| { |
| "architecture": { |
| "attention_variant": "GQA", |
| "family": "image_text_to_text", |
| "is_moe": false, |
| "mixer": "attention", |
| "positional": "rope", |
| "sliding_window": 1024, |
| "tie_word_embeddings": true, |
| "towers": [ |
| { |
| "attention_variant": "GQA", |
| "config_class": "Gemma4UnifiedTextConfig", |
| "hidden_size": 2304, |
| "name": "text", |
| "num_attention_heads": 8, |
| "num_hidden_layers": 30, |
| "positional": "rope" |
| } |
| ], |
| "view": "multimodal" |
| }, |
| "capabilities": { |
| "attention_backends": [ |
| "eager", |
| "sdpa", |
| "flash_attention", |
| "flex_attention" |
| ], |
| "attention_patterns": [ |
| "sliding", |
| "causal" |
| ], |
| "attention_schedule": [ |
| "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" |
| ], |
| "task_heads": [ |
| "causal_lm", |
| "image_text_to_text", |
| "multimodal_lm", |
| "pretraining" |
| ], |
| "tensor_parallel": true |
| }, |
| "components": [ |
| { |
| "children": [ |
| "language_model" |
| ], |
| "class_name": "Gemma4UnifiedModel", |
| "id": "model", |
| "kind": "model", |
| "path_pattern": "model" |
| }, |
| { |
| "children": [ |
| "language_model.embed_tokens", |
| "decoder_layers", |
| "language_model.norm", |
| "language_model.rotary_emb" |
| ], |
| "class_name": "Gemma4UnifiedTextModel", |
| "id": "language_model", |
| "kind": "model", |
| "path_pattern": "model.language_model" |
| }, |
| { |
| "attributes": { |
| "embedding_dim": "config.hidden_size", |
| "num_embeddings": "config.vocab_size", |
| "tied_lm_head": true |
| }, |
| "class_name": "Gemma4UnifiedTextScaledWordEmbedding", |
| "id": "language_model.embed_tokens", |
| "kind": "embedding", |
| "path_pattern": "model.language_model.embed_tokens" |
| }, |
| { |
| "attributes": { |
| "norm_type": "rms" |
| }, |
| "class_name": "Gemma4UnifiedRMSNorm", |
| "id": "language_model.norm", |
| "kind": "normalization", |
| "path_pattern": "model.language_model.norm" |
| }, |
| { |
| "attributes": { |
| "head_dim": "config.head_dim", |
| "scheme": "rope" |
| }, |
| "class_name": "Gemma4UnifiedTextRotaryEmbedding", |
| "id": "language_model.rotary_emb", |
| "kind": "position", |
| "path_pattern": "model.language_model.rotary_emb" |
| } |
| ], |
| "config": { |
| "class_name": "Gemma4UnifiedConfig", |
| "model_type": "gemma4_unified", |
| "module": "transformers.models.gemma4_unified.configuration_gemma4_unified", |
| "referenced_fields": {}, |
| "salient_fields": { |
| "is_encoder_decoder": false, |
| "tie_word_embeddings": true |
| } |
| }, |
| "edges": [ |
| { |
| "kind": "data", |
| "source": "language_model.embed_tokens", |
| "target": "decoder_layers" |
| }, |
| { |
| "kind": "data", |
| "source": "decoder_layers", |
| "target": "language_model.norm" |
| }, |
| { |
| "kind": "data", |
| "source": "decoder_layer.self_attn", |
| "target": "decoder_layer.mlp" |
| }, |
| { |
| "kind": "data", |
| "source": "decoder_layer.mlp", |
| "target": "decoder_layer.input_layernorm" |
| }, |
| { |
| "kind": "data", |
| "source": "decoder_layer.input_layernorm", |
| "target": "decoder_layer.post_attention_layernorm" |
| }, |
| { |
| "kind": "data", |
| "source": "decoder_layer.post_attention_layernorm", |
| "target": "decoder_layer.pre_feedforward_layernorm" |
| }, |
| { |
| "kind": "data", |
| "source": "decoder_layer.pre_feedforward_layernorm", |
| "target": "decoder_layer.post_feedforward_layernorm" |
| }, |
| { |
| "kind": "data", |
| "source": "decoder_layer.self_attn.k_norm", |
| "target": "decoder_layer.self_attn.v_norm" |
| }, |
| { |
| "kind": "mask", |
| "source": "input:attention_mask", |
