Model Compatibility: Full Decision Tree and States

Complete compatibility decision tree for EMEP model pairs. Covers architecture, tokenizer, tensor shapes, and metadata checks down to COMPATIBLE, CONDITIONALLY_COMPATIBLE, and INCOMPATIBLE states.

Model compatibility is the gate through which every candidate merge must pass. This page defines every check dimension and the exact decision tree that produces one of three states: COMPATIBLE, CONDITIONALLY_COMPATIBLE, or INCOMPATIBLE.

Check Dimensions

The ModelCompatibilityAnalyzer evaluates every pair across these dimensions:

Dimension	What It Checks
Architecture family	Llama, Mistral, Qwen, Gemma, etc.
Model type	base, instruct, chat, code
Parameter count	exact match or declared tolerance
Layers	layer count and ordering
Hidden size	hidden dimension per layer
Intermediate size	MLP expansion dimension
Attention heads	count and grouping
KV heads	GQA/MQA head count
Embedding size	vocab embedding dimension
Vocabulary	tokenizer vocab size and overlap
Special tokens	BOS, EOS, PAD, UNK presence and IDs
Tokenizer family	BPE, SentencePiece, WordPiece
RoPE config	base, theta, scaling method
Context config	max context length and scaling
Parameter names	key naming conventions in state_dict
Tensor shapes	per-parameter shape equality
dtype	weight data type
Weight format	safetensors, bin, gguf, etc.
Revision	model commit or tag
Source	Hugging Face hub, mirror, local path
License	permissible merge license
Hash	SHA-256 of weight files

Compatibility Decision Tree

The following flowchart shows the full path from two input models to a final state.

CONDITIONALLY_COMPATIBLE Explained

A pair is CONDITIONALLY_COMPATIBLE when the core architecture matches but a non-fatal discrepancy exists. The MergeStrategy must handle the discrepancy explicitly.

Condition	Required Action
Different revision	Strategy must accept revision drift
RoPE mismatch	Strategy must interpolate or override RoPE config
Context mismatch	Strategy must clamp or scale context
Tokenizer partial match	Strategy must remap or retrain embeddings
dtype mismatch	Strategy must cast or quantize
Type mismatch (base vs instruct)	Strategy must handle task vector sign
Franken-Merge selected	Strategy must provide structural adapters

INCOMPATIBLE Explained

A pair is INCOMPATIBLE when a fatal mismatch blocks all standard strategies. Cross-family merges without franken-merge adapters, mismatched layer counts, and mismatched tensor shapes are all fatal. The MergeEngine rejects these pairs before strategy selection.

Cross-Links

• Architecture Compatibility for the architecture-only subset of this tree.
• Tokenizer Compatibility for the tokenizer branch in detail.
• Tensor Shape Validation for the tensor shape and dtype branch.
• Merge Engine for how the compatibility result feeds into the merge pipeline.