Layer window & partial shards

A node's contiguous slice of the model — downloadable as a mini-GGUF instead of the full checkpoint.

A layer window is the contiguous range of transformer layers a ring node serves. A node doesn't need the full checkpoint to serve one — a stage mini-GGUF carries only the window's tensors: for the 122B, 254 MB of 26 tensors (of 338 total) versus the 77.6 GB full model, or ~1.5 GB for a one-layer window on a phone.

rank 3  layers [39,48]  vram=3.4GiB  kv=0.9GiB  backend=opencl
shard: mini-GGUF with exactly those blk.39-48 tensors → download → load

• A node polls the coverage/demand map to see which windows are under-served and what each pays.
• It picks the highest-reward uncovered window that fits its budget, downloads exactly that, and joins.
• Coverage self-heals: when a node churns away, its window becomes scarce — and therefore lucrative — again.
• Verified end-to-end with a NAT-bound phone: poll → self-enroll → partial download → Adreno GPU load → ring inference completed, contribution credited.

The expert swarm reuses this exact market at the finer (layer, expert-range) grain — same map, same self-enrollment, same rewards, smaller units.