{
  "architecture": {
    "attention_variant": "MHA",
    "family": "causal_lm",
    "is_moe": false,
    "mixer": "attention",
    "positional": "rope",
    "tie_word_embeddings": true,
    "view": "decoder"
  },
  "capabilities": {
    "attention_backends": [
      "eager",
      "sdpa",
      "flash_attention",
      "flex_attention"
    ],
    "attention_patterns": [
      "causal"
    ],
    "attention_schedule": null,
    "task_heads": [
      "causal_lm"
    ],
    "tensor_parallel": true
  },
  "components": [
    {
      "children": [
        "embed_tokens",
        "decoder_layers",
        "norm",
        "rotary_emb"
      ],
      "class_name": "CohereModel",
      "id": "model",
      "kind": "model",
      "path_pattern": "model"
    },
    {
      "attributes": {
        "embedding_dim": "config.hidden_size",
        "num_embeddings": "config.vocab_size",
        "tied_lm_head": true
      },
      "class_name": "Embedding",
      "id": "embed_tokens",
      "kind": "embedding",
      "path_pattern": "model.embed_tokens"
    },
    {
      "attributes": {
        "norm_type": "layer"
      },
      "class_name": "CohereLayerNorm",
      "id": "norm",
      "kind": "normalization",
      "path_pattern": "model.norm"
    },
    {
      "attributes": {
        "rope_theta": 500000.0,
        "scheme": "rope"
      },
      "class_name": "CohereRotaryEmbedding",
      "id": "rotary_emb",
      "kind": "position",
      "path_pattern": "model.rotary_emb"
    }
  ],
  "config": {
    "class_name": "CohereConfig",
    "model_type": "cohere",
    "module": "transformers.models.cohere.configuration_cohere",
    "referenced_fields": {
      "num_hidden_layers": 40
    },
    "salient_fields": {
      "hidden_act": "silu",
      "hidden_size": 8192,
      "intermediate_size": 22528,
      "is_encoder_decoder": false,
      "max_position_embeddings": 8192,
      "num_attention_heads": 64,
      "num_hidden_layers": 40,
      "num_key_value_heads": 64,
      "tie_word_embeddings": true,
      "vocab_size": 256000
    }
  },
  "dataflow": {
    "input": {
      "name": "input_ids",
      "shape": [
        "B",
        "S"
      ]
    },
    "output": {
      "shape": [
        "B",
        "S",
        "config.hidden_size"
      ]
    },
    "shapes": {
      "decoder_layer.input_layernorm": {
        "in": [
          "B",
          "S",
          "config.hidden_size"
        ],
        "out": [
          "B",
          "S",
          "config.hidden_size"
        ]
      },
      "decoder_layer.mlp": {
        "in": [
          "B",
          "S",
          "config.hidden_size"
        ],
        "out": [
          "B",
          "S",
          "config.hidden_size"
        ]
      },
      "decoder_layer.self_attn": {
        "in": null,
        "out": [
          "B",
          "S",
          "config.hidden_size"
        ]
      },
      "decoder_layers": {
        "in": [
          "B",
          "S",
          "config.hidden_size"
        ],
        "out": [
          "B",
          "S",
          "config.hidden_size"
        ]
      },
      "embed_tokens": {
        "in": [
          "B",
          "S"
        ],
        "out": [
          "B",
          "S",
          "config.hidden_size"
        ]
      },
      "norm": {
        "in": [
          "B",
          "S",
          "config.hidden_size"
        ],
        "out": [
          "B",
          "S",
          "config.hidden_size"
        ]
      },
      "rotary_emb": {
        "in": [
          "B",
          "S",
          "config.hidden_size"
        ],
        "out": [
          "B",
          "S",
          128
        ]
      }
    },
    "source": "observed_forward_meta"
  },
  "edges": [
    {
      "kind": "data",
      "source": "embed_tokens",
      "target": "decoder_layers"
    },
    {
      "kind": "data",
      "source": "decoder_layers",
      "target": "norm"
    },
    {
      "kind": "mask",
      "source": "input:attention_mask",
      "target": "decoder_layer.self_attn"
    },
    {
      "kind": "residual",
      "source": "decoder_layer",
      "target": "decoder_layer.self_attn"
    },
    {
      "kind": "residual",
      "source": "decoder_layer",
      "target": "decoder_layer.mlp"
    },
    {
      "kind": "position",
      "source": "rotary_emb",
      "target": "decoder_layer.self_attn"
