| { |
| "architecture": { |
| "attention_variant": "GQA", |
| "family": "image_text_to_text", |
| "is_moe": false, |
| "mixer": "attention", |
| "positional": "rope", |
| "sliding_window": 4096, |
| "tie_word_embeddings": true, |
| "towers": [ |
| { |
| "attention_variant": "GQA", |
| "config_class": "Gemma3TextConfig", |
| "hidden_size": 2304, |
| "name": "text", |
| "num_attention_heads": 8, |
| "num_hidden_layers": 26, |
| "positional": "rope" |
| }, |
| { |
| "attention_variant": "MHA", |
| "config_class": "SiglipVisionConfig", |
| "hidden_size": 768, |
| "name": "vision", |
| "num_attention_heads": 12, |
| "num_hidden_layers": 12 |
| } |
| ], |
| "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" |
| ], |
| "task_heads": [ |
| "causal_lm", |
| "image_text_to_text", |
| "multimodal_lm", |
| "pretraining", |
| "sequence_classification" |
| ], |
| "tensor_parallel": true |
| }, |
| "components": [ |
| { |
| "children": [ |
| "vision_tower", |
| "multi_modal_projector", |
| "language_model" |
| ], |
| "class_name": "Gemma3Model", |
| "id": "model", |
| "kind": "model", |
| "path_pattern": "model" |
| }, |
| { |
| "children": [ |
| "vision_tower.embeddings", |
| "vision_tower.encoder", |
| "vision_tower.post_layernorm", |
| "vision_tower.head" |
| ], |
| "class_name": "SiglipVisionModel", |
| "id": "vision_tower", |
| "kind": "model", |
| "path_pattern": "model.vision_tower" |
| }, |
| { |
| "attributes": { |
| "embedding_dim": "config.hidden_size" |
| }, |
| "children": [ |
| "vision_tower.embeddings.patch_embedding", |
| "vision_tower.embeddings.position_embedding" |
| ], |
| "class_name": "SiglipVisionEmbeddings", |
| "id": "vision_tower.embeddings", |
| "kind": "embedding", |
| "path_pattern": "model.vision_tower.embeddings" |
| }, |
| { |
| "class_name": "Conv2d", |
| "id": "vision_tower.embeddings.patch_embedding", |
| "kind": "convolution", |
| "path_pattern": "model.vision_tower.embeddings.patch_embedding" |
| }, |
| { |
| "attributes": {}, |
| "class_name": "Embedding", |
| "id": "vision_tower.embeddings.position_embedding", |
| "kind": "position", |
| "path_pattern": "model.vision_tower.embeddings.position_embedding" |
| }, |
| { |
| "children": [ |
| "encoder_layers" |
| ], |
| "class_name": "SiglipEncoder", |
| "id": "vision_tower.encoder", |
| "kind": "encoder", |
| "path_pattern": "model.vision_tower.encoder" |
| }, |
| { |
| "attributes": { |
| "norm_type": "layer" |
| }, |
| "class_name": "LayerNorm", |
| "id": "vision_tower.post_layernorm", |
| "kind": "normalization", |
| "path_pattern": "model.vision_tower.post_layernorm" |
| }, |
| { |
| "attributes": { |
| "head_dim": 64, |
| "n_heads": 12, |
| "n_kv_heads": 12, |
| "rope": false, |
| "variant": "MHA" |
| }, |
| "children": [ |
| "vision_tower.head.attention", |
| "vision_tower.head.layernorm", |
| "vision_tower.head.mlp" |
| ], |
| "class_name": "SiglipMultiheadAttentionPoolingHead", |
| "id": "vision_tower.head", |
| "kind": "attention", |
| "path_pattern": "model.vision_tower.head" |
| }, |
| { |
| "attributes": { |
| "head_dim": 64, |
| "n_heads": 12, |
| "n_kv_heads": 12, |
