{
  "modelType": "embedding",
  "output": {
    "baseDir": "models/local",
    "modelBaseId": "esmc-300m-f32-af32",
    "textOnly": true,
    "fast": false
  },
  "architecture": {
    "numLayers": 30,
    "hiddenSize": 960,
    "intermediateSize": 2560,
    "numAttentionHeads": 15,
    "numKeyValueHeads": 15,
    "headDim": 64,
    "vocabSize": 64,
    "maxSeqLen": 2048,
    "ropeTheta": 10000
  },
  "quantization": {
    "weights": "f32",
    "embeddings": "f32",
    "lmHead": "f32",
    "q4kLayout": "row",
    "computePrecision": "f32"
  },
  "sourceTensors": {
    "requireAll": true,
    "rules": [
      {
        "kind": "rename",
        "match": "^esmc\\.embed\\.weight$",
        "replace": "model.embed_tokens.weight",
        "expectedMatches": 1,
        "role": "embedding",
        "group": "embed"
      },
      {
        "kind": "rename",
        "match": "^esmc\\.transformer\\.norm\\.weight$",
        "replace": "model.norm.weight",
        "expectedMatches": 1,
        "role": "norm",
        "group": "head"
      },
      {
        "kind": "rename",
        "match": "^esmc\\.transformer\\.blocks\\.(\\d+)\\.attn\\.layernorm_qkv\\.layer_norm_weight$",
        "replace": "model.layers.$1.input_layernorm.weight",
        "expectedMatches": 30
      },
      {
        "kind": "rename",
        "match": "^esmc\\.transformer\\.blocks\\.(\\d+)\\.attn\\.layernorm_qkv\\.layer_norm_bias$",
        "replace": "model.layers.$1.input_layernorm.bias",
        "expectedMatches": 30
      },
      {
        "kind": "split",
        "match": "^esmc\\.transformer\\.blocks\\.(\\d+)\\.attn\\.layernorm_qkv\\.weight$",
        "axis": 0,
        "expectedMatches": 30,
        "parts": [
          { "replace": "model.layers.$1.self_attn.q_proj.weight", "size": 960 },
          { "replace": "model.layers.$1.self_attn.k_proj.weight", "size": 960 },
          { "replace": "model.layers.$1.self_attn.v_proj.weight", "size": 960 }
        ]
      },
      {
        "kind": "rename",
        "match": "^esmc\\.transformer\\.blocks\\.(\\d+)\\.attn\\.q_ln\\.weight$",
        "replace": "model.layers.$1.self_attn.q_norm.weight",
        "expectedMatches": 30
      },
      {
        "kind": "rename",
        "match": "^esmc\\.transformer\\.blocks\\.(\\d+)\\.attn\\.k_ln\\.weight$",
        "replace": "model.layers.$1.self_attn.k_norm.weight",
        "expectedMatches": 30
      },
      {
        "kind": "rename",
        "match": "^esmc\\.transformer\\.blocks\\.(\\d+)\\.attn\\.out_proj\\.weight$",
        "replace": "model.layers.$1.self_attn.o_proj.weight",
        "expectedMatches": 30
      },
      {
        "kind": "rename",
        "match": "^esmc\\.transformer\\.blocks\\.(\\d+)\\.ffn\\.layer_norm_weight$",
        "replace": "model.layers.$1.post_attention_layernorm.weight",
        "expectedMatches": 30
      },
      {
        "kind": "rename",
        "match": "^esmc\\.transformer\\.blocks\\.(\\d+)\\.ffn\\.layer_norm_bias$",
        "replace": "model.layers.$1.post_attention_layernorm.bias",
        "expectedMatches": 30
      },
      {
        "kind": "split",
        "match": "^esmc\\.transformer\\.blocks\\.(\\d+)\\.ffn\\.fc1_weight$",
        "axis": 0,
        "expectedMatches": 30,
        "parts": [
          { "replace": "model.layers.$1.mlp.gate_proj.weight", "size": 2560 },
          { "replace": "model.layers.$1.mlp.up_proj.weight", "size": 2560 }
        ]
      },
      {
        "kind": "rename",
        "match": "^esmc\\.transformer\\.blocks\\.(\\d+)\\.ffn\\.fc2_weight$",
        "replace": "model.layers.$1.mlp.down_proj.weight",
        "expectedMatches": 30
      },
      {
        "kind": "ignore",
        "match": "^esmc\\.transformer\\.blocks\\.\\d+\\.(?:attn\\.(?:layernorm_qkv|out_proj)|ffn)\\._extra_state$",
        "expectedMatches": 90,
        "reason": "PyTorch module extra-state markers contain no inference parameters."
      },
      {
        "kind": "ignore",
        "match": "^lm_head\\.",
        "expectedMatches": 6,
        "reason": "The initial ESMC lane exposes embeddings. Its dense-GELU-LayerNorm masked-token head requires a separately qualified output pipeline."