    },
    {
      "kind": "data",
      "provenance": "intra_module",
      "source": "decoder_layer.self_attn",
      "target": "decoder_layer.self_attn.q_proj"
    },
    {
      "kind": "data",
      "provenance": "intra_module",
      "source": "decoder_layer.self_attn",
      "target": "decoder_layer.self_attn.k_proj"
    },
    {
      "kind": "data",
      "provenance": "intra_module",
      "source": "decoder_layer.self_attn",
      "target": "decoder_layer.self_attn.v_proj"
    },
    {
      "kind": "data",
      "provenance": "intra_module",
      "source": "decoder_layer.self_attn.q_proj",
      "target": "decoder_layer.self_attn.o_proj"
    },
    {
      "kind": "data",
      "provenance": "intra_module",
      "source": "decoder_layer.self_attn.k_proj",
      "target": "decoder_layer.self_attn.o_proj"
    },
    {
      "kind": "data",
      "provenance": "intra_module",
      "source": "decoder_layer.self_attn.v_proj",
      "target": "decoder_layer.self_attn.o_proj"
    },
    {
      "kind": "data",
      "provenance": "intra_module",
      "source": "decoder_layer.mlp",
      "target": "decoder_layer.mlp.gate_proj"
    },
    {
      "kind": "data",
      "provenance": "intra_module",
      "source": "decoder_layer.mlp",
      "target": "decoder_layer.mlp.up_proj"
    },
    {
      "kind": "data",
      "provenance": "intra_module",
      "source": "decoder_layer.mlp.gate_proj",
      "target": "decoder_layer.mlp.down_proj"
    },
    {
      "kind": "data",
      "provenance": "intra_module",
      "source": "decoder_layer.mlp.up_proj",
      "target": "decoder_layer.mlp.down_proj"
    },
    {
      "kind": "cache_read",
      "source": "state:kv_cache",
      "target": "decoder_layer.self_attn"
    },
    {
      "kind": "cache_write",
      "source": "decoder_layer.self_attn",
      "target": "state:kv_cache"
    },
    {
      "kind": "data",
      "provenance": "observed_forward",
      "source": "decoder_layer.input_layernorm",
      "target": "decoder_layer.self_attn"
    },
    {
      "kind": "data",
      "provenance": "observed_forward",
      "source": "decoder_layer.self_attn",
      "target": "decoder_layer.mlp"
    }
  ],
  "extends": "llama",
  "model_type": "cohere",
  "patches": [
    {
      "added": {
        "methods": [
          "__init__",
          "forward"
        ]
      },
      "component_kind": "normalization",
      "parent_class": "Module",
      "relation": "new",
      "target_class": "CohereLayerNorm"
    },
    {
      "component_kind": "position",
      "overridden": {
        "methods": [
          "forward"
        ]
      },
      "parent_class": "LlamaRotaryEmbedding",
      "relation": "inherits",
      "target_class": "CohereRotaryEmbedding"
    },
    {
      "component_kind": "feed_forward",
      "overridden": {
        "methods": [
          "__init__"
        ]
      },
      "parent_class": "LlamaMLP",
      "relation": "inherits",
      "target_class": "CohereMLP"
    },
    {
      "component_kind": "attention",
      "overridden": {
        "methods": [
          "__init__",
          "forward"
        ]
      },
      "parent_class": "LlamaAttention",
      "relation": "inherits",
      "target_class": "CohereAttention"
    },
    {
      "added": {
        "methods": [
          "__init__",
          "forward"
        ]
      },
      "component_kind": "transformer_block",
      "parent_class": "GradientCheckpointingLayer",
      "relation": "new",
      "target_class": "CohereDecoderLayer"
    },
    {
      "component_kind": "model",
      "overridden": {
        "methods": [
          "__init__"
        ]
      },
      "parent_class": "LlamaModel",
      "relation": "inherits",
      "target_class": "CohereModel"
    },
    {
      "component_kind": "lm_head",
      "overridden": {
        "methods": [
          "__init__",
          "forward"
        ]
      },
      "parent_class": "LlamaForCausalLM",
      "relation": "inherits",
      "target_class": "CohereForCausalLM"
    }
  ],
  "provenance": {
    "config_class": "CohereConfig",
    "config_module": "transformers.models.cohere.configuration_cohere",
    "model_class": "CohereModel",
    "model_module": "transformers.models.cohere.modeling_cohere"
  },
  "repeats": [
    {
      "body": "decoder_layer",
      "container_path_pattern": "model.layers",