| "target": "decoder_layer.self_attn" |
| }, |
| { |
| "kind": "residual", |
| "source": "decoder_layer", |
| "target": "decoder_layer.self_attn" |
| }, |
| { |
| "kind": "residual", |
| "source": "decoder_layer", |
| "target": "decoder_layer.mlp" |
| }, |
| { |
| "kind": "position", |
| "source": "language_model.rotary_emb", |
| "target": "decoder_layer.self_attn" |
| }, |
| { |
| "kind": "data", |
| "provenance": "intra_module", |
| "source": "decoder_layer.self_attn", |
| "target": "decoder_layer.self_attn.q_proj" |
| }, |
| { |
| "kind": "data", |
| "provenance": "intra_module", |
| "source": "decoder_layer.self_attn", |
| "target": "decoder_layer.self_attn.k_proj" |
| }, |
| { |
| "kind": "data", |
| "provenance": "intra_module", |
| "source": "decoder_layer.self_attn", |
| "target": "decoder_layer.self_attn.v_proj" |
| }, |
| { |
| "kind": "data", |
| "provenance": "intra_module", |
| "source": "decoder_layer.self_attn.q_proj", |
| "target": "decoder_layer.self_attn.q_norm" |
| }, |
| { |
| "kind": "data", |
| "provenance": "intra_module", |
| "source": "decoder_layer.self_attn.k_proj", |
| "target": "decoder_layer.self_attn.k_norm" |
| }, |
| { |
| "kind": "data", |
| "provenance": "intra_module", |
| "source": "decoder_layer.self_attn.q_norm", |
| "target": "decoder_layer.self_attn.o_proj" |
| }, |
| { |
| "kind": "data", |
| "provenance": "intra_module", |
| "source": "decoder_layer.self_attn.k_norm", |
| "target": "decoder_layer.self_attn.o_proj" |
| }, |
| { |
| "kind": "data", |
| "provenance": "intra_module", |
| "source": "decoder_layer.self_attn.v_proj", |
| "target": "decoder_layer.self_attn.o_proj" |
| }, |
| { |
| "kind": "data", |
| "provenance": "intra_module", |
| "source": "decoder_layer.mlp", |
| "target": "decoder_layer.mlp.gate_proj" |
| }, |
| { |
| "kind": "data", |
| "provenance": "intra_module", |
| "source": "decoder_layer.mlp", |
| "target": "decoder_layer.mlp.up_proj" |
| }, |
| { |
| "kind": "data", |
| "provenance": "intra_module", |
| "source": "decoder_layer.mlp.gate_proj", |
| "target": "decoder_layer.mlp.down_proj" |
| }, |
| { |
| "kind": "data", |
| "provenance": "intra_module", |
| "source": "decoder_layer.mlp.up_proj", |
| "target": "decoder_layer.mlp.down_proj" |
| } |
| ], |
| "extends": "gemma4", |
| "model_type": "gemma4_unified", |
| "patches": [ |
| { |
| "component_kind": null, |
| "parent_class": "Gemma4ImageProcessorKwargs", |
| "relation": "new", |
| "target_class": "Gemma4UnifiedImageProcessorKwargs" |
| }, |
| { |
| "added": { |
| "methods": [ |
| "_preprocess" |
| ] |
| }, |
| "component_kind": null, |
| "parent_class": "Gemma4ImageProcessor", |
| "relation": "new", |
| "target_class": "Gemma4UnifiedImageProcessor" |
| }, |
| { |
| "component_kind": null, |
| "parent_class": "Gemma4VideoProcessorKwargs", |
| "relation": "new", |
| "target_class": "Gemma4UnifiedVideoProcessorKwargs" |
| }, |
| { |
| "added": { |
| "methods": [ |
| "_preprocess" |
| ] |
| }, |
| "component_kind": null, |
| "parent_class": "Gemma4VideoProcessor", |
| "relation": "new", |
| "target_class": "Gemma4UnifiedVideoProcessor" |
| }, |
| { |
| "component_kind": null, |
| "parent_class": "Gemma4ProcessorKwargs", |
| "relation": "new", |
| "target_class": "Gemma4UnifiedProcessorKwargs" |
| }, |
| { |
| "added": { |
| "methods": [ |
| "_compute_audio_num_tokens", |
| "replace_audio_token" |
| ] |
| }, |
| "component_kind": null, |
| "parent_class": "Gemma4Processor", |
| "relation": "new", |
| "target_class": "Gemma4UnifiedProcessor" |
| }, |
| { |
| "added": { |
| "attrs": [ |
| "audio_embed_dim", |
| "initializer_range", |
| "model_type", |
