diff --git a/.devops/llama-cli-intel.Dockerfile b/.devops/llama-cli-intel.Dockerfile index 6789e17af..bd816f9f5 100644 --- a/.devops/llama-cli-intel.Dockerfile +++ b/.devops/llama-cli-intel.Dockerfile @@ -2,7 +2,7 @@ ARG ONEAPI_VERSION=2024.1.1-devel-ubuntu22.04 FROM intel/oneapi-basekit:$ONEAPI_VERSION as build -ARG LLAMA_SYCL_F16=OFF +ARG GGML_SYCL_F16=OFF RUN apt-get update && \ apt-get install -y git @@ -10,11 +10,11 @@ WORKDIR /app COPY . . -RUN if [ "${LLAMA_SYCL_F16}" = "ON" ]; then \ - echo "LLAMA_SYCL_F16 is set" && \ - export OPT_SYCL_F16="-DLLAMA_SYCL_F16=ON"; \ +RUN if [ "${GGML_SYCL_F16}" = "ON" ]; then \ + echo "GGML_SYCL_F16 is set" && \ + export OPT_SYCL_F16="-DGGML_SYCL_F16=ON"; \ fi && \ - cmake -B build -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ${OPT_SYCL_F16} && \ + cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ${OPT_SYCL_F16} && \ cmake --build build --config Release --target llama-cli FROM intel/oneapi-basekit:$ONEAPI_VERSION as runtime diff --git a/.devops/llama-server-intel.Dockerfile b/.devops/llama-server-intel.Dockerfile index 3bf1670ec..8f8fef8c0 100644 --- a/.devops/llama-server-intel.Dockerfile +++ b/.devops/llama-server-intel.Dockerfile @@ -2,7 +2,7 @@ ARG ONEAPI_VERSION=2024.1.1-devel-ubuntu22.04 FROM intel/oneapi-basekit:$ONEAPI_VERSION as build -ARG LLAMA_SYCL_F16=OFF +ARG GGML_SYCL_F16=OFF RUN apt-get update && \ apt-get install -y git libcurl4-openssl-dev @@ -10,11 +10,11 @@ WORKDIR /app COPY . . -RUN if [ "${LLAMA_SYCL_F16}" = "ON" ]; then \ - echo "LLAMA_SYCL_F16 is set" && \ - export OPT_SYCL_F16="-DLLAMA_SYCL_F16=ON"; \ +RUN if [ "${GGML_SYCL_F16}" = "ON" ]; then \ + echo "GGML_SYCL_F16 is set" && \ + export OPT_SYCL_F16="-DGGML_SYCL_F16=ON"; \ fi && \ - cmake -B build -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_CURL=ON ${OPT_SYCL_F16} && \ + cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_CURL=ON ${OPT_SYCL_F16} && \ cmake --build build --config Release --target llama-server FROM intel/oneapi-basekit:$ONEAPI_VERSION as runtime diff --git a/AUTHORS b/AUTHORS index b029f13da..1bd36158a 100644 --- a/AUTHORS +++ b/AUTHORS @@ -1,8 +1,9 @@ -# date: Tue Apr 9 09:17:14 EEST 2024 +# date: Wed Jun 26 19:36:34 EEST 2024 # this file is auto-generated by scripts/gen-authors.sh 0cc4m 0xspringtime <110655352+0xspringtime@users.noreply.github.com> +20kdc 2f38b454 3ooabkhxtn <31479382+3ooabkhxtn@users.noreply.github.com> 44670 <44670@users.noreply.github.com> @@ -11,14 +12,18 @@ AT Aarni Koskela Aaron Miller Aaryaman Vasishta +Abheek Gulati Abhilash Majumder <30946547+abhilash1910@users.noreply.github.com> Abhishek Gopinath K <31348521+overtunned@users.noreply.github.com> Adithya Balaji AdithyanI Adrian Adrian Hesketh +Ahmet Zeer AidanBeltonS <87009434+AidanBeltonS@users.noreply.github.com> Aisuko +Akarshan Biswas +Albert Jin Alberto <57916483+albbus-stack@users.noreply.github.com> Alex Alex Azarov @@ -35,19 +40,24 @@ Ali Nehzat Ali Tariq Alon AlpinDale <52078762+AlpinDale@users.noreply.github.com> +Amir AmirAli Mirian <37371367+amiralimi@users.noreply.github.com> Ananta Bastola Anas Ahouzi <112881240+aahouzi@users.noreply.github.com> András Salamon Andrei Andrew Canis +Andrew Downing Andrew Duffy Andrew Godfrey +Andy Tai Arik Poznanski Artem +Artem Zinnatullin Artyom Lebedev Asbjørn Olling Ásgeir Bjarni Ingvarsson +Ashish <1856117+ashishdatta@users.noreply.github.com> Ashok Gelal <401055+ashokgelal@users.noreply.github.com> Ashraful Islam Atsushi Tatsuma @@ -57,35 +67,46 @@ BADR Bach Le Bailey Chittle <39804642+bachittle@users.noreply.github.com> BarfingLemurs <128182951+BarfingLemurs@users.noreply.github.com> +Bartowski Behnam M <58621210+ibehnam@users.noreply.github.com> +Ben Ashbaugh Ben Garney Ben Siraphob Ben Williams +Benjamin Findley <39356821+Kartoffelsaft@users.noreply.github.com> Benjamin Lecaillon <84293038+blecaillon@users.noreply.github.com> Bernat Vadell +Bingan <70050083+binganao@users.noreply.github.com> Bodo Graumann Bono Lv Borislav Stanimirov Branden Butler Brian Bruce MacDonald +Bryan Honof CJ Pais CRD716 +Calvin Laurenson Cameron Cameron Kaiser +Carolinabanana <140120812+Carolinabanana@users.noreply.github.com> Casey Primozic Casey Primozic CausalLM <148736309+CausalLM@users.noreply.github.com> Cebtenzzre Chad Brewbaker +Chao Jiang Cheng Shao +Chris Elrod Chris Kuehl Christian Demsar Christian Demsar Christian Falch <875252+chrfalch@users.noreply.github.com> Christian Kögler +Christian Zhou-Zheng <59622928+christianazinn@users.noreply.github.com> Clark Saben <76020733+csaben@users.noreply.github.com> Clint Herron +CrispStrobe <154636388+CrispStrobe@users.noreply.github.com> Cuong Trinh Manh DAN™ Damian Stewart @@ -95,8 +116,12 @@ Daniel Bevenius Daniel Drake Daniel Hiltgen Daniel Illescas Romero +Daniele <57776841+daniandtheweb@users.noreply.github.com> DannyDaemonic Dat Quoc Nguyen <2412555+datquocnguyen@users.noreply.github.com> +Dave +Dave Airlie +Dave Airlie Dave Della Costa David Friehs David Kennedy @@ -104,10 +129,13 @@ David Pflug David Renshaw David Sommers <12738+databyte@users.noreply.github.com> David Yang +Dawid Potocki Dawid Wysocki <62249621+TortillaZHawaii@users.noreply.github.com> Dean Deins +Deven Mistry <31466137+deven367@users.noreply.github.com> Didzis Gosko +Djip007 Don Mahurin DooWoong Lee (David) Doomsdayrs <38189170+Doomsdayrs@users.noreply.github.com> @@ -116,8 +144,11 @@ Dr. Tom Murphy VII Ph.D <499244+tom7@users.noreply.github.com> Ebey Abraham Ed Lee Ed Lepedus +Eddie-Wang Edward Taylor +Elaine Elbios <141279586+Elbios@users.noreply.github.com> +Elton Kola Engininja2 <139037756+Engininja2@users.noreply.github.com> Equim Eric Sommerlade @@ -143,37 +174,47 @@ Firat Folko-Ven <71110216+Folko-Ven@users.noreply.github.com> Foul-Tarnished <107711110+Foul-Tarnished@users.noreply.github.com> Francisco Melo <43780565+francis2tm@users.noreply.github.com> +Frank Mai FrankHB +Fred Douglas <43351173+fredlas@users.noreply.github.com> Frederik Vogel Gabe Goodhart GainLee Galunid Gary Linscott Gary Mulder +Gavin Zhao Genkagaku.GPT Georgi Gerganov Gilad S +Giuseppe Scrivano GiviMAD Govlzkoy Guillaume "Vermeille" Sanchez Guillaume Wenzek Guoteng <32697156+SolenoidWGT@users.noreply.github.com> Gustavo Rocha Dias <91472747+gustrd@users.noreply.github.com> +Haggai Nuchi Halalaluyafail3 <55773281+Halalaluyafail3@users.noreply.github.com> +Hamdoud Hakem <90524568+hamdoudhakem@users.noreply.github.com> +HanishKVC Haohui Mai Haoxiang Fei Harald Fernengel Hatsune Miku <129688334+at8u@users.noreply.github.com> +HatsuneMikuUwU33 <173229399+HatsuneMikuUwU33@users.noreply.github.com> Henk Poley Henri Vasserman Henrik Forstén Herman Semenov Hesen Peng Hoang Nguyen +Hong Bo PENG Hongyu Ouyang <96765450+casavaca@users.noreply.github.com> Howard Su Hua Jiang Huawei Lin +Hugo Roussel Ian Bull Ian Bull Ian Scrivener @@ -190,8 +231,10 @@ Ivan Stepanov JH23X <165871467+JH23X@users.noreply.github.com> Jack Mousseau JackJollimore <130917767+JackJollimore@users.noreply.github.com> +Jaemin Son Jag Chadha Jakub N +James A Capozzoli <157492257+jac-jim@users.noreply.github.com> James Reynolds Jan Boon Jan Boon @@ -205,12 +248,17 @@ Jean-Michaël Celerier Jed Fox Jeffrey Quesnelle Jesse Jojo Johnson +Jeximo Jhen-Jie Hong Jiahao Li Jian Liao JidongZhang-THU <1119708529@qq.com> Jinwoo Jeong <33892306+williamjeong2@users.noreply.github.com> Jiří Podivín <66251151+jpodivin@users.noreply.github.com> +Jiří Sejkora +Joan Fontanals +Joan Fontanals +Johan Johannes Gäßler Johannes Rudolph John <78893154+cmp-nct@users.noreply.github.com> @@ -221,15 +269,19 @@ Jonas Wunderlich <32615971+jonas-w@users.noreply.github.com> Jorge A <161275481+jorgealias@users.noreply.github.com> Jose Maldonado <63384398+yukiteruamano@users.noreply.github.com> Joseph Stahl <1269177+josephst@users.noreply.github.com> +Josh Ramer Joyce Juan Calderon-Perez <835733+gaby@users.noreply.github.com> Judd Julius Arkenberg Jun Jie <71215065+junnjiee16@users.noreply.github.com> +Junyang Lin Juraj Bednar Justin Parker Justin Suess +Justina Cho Justine Tunney +Justine Tunney Juuso Alasuutari KASR Kamil Tomšík @@ -242,6 +294,7 @@ Kawrakow <48489457+ikawrakow@users.noreply.github.com> Keiichi Tabata Kenvix ⭐ Kerfuffle <44031344+KerfuffleV2@users.noreply.github.com> +Kevin Gibbons