| "rope": false, |
| "variant": "MHA" |
| }, |
| "children": [ |
| "vision_tower.head.attention.out_proj" |
| ], |
| "class_name": "MultiheadAttention", |
| "id": "vision_tower.head.attention", |
| "kind": "attention", |
| "path_pattern": "model.vision_tower.head.attention" |
| }, |
| { |
| "class_name": "NonDynamicallyQuantizableLinear", |
| "id": "vision_tower.head.attention.out_proj", |
| "kind": "projection", |
| "path_pattern": "model.vision_tower.head.attention.out_proj" |
| }, |
| { |
| "attributes": { |
| "norm_type": "layer" |
| }, |
| "class_name": "LayerNorm", |
| "id": "vision_tower.head.layernorm", |
| "kind": "normalization", |
| "path_pattern": "model.vision_tower.head.layernorm" |
| }, |
| { |
| "attributes": { |
| "activation": "gelu_pytorch_tanh", |
| "hidden_size": 768, |
| "intermediate_size": 3072 |
| }, |
| "children": [ |
| "vision_tower.head.mlp.activation_fn", |
| "vision_tower.head.mlp.fc1", |
| "vision_tower.head.mlp.fc2" |
| ], |
| "class_name": "SiglipMLP", |
| "id": "vision_tower.head.mlp", |
| "kind": "feed_forward", |
| "path_pattern": "model.vision_tower.head.mlp" |
| }, |
| { |
| "attributes": { |
| "activation": "gelu_pytorch_tanh", |
| "hidden_size": 768, |
| "intermediate_size": 3072 |
| }, |
| "class_name": "GELUTanh", |
| "id": "vision_tower.head.mlp.activation_fn", |
| "kind": "feed_forward", |
| "path_pattern": "model.vision_tower.head.mlp.activation_fn" |
| }, |
| { |
| "attributes": { |
| "in_features": "config.hidden_size", |
| "out_features": "config.intermediate_size" |
| }, |
| "class_name": "Linear", |
| "id": "vision_tower.head.mlp.fc1", |
| "kind": "projection", |
| "path_pattern": "model.vision_tower.head.mlp.fc1" |
| }, |
| { |
| "attributes": { |
| "in_features": "config.intermediate_size", |
| "out_features": "config.hidden_size" |
| }, |
| "class_name": "Linear", |
| "id": "vision_tower.head.mlp.fc2", |
| "kind": "projection", |
| "path_pattern": "model.vision_tower.head.mlp.fc2" |
| }, |
| { |
| "children": [ |
| "multi_modal_projector.mm_soft_emb_norm" |
| ], |
| "class_name": "Gemma3MultiModalProjector", |
| "id": "multi_modal_projector", |
| "kind": "module", |
| "path_pattern": "model.multi_modal_projector" |
| }, |
| { |
| "attributes": { |
| "norm_type": "rms" |
| }, |
| "class_name": "Gemma3RMSNorm", |
| "id": "multi_modal_projector.mm_soft_emb_norm", |
| "kind": "normalization", |
| "path_pattern": "model.multi_modal_projector.mm_soft_emb_norm" |
| }, |
| { |
| "children": [ |
| "language_model.embed_tokens", |
| "decoder_layers", |
| "language_model.norm", |
| "language_model.rotary_emb" |
| ], |
| "class_name": "Gemma3TextModel", |
| "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": "Gemma3TextScaledWordEmbedding", |
| "id": "language_model.embed_tokens", |
| "kind": "embedding", |
| "path_pattern": "model.language_model.embed_tokens" |
| }, |
| { |
| "attributes": { |
| "norm_type": "rms" |
| }, |
| "class_name": "Gemma3RMSNorm", |
| "id": "language_model.norm", |
| "kind": "normalization", |
| "path_pattern": "model.language_model.norm" |
| }, |
| { |
| "attributes": { |
| "head_dim": "config.head_dim", |
