Module placement 

Job placement algorithm for multi-node adapter training (GPU-SHARE Phase 3, §3.3).

Scores each node for each adapter job and assigns greedily:

score = (free_vram / adapter_budget) × gpu_flops_factor × (1 / current_load)

Where gpu_flops_factor normalizes different GPU types:

• RTX 4090: 1.0 (reference)
• Jetson Orin: 0.06 (8 SMs vs 128)
• CPU (Intel): 0.01

Structs

AdapterJob

A pending adapter job to place on a cluster node.

NodeLoad

Current load state of a node for placement scoring.

PlacementDecision

Result of placing an adapter on a node.

Functions

place_adapters

Place adapter jobs across cluster nodes greedily.

score_node

Score a node for a given adapter budget.