diff --git a/configs/training/Qwen1.5B/Qwen1.5b.toml b/configs/training/Qwen1.5B/Qwen1.5b.toml index 24ff7102d23..19d230b0a71 100644 --- a/configs/training/Qwen1.5B/Qwen1.5b.toml +++ b/configs/training/Qwen1.5B/Qwen1.5b.toml @@ -10,7 +10,7 @@ attn_impl = "flash_attention_2" batch_size = 512 warmup_steps = 1 total_steps = 100000000000000 -step_per_rollout = 8 +step_per_rollout = 4 [optim.optim] lr = 1e-6