VLA Training Throughput (pi_0) on ManiSkill
436.32Training ThroughputTrain Async
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Train AsyncNumber of GPUs=322026.03 | 436.32 | 17.84 | |
| ColocatedNumber of GPUs=322026.03 | 370.26 | — | |
| StreamerNumber of GPUs=322026.03 | 280.07 | — | |
| Rollout AsyncNumber of GPUs=322026.03 | 275.36 | — | |
| Disaggregated (1:1)Number of GPUs=322026.03 | 257.21 | — | |
| Train AsyncNumber of GPUs=162026.03 | 244.61 | 5.33 | |
| ColocatedNumber of GPUs=162026.03 | 232.23 | — | |
| Disaggregated (1:1)Number of GPUs=162026.03 | 150.59 | — | |
| StreamerNumber of GPUs=162026.03 | 141.49 | — | |
| Rollout AsyncNumber of GPUs=162026.03 | 139.32 | — | |
| ColocatedNumber of GPUs=82026.03 | 132.56 | — | |
| Train AsyncNumber of GPUs=82026.03 | 126.65 | -4.46 | |
| StreamerNumber of GPUs=82026.03 | 71.61 | — | |
| Rollout AsyncNumber of GPUs=82026.03 | 69.07 | — | |
| Disaggregated (1:1)Number of GPUs=82026.03 | 60.64 | — |