๐Ÿค—ํ—ˆ๊น…ํŽ˜์ด์Šค, ํ•™์Šต์šฉ ํด๋Ÿฌ์Šคํ„ฐ๋ฅผ ์„œ๋น„์Šค ํ˜•ํƒœ๋กœ ์ œ๊ณตํ•˜๋Š” TCaaS(Training Cluster as a Service) ์ถœ์‹œ

image

ํ—ˆ๊น…ํŽ˜์ด์Šค์—์„œ ๋ณด์œ ํ•œ ํ•™์Šต์šฉ ํด๋Ÿฌ์Šคํ„ฐ๋ฅผ ์„œ๋น„์Šค ํ˜•ํƒœ๋กœ ์ œ๊ณตํ•˜๋Š” Training Cluster as a Service ๋ฅผ ์‹œ์ž‘ํ•˜์˜€์Šต๋‹ˆ๋‹ค. (์•„์ง์€ ๋Œ€๊ธฐ ๋ชฉ๋ก(waitlist)์— ๋“ฑ๋กํ•˜์—ฌ ์ˆœ์ฐจ์ ์œผ๋กœ ์„œ๋น„์Šค๋ฅผ ์ œ๊ณต๋ฐ›๋Š” ํ˜•ํƒœ๋กœ์ธ ๊ฒƒ ๊ฐ™์Šต๋‹ˆ๋‹ค.)


์•„๋ž˜ ๋งํฌ์—์„œ ๋‚ด์šฉ ํ™•์ธ ๋ฐ ๋Œ€๊ธฐ ์‹ ์ฒญ์„ ํ•˜์‹ค ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์•„์ง์€ ๋ณ„๋‹ค๋ฅธ ๋‚ด์šฉ์€ ์—†์Šต๋‹ˆ๋‹ค. :sweat_smile:


ํ˜„์žฌ๋Š” ๋ชจ๋ธ์˜ ํ˜•ํƒœ์™€ ํฌ๊ธฐ, ๊ทธ๋ฆฌ๊ณ  ๋ฐ์ดํ„ฐ์…‹์˜ ํฌ๊ธฐ์™€ ์›ํ•˜๋Š” ์ธํ”„๋ผ ๊ทœ๋ชจ๋ฅผ ์„ ํƒํ•˜๋ฉด ํ•™์Šต์— ํ•„์š”ํ•œ ์‹œ๊ฐ„๊ณผ ํ•จ๊ป˜ ๋Œ€๋žต์ ์ธ ๊ฐ€๊ฒฉ์„ ํ•™์ธํ•  ์ˆ˜ ์žˆ๋Š” ํŽ˜์ด์ง€๋งŒ ์ œ๊ณตํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.

์•„๋ž˜๋Š” ํ™ˆํŽ˜์ด์ง€์˜ ์„ค๋ช…์ž…๋‹ˆ๋‹ค.


https://huggingface.co/datasets/victor/design-stuff/resolve/main/cubes-light-3.mov

์ž์‹ ๋งŒ์˜ ํŒŒ์šด๋ฐ์ด์…˜ ๋ชจ๋ธ ํ•™์Šต / Train your own foundation model

ํŠน์ • ๋„๋ฉ”์ธ ๋ฐ ๋น„์ฆˆ๋‹ˆ์Šค ์š”๊ตฌ ์‚ฌํ•ญ์— ์ตœ์ ํ™”๋œ ๋ชจ๋ธ์„ ์ง์ ‘ ํ•™์Šตํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ•™์Šตํ•œ ๋ชจ๋ธ์€ ์›ํ•˜๋Š”๋Œ€๋กœ ์ž์œ ๋กญ๊ฒŒ ์‚ฌ์šฉ์ด ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค.

A model that is optimized for your specific domain and business needs. Youโ€™re free to use it however you want.

๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ํ†ต์ œ๊ถŒ ์œ ์ง€ / Keep control of your data

ํ—ˆ๊น…ํŽ˜์ด์Šค๋Š” ํ•™์Šต ๋ฐ์ดํ„ฐ๋ฅผ ์ €์žฅํ•˜์ง€ ์•Š์œผ๋ฉฐ, ์„œ๋น„์Šค ์‚ฌ์šฉ์ž๋Š” ํ•™์Šต ๊ฒฐ๊ณผ์™€ ๋กœ๊ทธ, ์ฒดํฌํฌ์ธํŠธ ์ „์ฒด์— ์ ‘๊ทผํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

We donโ€™t store the training data and you get access to the whole training output, logs and checkpoints.

์ธํ”„๋ผ ์ „๋ฌธ๊ฐ€ ์ง€์› / Infra Expert support

BLOOM, StarCoder ๋“ฑ๊ณผ ๊ฐ™์€ LLM ํ•™์Šต์— ๊ธฐ์—ฌํ–ˆ๋˜, ๋Œ€๊ทœ๋ชจ์˜ ํ•™์Šต ๊ฒฝํ—˜์ด ์žˆ์Šต๋‹ˆ๋‹ค.

We have experience with large-scale training, having contributed to LLMs like BLOOM, StarCoder, and more.


๋” ์ฝ์–ด๋ณด๊ธฐ