Overlap Scheduler — TensorRT LLM (original) (raw)

To maximize GPU utilization, the scheduler overlaps CPU tasks (e.g., checking sampling stop criteria, updating responses, scheduling the next batch) with GPU computation.

How It Works#

At step n, the system launches GPU computation for step n+1 without waiting for CPU tasks (e.g., stop criteria checks) from step n to complete. This allows:

Tradeoff#

The optimization introduces one extra decoding step but significantly improves throughput.

Usage#

Enabled by default. To disable, set disable_overlap_scheduler=True in the configuration.

References#