Post
53
Big update to 🕸️ DaisyChain-Web - the browser demo where your spare devices pretrain a language model together, peer-to-peer. 🌼
Since launch, the demo has grown from a proof-of-concept into something much more real:
- Block-scaled INT8 quantization
- Batched attention GEMM
- Fused dequant+ReLU epilogue
- Weight-tied unembedding
- WebSocket relay fallback
- Server keepalive ping/pong every 30s
- disconnected-state redial
- Visibility/network-change reconnect (Phones that lock the screen or hop wifi↔cellular reconnect on resume.)
- DAISY_RTC_CONFIG - operators can supply their own TURN/ICE config via env var without touching client code.
- Split-K f32 backward
- Gather-fused attention
Net effect of this push: compute step 821ms -> 420ms (1.95×); full 2-device run 177s -> 131s.
🌼 Try the demo: Quazim0t0/DaisyChain-Web
📦 Full project: DaisyChainAI/DaisyChain-Train
_____
⚡Also:
🌌 Wheeler–DeWitt‑62M - 2B Tokens Pretrain on a 3060 GPU.
📦 Model: Quazim0t0/Wheeler-DeWitt-62M
🧠 Demo Chat: https://huggingface.co/spaces/Quazim0t0/Wheeler-Chat
A 62.9M‑parameter research language model whose per‑layer channel mixer is the Wheeler–DeWitt equation of canonical quantum gravity, with a fractal (Cantor‑set) RoPE frequency spectrum.
- Elo / Bradley-Terry key rating - keys accumulate a persistent "reputation" score that biases future attention logits, carried through the KV cache.
- Channel mixer (the headline): WheelerDeWittBlock = replaces the MLP with a leapfrog integration of the Wheeler-DeWitt wave equation over 64 minisuperspace modes under a Lorentzian DeWitt supermetric (4 wave steps, learnable lapse), with a Hamiltonian-constraint aux loss ⟨H²⟩ pushing each layer toward HΨ=0.
- Positional encoding: Fractal RoPE - RoPE frequencies placed on a Cantor set (γ=1.0) instead of the geometric ladder; baked in from scratch.
Since launch, the demo has grown from a proof-of-concept into something much more real:
- Block-scaled INT8 quantization
- Batched attention GEMM
- Fused dequant+ReLU epilogue
- Weight-tied unembedding
- WebSocket relay fallback
- Server keepalive ping/pong every 30s
- disconnected-state redial
- Visibility/network-change reconnect (Phones that lock the screen or hop wifi↔cellular reconnect on resume.)
- DAISY_RTC_CONFIG - operators can supply their own TURN/ICE config via env var without touching client code.
- Split-K f32 backward
- Gather-fused attention
Net effect of this push: compute step 821ms -> 420ms (1.95×); full 2-device run 177s -> 131s.
🌼 Try the demo: Quazim0t0/DaisyChain-Web
📦 Full project: DaisyChainAI/DaisyChain-Train
_____
⚡Also:
🌌 Wheeler–DeWitt‑62M - 2B Tokens Pretrain on a 3060 GPU.
📦 Model: Quazim0t0/Wheeler-DeWitt-62M
🧠 Demo Chat: https://huggingface.co/spaces/Quazim0t0/Wheeler-Chat
A 62.9M‑parameter research language model whose per‑layer channel mixer is the Wheeler–DeWitt equation of canonical quantum gravity, with a fractal (Cantor‑set) RoPE frequency spectrum.
- Elo / Bradley-Terry key rating - keys accumulate a persistent "reputation" score that biases future attention logits, carried through the KV cache.
- Channel mixer (the headline): WheelerDeWittBlock = replaces the MLP with a leapfrog integration of the Wheeler-DeWitt wave equation over 64 minisuperspace modes under a Lorentzian DeWitt supermetric (4 wave steps, learnable lapse), with a Hamiltonian-constraint aux loss ⟨H²⟩ pushing each layer toward HΨ=0.
- Positional encoding: Fractal RoPE - RoPE frequencies placed on a Cantor set (γ=1.0) instead of the geometric ladder; baked in from scratch.