Skip to content

YoungHaKim7/Machine_Learning_Rust

Repository files navigation

link















์ตœ์‹ ๋‰ด์Šค(์™ธ๋ถ€๋งํฌ)


C++ examples for the Vulkan graphics API

๐Ÿ‰ Making Rust a first-class language and ecosystem for GPU shaders ๐Ÿšง|๐Ÿ”|


  • GPU๋Š” ์—ฐ์‚ฐ ์†๋„๊ฐ€ ๋ฉ”๋ชจ๋ฆฌ ์ ‘๊ทผ ์†๋„๋ณด๋‹ค ์›”๋“ฑํžˆ ๋นจ๋ผ์„œ, ๋ฉ”๋ชจ๋ฆฌ ๊ณ„์ธต ๊ตฌ์กฐ๊ฐ€ ์„ฑ๋Šฅ์˜ ๋ณ‘๋ชฉ์„ ์ผ์œผํ‚ด
  • ์—ฐ์‚ฐ ์ง‘์•ฝ๋„(Arithmetic Intensity, AI) ์— ๋”ฐ๋ผ ์—ฐ์‚ฐ์ด ๋ฉ”๋ชจ๋ฆฌ ๋ฐ”์šด๋“œ, ๊ณ„์‚ฐ ๋ฐ”์šด๋“œ ์ƒํƒœ๋กœ ๊ตฌ๋ถ„๋˜๋ฉฐ, A100 GPU์˜ ์ž„๊ณ„์ ์€ ์•ฝ 13 FLOPs/Byte์ž„
  • ์„ฑ๋Šฅ ์ตœ์ ํ™” ์ฃผ์š” ์ „๋žต์œผ๋กœ ์—ฐ์‚ฐ ํ•ฉ์น˜๊ธฐ(Fusioโ€ฆ

๋จธ์‹ ๋Ÿฌ๋‹ ์ „์šฉ TPU(TPU ์‹ฌ์ธต ๋ถ„์„)|๐Ÿ”|

  • 250623TPU ์‹ฌ์ธต ๋ถ„์„
  • TPU๋Š” ๊ตฌ๊ธ€์ด ๊ฐœ๋ฐœํ•œ ๋Œ€๊ทœ๋ชจ AI ํ•™์Šต ๋ฐ ์ถ”๋ก ์šฉ ๋งž์ถคํ˜• ์นฉ์œผ๋กœ, GPU์™€๋Š” ๋‹ค๋ฅธ ์„ค๊ณ„ ์ฒ ํ•™์„ ๊ฐ€์ง€๊ณ  ์žˆ์Œ
  • ํ™•์žฅ์„ฑ๊ณผ ์—๋„ˆ์ง€ ํšจ์œจ์„ฑ์„ ๊ฐ•์กฐํ•˜๋ฉฐ, ํ•˜๋“œ์›จ์–ด(์˜ˆ: ์‹œ์Šคํ…œ ์˜จ์นฉ ๊ตฌ์„ฑ, ๋Œ€ํ˜• ์˜จ์นฉ ๋ฉ”๋ชจ๋ฆฌ)์™€ ์†Œํ”„ํŠธ์›จ์–ด(** XLA ์ปดํŒŒ์ผ๋Ÿฌ**)๋ฅผ ํ•จ๊ป˜ ์„ค๊ณ„ํ•จ
  • ํ•ต์‹ฌ ๊ตฌ์กฐ๋Š” ์‹œ์Šคํ†จ๋ฆญ ์–ด๋ ˆ์ด์™€ ํŒŒโ€ฆ

Cuda(nvidia์˜ ์šฉ๋„๋ณ„ ์ •๋ฆฌ|๐Ÿ”|

cuPYNUMERIC NUMERICAL COMPUTING
cuLITHO COMPUTATIONAL
LITHOGRAPHY
AERIAL 5G/6G SIGNAL PROCESSING
cuOPT DECISION OPTIMIZATION
PARABRICK GENE SEQUENCING
MONAI MEDICAL IMAGING
๋ฌธ์„œ https://docs.monai.io/en/stable/networks.html
EARTH-2 WEATHER ANALYTICS
cuQUANTUM
CUDA-Q
QUANTUM COMPUTING
cuEQUIVARIANCE
cuTENSOR
QUANTUM CHEMISTRY

  • DeepSeek ํŒ€์ด ๋‚ด๋ถ€ ์ถ”๋ก  ์—”์ง„(DeepSeek Inference Engine)์„ ์˜คํ”ˆ์†Œ์Šค๋กœ ํ™˜์›ํ•˜๊ธฐ ์œ„ํ•œ ๊ณ„ํš์„ ๊ณต๊ฐœํ•จ
  • ๊ธฐ์กด์˜ ์ถ”๋ก  ์—”์ง„์€ vLLM ๊ธฐ๋ฐ˜์ด๋ฉฐ, DeepSeek-V3 ๋ฐ R1 ๋ชจ๋ธ์˜ ๋ฐฐํฌ ์ˆ˜์š” ์ฆ๊ฐ€์— ๋”ฐ๋ผ ๊ณต์œ ๋ฅผ ๊ณ ๋ ค์ค‘
  • ๊ธฐ์กด ์ฝ”๋“œ์™€ ์ธํ”„๋ผ ์ข…์†์„ฑ, ์œ ์ง€๋ณด์ˆ˜ ๋ถ€๋‹ด ๋“ฑ์œผ๋กœ ์ „์ฒด ๊ณต๊ฐœ๋Š” ์–ด๋ ค์›€, ๋Œ€์‹  **๋ชจ๋“ˆํ™” ๋ฐ ๊ธฐโ€ฆ
  • 3FS๋Š” DeepSeek๊ฐ€ ๊ฐœ๋ฐœํ•œ ๊ณ ์„ฑ๋Šฅ ์˜คํ”ˆ์†Œ์Šค ๋ถ„์‚ฐ ํŒŒ์ผ ์‹œ์Šคํ…œ์œผ๋กœ, ๋Œ€๊ทœ๋ชจ ๋ฐ์ดํ„ฐ ์ฒ˜๋ฆฌ์™€ ๋†’์€ ์ฒ˜๋ฆฌ๋Ÿ‰์„ ์ง€์›ํ•จ
  • ์ผ๋ฐ˜์ ์ธ ํŒŒ์ผ ์‹œ์Šคํ…œ์ฒ˜๋Ÿผ ๋™์ž‘ํ•˜์ง€๋งŒ, ์‹ค์ œ๋กœ๋Š” ์—ฌ๋Ÿฌ ๋จธ์‹ ์— ๋ฐ์ดํ„ฐ๋ฅผ ๋ถ„์‚ฐ ์ €์žฅํ•˜๋ฉฐ ์‚ฌ์šฉ์ž๋Š” ์ด๋ฅผ ์˜์‹ํ•˜์ง€ ์•Š์•„๋„ ๋˜๋Š” ์ถ”์ƒํ™” ๊ตฌ์กฐ๋ฅผ ๊ฐ€์ง
  • **4๊ฐ€์ง€ ์ฃผ์š” ๊ตฌ์„ฑ ์š”์†Œ (Meta, Mgmtd, Stoโ€ฆ

โ–ฒDeepSeek, 3FS ํŒŒ์ผ์‹œ์Šคํ…œ ๊ณผ Smallpond ๋ฐ์ดํ„ฐ ์ฒ˜๋ฆฌ ํ”„๋ ˆ์ž„์›Œํฌ ์˜คํ”ˆ์†Œ์Šค ๊ณต๊ฐœ (5 of 5) (github.com/deepseek-ai)|๐Ÿ”|

  • 250228
  • Fire-Flyer File System(3FS)๋Š” AI ํ•™์Šต ๋ฐ ์ถ”๋ก  ์›Œํฌ๋กœ๋“œ๋ฅผ ์ฒ˜๋ฆฌํ•˜๊ธฐ ์œ„ํ•ด ์„ค๊ณ„๋œ ๊ณ ์„ฑ๋Šฅ ๋ถ„์‚ฐ ํŒŒ์ผ ์‹œ์Šคํ…œ ์ตœ์‹  SSD ๋ฐ RDMA ๋„คํŠธ์›Œํฌ๋ฅผ ํ™œ์šฉํ•˜์—ฌ ๊ณต์œ  ์Šคํ† ๋ฆฌ์ง€ ๊ณ„์ธต์„ ์ œ๊ณตํ•˜๊ณ , ๋ถ„์‚ฐ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ๊ฐœ๋ฐœ์„ ๋‹จ์ˆœํ™”ํ•จ
  • DeepSeek V3/R1์—์„œ ์‚ฌ์šฉํ–ˆ๋˜ ์ „๋žต ๋ฐ ์ฝ”๋“œ๋“ค
    • DualPipe : ๊ณ„์‚ฐ-ํ†ต์‹  ์˜ค๋ฒ„๋žฉ์„ ์œ„ํ•œ ์–‘๋ฐฉํ–ฅ ํŒŒ์ดํ”„๋ผ์ธ ๋ณ‘๋ ฌํ™” ์•Œ๊ณ ๋ฆฌ๋“ฌ
    • EPLB: Expert-Parallel ๋กœ๋“œ๋ฐธ๋Ÿฐ์„œ
    • Profile-Data: DeepSeek ์ธํ”„๋ผ์˜ ๋ฐ์ดํ„ฐ ํ”„๋กœํŒŒ์ผ๋ง์œผ๋กœ ๊ณ„์‚ฐ-ํ†ต์‹  ์˜ค๋ฒ„๋žฉ์„ ๋ถ„์„

