| { |
| "model_type": "afmoe", |
| "architectures": [ |
| "MOE" |
| ], |
| "profile_name": "1b_3e_8l_t4x2", |
| "vocab_size": 200024, |
| "text_embed_dim": 1024, |
| "vision_embed_dim": 1024, |
| "hidden_dim": 1024, |
| "ffn_dim": 6144, |
| "num_layers": 8, |
| "num_heads": 16, |
| "num_kv_heads": 4, |
| "num_experts": 3, |
| "top_k": 2, |
| "max_position_embeddings": 16384, |
| "router_aux_loss_coef": 0.01, |
| "share_experts_across_layers": false, |
| "gradient_checkpointing": true, |
| "num_agents": 4, |
| "moe_capacity_factor": 1.0, |
| "moe_hierarchy_groups": 1, |
| "moe_hierarchy_top_k": 1, |
| "num_shared_experts": 0, |
| "load_balancing_mode": "aux_free", |
| "router_bias_update_rate": 0.01, |
| "kv_latent_dim": 128, |
| "kv_cache_dtype": "int4", |
| "rope_training_context": 16384, |
| "rope_ntk_alpha": 1.0, |
| "rope_yarn_scale": 1.0, |
| "ring_attention_chunk_size": 0, |
| "prefill_chunk_size": 256, |
| "use_q_former_projector": true, |
| "q_former_queries": 8, |
| "q_former_layers": 1, |
| "tokenizer_name": "ai-tokenizer:GPT-5" |
| } |
