{
  "modelType": "embedding",
  "output": {
    "baseDir": "models/local",
    "modelBaseId": "google-embeddinggemma-300m-q4k-ehf16-af32"
  },
  "quantization": {
    "weights": "q4k",
    "embeddings": "f16",
    "lmHead": "f16",
    "q4kLayout": "row",
    "computePrecision": "f32"
  },
  "inference": {
    "attention": {
      "queryPreAttnScalar": 256,
      "attnLogitSoftcapping": null,
      "slidingWindow": null,
      "queryKeyNorm": true,
      "valueNorm": false,
      "causal": false,
      "attentionBias": false,
      "attentionOutputGate": false
    },
    "normalization": {
      "rmsNormEps": 0.000001,
      "rmsNormWeightOffset": true,
      "postAttentionNorm": true,
      "preFeedforwardNorm": true,
      "postFeedforwardNorm": true
    },
    "ffn": {
      "activation": "gelu",
      "gatedActivation": true,
      "branchMode": "auto",
      "useDoubleWideMlp": false,
      "swigluLimit": null
    },
    "rope": {
      "ropeTheta": 1000000,
      "ropeLocalTheta": 10000,
      "mropeInterleaved": false,
      "mropeSection": null,
      "partialRotaryFactor": null,
      "ropeLocalPartialRotaryFactor": null,
      "ropeFrequencyBaseDim": null,
      "ropeLocalFrequencyBaseDim": null,
      "ropeScalingType": null,
      "ropeScalingFactor": 1,
      "ropeLocalScalingType": null,
      "ropeLocalScalingFactor": 1,
      "yarnBetaFast": null,
      "yarnBetaSlow": null,
      "yarnOriginalMaxPos": null,
      "ropeLocalYarnBetaFast": null,
      "ropeLocalYarnBetaSlow": null,
      "ropeLocalYarnOriginalMaxPos": null,
      "longropeShortFactor": null,
      "longropeLongFactor": null,
      "longropeOriginalMaxPos": null
    },
    "output": {
      "finalLogitSoftcapping": null,
      "tieWordEmbeddings": false,
      "scaleEmbeddings": true,
      "embeddingTranspose": false,
      "embeddingVocabSize": null,
      "embeddingPostprocessor": {
        "poolingMode": "mean",
        "includePrompt": true,
        "projections": [
          {
            "weightTensor": "embedding_postprocessor.projections.0.weight",
            "biasTensor": null,
            "inputSize": 768,
            "outputSize": 3072,
            "activation": "identity"
          },
          {
            "weightTensor": "embedding_postprocessor.projections.1.weight",
            "biasTensor": null,
            "inputSize": 3072,
            "outputSize": 768,
            "activation": "identity"
          }
        ],
        "normalize": "l2"
      },
      "embeddingScale": null,
      "logitInputScale": 1
    },
    "layerPattern": {
      "type": "every_n",
      "globalPattern": null,
      "period": 6,
      "offset": null,
      "layerTypes": null,
      "residualBranchScale": 1
    },
    "chatTemplate": {
      "type": null,
      "enabled": false
    }
  },
  "session": {
    "compute": {
      "defaults": {
        "activationDtype": "f32",
        "mathDtype": "f32",
        "accumDtype": "f32",
        "outputDtype": "f32"
      }
    },
    "kvcache": {
      "kvDtype": "f32",
      "layout": "contiguous",
      "pageSize": 256,
      "tiering": {
        "mode": "off",
        "hotWindow": 1024,
        "coldPageSize": 256,
        "coldDtype": "f16",
        "compression": {
          "mode": "none",
          "blockSize": 1,
          "bitWidth": 4,
          "prodMode": false
        },
        "gating": {
          "mode": "auto",
          "minAluBwRatio": 0
        }
      },
      "quantization": {
        "mode": "none",
        "bitWidth": 4,
        "prodMode": false
      }
    },
    "decodeLoop": {
      "batchSize": 4,
      "stopCheckMode": "batch",
      "readbackInterval": 1,
      "readbackMode": "sequential",
      "ringTokens": 1,
      "ringStop": 1,
      "ringStaging": 1,
      "disableCommandBatching": false
    },
    "perLayerInputs": {
      "materialization": "auto",
      "rowCache": {
        "mode": "lru",
        "maxRows": 256,
        "maxBytes": 134217728,
        "decodedDtype": "f32"
      },
      "prefetch": {
        "mode": "next_token",
        "rowsAhead": 1
      },
      "gpuUpload": {
        "mode": "per_step_slices",
        "stagingRows": 2
      },
      "hotCache": {
        "mode": "prepared_tokens",
        "maxTokens": 1024,
        "maxBytes": 134217728,
        "outputDtype": "f32"
      }
    },
    "prefillChunkSubmitMode": "sync",
