{ "id": "litert/gemma-4-12b-it", "runtime": { "modelType": "gemma4", "architecture": { "vocabSize": 262144, "hiddenSize": 3840, "intermediateSize": 15360, "numLayers": 47, "numAttentionHeads": 16, "numKeyValueHeads": 8, "numGlobalKeyValueHeads": 1, "headDim": 256, "globalHeadDim": 512, "hiddenSizePerLayerInput": 0, "vocabSizePerLayerInput": 0, "numKvSharedLayers": 0, "attentionKeyEqualsValue": true, "maxSeqLen": 131072 }, "rawConfig": { "architectures": [ "Gemma4ForConditionalGeneration" ], "model_type": "gemma4_unified", "eos_token_id": [ 1, 106 ], "text_config": { "model_type": "gemma4_unified_text", "vocab_size": 262144, "hidden_size": 3840, "intermediate_size": 15360, "num_hidden_layers": 47, "num_attention_heads": 16, "num_key_value_heads": 8, "num_global_key_value_heads": 1, "head_dim": 256, "global_head_dim": 512, "hidden_activation": "gelu_pytorch_tanh", "max_position_embeddings": 131072, "sliding_window": 1024, "rms_norm_eps": 0.000001, "rope_theta": 1000000, "rope_local_base_freq": 10000, "rope_scaling": { "rope_type": "default", "factor": 8 }, "attention_k_eq_v": true, "attention_bias": false, "hidden_size_per_layer_input": 0, "eos_token_id": 1, "pad_token_id": 0, "bos_token_id": 2, "layer_types": [ "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention", "full_attention", "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention", "full_attention", "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention", "full_attention", "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention", "full_attention", "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention", "full_attention", "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention", "full_attention", "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention", "full_attention", "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention" ] } }, "manifestConfig": {}, "manifestInference": { "schema": "doppler.execution/v1", "unsupported": { "code": "gemma4-12b-litert-direct-source-unverified", "message": "The current LiteRT-LM artifact exposes 47 repeated decoder-weight groups while the current source model contract exposes 48 text layers; Doppler's split-section adapter maps generic TFLite constants by shape/order and currently produces saturated logits on this artifact.", "recommendation": "Use a supported RDRR/QAT conversion path, or add an explicit LiteRT graph contraction map with norm/projection parity receipts before enabling direct-source execution." }, "attention": { "queryPreAttnScalar": 1, "attnLogitSoftcapping": null, "slidingWindow": 1024, "queryKeyNorm": true, "queryKeyNormLayers": null, "queryKeyNormWeightLayers": [ 2, 5, 9, 10, 11, 13, 15, 16, 17, 18, 23, 25, 30, 31, 33, 35, 36, 38, 39, 40, 41, 42, 43, 44, 46 ], "valueNorm": true, "causal": true, "attentionBias": false, "attentionOutputGate": false }, "normalization": { "rmsNormEps": 0.000001, "rmsNormWeightOffset": false, "postAttentionNorm": true, "preFeedforwardNorm": true, "postFeedforwardNorm": true }, "ffn": { "activation": "gelu", "gatedActivation": true, "branchMode": "auto", "useDoubleWideMlp": false, "swigluLimit": null }, "rope": { "ropeTheta": 1000000, "ropeLocalTheta": 10000, "ropeInterleaved": false, "mropeInterleaved": false, "mropeSection": null, "partialRotaryFactor": 0.25, "ropeLocalPartialRotaryFactor": null, "ropeFrequencyBaseDim": 512, "ropeLocalFrequencyBaseDim": null, "ropeScalingType": null, "ropeScalingFactor": 1, "ropeLocalScalingType": null, "ropeLocalScalingFactor": 1, "yarnBetaFast": null, "yarnBetaSlow": null, "yarnOriginalMaxPos": null, "ropeLocalYarnBetaFast": null, "ropeLocalYarnBetaSlow": null, "ropeLocalYarnOriginalMaxPos": null }, "output": { "finalLogitSoftcapping": 30, "tieWordEmbeddings": true, "scaleEmbeddings": true, "embeddingTranspose": false, "embeddingVocabSize": 262144, "embeddingPostprocessor": null }, "layerPattern": { "type": "custom", "globalPattern": null, "period": null, "offset": null, "layerTypes": [ "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention", "full_attention", "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention", "full_attention", "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention", "full_attention", "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention", "full_attention", "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention", "full_attention", "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention", "full_attention", "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention", "full_attention", "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention", "sliding_attention" ] }, "chatTemplate": { "type": "gemma4", "enabled": true }, "pipeline": null, "session": { "compute": { "defaults": { "activationDtype": "f32", "mathDtype": "f32", "accumDtype": "f32", "outputDtype": "f32" } }, "kvcache": { "kvDtype": "f16", "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": 1, "stopCheckMode": "batch", "readbackInterval": 1, "readbackMode": "overlapped", "ringTokens": 2, "ringStop": 1, "ringStaging": 2, "disableCommandBatching": false }, "perLayerInputs": { "materialization": "gpu_split_tables", "rowCache": { "mode": "lru", "maxRows": 512, "maxBytes": 268435456, "decodedDtype": "f32" }, "prefetch": { "mode": "next_token", "rowsAhead": 1 }, "gpuUpload": { "mode": "per_step_slices", "stagingRows": 2 }, "hotCache": { "mode": "prepared_tokens", "maxTokens": 4096, "maxBytes": 268435456, "outputDtype": "f32" } }, "prefillChunkSubmitMode": "sync", "useFlashPrefillAttention": false, "useWideTileQ4KPrefill": false, "retainQ4KMaterialization": false } }, "tokenizer": { "litertlm": { "type": "sentencepiece", "sentencepieceModel": "TOKENIZER_MODEL", "vocabSize": 262144, "padTokenId": 0, "eosTokenId": 1, "eosTokens": [ 1, 106 ], "bosTokenId": 2, "unkTokenId": 3, "addBosToken": false, "addEosToken": false } } }, "package": { "litertlm": { "tfliteModelType": "tf_lite_prefill_decode", "embedderTFLiteModelType": "tf_lite_embedder", "graphAdapter": "gemma4_unified_12b_split_int4", "fixedInt4Scale": 0.0625, "executionTemplateProfileId": "litert/gemma-4-e2b-it", "fixedInt4StorageEncoding": "offset_binary", "unsupported": { "code": "gemma4-12b-litert-direct-source-unverified", "message": "The current LiteRT-LM artifact exposes 47 repeated decoder-weight groups while the current source model contract exposes 48 text layers; Doppler's split-section adapter maps generic TFLite constants by shape/order and currently produces saturated logits on this artifact.", "recommendation": "Use a supported RDRR/QAT conversion path, or add an explicit LiteRT graph contraction map with norm/projection parity receipts before enabling direct-source execution." }, "layerScalarLayers": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 22, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 46 ], "missingLayerScalarValue": 1 } } }