{
  "id": "phi3-mini",
  "name": "Phi-3 Mini Block",
  "description": "Phi-3 Mini 3.8B decoder block — RMSNorm, full MHA (32 heads, 3072D), SwiGLU FFN (8192), RoPE",
  "components": [
    {
      "id": "input_tokens",
      "type": "input",
      "name": "tokens",
      "params": { "shape": [1, 2048] },
      "position": { "x": 400, "y": 40 }
    },
    {
      "id": "embed",
      "type": "embedding",
      "name": "embed",
      "scope": "embeddings",
      "params": { "numEmbeddings": 32064, "embeddingDim": 3072 },
      "position": { "x": 400, "y": 140 },
      "notes": "Phi-3 Mini: vocab_size=32064, hidden_size=3072"
    },
    {
      "id": "rope",
      "type": "rope",
      "name": "rope",
      "scope": "layer.0.attention",
      "params": { "headDim": 96 },
      "position": { "x": 600, "y": 240 },
      "notes": "RoPE positional encoding — head_dim=96 (3072/32). Phi-3 uses rope_theta=10000"
    },
    {
      "id": "norm_attn",
      "type": "rmsNorm",
      "name": "norm_attn",
      "scope": "layer.0.attention",
      "params": { "normalizedShape": 3072 },
      "position": { "x": 400, "y": 240 },
      "notes": "Pre-attention RMSNorm. Phi-3 uses RMSNorm throughout (no bias)"
    },
    {
      "id": "attn",
      "type": "groupedQueryAttention",
      "name": "attn",
      "scope": "layer.0.attention",
      "params": { "embedDim": 3072, "numHeads": 32, "numKVHeads": 32 },
      "position": { "x": 400, "y": 340 },
      "notes": "Phi-3 Mini uses full MHA (32H/32KV = no GQA). causal attention with RoPE."
    },
    {
      "id": "residual1",
      "type": "add",
      "name": "residual_1",
      "scope": "layer.0.attention",
      "params": {},
      "position": { "x": 400, "y": 460 },
      "notes": "Post-attention residual stream"
    },
    {
      "id": "norm_ffn",
      "type": "rmsNorm",
      "name": "norm_ffn",
      "scope": "layer.0.ffn",
      "params": { "normalizedShape": 3072 },
      "position": { "x": 400, "y": 560 },
      "notes": "Pre-FFN RMSNorm"
    },
    {
      "id": "ffn",
      "type": "swiglu",
      "name": "ffn",
      "scope": "layer.0.ffn",
      "params": { "embedDim": 3072, "intermediateSize": 8192 },
      "position": { "x": 400, "y": 660 },
      "notes": "SwiGLU FFN — intermediate=8192 (~2.67× hidden). Phi-3's key efficiency: smaller model with large FFN ratio."
    },
    {
      "id": "residual2",
      "type": "add",
      "name": "residual_2",
      "scope": "layer.0.ffn",
      "params": {},
      "position": { "x": 400, "y": 780 }
    },
    {
      "id": "norm_out",
      "type": "rmsNorm",
      "name": "norm_out",
      "scope": "head",
      "params": { "normalizedShape": 3072 },
      "position": { "x": 400, "y": 880 },
      "notes": "Final RMSNorm before lm_head"
    },
    {
      "id": "lm_head",
      "type": "linear",
      "name": "lm_head",
      "scope": "head",
      "params": { "outFeatures": 32064 },
      "position": { "x": 400, "y": 980 },
      "notes": "Language model head. Phi-3 Mini: 32 layers × (attn ~9.5M + ffn ~50M) ≈ 3.8B total"
    },
    {
      "id": "output",
      "type": "output",
      "name": "output",
      "params": {},
      "position": { "x": 400, "y": 1060 }
    }
  ],
  "connections": [
    { "id": "c1",  "from": "input_tokens", "to": "embed" },
    { "id": "c2",  "from": "embed",        "to": "norm_attn" },
    { "id": "c3",  "from": "norm_attn",    "to": "attn" },
    { "id": "c4",  "from": "rope",         "to": "attn" },
    { "id": "c5",  "from": "attn",         "to": "residual1" },
    { "id": "c6",  "from": "embed",        "to": "residual1" },
    { "id": "c7",  "from": "residual1",    "to": "norm_ffn" },
    { "id": "c8",  "from": "norm_ffn",     "to": "ffn" },
    { "id": "c9",  "from": "ffn",          "to": "residual2" },
    { "id": "c10", "from": "residual1",    "to": "residual2" },
    { "id": "c11", "from": "residual2",    "to": "norm_out" },
    { "id": "c12", "from": "norm_out",     "to": "lm_head" },
    { "id": "c13", "from": "lm_head",      "to": "output" }
  ]
}