Kevin Ji <1146876+kevinji@users.noreply.github.com> Kevin Kwok Kevin Lo @@ -257,6 +310,7 @@ Laura Lee <44310445+lx200916@users.noreply.github.com> Lee Drake Leng Yue +Leon Knauer LeonEricsson <70749762+LeonEricsson@users.noreply.github.com> Leonardo Neumann Li Tan @@ -265,20 +319,26 @@ LoganDark LostRuins <39025047+LostRuins@users.noreply.github.com> Luciano Luo Tian +Lyle Dean M. Yusuf Sarıgöz Maarten ter Huurne Mack Straight Maël Kerbiriou MaggotHATE +Manuel <44313466+makuche@users.noreply.github.com> Marc Köhlbrugge Marco Matthies <71844+marcom@users.noreply.github.com> Marcus Dunn <51931484+MarcusDunn@users.noreply.github.com> Marian Cepok Mark Fairbairn Marko Tasic +Markus Tavenrath +Martin Delille Martin Krasser Martin Schwaighofer Marvin Gießing +Masaya, Kato <62578291+msy-kato@users.noreply.github.com> +MasterYi1024 <39848311+MasterYi1024@users.noreply.github.com> Mateusz Charytoniuk Matheus C. França Matheus Gabriel Alves Silva @@ -287,8 +347,11 @@ Mathijs de Bruin Matt Clayton <156335168+mattjcly@users.noreply.github.com> Matt Pulver Matteo Boschini <12133566+mbosc@users.noreply.github.com> +Mattheus Chediak Matthew Tejo Matvey Soloviev +Max Krasnyansky +Max Krasnyansky Maxime <672982+maximegmd@users.noreply.github.com> Maximilian Winter Meng Zhang @@ -300,32 +363,41 @@ Michael Kesper Michael Klimenko Michael Podvitskiy Michael Potter +Michael de Gans Michaël de Vries Mihai Mike +Mikko Juola Minsoo Cheong <54794500+mscheong01@users.noreply.github.com> Mirko185 Mirror Azure <54669636+MirrorAzure@users.noreply.github.com> Miwa / Ensan <63481257+ensan-hcl@users.noreply.github.com> Mohammadreza Hendiani +Mohammadreza Hendiani Murilo Santana Musab Gultekin Nam D. Tran <42194884+namtranase@users.noreply.github.com> +Nathan Epstein NawafAlansari <72708095+NawafAlansari@users.noreply.github.com> Nebula +Neo Zhang <14088817+arthw@users.noreply.github.com> +Neo Zhang Neo Zhang Jianyu Neuman Vong Nexesenex <124105151+Nexesenex@users.noreply.github.com> Niall Coates <1349685+Niall-@users.noreply.github.com> Nicolai Weitkemper +Nicolás Pérez Nigel Bosch Niklas Korz +Nikolas <127742645+nneubacher@users.noreply.github.com> Nindaleth Oleksandr Nikitin Oleksii Maryshchenko Olivier Chafik Ondřej Čertík Ouadie EL FAROUKI +Patrice Ferlet Paul Tsochantaris Pavol Rusnak Pedro Cuenca @@ -343,9 +415,14 @@ RJ Adriaansen Radoslav Gerganov Radosław Gryta Rahul Vivek Nair <68507071+RahulVivekNair@users.noreply.github.com> +Raj Hammeer Singh Hada +Ralph Soika Rand Xie Randall Fitzgerald Reinforce-II +Ren Xuancheng +Rene Leonhardt <65483435+reneleonhardt@users.noreply.github.com> +RhinoDevel Riceball LEE Richard Kiss Richard Roberson @@ -373,6 +450,7 @@ Rowan Hart Rune <43761327+Rune-AI@users.noreply.github.com> Ryan Landay Ryder Wishart +Ryuei Rőczey Barnabás <31726601+An0nie@users.noreply.github.com> SakuraUmi Salvador E. Tropea @@ -386,6 +464,7 @@ SebastianApel <13675545+SebastianApel@users.noreply.github.com> Senemu <10880819+Senemu@users.noreply.github.com> Sergey Alirzaev Sergio López +Sertaç Özercan <852750+sozercan@users.noreply.github.com> SeungWon Jeong <65549245+redlion0929@users.noreply.github.com> ShadovvBeast Shakhar Dasgupta @@ -394,6 +473,7 @@ Shijie <821898965@qq.com> Shintarou Okada Shouzheng Liu <61452103+lshzh-ww@users.noreply.github.com> Shouzheng Liu +Shuichi Tsutsumi Sigbjørn Skjæret Simon Willison Siwen Yu @@ -405,11 +485,14 @@ Someone Someone Serge Sourab Mangrulkar <13534540+pacman100@users.noreply.github.com> Spencer Sutton +Srihari-mcw <96763064+Srihari-mcw@users.noreply.github.com> Srinivas Billa Stefan Sydow +Steffen Röcker Stephan Walter Stephen Nichols Steve Grubb +Steven Prichard Steven Roussey Steward Garcia <57494570+FSSRepo@users.noreply.github.com> Suaj Carrot <72162667+SuajCarrot@users.noreply.github.com> @@ -434,16 +517,19 @@ Tom C Tom Jobbins <784313+TheBloke@users.noreply.github.com> Tomas Tomáš Pazdiora +Tristan Druyen Tristan Ross Tungsten842 <886724vf@anonaddy.me> Tungsten842 Tushar UEXTM.com <84163508+uextm@users.noreply.github.com> +Ulrich Drepper Uzo Nweke Vaibhav Srivastav Val Kharitonov Valentin Konovalov Valentyn Bezshapkin <61702053+valentynbez@users.noreply.github.com> +Victor Nogueira Victor Z. Peng Vlad Vladimir @@ -455,7 +541,9 @@ Weird Constructor Welby Seely Wentai Zhang WillCorticesAI <150854901+WillCorticesAI@users.noreply.github.com> +William Tambellini Willy Tarreau +Wouter <9594229+DifferentialityDevelopment@users.noreply.github.com> Wu Jian Ping Wu Jian Ping Xiake Sun @@ -466,6 +554,8 @@ Xiaoyi Chen Xingchen Song(宋星辰) Xuan Son Nguyen Yann Follet <131855179+YannFollet@users.noreply.github.com> +Yaroslav +Yazan Agha-Schrader Yiming Cui Yishuo Wang Yueh-Po Peng <94939112+y10ab1@users.noreply.github.com> @@ -477,6 +567,7 @@ Zane Shannon Zay <95888118+isaiahbjork@users.noreply.github.com> Zenix Zhang Peiyuan +Zheng.Deng <32841220+dengzheng-cloud@users.noreply.github.com> ZhouYuChen Ziad Ben Hadj-Alouane Ziang Wu <97337387+ZiangWu-77@users.noreply.github.com> @@ -484,14 +575,18 @@ Zsapi a-n-n-a-l-e-e <150648636+a-n-n-a-l-e-e@users.noreply.github.com> adel boussaken afrideva <95653597+afrideva@users.noreply.github.com> +agray3 akawrykow <142945436+akawrykow@users.noreply.github.com> alexpinel <93524949+alexpinel@users.noreply.github.com> alonfaraj +alwqx +amd-lalithnc andrijdavid anon998 <131767832+anon998@users.noreply.github.com> anzz1 apaz apcameron <37645737+apcameron@users.noreply.github.com> +arch-btw <57669023+arch-btw@users.noreply.github.com> arcrank arlo-phoenix <140345165+arlo-phoenix@users.noreply.github.com> at8u <129688334+at8u@users.noreply.github.com> @@ -514,13 +609,17 @@ cocktailpeanut <121128867+cocktailpeanut@users.noreply.github.com> coezbek comex compilade <113953597+compilade@users.noreply.github.com> +compilade +cpumaxx <163466046+cpumaxx@users.noreply.github.com> crasm crasm daboe01 david raistrick +ddh0 ddpasa <112642920+ddpasa@users.noreply.github.com> deepdiffuser <112834445+deepdiffuser@users.noreply.github.com> divinity76 +dm4 dotpy314 <33351922+dotpy314@users.noreply.github.com> drbh ds5t5 <145942675+ds5t5@users.noreply.github.com> @@ -529,6 +628,7 @@ eastriver ebraminio eiery <19350831+eiery@users.noreply.github.com> eric8607242 +fairydreaming <166155368+fairydreaming@users.noreply.github.com> fraxy-v <65565042+fraxy-v@users.noreply.github.com> github-actions[bot] gliptic @@ -539,6 +639,7 @@ h-h-h-h <13482553+h-h-h-h@users.noreply.github.com> hankcs hoangmit hongbo.mo <352280764@qq.com> +hopkins385 <98618192+hopkins385@users.noreply.github.com> howlger howlger hutli <6594598+hutli@users.noreply.github.com> @@ -549,14 +650,22 @@ hydai iSma iacore <74560659+iacore@users.noreply.github.com> igarnier +intelmatt <61025942+intelmatt@users.noreply.github.com> iohub jacobi petrucciani <8117202+jpetrucciani@users.noreply.github.com> +jaime-m-p <167997752+jaime-m-p@users.noreply.github.com> jameswu2014 <545426914@qq.com> +jiez <373447296@qq.com> jneem +joecryptotoo <80373433+joecryptotoo@users.noreply.github.com> johnson442 <56517414+johnson442@users.noreply.github.com> +jojorne jon-chuang <9093549+jon-chuang@users.noreply.github.com> jp-x-g +jukofyork <69222624+jukofyork@users.noreply.github.com> +junchao-loongson <68935141+junchao-loongson@users.noreply.github.com> jwj7140 <32943891+jwj7140@users.noreply.github.com> +k.h.lai kaizau kalomaze <66376113+kalomaze@users.noreply.github.com> kang @@ -575,11 +684,15 @@ ldwang le.chang leejet limitedAtonement +liuwei-git <14815172+liuwei-git@users.noreply.github.com> lon <114724657+longregen@users.noreply.github.com> +loonerin <132926317+loonerin@users.noreply.github.com> +luoyu-intel m3ndax maddes8cht <55592906+maddes8cht@users.noreply.github.com> makomk manikbhandari +maor-ps <154728172+maor-ps@users.noreply.github.com> mdrokz mgroeber9110 <45620825+mgroeber9110@users.noreply.github.com> minarchist @@ -593,15 +706,19 @@ ngc92 <7938269+ngc92@users.noreply.github.com> nhamanasu <45545786+nhamanasu@users.noreply.github.com> niansa/tuxifan niansa/tuxifan +nickp27 ningshanwutuobang nold nopperl <54780682+nopperl@users.noreply.github.com> nusu-github <29514220+nusu-github@users.noreply.github.com> olexiyb +omahs <73983677+omahs@users.noreply.github.com> oobabooga <112222186+oobabooga@users.noreply.github.com> opparco