ewc_moe_atari — GB10/DGX Spark Fork

Fork of RespectableGlioma/ewc_moe_atari

This fork adapts the out-of-core Mixture of Experts continual learning architecture for the GB10/DGX Spark unified memory platform.

---

What This Fork Adds

The original codebase assumes a discrete GPU memory model — CPU RAM and GPU VRAM as separate pools. On GB10 (Grace Blackwell Superchip), CPU and GPU share a single physical LPDDR5X pool. This fork adapts the expert lifecycle management to that architecture.

Changes to existing code:

• core/tiered_store.py — page cache release on expert eviction (posix_fadvise)
• core/expert_manager.py — memory check before expert load (/proc/meminfo)

New module:

• gb10/ — hardware signal readers and QuadSwarm multi-objective optimizer

See CHANGELOG.md for the full technical detail.

---

GB10 Module

gb10/
├── memory.py        — /proc/meminfo EMV + PSI memory pressure
├── thermal.py       — ACPI thermal zones + CX7 NIC temperature
├── fitness.py       — four-signal expert load cost function
└── swarm_router.py  — QuadSwarm multi-objective optimizer (Optuna v4.9)

See gb10/README.md for setup, usage, and querying results.

---

Original Project

The original ewc_moe_atari implements continual learning across multiple Atari games using a hierarchical out-of-core Mixture of Experts architecture. A small meta-agent routes observations to specialized expert networks stored on disk, allowing scaling to many games without proportional memory growth.

See Claude.md for the full architecture description.

---

Issues and Data Sharing

https://github.com/parallelArchitect/ewc_moe_atari/issues