Performance

Use these guides when latency, throughput, memory, placement, or large-model training efficiency becomes the bottleneck.

Guide	What you get
RTC	Hide policy inference latency by overlapping it with action-chunk execution, in simulation and on real robots.
Env Decoupled Mode	Decouple Env Workers from Rollout Workers for dynamic embodied rollout scheduling.
Observation Compression	Losslessly compress image observations on the Env to Rollout channel to save bandwidth.
LoRA	Train with LoRA adapters.
Auto Placement	Auto-select the best placement for a workload.
Dynamic Scheduling	Dynamically schedule resources during training.
Profiling	System-level profiling of Ray worker processes.
5D Parallelism	Configure 5D parallelism for large models.