{ "id": "litert/gemma-4-e4b-it", "runtime": { "modelType": "gemma4", "architecture": { "numLayers": 42, "hiddenSize": 2560, "intermediateSize": 10240, "numAttentionHeads": 8, "numKeyValueHeads": 2, "headDim": 256, "vocabSize": 262144, "maxSeqLen": 131072, "linearNumKeyHeads": null, "linearNumValueHeads": null, "linearKeyHeadDim": null, "linearValueHeadDim": null, "linearConvKernelDim": null, "hiddenSizePerLayerInput": 256, "vocabSizePerLayerInput": 262144, "globalHeadDim": 512, "numKvSharedLayers": 18 }, "rawConfig": { "architectures": [ "Gemma4ForConditionalGeneration" ], "model_type": "gemma4", "text_config": { "model_type": "gemma4_text", "hidden_size": 2560, "intermediate_size": 10240, "num_hidden_layers": 42, "num_attention_heads": 8, "num_key_value_heads": 2, "num_kv_shared_layers": 18, "head_dim": 256, "global_head_dim": 512, "hidden_size_per_layer_input": 256, "vocab_size_per_layer_input": 262144, "hidden_activation": "gelu_pytorch_tanh", "sliding_window": 512, "rms_norm_eps": 0.000001, "rope_parameters": { "full_attention": { "partial_rotary_factor": 0.25, "rope_theta": 1000000, "rope_type": "proportional" }, "sliding_attention": { "rope_theta": 10000, "rope_type": "default" } }, "attention_k_eq_v": false, "attention_bias": false, "use_double_wide_mlp": false, "tie_word_embeddings": true, "final_logit_softcapping": 30, "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" ] }, "eos_token_id": [ 1, 106 ], "image_token_id": 258880, "audio_token_id": 258881, "video_token_id": 258884, "vision_config": { "model_type": "gemma4_vision", "hidden_size": 768, "num_attention_heads": 12, "num_hidden_layers": 16, "num_key_value_heads": 12, "head_dim": 64, "global_head_dim": 64, "intermediate_size": 3072, "patch_size": 16, "pooling_kernel_size": 3, "position_embedding_size": 10240, "default_output_length": 280, "rms_norm_eps": 0.000001, "hidden_activation": "gelu_pytorch_tanh", "use_clipped_linears": true, "standardize": false, "rope_parameters": { "rope_theta": 100 } }, "audio_config": { "model_type": "gemma4_audio", "audio_architecture": "gemma4", "num_hidden_layers": 12, "hidden_size": 1024, "num_attention_heads": 8, "conv_kernel_size": 5, "subsampling_conv_channels": [ 128, 32 ], "output_proj_dims": 1536, "attention_context_left": 13, "attention_context_right": 0, "attention_chunk_size": 12, "attention_logit_cap": 50, "attention_invalid_logits_value": -1000000000, "residual_weight": 0.5, "rms_norm_eps": 0.000001, "hidden_act": "silu", "use_clipped_linears": true } }, "manifestConfig": { "visionConfig": { "vision_architecture": "gemma4", "hidden_size": 768, "num_attention_heads": 12, "num_hidden_layers": 16, "num_key_value_heads": 12, "head_dim": 64, "intermediate_size": 3072, "patch_size": 16, "pooling_kernel_size": 3, "position_embedding_size": 10240, "default_output_length": 280, "rms_norm_eps": 0.000001, "hidden_activation": "gelu_pytorch_tanh", "use_clipped_linears": true, "standardize": false, "rope_parameters": { "rope_theta": 100 } }, "audioConfig": { "audio_architecture": "gemma4", "num_hidden_layers": 12, "hidden_size": 1024, "num_attention_heads": 8, "conv_kernel_size": 5, "subsampling_conv_channels": [ 128, 32 ], "output_proj_dims": 1536, "attention_context_left": 13, "attention_context_right": 0, "attention_chunk_size": 12, "attention_logit_cap": 50, "attention_invalid_logits_value": -1000000000, "residual_weight": 0.5, "rms_norm_eps": 0.000001, "hidden_act": "silu", "use_clipped_linears": true } }, "manifestInference": { "schema": "doppler.execution/v1", "attention": { "queryPreAttnScalar": 1, "attnLogitSoftcapping": null, "slidingWindow": 512, "queryKeyNorm": true, "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": null, "embeddingPostprocessor": null }, "layerPattern": { "type": "every_n", "globalPattern": null, "period": 6, "offset": 5, "layerTypes": null }, "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": "range_backed", "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" } } } }, "tokenizer": { "task": { "sentencepieceModel": "TOKENIZER_MODEL", "vocabSize": 262144, "padTokenId": 0, "eosTokenId": 1, "eosTokens": [ 1, 106 ], "bosTokenId": 2, "unkTokenId": 3, "addBosToken": false, "addEosToken": false }, "litertlm": { "sentencepieceModel": "TOKENIZER_MODEL", "vocabSize": 262144, "padTokenId": 0, "eosTokenId": 1, "eosTokens": [ 1, 106 ], "bosTokenId": 2, "unkTokenId": 3, "addBosToken": false, "addEosToken": false } } }, "package": { "task": { "tfliteEntry": "TF_LITE_PREFILL_DECODE", "tokenizerModelEntry": "TOKENIZER_MODEL", "metadataEntry": "METADATA", "unsupported": { "code": "gemma4-e4b-litert-direct-source-unverified", "message": "The public Gemma 4 E4B LiteRT web package is registered for discovery, but Doppler has no E4B direct-source parity receipt yet and external reports show tokenizer/detokenizer mismatch risk on this package family.", "recommendation": "Use this profile only to keep package detection explicit. Enable execution after adding a source-package parity receipt, tokenizer round-trip receipt, and RDRR conversion path." } }, "litertlm": { "tfliteModelType": "TF_LITE_PREFILL_DECODE", "tokenizerSectionType": "SP_Tokenizer", "metadataSectionType": "LlmMetadataProto", "executionTemplateProfileId": "litert/gemma-4-e2b-it", "unsupported": { "code": "gemma4-e4b-litert-direct-source-unverified", "message": "The public Gemma 4 E4B LiteRT-LM package is registered for discovery, but Doppler has no E4B direct-source parity receipt or RDRR conversion path yet.", "recommendation": "Add an explicit E4B LiteRT graph contraction map, tokenizer receipt, and text-generation parity receipt before enabling direct-source execution." } } } }