DeepSeek, DeepGEMM ์˜คํ”ˆ์†Œ์Šค ๊ณต๊ฐœ (3 of 5) (github.com/deepseek-ai)|๐Ÿ”|

  • https://news.hada.io/topic?id=19444 3P by xguru 2์ผ์ „ | โ˜… favorite
  • FP8 ํ–‰๋ ฌ ๊ณฑ์…ˆ(GEMM) ์„ ํšจ์œจ์ ์œผ๋กœ ์ˆ˜ํ–‰ํ•˜๋Š” ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋กœ, DeepSeek-V3์—์„œ ์ œ์•ˆ๋œ ๋ฏธ์„ธ ์กฐ์ • ์Šค์ผ€์ผ๋ง(fine-grained scaling) ๋ฐฉ์‹์„ ์ง€์›ํ•จ ์ผ๋ฐ˜ GEMM๊ณผ Mix-of-Experts(MoE) ๊ทธ๋ฃนํ™” GEMM์„ ๋ชจ๋‘ ์ง€์› CUDA ๊ธฐ๋ฐ˜์œผ๋กœ ๊ตฌํ˜„๋˜์—ˆ์œผ๋ฉฐ, ์„ค์น˜ ์‹œ ๋ณ„๋„ ์ปดํŒŒ์ผ ์—†์ด ๊ฒฝ๋Ÿ‰ Just-In-Time(JIT) ๋ชจ๋“ˆ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋Ÿฐํƒ€์ž„์—์„œ ์ปค๋„์„ ์ปดํŒŒ์ผํ•จ ํ˜„์žฌ NVIDIA Hopper ํ…์„œ ์ฝ”์–ด ์ „์šฉ์œผ๋กœ ์ง€์› FP8 ํ…์„œ ์ฝ”์–ด์˜ ๋ถ€์ •ํ™•ํ•œ ๋ˆ„์  ์—ฐ์‚ฐ์„ ๋ณด์™„ํ•˜๊ธฐ ์œ„ํ•ด CUDA ์ฝ”์–ด ๊ธฐ๋ฐ˜ ์ด์ค‘ ๋ˆ„์ (promotion) ์‚ฌ์šฉ CUTLASS ๋ฐ CuTe์˜ ์ผ๋ถ€ ๊ฐœ๋…์„ ํ™œ์šฉํ•˜์ง€๋งŒ, ๋ณต์žกํ•œ ํ…œํ”Œ๋ฆฟ ์˜์กด์„ฑ์„ ์ค„์—ฌ ์•ฝ 300์ค„์˜ ์ปค๋„ ์ฝ”๋“œ๋งŒ ํฌํ•จํ•˜๋Š” ๋‹จ์ˆœํ•œ ์„ค๊ณ„ Hopper FP8 ํ–‰๋ ฌ ์—ฐ์‚ฐ ๋ฐ ์ตœ์ ํ™” ๊ธฐ๋ฒ•์„ ํ•™์Šตํ•˜๊ธฐ์— ์ ํ•ฉ ๊ฒฝ๋Ÿ‰ ์„ค๊ณ„์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ๋‹ค์–‘ํ•œ ํ–‰๋ ฌ ํฌ๊ธฐ์—์„œ ์ „๋ฌธ๊ฐ€ ์ˆ˜์ค€์œผ๋กœ ํŠœ๋‹๋œ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ์™€ ์œ ์‚ฌํ•˜๊ฑฐ๋‚˜ ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ณด์ž„

DeepSeek, DeepEP ์˜คํ”ˆ์†Œ์Šค ๊ณต๊ฐœ (2 of 5) (github.com/deepseek-ai)|๐Ÿ”|

  • https://news.hada.io/topic?id=19421 3P by xguru 3์ผ์ „ | โ˜… favorite | ๋Œ“๊ธ€๊ณผ ํ† ๋ก 
  • Mixture-of-Experts(MoE) ๋ฐ Expert Parallelism(EP)์„ ์œ„ํ•œ ๊ณ ์„ฑ๋Šฅ ํ†ต์‹  ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ GPU ๊ธฐ๋ฐ˜ All-to-All ์ปค๋„์„ ์ œ๊ณตํ•˜์—ฌ MoE ๋””์ŠคํŒจ์น˜ ๋ฐ ๊ฒฐํ•ฉ ์—ฐ์‚ฐ์„ ๊ณ ์†์œผ๋กœ ์ฒ˜๋ฆฌ FP8๊ณผ ๊ฐ™์€ ์ €์ •๋ฐ€ ์—ฐ์‚ฐ ์ง€์› DeepSeek-V3 ๋…ผ๋ฌธ์—์„œ ์ œ์•ˆํ•œ ๊ทธ๋ฃน ์ œํ•œ ๊ฒŒ์ดํŒ…(group-limited gating) ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์ ์šฉํ•˜์—ฌ ๋น„๋Œ€์นญ ๋„๋ฉ”์ธ ๋Œ€์—ญํญ ํฌ์›Œ๋”ฉ์„ ์ตœ์ ํ™” ์˜ˆ: NVLink โ†’ RDMA ๋ฐ์ดํ„ฐ ์ „์†ก ์ตœ์ ํ™” ํ›ˆ๋ จ ๋ฐ ์ถ”๋ก  ํ”„๋ฆฌํ•„๋ง(prefilling) ์ž‘์—…์— ์ ํ•ฉํ•œ ๋†’์€ ์ฒ˜๋ฆฌ๋Ÿ‰ ์ œ๊ณต ์ง€์—ฐ์‹œ๊ฐ„์— ๋ฏผ๊ฐํ•œ ์ถ”๋ก  ๋””์ฝ”๋”ฉ์„ ์œ„ํ•ด RDMA ์ „์šฉ ์ €์ง€์—ฐ ์ปค๋„ ํฌํ•จ ํ†ต์‹ -์—ฐ์‚ฐ ์˜ค๋ฒ„๋žฉ ๊ธฐ๋ฒ• ์ œ๊ณต (SM ๋ฆฌ์†Œ์Šค๋ฅผ ์ ์œ ํ•˜์ง€ ์•Š์Œ)

DeepSeek, FlashMLA ์˜คํ”ˆ์†Œ์Šค ๊ณต๊ฐœ (1 of 5) (github.com/deepseek-ai)

  • https://news.hada.io/topic?id=19401 5P by xguru 4์ผ์ „ | โ˜… favorite | ๋Œ“๊ธ€ 2๊ฐœ
  • Hopper GPU๋ฅผ ์œ„ํ•œ ํšจ์œจ์ ์ธ MLA ๋””์ฝ”๋”ฉ ์ปค๋„ ๊ฐ€๋ณ€ ๊ธธ์ด ์‹œํ€€์Šค ์„œ๋น™์„ ์œ„ํ•ด ์ตœ์ ํ™” ๋จ ํ˜„์žฌ ๋ฆด๋ฆฌ์ฆˆ ๋œ ๊ฒƒ BF16 64 ๋ธ”๋ก์‚ฌ์ด์ฆˆ Paged kvcache ๋ฒค์น˜๋งˆํฌ: CUDA 12.6์„ ์‚ฌ์šฉํ•˜์—ฌ H800 SXM5์—์„œ ๋ฉ”๋ชจ๋ฆฌ ๋ฐ”์šด๋“œ ๊ตฌ์„ฑ์—์„œ ์ตœ๋Œ€ 3000GB/s, ์—ฐ์‚ฐ ๋ฐ”์šด๋“œ ๊ตฌ์„ฑ์—์„œ 580 TFLOPS๋ฅผ ๋‹ฌ์„ฑ FlashAttention 2&3 ์™€ cutlass ์—์„œ ์˜๊ฐ์„ ๋ฐ›์Œ DeepSeek Open Infra ๋กœ ๊ณต๊ฐœ๋˜๋Š” 5๊ฐœ ์˜คํ”ˆ์†Œ์Šค ์ค‘ ์ฒซ๋ฒˆ์งธ ์ž„
  • https://github.com/deepseek-ai/FlashMLA

(250203)๋จธ์‹ ๋Ÿฌ๋‹ ๋ชจ๋“  ๋ชจ๋ธ ์„ค๋ช… ๐Ÿ‘ ๊ตฟ|๐Ÿ”|


huggingface.co ๋ชจ๋ธ ๋‹ค์šด ๋ฐ›๋Š” ๋ฐฉ๋ฒ•|๐Ÿ”|

pip3 install huggingface-hub

huggingface-cli download TheBloke/LLaMA-13b-GGUF llama-13b.Q4_K_M.gguf --local-dir . --local-dir-use-symlinks False

