| { |
| "architecture": { |
| "family": "causal_lm", |
| "is_moe": false, |
| "layer_summary": { |
| "mamba": 32 |
| }, |
| "positional": "rope", |
| "tie_word_embeddings": false, |
| "view": "decoder" |
| }, |
| "capabilities": { |
| "kernels": { |
| "RMSNorm": [ |
| "kernels-community/liger-kernels", |
| "kernels-community/rmsnorm", |
| "kernels-community/mlx_rmsnorm" |
| ] |
| }, |
| "task_heads": [ |
| "causal_lm" |
| ], |
| "tensor_parallel": false |
| }, |
| "components": [ |
| { |
| "children": [ |
| "embed_tokens", |
| "decoder_layers", |
| "final_layernorm", |
| "rotary_emb" |
| ], |
| "class_name": "BambaModel", |
| "id": "model", |
| "kind": "model", |
| "path_pattern": "model" |
| }, |
| { |
| "attributes": { |
| "embedding_dim": "config.hidden_size", |
| "num_embeddings": "config.vocab_size", |
| "tied_lm_head": false |
| }, |
| "class_name": "Embedding", |
| "id": "embed_tokens", |
| "kind": "embedding", |
| "path_pattern": "model.embed_tokens" |
| }, |
| { |
| "attributes": { |
| "kernel": "RMSNorm", |
| "norm_type": "rms" |
| }, |
| "class_name": "BambaRMSNorm", |
| "id": "final_layernorm", |
| "kind": "normalization", |
| "path_pattern": "model.final_layernorm" |
| }, |
| { |
| "attributes": { |
| "rope_theta": 10000.0, |
| "scheme": "rope" |
| }, |
| "class_name": "BambaRotaryEmbedding", |
| "id": "rotary_emb", |
| "kind": "position", |
| "path_pattern": "model.rotary_emb" |
| } |
| ], |
| "config": { |
| "class_name": "BambaConfig", |
| "model_type": "bamba", |
| "module": "transformers.models.bamba.configuration_bamba", |
| "referenced_fields": { |
| "num_hidden_layers": 32 |
| }, |
| "salient_fields": { |
| "hidden_act": "silu", |
| "hidden_size": 4096, |
| "intermediate_size": 14336, |
| "is_encoder_decoder": false, |
| "max_position_embeddings": 262144, |
| "num_attention_heads": 32, |
| "num_hidden_layers": 32, |
| "num_key_value_heads": 8, |
| "tie_word_embeddings": false, |
| "vocab_size": 128000 |
| } |
| }, |
| "edges": [ |
| { |
| "kind": "data", |
| "source": "embed_tokens", |
| "target": "decoder_layers" |
| }, |
| { |
| "kind": "data", |
| "source": "decoder_layers", |
| "target": "final_layernorm" |
| }, |
| { |
| "kind": "data", |
| "source": "decoder_layer.feed_forward", |
| "target": "decoder_layer.input_layernorm" |
| }, |
| { |
| "kind": "data", |
| "source": "decoder_layer.input_layernorm", |
| "target": "decoder_layer.pre_ff_layernorm" |
| }, |
| { |
| "kind": "residual", |
| "source": "decoder_layer", |
| "target": "decoder_layer.feed_forward" |
| }, |
| { |
| "kind": "data", |
| "provenance": "intra_module", |
| "source": "decoder_layer.feed_forward", |
| "target": "decoder_layer.feed_forward.gate_proj" |
| }, |
| { |
| "kind": "data", |
| "provenance": "intra_module", |
| "source": "decoder_layer.feed_forward", |
| "target": "decoder_layer.feed_forward.up_proj" |
| }, |
| { |
| "kind": "data", |
| "provenance": "intra_module", |
| "source": "decoder_layer.feed_forward.gate_proj", |
| "target": "decoder_layer.feed_forward.down_proj" |
| }, |
| { |
| "kind": "data", |
| "provenance": "intra_module", |
| "source": "decoder_layer.feed_forward.up_proj", |
| "target": "decoder_layer.feed_forward.down_proj" |
| } |
| ], |
| "extends": "llama", |
| "model_type": "bamba", |
| "patches": [ |
| { |
| "added": { |
| "attrs": [ |
| "cu_seq_lens_k", |
| "cu_seq_lens_q", |
| "max_length_k", |
| "max_length_q", |
| "seq_idx" |
| ] |
| }, |
| "component_kind": "attention", |
| "parent_class": "TypedDict", |
| "relation": "new", |
| "target_class": "BambaFlashAttentionKwargs" |
| }, |
| { |
| "added": { |
| "methods": [ |
| "__init__", |
| "cuda_kernels_forward", |
| "forward", |
| "torch_forward" |
| ] |
| }, |
| "component_kind": null, |
| "parent_class": "Module", |
| "relation": "new", |
| "target_class": "BambaMixer" |
| }, |
| { |
| "component_kind": "transformer_block", |
| "overridden": { |
| "methods": [ |
| "__init__", |
| "forward" |
| ] |
| }, |
| "parent_class": "JambaAttentionDecoderLayer", |
| "parent_model": "jamba", |
| "relation": "inherits", |
| "target_class": "BambaDecoderLayer" |
| }, |
| { |
| "added": { |
| "attrs": [ |
| "_can_record_outputs", |
| "_is_stateful", |
| "_no_split_modules", |
| "_skip_keys_device_placement", |
| "_supports_flash_attn", |
| "_supports_sdpa", |
| "base_model_prefix", |
| "config", |
| "supports_gradient_checkpointing" |
| ], |