| "scheme": "rope" |
| }, |
| "class_name": "Gemma3RotaryEmbedding", |
| "id": "language_model.rotary_emb", |
| "kind": "position", |
| "path_pattern": "model.language_model.rotary_emb" |
| } |
| ], |
| "config": { |
| "class_name": "Gemma3Config", |
| "model_type": "gemma3", |
| "module": "transformers.models.gemma3.configuration_gemma3", |
| "referenced_fields": {}, |
| "salient_fields": { |
| "is_encoder_decoder": false, |
| "tie_word_embeddings": true |
| } |
| }, |
| "dataflow": { |
| "input": { |
| "name": "input_ids", |
| "shape": [ |
| "B", |
| "S" |
| ] |
| }, |
| "output": { |
| "shape": [ |
| "B", |
| "S", |
| 2304 |
| ] |
| }, |
| "shapes": { |
| "decoder_layer.input_layernorm": { |
| "in": [ |
| "B", |
| "S", |
| 2304 |
| ], |
| "out": [ |
| "B", |
| "S", |
| 2304 |
| ] |
| }, |
| "decoder_layer.mlp": { |
| "in": [ |
| "B", |
| "S", |
| 2304 |
| ], |
| "out": [ |
| "B", |
| "S", |
| 2304 |
| ] |
| }, |
| "decoder_layer.post_attention_layernorm": { |
| "in": [ |
| "B", |
| "S", |
| 2304 |
| ], |
| "out": [ |
| "B", |
| "S", |
| 2304 |
| ] |
| }, |
| "decoder_layer.post_feedforward_layernorm": { |
| "in": [ |
| "B", |
| "S", |
| 2304 |
| ], |
| "out": [ |
| "B", |
| "S", |
| 2304 |
| ] |
| }, |
| "decoder_layer.pre_feedforward_layernorm": { |
| "in": [ |
| "B", |
| "S", |
| 2304 |
| ], |
| "out": [ |
| "B", |
| "S", |
| 2304 |
| ] |
| }, |
| "decoder_layer.self_attn": { |
| "in": null, |
| "out": [ |
| "B", |
| "S", |
| 2304 |
| ] |
| }, |
| "language_model": { |
| "in": null, |
| "out": [ |
| "B", |
| "S", |
| 2304 |
| ] |
| } |
| }, |
| "source": "observed_forward_meta" |
| }, |
| "edges": [ |
| { |
| "kind": "data", |
| "source": "vision_tower.embeddings", |
| "target": "vision_tower.encoder" |
| }, |
| { |
| "kind": "data", |
| "source": "vision_tower.encoder", |
| "target": "vision_tower.post_layernorm" |
| }, |
| { |
| "kind": "data", |
| "source": "vision_tower.post_layernorm", |
| "target": "vision_tower.head" |
| }, |
| { |
| "kind": "data", |
| "source": "encoder_layer.layer_norm1", |
| "target": "encoder_layer.self_attn" |
| }, |
| { |
| "kind": "data", |
| "source": "encoder_layer.self_attn", |
| "target": "encoder_layer.layer_norm2" |
| }, |
| { |
| "kind": "data", |
| "source": "encoder_layer.layer_norm2", |
| "target": "encoder_layer.mlp" |
| }, |
| { |
| "kind": "data", |
| "source": "vision_tower.head.attention", |
| "target": "vision_tower.head.layernorm" |
| }, |
| { |
| "kind": "data", |
| "source": "vision_tower.head.layernorm", |
| "target": "vision_tower.head.mlp" |
| }, |
| { |
| "kind": "data", |
| "source": "language_model.embed_tokens", |
| "target": "decoder_layers" |
| }, |
| { |
| "kind": "data", |
| "source": "decoder_layers", |
| "target": "language_model.norm" |
| }, |
| { |
| "kind": "mask", |
| "source": "input:attention_mask", |
| "target": "encoder_layer.self_attn" |
| }, |
| { |
| "kind": "mask", |
| "source": "input:attention_mask", |
| "target": "vision_tower.head" |
| }, |
| { |
| "kind": "mask", |
| "source": "input:attention_mask", |
| "target": "vision_tower.head.attention" |
| }, |
| { |
| "kind": "mask", |
