{"id":27534,"date":"2023-09-14T10:53:59","date_gmt":"2023-09-14T01:53:59","guid":{"rendered":"http:\/\/nori.company\/?p=27534"},"modified":"2023-09-14T10:53:59","modified_gmt":"2023-09-14T01:53:59","slug":"exllamav2-%ec%9d%bc%eb%b0%98-gpu%ec%97%90%ec%84%9c-%eb%a1%9c%ec%bb%ac-llms%eb%a5%bc-%ec%8b%a4%ed%96%89%ed%95%98%ea%b8%b0-%ec%9c%84%ed%95%9c-%eb%b9%a0%eb%a5%b8-%ec%b6%94%eb%a1%a0-%eb%9d%bc%ec%9d%b4","status":"publish","type":"post","link":"https:\/\/nori.company\/?p=27534","title":{"rendered":"ExLlamaV2: \uc77c\ubc18 GPU\uc5d0\uc11c \ub85c\uceec LLMs\ub97c \uc2e4\ud589\ud558\uae30 \uc704\ud55c \ube60\ub978 \ucd94\ub860 \ub77c\uc774\ube0c\ub7ec\ub9ac"},"content":{"rendered":"<ul>\n<li>3090\/4090 \uac19\uc740 GPU\uc5d0\uc11c \ub85c\uceec LLM\uc744 \uc2e4\ud589\ud558\ub3c4\ub85d \uc124\uacc4\ub41c \ucd94\ub860 \ub77c\uc774\ube0c\ub7ec\ub9ac<\/li>\n<li>\ucd08\uae30 \ub9b4\ub9ac\uc988\uc774\uba70 \ucf54\ub4dc\ub294 \uc544\uc9c1 \ud14c\uc2a4\ud2b8 \ub2e8\uacc4\uc5d0 \uc788\uc73c\uba70, \uc77c\ubd80 \uc8fc\uc694 \uae30\ub2a5\uc740 \uc544\uc9c1 \uad6c\ud604\ub418\uc9c0 \uc54a\uc74c<\/li>\n<li>V1\uc5d0 \ube44\ud574 ExLlamaV2\ub294 \ub354 \ube60\ub974\uace0 \ub354 \ub098\uc740 \ucee4\ub110, \ub354 \uae54\ub054\ud558\uace0 \ub2e4\uc591\ud55c \ucf54\ub4dc\ubca0\uc774\uc2a4, \uadf8\ub9ac\uace0 \uc0c8\ub85c\uc6b4 \uc591\uc790 \ud615\uc2dd\uc744 \uc9c0\uc6d0<\/li>\n<li...\n\n\n\ucd9c\ucc98 : <a href=\"https:\/\/news.hada.io\/topic?id=10866\" target=\"_blank\" rel=\"noopener\">GeekNews &#8211; \uac1c\ubc1c\/\uae30\uc220\/\uc2a4\ud0c0\ud2b8\uc5c5 \ub274\uc2a4 \uc11c\ube44\uc2a4 <\/a><\/ul>\n","protected":false},"excerpt":{"rendered":"<p>3090\/4090 \uac19\uc740 GPU\uc5d0\uc11c \ub85c\uceec LLM\uc744 \uc2e4\ud589\ud558\ub3c4\ub85d \uc124\uacc4\ub41c \ucd94\ub860 \ub77c\uc774\ube0c\ub7ec\ub9ac \ucd08\uae30 \ub9b4\ub9ac\uc988\uc774\uba70 \ucf54\ub4dc\ub294 \uc544\uc9c1 \ud14c\uc2a4\ud2b8 \ub2e8\uacc4\uc5d0 \uc788\uc73c\uba70, \uc77c\ubd80 \uc8fc\uc694 \uae30\ub2a5\uc740 \uc544\uc9c1&hellip;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"jetpack_post_was_ever_published":false,"_jetpack_newsletter_access":"","_jetpack_dont_email_post_to_subs":false,"_jetpack_newsletter_tier_id":0,"_jetpack_memberships_contains_paywalled_content":false,"_jetpack_memberships_contains_paid_content":false,"footnotes":"","jetpack_publicize_message":"","jetpack_publicize_feature_enabled":true,"jetpack_social_post_already