| "rms_norm_eps" |
| ], |
| "methods": [ |
| "audio_samples_per_token", |
| "hidden_size", |
| "output_proj_dims" |
| ] |
| }, |
| "component_kind": "config", |
| "parent_class": "PreTrainedConfig", |
| "relation": "new", |
| "target_class": "Gemma4UnifiedAudioConfig" |
| }, |
| { |
| "added": { |
| "attrs": [ |
| "base_model_ep_plan", |
| "base_model_tp_plan", |
| "enable_moe_block", |
| "hidden_size_per_layer_input", |
| "max_position_embeddings", |
| "model_type", |
| "moe_intermediate_size", |
| "num_experts", |
| "sliding_window", |
| "top_k_experts", |
| "use_bidirectional_attention", |
| "vocab_size_per_layer_input" |
| ] |
| }, |
| "component_kind": "config", |
| "parent_class": "Gemma4TextConfig", |
| "relation": "new", |
| "target_class": "Gemma4UnifiedTextConfig" |
| }, |
| { |
| "added": { |
| "attrs": [ |
| "initializer_range", |
| "mm_embed_dim", |
| "mm_posemb_size", |
| "model_type", |
| "output_proj_dims", |
| "patch_size", |
| "pooling_kernel_size", |
| "rms_norm_eps" |
| ], |
| "methods": [ |
| "model_patch_size" |
| ] |
| }, |
| "component_kind": "config", |
| "parent_class": "PreTrainedConfig", |
| "relation": "new", |
| "target_class": "Gemma4UnifiedVisionConfig" |
| }, |
| { |
| "added": { |
| "attrs": [ |
| "audio_config", |
| "sub_configs", |
| "text_config", |
| "vision_config" |
| ] |
| }, |
| "component_kind": "config", |
| "parent_class": "Gemma4Config", |
| "relation": "new", |
| "target_class": "Gemma4UnifiedConfig" |
| }, |
| { |
| "added": { |
| "attrs": [ |
| "attention_mask", |
| "pooler_output" |
| ] |
| }, |
| "component_kind": null, |
| "parent_class": "ModelOutput", |
| "relation": "new", |
| "target_class": "Gemma4UnifiedAudioModelOutput" |
| }, |
| { |
| "added": { |
| "attrs": [ |
| "pooler_output" |
| ] |
| }, |
| "component_kind": null, |
| "parent_class": "ModelOutput", |
| "relation": "new", |
| "target_class": "Gemma4UnifiedVisionModelOutput" |
| }, |
| { |
| "component_kind": "transformer_block", |
| "overridden": { |
| "methods": [ |
| "__init__", |
| "forward" |
| ] |
| }, |
| "parent_class": "Gemma2DecoderLayer", |
| "parent_model": "gemma2", |
| "relation": "inherits", |
| "target_class": "Gemma4UnifiedTextDecoderLayer" |
| }, |
| { |
| "component_kind": "model", |
| "deleted": { |
| "methods": [ |
| "_resize_per_layer_embeddings", |
| "get_per_layer_input_embeddings", |
| "resize_token_embeddings", |
| "set_per_layer_input_embeddings" |
| ] |
| }, |
| "overridden": { |
| "attrs": [ |
| "_no_split_modules" |
| ], |
| "methods": [ |
| "_init_weights" |
| ] |
| }, |
| "parent_class": "Gemma4PreTrainedModel", |
| "relation": "inherits", |
| "target_class": "Gemma4UnifiedPreTrainedModel" |
| }, |
| { |
| "added": { |
| "attrs": [ |
| "_can_record_outputs", |
| "config", |
| "input_modalities" |
| ] |
| }, |
| "component_kind": "model", |
| "overridden": { |
| "methods": [ |
| "__init__", |
| "forward" |
| ] |
| }, |
| "parent_class": "LlamaModel", |
| "parent_model": "llama", |
| "relation": "inherits", |
| "target_class": "Gemma4UnifiedTextModel" |
| }, |
| { |
| "component_kind": "lm_head", |
| "overridden": { |
| "methods": [ |
| "forward" |
| ] |
| }, |
| "parent_class": "Gemma4ForCausalLM", |
| "relation": "inherits", |
| "target_class": "Gemma4UnifiedForCausalLM" |
| }, |
| { |
| "added": { |
| "methods": [ |
| "__init__", |
| "forward" |
| ] |
| }, |
| "component_kind": null, |
| "parent_class": "Module", |
| "relation": "new", |
| "target_class": "Gemma4UnifiedVisionEmbedder" |
| }, |
| { |
| "component_kind": null, |
| "overridden": { |
| "methods": [ |
| "__init__", |
| "forward" |
| ] |