(241217)๋“œ๋””์–ด ์˜ฌ๋ผ์˜ด ์ด๊ฑธ ๋Ÿฌ์ŠคํŠธ ์ฝ”๋“œ๋กœ ๋งŒ๋“ค๋ฉด ๋Œ€๋ฐ•์ด์š” ใ…‹ใ…‹**GNโบ: C++์™€ CUDA๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ฒ˜์Œ๋ถ€ํ„ฐ LLM ์ถ”๋ก  ์—”์ง„ ๋งŒ๋“ค๊ธฐ**|๐Ÿ”|

  • https://andrewkchan.dev/posts/yalm.html
    • https://github.com/andrewkchan/yalm
    • C++์™€ CUDA๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ ์—†์ด LLM ์ถ”๋ก  ์—”์ง„์„ ๊ตฌ์ถ•ํ•˜๋Š” ๋ฐฉ๋ฒ•
    • ์ด๋ฅผ ํ†ตํ•ด LLM ์ถ”๋ก ์˜ ์ „์ฒด ์Šคํƒ์„ ์ดํ•ดํ•˜๊ณ , ๋‹ค์–‘ํ•œ ์ตœ์ ํ™”๊ฐ€ ์ถ”๋ก  ์†๋„์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ์‹ค๊ฐํ•  ์ˆ˜ ์žˆ์Œ
    • ๋ชฉํ‘œ : ๋‹จ์ผ CPU + GPU ์„œ๋ฒ„์—์„œ ๋‹จ์ผ ๋ฐฐ์น˜๋กœ ๋น ๋ฅด๊ฒŒ ์ถ”๋ก ํ•  ์ˆ˜ ์žˆ๋„๋ก ๋ชจ๋ธ์„ ๊ตฌํ˜„ํ•˜๊ณ  llama.cpp๋ณด๋‹ค ๋น ๋ฅธ ํ† ํฐ ์ฒ˜...

Run LLaMA inference on CPU, with Rust ๐Ÿฆ€๐Ÿš€๐Ÿฆ™|๐Ÿ”|


Fast ML inference & training for ONNX models in Rust(์ปดํ“จํ„ฐ ๋น„์ ผ ์ฐพ๋‹ค๊ฐ€ ์•Œ๊ฒŒ ๋จyolo)|๐Ÿ”|


Artificial_Intelligence(NLP, Natural Language Processing models and pipelines.)|๐Ÿ”|

Rust MachineLearningrustmascot|๐Ÿ”|

dfdx: shape checked deep learning in rust|๐Ÿ”|

Minimalist ML framework for Rust|๐Ÿ”|

https://github.com/huggingface/candle



ollama ์“ธ๋งŒํ•œ๊ฑฐ|๐Ÿ”|

# llama3.3(๊ฐ€์ •์šฉ ์ปดํ“จํ„ฐ๋กœ 405B๋ชจ๋ธ์„ ๊ฒฝํ—˜ ๊ฐ€๋Šฅ ์ง€๊ธˆ์€ ์•„์ฃผ ๋А๋ฆฌ๋‹ค. 241212
# New state of the art 70B model. Llama 3.3 70B offers similar performance compared to Llama 3.1 405B model.
ollama run llama3.3

# ollam ํ”„๋กœ์„ธ์„œ ์ž˜ ์‹คํ–‰ ๋˜๋Š”์ง€ ํ™•์ธ
# pgrep ollama
6327
6521

# ollam "/bye" ๋กœ ์ข…๋ฃŒ ์‹œํ‚ค๊ณ  ์„œ๋น„์Šค ์ข…๋ฃŒ ์‹œํ‚ค๊ธฐ 
$ systemctl stop ollama.service

# 4.7GB
ollama run llama3.1

# 26GB
ollama run mixtral:8x7b

# 39GB
ollama run llama3.1:70b

# 79GB
ollama run mixtral:8x22b

(C++์ฝ”๋“œ๋กœ ๋จธ์‹ ๋Ÿฌ๋‹ ์ž˜ ์„ค๋ช…๋จ.)Snake learns with NEUROEVOLUTION (implementing NEAT from scratch in C++) |Tech With Nikola|๐Ÿ”|


1bit์— ์ง‘์ค‘ํ•˜์ž NVIDIA๋„ ์ด์ œ ๋์ด๋„ค|๐Ÿ”|

Blackwell Hopper
Supported Tensor Core precisions FP64, TF32, BF16, FP16, FP8, INT8, FP6, FP4 FP64, TF32, BF16, FP16, FP8, INT8
Supported CUDA* Core precisions FP64, FP32, FP16, BF16 FP64, FP32, FP16, BF16, INT8

Microsoft, CPU์—์„œ ์‹คํ–‰๊ฐ€๋Šฅํ•œ ์ดˆ๊ณ ํšจ์œจ AI ๋ชจ๋ธ BitNet ๊ฐœ๋ฐœ|๐Ÿ”|

  • Microsoft ์—ฐ๊ตฌ์ง„์ด BitNet b1.58 2B4T๋ผ๋Š” ์ดˆํšจ์œจ์ ์ธ AI ๋ชจ๋ธ์„ ๊ฐœ๋ฐœํ–ˆ์Œ
  • 1๋น„ํŠธ ์–‘์žํ™”๋ฅผ ํ†ตํ•ด ๋†’์€ ์†๋„์™€ ๋‚ฎ์€ ๋ฉ”๋ชจ๋ฆฌ ์‚ฌ์šฉ๋Ÿ‰ ๋‹ฌ์„ฑํ•˜์—ฌ CPU์—์„œ๋„ ์‹คํ–‰ ๊ฐ€๋Šฅํ•˜๋ฉฐ MIT ๋ผ์ด์„ ์Šค๋กœ ๊ณต๊ฐœ๋จ
  • Applโ€ฆ
  • https://news.hada.io/topic?id=20406

NVIDIA์นฉ ์ž์„ธํžˆ ์•Œ์•„๋ณด๊ธฐ(240617)|๐Ÿ”|

NVIDIA๋Š” 16-bit Float(FP16/BF16) ๋ถ€๋™ ์†Œ์ˆ˜์ ์— ์ตœ์ ํ™” ๋˜์–ด์žˆ์–ด์„œ|๐Ÿ”|

  • ์™„์ „ํžˆ ๋‹ค๋ฅธ ๋ฐฉ์‹์œผ๋กœ ์ ‘๊ทผํ•˜๊ณ  ์žˆ๋‹ค.
    • Develop optimized kernels for 1-bit operations
    • Use FPGAs or ASICs for 1-bit operations

BitNet b1.58(This Work). vs 16-bit Float(FP16/BF16)|๐Ÿ”|

  • 9min 46s ์ฐธ๊ณ 

Why BitNet b1.58?|๐Ÿ”|

  • Each cell only three values:
    • { -1, 0 ,1 }
    • How many bits are needed to differentiate three equally likely states?

$$Log_2(3) = 1.58$$

(24๋…„ 04์›”๊ฒฝ์ฏค)GNโบ: 1๋น„ํŠธ LLM ์‹œ๋Œ€: ๋น„์šฉ ํšจ์œจ์ ์ธ ์ปดํ“จํŒ…์„ ์œ„ํ•œ ์‚ผ์ง„ ํŒŒ๋ผ๋ฏธํ„ฐ (arxiv.org)|๐Ÿ”|


๋ฒกํ„ฐ DB์˜ ๊ฐœ๋…์žก๊ธฐ & LLM์˜ ์ •์˜|๐Ÿ”|

  • ์ถœ์ฒ˜ : http://www.itdaily.kr/news/articleView.html?idxno=220008

  • LLM์€

    • ๋”ฅ๋Ÿฌ๋‹์˜ ํ•œ ์ข…๋ฅ˜๋กœ ๋ฒกํ„ฐ ๋ฐ์ดํ„ฐ๋กœ ์ด๋ฃจ์–ด์ง„ ์–ธ์–ด๋ชจ๋ธ์ด๋ผ๋Š” ๊ฒƒ์ด๋‹ค. ์‹ค์ œ ๋ฒกํ„ฐ ๋ฐ์ดํ„ฐ๋กœ ์ด๋ค„์ง„ LLM์—๋Š” ์ด๋ฏธ ๋ฐ์ดํ„ฐ๋“ค์„ ์ €์žฅํ•˜๋Š” ๋ฒกํ„ฐ DB๊ฐ€ ๋‚ด์žฅ๋ผ ์žˆ๋‹ค.
  • ๋ฒกํ„ฐ DB๋Š”

    • ์œ ์‚ฌํ•œ ๋ฒกํ„ฐ๊ฐ’๋ผ๋ฆฌ ๊ตฐ์ง‘์„ ํ˜•์„ฑํ•œ ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์„ ํ•™์Šต์‹œํ‚ค๊ธฐ ์œ„ํ•ด์„  ๋ฒกํ„ฐ ๋ฐ์ดํ„ฐ์— ํŠนํ™”๋œ ๋ฐ์ดํ„ฐ ์ €์žฅ์†Œ๊ฐ€ ํ•„์š”ํ•˜๋‹ค. ์ด๋Ÿฌํ•œ ์š”๊ตฌ์— ๋”ฐ๋ผ ์ •ํ˜•ํ™”๋œ ๋ฐ์ดํ„ฐ๊ฐ€ ์•„๋‹Œ ๋น„์ •ํ˜• ๋ฐ์ดํ„ฐ๋ฅผ ๋น ๋ฅด๊ฒŒ ๋ฒกํ„ฐํ™”(์ž„๋ฒ ๋”ฉ, Embedding)ํ•ด ์ €์žฅํ•˜๊ณ  ์ฝ์„ ์ˆ˜ ์žˆ๋Š” ๋ฒกํ„ฐ DB๊ฐ€ ๋“ฑ์žฅํ–ˆ๋‹ค.
  • ๋ฒกํ„ฐDB์˜ ์žฅ์  4๊ฐ€์ง€

    • โ–ณ๋ฒกํ„ฐ ๋ฐ์ดํ„ฐ ์ฒ˜๋ฆฌ
    • โ–ณ๋ฒกํ„ฐ ๊ฒ€์ƒ‰ ๋ฐ ์œ ์‚ฌ์„ฑ ๋ถ„์„
    • โ–ณ๋Œ€์šฉ๋Ÿ‰ ๋ฒกํ„ฐ ๋ฐ์ดํ„ฐ ์ฒ˜๋ฆฌ
    • โ–ณ๋ฒกํ„ฐ ๋ฐ์ดํ„ฐ ๊ฐฑ์‹  ๋“ฑ์— ํŠนํ™”๋œ ์žฅ์ ์„ ๊ฐ–๊ณ  ์žˆ๋‹ค.