ostix360 <55257054+ostix360@users.noreply.github.com> +pengxin99 perserk +pmysl postmasters pudepiedj qingfengfenga <41416092+qingfengfenga@users.noreply.github.com> @@ -614,16 +731,19 @@ rhuddleston rimoliga <53384203+rimoliga@users.noreply.github.com> runfuture sandyiscool +sasha0552 semidark sharpHL <132747147+sharpHL@users.noreply.github.com> shibe2 singularity <12184989+singularity-s0@users.noreply.github.com> sjinzh +sjxx <63994076+ylsdamxssjxxdd@users.noreply.github.com> slaren <2141330+slaren@users.noreply.github.com> slaren snadampal <87143774+snadampal@users.noreply.github.com> staviq stduhpf +strawberrymelonpanda <152940198+strawberrymelonpanda@users.noreply.github.com> swittk takov751 <40316768+takov751@users.noreply.github.com> tarcey @@ -636,12 +756,16 @@ uint256_t uint256_t unbounded valiray <133289098+valiray@users.noreply.github.com> +vik +viric vodkaslime <646329483@qq.com> vvhg1 <94630311+vvhg1@users.noreply.github.com> vxiiduu <73044267+vxiiduu@users.noreply.github.com> wbpxre150 <100937007+wbpxre150@users.noreply.github.com> whoreson <139810751+whoreson@users.noreply.github.com> +woachk <24752637+woachk@users.noreply.github.com> wonjun Jang +woodx <124784234+woodx9@users.noreply.github.com> wzy <32936898+Freed-Wu@users.noreply.github.com> xaedes xaedes @@ -649,7 +773,10 @@ xloem <0xloem@gmail.com> yangli2 yuiseki zakkor +zhangkaihuo zhouwg <6889919+zhouwg@users.noreply.github.com> +zhouwg zrm +Ștefan-Gabriel Muscalu 源文雨 <41315874+fumiama@users.noreply.github.com> Нияз Гарифзянов <112617865+garrnizon@users.noreply.github.com> diff --git a/common/common.cpp b/common/common.cpp index e1931b115..486ef15b2 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -1027,6 +1027,10 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa params.input_suffix = argv[i]; return true; } + if (arg == "--spm-infill") { + params.spm_infill = true; + return true; + } if (arg == "--grammar") { CHECK_ARG sparams.grammar = argv[i]; @@ -1410,6 +1414,8 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param options.push_back({ "main infill", " --in-prefix-bos", "prefix BOS to user inputs, preceding the `--in-prefix` string" }); options.push_back({ "main infill", " --in-prefix STRING", "string to prefix user inputs with (default: empty)" }); options.push_back({ "main infill", " --in-suffix STRING", "string to suffix after user inputs with (default: empty)" }); + options.push_back({ "server infill", + " --spm-infill", "use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. (default: %s)", params.spm_infill ? "enabled" : "disabled" }); options.push_back({ "sampling" }); options.push_back({ "*", " --samplers SAMPLERS", "samplers that will be used for generation in the order, separated by \';\'\n" @@ -2619,6 +2625,7 @@ std::string llama_chat_apply_template(const struct llama_model * model, const std::vector & msgs, bool add_ass) { int alloc_size = 0; + bool fallback = false; // indicate if we must fallback to default chatml std::vector chat; for (auto & msg : msgs) { chat.push_back({msg.role.c_str(), msg.content.c_str()}); @@ -2631,10 +2638,26 @@ std::string llama_chat_apply_template(const struct llama_model * model, // run the first time to get the total output length int32_t res = llama_chat_apply_template(model, ptr_tmpl, chat.data(), chat.size(), add_ass, buf.data(), buf.size()); + // error: chat template is not supported + if (res < 0) { + if (ptr_tmpl != nullptr) { + // if the custom "tmpl" is not supported, we throw an error + // this is a bit redundant (for good), since we're not sure if user validated the custom template with llama_chat_verify_template() + throw std::runtime_error("this custom template is not supported"); + } else { + // If the built-in template is not supported, we default to chatml + res = llama_chat_apply_template(nullptr, "chatml", chat.data(), chat.size(), add_ass, buf.data(), buf.size()); + fallback = true; + } + } + // if it turns out that our buffer is too small, we resize it if ((size_t) res > buf.size()) { buf.resize(res); - res = llama_chat_apply_template(model, ptr_tmpl, chat.data(), chat.size(), add_ass, buf.data(), buf.size()); + res = llama_chat_apply_template( + fallback ? nullptr : model, + fallback ? "chatml" : ptr_tmpl, + chat.data(), chat.size(), add_ass, buf.data(), buf.size()); } std::string formatted_chat(buf.data(), res); @@ -2805,125 +2828,87 @@ float llama_embd_similarity_cos(const float * embd1, const float * embd2, int n) // static llama_control_vector_data llama_control_vector_load_one(const llama_control_vector_load_info & load_info) { - int32_t n_tensors; - - size_t n_bytes = 0; - - uint32_t max_direction_layer = 0; - llama_control_vector_data result = { -1, {} }; - // calculate size of ctx needed for tensors, ensure tensors are f32, and find max layer - { - struct ggml_init_params meta_params = { - /* .mem_size = */ ggml_tensor_overhead() * 128 + ggml_graph_overhead(), - /* .mem_buffer = */ nullptr, - /* .no_alloc = */ true, - }; - ggml_context * meta_ctx = ggml_init(meta_params); - struct gguf_init_params meta_gguf_params = { - /* .no_alloc = */ true, - /* .ctx = */ &meta_ctx, - }; - struct gguf_context * meta_ctx_gguf = gguf_init_from_file(load_info.fname.c_str(), meta_gguf_params); - if (!meta_ctx_gguf) { - fprintf(stderr, "%s: failed to load control vector from %s\n", __func__, load_info.fname.c_str()); - ggml_free(meta_ctx); - return result; - } - - n_tensors = gguf_get_n_tensors(meta_ctx_gguf); - for (int i = 0; i < n_tensors; i++) { - std::string name = gguf_get_tensor_name(meta_ctx_gguf, i); - - // split on '.' - size_t dotpos = name.find('.'); - if (dotpos != std::string::npos && name.substr(0, dotpos) == "direction") { - try { - uint32_t layer = std::stoi(name.substr(dotpos + 1)); - if (layer == 0) { - fprintf(stderr, "%s: direction tensor invalid in %s\n", __func__, load_info.fname.c_str()); - ggml_free(meta_ctx); - gguf_free(meta_ctx_gguf); - return result; - } - if (layer > max_direction_layer) { - max_direction_layer = layer; - } - } catch (...) { - fprintf(stderr, "%s: direction tensor invalid in %s\n", __func__, load_info.fname.c_str()); - ggml_free(meta_ctx); - gguf_free(meta_ctx_gguf); - return result; - } - } - - struct ggml_tensor * tensor_meta = ggml_get_tensor(meta_ctx, name.c_str()); - if (tensor_meta->type != GGML_TYPE_F32 || ggml_n_dims(tensor_meta) != 1) { - fprintf(stderr, "%s: direction tensor invalid in %s\n", __func__, load_info.fname.c_str()); - ggml_free(meta_ctx); - gguf_free(meta_ctx_gguf); - return result; - } - if (result.n_embd == -1) { - result.n_embd = ggml_nelements(tensor_meta); - } else if (ggml_nelements(tensor_meta) != result.n_embd) { - fprintf(stderr, "%s: direction tensor sizes mismatched in %s\n", __func__, load_info.fname.c_str()); - ggml_free(meta_ctx); - gguf_free(meta_ctx_gguf); - return result; - } - n_bytes += ggml_nbytes(tensor_meta); - } - ggml_free(meta_ctx); - gguf_free(meta_ctx_gguf); + ggml_context * ctx = nullptr; + struct gguf_init_params meta_gguf_params = { + /* .no_alloc = */ false, + /* .ctx = */ &ctx, + }; + struct gguf_context * ctx_gguf = gguf_init_from_file(load_info.fname.c_str(), meta_gguf_params); + if (!ctx_gguf) { + fprintf(stderr, "%s: failed to load control vector file from %s\n", __func__, load_info.fname.c_str()); + return result; } + int32_t n_tensors = gguf_get_n_tensors(ctx_gguf); if (n_tensors == 0) { fprintf(stderr, "%s: no direction tensors found in %s\n", __func__, load_info.fname.c_str()); - return result; } - // load and scale tensors into final control vector context - struct ggml_init_params ggml_params = { - /* .mem_size = */ ggml_tensor_overhead() * n_tensors + n_bytes, - /* .mem_buffer = */ nullptr, - /* .no_alloc = */ false, - }; - struct ggml_context * ctx = ggml_init(ggml_params); + for (int i = 0; i < n_tensors; i++) { + std::string name = gguf_get_tensor_name(ctx_gguf, i); - struct gguf_init_params params = { - /*.no_alloc = */ false, - /*.ctx = */ &ctx, - }; - struct gguf_context * ctx_gguf = gguf_init_from_file(load_info.fname.c_str(), params); - if (!ctx_gguf) { - fprintf(stderr, "%s: failed to load control vector from %s\n", __func__, load_info.fname.c_str()); - ggml_free(ctx); - return result; - } + int layer_idx = -1; - // do