| "source": "input:attention_mask", |
| "target": "decoder_layer.self_attn" |
| }, |
| { |
| "kind": "residual", |
| "source": "encoder_layer", |
| "target": "encoder_layer.self_attn" |
| }, |
| { |
| "kind": "residual", |
| "source": "encoder_layer", |
| "target": "encoder_layer.mlp" |
| }, |
| { |
| "kind": "residual", |
| "source": "decoder_layer", |
| "target": "decoder_layer.self_attn" |
| }, |
| { |
| "kind": "residual", |
| "source": "decoder_layer", |
| "target": "decoder_layer.mlp" |
| }, |
| { |
| "kind": "position", |
| "source": "vision_tower.embeddings.position_embedding", |
| "target": "vision_tower.embeddings" |
| }, |
| { |
| "kind": "position", |
| "source": "language_model.rotary_emb", |
| "target": "decoder_layer.self_attn" |
| }, |
| { |
| "kind": "data", |
| "provenance": "intra_module", |
| "source": "vision_tower.head.attention", |
| "target": "vision_tower.head.attention.out_proj" |
| }, |
| { |
| "kind": "data", |
| "provenance": "intra_module", |
| "source": "vision_tower.head.mlp", |
| "target": "vision_tower.head.mlp.fc1" |
| }, |
| { |
| "kind": "data", |
| "provenance": "intra_module", |
| "source": "vision_tower.head.mlp.fc1", |
| "target": "vision_tower.head.mlp.fc2" |
| }, |
| { |
| "kind": "data", |
| "provenance": "intra_module", |
| "source": "encoder_layer.self_attn", |
| "target": "encoder_layer.self_attn.k_proj" |
| }, |
| { |
| "kind": "data", |
| "provenance": "intra_module", |
| "source": "encoder_layer.self_attn", |
| "target": "encoder_layer.self_attn.v_proj" |
| }, |
| { |
| "kind": "data", |
| "provenance": "intra_module", |
| "source": "encoder_layer.self_attn", |
| "target": "encoder_layer.self_attn.q_proj" |
| }, |
| { |
| "kind": "data", |
| "provenance": "intra_module", |
| "source": "encoder_layer.self_attn.k_proj", |
| "target": "encoder_layer.self_attn.out_proj" |
| }, |
| { |
| "kind": "data", |
| "provenance": "intra_module", |
| "source": "encoder_layer.self_attn.v_proj", |
| "target": "encoder_layer.self_attn.out_proj" |
| }, |
| { |
| "kind": "data", |
| "provenance": "intra_module", |
| "source": "encoder_layer.self_attn.q_proj", |
| "target": "encoder_layer.self_attn.out_proj" |
| }, |
| { |
| "kind": "data", |
| "provenance": "intra_module", |
| "source": "encoder_layer.mlp", |
| "target": "encoder_layer.mlp.fc1" |
| }, |
| { |
| "kind": "data", |
| "provenance": "intra_module", |
| "source": "encoder_layer.mlp.fc1", |
| "target": "encoder_layer.mlp.fc2" |
| }, |
| { |
| "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" |
| }, |
| { |
| "kind": "data", |
| "provenance": "observed_forward", |
| "source": "decoder_layer.input_layernorm", |
| "target": "decoder_layer.self_attn" |
| }, |
| { |
| "kind": "data", |
| "provenance": "observed_forward", |
| "source": "decoder_layer.self_attn", |
| "target": "decoder_layer.post_attention_layernorm" |
| }, |
| { |
| "kind": "data", |
| "provenance": "observed_forward", |
| "source": "decoder_layer.post_attention_layernorm", |
| "target": "decoder_layer.pre_feedforward_layernorm" |
| }, |
| { |
| "kind": "data", |
| "provenance": "observed_forward", |
