pyvene.models.mllama.modelings_intervenable_mllama.mllama_type_to_dimension_mapping

mllama_type_to_dimension_mapping = {'global.attention_input': ('vision_config.hidden_size',), 'global.attention_output': ('vision_config.hidden_size',), 'global.attention_value_output': ('vision_config.hidden_size',), 'global.block_input': ('vision_config.hidden_size',), 'global.block_output': ('vision_config.hidden_size',), 'global.head_attention_value_output': ('vision_config.hidden_size/vision_config.num_attention_heads',), 'global.head_key_output': ('vision_config.hidden_size/vision_config.num_attention_heads',), 'global.head_query_output': ('vision_config.hidden_size/vision_config.num_attention_heads',), 'global.head_value_output': ('vision_config.hidden_size/vision_config.num_attention_heads',), 'global.key_output': ('vision_config.hidden_size',), 'global.mlp_activation': ('vision_config.intermediate_size',), 'global.mlp_input': ('vision_config.hidden_size',), 'global.mlp_output': ('vision_config.hidden_size',), 'global.n_head': ('vision_config.num_attention_heads',), 'global.n_kv_head': ('vision_config.num_key_value_heads',), 'global.query_output': ('vision_config.hidden_size',), 'global.value_output': ('vision_config.hidden_size',), 'language.attention_input': ('text_config.hidden_size',), 'language.attention_output': ('text_config.hidden_size',), 'language.attention_value_output': ('text_config.hidden_size',), 'language.block_input': ('text_config.hidden_size',), 'language.block_output': ('text_config.hidden_size',), 'language.head_attention_value_output': ('text_config.hidden_size/text_config.num_attention_heads',), 'language.head_key_output': ('text_config.hidden_size/text_config.num_attention_heads',), 'language.head_query_output': ('text_config.hidden_size/text_config.num_attention_heads',), 'language.head_value_output': ('text_config.hidden_size/text_config.num_attention_heads',), 'language.key_output': ('text_config.hidden_size',), 'language.mlp_activation': ('text_config.intermediate_size',), 'language.mlp_input': ('text_config.hidden_size',), 'language.mlp_output': ('text_config.hidden_size',), 'language.n_head': ('text_config.num_attention_heads',), 'language.n_kv_head': ('text_config.num_key_value_heads',), 'language.query_output': ('text_config.hidden_size',), 'language.value_output': ('text_config.hidden_size',), 'vision.attention_input': ('vision_config.hidden_size',), 'vision.attention_output': ('vision_config.hidden_size',), 'vision.attention_value_output': ('vision_config.hidden_size',), 'vision.block_input': ('vision_config.hidden_size',), 'vision.block_output': ('vision_config.hidden_size',), 'vision.head_attention_value_output': ('vision_config.hidden_size/vision_config.num_attention_heads',), 'vision.head_key_output': ('vision_config.hidden_size/vision_config.num_attention_heads',), 'vision.head_query_output': ('vision_config.hidden_size/vision_config.num_attention_heads',), 'vision.head_value_output': ('vision_config.hidden_size/vision_config.num_attention_heads',), 'vision.key_output': ('vision_config.hidden_size',), 'vision.mlp_activation': ('vision_config.intermediate_size',), 'vision.mlp_input': ('vision_config.hidden_size',), 'vision.mlp_output': ('vision_config.hidden_size',), 'vision.n_head': ('vision_config.num_attention_heads',), 'vision.n_kv_head': ('vision_config.num_key_value_heads',), 'vision.query_output': ('vision_config.hidden_size',), 'vision.value_output': ('vision_config.hidden_size',)}

mllama model with LM head