      "count": 40,
      "count_expr": "config.num_hidden_layers",
      "count_source": "config",
      "id": "decoder_layers",
      "index_symbol": "i",
      "item_path_pattern": "model.layers.{i}",
      "kind": "symbolic_repeat",
      "provenance": {
        "class_name": "CohereDecoderLayer",
        "container_path_pattern": "model.layers",
        "item_path_pattern": "model.layers.{i}",
        "source": "module_tree_repeat_collapse"
      },
      "repeated_class_name": "CohereDecoderLayer"
    }
  ],
  "schema_version": "architecture-template-v0",
  "templates": [
    {
      "children": [
        "decoder_layer.self_attn",
        "decoder_layer.mlp",
        "decoder_layer.input_layernorm"
      ],
      "class_name": "CohereDecoderLayer",
      "id": "decoder_layer",
      "kind": "transformer_block",
      "path_pattern": "model.layers.{i}"
    },
    {
      "attributes": {
        "head_dim": 128,
        "n_heads": 64,
        "n_kv_heads": 64,
        "pattern": "causal",
        "rope": true,
        "variant": "MHA"
      },
      "children": [
        "decoder_layer.self_attn.q_proj",
        "decoder_layer.self_attn.k_proj",
        "decoder_layer.self_attn.v_proj",
        "decoder_layer.self_attn.o_proj"
      ],
      "class_name": "CohereAttention",
      "id": "decoder_layer.self_attn",
      "kind": "attention",
      "path_pattern": "model.layers.{i}.self_attn"
    },
    {
      "attributes": {
        "in_features": "config.hidden_size",
        "out_features": "config.hidden_size",
        "tp": "colwise"
      },
      "class_name": "Linear",
      "id": "decoder_layer.self_attn.q_proj",
      "kind": "projection",
      "path_pattern": "model.layers.{i}.self_attn.q_proj"
    },
    {
      "attributes": {
        "in_features": "config.hidden_size",
        "out_features": "config.hidden_size",
        "tp": "colwise"
      },
      "class_name": "Linear",
      "id": "decoder_layer.self_attn.k_proj",
      "kind": "projection",
      "path_pattern": "model.layers.{i}.self_attn.k_proj"
    },
    {
      "attributes": {
        "in_features": "config.hidden_size",
        "out_features": "config.hidden_size",
        "tp": "colwise"
      },
      "class_name": "Linear",
      "id": "decoder_layer.self_attn.v_proj",
      "kind": "projection",
      "path_pattern": "model.layers.{i}.self_attn.v_proj"
    },
    {
      "attributes": {
        "in_features": "config.hidden_size",
        "out_features": "config.hidden_size",
        "tp": "rowwise"
      },
      "class_name": "Linear",
      "id": "decoder_layer.self_attn.o_proj",
      "kind": "projection",
      "path_pattern": "model.layers.{i}.self_attn.o_proj"
    },
    {
      "attributes": {
        "activation": "silu",
        "hidden_size": 8192,
        "intermediate_size": 22528
      },
      "children": [
        "decoder_layer.mlp.gate_proj",
        "decoder_layer.mlp.up_proj",
        "decoder_layer.mlp.down_proj"
      ],
      "class_name": "CohereMLP",
      "id": "decoder_layer.mlp",
      "kind": "feed_forward",
      "path_pattern": "model.layers.{i}.mlp"
    },
    {
      "attributes": {
        "in_features": "config.hidden_size",
        "out_features": "config.intermediate_size",
        "tp": "colwise"
      },
      "class_name": "Linear",
      "id": "decoder_layer.mlp.gate_proj",
      "kind": "projection",
      "path_pattern": "model.layers.{i}.mlp.gate_proj"
    },
    {
      "attributes": {
        "in_features": "config.hidden_size",
        "out_features": "config.intermediate_size",
        "tp": "colwise"
      },
      "class_name": "Linear",
      "id": "decoder_layer.mlp.up_proj",
      "kind": "projection",
      "path_pattern": "model.layers.{i}.mlp.up_proj"
    },
    {
      "attributes": {
        "in_features": "config.intermediate_size",
        "out_features": "config.hidden_size",
        "tp": "rowwise"
      },
      "class_name": "Linear",
      "id": "decoder_layer.mlp.down_proj",
      "kind": "projection",
      "path_pattern": "model.layers.{i}.mlp.down_proj"
    },
    {
      "attributes": {
        "norm_type": "layer"
      },
      "class_name": "CohereLayerNorm",
      "id": "decoder_layer.input_layernorm",
      "kind": "normalization",
      "path_pattern": "model.layers.{i}.input_layernorm"
    }
  ]
}