    1. โ–ณ๋ฒกํ„ฐ ๋ฐ์ดํ„ฐ ์ฒ˜๋ฆฌ
    • ๋ฒกํ„ฐ ๋ฐ์ดํ„ฐ๋ฅผ ์ฒ˜๋ฆฌํ•˜๋Š” ๋ฐ ํŠนํ™”๋ผ ์žˆ๋‹ค. ์ƒ์„ฑํ˜• AI์—์„œ๋Š” ์ฃผ๋กœ ๋ฒกํ„ฐ ๋ฐ์ดํ„ฐ๊ฐ€ ์ด์šฉ๋œ๋‹ค. LLM ๋‚ด ํ† ํฐ์ด ๋ฒกํ„ฐํ™”๋ผ ๋‚ด์žฅ๋œ ๋ฒกํ„ฐ DB์— ์ €์žฅ๋ผ ์žˆ๊ธฐ ๋•Œ๋ฌธ์ด๋‹ค. ์ƒ์„ฑํ˜• AI์—์„œ๋Š” ์ฃผ๋กœ ์ด๋ฏธ์ง€, ์Œ์„ฑ, ํ…์ŠคํŠธ ๋“ฑ ๋น„์ •ํ˜• ๋ฐ์ดํ„ฐ๊ฐ€ ๋ฒกํ„ฐ ํ˜•ํƒœ๋กœ ๋ณ€ํ™˜๋ผ ์ฒ˜๋ฆฌ๋˜๋Š”๋ฐ, ๋ฒกํ„ฐ DB๋Š” ๋ฒกํ„ฐ ๋ฐ์ดํ„ฐ๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ์ €์žฅํ•˜๊ณ  ๊ด€๋ฆฌํ•  ์ˆ˜ ์žˆ๋‹ค.
    1. โ–ณ๋ฒกํ„ฐ ๊ฒ€์ƒ‰ ๋ฐ ์œ ์‚ฌ์„ฑ ๋ถ„์„
    • ๋ฒกํ„ฐ ๊ฒ€์ƒ‰ ๋ฐ ์œ ์‚ฌ์„ฑ ๋ถ„์„์— ํŠนํ™”๋œ ๊ธฐ๋Šฅ์„ ์ œ๊ณตํ•œ๋‹ค๋Š” ์ ์ด๋‹ค. ์ด์— ๋Œ€ํ•ด EDB ์ธก ๊ด€๊ณ„์ž๋Š” โ€œ์ƒ์„ฑํ˜• AI์—์„œ๋Š” ๋ฒกํ„ฐ ๋ฐ์ดํ„ฐ ๊ฐ„์˜ ์œ ์‚ฌ์„ฑ์„ ๋ถ„์„ํ•˜๊ฑฐ๋‚˜ ๋ฒกํ„ฐ ๋ฐ์ดํ„ฐ ๊ฒ€์ƒ‰์ด ํ•„์š”ํ•œ ๊ฒฝ์šฐ๊ฐ€ ๋งŽ๋‹ค. ๋ฒกํ„ฐ DB๋Š” ์ด๋Ÿฌํ•œ ์š”๊ตฌ์‚ฌํ•ญ์— ํŠนํ™”๋œ ๊ธฐ๋Šฅ์„ ์ œ๊ณตํ•ด ๋ฒกํ„ฐ ๋ฐ์ดํ„ฐ ๊ฐ„ ์œ ์‚ฌ์„ฑ์„ ํšจ๊ณผ์ ์œผ๋กœ ๊ณ„์‚ฐํ•˜๊ณ  ๊ฒ€์ƒ‰ํ•  ์ˆ˜ ์žˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ƒ์„ฑํ˜• AI์—์„œ๋Š” ๋ฒกํ„ฐ DB๋ฅผ ํ†ตํ•ด ๋น ๋ฅด๊ณ  ์ •ํ™•ํ•œ ๋ฒกํ„ฐ ๊ฒ€์ƒ‰ ๋ฐ ์œ ์‚ฌ์„ฑ ๋ถ„์„์ด ๊ฐ€๋Šฅํ•˜๋‹คโ€๊ณ  ์„ค๋ช…ํ–ˆ๋‹ค.
    1. โ–ณ๋Œ€์šฉ๋Ÿ‰ ๋ฒกํ„ฐ ๋ฐ์ดํ„ฐ ์ฒ˜๋ฆฌ
    • ์„ธ ๋ฒˆ์งธ๋กœ๋Š” ๋Œ€์šฉ๋Ÿ‰ ๋ฒกํ„ฐ ๋ฐ์ดํ„ฐ ์ฒ˜๋ฆฌ๊ฐ€ ๊ฐ€๋Šฅํ•˜๋‹ค๋Š” ์ ์ด๋‹ค. ์ƒ์„ฑํ˜• AI ๋ชจ๋ธ์€ ๋Œ€์šฉ๋Ÿ‰์˜ ๋ฒกํ„ฐ ๋ฐ์ดํ„ฐ๋ฅผ ๋‹ค๋ฃฌ๋‹ค. ๋ฒกํ„ฐ DB๋Š” ๋Œ€์šฉ๋Ÿ‰์˜ ๋ฒกํ„ฐ ๋ฐ์ดํ„ฐ๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ์ €์žฅํ•˜๊ณ  ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ๋Š” ๊ธฐ๋Šฅ์„ ์ œ๊ณตํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ƒ์„ฑํ˜• AI ๋ชจ๋ธ์˜ ๋Œ€๊ทœ๋ชจ ๋ฐ์ดํ„ฐ์— ๋Œ€์‘ํ•  ์ˆ˜ ์žˆ๋‹ค. ๋„๋ฆฌ ์“ฐ์ด๋Š” ๊ด€๊ณ„ํ˜• ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค(RDB)๋Š” ๋ฒกํ„ฐ ๋ฐ์ดํ„ฐ์˜ ๊ธธ์ด๋‚˜ ํ˜•ํƒœ์˜ ๋‹ค์–‘์„ฑ, ์ฟผ๋ฆฌ ๋ฐ ์ธ๋ฑ์‹ฑ, ๋ฐ์ดํ„ฐ ๋ชจ๋ธ ๋ฌด๊ฒฐ์„ฑ ์ œ์•ฝ ๋“ฑ ๋•Œ๋ฌธ์— ๋ฒกํ„ฐ ๋ฐ์ดํ„ฐ๋ฅผ ์ €์žฅํ•˜๊ธฐ์— ์ ํ•ฉํ•˜์ง€ ์•Š๋‹ค.
    1. โ–ณ๋ฒกํ„ฐ ๋ฐ์ดํ„ฐ ๊ฐฑ์‹ 
    • ๋ฒกํ„ฐ DB๋Š” ๋ฒกํ„ฐ ๋ฐ์ดํ„ฐ์˜ ๊ฐฑ์‹  ๋ฐ ์ฟผ๋ฆฌ์— ๋Œ€ํ•œ ๊ธฐ๋Šฅ์„ ์ œ๊ณตํ•œ๋‹ค. ๋ฒกํ„ฐ DB๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ๋‹จ๋…์œผ๋กœ ์“ฐ์ด์ง€ ์•Š๋Š”๋‹ค. ๋Œ€๋ถ€๋ถ„์˜ ๊ฒฝ์šฐ LLM์„ ๋ฒกํ„ฐ DB๊ฐ€ ์—ฐ๊ฒฐ๋œ ๋žญ์ฒด์ธ(LangChain)์ด๋ผ๋Š” ์–ธ์–ด ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•˜๊ณ  ์ €์žฅํ•˜๋Š” ํ”Œ๋žซํผ์„ ์—ฐ๊ฒฐํ•ด ์ด์šฉํ•˜๋Š” ๊ตฌ์กฐ๋‹ค. ๋ฐ์ดํ„ฐ ์†Œ์Šค. ๋‹จ์–ด ์ž„๋ฒ ๋”ฉ, ๋ฒกํ„ฐ DB ๋“ฑ์„ LLM๊ณผ ์—ฐ๊ฒฐํ•˜๋Š” ๋งค๊ฐœ๋ผ๊ณ  ๋ณผ ์ˆ˜ ์žˆ๋‹ค. ๊ฐœ๋ณ„๋กœ ๊ตฌ์ถ•๋œ ๋ฒกํ„ฐ DB์— ์ตœ์‹ ์˜ ๋ฐ์ดํ„ฐ๋ฅผ ์ž„๋ฒ ๋”ฉํ•ด ์ €์žฅํ•˜๋ฉด LLM ์žฌํ•™์Šต ํ•˜์ง€ ์•Š๊ณ ๋„ ์ตœ์‹  ๋ฐ์ดํ„ฐ๋ฅผ LLM์— ์ ์šฉํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋œ๋‹ค.
      • โ€œ์ƒ์„ฑํ˜• AI์˜ ์น˜๋ช…์ ์ธ ๋ฌธ์ œ๋กœ ๊ผฝํžˆ๋Š” ํ™˜๊ฐ ํ˜„์ƒ์€ ์ƒ๋‹น ๋ถ€๋ถ„ ๋ฐ์ดํ„ฐ ์ตœ์‹ ํ™” ๋ฌธ์ œ ๋•Œ๋ฌธ์— ๋ฐœ์ƒํ•œ๋‹ค. LLM์€ ํŠน์ • ์‹œ์ ๊นŒ์ง€ ํ•™์Šต๋œ ๋ฐ์ดํ„ฐ๋กœ ๊ตฌ์ถ•๋œ๋‹ค. ๋•Œ๋ฌธ์— ํ•™์Šต ๋ฐ์ดํ„ฐ๋ฅผ ๊พธ์ค€ํžˆ ์ตœ์‹ ํ™”ํ•ด์•ผ ํ•œ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํ•™์Šต ๋ฐ์ดํ„ฐ๋ฅผ ์ตœ์‹ ํ™” ํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ์ธํ”„๋ผ ๋น„์šฉ, GPU ๋น„์šฉ, ์ธ๋ ฅ ํˆฌ์ž… ๋“ฑ ๋งŽ์€ ๋น„์šฉ๊ณผ ์‹œ๊ฐ„์ด ํ•„์š”ํ•˜๋‹ค. ์ด๋Ÿฐ ๋ฌธ์ œ์˜ ์ƒ๋‹น๋ถ€๋ถ„์„ ๋ฒกํ„ฐ DB๋กœ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ๋‹คโ€
  • ๋ฒกํ„ฐDB์˜ ์ฐจ์ด์ 