_shared":true,"jetpack_social_options":{"image_generator_settings":{"template":"highway","default_image_id":0,"font":"","enabled":false},"version":2}},"categories":[138],"tags":[163],"class_list":["post-27534","post","type-post","status-publish","format-standard","hentry","category-it","tag-geeknews-----"],"jetpack_publicize_connections":[],"jetpack_featured_media_url":"","jetpack_sharing_enabled":true,"jetpack_shortlink":"https:\/\/wp.me\/pa3PO4-7a6","jetpack-related-posts":[{"id":25883,"url":"https:\/\/nori.company\/?p=25883","url_meta":{"origin":27534,"position":0},"title":"MK-1","author":"\ub178\ub9ac\ucef4\ud37c\ub2c8","date":"2023\ub144 8\uc6d4 7\uc77c","format":false,"excerpt":"MK-1\uc740 OpenAI, Anthropic, Google\uacfc \uac19\uc740 \uc5d8\ub9ac\ud2b8 AI \uac15\uc790\ub4e4\uacfc \ub3d9\ub4f1\ud558\uac70\ub098 \ub354 \ub098\uc740 \ub2a5\ub825\uc744 \uac00\uc9c4 AI \ubaa8\ub378\uc744 \uc81c\uacf5\ud558\ub824\ub294 \uc0c8\ub85c\uc6b4 \ud68c\uc0ac\uc785\ub2c8\ub2e4. \ud68c\uc0ac\uc758 \uccab \ubc88\uc9f8 \uc81c\ud488\uc778 MKML\uc740 \uba87 \uc904\uc758 Python \ucf54\ub4dc\ub9cc\uc73c\ub85c GPU\uc5d0\uc11c \ud070 \uc5b8\uc5b4 \ubaa8\ub378(LLM) \ucd94\ub860 \ube44\uc6a9\uc744 2\ubc30 \uc904\uc77c \uc218 \uc788\ub294 \ucd94\ub860 \ub7f0\ud0c0\uc784\uc785\ub2c8\ub2e4. MKML\uc740 Hugg... \ucd9c\ucc98 : GeekNews - \uac1c\ubc1c\/\uae30\uc220\/\uc2a4\ud0c0\ud2b8\uc5c5 \ub274\uc2a4 \uc11c\ube44\uc2a4","rel":"","context":"&quot;IT&quot;\uc5d0\uc11c","block_context":{"text":"IT","link":"https:\/\/nori.company\/?cat=138"},"img":{"alt_text":"","src":"","width":0,"height":0},"classes":[]},{"id":25321,"url":"https:\/\/nori.company\/?p=25321","url_meta":{"origin":27534,"position":1},"title":"\uce74\uce74\uc624\ube0c\ub808\uc778, \u201cAI \uac1c\ubc1c \ud6a8\uc728\uc131 \ub192\uc5ec, \uc778\uacf5\uc9c0\ub2a5 \ubbf8\ub798 \ubc1d\ud78c\ub2e4\u201d&#8230;\uac70\ub300 AI \ubaa8\ub378 \uc27d\uac8c \uac1c\ubc1c\ud558\ub294 \u2018\ud2b8\ub77c\uc774\ub358\ud2b8\u2019 \uc804\uaca9 \uacf5\uac1c","author":"\ub178\ub9ac\ucef4\ud37c\ub2c8","date":"2023\ub144 7\uc6d4 26\uc77c","format":false,"excerpt":"\uce74\uce74\uc624\ube0c\ub808\uc778\uc774 AI \ubaa8\ub378\uc758 \ud6c8\ub828 \ubc0f \ucd94\ub860 \uc18d\ub3c4\ub97c \ud5a5\uc0c1\uc2dc\ud0a4\ub294 \uc131\ub2a5 \ub77c\uc774\ube0c\ub7ec\ub9ac \u2018\ud2b8\ub77c\uc774\ub358\ud2b8(Trident)\u2019\ub97c 26\uc77c \uacf5\uac1c\ud588\ub2e4.