@affectively/aether
v5.0.0
Published
Aether — distributed inference runtime
Maintainers
Readme
Aether
Distributed inference runtime for the AFFECTIVELY platform.
Aether is the medium through which hidden states flow between layer nodes across the network. It orchestrates transformer layer computation across Cloud Run nodes, Cloudflare Workers, and local processes.
Architecture
- Coordinator — orchestrates layer nodes, manages KV cache, tokenization, and sampling
- Layer Node — processes a subset of transformer layers with weight caching and dequantization
- gRPC Client — HTTP/2 binary transport between coordinator and layer nodes
- Local Client — in-process computation eliminating network round-trips
- Generic Pool — dynamic model loading with LRU weight eviction
Log Namespaces
All Aether components use structured log prefixes:
| Prefix | Component |
|--------|-----------|
| [Aether:Coordinator] | Main coordinator |
| [Aether:Node] | Layer node |
| [Aether:Layer N] | Per-layer debug |
| [Aether:gRPC] | Inter-node client |
| [Aether:Generic] | Generic warm pool |
| [Aether:Local] | In-process node |
Logs route through DashRelay to the aether-logs room for real-time monitoring in Forge.
License
Copyright (c) 2026 Taylor William Buley (https://buley.fyi). Personal, non-commercial, and academic use only. See LICENSE.