not store data for layer 0 (it's not used) - result.data.resize(result.n_embd * max_direction_layer); - - for (uint32_t il = 1; il <= max_direction_layer; il++) { - const std::string name = "direction." + std::to_string(il); - const ggml_tensor * tensor = ggml_get_tensor(ctx, name.c_str()); - - float * dst = result.data.data() + result.n_embd * (il - 1); - - if (tensor) { - const float * src = (const float *) tensor->data; - for (int j = 0; j < result.n_embd; j++) { - dst[j] = src[j] * load_info.strength; - } - } else { - for (int j = 0; j < result.n_embd; j++) { - dst[j] = 0.0f; + // split on '.' + size_t dotpos = name.find('.'); + if (dotpos != std::string::npos && name.substr(0, dotpos) == "direction") { + try { + layer_idx = std::stoi(name.substr(dotpos + 1)); + } catch (...) { + layer_idx = -1; } } + if (layer_idx < 0) { + fprintf(stderr, "%s: invalid/unparsable direction tensor layer index in %s\n", __func__, load_info.fname.c_str()); + result.n_embd = -1; + break; + } else if (layer_idx == 0) { + fprintf(stderr, "%s: invalid (zero) direction tensor layer index in %s\n", __func__, load_info.fname.c_str()); + result.n_embd = -1; + break; + } + + struct ggml_tensor * tensor = ggml_get_tensor(ctx, name.c_str()); + if (tensor->type != GGML_TYPE_F32) { + fprintf(stderr, "%s: invalid (non-F32) direction tensor type in %s\n", __func__, load_info.fname.c_str()); + result.n_embd = -1; + break; + } + if (ggml_n_dims(tensor) != 1) { + fprintf(stderr, "%s: invalid (non-1D) direction tensor shape in %s\n", __func__, load_info.fname.c_str()); + result.n_embd = -1; + break; + } + + if (result.n_embd == -1) { + result.n_embd = ggml_nelements(tensor); + } else if (ggml_nelements(tensor) != result.n_embd) { + fprintf(stderr, "%s: direction tensor in %s does not match previous dimensions\n", __func__, load_info.fname.c_str()); + result.n_embd = -1; + break; + } + + // extend if necessary - do not store data for layer 0 (it's not used) + result.data.resize(std::max(result.data.size(), static_cast(result.n_embd * layer_idx)), 0.0f); + + const float * src = (const float *) tensor->data; + float * dst = result.data.data() + result.n_embd * (layer_idx - 1); // layer 1 at [0] + for (int j = 0; j < result.n_embd; j++) { + dst[j] += src[j] * load_info.strength; // allows multiple directions for same layer in same file + } + } + if (result.n_embd == -1) { + fprintf(stderr, "%s: skipping %s due to invalid direction tensors\n", __func__, load_info.fname.c_str()); + result.data.clear(); + } + + gguf_free(ctx_gguf); + ggml_free(ctx); + return result; } @@ -2934,16 +2919,19 @@ llama_control_vector_data llama_control_vector_load(const std::vector & chat, diff --git a/common/json-schema-to-grammar.cpp b/common/json-schema-to-grammar.cpp index 2f233e2e7..881eb49e3 100644 --- a/common/json-schema-to-grammar.cpp +++ b/common/json-schema-to-grammar.cpp @@ -316,7 +316,7 @@ std::unordered_map GRAMMAR_LITERAL_ESCAPES = { }; std::unordered_set NON_LITERAL_SET = {'|', '.', '(', ')', '[', ']', '{', '}', '*', '+', '?'}; -std::unordered_set ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = {'[', ']', '(', ')', '|', '{', '}', '*', '+', '?'}; +std::unordered_set ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = {'^', '$', '.', '[', ']', '(', ')', '|', '{', '}', '*', '+', '?'}; template std::string join(Iterator begin, Iterator end, const std::string & separator) { @@ -720,7 +720,7 @@ private: } prop_names.push_back(prop_name); } - if (!(additional_properties.is_boolean() && !additional_properties.get())) { + if ((additional_properties.is_boolean() && additional_properties.get()) || additional_properties.is_object()) { std::string sub_name = name + (name.empty() ? "" : "-") + "additional"; std::string value_rule = additional_properties.is_object() ? visit(additional_properties, sub_name + "-value") diff --git a/convert-hf-to-gguf-update.py b/convert-hf-to-gguf-update.py index 67598b561..2758214fa 100755 --- a/convert-hf-to-gguf-update.py +++ b/convert-hf-to-gguf-update.py @@ -85,6 +85,7 @@ models = [ {"name": "smaug-bpe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/abacusai/Smaug-Llama-3-70B-Instruct", }, {"name": "poro-chat", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LumiOpen/Poro-34B-chat", }, {"name": "jina-v2-code", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-code", }, + {"name": "viking", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LumiOpen/Viking-7B", }, # Also used for Viking 13B and 33B ] diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index c26fad930..5bcc849db 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -487,6 +487,9 @@ class Model: if chkhsh == "7967bfa498ade6b757b064f31e964dddbb80f8f9a4d68d4ba7998fcf281c531a": # ref: https://huggingface.co/jinaai/jina-embeddings-v2-base-code res = "jina-v2-code" + if chkhsh == "7fc505bd3104ca1083b150b17d088b59534ede9bde81f0dd2090967d7fe52cee": + # ref: https://huggingface.co/LumiOpen/Viking-7B + res = "viking" if res is None: logger.warning("\n") @@ -2337,6 +2340,46 @@ class GemmaModel(Model): return [(self.map_tensor_name(name), data_torch)] +@Model.register("Gemma2ForCausalLM") +class Gemma2Model(Model): + model_arch = gguf.MODEL_ARCH.GEMMA2 + + def set_vocab(self): + self._set_vocab_llama_hf() + self.gguf_writer.add_add_space_prefix(False) + + def set_gguf_parameters(self): + hparams = self.hparams + block_count = hparams["num_hidden_layers"] + + self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name) + self.gguf_writer.add_context_length(hparams["max_position_embeddings"]) + self.gguf_writer.add_embedding_length(hparams["hidden_size"]) + self.gguf_writer.add_block_count(block_count) + self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"]) + self.gguf_writer.add_head_count(hparams["num_attention_heads"]) + self.gguf_writer.add_head_count_kv(self.hparams["num_key_value_heads"] if "num_key_value_heads" in hparams else hparams["num_attention_heads"]) + self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"]) + self.gguf_writer.add_key_length(hparams["head_dim"]) + self.gguf_writer.add_value_length(hparams["head_dim"]) + self.gguf_writer.add_file_type(self.ftype) + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + del bid # unusem + + # lm_head is not used in llama.cpp, while autoawq will include this tensor in model + # To prevent errors, skip loading lm_head.weight. + if name == "lm_head.weight": + logger.debug(f"Skipping get tensor {name!r} in safetensors so that convert can end normally.") + return [] + + # ref: https://github.com/huggingface/transformers/blob/fc37f38915372c15992b540dfcbbe00a916d4fc6/src/transformers/models/gemma/modeling_gemma.py#L89 + if name.endswith("norm.weight"): + data_torch = data_torch + 1 + + return [(self.map_tensor_name(name), data_torch)] + + @Model.register("Starcoder2ForCausalLM") class StarCoder2Model(Model): model_arch = gguf.MODEL_ARCH.STARCODER2 diff --git a/examples/infill/README.md b/examples/infill/README.md index 74f42d2fc..810a0c5e7 100644 --- a/examples/infill/README.md +++ b/examples/infill/README.md @@ -15,6 +15,7 @@ In this section, we cover the most commonly used options for running the `infill - `-i, --interactive`: Run the program in interactive mode, allowing you to provide input directly and receive real-time responses. - `-n N, --n-predict N`: Set the number of tokens to predict when generating text. Adjusting this value can influence the length of the generated text. - `-c N, --ctx-size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference. +- `--spm-infill`: Use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. ## Input Prompts diff --git a/examples/infill/infill.cpp b/examples/infill/infill.cpp index 64c6ab9cf..1556a2fb7 100644 --- a/examples/infill/infill.cpp +++ b/examples/infill/infill.cpp @@ -211,6 +211,7 @@ int main(int argc, char ** argv) { suff_rm_leading_spc = false; } std::vector embd_inp; + std::vector embd_end; std::vector inp_pfx = ::llama_tokenize(ctx, params.input_prefix, false); std::vector inp_sfx = ::llama_tokenize(ctx, params.input_suffix, false); const int space_token = 29871; @@ -218,12 +219,13 @@ int main(int argc, char ** argv) { inp_sfx.erase(inp_sfx.begin()); } inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(model)); - if (add_bos) { - inp_pfx.insert(inp_pfx.begin(), llama_token_bos(model)); - } inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model)); - embd_inp = inp_pfx; - embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end()); + embd_inp = params.spm_infill ? inp_sfx : inp_pfx; + embd_end = params.spm_infill ? inp_pfx : inp_sfx; + if (add_bos) { + embd_inp.insert(embd_inp.begin(), llama_token_bos(model)); + } + embd_inp.insert(embd_inp.end(), embd_end.begin(), embd_end.end()); const llama_token middle_token = llama_token_middle(model); if (middle_token >= 0) { @@ -527,14 +529,14 @@ int main(int argc, char ** argv) { inp_sfx.erase(inp_sfx.begin()); } inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(model)); - if (add_bos) { - inp_pfx.insert(inp_pfx.begin(), llama_token_bos(model)); - } inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model)); - embd_inp = inp_pfx; - embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end()); + embd_inp = params.spm_infill ? inp_sfx : inp_pfx; + embd_end = params.spm_infill ? inp_pfx : inp_sfx; + if (add_bos) { + embd_inp.insert(embd_inp.begin(), llama_token_bos(model)); + } + embd_inp.insert(embd_inp.end(), embd_end.begin(), embd_end.end()); - const llama_token middle_token = llama_token_middle(model); if (middle_token >= 0) { embd_inp.push_back(middle_token); } diff --git a/examples/json_schema_to_grammar.py b/examples/json_schema_to_grammar.py index 92f6e3d47..072a230f7 100755 --- a/examples/json_schema_to_grammar.py +++ b/examples/json_schema_to_grammar.py @@ -231,7 +231,7 @@ GRAMMAR_RANGE_LITERAL_ESCAPE_RE = re.compile(r'[\r\n"\]\-\\]') GRAMMAR_LITERAL_ESCAPES = {'\r': '\\r', '\n': '\\n', '"': '\\"', '-': '\\-', ']': '\\]'} NON_LITERAL_SET = set('|.()[]{}*+?') -ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = set('[]()|{}*+?') +ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = set('^$.[]()|{}*+?') class SchemaConverter: @@ -602,7 +602,7 @@ class SchemaConverter: else: add_component(t, is_required=True) - return self._add_rule(rule_name, self._build_object_rule(properties, required, hybrid_name, additional_properties=[])) + return self._add_rule(rule_name, self._build_object_rule(properties, required, hybrid_name, additional_properties=None)) elif schema_type in (None, 'array') and ('items' in schema or 'prefixItems' in schema): items = schema.get('items') or schema['prefixItems'] @@ -691,7 +691,7 @@ class SchemaConverter: required_props = [k for k in sorted_props if k in required] optional_props = [k for k in sorted_props if k not in required] - if additional_properties != False: + if additional_properties is not None and additional_properties != False: sub_name = f'{name}{"-" if name else ""}additional' value_rule = self.visit(additional_properties, f'{sub_name}-value') if isinstance(additional_properties, dict) else \ self._add_primitive('value', PRIMITIVE_RULES['value']) diff --git a/examples/llama.android/llama/src/main/cpp/llama-android.cpp b/examples/llama.android/llama/src/main/cpp/llama-android.cpp index 874158ef0..92a6b16b1 100644 --- a/examples/llama.android/llama/src/main/cpp/llama-android.cpp +++ b/examples/llama.android/llama/src/main/cpp/llama-android.cpp @@ -5,7 +5,7 @@ #include #include #include "llama.h" -#include "common/common.h" +#include "common.h" // Write C++ code here. // diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp index 95fbe3d02..d6882eec3 100644 --- a/examples/llava/clip.cpp +++ b/examples/llava/clip.cpp @@ -1121,20 +1121,20 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { } if (n < 32) hparams.image_grid_pinpoints[n] = 0; - } catch (std::runtime_error & e) { + } catch (std::runtime_error & /*e*/) { hparams.image_grid_pinpoints[0]=0; } try { int idx = get_key_idx(ctx, KEY_MM_PATCH_MERGE_TYPE); strcpy(hparams.mm_patch_merge_type, gguf_get_val_str(ctx, idx)); - } catch (std::runtime_error & e) { + } catch (std::runtime_error & /*e*/) { strcpy(hparams.mm_patch_merge_type, "flat"); } try { hparams.image_crop_resolution = get_u32(ctx, KEY_IMAGE_CROP_RESOLUTION); // llava-1.6 - } catch(const std::exception& e) { + } catch(const std::exception& /*e*/) { hparams.image_crop_resolution = hparams.image_size; } @@ -1173,7 +1173,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { try { vision_model.class_embedding = get_tensor(new_clip->ctx_data, TN_CLASS_EMBD); new_clip->has_class_embedding = true; - } catch (const std::exception& e) { + } catch (const std::exception& /*e*/) { new_clip->has_class_embedding = false; } @@ -1181,7 +1181,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { vision_model.pre_ln_w = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "weight")); vision_model.pre_ln_b = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "bias")); new_clip->has_pre_norm = true; - } catch (std::exception & e) { + } catch (std::exception & /*e*/) { new_clip->has_pre_norm = false; } @@ -1189,21 +1189,21 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { vision_model.post_ln_w = get_tensor(new_clip->ctx_data, format(TN_LN_POST, "v", "weight")); vision_model.post_ln_b = get_tensor(new_clip->ctx_data, format(TN_LN_POST, "v", "bias")); new_clip->has_post_norm = true; - } catch (std::exception & e) { + } catch (std::exception & /*e*/) { new_clip->has_post_norm = false; } try { vision_model.patch_bias = get_tensor(new_clip->ctx_data, TN_PATCH_BIAS); new_clip->has_patch_bias = true; - } catch (std::exception & e) { + } catch (std::exception & /*e*/) { new_clip->has_patch_bias = false; } try { vision_model.patch_embeddings = get_tensor(new_clip->ctx_data, TN_PATCH_EMBD); vision_model.position_embeddings = get_tensor(new_clip->ctx_data, format(TN_POS_EMBD, "v")); - } catch(const std::exception& e) { + } catch(const std::exception& /*e*/) { LOG_TEE("%s: failed to load vision model tensors\n", __func__); } @@ -1215,26 +1215,26 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { // Yi-type llava vision_model.mm_1_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 1, "weight")); vision_model.mm_1_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 1, "bias")); - } catch (std::runtime_error & e) { } + } catch (std::runtime_error & /*e*/) { } try { // missing in Yi-type llava vision_model.mm_2_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "weight")); vision_model.mm_2_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "bias")); - } catch (std::runtime_error & e) { } + } catch (std::runtime_error & /*e*/) { } try { // Yi-type llava vision_model.mm_3_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 3, "weight")); vision_model.mm_3_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 3, "bias")); - } catch (std::runtime_error & e) { } + } catch (std::runtime_error & /*e*/) { } try { // Yi-type llava vision_model.mm_4_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 4, "weight")); vision_model.mm_4_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 4, "bias")); - } catch (std::runtime_error & e) { } + } catch (std::runtime_error & /*e*/) { } try { vision_model.image_newline = get_tensor(new_clip->ctx_data, TN_IMAGE_NEWLINE); // LOG_TEE("%s: image_newline tensor (llava-1.6) found\n", __func__); - } catch (std::runtime_error & e) { } + } catch (std::runtime_error & /*e*/) { } } else if (new_clip->proj_type == PROJECTOR_TYPE_LDP) { // MobileVLM projection vision_model.mm_model_mlp_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 1, "weight")); diff --git a/examples/main/main.cpp b/examples/main/main.cpp index 9d46a0760..42ce45d53 100644 --- a/examples/main/main.cpp +++ b/examples/main/main.cpp @@ -811,7 +811,7 @@ int main(int argc, char ** argv) { is_antiprompt = true; } - chat_add_and_format(model, chat_msgs, "system", assistant_ss.str()); + chat_add_and_format(model, chat_msgs, "assistant", assistant_ss.str()); is_interacting = true; printf("\n"); } diff --git a/examples/server/README.md b/examples/server/README.md index e7fb0bf64..4fab006bb 100644 --- a/examples/server/README.md +++ b/examples/server/README.md @@ -73,6 +73,7 @@ The project is under active development, and we are [looking for feedback and co - `-fa`, `--flash-attn` : enable flash attention (default: disabled). - `-ctk TYPE`, `--cache-type-k TYPE` : KV cache data type for K (default: `f16`, options `f32`, `f16`, `q8_0`, `q4_0`, `q4_1`, `iq4_nl`, `q5_0`, or `q5_1`) - `-ctv TYPE`, `--cache-type-v TYPE` : KV cache type for V (default `f16`, see `-ctk` for options) +- `--spm-infill` : Use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. **If compiled with `LLAMA_SERVER_SSL=ON`** - `--ssl-key-file FNAME`: path to file a PEM-encoded SSL private key diff --git a/examples/server/public/json-schema-to-grammar.mjs b/examples/server/public/json-schema-to-grammar.mjs index 06d76edde..7267f3f9c 100644 --- a/examples/server/public/json-schema-to-grammar.mjs +++ b/examples/server/public/json-schema-to-grammar.mjs @@ -259,7 +259,7 @@ const GRAMMAR_RANGE_LITERAL_ESCAPE_RE = /[\n\r"\]\-\\]/g; const GRAMMAR_LITERAL_ESCAPES = { '\r': '\\r', '\n': '\\n', '"': '\\"', '-': '\\-', ']': '\\]' }; const NON_LITERAL_SET = new Set('|.