Module gpu 

GPU resource management and multi-node training infrastructure.

Implements the GPU sharing spec across three phases:

Phase 1: VRAM Guard + Sequential Queue

• Ledger: flock-based VRAM reservation with lease expiry (GPU-SHARE-001)
• Guard: Pre-allocation VRAM check + post-init tracking (GPU-SHARE-002)
• Wait: Polling wait queue with timeout (GPU-SHARE-003)
• Profiler: Brick-phase profiler for GPU sharing ops (GPU-SHARE-005)
• MPS: Experimental CUDA MPS support (opt-in, §1.5)

Phase 2: Multi-Adapter Single-Process

See crate::finetune::multi_adapter_pipeline.

Phase 3: Multi-Node via Forjar

• Cluster: Cluster YAML config schema + validation (§3.2)
• Placement: Greedy job placement with FLOPS scoring (§3.3)
• Coordinator: Checkpoint polling + leaderboard (§3.4)

Modules

cluster

Cluster configuration for multi-node GPU training (GPU-SHARE Phase 3, §3.2).

coordinator

Checkpoint coordination for multi-node adapter training (GPU-SHARE Phase 3, §3.4).

error

GPU sharing error types (GPU-SHARE-001/002/003).

guard

VRAM Guard (GPU-SHARE-002).

ledger

VRAM Reservation Ledger (GPU-SHARE-001).

mps

Experimental CUDA MPS (Multi-Process Service) support (GPU-SHARE §1.5).

placement

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

profiler

Brick-phase profiler for GPU sharing operations (GPU-SHARE-005).

wait

Wait-for-VRAM polling queue (GPU-SHARE-003).