| { |
| "bos_token_id": 100257, |
| "embedding_dropout": 0, |
| "eos_token_id": 100257, |
| "hidden_size": 1536, |
| "init_method": "normal", |
| "initializer_range": 0.02, |
| "layer_norm_epsilon": 1e-05, |
| "m_emb": null, |
| "m_residual": null, |
| "m_width": null, |
| "max_position_embeddings": 4096, |
| "mlp_blocks": [ |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| }, |
| { |
| "activation_function": "swiglu", |
| "add_bias": false, |
| "dropout": 0, |
| "intermediate_size": 256, |
| "mlp_type": "MoE", |
| "normalized_topk": true, |
| "num_experts": 128, |
| "num_experts_per_tok": 8, |
| "shared_expert_gating": false, |
| "shared_intermediate_size": 1024, |
| "use_interleaved_weights": true |
| } |
| ], |
| "model_type": "gpt_base", |
| "normalization_function": "rmsnorm", |
| "num_layers": 40, |
| "pad_token_id": 100256, |
| "position_embedding_type": "nope", |
| "rope_dim": null, |
| "rope_scaling": null, |
| "rope_theta": 10000, |
| "router_aux_loss_coef": 0.001, |
| "sequence_mixer_blocks": [ |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| }, |
| { |
| "A_init_max": 16, |
| "A_init_min": 0, |
| "allow_neg_eigval": false, |
| "attention_multiplier": null, |
| "dt_init_floor": 0.0001, |
| "dt_init_max": 0.1, |
| "dt_init_min": 0.001, |
| "k_head_dim": 256, |
| "kernel_size": 4, |
| "num_k_heads": 3, |
| "num_v_heads": 3, |
| "sequence_mixer_type": "gated_deltanet", |
| "use_gate": true, |
| "v_head_dim": 512 |
| } |
| ], |
| "transformers_version": "4.57.3", |
| "use_cache": true, |
| "vocab_size": 100352 |
| } |
|
|