| "source": "decoder_layer.pre_feedforward_layernorm", |
| "target": "decoder_layer.mlp" |
| }, |
| { |
| "kind": "data", |
| "provenance": "observed_forward", |
| "source": "decoder_layer.mlp", |
| "target": "decoder_layer.post_feedforward_layernorm" |
| } |
| ], |
| "extends": "gemma2", |
| "model_type": "gemma3", |
| "patches": [ |
| { |
| "added": { |
| "attrs": [ |
| "attn_logit_softcapping", |
| "base_model_tp_plan", |
| "default_theta", |
| "final_logit_softcapping", |
| "layer_types", |
| "max_position_embeddings", |
| "model_type", |
| "rope_parameters", |
| "use_bidirectional_attention", |
| "vocab_size" |
| ], |
| "methods": [ |
| "__post_init__", |
| "convert_rope_params_to_dict" |
| ] |
| }, |
| "component_kind": "config", |
| "parent_class": "Gemma2Config", |
| "relation": "new", |
| "target_class": "Gemma3TextConfig" |
| }, |
| { |
| "added": { |
| "attrs": [ |
| "attribute_map", |
| "boi_token_index", |
| "eoi_token_index", |
| "image_token_index", |
| "initializer_range", |
| "mm_tokens_per_image", |
| "model_type", |
| "sub_configs", |
| "text_config", |
| "tie_word_embeddings", |
| "vision_config" |
| ], |
| "methods": [ |
| "__post_init__" |
| ] |
| }, |
| "component_kind": "config", |
| "parent_class": "PreTrainedConfig", |
| "relation": "new", |
| "target_class": "Gemma3Config" |
| }, |
| { |
| "added": { |
| "methods": [ |
| "__init__", |
| "forward" |
| ] |
| }, |
| "component_kind": "embedding", |
| "parent_class": "Embedding", |
| "relation": "new", |
| "target_class": "Gemma3TextScaledWordEmbedding" |
| }, |
| { |
| "component_kind": "feed_forward", |
| "overridden": { |
| "methods": [ |
| "__init__" |
| ] |
| }, |
| "parent_class": "Gemma2MLP", |
| "relation": "inherits", |
| "target_class": "Gemma3MLP" |
| }, |
| { |
| "component_kind": "normalization", |
| "overridden": { |
| "methods": [ |
| "__init__" |
| ] |
| }, |
| "parent_class": "Gemma2RMSNorm", |
| "relation": "inherits", |
| "target_class": "Gemma3RMSNorm" |
| }, |
| { |
| "component_kind": "position", |
| "overridden": { |
| "methods": [ |
| "__init__", |
| "compute_default_rope_parameters", |
| "forward" |
| ] |
| }, |
| "parent_class": "Gemma2RotaryEmbedding", |
| "relation": "inherits", |
| "target_class": "Gemma3RotaryEmbedding" |
| }, |
| { |
| "component_kind": "attention", |
| "overridden": { |
| "methods": [ |
| "__init__", |
| "forward" |
| ] |
| }, |
| "parent_class": "Gemma2Attention", |
| "relation": "inherits", |
| "target_class": "Gemma3Attention" |
| }, |
| { |
| "added": { |
| "methods": [ |
| "__init__", |
| "forward" |
| ] |
| }, |
| "component_kind": "transformer_block", |
| "parent_class": "GradientCheckpointingLayer", |
| "relation": "new", |
| "target_class": "Gemma3DecoderLayer" |
| }, |
| { |
| "added": { |
| "attrs": [ |
| "input_modalities" |
| ] |
| }, |
| "component_kind": "model", |
| "overridden": { |
| "attrs": [ |
| "_no_split_modules", |
| "base_model_prefix" |
| ], |
| "methods": [ |
| "_init_weights" |
| ] |
| }, |
| "parent_class": "Gemma2PreTrainedModel", |
| "relation": "inherits", |
| "target_class": "Gemma3PreTrainedModel" |
| }, |
| { |
