checkpoint_ssm_layers: true d_ff: 3072 d_inner: 2048 d_model: 1024 dtype: float16 gradient_checkpointing: true load_balancing_coef: 0.01 max_seq_len: 2048 n_layer: 24 num_experts: 4 router_z_loss_coef: 0.001 ssm_d_state: 16 ssm_expand: 2 top_k: 1 use_cpu_offload: false use_flash_attention: false vocab_size: 261019