    "useFlashPrefillAttention": false,
    "useWideTileQ4KPrefill": false,
    "retainQ4KMaterialization": false
  },
  "execution": {
    "kernels": {
      "embed": {
        "kernel": "gather.wgsl",
        "entry": "main",
        "digest": "sha256:4b12653c53247b32ebde7f6cf6a989d6248977e3816c761540b990b5f9818cb6"
      },
      "rmsnorm": {
        "kernel": "rmsnorm.wgsl",
        "entry": "main",
        "digest": "sha256:284d3efb0ad0991fc57ece5f634cbc8a931fcab36f193b2b6832561d4cc79ef1"
      },
      "tiled": {
        "kernel": "matmul_f16w_f32a.wgsl",
        "entry": "main",
        "digest": "sha256:027a8f1cd9713cbe0b0ada160bd175e0542bb90896ad85441b023522d9a1befc"
      },
      "rope": {
        "kernel": "rope.wgsl",
        "entry": "main",
        "digest": "sha256:b2da9d396668981dab9794c2973b668279f768994466b083d2105730555e1a5b"
      },
      "attn_small": {
        "kernel": "attention_small.wgsl",
        "entry": "main",
        "digest": "sha256:6752ddd7ab53e6235c9b5b1a9515141c0d111df7fac9f4c0d7a38f9943490ed4"
      },
      "residual": {
        "kernel": "residual.wgsl",
        "entry": "main",
        "digest": "sha256:abd19bc08ad668a7cad562e2bc0a4be1aa8827ad4300d8d474b7fc2af4c27117"
      },
      "gelu": {
        "kernel": "gelu.wgsl",
        "entry": "main",
        "digest": "sha256:7f8900b69de5107e4cf424cac0d4de1539591ece34e977ff30f6f7364460ab34",
        "constants": {
          "HAS_GATE": true
        }
      }
    },
    "preLayer": [
      [
        "embed",
        "embed",
        "embed_tokens"
      ]
    ],
    "decode": [
      [
        "input_norm",
        "rmsnorm"
      ],
      [
        "q_proj",
        "tiled",
        "layer.{L}.self_attn.q_proj"
      ],
      [
        "k_proj",
        "tiled",
        "layer.{L}.self_attn.k_proj"
      ],
      [
        "v_proj",
        "tiled",
        "layer.{L}.self_attn.v_proj"
      ],
      [
        "rope_q",
        "rope"
      ],
      [
        "rope_k",
        "rope"
      ],
      [
        "attention",
        "attn_small"
      ],
      [
        "o_proj",
        "tiled",
        "layer.{L}.self_attn.o_proj"
      ],
      [
        "attn_residual",
        "residual"
      ],
      [
        "post_attn_norm",
        "rmsnorm"
      ],
      [
        "gate_proj",
        "tiled",
        "layer.{L}.mlp.gate_proj"
      ],
      [
        "up_proj",
        "tiled",
        "layer.{L}.mlp.up_proj"
      ],
      [
        "activation",
        "gelu"
      ],
      [
        "down_proj",
        "tiled",
        "layer.{L}.mlp.down_proj"
      ],
      [
        "ffn_residual",
        "residual"
      ]
    ],
    "prefill": [
      [
        "input_norm",
        "rmsnorm"
      ],
      [
        "q_proj",
        "tiled",
        "layer.{L}.self_attn.q_proj"
      ],
      [
        "k_proj",
        "tiled",
        "layer.{L}.self_attn.k_proj"
      ],
      [
        "v_proj",
        "tiled",
        "layer.{L}.self_attn.v_proj"
      ],
      [
        "rope_q",
        "rope"
      ],
      [
        "rope_k",
        "rope"
      ],
      [
        "attention",
        "attn_small"
      ],
      [
        "o_proj",
        "tiled",
        "layer.{L}.self_attn.o_proj"
      ],
      [
        "attn_residual",
        "residual"
      ],
      [
        "post_attn_norm",
        "rmsnorm"
      ],
      [
        "gate_proj",
        "tiled",
        "layer.{L}.mlp.gate_proj"
      ],
      [
        "up_proj",
        "tiled",
        "layer.{L}.mlp.up_proj"
      ],
      [
        "activation",
        "gelu"
      ],
      [
        "down_proj",
        "tiled",
        "layer.{L}.mlp.down_proj"
      ],
      [
        "ffn_residual",
        "residual"
      ]
    ],
    "postLayer": [
      [
        "final_norm",
        "rmsnorm"
      ]
    ],
    "policies": {
      "unsupportedPrecision": "error",
      "dtypeTransition": "require_cast_step",
      "unresolvedKernel": "error"
    }
  },
  "largeWeights": {
    "gpuResidentOverrides": null
  },
  "manifest": {
    "artifactIdentity": {
      "sourceCheckpointId": "google/embeddinggemma-300m",
      "weightPackId": "google-embeddinggemma-300m-q4k-ehf16-af32-wp-catalog-v1",
      "manifestVariantId": "google-embeddinggemma-300m-q4k-ehf16-af32-mv-exec-v1",
      "artifactCompleteness": "complete"
    }
  }
}