| "added": { |
| "attrs": [ |
| "config", |
| "input_modalities" |
| ] |
| }, |
| "component_kind": "model", |
| "overridden": { |
| "methods": [ |
| "__init__", |
| "forward" |
| ] |
| }, |
| "parent_class": "Gemma2Model", |
| "relation": "inherits", |
| "target_class": "Gemma3TextModel" |
| }, |
| { |
| "added": { |
| "attrs": [ |
| "config" |
| ] |
| }, |
| "component_kind": "lm_head", |
| "overridden": { |
| "methods": [ |
| "__init__" |
| ] |
| }, |
| "parent_class": "Gemma2ForCausalLM", |
| "relation": "inherits", |
| "target_class": "Gemma3ForCausalLM" |
| }, |
| { |
| "added": { |
| "methods": [ |
| "__init__", |
| "forward" |
| ] |
| }, |
| "component_kind": null, |
| "parent_class": "Module", |
| "relation": "new", |
| "target_class": "Gemma3MultiModalProjector" |
| }, |
| { |
| "component_kind": "model", |
| "overridden": { |
| "attrs": [ |
| "accepts_loss_kwargs" |
| ], |
| "methods": [ |
| "__init__", |
| "forward", |
| "get_image_features" |
| ] |
| }, |
| "parent_class": "PaliGemmaModel", |
| "parent_model": "paligemma", |
| "relation": "inherits", |
| "target_class": "Gemma3Model" |
| }, |
| { |
| "added": { |
| "attrs": [ |
| "accepts_loss_kwargs" |
| ] |
| }, |
| "component_kind": "lm_head", |
| "overridden": { |
| "methods": [ |
| "create_masks_for_generate", |
| "forward", |
| "prepare_inputs_for_generation" |
| ] |
| }, |
| "parent_class": "PaliGemmaForConditionalGeneration", |
| "parent_model": "paligemma", |
| "relation": "inherits", |
| "target_class": "Gemma3ForConditionalGeneration" |
| }, |
| { |
| "added": { |
| "attrs": [ |
| "config", |
| "input_modalities" |
| ] |
| }, |
| "component_kind": "head", |
| "parent_class": "GenericForSequenceClassification", |
| "relation": "new", |
| "target_class": "Gemma3TextForSequenceClassification" |
| }, |
| { |
| "added": { |
| "methods": [ |
| "forward" |
| ] |
| }, |
| "component_kind": "head", |
| "parent_class": "GenericForSequenceClassification", |
| "relation": "new", |
| "target_class": "Gemma3ForSequenceClassification" |
| } |
| ], |
| "provenance": { |
| "config_class": "Gemma3Config", |
| "config_module": "transformers.models.gemma3.configuration_gemma3", |
| "model_class": "Gemma3Model", |
| "model_module": "transformers.models.gemma3.modeling_gemma3" |
| }, |
| "repeats": [ |
| { |
| "body": "encoder_layer", |
| "container_path_pattern": "model.vision_tower.encoder.layers", |
| "count": 12, |
| "count_expr": "12", |
| "count_source": "module_tree", |
| "id": "encoder_layers", |
| "index_symbol": "i", |
| "item_path_pattern": "model.vision_tower.encoder.layers.{i}", |
| "kind": "symbolic_repeat", |
| "provenance": { |
| "class_name": "SiglipEncoderLayer", |
| "container_path_pattern": "model.vision_tower.encoder.layers", |
| "item_path_pattern": "model.vision_tower.encoder.layers.{i}", |
| "source": "module_tree_repeat_collapse" |
| }, |
| "repeated_class_name": "SiglipEncoderLayer" |
| }, |
| { |
| "body": "decoder_layer", |
| "container_path_pattern": "model.language_model.layers", |
| "count": 26, |
| "count_expr": "26", |
| "count_source": "module_tree", |
| "id": "decoder_layers", |
