| // 02 — everything in 01, plus `capabilities`: what the model can do / run with. |
| // (structure/dataflow trimmed here to keep the focus on the new block — see 01 for the full diagram.) |
| { |
| "schema_version": "architecture-template-v0", |
| "model_type": "mistral", |
| "architecture": { |
| "view": "decoder", "family": "causal_lm", "attention_variant": "GQA", |
| "positional": "rope", "is_moe": false, "sliding_window": 4096, "tie_word_embeddings": false |
| }, |
|
|
| // ▼▼▼ the tier-02 addition ▼▼▼ |
| "capabilities": { |
| "attention_backends": ["eager", "sdpa", "flash_attention", "flex_attention"], // "can run with", not "installed" |
| "attention_patterns": ["sliding"], // distinct per-layer mask kinds |
| "attention_schedule": null, // raw config.layer_types when non-uniform, else null |
| "task_heads": ["causal_lm", "question_answering", "sequence_classification", "token_classification"], |
| "tensor_parallel": true, // → projection nodes carry attributes.tp (colwise/rowwise) |
| "kernels": { // kernelizable layers → compatible Hub kernel repos |
| "RMSNorm": ["kernels-community/liger-kernels", "kernels-community/rmsnorm", "kernels-community/mlx_rmsnorm"] |
| } |
| }, |
| // ▲▲▲ nodes point in via attributes.kernel: "RMSNorm" ▲▲▲ |
|
|
| "config": { |
| "class_name": "MistralConfig", "module": "transformers.models.mistral.configuration_mistral", "model_type": "mistral", |
| "referenced_fields": { "num_hidden_layers": 32 }, |
| "salient_fields": { "hidden_size": 4096, "intermediate_size": 14336, "num_key_value_heads": 8, "head_dim": 128 } |
| }, |
| "components": [ |
| { "id": "model", "kind": "model", "class_name": "MistralModel", "path_pattern": "model", |
| "children": ["embed_tokens", "decoder_layers", "norm"] }, |
| { "id": "norm", "kind": "normalization", "class_name": "MistralRMSNorm", "path_pattern": "model.norm", |
| "attributes": { "norm_type": "rms", "kernel": "RMSNorm" } } // ← joins to capabilities.kernels["RMSNorm"] |
| // … embed_tokens, rotary_emb (see 01) |
| ], |
| "templates": [ |
| { "id": "decoder_layer.self_attn", "kind": "attention", "class_name": "MistralAttention", |
| "path_pattern": "model.layers.{i}.self_attn", |
| "attributes": { "variant": "GQA", "n_heads": 32, "n_kv_heads": 8, "head_dim": 128, "pattern": "sliding" } } |
| // … decoder_layer + its projection / norm / mlp children (see 01) |
| ], |
| "repeats": [ |
| { "id": "decoder_layers", "kind": "symbolic_repeat", "body": "decoder_layer", |
| "count_expr": "config.num_hidden_layers", "count": 32, "count_source": "config", "index_symbol": "i", |
| "container_path_pattern": "model.layers", "item_path_pattern": "model.layers.{i}", |
| "repeated_class_name": "MistralDecoderLayer" } |
| ], |
| "edges": [ |
| { "source": "embed_tokens", "target": "decoder_layers", "kind": "data" }, |
| { "source": "decoder_layers", "target": "norm", "kind": "data" } |
| ], |
| "dataflow": { |
| "source": "observed_forward_meta", |
| "input": { "name": "input_ids", "shape": ["B", "S"] }, |
| "output": { "shape": ["B", "S", "config.hidden_size"] }, |
| "shapes": { "embed_tokens": { "in": ["B", "S"], "out": ["B", "S", "config.hidden_size"] } } |
| }, |
| "provenance": { |
| "config_class": "MistralConfig", "config_module": "transformers.models.mistral.configuration_mistral", |
| "model_class": "MistralModel", "model_module": "transformers.models.mistral.modeling_mistral" |
| } |
| } |
|
|