()[]{}*+?'); -const ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = new Set('[]()|{}*+?'); +const ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = new Set('^$.[]()|{}*+?'); export class SchemaConverter { constructor(options) { @@ -751,7 +751,7 @@ export class SchemaConverter { const requiredProps = sortedProps.filter(k => required.has(k)); const optionalProps = sortedProps.filter(k => !required.has(k)); - if (additionalProperties !== false) { + if (additionalProperties) { const subName = `${name ?? ''}${name ? '-' : ''}additional`; const valueRule = additionalProperties != null && typeof additionalProperties === 'object' ? this.visit(additionalProperties, `${subName}-value`) diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 54c60d9b9..e03c8de21 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -2021,6 +2021,7 @@ struct server_context { slot.t_start_generation = 0; if (slot.infill) { + const bool add_bos = llama_should_add_bos_token(model); bool suff_rm_leading_spc = true; if (params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1) { params.input_suffix.erase(0, 1); @@ -2036,16 +2037,21 @@ struct server_context { } prefix_tokens.insert(prefix_tokens.begin(), llama_token_prefix(model)); - prefix_tokens.insert(prefix_tokens.begin(), llama_token_bos(model)); // always add BOS - prefix_tokens.insert(prefix_tokens.end(), llama_token_suffix(model)); - prefix_tokens.insert(prefix_tokens.end(), suffix_tokens.begin(), suffix_tokens.end()); + suffix_tokens.insert(suffix_tokens.begin(), llama_token_suffix(model)); + + auto embd_inp = params.spm_infill ? suffix_tokens : prefix_tokens; + auto embd_end = params.spm_infill ? prefix_tokens : suffix_tokens; + if (add_bos) { + embd_inp.insert(embd_inp.begin(), llama_token_bos(model)); + } + embd_inp.insert(embd_inp.end(), embd_end.begin(), embd_end.end()); const llama_token middle_token = llama_token_middle(model); if (middle_token >= 0) { - prefix_tokens.push_back(middle_token); + embd_inp.push_back(middle_token); } - prompt_tokens = prefix_tokens; + prompt_tokens = embd_inp; } else { prompt_tokens = tokenize(slot.prompt, system_prompt.empty()); // add BOS if there isn't system prompt } diff --git a/ggml/src/ggml-cuda/mmq.cuh b/ggml/src/ggml-cuda/mmq.cuh index f97bbd6e7..0b357fc0a 100644 --- a/ggml/src/ggml-cuda/mmq.cuh +++ b/ggml/src/ggml-cuda/mmq.cuh @@ -2476,7 +2476,7 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a const dim3 block_nums_mmq(nsm, 1, 1); - ggml_cuda_pool & pool = ctx.pool(); + ggml_cuda_pool & pool = ctx.pool(id); ggml_cuda_pool_alloc tmp_fixup(pool, block_nums_mmq.x * mmq_x*mmq_y); if (args.ne01 % mmq_y == 0) { diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index 222a2d137..cf3d09e70 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -150,6 +150,7 @@ class MODEL_ARCH(IntEnum): INTERNLM2 = auto() MINICPM = auto() GEMMA = auto() + GEMMA2 = auto() STARCODER2 = auto() MAMBA = auto() XVERSE = auto() @@ -180,10 +181,13 @@ class MODEL_TENSOR(IntEnum): ATTN_NORM = auto() ATTN_NORM_2 = auto() ATTN_OUT_NORM = auto() + ATTN_POST_NORM = auto() ATTN_ROT_EMBD = auto() FFN_GATE_INP = auto() FFN_GATE_INP_SHEXP = auto() FFN_NORM = auto() + FFN_PRE_NORM = auto() + FFN_POST_NORM = auto() FFN_GATE = auto() FFN_DOWN = auto() FFN_UP = auto() @@ -270,6 +274,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.INTERNLM2: "internlm2", MODEL_ARCH.MINICPM: "minicpm", MODEL_ARCH.GEMMA: "gemma", + MODEL_ARCH.GEMMA2: "gemma2", MODEL_ARCH.STARCODER2: "starcoder2", MODEL_ARCH.MAMBA: "mamba", MODEL_ARCH.XVERSE: "xverse", @@ -303,9 +308,12 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = { MODEL_TENSOR.ATTN_Q_NORM: "blk.{bid}.attn_q_norm", MODEL_TENSOR.ATTN_K_NORM: "blk.{bid}.attn_k_norm", MODEL_TENSOR.ATTN_OUT_NORM: "blk.{bid}.attn_output_norm", + MODEL_TENSOR.ATTN_POST_NORM: "blk.{bid}.post_attention_norm", MODEL_TENSOR.FFN_GATE_INP: "blk.{bid}.ffn_gate_inp", MODEL_TENSOR.FFN_GATE_INP_SHEXP: "blk.{bid}.ffn_gate_inp_shexp", MODEL_TENSOR.FFN_NORM: "blk.{bid}.ffn_norm", + MODEL_TENSOR.FFN_PRE_NORM: "blk.{bid}.ffn_norm", + MODEL_TENSOR.FFN_POST_NORM: "blk.{bid}.post_ffw_norm", MODEL_TENSOR.FFN_GATE: "blk.{bid}.ffn_gate", MODEL_TENSOR.FFN_DOWN: "blk.{bid}.ffn_down", MODEL_TENSOR.FFN_UP: "blk.{bid}.ffn_up", @@ -751,6 +759,21 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_UP, MODEL_TENSOR.FFN_NORM, ], + MODEL_ARCH.GEMMA2: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_POST_NORM, + MODEL_TENSOR.FFN_PRE_NORM, + MODEL_TENSOR.FFN_POST_NORM, + ], MODEL_ARCH.STARCODER2: [ MODEL_TENSOR.TOKEN_EMBD, MODEL_TENSOR.OUTPUT_NORM, diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py index 7b047f241..0bed43939 100644 --- a/gguf-py/gguf/tensor_mapping.py +++ b/gguf-py/gguf/tensor_mapping.py @@ -187,6 +187,10 @@ class TensorNameMap: "transformer.blocks.{bid}.norm_attn_norm.norm_2", # dbrx ), + MODEL_TENSOR.ATTN_POST_NORM: ( + "model.layers.{bid}.post_attention_layernorm", # gemma2 + ), + # Rotary embeddings MODEL_TENSOR.ATTN_ROT_EMBD: ( "model.layers.{bid}.self_attn.rotary_emb.inv_freq", # llama-hf @@ -210,6 +214,16 @@ class TensorNameMap: "transformer.decoder_layer.{bid}.rms_norm_2", # Grok ), + # Post feed-forward norm + MODEL_TENSOR.FFN_PRE_NORM: ( + "model.layers.{bid}.pre_feedforward_layernorm", # gemma2 + ), + + # Post feed-forward norm + MODEL_TENSOR.FFN_POST_NORM: ( + "model.layers.{bid}.post_feedforward_layernorm", # gemma2 + ), + MODEL_TENSOR.FFN_GATE_INP: ( "layers.{bid}.feed_forward.gate", # mixtral "model.layers.{bid}.block_sparse_moe.gate", # mixtral diff --git a/include/llama.h b/include/llama.h index f626af3e9..8a2e262eb 100644 --- a/include/llama.h +++ b/include/llama.h @@ -88,6 +88,7 @@ extern "C" { LLAMA_VOCAB_PRE_TYPE_DBRX = 13, LLAMA_VOCAB_PRE_TYPE_SMAUG = 14, LLAMA_VOCAB_PRE_TYPE_PORO = 15, + LLAMA_VOCAB_PRE_TYPE_VIKING = 16, }; // note: these values should be synchronized with ggml_rope diff --git a/src/llama.cpp b/src/llama.cpp index 78218897a..5ccdc8054 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -241,6 +241,7 @@ enum llm_arch { LLM_ARCH_INTERNLM2, LLM_ARCH_MINICPM, LLM_ARCH_GEMMA, + LLM_ARCH_GEMMA2, LLM_ARCH_STARCODER2, LLM_ARCH_MAMBA, LLM_ARCH_XVERSE, @@ -281,6 +282,7 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_INTERNLM2, "internlm2" }, { LLM_ARCH_MINICPM, "minicpm" }, { LLM_ARCH_GEMMA, "gemma" }, + { LLM_ARCH_GEMMA2, "gemma2" }, { LLM_ARCH_STARCODER2, "starcoder2" }, { LLM_ARCH_MAMBA, "mamba" }, { LLM_ARCH_XVERSE, "xverse" }, @@ -502,10 +504,12 @@ enum llm_tensor { LLM_TENSOR_ATTN_NORM, LLM_TENSOR_ATTN_NORM_2, LLM_TENSOR_ATTN_OUT_NORM, + LLM_TENSOR_ATTN_POST_NORM, LLM_TENSOR_ATTN_ROT_EMBD, LLM_TENSOR_FFN_GATE_INP, LLM_TENSOR_FFN_GATE_INP_SHEXP, LLM_TENSOR_FFN_NORM, + LLM_TENSOR_FFN_POST_NORM, LLM_TENSOR_FFN_GATE, LLM_TENSOR_FFN_DOWN, LLM_TENSOR_FFN_UP, @@ -1028,6 +1032,24 @@ static const std::map> LLM_TENSOR_NA { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, }, }, + { + LLM_ARCH_GEMMA2, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_ATTN_POST_NORM, "blk.%d.post_attention_norm" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + { LLM_TENSOR_FFN_POST_NORM, "blk.