| "index_symbol": "i", |
| "item_path_pattern": "model.language_model.layers.{i}", |
| "kind": "symbolic_repeat", |
| "provenance": { |
| "class_name": "Gemma3DecoderLayer", |
| "container_path_pattern": "model.language_model.layers", |
| "item_path_pattern": "model.language_model.layers.{i}", |
| "source": "module_tree_repeat_collapse" |
| }, |
| "repeated_class_name": "Gemma3DecoderLayer" |
| } |
| ], |
| "schema_version": "architecture-template-v0", |
| "templates": [ |
| { |
| "children": [ |
| "encoder_layer.layer_norm1", |
| "encoder_layer.self_attn", |
| "encoder_layer.layer_norm2", |
| "encoder_layer.mlp" |
| ], |
| "class_name": "SiglipEncoderLayer", |
| "id": "encoder_layer", |
| "kind": "transformer_block", |
| "path_pattern": "model.vision_tower.encoder.layers.{i}" |
| }, |
| { |
| "attributes": { |
| "norm_type": "layer" |
| }, |
| "class_name": "LayerNorm", |
| "id": "encoder_layer.layer_norm1", |
| "kind": "normalization", |
| "path_pattern": "model.vision_tower.encoder.layers.{i}.layer_norm1" |
| }, |
| { |
| "attributes": { |
| "head_dim": 64, |
| "n_heads": 12, |
| "n_kv_heads": 12, |
| "rope": false, |
| "variant": "MHA" |
| }, |
| "children": [ |
| "encoder_layer.self_attn.k_proj", |
| "encoder_layer.self_attn.v_proj", |
| "encoder_layer.self_attn.q_proj", |
| "encoder_layer.self_attn.out_proj" |
| ], |
| "class_name": "SiglipAttention", |
| "id": "encoder_layer.self_attn", |
| "kind": "attention", |
| "path_pattern": "model.vision_tower.encoder.layers.{i}.self_attn" |
| }, |
| { |
| "attributes": { |
| "in_features": "config.hidden_size", |
| "out_features": "config.hidden_size" |
| }, |
| "class_name": "Linear", |
| "id": "encoder_layer.self_attn.k_proj", |
| "kind": "projection", |
| "path_pattern": "model.vision_tower.encoder.layers.{i}.self_attn.k_proj" |
| }, |
| { |
| "attributes": { |
| "in_features": "config.hidden_size", |
| "out_features": "config.hidden_size" |
| }, |
| "class_name": "Linear", |
| "id": "encoder_layer.self_attn.v_proj", |
| "kind": "projection", |
| "path_pattern": "model.vision_tower.encoder.layers.{i}.self_attn.v_proj" |
| }, |
| { |
| "attributes": { |
| "in_features": "config.hidden_size", |
| "out_features": "config.hidden_size" |
| }, |
| "class_name": "Linear", |
| "id": "encoder_layer.self_attn.q_proj", |
| "kind": "projection", |
| "path_pattern": "model.vision_tower.encoder.layers.{i}.self_attn.q_proj" |
| }, |
| { |
| "attributes": { |
| "in_features": "config.hidden_size", |
| "out_features": "config.hidden_size" |
| }, |
| "class_name": "Linear", |
| "id": "encoder_layer.self_attn.out_proj", |
| "kind": "projection", |
| "path_pattern": "model.vision_tower.encoder.layers.{i}.self_attn.out_proj" |
| }, |
| { |
| "attributes": { |
| "norm_type": "layer" |
| }, |
| "class_name": "LayerNorm", |
| "id": "encoder_layer.layer_norm2", |
| "kind": "normalization", |
| "path_pattern": "model.vision_tower.encoder.layers.{i}.layer_norm2" |
| }, |
| { |
| "attributes": { |
| "activation": "gelu_pytorch_tanh", |
| "hidden_size": 768, |
| "intermediate_size": 3072 |
| }, |
| "children": [ |
| "encoder_layer.mlp.activation_fn", |