\u2018\ud2b8\ub77c\uc774\ub358\ud2b8\u2019\ub294 \uc624\ud508AI\uc758 GPU(\uadf8\ub798\ud53d \ucc98\ub9ac\uc7a5\uce58, Graphic Processing Unit) \ud504\ub85c\uadf8\ub798\ubc0d \uc5b8\uc5b4\uc778 \u2018\ud2b8\ub9ac\ud1a4(Triton)\u2019\uc744 \uae30\ubc18\uc73c\ub85c \uad6c\ud604\ud55c \ub77c\uc774\ube0c\ub7ec\ub9ac(\uc7ac\uc0ac\uc6a9 \uac00\ub2a5\ud55c \ucf54\ub4dc\uc758 \uc9d1\ud569)\ub2e4. \u2018\ud2b8\ub77c\uc774\ub358\ud2b8\u2019\ub294 AI \ubaa8\ub378 \uac1c\ubc1c\uc5d0 \ud544\uc694\ud55c \uc5f0\uc0b0\ub4e4\uc744 \ucd5c\uc801\ud654\ub41c \ucee4\ub110(Kernel)\ub85c \uad6c\ud604\ud558\uace0, \uba54\ud0c0\uac00 \uac1c\ubc1c\ud55c \uba38\uc2e0\ub7ec\ub2dd \ud504\ub808\uc784\uc6cc\ud06c \u2018\ud30c\uc774\ud1a0\uce58(PyTorch)\u2019\uc640 \uac19\uc740 \uc0ac\uc6a9 \ubc29\ubc95\uc744 \uc81c\uacf5\ud55c\ub2e4. \uc778\uacf5\uc9c0\ub2a5 \uac1c\ubc1c\uc790\ub294 \ubcf5\uc7a1\ud558\uace0 \uc5b4\u2026","rel":"","context":"&quot;AI&quot;\uc5d0\uc11c","block_context":{"text":"AI","link":"https:\/\/nori.company\/?cat=155"},"img":{"alt_text":"","src":"","width":0,"height":0},"classes":[]},{"id":22735,"url":"https:\/\/nori.company\/?p=22735","url_meta":{"origin":27534,"position":2},"title":"Simple LLM Finetuner &#8211; \ucd08\ubcf4\uc790 \uce5c\ud654\uc801\uc778 UI\ub97c \uac00\uc9c4 LLM \ud29c\ub2dd \ub3c4\uad6c","author":"\ub178\ub9ac\ucef4\ud37c\ub2c8","date":"2023\ub144 4\uc6d4 27\uc77c","format":false,"excerpt":"\ub2e4\uc591\ud55c LLM\uc5d0 \ub300\ud574\uc11c LoRA \uba54\uc18c\ub4dc(PEFT \ub77c\uc774\ube0c\ub7ec\ub9ac \uc774\uc6a9)\ub85c \uc27d\uac8c \ud30c\uc778\ud29c\ub2dd \uac00\ub2a5 \ub370\uc774\ud130\uc14b \uad00\ub9ac, \ud30c\ub77c\ubbf8\ud130 \ucee4\uc2a4\ud130\ub9c8\uc774\uc9d5, \ud6c8\ub828 \ubc0f \ubaa8\ub378\uc758 \ucd94\ub860 \ub2a5\ub825 \ud3c9\uac00 \uc77c\ubc18 \uc18c\ube44\uc790\uc6a9 NVIDIA GPU(16G VRAM \uc774\uc0c1)\ub85c\ub3c4 \uac00\ub2a5 Linux \ub610\ub294 WSL ... \ucd9c\ucc98 : GeekNews - \uac1c\ubc1c\/\uae30\uc220\/\uc2a4\ud0c0\ud2b8\uc5c5 \ub274\uc2a4 \uc11c\ube44\uc2a4","rel":"","context":"&quot;IT&quot;\uc5d0\uc11c","block_context":{"text":"IT","link":"https:\/\/nori.company\/?cat=138"},"img":{"alt_text":"","src":"","width":0,"height":0},"classes":[]},{"id":25769,"url":"https:\/\/nori.company\/?p=25769","url_meta":{"origin":27534,"position":3},"title":"VkFFT &#8211; \ud6a8\uc728\uc801\uc778 GPU \uac00\uc18d FFT \ub77c\uc774\ube0c\ub7ec\ub9ac \uc624\ud508\uc18c\uc2a4","author":"\ub178\ub9ac\ucef4\ud37c\ub2c8","date":"2023\ub144 