VLA Training Throughput (pi_0.5) on LIBERO
1,120.91Training Throughput (pi_0.5)Streamer
Evaluation results
| Method | Links | ||
|---|---|---|---|
| StreamerNumber of GPUs=322026.03 | 1,120.91 | 59.25 | |
| Rollout AsyncNumber of GPUs=322026.03 | 1,041.36 | — | |
| Train AsyncNumber of GPUs=322026.03 | 737.46 | — | |
| StreamerNumber of GPUs=162026.03 | 713.38 | 30.29 | |
| ColocatedNumber of GPUs=322026.03 | 703.85 | — | |
| Rollout AsyncNumber of GPUs=162026.03 | 686.8 | — | |
| ColocatedNumber of GPUs=162026.03 | 547.55 | — | |
| Disaggregated (1:1)Number of GPUs=322026.03 | 457.23 | — | |
| Train AsyncNumber of GPUs=162026.03 | 441.49 | — | |
| StreamerNumber of GPUs=82026.03 | 383.4 | 32.5 | |
| Rollout AsyncNumber of GPUs=82026.03 | 369.56 | — | |
| Disaggregated (1:1)Number of GPUs=162026.03 | 307.84 | — | |
| ColocatedNumber of GPUs=82026.03 | 289.23 | — | |
| Train AsyncNumber of GPUs=82026.03 | 229.68 | — | |
| Disaggregated (1:1)Number of GPUs=82026.03 | 162.75 | — |