| "encoder_layer.mlp.fc1", |
| "encoder_layer.mlp.fc2" |
| ], |
| "class_name": "SiglipMLP", |
| "id": "encoder_layer.mlp", |
| "kind": "feed_forward", |
| "path_pattern": "model.vision_tower.encoder.layers.{i}.mlp" |
| }, |
| { |
| "attributes": { |
| "activation": "gelu_pytorch_tanh", |
| "hidden_size": 768, |
| "intermediate_size": 3072 |
| }, |
| "class_name": "GELUTanh", |
| "id": "encoder_layer.mlp.activation_fn", |
| "kind": "feed_forward", |
| "path_pattern": "model.vision_tower.encoder.layers.{i}.mlp.activation_fn" |
| }, |
| { |
| "attributes": { |
| "in_features": "config.hidden_size", |
| "out_features": "config.intermediate_size" |
| }, |
| "class_name": "Linear", |
| "id": "encoder_layer.mlp.fc1", |
| "kind": "projection", |
| "path_pattern": "model.vision_tower.encoder.layers.{i}.mlp.fc1" |
| }, |
| { |
| "attributes": { |
| "in_features": "config.intermediate_size", |
| "out_features": "config.hidden_size" |
| }, |
| "class_name": "Linear", |
| "id": "encoder_layer.mlp.fc2", |
| "kind": "projection", |
| "path_pattern": "model.vision_tower.encoder.layers.{i}.mlp.fc2" |
| }, |
| { |
| "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": "Gemma3DecoderLayer", |
| "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": 4096, |
| "variant": "GQA" |
| }, |
| "children": [ |
| "decoder_layer.self_attn.q_proj", |
| "decoder_layer.self_attn.k_proj", |
| "decoder_layer.self_attn.v_proj", |
| "decoder_layer.self_attn.o_proj", |
| "decoder_layer.self_attn.q_norm", |
| "decoder_layer.self_attn.k_norm" |
| ], |
| "class_name": "Gemma3Attention", |
| "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": { |
| "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": { |
| "norm_type": "rms" |
| }, |
| "class_name": "Gemma3RMSNorm", |
| "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": "Gemma3RMSNorm", |
| "id": "decoder_layer.self_attn.k_norm", |
| "kind": "normalization", |
| "path_pattern": "model.language_model.layers.{i}.self_attn.k_norm" |
| }, |
| { |
| "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": "Gemma3MLP", |
| "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": "Gemma3RMSNorm", |
| "id": "decoder_layer.input_layernorm", |
| "kind": "normalization", |
| "path_pattern": "model.language_model.layers.{i}.input_layernorm" |
| }, |
| { |
| "attributes": { |
| "norm_type": "rms" |
| }, |
| "class_name": "Gemma3RMSNorm", |
| "id": "decoder_layer.post_attention_layernorm", |
| "kind": "normalization", |
| "path_pattern": "model.language_model.layers.{i}.post_attention_layernorm" |
| }, |
| { |
| "attributes": { |
| "norm_type": "rms" |
| }, |
| "class_name": "Gemma3RMSNorm", |
| "id": "decoder_layer.pre_feedforward_layernorm", |
| "kind": "normalization", |
| "path_pattern": "model.language_model.layers.{i}.pre_feedforward_layernorm" |
| }, |
| { |
| "attributes": { |
| "norm_type": "rms" |
| }, |
| "class_name": "Gemma3RMSNorm", |
| "id": "decoder_layer.post_feedforward_layernorm", |
| "kind": "normalization", |
| "path_pattern": "model.language_model.layers.{i}.post_feedforward_layernorm" |
| } |
| ] |
| } |
|
|