    • ๋ฒกํ„ฐ DB๋Š” ์‚ฌ์‹ค ํƒ€ DBMS์™€ ํฐ ์ฐจ์ด๊ฐ€ ์—†๋‹ค. ๋‹ค๋งŒ ๋‹ค๋ฃจ๋Š” ๋ฐ์ดํ„ฐ์˜ ์„ฑ๊ฒฉ๊ณผ ์ฒ˜๋ฆฌ ๋ฐฉ๋ฒ•์ด ๋‹ค๋ฅด๋‹ค. ๋ฒกํ„ฐ DB๋Š” ์ฃผ๋กœ ์‹ค์ˆ˜(Real Number) ํ˜•ํƒœ์˜ ๋ฐ์ดํ„ฐ๊ฐ€ ํฌํ•จ๋œ๋‹ค. ๋˜ ์‹ค์ˆ˜ ํ˜•ํƒœ์˜ ๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์œ ์‚ฌ๋„๊ฐ€ ๋†’์€ ๊ฒฐ๊ณผ๋ฅผ ์ถ”์ถœํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ๋ฐฉ๋ฒ•์„ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ๋‹คโ€๋ฉด์„œ โ€œ๋ฒกํ„ฐํ™”๋œ ๋ฐ์ดํ„ฐ ๊ฐ„์˜ ์œ ์‚ฌ๋„๋ฅผ ์ธก์ •ํ•˜๋Š” ๋ฐ์—๋Š” ์ฃผ๋กœ โ€˜์ฝ”์‚ฌ์ธ ์œ ์‚ฌ๋„(Cosine Similarity)โ€™์™€ โ€˜์œ ํฌ๋ฆฌ๋“œ ๊ฑฐ๋ฆฌ(Euclidean Distance)โ€™ ๋“ฑ์˜ ์ธก์ • ๋ฐฉ๋ฒ•์ด ํ™œ์šฉ๋œ๋‹ค. ์ฝ”์‚ฌ์ธ ์œ ์‚ฌ๋„๋Š” ๋‘ ๋ฒกํ„ฐ ์‚ฌ์ž‡๊ฐ์„ ํ†ตํ•ด ๋ฒกํ„ฐ ๋ฐ์ดํ„ฐ๊ฐ€ ์–ผ๋งˆ๋‚˜ ์œ ์‚ฌ๋„๊ฐ€ ์žˆ๋Š”์ง€ ์ธก์ •ํ•˜๋Š” ๋ฐฉ๋ฒ•์ด๋ฉฐ, ์œ ํฌ๋ฆฌ๋“œ ๊ฑฐ๋ฆฌ๋Š” ํ‰๋ฉด์—์„œ์˜ ๋‘ ๋ฒกํ„ฐ๊ฐ’ ์‚ฌ์ด์˜ ์ง์„ ๊ฑฐ๋ฆฌ๋ฅผ ์ธก์ •ํ•ด ๊ฐ’ ์‚ฌ์ด์˜ ์œ ์‚ฌ๋„๋ฅผ ํŒŒ์•…ํ•˜๋Š” ๋ฐฉ๋ฒ•์ด๋‹คโ€๋ผ๊ณ  ์„ค๋ช…ํ–ˆ๋‹ค.
    • โ€œ๋ฒกํ„ฐ DB๋Š” ์˜๋ฏธ ๊ธฐ๋ฐ˜์˜ ์ฟผ๋ฆฌ๋ฅผ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•œ๋‹ค. ๊ธฐ์กด DBMS๋Š” ๊ฐ ์Šคํ‚ค๋งˆ์˜ ๊ด€๊ณ„๋ฅผ ํ†ตํ•ด ๋ฐ์ดํ„ฐ๋ฅผ ํšจ์œจ์ ์œผ๋กœ ์ถ”์ถœํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘์ง€๋งŒ, ๋ฒกํ„ฐ DB๋Š” ์ˆ˜์น˜ํ™”๋œ ๋ฒกํ„ฐ์˜ ์˜๋ฏธ๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ์ถ”์ถœํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘”๋‹ค. ์ˆ˜์น˜ํ™”๋œ ๋ฐ์ดํ„ฐ ์ €์žฅ๊ณผ ์œ ์‚ฌ๋„ ์ธก์ •์„ ์œ„ํ•œ ๋‹ค์–‘ํ•œ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์ œ๊ณตํ•˜๋Š” DB๊ฐ€ ๋ฐ”๋กœ ๋ฒกํ„ฐ DB๋‹คโ€

์ดˆ๋ณด์ž๋ฅผ ์œ„ํ•œ Vector Embeddings ๊ฐ€์ด๋“œ (timescale.com)|๐Ÿ”|

  • https://news.hada.io/topic?id=15094&utm_source=weekly&utm_medium=email&utm_campaign=202423