| }, |
| "parent_class": "Gemma4MultimodalEmbedder", |
| "relation": "inherits", |
| "target_class": "Gemma4UnifiedMultimodalEmbedder" |
| }, |
| { |
| "component_kind": "model", |
| "deleted": { |
| "methods": [ |
| "get_per_layer_input_embeddings", |
| "set_per_layer_input_embeddings" |
| ] |
| }, |
| "overridden": { |
| "methods": [ |
| "__init__", |
| "forward", |
| "get_audio_features", |
| "get_image_features", |
| "get_video_features" |
| ] |
| }, |
| "parent_class": "Gemma4Model", |
| "relation": "inherits", |
| "target_class": "Gemma4UnifiedModel" |
| }, |
| { |
| "component_kind": "lm_head", |
| "deleted": { |
| "methods": [ |
| "get_per_layer_input_embeddings", |
| "set_per_layer_input_embeddings" |
| ] |
| }, |
| "overridden": { |
| "methods": [ |
| "forward", |
| "prepare_inputs_for_generation" |
| ] |
| }, |
| "parent_class": "Gemma4ForConditionalGeneration", |
| "relation": "inherits", |
| "target_class": "Gemma4UnifiedForConditionalGeneration" |
| } |
| ], |
| "provenance": { |
| "config_class": "Gemma4UnifiedConfig", |
| "config_module": "transformers.models.gemma4_unified.configuration_gemma4_unified", |
| "model_class": "Gemma4UnifiedModel", |
| "model_module": "transformers.models.gemma4_unified.modeling_gemma4_unified" |
| }, |
| "repeats": [ |
| { |
| "body": "decoder_layer", |
| "container_path_pattern": "model.language_model.layers", |
| "count": 30, |
| "count_expr": "30", |
| "count_source": "module_tree", |
| "id": "decoder_layers", |
| "index_symbol": "i", |
| "item_path_pattern": "model.language_model.layers.{i}", |
| "kind": "symbolic_repeat", |
| "provenance": { |
| "class_name": "Gemma4UnifiedTextDecoderLayer", |
| "container_path_pattern": "model.language_model.layers", |
| "item_path_pattern": "model.language_model.layers.{i}", |
| "source": "module_tree_repeat_collapse" |
| }, |
| "repeated_class_name": "Gemma4UnifiedTextDecoderLayer" |
| } |
| ], |
| "schema_version": "architecture-template-v0", |
| "templates": [ |
| { |
| "children": [ |
| "decoder_layer.self_attn", |
| "decoder_layer.mlp", |
| "decoder_layer.input_layernorm", |
| "decoder_layer.post_attention_layernorm", |
| "decoder_layer.pre_feedforward_layernorm", |
| "decoder_layer.post_feedforward_layernorm" |
| ], |
| "class_name": "Gemma4UnifiedTextDecoderLayer", |
| "id": "decoder_layer", |
| "kind": "transformer_block", |
| "path_pattern": "model.language_model.layers.{i}" |
| }, |
| { |
| "attributes": { |
| "head_dim": 256, |
| "n_heads": 8, |
| "n_kv_heads": 4, |
| "qk_norm": true, |
| "rope": true, |
| "sliding_window": 1024, |
| "variant": "GQA" |
| }, |
| "children": [ |
| "decoder_layer.self_attn.q_proj", |
| "decoder_layer.self_attn.q_norm", |
| "decoder_layer.self_attn.k_norm", |
| "decoder_layer.self_attn.v_norm", |
| "decoder_layer.self_attn.k_proj", |
| "decoder_layer.self_attn.v_proj", |
| "decoder_layer.self_attn.o_proj" |
| ], |
| "class_name": "Gemma4UnifiedTextAttention", |
| "id": "decoder_layer.self_attn", |
| "kind": "attention", |
| "path_pattern": "model.language_model.layers.{i}.self_attn" |
| }, |
| { |
| "attributes": { |
| "in_features": "config.hidden_size", |
| "out_features": 2048 |
| }, |
| "class_name": "Linear", |
| "id": "decoder_layer.self_attn.q_proj", |
| "kind": "projection", |
| "path_pattern": "model.language_model.layers.{i}.self_attn.q_proj" |
| }, |
| { |
| "attributes": { |
| "norm_type": "rms" |
| }, |
| "class_name": "Gemma4UnifiedRMSNorm", |
| "id": "decoder_layer.self_attn.q_norm", |
| "kind": "normalization", |
| "path_pattern": "model.language_model.layers.{i}.self_attn.q_norm" |
| }, |
| { |
| "attributes": { |
