| // 03 — everything in 01+02, plus modular inheritance: `extends` + `patches`. |
| // (structure/capabilities trimmed here — see 01/02.) Standalone models have extends=null and no patches. |
| { |
| "schema_version": "architecture-template-v0", |
| "model_type": "mistral", |
|
|
| // ▼▼▼ the tier-03 addition ▼▼▼ |
| "extends": "llama", // dominant parent model_type (from modular_mistral.py) |
| "patches": [ // per-class changes vs the parent; trivial (unchanged) classes are omitted |
| { "relation": "inherits", "target_class": "MistralAttention", "component_kind": "attention", |
| "parent_class": "LlamaAttention", "overridden": { "methods": ["__init__", "forward"] } }, |
| { "relation": "inherits", "target_class": "MistralMLP", "component_kind": "feed_forward", |
| "parent_class": "LlamaMLP", "overridden": { "methods": ["__init__"] } }, |
| { "relation": "new", "target_class": "MistralForQuestionAnswering", "component_kind": null, |
| "parent_class": "MistralPreTrainedModel" } |
| // … relation: "inherits" | "new"; buckets: overridden / added / deleted, each { methods?, attrs? } |
| ], |
| // ▲▲▲ the diff_size metric is NOT here — it lives per-node in modular_graph.json (see bottom) ▲▲▲ |
|
|
| "architecture": { "view": "decoder", "family": "causal_lm", "attention_variant": "GQA", "positional": "rope" }, |
| "capabilities": { "attention_backends": ["eager", "sdpa", "flash_attention", "flex_attention"], |
| "kernels": { "RMSNorm": ["kernels-community/rmsnorm"] } }, |
| "config": { |
| "class_name": "MistralConfig", "module": "transformers.models.mistral.configuration_mistral", "model_type": "mistral", |
| "referenced_fields": { "num_hidden_layers": 32 } |
| }, |
| "components": [ |
| { "id": "model", "kind": "model", "class_name": "MistralModel", "path_pattern": "model", |
| "children": ["embed_tokens", "decoder_layers", "norm"] } |
| // … see 01 |
| ], |
| "templates": [ |
| { "id": "decoder_layer.self_attn", "kind": "attention", "class_name": "MistralAttention", |
| "path_pattern": "model.layers.{i}.self_attn", "attributes": { "variant": "GQA", "pattern": "sliding" } } |
| // … 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" } ], |
| "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" |
| } |
| } |
|
|
| // Companion library-wide file `modular_graph.json` (one for the whole library, not per model): |
| // { "schema_version": "modular-graph-v0", |
| // "roots": ["llama", "vit", "clip"], |
| // "nodes": { "mistral": { "extends": "llama", "children": ["mixtral"], "root": "llama", |
| // "depth": 1, "is_modular": true, "diff_size": 9 } } } |
| // → models sharing a `root` are the same lineage (align cleanly for comparison); diff_size = modularity metric. |
|
|