| "methods": [ |
| "_init_weights" |
| ] |
| }, |
| "component_kind": "model", |
| "parent_class": "PreTrainedModel", |
| "relation": "new", |
| "target_class": "BambaPreTrainedModel" |
| }, |
| { |
| "added": { |
| "methods": [ |
| "__init__", |
| "_update_mamba_mask", |
| "forward" |
| ] |
| }, |
| "component_kind": "model", |
| "parent_class": "BambaPreTrainedModel", |
| "relation": "new", |
| "target_class": "BambaModel" |
| }, |
| { |
| "added": { |
| "methods": [ |
| "prepare_inputs_for_generation" |
| ] |
| }, |
| "component_kind": "lm_head", |
| "overridden": { |
| "methods": [ |
| "__init__", |
| "forward" |
| ] |
| }, |
| "parent_class": "LlamaForCausalLM", |
| "relation": "inherits", |
| "target_class": "BambaForCausalLM" |
| } |
| ], |
| "provenance": { |
| "config_class": "BambaConfig", |
| "config_module": "transformers.models.bamba.configuration_bamba", |
| "model_class": "BambaModel", |
| "model_module": "transformers.models.bamba.modeling_bamba" |
| }, |
| "repeats": [ |
| { |
| "body": "decoder_layer", |
| "container_path_pattern": "model.layers", |
| "count": 32, |
| "count_expr": "config.num_hidden_layers", |
| "count_source": "config", |
| "id": "decoder_layers", |
| "index_symbol": "i", |
| "item_path_pattern": "model.layers.{i}", |
| "kind": "symbolic_repeat", |
| "provenance": { |
| "class_name": "BambaDecoderLayer", |
| "container_path_pattern": "model.layers", |
| "item_path_pattern": "model.layers.{i}", |
| "source": "module_tree_repeat_collapse" |
| }, |
| "repeated_class_name": "BambaDecoderLayer" |
| } |
| ], |
| "schema_version": "architecture-template-v0", |
| "templates": [ |
| { |
| "children": [ |
| "decoder_layer.feed_forward", |
| "decoder_layer.input_layernorm", |
| "decoder_layer.pre_ff_layernorm", |
| "decoder_layer.mamba" |
| ], |
| "class_name": "BambaDecoderLayer", |
| "id": "decoder_layer", |
| "kind": "transformer_block", |
| "path_pattern": "model.layers.{i}" |
| }, |
| { |
| "attributes": { |
| "activation": "silu", |
| "hidden_size": 4096, |
| "intermediate_size": 14336 |
| }, |
| "children": [ |
| "decoder_layer.feed_forward.gate_proj", |
| "decoder_layer.feed_forward.up_proj", |
| "decoder_layer.feed_forward.down_proj" |
| ], |
| "class_name": "BambaMLP", |
| "id": "decoder_layer.feed_forward", |
| "kind": "feed_forward", |
| "path_pattern": "model.layers.{i}.feed_forward" |
| }, |
| { |
| "attributes": { |
| "in_features": "config.hidden_size", |
| "out_features": "config.intermediate_size" |
| }, |
| "class_name": "Linear", |
| "id": "decoder_layer.feed_forward.gate_proj", |
| "kind": "projection", |
| "path_pattern": "model.layers.{i}.feed_forward.gate_proj" |
| }, |
| { |
| "attributes": { |
| "in_features": "config.hidden_size", |
| "out_features": "config.intermediate_size" |
| }, |
| "class_name": "Linear", |
| "id": "decoder_layer.feed_forward.up_proj", |
| "kind": "projection", |
| "path_pattern": "model.layers.{i}.feed_forward.up_proj" |
| }, |
| { |
| "attributes": { |
| "in_features": "config.intermediate_size", |
| "out_features": "config.hidden_size" |
| }, |
| "class_name": "Linear", |
| "id": "decoder_layer.feed_forward.down_proj", |
| "kind": "projection", |
| "path_pattern": "model.layers.{i}.feed_forward.down_proj" |
| }, |
| { |
| "attributes": { |
| "kernel": "RMSNorm", |
| "norm_type": "rms" |
| }, |
| "class_name": "BambaRMSNorm", |
| "id": "decoder_layer.input_layernorm", |
| "kind": "normalization", |
| "path_pattern": "model.layers.{i}.input_layernorm" |
| }, |
| { |
| "attributes": { |
| "kernel": "RMSNorm", |
| "norm_type": "rms" |
| }, |
| "class_name": "BambaRMSNorm", |
| "id": "decoder_layer.pre_ff_layernorm", |
| "kind": "normalization", |
| "path_pattern": "model.layers.{i}.pre_ff_layernorm" |
| }, |
| { |
| "children": [ |
| "decoder_layer.mamba.conv1d", |
| "decoder_layer.mamba.norm" |
| ], |
| "class_name": "BambaMixer", |
| "id": "decoder_layer.mamba", |
| "kind": "module", |
| "path_pattern": "model.layers.{i}.mamba" |
| }, |
| { |
| "class_name": "Conv1d", |
| "id": "decoder_layer.mamba.conv1d", |
| "kind": "convolution", |
| "path_pattern": "model.layers.{i}.mamba.conv1d" |
| }, |
| { |
| "attributes": { |
| "norm_type": "rms" |
| }, |
| "class_name": "BambaRMSNormGated", |
| "id": "decoder_layer.mamba.norm", |
| "kind": "normalization", |
| "path_pattern": "model.layers.{i}.mamba.norm" |
| } |
| ] |
| } |
|
|