| "norm_type": "rms" |
| }, |
| "class_name": "Gemma4UnifiedRMSNorm", |
| "id": "decoder_layer.self_attn.k_norm", |
| "kind": "normalization", |
| "path_pattern": "model.language_model.layers.{i}.self_attn.k_norm" |
| }, |
| { |
| "attributes": { |
| "norm_type": "rms" |
| }, |
| "class_name": "Gemma4UnifiedRMSNorm", |
| "id": "decoder_layer.self_attn.v_norm", |
| "kind": "normalization", |
| "path_pattern": "model.language_model.layers.{i}.self_attn.v_norm" |
| }, |
| { |
| "attributes": { |
| "in_features": "config.hidden_size", |
| "out_features": 1024 |
| }, |
| "class_name": "Linear", |
| "id": "decoder_layer.self_attn.k_proj", |
| "kind": "projection", |
| "path_pattern": "model.language_model.layers.{i}.self_attn.k_proj" |
| }, |
| { |
| "attributes": { |
| "in_features": "config.hidden_size", |
| "out_features": 1024 |
| }, |
| "class_name": "Linear", |
| "id": "decoder_layer.self_attn.v_proj", |
| "kind": "projection", |
| "path_pattern": "model.language_model.layers.{i}.self_attn.v_proj" |
| }, |
| { |
| "attributes": { |
| "in_features": 2048, |
| "out_features": "config.hidden_size" |
| }, |
| "class_name": "Linear", |
| "id": "decoder_layer.self_attn.o_proj", |
| "kind": "projection", |
| "path_pattern": "model.language_model.layers.{i}.self_attn.o_proj" |
| }, |
| { |
| "attributes": { |
| "hidden_size": 2304, |
| "intermediate_size": 9216 |
| }, |
| "children": [ |
| "decoder_layer.mlp.gate_proj", |
| "decoder_layer.mlp.up_proj", |
| "decoder_layer.mlp.down_proj" |
| ], |
| "class_name": "Gemma4UnifiedTextMLP", |
| "id": "decoder_layer.mlp", |
| "kind": "feed_forward", |
| "path_pattern": "model.language_model.layers.{i}.mlp" |
| }, |
| { |
| "attributes": { |
| "in_features": "config.hidden_size", |
| "out_features": "config.intermediate_size" |
| }, |
| "class_name": "Linear", |
| "id": "decoder_layer.mlp.gate_proj", |
| "kind": "projection", |
| "path_pattern": "model.language_model.layers.{i}.mlp.gate_proj" |
| }, |
| { |
| "attributes": { |
| "in_features": "config.hidden_size", |
| "out_features": "config.intermediate_size" |
| }, |
| "class_name": "Linear", |
| "id": "decoder_layer.mlp.up_proj", |
| "kind": "projection", |
| "path_pattern": "model.language_model.layers.{i}.mlp.up_proj" |
| }, |
| { |
| "attributes": { |
| "in_features": "config.intermediate_size", |
| "out_features": "config.hidden_size" |
| }, |
| "class_name": "Linear", |
| "id": "decoder_layer.mlp.down_proj", |
| "kind": "projection", |
| "path_pattern": "model.language_model.layers.{i}.mlp.down_proj" |
| }, |
| { |
| "attributes": { |
| "norm_type": "rms" |
| }, |
| "class_name": "Gemma4UnifiedRMSNorm", |
| "id": "decoder_layer.input_layernorm", |
| "kind": "normalization", |
| "path_pattern": "model.language_model.layers.{i}.input_layernorm" |
| }, |
| { |
| "attributes": { |
| "norm_type": "rms" |
| }, |
| "class_name": "Gemma4UnifiedRMSNorm", |
| "id": "decoder_layer.post_attention_layernorm", |
| "kind": "normalization", |
| "path_pattern": "model.language_model.layers.{i}.post_attention_layernorm" |
| }, |
| { |
| "attributes": { |
| "norm_type": "rms" |
| }, |
| "class_name": "Gemma4UnifiedRMSNorm", |
| "id": "decoder_layer.pre_feedforward_layernorm", |
| "kind": "normalization", |
| "path_pattern": "model.language_model.layers.{i}.pre_feedforward_layernorm" |
| }, |
| { |
| "attributes": { |
| "norm_type": "rms" |
| }, |
| "class_name": "Gemma4UnifiedRMSNorm", |
| "id": "decoder_layer.post_feedforward_layernorm", |
| "kind": "normalization", |
| "path_pattern": "model.language_model.layers.{i}.post_feedforward_layernorm" |
| } |
| ] |
| } |
|
|
