{
  "gpt2a": {
    "tag": "gpt2a",
    "repo": "onnx-community/gpt2-alpaca-gpt4-ONNX",
    "kind": "text-instruction",
    "notas": "GPT-2 124M afinado em Alpaca-GPT4 (epoca <= 2023/24)",
    "matrix": "transformer.wte.weight",
    "shape": [
      50260,
      768
    ],
    "dtype": "float16",
    "K_experts": 32,
    "K_xi": 8,
    "d_src": 768,
    "n_adapters": 4,
    "inertia": 4198.662109375,
    "samples": 6000,
    "int4": true,
    "adapter_rank": 48,
    "epoca": "<= 2024/2025",
    "runtime_mode": "embedding"
  },
  "gpt2g": {
    "tag": "gpt2g",
    "repo": "onnx-community/gpt2-ONNX-GQA",
    "kind": "text-gqa",
    "notas": "GPT-2 com atencao agrupada (GQA) exportada ONNX/OGA",
    "matrix": "transformer.wte.weight",
    "shape": [
      50257,
      768
    ],
    "dtype": "1",
    "K_experts": 32,
    "K_xi": 8,
    "d_src": 768,
    "n_adapters": 4,
    "inertia": 4229.31640625,
    "samples": 6000,
    "int4": true,
    "adapter_rank": 48,
    "epoca": "<= 2024/2025",
    "gqa_signature": {
      "gqa_nodes": 12,
      "ops_top": [
        [
          "Add",
          97
        ],
        [
          "MatMul",
          73
        ],
        [
          "LayerNormalization",
          25
        ],
        [
          "GroupQueryAttention",
          12
        ],
        [
          "Gelu",
          12
        ],
        [
          "Gather",
          3
        ],
        [
          "Constant",
          2
        ],
        [
          "Cast",
          2
        ]
      ],
      "kv_candidate_shapes": []
    },
    "runtime_mode": "embedding"
  },
  "dinov2": {
    "tag": "dinov2",
    "repo": "onnx-community/dinov2-small-with-attentions",
    "kind": "vision",
    "notas": "ViT-S/14 DINOv2 (autoatencao auto-supervisionada, 2023/24)",
    "matrix": "onnx::MatMul_1868",
    "shape": [
      384,
      1536
    ],
    "dtype": "float16",
    "K_experts": 48,
    "K_xi": 8,
    "d_src": 384,
    "n_adapters": 4,
    "inertia": 1218.400146484375,
    "samples": 1536,
    "int4": true,
    "adapter_rank": 48,
    "epoca": "<= 2024/2025",
    "visao": {
      "patch": 14,
      "image_size": 518,
      "uso": "prototipos int4 desquantizados (padrao granite)"
    }
  }
}