    • 26P by xguru 24.05.31.
  • ๋ฒกํ„ฐ ์ž„๋ฒ ๋”ฉ์˜ ์ข…๋ฅ˜

    • ๋‹จ์–ด ์ž„๋ฒ ๋”ฉ: NLP์—์„œ ๋‹จ์–ด๋ฅผ ํ‘œํ˜„ํ•˜๋ฉฐ, ๋‹จ์–ด ๊ฐ„์˜ ์˜๋ฏธ์  ๊ด€๊ณ„๋ฅผ ์บก์ฒ˜ํ•จ. ์–ธ์–ด ๋ฒˆ์—ญ, ๋‹จ์–ด ์œ ์‚ฌ์„ฑ, ๊ฐ์ • ๋ถ„์„ ๋“ฑ์— ์‚ฌ์šฉ๋จ.
    • ๋ฌธ์žฅ ์ž„๋ฒ ๋”ฉ: ๋ฌธ์žฅ์˜ ์˜๋ฏธ์™€ ๋ฌธ๋งฅ์„ ์บก์ฒ˜ํ•˜๋ฉฐ, ์ •๋ณด ๊ฒ€์ƒ‰, ํ…์ŠคํŠธ ๋ถ„๋ฅ˜, ๊ฐ์ • ๋ถ„์„ ๋“ฑ์— ์‚ฌ์šฉ๋จ.
    • ๋ฌธ์„œ ์ž„๋ฒ ๋”ฉ: ๋ณด๊ณ ์„œ๋‚˜ ๊ธฐ์‚ฌ ๊ฐ™์€ ๋ฌธ์„œ์˜ ๋‚ด์šฉ์„ ์บก์ฒ˜ํ•˜๋ฉฐ, ์ถ”์ฒœ ์‹œ์Šคํ…œ, ์ •๋ณด ๊ฒ€์ƒ‰, ๋ฌธ์„œ ์œ ์‚ฌ์„ฑ ๋ฐ ๋ถ„๋ฅ˜ ๋“ฑ์— ์‚ฌ์šฉ๋จ.
    • ๊ทธ๋ž˜ํ”„ ์ž„๋ฒ ๋”ฉ: ๊ทธ๋ž˜ํ”„์˜ ๋…ธ๋“œ์™€ ์—ฃ์ง€๋ฅผ ๋ฒกํ„ฐ ๊ณต๊ฐ„์— ํ‘œํ˜„ํ•˜๋ฉฐ, ๋…ธ๋“œ ๋ถ„๋ฅ˜, ์ปค๋ฎค๋‹ˆํ‹ฐ ์ธ์‹, ๋งํฌ ์˜ˆ์ธก ๋“ฑ์— ์‚ฌ์šฉ๋จ.
    • ์ด๋ฏธ์ง€ ์ž„๋ฒ ๋”ฉ: ์ด๋ฏธ์ง€์˜ ๋‹ค์–‘ํ•œ ์ธก๋ฉด์„ ํ‘œํ˜„ํ•˜๋ฉฐ, ์ฝ˜ํ…์ธ  ๊ธฐ๋ฐ˜ ์ถ”์ฒœ ์‹œ์Šคํ…œ, ์ด๋ฏธ์ง€ ๋ฐ ๊ฐ์ฒด ์ธ์‹, ์ด๋ฏธ์ง€ ๊ฒ€์ƒ‰ ์‹œ์Šคํ…œ ๋“ฑ์— ์‚ฌ์šฉ๋จ.
    • ์ œํ’ˆ ์ž„๋ฒ ๋”ฉ: ๋””์ง€ํ„ธ ์ œํ’ˆ์ด๋‚˜ ๋ฌผ๋ฆฌ์  ์ œํ’ˆ์„ ํ‘œํ˜„ํ•˜๋ฉฐ, ์ œํ’ˆ ์ถ”์ฒœ ๋ฐ ๋ถ„๋ฅ˜ ์‹œ์Šคํ…œ, ์ œํ’ˆ ๊ฒ€์ƒ‰ ๋“ฑ์— ์‚ฌ์šฉ๋จ.
    • ์˜ค๋””์˜ค ์ž„๋ฒ ๋”ฉ: ์˜ค๋””์˜ค ์‹ ํ˜ธ์˜ ๋ฆฌ๋“ฌ, ํ†ค, ํ”ผ์น˜ ๋“ฑ์„ ํ‘œํ˜„ํ•˜๋ฉฐ, ๊ฐ์ • ๊ฐ์ง€, ์Œ์„ฑ ์ธ์‹, ์Œ์•… ์ถ”์ฒœ ๋“ฑ์— ์‚ฌ์šฉ๋จ.
  • ์‹ ๊ฒฝ๋ง์ด ์ž„๋ฒ ๋”ฉ์„ ์ƒ์„ฑํ•˜๋Š” ๋ฐฉ๋ฒ•

    • ํ‘œํ˜„ ํ•™์Šต: ์‹ ๊ฒฝ๋ง์ด ๊ณ ์ฐจ์› ๋ฐ์ดํ„ฐ๋ฅผ ์ €์ฐจ์› ๊ณต๊ฐ„์œผ๋กœ ๋งคํ•‘ํ•˜์—ฌ ์ค‘์š”ํ•œ ํŠน์„ฑ์„ ๋ณด์กดํ•จ.
    • ํ›ˆ๋ จ ๊ณผ์ •: ์‹ ๊ฒฝ๋ง์ด ๋ฐ์ดํ„ฐ๋ฅผ ์˜๋ฏธ ์žˆ๋Š” ์ž„๋ฒ ๋”ฉ์œผ๋กœ ๋ณ€ํ™˜ํ•˜๋„๋ก ํ•™์Šตํ•จ. ์ด๋Š” ๋‰ด๋Ÿฐ์˜ ๊ฐ€์ค‘์น˜์™€ ๋ฐ”์ด์–ด์Šค๋ฅผ ์กฐ์ •ํ•˜๋Š” ๊ณผ์ •์—์„œ ์ด๋ฃจ์–ด์ง.
    • ์˜ˆ์‹œ: ์˜ํ™” ๋ฆฌ๋ทฐ์˜ ๊ธ์ •/๋ถ€์ • ๋ถ„๋ฅ˜๋ฅผ ์œ„ํ•œ ์‹ ๊ฒฝ๋ง์—์„œ ๋‹จ์–ด ์ž„๋ฒ ๋”ฉ์ด ํ•™์Šต๋จ. "good"๊ณผ "excellent" ๊ฐ™์€ ๋‹จ์–ด๋Š” ์œ ์‚ฌํ•œ ์ž„๋ฒ ๋”ฉ์„ ๊ฐ€์ง€๊ฒŒ ๋จ.
  • ๋ฒกํ„ฐ ์ž„๋ฒ ๋”ฉ์˜ ์ž‘๋™ ์›๋ฆฌ

    • ๋ฒกํ„ฐ ๊ณต๊ฐ„: ๊ฐ์ฒด๋‚˜ ํŠน์ง•์„ ๋‹ค์ฐจ์› ๋ฒกํ„ฐ ๊ณต๊ฐ„์˜ ์ ์œผ๋กœ ํ‘œํ˜„ํ•˜๋ฉฐ, ์œ ์‚ฌํ•œ ํ•ญ๋ชฉ์€ ๊ฐ€๊นŒ์ด ์œ„์น˜ํ•จ.
    • ๊ฑฐ๋ฆฌ ์ธก์ •: ์œ ํด๋ฆฌ๋“œ ๊ฑฐ๋ฆฌ, ์ฝ”์‚ฌ์ธ ์œ ์‚ฌ๋„ ๋“ฑ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ฒกํ„ฐ ๊ฐ„์˜ ๊ด€๊ณ„๋ฅผ ์ •๋Ÿ‰ํ™”ํ•จ.
    • ์˜ˆ์‹œ: "cat"๊ณผ "dog"์˜ ๋ฒกํ„ฐ๋Š” "cat"๊ณผ "car"์˜ ๋ฒกํ„ฐ๋ณด๋‹ค ๋” ๊ฐ€๊นŒ์ด ์œ„์น˜ํ•จ.
  • ๋ฒกํ„ฐ ์ž„๋ฒ ๋”ฉ์„ ํ™œ์šฉํ•œ ๊ฐœ๋ฐœ

    • ์ฑ—๋ด‡: ์‚ฌ์šฉ์ž ์ฟผ๋ฆฌ์— ๋” ์ž˜ ์‘๋‹ตํ•˜๊ณ , ๋ฌธ๋งฅ์ ์œผ๋กœ ๊ด€๋ จ๋œ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๋ฉฐ, ์ผ๊ด€๋œ ๋Œ€ํ™”๋ฅผ ์œ ์ง€ํ•จ.
    • ์‹œ๋งจํ‹ฑ ๊ฒ€์ƒ‰ ์—”์ง„: ํ‚ค์›Œ๋“œ ๋งค์นญ ๋Œ€์‹  ์˜๋ฏธ์  ์œ ์‚ฌ์„ฑ์— ๊ธฐ๋ฐ˜ํ•œ ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ๋ฅผ ์ œ๊ณตํ•จ.
    • ํ…์ŠคํŠธ ๋ถ„๋ฅ˜ ์‹œ์Šคํ…œ: ๋ฌธ์„œ๋ฅผ ๊ตฌ๋ฌธ๊ณผ ๋‹จ์–ด์— ๋”ฐ๋ผ ๋ถ„๋ฅ˜ํ•จ.
    • ์ถ”์ฒœ ์‹œ์Šคํ…œ: ํ‚ค์›Œ๋“œ์™€ ์„ค๋ช…์˜ ์œ ์‚ฌ์„ฑ์— ๋”ฐ๋ผ ์ฝ˜ํ…์ธ ๋ฅผ ์ถ”์ฒœํ•จ.
  • ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ๋ฒกํ„ฐ ์ž„๋ฒ ๋”ฉ ์ƒ์„ฑ ๋ฐฉ๋ฒ•

    • ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘: ํ…์ŠคํŠธ, ์˜ค๋””์˜ค, ์ด๋ฏธ์ง€, ์‹œ๊ณ„์—ด ๋ฐ์ดํ„ฐ ๋“ฑ ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•จ.
    • ๋ฐ์ดํ„ฐ ์ „์ฒ˜๋ฆฌ: ํ† ํฐํ™”, ๋…ธ์ด์ฆˆ ์ œ๊ฑฐ, ์ด๋ฏธ์ง€ ํฌ๊ธฐ ์กฐ์ •, ์ •๊ทœํ™” ๋“ฑ ๋ฐ์ดํ„ฐ๋ฅผ ๋ถ„์„์— ์ ํ•ฉํ•˜๊ฒŒ ์ฒ˜๋ฆฌํ•จ.
    • ๋ฐ์ดํ„ฐ ๋ถ„ํ• : ํ…์ŠคํŠธ๋ฅผ ๋ฌธ์žฅ์ด๋‚˜ ๋‹จ์–ด๋กœ, ์ด๋ฏธ์ง€๋ฅผ ์„ธ๊ทธ๋จผํŠธ๋กœ, ์‹œ๊ณ„์—ด ๋ฐ์ดํ„ฐ๋ฅผ ๊ฐ„๊ฒฉ์œผ๋กœ ๋‚˜๋ˆ”.
    • ๋ฒกํ„ฐํ™”: ๊ฐ ๋ฐ์ดํ„ฐ ์กฐ๊ฐ์„ ๋ฒกํ„ฐ๋กœ ๋ณ€ํ™˜ํ•จ. ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ๋Š” OpenAI์˜ ํ…์ŠคํŠธ ์ž„๋ฒ ๋”ฉ ๋ชจ๋ธ, ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ๋Š” CNN ๋ชจ๋ธ, ์˜ค๋””์˜ค ๋ฐ์ดํ„ฐ๋Š” ์ŠคํŽ™ํŠธ๋กœ๊ทธ๋žจ ๋“ฑ์„ ์‚ฌ์šฉํ•จ.
  • ๋ฒกํ„ฐ ์ž„๋ฒ ๋”ฉ ์ €์žฅ ๋ฐฉ๋ฒ•

    • ๋ฒกํ„ฐ ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค: ๋ฒกํ„ฐ ๋ฐ์ดํ„ฐ๋ฅผ ํšจ์œจ์ ์œผ๋กœ ์ €์žฅํ•˜๊ณ  ๊ฒ€์ƒ‰ํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค ์‚ฌ์šฉ.
    • PostgreSQL: ๋ฒกํ„ฐ ๋ฐ์ดํ„ฐ๋ฅผ ๋‹ค๋ฅธ ๊ด€๊ณ„ํ˜• ๋ฐ์ดํ„ฐ์™€ ํ•จ๊ป˜ ์ €์žฅํ•  ์ˆ˜ ์žˆ์Œ. pgvector ํ™•์žฅ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ฒกํ„ฐ๋ฅผ ์ €์žฅํ•˜๊ณ  ์ฟผ๋ฆฌํ•  ์ˆ˜ ์žˆ์Œ.
  • ๊ทธ ์™ธ์— ์ข‹์€๊ธ€



์—ญ์‹œ ๊ฐ“ c์–ธ์–ด|๐Ÿ”|

  • llm.c, ์ด์ œ ๋ฉ€ํ‹ฐGPU ํŠธ๋ ˆ์ด๋‹์„ ์ง€์›ํ•˜๋ฉฐ PyTorch๋ณด๋‹ค ~7% ๋น ๋ฆ„
  • Andrej Karpathy๊ฐ€ ์ˆœ์ˆ˜ C/CUDA๋กœ ๋งŒ๋“  ๊ฐ„๋‹จํ•œ LLM ํ›ˆ๋ จ ์ฝ”๋“œ
  • ์ด์ œ ๋ฉ€ํ‹ฐ GPU ํŠธ๋ ˆ์ด๋‹์„ bfloat16์œผ๋กœ Flash Attention๊ณผ ํ•จ๊ป˜ ์ˆ˜ํ–‰
  • ~3000 ๋ผ์ธ์˜ C/CUDA ์ฝ”๋“œ๋กœ ๊ตฌํ˜„๋˜์—ˆ์œผ๋ฉฐ, ์ „๋ฐ˜์ ์œผ๋กœ PyTorch๋ณด๋‹ค 7% ์ •๋„๊นŒ์ง€ ๋น ๋ฆ„
  • ์ง€๊ธˆ๊นŒ์ง€ ์ž‘์—…ํ•œ ๋‚ด์šฉ๋“ค
    • ํ˜ผํ•ฉ ์ •๋ฐ€๋„ ํ›ˆ๋ จ(bfloat16)
    • ์ •๊ทœํ™”๋œ...

ํŒŒ์ดํ† ์น˜ bye bye ๐Ÿ‘‹ ์กด๋‚˜๊ฒŒ ๊ตฌ๋ฆฐ ํŒŒ์ดํ† ์น˜ ใ…‹ใ…‹ใ…‹ ๊ทธ๋™์•ˆ ์ฐธ๊ณ  ์“ฐ๋А๋ผ ํž˜๋“ค์—ˆ๋‹ค ใ…‹ใ…‹ ๋”๋Ÿฝ๊ณ  ์น˜์‚ฌํ•ด์„œ ๋” ๊ณต๋ถ€ํ•ด์„œ ๋Ÿฌ์ŠคํŠธ๋กœ ๋งŒ๋“ค์–ด ๋ณด์ž ใ…‹ใ…‹|๐Ÿ”|

๋ฐ”๋กœ ํ•ด๋ด์•ผ์ง€|๐Ÿ”|


MachineLearning_Tutorial|๐Ÿ”|


LLM -> LMM์œผ๋กœ ํŒจ๋Ÿฌ๋‹ค์ž„ ์ „ํ™˜ ์ค‘~~|๐Ÿ”|


Jupyter ๋…ธํŠธ๋ถ ๋Ÿฌ์ŠคํŠธ๋กœ ๋น ๋ฅด๊ฒŒ ๋Œ๋ฆฌ๊ธฐ|๐Ÿ”|

$ cargo install --locked evcxr_jupyter
  • Then, use its binary to automatically install it inside Jupyter:
$ evcxr_jupyter --install

Rust+WASM์œผ๋กœ ์ด๊ธฐ์ข… Edge์—์„œ ๋น ๋ฅด๊ณ  ํฌํ„ฐ๋ธ”ํ•œ Llama2 ์ถ”๋ก  ์‹คํ–‰ํ•˜๊ธฐ (secondstate.io)|๐Ÿ”|


m1 macOS pytorch install

https://pytorch.org/get-started/locally/


h2oGPT - ์™„์ „ํ•œ ์˜คํ”ˆ์†Œ์Šค GPT (github.com/h2oai)|๐Ÿ”|

llama2๋ฅผ ํŒŒ์ธ ํŠœ๋‹ ํ•˜๊ณ  ์žˆ๋Š” ์‚ฌ๋žŒ๋“ค|๐Ÿ”|

JS๊ฐ•์˜ No Black Box Machine Learning Course โ€“ Learn Without Libraries|๐Ÿ”|

https://youtu.be/vDDjtwQDw2k?si=exYH6L2aHAYEqGTJ


AlphaGo - The Movie | Full award-winning documentary|๐Ÿ”|

https://youtu.be/WXuK6gekU1Y?si=D9ZPN7Lxc6icN2g9


Building a neural network FROM SCRATCH (no Tensorflow/Pytorch, just numpy & math)|๐Ÿ”|

  • C์–ธ์–ด๋กœ Tesorflow/Pythorch ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ ์•ˆ ์“ฐ๊ณ  ์‹ ๊ฒฝ๋ง ๊ตฌ์ถ•ํ•˜๊ธฐ ๊ผญ ํ•ด๋ณด์žโค

https://youtu.be/w8yWXqWQYmU


์‹ ๊ฒฝ๋ง ์ˆ˜ํ•™ ๊ทธ๋ฆผ์œผ๋กœ ๋‹ค ์ดํ•ดํ•˜๊ธฐ - ๋ณต์žกํ•œ ์‹ ๊ฒฝ๋ง๋„ ๋‹ค ์ดํ•ด๋œ๋‹ค !!! ์ตœ๊ณ Why Neural Networks can learn (almost) anything | Emergent Garden|๐Ÿ”|

https://youtu.be/0QczhVg5HaI


Dalai - Automatically install, run, and play with LLaMA on your computer|๐Ÿ”|

  • What is Dalai?

It lets you one-click install LLaMA on your machine. No need to bother building cpp files, cloning GitHub, and downloading files and stuff. Everything is automated. Dalai is a tool in the Large Language Model Tools category of a tech stack. Dalai is an open source tool with GitHub stars and GitHub forks. Hereโ€™s a link to Dalai's open source repository on GitHub

https://cocktailpeanut.github.io/dalai/#/

https://stackshare.io/dalai?utm_source=weekly_digest&utm_medium=email&utm_campaign=03292023&utm_content=new_tool

  • ํ•œ๊ตญ์— ๋ˆ„๊ตฐ๊ฐ€ ์˜ฌ๋ฆฐ ๊ฒŒ์‹œํŒ ๊ธ€

https://www.ddengle.com/board_free/19129866

The Pile is a large, diverse, open source language modelling data set|๐Ÿ”|

https://github.com/EleutherAI/the-pile


brew install libtorch(macOS)|๐Ÿ”|

  • pytorch ์‹คํ–‰์ „ ์ด๊ฑฐ ๋จผ์ € ์‹คํ–‰ํ•  ๊ฒƒ !!!
export LIBTORCH='/opt/homebrew/Cellar/pytorch/1.13.1'

export LD_LIBRARY_PATH=$LIBTORCH:$LD_LIBRARY_PATH


echo $LD_LIBRARY_PATH
/opt/homebrew/Cellar/pytorch/1.13.1:

Rust Artificial Intelligence (The Simple Way)|๐Ÿ”|

https://youtu.be/StMP7g-0wK4

https://github.com/guillaume-be/rust-bert


The AI community building the future.|๐Ÿ”|

https://huggingface.co/



How to Build a Machine Learning Model in Rust|๐Ÿ”|

https://www.freecodecamp.org/news/how-to-build-a-machine-learning-model-in-rust/

Rust Machine Learning Book|๐Ÿ”|


Unicode (Vim Plug-in)|๐Ÿ”|

https://github.com/chrisbra/unicode.vim

  • Ex commands:
:UnicodeTable    - Print Unicode Table in new window
:Digraphs        - Search for specific digraph char
:UnicodeSearch   - Search for specific unicode char
:UnicodeSearch!  - Search for specific unicode char (and add at current cursor position)
:UnicodeName     - Identify character under cursor (like ga command)
:DownloadUnicode - Download (or update) Unicode data
:UnicodeCache    - Create cache file


  • Normal mode commands:

<C-X><C-G>  - Complete Digraph char
<C-X><C-Z>  - Complete Unicode char
<F4>        - Combine characters into digraphs
Scripting Functions:
unicode#FindUnicodeBy() - Find unicode characters
unicode#FindDigraphBy() - Find Digraph char
unicode#Digraph()       - Returns digraph char
unicode#UnicodeName()   - Identifies unicode character (by value)

Natural Language Processing for Rust|๐Ÿ”|

https://github.com/lexi-sh/rs-natural



This repository is a list of machine learning libraries written in Rust. It's a compilation of GitHub repositories, blogs, books, movies, discussions, papers, etc. ๐Ÿฆ€|๐Ÿ”|

https://github.com/vaaaaanquish/Awesome-Rust-MachineLearning



Here is an example of a simple machine learning program written in the Rust programming language:|๐Ÿ”|

  • GPT3

https://chat.openai.com/chat

This program sets up a game context and creates a new MainState struct to hold the game's state. It then runs the game loop, which handles events and updates the game state accordingly. The MainState struct and its associated methods and the event handling methods can be customized to implement the specific game logic and mechanics.


use ndarray::{Array, Array2};
use rand::{thread_rng, Rng};
use rusty_machine::linalg::{Matrix, Vector};
use rusty_machine::learning::k_means::KMeansClassifier;
use rusty_machine::learning::UnSupModel;

// Create some random data to cluster
let mut rng = thread_rng();
let data: Vec<Vector<f64>> = (0..100)
    .map(|_| Vector::new(vec![rng.gen(), rng.gen()]))
    .collect();
let data = Matrix::new(100, 2, data);

// Create a new k-means model with 2 clusters
let mut model = KMeansClassifier::new(2);

// Train the model on the data
model.train(&data)?;

// Predict the cluster assignments for the data
let assignments = model.predict(&data)?;

// Print the cluster assignments
println!("Cluster assignments: {:?}", assignments);
This program creates some random data and uses the KMeansClassifier type from the rusty_machine crate to train a k-means clustering model on the data. It then uses the trained model to predict the cluster assignments for the data and prints the results. The KMeansClassifier type and the train and predict methods can be customized to implement different machine learning algorithms and apply them to different types of data.




  • ์€๊ทผํžˆ ์ฝ”๋“œ ์•ˆ ๋งž๋Š”๋‹ค. ใ…‹

https://athemathmo.github.io/rusty-machine/doc/rusty_machine/index.html



How to Build a Machine Learning Model in Rust|๐Ÿ”|

https://www.freecodecamp.org/news/how-to-build-a-machine-learning-model-in-rust/


Machine_Learning_Rust|๐Ÿ”|

๋จธ์‹ ๋Ÿฌ๋‹๊ธฐ์ดˆ|๐Ÿ”|



์ž„๋ฒ ๋”ฉ๊ณผ ์ •๋ณด ๊ฒ€์ƒ‰ ์ „ ๊ณผ์ • โ€” ์ž„๋ฒ ๋”ฉ ๊ฐœ๋…๊ณผ ํ•œ๊ณ„, ๋ฐ์ดํ„ฐ์…‹ ์ƒ์„ฑยท๋ผ๋ฒจ๋ง, ๊ฐ์ข… ์˜คํ”„ ๋” ์…ธํ”„ ๋ชจ๋ธ ํ‰๊ฐ€, ํ•˜์ด๋ธŒ๋ฆฌ๋“œยท๋ฆฌ๋žญํ‚น, ์ž„๋ฒ ๋”ฉ ๋ชจ๋ธ ํŒŒ์ธํŠœ๋‹ ๋ฐ ์ตœ์ ํ™”, ํ•ด์„ ๊ฐ€๋Šฅ์„ฑ๊นŒ์ง€ โ€“์„ ์ฃผ์šฑ ์‚ดํŽด๋ณด๋Š” ๊ธ€์ž…๋‹ˆ๋‹ค.

๋‹ค๋ฃจ๋Š” ์ฃผ์ œ๋“ค

  • ์ž„๋ฒ ๋”ฉ๊ณผ ๊ทธ ์ผ๋ฐ˜ํ™” ๊ฐ€๋Šฅ์„ฑ(Generalizability)์— ๋Œ€ํ•œ ๋…ผ์˜
  • ์ธ๊ฐ„๊ณผ + ...

Compiling CUDA with clang|๐Ÿ”|

clang++ axpy.cu -o axpy --cuda-gpu-arch=<GPU arch> \
    -L<CUDA install path>/<lib64 or lib>             \
    -lcudart_static -ldl -lrt -pthread

./axpy
y[0] = 2
y[1] = 4
y[2] = 6
y[3] = 8
#include <iostream>

__global__ void axpy(float a, float* x, float* y) {
  y[threadIdx.x] = a * x[threadIdx.x];
}

int main(int argc, char* argv[]) {
  const int kDataLen = 4;

  float a = 2.0f;
  float host_x[kDataLen] = {1.0f, 2.0f, 3.0f, 4.0f};
  float host_y[kDataLen];

  // Copy input data to device.
  float* device_x;
  float* device_y;
  cudaMalloc(&device_x, kDataLen * sizeof(float));
  cudaMalloc(&device_y, kDataLen * sizeof(float));
  cudaMemcpy(device_x, host_x, kDataLen * sizeof(float),
             cudaMemcpyHostToDevice);

  // Launch the kernel.
  axpy<<<1, kDataLen>>>(a, device_x, device_y);

  // Copy output data to host.
  cudaDeviceSynchronize();
  cudaMemcpy(host_y, device_y, kDataLen * sizeof(float),
             cudaMemcpyDeviceToHost);

  // Print the results.
  for (int i = 0; i < kDataLen; ++i) {
    std::cout << "y[" << i << "] = " << host_y[i] << "\n";
  }

  cudaDeviceReset();
  return 0;
}

์ตœ์‹ ๋‰ด์Šค ๋ชจ์Œ

Rust๋กœ ๋งŒ๋“  ๋จธ์‹ ๋Ÿฌ๋‹ ๊ด€๋ จ ์ž๋ฃŒ ๋ชจ์Œ

Porting GPU shaders to Rust 30x faster with AI

  • CubeCL์€ Rust์—์„œ GPU ์ปค๋„์„ ์ž‘์„ฑํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•ด์ฃผ๋Š” ๊ณ ์„ฑ๋Šฅ ๋ฉ€ํ‹ฐํ”Œ๋žซํผ ์–ธ์–ด ํ™•์žฅ
  • ํ•จ์ˆ˜, ์ œ๋„ค๋ฆญ, ๊ตฌ์กฐ์ฒด๋ฅผ ์™„๋ฒฝํ•˜๊ฒŒ ์ง€์›ํ•˜๋ฉฐ, ํŠน์„ฑ, ๋ฉ”์„œ๋“œ, ํƒ€์ž… ์ถ”๋ก ์€ ๋ถ€๋ถ„์ ์œผ๋กœ ์ง€์›
  • WGPU, CUDA, ROCm ๊ธฐ๋ฐ˜ ๋Ÿฐํƒ€์ž„์„ ์ง€์›ํ•˜๋ฉฐ, SIMD ๋ช…๋ น์–ด๋ฅผ ํ™œ์šฉํ•œ ์ตœ์ ํ™”๋œ JIT CPU ๋Ÿฐํƒ€์ž„๋„ ๊ฐœ๋ฐœ์ค‘
  • *โ€ฆ
  • Llama.cpp๊ฐ€ ์ด์ œ libmtmd๋ฅผ ํ†ตํ•ด ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ์ž…๋ ฅ(๋น„์ „ ํฌํ•จ)์„ ์ง€์›ํ•จ
    • llama-mtmd-cli ๋˜๋Š” llama-server๋ฅผ ํ†ตํ•œ OpenAI ํ˜ธํ™˜ /chat/completions API
  • Gemma 3, SmolVLM, Pixtral, Qwen 2/2.5, Mistra Small, InternVL ๋“ฑ ๋ชจ๋ธ์—์„œ ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ๊ธฐ๋Šฅ ์ฆ‰์‹œ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•จ
    • Pre-quantized ๋ชจ๋ธโ€ฆ

About

Machine_Learning_Rust & My Youtube Channel - GlobalYoung https://www.youtube.com/@GlobalYoung7

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published