      }
    ]
  },
  "manifest": {
    "hashAlgorithm": "blake3",
    "artifactIdentity": {
      "sourceCheckpointId": "biohub/ESMC-300M@a59b831785f907e96e6a246b1d142bfb76df31ee",
      "sourceRepo": "biohub/ESMC-300M",
      "sourceRevision": "a59b831785f907e96e6a246b1d142bfb76df31ee",
      "sourceFormat": "safetensors",
      "conversionConfigPath": "src/config/conversion/esmc/esmc-300m-f32-af32.json",
      "weightPackId": "esmc-300m-f32-af32-wp-v1",
      "manifestVariantId": "esmc-300m-f32-af32-mv-exec-v1",
      "modalitySet": ["protein_sequence"],
      "materializationProfile": "f32-weights-f32-activations",
      "artifactCompleteness": "complete"
    },
    "eosTokenId": 2
  },
  "inference": {
    "supportsEmbedding": true,
    "supportsSequence": true,
    "sequence": {
      "alphabet": "amino_acid",
      "tokenEmbeddings": true,
      "pooledEmbedding": {
        "mode": "mean",
        "excludeTokenIds": [0, 1, 2, 3, 31, 32]
      },
      "logits": false
    },
    "attention": {
      "queryPreAttnScalar": 64,
      "attnLogitSoftcapping": null,
      "slidingWindow": null,
      "queryKeyNorm": true,
      "queryKeyNormType": "layernorm",
      "queryKeyNormAxis": "projection",
      "queryKeyNormLayers": null,
      "queryKeyNormWeightLayers": null,
      "valueNorm": false,
      "causal": false,
      "attentionBias": true,
      "attentionOutputGate": false,
      "outputGateType": null
    },
    "normalization": {
      "type": "layernorm",
      "rmsNormEps": 0.00001,
      "rmsNormWeightOffset": false,
      "postAttentionNorm": true,
      "preFeedforwardNorm": false,
      "postFeedforwardNorm": false,
      "finalNormBiasTensor": null
    },
    "ffn": {
      "activation": "silu",
      "gatedActivation": true,
      "branchMode": "dense",
      "useDoubleWideMlp": false,
      "swigluLimit": null
    },
    "rope": {
      "ropeTheta": 10000,
      "ropeLocalTheta": null,
      "ropeInterleaved": false,
      "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": false,
      "embeddingScale": null,
      "logitInputScale": 1,
      "embeddingTranspose": false,
      "embeddingVocabSize": null,
      "embeddingPostprocessor": null,
      "lmHeadBiasTensor": null
    },
    "layerPattern": {
      "type": "uniform",
      "globalPattern": null,
      "period": null,
      "offset": null,
      "layerTypes": null,
      "residualBranchScale": 1.0954451150103321
    },
    "chatTemplate": { "type": null, "enabled": false }
  },
  "session": {
    "compute": {
      "defaults": {
        "activationDtype": "f32",
        "mathDtype": "f32",
        "accumDtype": "f32",
        "outputDtype": "f32"
      }
    },
    "kvcache": {
      "kvDtype": "f32",
      "layout": "contiguous",
      "maxSeqLen": 2048,
      "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": 1,
      "stopCheckMode": "per-token",
      "readbackInterval": 1,
      "readbackMode": "sequential",
      "submitLatencyThresholdMs": null,
      "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" }
    },
    "skipEmbeddingKVCacheWrites": true,
    "prefillChunkSubmitMode": "sync",
    "useFlashPrefillAttention": false,
    "useWideTileQ4KPrefill": false,
    "retainQ4KMaterialization": false
  },
  "execution": {
    "inlineKernelPath": true,
    "policies": {
      "unsupportedPrecision": "error",
      "dtypeTransition": "require_cast_step",
      "unresolvedKernel": "error"
    },
    "kernels": {
      "embed": { "kernel": "gather.wgsl", "entry": "main", "digest": "sha256:4b12653c53247b32ebde7f6cf6a989d6248977e3816c761540b990b5f9818cb6" },
      "layernorm": { "kernel": "layernorm.wgsl", "entry": "main", "digest": "sha256:8ff885017bfcfc8fdb8d16e09c358c000830bfd5fbf3d87187694edb0cb44f7c" },
      "tiled": { "kernel": "matmul_f32.wgsl", "entry": "main", "digest": "sha256:b5bb8e3d8014136e33de7935dd2a1f074c988044fe05cf5b559718c6f061eaa8" },
      "bias_add": { "kernel": "bias_add.wgsl", "entry": "main", "digest": "sha256:e7d3b4a5f9efc8b0569c56bcdeb63e921fa6d1d006901deabc734cab40346715" },
      "rope": { "kernel": "rope.wgsl", "entry": "main", "digest": "sha256:b2da9d396668981dab9794c2973b668279f768994466b083d2105730555e1a5b" },
      "attn_small": { "kernel": "attention_small.wgsl", "entry": "main", "digest": "sha256:6752ddd7ab53e6235c9b5b1a9515141c0d111df7fac9f4c0d7a38f9943490ed4", "precision": { "kvDtype": "f32" } },
      "residual": { "kernel": "residual.wgsl", "entry": "main", "digest": "sha256:abd19bc08ad668a7cad562e2bc0a4be1aa8827ad4300d8d474b7fc2af4c27117" },
      "silu": { "kernel": "silu.wgsl", "entry": "main", "digest": "sha256:d6b21e62031ac0d748617f7ebfb43834552a5b8f9590c18dc19cd0cd41a2fbca", "constants": { "HAS_GATE": true } }
    },
    "preLayer": [["embed", "embed", "embed_tokens"]],
    "decode": [
      ["input_norm", "layernorm"],
      ["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", "layernorm"],
      ["gate_proj", "tiled", "layer.{L}.mlp.gate_proj"],
      ["up_proj", "tiled", "layer.{L}.mlp.up_proj"],
      ["activation", "silu"],
      ["down_proj", "tiled", "layer.{L}.mlp.down_proj"],
      ["ffn_residual", "residual"]
    ],
    "prefill": [
      ["input_norm", "layernorm"],
      ["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", "layernorm"],
      ["gate_proj", "tiled", "layer.{L}.mlp.gate_proj"],
      ["up_proj", "tiled", "layer.{L}.mlp.up_proj"],
      ["activation", "silu"],
      ["down_proj", "tiled", "layer.{L}.mlp.down_proj"],
      ["ffn_residual", "residual"]
    ],
    "postLayer": [["final_norm", "layernorm"]],
    "sampling": null
  },
  "largeWeights": { "gpuResidentOverrides": null }
}