Weight Synchronization

Flux synchronizes weights between training (Megatron) and inference (SGLang) efficiently.

Sync Methods

Method	Latency	Use Case
CUDA IPC	~10ms	Same node
NCCL	~100ms	Cross-node
HTTP	~1s	Fallback

Delta Compression

Only transfer changed weights:

weight_sync:
  method: delta
  compression: true
  snapshot_interval: 10

Typical compression: 60-80% bandwidth reduction.