%d.post_ffw_norm" }, + }, + }, { LLM_ARCH_STARCODER2, { @@ -2066,6 +2088,9 @@ enum e_model { MODEL_8x22B, MODEL_16x12B, MODEL_10B_128x3_66B, + MODEL_57B_A14B, + MODEL_9B, + MODEL_27B, }; static const size_t kiB = 1024; @@ -2242,6 +2267,7 @@ struct llama_layer { struct ggml_tensor * attn_q_a_norm; struct ggml_tensor * attn_kv_a_norm; struct ggml_tensor * attn_sub_norm; + struct ggml_tensor * attn_post_norm; struct ggml_tensor * ffn_sub_norm; // attention @@ -2265,6 +2291,7 @@ struct llama_layer { // normalization struct ggml_tensor * ffn_norm; struct ggml_tensor * ffn_norm_b; + struct ggml_tensor * ffn_post_norm; struct ggml_tensor * layer_out_norm; struct ggml_tensor * layer_out_norm_b; struct ggml_tensor * ffn_norm_exps; @@ -4320,6 +4347,9 @@ static const char * llama_model_type_name(e_model type) { case MODEL_8x22B: return "8x22B"; case MODEL_16x12B: return "16x12B"; case MODEL_10B_128x3_66B: return "10B+128x3.66B"; + case MODEL_57B_A14B: return "57B.A14B"; + case MODEL_9B: return "9B"; + case MODEL_27B: return "27B"; default: return "?B"; } } @@ -4641,6 +4671,7 @@ static void llm_load_hparams( ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); switch (hparams.n_layer) { case 24: model.type = e_model::MODEL_A2_7B; break; + case 28: model.type = e_model::MODEL_57B_A14B; break; default: model.type = e_model::MODEL_UNKNOWN; } } break; @@ -4721,6 +4752,16 @@ static void llm_load_hparams( default: model.type = e_model::MODEL_UNKNOWN; } } break; + case LLM_ARCH_GEMMA2: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + + switch (hparams.n_layer) { + case 42: model.type = e_model::MODEL_9B; break; + case 46: model.type = e_model::MODEL_27B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; case LLM_ARCH_STARCODER2: { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); @@ -5130,6 +5171,9 @@ static void llm_load_vocab( } else if ( tokenizer_pre == "poro-chat") { vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_PORO; + } else if ( + tokenizer_pre == "viking") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_VIKING; } else { throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str())); } @@ -5224,10 +5268,10 @@ static void llm_load_vocab( if (gen_name.find("code") != std::string::npos) { if (model.arch == LLM_ARCH_LLAMA && 32010 < vocab.id_to_token.size() - && vocab.id_to_token[32007].text == "
"
-              && vocab.id_to_token[32008].text == ""
-              && vocab.id_to_token[32009].text == ""
-              && vocab.id_to_token[32010].text == "") {
+              && vocab.id_to_token[32007].text.find("
") != std::string::npos
+              && vocab.id_to_token[32008].text.find("") != std::string::npos
+              && vocab.id_to_token[32009].text.find("") != std::string::npos
+              && vocab.id_to_token[32010].text.find("") != std::string::npos) {
                 vocab.special_prefix_id = 32007;
                 vocab.special_suffix_id = 32008;
                 vocab.special_middle_id = 32009;
@@ -6585,6 +6629,40 @@ static bool llm_load_tensors(
                         layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
                     }
                 } break;
+            case LLM_ARCH_GEMMA2:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                    // output
+                    model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                    model.output      = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); // same as tok_embd, duplicated to allow offloading
+
+                    const int64_t n_ff          = hparams.n_ff;
+                    const int64_t n_embd_head_k = hparams.n_embd_head_k;
+                    const int64_t n_embd_k_gqa  = hparams.n_embd_k_gqa();
+                    const int64_t n_embd_v_gqa  = hparams.n_embd_v_gqa();
+
+                    for (uint32_t i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+
+                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * hparams.n_head});
+                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa});
+                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa});
+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * hparams.n_head, n_embd});
+                        layer.attn_post_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd});
+
+                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+                        layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
+                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
+                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
+                        layer.ffn_post_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd});
+                    }
+                } break;
             case LLM_ARCH_STARCODER2:
                 {
                     model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
@@ -10996,6 +11074,125 @@ struct llm_build_context {
         return gf;
     }
 
+    struct ggml_cgraph * build_gemma2() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        const int64_t n_embd_head_k = hparams.n_embd_head_k;
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
+        cb(inpL, "inp_scaled", -1);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+
+                struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+
+                struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_rope_ext(
+                        ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head,    n_tokens), inp_pos, nullptr,
+                        n_embd_head_k, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow);
+                cb(Qcur, "Qcur", il);
+
+                Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k)));
+                cb(Qcur, "Qcur_scaled", il);
+
+                Kcur = ggml_rope_ext(
+                        ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv, n_tokens), inp_pos, nullptr,
+                        n_embd_head_k, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow);
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
+                        model.layers[il].wo, NULL,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il);
+            }
+
+            cur = llm_build_norm(ctx0, cur, hparams,
+                    model.layers[il].attn_post_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_post_norm", il);
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+            }
+
+            struct ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
+            cb(sa_out, "sa_out", il);
+
+            cur = llm_build_norm(ctx0, sa_out, hparams,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "ffn_norm", il);
+
+            // feed-forward network
+            {
+                cur = llm_build_ffn(ctx0, cur,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        model.layers[il].ffn_gate, NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        LLM_FFN_GELU, LLM_FFN_PAR, cb, il);
+                cb(cur, "ffn_out", il);
+            }
+
+            cur = llm_build_norm(ctx0, cur, hparams,
+                model.layers[il].ffn_post_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+            cb(cur, "ffn_post_norm", -1);
+
+            cur = ggml_add(ctx0, cur, sa_out);
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = ggml_mul_mat(ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+
     struct ggml_cgraph * build_starcoder2() {
         struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
 
@@ -12376,6 +12573,10 @@ static struct ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_gemma();
             } break;
+        case LLM_ARCH_GEMMA2:
+            {
+                result = llm.build_gemma2();
+            } break;
         case LLM_ARCH_STARCODER2:
             {
                 result = llm.build_starcoder2();
@@ -14003,6 +14204,12 @@ struct llm_tokenizer_bpe {
                     " ?[^(\\s|.,!?…。，、।۔،)]+",
                 };
                 break;
+            case LLAMA_VOCAB_PRE_TYPE_VIKING:
+                regex_exprs = {
+                    "\\p{N}",
+                    " ?[^(\\s|.,!?…。，、।۔،)]+",
+                };
+                break;
             default:
                 // default regex for BPE tokenization pre-processing
                 regex_exprs = {
@@ -17915,6 +18122,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
         case LLM_ARCH_PHI2:
         case LLM_ARCH_PHI3:
         case LLM_ARCH_GEMMA:
+        case LLM_ARCH_GEMMA2:
         case LLM_ARCH_STARCODER2:
         case LLM_ARCH_GPTNEOX:
             return LLAMA_ROPE_TYPE_NEOX;
@@ -19752,7 +19960,10 @@ static int32_t llama_chat_apply_template_internal(
     std::string & dest, bool add_ass) {
     // Taken from the research: https://github.com/ggerganov/llama.cpp/issues/5527
     std::stringstream ss;
-    if (tmpl == "chatml" || tmpl.find("<|im_start|>") != std::string::npos) {
+    auto tmpl_contains = [&tmpl](std::string haystack) -> bool {
+        return tmpl.find(haystack) != std::string::npos;
+    };
+    if (tmpl == "chatml" || tmpl_contains("<|im_start|>")) {
         // chatml template
         for (auto message : chat) {
             ss << "<|im_start|>" << message->role << "\n" << message->content << "<|im_end|>\n";
@@ -19760,16 +19971,16 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "<|im_start|>assistant\n";
         }
-    } else if (tmpl == "llama2" || tmpl == "mistral" || tmpl.find("[INST]") != std::string::npos) {
+    } else if (tmpl == "llama2" || tmpl == "mistral" || tmpl_contains("[INST]")) {
         // llama2 template and its variants
         // [variant] support system message
-        bool support_system_message = tmpl.find("<>") != std::string::npos || tmpl == "mistral";
+        bool support_system_message = tmpl_contains("<>") || tmpl == "mistral";
         // [variant] space before + after response
-        bool space_around_response = tmpl.find("' ' + eos_token") != std::string::npos;
+        bool space_around_response = tmpl_contains("' ' + eos_token");
         // [variant] add BOS inside history
-        bool add_bos_inside_history = tmpl.find("bos_token + '[INST]") != std::string::npos;
+        bool add_bos_inside_history = tmpl_contains("bos_token + '[INST]");
         // [variant] trim spaces from the input message
-        bool strip_message = tmpl.find("content.strip()") != std::string::npos;
+        bool strip_message = tmpl_contains("content.strip()");
         // construct the prompt
         bool is_inside_turn = true; // skip BOS at the beginning
         ss << "[INST] ";
@@ -19795,7 +20006,7 @@ static int32_t llama_chat_apply_template_internal(
             }
         }
         // llama2 templates seem to not care about "add_generation_prompt"
-    } else if (tmpl == "phi3" || (tmpl.find("<|assistant|>") != std::string::npos && tmpl.find("<|end|>") != std::string::npos)) {
+    } else if (tmpl == "phi3" || (tmpl_contains("<|assistant|>") && tmpl_contains("<|end|>"))) {
         // Phi 3
         for (auto message : chat) {
             std::string role(message->role);
@@ -19804,7 +20015,7 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "<|assistant|>\n";
         }
-    } else if (tmpl == "zephyr" || tmpl.find("<|user|>") != std::string::npos) {
+    } else if (tmpl == "zephyr" || tmpl_contains("<|user|>")) {
         // zephyr template
         for (auto message : chat) {
             ss << "<|" << message->role << "|>" << "\n" << message->content << "<|endoftext|>\n";
@@ -19812,7 +20023,7 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "<|assistant|>\n";
         }
-    } else if (tmpl == "monarch" || tmpl.find("bos_token + message['role']") != std::string::npos) {
+    } else if (tmpl == "monarch" || tmpl_contains("bos_token + message['role']")) {
         // mlabonne/AlphaMonarch-7B template (the  is included inside history)
         for (auto message : chat) {
             std::string bos = (message == chat.front()) ? "" : ""; // skip BOS for first message
@@ -19821,7 +20032,7 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "assistant\n";
         }
-    } else if (tmpl == "gemma" || tmpl.find("") != std::string::npos) {
+    } else if (tmpl == "gemma" || tmpl == "gemma2" || tmpl_contains("")) {
         // google/gemma-7b-it
         std::string system_prompt = "";
         for (auto message : chat) {
@@ -19843,7 +20054,7 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "model\n";
         }
-    } else if (tmpl == "orion" || tmpl.find("'\\n\\nAssistant: ' + eos_token") != std::string::npos) {
+    } else if (tmpl == "orion" || tmpl_contains("'\\n\\nAssistant: ' + eos_token")) {
         // OrionStarAI/Orion-14B-Chat
         std::string system_prompt = "";
         for (auto message : chat) {
@@ -19863,7 +20074,7 @@ static int32_t llama_chat_apply_template_internal(
                 ss << message->content << "";
             }
         }
-    } else if (tmpl == "openchat" || tmpl.find("GPT4 Correct ") != std::string::npos) {
+    } else if (tmpl == "openchat" || tmpl_contains("GPT4 Correct ")) {
         // openchat/openchat-3.5-0106,
         for (auto message : chat) {
             std::string role(message->role);
@@ -19877,13 +20088,13 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "GPT4 Correct Assistant:";
         }
-    } else if (tmpl == "vicuna" || tmpl == "vicuna-orca" || (tmpl.find("USER: ") != std::string::npos && tmpl.find("ASSISTANT: ") != std::string::npos)) {
+    } else if (tmpl == "vicuna" || tmpl == "vicuna-orca" || (tmpl_contains("USER: ") && tmpl_contains("ASSISTANT: "))) {
         // eachadea/vicuna-13b-1.1 (and Orca variant)
         for (auto message : chat) {
             std::string role(message->role);
             if (role == "system") {
                 // Orca-Vicuna variant uses a system prefix
-                if (tmpl == "vicuna-orca" || tmpl.find("SYSTEM: ") != std::string::npos) {
+                if (tmpl == "vicuna-orca" || tmpl_contains("SYSTEM: ")) {
                     ss << "SYSTEM: " << message->content << "\n";
                 } else {
                     ss << message->content << "\n\n";
@@ -19897,7 +20108,7 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "ASSISTANT:";
         }
-    } else if (tmpl == "deepseek" || (tmpl.find("### Instruction:") != std::string::npos && tmpl.find("<|EOT|>") != std::string::npos)) {
+    } else if (tmpl == "deepseek" || (tmpl_contains("### Instruction:") && tmpl_contains("<|EOT|>"))) {
         // deepseek-ai/deepseek-coder-33b-instruct
         for (auto message : chat) {
             std::string role(message->role);
@@ -19912,7 +20123,7 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "### Response:\n";
         }
-    } else if (tmpl == "command-r" || (tmpl.find("<|START_OF_TURN_TOKEN|>") != std::string::npos && tmpl.find("<|USER_TOKEN|>") != std::string::npos)) {
+    } else if (tmpl == "command-r" || (tmpl_contains("<|START_OF_TURN_TOKEN|>") && tmpl_contains("<|USER_TOKEN|>"))) {
         // CohereForAI/c4ai-command-r-plus
         for (auto message : chat) {
             std::string role(message->role);
@@ -19927,7 +20138,7 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>";
         }
-    } else if (tmpl == "llama3" || (tmpl.find("<|start_header_id|>") != std::string::npos && tmpl.find("<|end_header_id|>") != std::string::npos)) {
+    } else if (tmpl == "llama3" || (tmpl_contains("<|start_header_id|>") && tmpl_contains("<|end_header_id|>"))) {
         // Llama 3
         for (auto message : chat) {
             std::string role(message->role);
@@ -19936,6 +20147,33 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "<|start_header_id|>assistant<|end_header_id|>\n\n";
         }
+    } else if (tmpl == "minicpm" || tmpl_contains(u8"<用户>")) {
+        // MiniCPM-3B-OpenHermes-2.5-v2-GGUF
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "user") {
+                ss << u8"<用户>";
+                ss << trim(message->content);
+                ss << "";
+            } else {
+                ss << trim(message->content);
+            }
+        }
+    } else if (tmpl == "deepseek2" || tmpl_contains("'Assistant: ' + message['content'] + eos_token")) {
+        // DeepSeek-V2
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                ss << message->content << "\n\n";
+            } else if (role == "user") {
+                ss << "User: " << message->content << "\n\n";
+            } else if (role == "assistant") {
+                ss << "Assistant: " << message->content << u8"<｜end▁of▁sentence｜>";
+            }
+        }
+        if (add_ass) {
+            ss << "Assistant:";
+        }
     } else {
         // template not supported
         return -1;