mergekit/_data/architectures/mistral.json

{
    "model_type": "mistral",
    "architectures": [
        "MistralForCausalLM"
    ],
    "pre_weights": [
        {
            "name": "model.embed_tokens.weight",
            "is_embed": true,
            "output_space": "h_0"
        }
    ],
    "num_layers_config_key": "num_hidden_layers",
    "layer_templates": {
        "weights": [
            {
                "name": "model.layers.${layer_index}.input_layernorm.weight",
                "input_space": "h_${layer_index}"
            },
            {
                "name": "model.layers.${layer_index}.self_attn.q_proj.weight",
                "input_space": "h_${layer_index}",
                "output_space": "attn_qk_${layer_index}"
            },
            {
                "name": "model.layers.${layer_index}.self_attn.k_proj.weight",
                "input_space": "h_${layer_index}",
                "output_space": "attn_qk_${layer_index}"
            },
            {
                "name": "model.layers.${layer_index}.self_attn.v_proj.weight",
                "input_space": "h_${layer_index}",
                "output_space": "attn_v_${layer_index}"
            },
            {
                "name": "model.layers.${layer_index}.self_attn.o_proj.weight",
                "input_space": "attn_v_${layer_index}",
                "output_space": "post_attn_${layer_index}"
            },
            {
                "name": "model.layers.${layer_index}.post_attention_layernorm.weight",
                "input_space": "h_a_${layer_index}"
            },
            {
                "name": "model.layers.${layer_index}.mlp.up_proj.weight",
                "input_space": "h_a_${layer_index}",
                "output_space": "up_${layer_index}"
            },
            {
                "name": "model.layers.${layer_index}.mlp.gate_proj.weight",
                "input_space": "h_a_${layer_index}",
                "output_space": "up_${layer_index}"
            },
            {
                "name": "model.layers.${layer_index}.mlp.down_proj.weight",
                "input_space": "up_${layer_index}",
                "output_space": "post_mlp_${layer_index}"
            }
        ],
        "procedural_spaces": [
            {
                "name": "h_a_${layer_index}",
                "type": "residual",
                "inputs": [
                    "h_${layer_index}",
                    "post_attn_${layer_index}"
                ]
            },
            {
                "name": "h_${layer_index+1}",
                "type": "residual",
                "inputs": [
                    "h_a_${layer_index}",
                    "post_mlp_${layer_index}"
                ]
            }
        ]
    },
    "post_weights": [
        {
            "name": "model.norm.weight",
            "input_space": "h_${num_layers}"
        },
        {
            "name": "lm_head.weight",
            "input_space": "h_${num_layers}",
            "is_embed": true
        }
    ]
}