defaults: ../../grpo_math_1B.yaml checkpointing: checkpoint_dir: results/grpo-qwen3.5-35ba3b-2n8g-megatron-ep16tp2cp2 policy: model_name: Qwen/Qwen3.5-35B-A3B-Base train_micro_batch_size: 1 logprob_batch_size: 1 max_total_sequence_length: 4096 make_sequence_length_divisible_by: 8 automodel_cfg: enabled: false megatron_cfg: enabled: true tensor_model_parallel_size: 2 context_parallel_size: 2 sequence_parallel: true expert_model_parallel_size: 16 apply_rope_fusion: false activation_checkpointing: true defer_fp32_logits: true generation: vllm_cfg: tensor_parallel_size: 4 # CUDA graphs instead of enforce_eager (~5x faster generation), see #4569 logger: wandb_enabled: true tensorboard_enabled: true wandb: project: nemo-rl name: grpo-qwen3.5-35ba3b-2n8g-megatron-ep16tp2cp2 cluster: gpus_per_node: 8 num_nodes: 2