← 최신 프로덕트

Duplicate 3 layers in a 24B LLM

logical deduction .22→.76. No training. I replicated Ng's RYS method and found that duplicating 3 specific layers in Qwen2.5-32B boosts reasoning by 17% and duplicating layers 12-14 in Devstral-24B improves logical deduction from 0.22→0.76 on BBH — no training, no weight changes, just routing hidden states through the same circuit twice. Tools included. Two AMD GPUs, one evening. I replicated David Ng's RYS method ( https://dnhkng.github.io/posts/rys/ ) on consumer AMD GPUs (RX 7900 XT + RX 6950 XT) and found something I didn't expect.

오픈소스

244 스타 · GitHub외부 사이트

github.com 방문외부 사이트

반응 (1)

비슷한 프로덕트

설명이 비슷한 순

  1. How I topped the HuggingFace open LLM leaderboard on two gaming GPUs

    dnhkng.github.io

    ML, Biotech, Hardware, and Coordination Problems. Sometimes I write about hard problems and how to solve them.

    Hacker News495점

  2. Minimal LLM Post-Training Experiments on an 8GB GPU (SFT

    github.com

    DPO, GRPO)

    Hacker News21점

  3. Cuq

    github.com

    Formal Verification of Rust GPU Kernels

    Hacker News94점

  4. Speeding up LLM inference 2x times (possibly)

    asciinema.org

    Here's a project I've been working on for the last few months.

    Hacker News419점

  5. Can your GPU run this LLM?

    github.com

    Calculate token/s & GPU memory requirement for any LLM. Supports llama.cpp/ggml/bnb/QLoRA quantization

    Hacker News15점