One GRPO step on a GB10, by phase
Median seconds per phase over 12 plain GRPO steps of Miles with Qwen3-0.6B on one NVIDIA GB10, with SGLang and FSDP colocated. SGLang rollout takes 57.2 s, the trainer recomputes log probabilities in 25.2 s and trains in 80.3 s, the weight update takes 0.65 s, and sleep and wake take 1.7 s.
0
25
50
75
100
median seconds per step
57.2 s
rollout
25.1 s
log probs
80.3 s
train
0.65 s
weights
1.70 s
sleep, wake