46:15+03:00Georgi Gerganovmetal : improve `MUL_MAT_ID` (#15541)
791c4e9239064a564de7b94ee2b401ae907235a8fca39842a7f73012eb42816ca4f26411782bd3da7c52025-08-26T16:05:55+08:00tc-mbmodel : support MiniCPM-V 4.5 (#15575)
7920115cfb7c1dec0753aecb34f3af4338bf9de8eeb3a617cb18121a85eed17ac5585788810c5d9c1e0 0d5a470223fc90b6b6807921d68011ff06ae7f9e2025-08-26T15:19:56+08:00Yunzhe JiaMerge branch 'master' into dev
79339842a7f73012eb42816ca4f26411782bd3da7c50fd90db5858e325358a5fbdaa4e327ee2a79d0d42025-08-26T09:08:08+02:00Sigbjørn Skjæretgguf-py : remove erroneous FFN_GATE entry (#15583)
7940fd90db5858e325358a5fbdaa4e327ee2a79d0d44c37636b3ea96f2574eeb7668b93fcc0e64b05dd2025-08-26T08:51:43+02:00Sigbjørn Skjæretmetal : remove contiguous assertion for src0 in IM2COL (#15577)
7954c37636b3ea96f2574eeb7668b93fcc0e64b05dd34bdbbd7c2b70b848718e95bc48010f6aecd28162025-08-26T13:15:33+07:00Yoshi_likes_e4Add a warning for special devices (#15563)
79634bdbbd7c2b70b848718e95bc48010f6aecd281674f52f77f28a5ad6d6075231afcb8d1ad763ca322025-08-25T23:42:44-05:00Jeff Bolzvulkan: Remove splitting for mul_mat_id (#15568)
79774f52f77f28a5ad6d6075231afcb8d1ad763ca32f7207b0415986dd7f48447149da7de3a823382762025-08-26T05:21:22+08:00QeeweewCUDA: Accelerate MXFP4 table lookup using `__byte_perm` (#15451)
798f7207b0415986dd7f48447149da7de3a823382764d917cd4f64cc37744e76d084659475819fb07282025-08-25T14:18:09-07:00lhezopencl: fix support ops condition for `rms_norm` (#15560)
7994d917cd4f64cc37744e76d084659475819fb0728886b97a5d693550c2da470c091d9d27bf38398f82025-08-25T17:56:59+02:00Ruben Ortlamvulkan: fix min subgroup 16 condition for mmid subgroup optimization (#15565)
800886b97a5d693550c2da470c091d9d27bf38398f8111f8d06f0b9169059779a35f655b343242ccbb62025-08-25T10:47:16-05:00Jeff Bolztests: Generate unique input values for count_equal (#15487)
801111f8d06f0b9169059779a35f655b343242ccbb65eff6ec9b1220b599a43b594b1110487ab6aca082025-08-25T11:27:34-04:00Ihar Hrachyshkametal: fix regression when no metal devices are present (#15531)
8025eff6ec9b1220b599a43b594b1110487ab6aca08dfd9b5f6c7586c88588f06a644c131bec071a0a12025-08-25T17:23:40+02:00Johannes GäßlerCUDA: MoE helper in device code, better tile sizes (#15525)
803dfd9b5f6c7586c88588f06a644c131bec071a0a15a6bc6b1a6cb665a944426c2055794950e524bf52025-08-25T15:00:43+02:00Daniel Beveniusmodel-conversion : set pooling type to none in logits.cpp (#15564)
8045a6bc6b1a6cb665a944426c2055794950e524bf56b64f74b55628e4193f4fb00313f07dbd85565282025-08-25T14:25:25+02:00Daniel Beveniusmodel-conversion : add model card template for embeddings [no ci] (#15557)
8056b64f74b55628e4193f4fb00313f07dbd85565280d5a470223fc90b6b6807921d68011ff06ae7f9e2025-08-25T13:56:43+03:00Georgi Gerganovbatched-bench : fix unified KV cache handling + pp timing (#15562)
8063a617cb18121a85eed17ac5585788810c5d9c1e0c42293848dd9f03eba9a6c1b85a600b398f065412025-08-13T10:58:59+08:00Yunzhe Jiaadd calrt_decode
807c42293848dd9f03eba9a6c1b85a600b398f065411cbf442f50a154aac0bf35d7b8b3436a2cdfd6f52025-08-06T15:46:06+08:00Yunzhe Jiasync with calrt API
8081cbf442f50a154aac0bf35d7b8b3436a2cdfd6f5ba67d6c03fed3193987a4ef13050e6e2a4451df42025-07-24T15:50:19+08:00Yunzhe Jia1.sync with calrt prefill,decode,kv_update op 2.add hw pattern untile function
809ba67d6c03fed3193987a4ef13050e6e2a4451df48f52c6f8abb2ed9ad967cc0bfb90acb0730111932025-06-26T17:42:32+08:00Yunzhe Jiallama-calrt: infer-op: copy data from output buffer to dst-tensor.data
8108f52c6f8abb2ed9ad967cc0bfb90acb073011193d2f325130c4ed77a3dcff2c23a926587fb08e2cc2025-06-26T15:10:45+08:00Yunzhe JiaSquashed commit of the following:
811d2f325130c4ed77a3dcff2c23a926587fb08e2cc6391ee7ffd97e723dfbdcc32f001e1b95a63cba62025-06-17T11:53:33+08:00Yunzhe Jiacalrt: add input_token in graph; input_token will be copy to calrt_in_buffer
8126391ee7ffd97e723dfbdcc32f001e1b95a63cba6b9850050a5b5147cb2e6720b98ed8c1e8e02eff92025-06-13T10:58:06+08:00Yunzhe Jiaupdate .gitignore to update cal_test.gguf
813b9850050a5b5147cb2e6720b98ed8c1e8e02eff9933b205c32551848589b8339437d778cb4f2a46d2025-06-06T17:25:52+08:00Yunzhe Jiacalrt: add call chain for using calrt::infer
814933b205c32551848589b8339437d778cb4f2a46d33fea9a8592fd338abe279b89350a3af2e0ff1f72025-06-12T17:47:19+08:00Yunzhe Jiagguf-py: add cal_test.py
81533fea9a8592fd338abe279b89350a3af2e0ff1f7806f05a353d322e97d97e02f910de2b408a8b4ab2025-06-06T17:25:52+08:00Yunzhe Jiacalrt: add load calbin
816806f05a353d322e97d97e02f910de2b408a8b4abec9e0301fef6476df83e94842c3b625501c955662025-05-30T09:45:37+08:00Yunzhe Jiaadd calrt demo
Reference in New Issue View Git Blame Copy Permalink