[{"data":1,"prerenderedAt":152},["ShallowReactive",2],{"project-40083109":3},{"id":4,"name":5,"line":6,"line_source":7,"title":8,"url":9,"domain":10,"maker":11,"posted_at":12,"year":13,"points":14,"comments":15,"rank_in_year":16,"shown_in_year":17,"top":18,"tier":19,"category":20,"subcategory":21,"format":22,"keywords":23,"reception":28,"sentiment":29,"stance_mix":30,"scored_comments":32,"aspects":33,"strip":36,"post_text":29,"what":38,"said":62,"maker_in_thread":92,"shape":100,"grid":102,"attempts":114,"same_week":115,"maker_shows":93,"maker_hidden":145,"reader_version":146,"card":147},"40083109","Unsloth","Finetune Llama-3 2x faster in a Colab notebook","title","Show HN: Finetune Llama-3 2x faster in a Colab notebook","https:\u002F\u002Fcolab.research.google.com\u002Fdrive\u002F135ced7oHytdxu3N2DNe1Z0kqjyYIkDXp?usp=sharing","colab.research.google.com","danielhanchen",1713496342,2024,45,6,1358,3515,39,"shown","ai","model finetuning library","library",[24,25,26,27],"llama-3 finetuning","lora finetuning","vram reduction","triton kernels","quiet",null,{"positive":31,"neutral":32,"mixed":31,"critical":31},0,2,{"idea":34,"performance":35},{"positive":31,"critical":31},{"positive":31,"critical":31},[31,37,31,37,31,31],3,{"problem":39,"audience":46,"pricing":29,"motivation":29,"time_to_build":29,"stack":53,"openness":61,"stage":61,"team_size":61},{"value":40,"source":41,"evidence":42},"Finetuning large language models is slow and uses too much VRAM.","inferred",[43],{"id":44,"by":11,"text":45},"40083110","Hey! If you're interested to try out finetuning Llama-3 8B (Meta's new 15 trillion token!! model), made a Colab to finetune Llama-3 2x faster and use 60% less VRAM, and supports…",{"value":47,"source":41,"evidence":48},"People finetuning Llama-3 models in Colab or Kaggle notebooks.",[49,50],{"id":44,"by":11,"text":45},{"id":51,"by":11,"text":52},"40103551","I have also have a Kaggle notebook: https:\u002F\u002Fwww.kaggle.com\u002Fcode\u002Fdanielhanchen\u002Fkaggle-llama-3-8b-unsloth-notebook Kaggle provides Tesla T4s 30 hours for free per week!!",[54],{"value":55,"source":56,"evidence":57},"Triton","maker",[58],{"id":59,"by":11,"text":60},"40083338","No catch at all!! There's 0 approximations, so everything is exact! We just have a custom backprop engine, rewrite everything into OpenAI's Triton language, and do all the…","unknown",{"praise":63,"objections":68,"requests":69,"questions":78,"alternatives":86},[64],{"value":65,"source":56,"evidence":66},"No accuracy loss from the speedups, since no approximations are used.",[67],{"id":59,"by":11,"text":60},[],[70],{"value":71,"source":72,"evidence":73},"Support for full parallel finetuning, not just 16bit and 4bit.","commenter",[74],{"id":75,"by":76,"text":77},"40083318","popinman322","Any news on when Unsloth's parallel full tuning will be available?",[79],{"value":80,"source":72,"evidence":81},"Is there a training speed or VRAM penalty for the speedups?",[82],{"id":83,"by":84,"text":85},"40083325","syntaxing","Hard to read on mobile, but if you don’t mind me asking, what’s the catch? Is there a penalty to training faster and less VRAM?",[87],{"value":88,"source":56,"evidence":89},"Hugging Face with Flash Attention 2.",[90,91],{"id":44,"by":11,"text":45},{"id":59,"by":11,"text":60},{"comments":93,"answered":32,"key":94},4,[95,96,99],{"id":44,"text":45},{"id":97,"text":98},"40083341","Full training hmm - for now finetuning in 16bit and 4bit are supported - if people are interested can work on it!",{"id":59,"text":60},{"commenters":37,"max_depth":37,"maker_replies":93,"first_comment_mins":31,"lifespan_hours":101},51,[103,106,108,110,112,113],[44,31,31,104,11,105],"project\u002Fpositive\u002Fperformance",1,[75,31,37,107,76,31],"project\u002Fneutral\u002Fidea",[97,31,31,109,11,105],"maker\u002Fneutral\u002Fidea",[83,31,37,111,84,31],"project\u002Fneutral\u002Fperformance",[59,31,31,104,11,105],[51,105,31,104,11,105],[],[116,121,126,130,135,140],{"id":117,"name":118,"points":119,"category":120},"40130304","Balanc3d",338,"games",{"id":122,"name":123,"points":124,"category":125},"40104641","Stepify",271,"productivity",{"id":127,"name":128,"points":129,"category":125},"40112958","OpenOrb",259,{"id":131,"name":132,"points":133,"category":134},"40132373","Getada",194,"dev_tools",{"id":136,"name":137,"points":138,"category":139},"40141980","Podlite",142,"languages",{"id":141,"name":142,"points":143,"category":144},"40089028","The Official MTA App",141,"consumer_social",false,"v2",{"id":148,"name":5,"line":6,"ai":145,"cat":20,"year":13,"points":14,"comments":15,"rec":28,"s":29,"top":18,"kw":149,"maker":11,"domain":10,"tier":19,"attempts":31,"strip":150,"url":9,"mix":151,"praise":65,"objection":-1,"sub":21},40083109,[24,25,26,27],"030300",[31,32,31,31],1794036653366]