{ "architecture": { "attn_pdrop": 0.0, "bos_token_id": 128000, "compute_dtype": "bf16", "embd_pdrop": 0.0, "eos_token_id": 128001, "feed_forward_prime": "swiglu", "force_flash": false, "hidden_size": 2048, "initializer_range": 0.02, "intermediate_size": 4224, "mini_batch_size": 1024, "num_attention_heads": 32, "num_hidden_layers": 24, "output_size": 128256, "param_dtype": "fp32", "post_norm": true, "pre_norm": true, "prime": true, "prime_start": 18, "qk_norm": true, "resid_pdrop": 0.0, "rms_norm_eps": 1e-06, "rope_theta": 500000.0, "seq_modeling_block": "SWA", "sliding_window_size": 8192, "state_dtype": "fp32", "suffix_len": 6, "swa_attention_backend": "auto", "tie_word_embeddings": true, "vocab_size": 128256 }, "format": "e2e-routed-router-v7", "router": { "adapter_rank": 128, "adapter_silu": true, "cadence": 4, "dead_branch_skip": false, "exit_ce_sites_per_chunk": 0, "exit_ce_vocab_chunk_size": 4096, "exit_norm": true, "input_norm_eps": 0.0078125, "kv_policy": "calculate", "masked_exit_cost": 0.0, "mode": "exit_only", "sites": [ { "end": 8, "schema": "binary", "start": 4 }, { "end": 12, "schema": "binary", "start": 8 }, { "end": 16, "schema": "binary", "start": 12 }, { "end": 20, "schema": "binary", "start": 16 }, { "end": 24, "schema": "binary", "start": 20 } ], "start_layer": 4, "temperature": 1.0, "ttt_only": false, "update_bias": 1.0 }, "training": { "accum_steps": 32, "data_seed": 42, "data_split": "train", "dataset_name": "books3", "dummy_dataset": false, "dummy_num_tokens": 1048576, "efficiency_coefficient": 0.5, "exit_ce_coefficient": 0.5, "gate_teacher_counterfactual": true, "global_batch_size": 32, "inner_learning_rate": 1.0, "inner_warmup_steps": 0, "mask_bos_targets": false, "mask_final_target": true, "model_seed": 42, "optimizer_inner": { "b1": null, "b2": null, "bf16_momentum": null, "clip_gradient": 1.0, "emb_wd": null, "end_lr": null, "init_lr": null, "lr": 1.0, "lr_decay_steps": null, "lr_warmup_steps": null, "optimizer_type": "sgd", "weight_decay": null }, "optimizer_outer": { "b1": 0.9, "b2": 0.95, "bf16_momentum": false, "clip_gradient": 1.0, "emb_wd": true, "end_lr": 0.0001, "init_lr": 0.0, "lr": 0.001, "lr_decay_steps": 1024, "lr_warmup_steps": 256, "optimizer_type": "adamw", "weight_decay": 0.1 }, "pure_min_realized": false, "router_warmup_steps": 256, "seq_length": 8192, "spec_inner": [ "language_model.**.suffix_blocks.feed_forward_prime.**" ], "spec_outer": [ "language_model.**.router_storage.**" ], "target_cost": null, "target_cost_realized": true, "total_steps": 1024, "train_mode": "meta", "z_coefficient": 0.0 } }