| { |
| "vocab_size": 206464, |
| "d_model": 256, |
| "n_layers": 4, |
| "max_seq_len": 512, |
| "mha_layer_indices": [ |
| 1, |
| 3 |
| ], |
| "dense_ffn_layer_indices": [ |
| 0 |
| ], |
| "hse_ffn_layer_indices": [ |
| 2 |
| ], |
| "arg_d_inner": 256, |
| "arg_d_state": 8, |
| "arg_d_rank": 8, |
| "arg_conv_kernel": 4, |
| "arg_local_window": 32, |
| "arg_local_n_heads": 4, |
| "arg_local_n_kv_heads": 1, |
| "arg_head_dim": 32, |
| "arg_scalar_gate": true, |
| "mha_n_heads": 4, |
| "mha_n_kv_heads": 1, |
| "mha_head_dim": 64, |
| "rope_theta": 500000.0, |
| "rope_scaling_type": null, |
| "rope_scaling_factor": 1.0, |
| "dense_d_ff": 512, |
| "hse_n_domains": 2, |
| "hse_n_experts_per_domain": 2, |
| "hse_top_k": 1, |
| "hse_d_ff_expert": 128, |
| "hse_capacity_factor": 1.25, |
| "hse_expert_dropout": 0.1, |
| "hse_aux_loss_alpha": 0.01, |
| "hse_gumbel_tau_init": 1.0, |
| "hse_gumbel_tau_final": 0.1, |
| "hse_gumbel_tau_warmup_tokens": 20000000000, |
| "norm_type": "dgn", |
| "dgn_n_groups": 4, |
| "norm_eps": 1e-06, |
| "init_std": 0.02, |
| "init_std_residual_scale": true, |
| "dropout": 0.0, |
| "tie_word_embeddings": false, |
| "use_cache": true, |
| "pad_token_id": 0, |
| "bos_token_id": 1, |
| "eos_token_id": 2, |
| "use_flash_attention": false, |
| "use_triton_kernels": false, |
| "use_torch_compile": false, |
| "compile_mode": "max-autotune", |
| "use_gradient_checkpointing": true, |
| "gradient_checkpointing_ratio": 0.5 |
| } |