Vision
Each Continuum node reports its resource envelope to the grid, enabling intelligent workload placement.
Phases
v0 (today): 80% RAM self-policing per node. Hardcoded 4096MB limit replaced with dynamic detection via sysctl/proc/meminfo.
v1: Each node reports resource envelope (RAM, VRAM, CPU, disk, network) to grid coordinator. Coordinator assigns per-node budgets based on fleet-wide capacity.
v2: Grid coordinator IS the scheduler — places workloads by capacity. 48GB node gets 27B model, 8GB node gets 0.8B. Like ECS task placement but for AI inference.
v3: Containerized forge jobs dispatched to nodes with available GPU. Factory becomes a distributed job scheduler across the mesh.
Context
The 80% RAM guard was added to fix MEMLEAK FATAL killing continuum-core on machines with >4GB RAM (issue #603). This is the local-only v0 — grid coordination is the long-term architecture.
Related: Foreman (per-node executor) and Ares (centralized authority) role split.
Vision
Each Continuum node reports its resource envelope to the grid, enabling intelligent workload placement.
Phases
v0 (today): 80% RAM self-policing per node. Hardcoded 4096MB limit replaced with dynamic detection via sysctl/proc/meminfo.
v1: Each node reports resource envelope (RAM, VRAM, CPU, disk, network) to grid coordinator. Coordinator assigns per-node budgets based on fleet-wide capacity.
v2: Grid coordinator IS the scheduler — places workloads by capacity. 48GB node gets 27B model, 8GB node gets 0.8B. Like ECS task placement but for AI inference.
v3: Containerized forge jobs dispatched to nodes with available GPU. Factory becomes a distributed job scheduler across the mesh.
Context
The 80% RAM guard was added to fix MEMLEAK FATAL killing continuum-core on machines with >4GB RAM (issue #603). This is the local-only v0 — grid coordination is the long-term architecture.
Related: Foreman (per-node executor) and Ares (centralized authority) role split.