8\uc6d4 4\uc77c","format":false,"excerpt":"NVidia \uc758 cuFFT \ub77c\uc774\ube0c\ub7ec\ub9ac \ubcf4\ub2e4 \ub354 \ub098\uc740 \uc131\ub2a5\uc758 \uc624\ud508\uc18c\uc2a4\ub97c \uc81c\uacf5\ud558\ub294 \uac83\uc744 \ubaa9\ud45c\ub85c \ud568 C\ub85c \uc791\uc131 \ub418\uc5c8\uc73c\uba70 Vulkan, CUDA, HIP, OpenCL, Level Zero, Metal \uc744 \ubc31\uc5d4\ub4dc\ub85c \uc9c0\uc6d0 \uba40\ud2f0\ub514\uba58\uc804 \uc9c0\uc6d0 : 1D\/2D\/3D\/ND (VKFFT_MAX_FFT_DIMENSIONS \ub85c \uc784\uc758\uc758 \ucc28\uc6d0\uc218 \uc9c0\uc815 \uac00\ub2a5) \uc708\ub3c4\uc6b0, \ub9ac\ub205\uc2a4,... \ucd9c\ucc98 : GeekNews - \uac1c\ubc1c\/\uae30\uc220\/\uc2a4\ud0c0\ud2b8\uc5c5 \ub274\uc2a4 \uc11c\ube44\uc2a4","rel":"","context":"&quot;IT&quot;\uc5d0\uc11c","block_context":{"text":"IT","link":"https:\/\/nori.company\/?cat=138"},"img":{"alt_text":"","src":"","width":0,"height":0},"classes":[]},{"id":26348,"url":"https:\/\/nori.company\/?p=26348","url_meta":{"origin":27534,"position":4},"title":"DoctorGPT &#8211; \ubbf8\uad6d \uc758\uc0ac \uba74\ud5c8 \uc2dc\ud5d8\uc744 \ud1b5\uacfc\ud558\ub294 LLM","author":"\ub178\ub9ac\ucef4\ud37c\ub2c8","date":"2023\ub144 8\uc6d4 18\uc77c","format":false,"excerpt":"Llama2 7b \ubaa8\ub378\uc744 \uc758\ud559 \ub300\ud654 \ub370\uc774\ud130\uc14b\uc73c\ub85c \ud30c\uc778 \ud29c\ub2dd \uac15\ud654\ud559\uc2b5 \ubc0f Constitutional AI\ub85c \ucd94\uac00 \uac1c\uc120 \ubaa8\ub378\uc740 3\uae30\uac00\ub85c \ub85c\uceec \ub514\ubc14\uc774\uc2a4\uc5d0\uc11c \ub3d9\uc791 iOS\/Android\/Web \uc5d0\uc11c \uc2e4\ud589 \uc9c0\uc6d0 \ubaa8\ub378 \ud6c8\ub828\uc744 \uc704\ud574\uc11c \ub85c\uceec GPU \ub610\ub294 Colab Pro \uc774\uc6a9 \ud544\uc694(\uc720\ub8cc \uc778\uc2a4\ud134\uc2a4\uc5d0\uc11c \uc57d 24\uc2dc\uac04 \uc18c\uc694)... \ucd9c\ucc98 : GeekNews - \uac1c\ubc1c\/\uae30\uc220\/\uc2a4\ud0c0\ud2b8\uc5c5 \ub274\uc2a4 \uc11c\ube44\uc2a4","rel":"","context":"&quot;IT&quot;\uc5d0\uc11c","block_context":{"text":"IT","link":"https:\/\/nori.company\/?cat=138"},"img":{"alt_text":"","src":"","width":0,"height":0},"classes":[]},{"id":21880,"url":"https:\/\/nori.company\/?p=21880","url_meta":{"origin":27534,"position":5},"title":"\ub9c8\uc774\ud06c\ub85c\uc18c\ud504\ud2b8\u00b7\ud150\uc13c\ud2b8\u00b7\ubc14\uc774\ub450, &#8216;\ucef4\ud4e8\ud130 \ube44\uc804 AI&#8217;\uc5d0 \uc624\ud508 \uc18c\uc2a4 GPU \uac00\uc18d \ub77c\uc774\ube0c\ub7ec\ub9ac &#8216;CV-\ucfe0\ub2e4&#8217; \ucc44\ud0dd","author":"\ub178\ub9ac\ucef4\ud37c\ub2c8","date":"2023\ub144 3\uc6d4 27\uc77c","format":false,"excerpt":"\ub9c8\uc774\ud06c\ub85c\uc18c\ud504\ud2b8(Microsoft), \ud150\uc13c\ud2b8(Tencent), \ubc14\uc774\ub450(Baidu)\uac00 \ucef4\ud4e8\ud130 \ube44\uc804 AI\ub97c \uc704\ud574 \uc5d4\ube44\ub514\uc544 'CV-\ucfe0\ub2e4(CV-CUDA)'\ub97c \ucc44\ud0dd\ud588\ub2e4.CV-\ucfe0\ub2e4\ub294 GPU\ub97c \ud65c\uc6a9\ud55c \ud074\ub77c\uc6b0\ub4dc \uaddc\ubaa8\uc758 \ucef4\ud4e8\ud130 \ube44\uc804\uc744 \uc704\ud55c \uc624\ud508 \uc18c\uc2a4 GPU \uac00\uc18d \ub77c\uc774\ube0c\ub7ec\ub9ac\ub85c, \uc804 \uc138\uacc4 \uae30\uc5c5\ub4e4\uc774 GPU\uc5d0\uc11c \uc5d4\ub4dc\ud22c\uc5d4\ub4dc AI \uae30\ubc18 \ucef4\ud4e8\ud130 \ube44\uc804 \ubc0f \uc774\ubbf8\uc9c0 \ucc98\ub9ac \ud30c\uc774\ud504\ub77c\uc778\uc744 \uad6c\ucd95\ud558\uace0 \ud655\uc7a5\ud560 \uc218 \uc788\ub3c4\ub85d \uc9c0\uc6d0\ud55c\ub2e4.\uc9c0\ub09c GTC 2023 \uae30\uc870\uc5f0\uc124\uc5d0\uc11c \uc820\uc2a8 \ud669\uc740 \"\uc778\ud130\ub137 \ud2b8\ub798\ud53d\uc758 80%\ub294 \ube44\ub514\uc624\uc774\uba70, \uc0ac\uc6a9\uc790\uac00\u2026","rel":"","context":"&quot;AI&quot;\uc5d0\uc11c","block_context":{"text":"AI","link":"https:\/\/nori.company\/?cat=155"},"img":{"alt_text":"","src":"","width":0,"height":0},"classes":[]}],"_links":{"self":[{"href":"https:\/\/nori.company\/index.php?rest_route=\/wp\/v2\/posts\/27534","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/nori.company\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/nori.company\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/nori.company\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/nori.company\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=27534"}],"version-history":[{"count":1,"href":"https:\/\/nori.company\/index.php?rest_route=\/wp\/v2\/posts\/27534\/revisions"}],"predecessor-version":[{"id":27535,"href":"https:\/\/nori.company\/index.php?rest_route=\/wp\/v2\/posts\/27534\/revisions\/27535"}],"wp:attachment":[{"href":"https:\/\/nori.company\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=27534"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/nori.company\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=27534"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/nori.company\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=27534"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}