Case Study: Per-Layer Merge Granularity

Ticket: GH-452 Module: aprender::online::per_layer_merge

Overview

Provides fine-grained per-tensor merge strategy overrides via YAML configuration. Different layers (attention, MLP, embeddings) can use different strategies and weights.

Key Components

• LayerMergeConfig — Ordered rules with default fallback
• LayerRule — Pattern + strategy + optional weights/scale
• MergeYamlConfig — Top-level YAML config (models, output, layers)
• match_layer_rule — First-match pattern lookup (substring + * wildcards)
• parse_merge_yaml — State-machine YAML parser (no serde_yaml dependency)
• validate_merge_config — Model count, strategy names, weight finiteness

Run

cargo run --example per_layer_merge

Falsification Tests