MergeEngine: Pipeline, Interface, and Failure Modes

MergeEngine specification for EMEP. Defines the full merge pipeline from model sources through strategy selection, tensor operations, numerical validation, and candidate registration. Includes failure modes.

MergeEngine is the central orchestrator for all model merging operations in EMEP. This page defines its purpose, interface, pipeline stages, and the exact flow from two or more source models to a registered candidate.

Purpose

MergeEngine takes validated, compatible models and produces a merged candidate. It delegates architecture and compatibility checks to ModelCompatibilityAnalyzer, tensor operations to TensorEngine, and strategy selection to MergeStrategy plug-ins.

Interface

def merge(
    sources: List[ModelManifest],
    strategy: MergeStrategy,
    config: MergeConfig,
    experiment_id: Optional[UUID] = None
) -> CandidateManifest

Parameter	Type	Description
sources	List[ModelManifest]	Two or more models to merge
strategy	MergeStrategy	Selected merge strategy plug-in
config	MergeConfig	Strategy-specific parameters
experiment_id	Optional[UUID]	Links merge to an experiment

Returns a CandidateManifest with status, artifact URI, and merge provenance.

Full Merge Pipeline

MergeStrategy Plug-In System

MergeEngine loads strategies as plug-ins. Each strategy implements:

class MergeStrategy(Protocol):
    def validate(self, sources: List[ModelManifest], config: MergeConfig) -> ValidationResult
    def execute(self, sources: List[ModelManifest], config: MergeConfig, tensor_engine: TensorEngine) -> MergedStateDict
    def required_compatibility(self) -> CompatibilityLevel

Strategies declare their required compatibility level. A strategy requiring COMPATIBLE will reject CONDITIONALLY_COMPATIBLE pairs. A strategy accepting CONDITIONALLY_COMPATIBLE must handle the discrepancy internally.

Failure Modes

Stage	Failure	Cause	Outcome
Import	Hash mismatch	Corruption or tampering	FAIL, alert
Arch Check	INCOMPATIBLE	Architecture mismatch	FAIL, no candidate
Tokenizer Check	INCOMPATIBLE	Tokenizer mismatch	FAIL, no candidate
Shape Check	INCOMPATIBLE	Tensor shape mismatch	FAIL, no candidate
Strategy Selection	No strategy accepts pair	Compatibility level too low	FAIL
Param Validation	Invalid config	Wrong parameter types or ranges	FAIL
Tensor Ops	OOM	GPU memory exhausted	FAIL, retry with smaller batch
Numerical Validation	NaN/Inf	Instability in merge math	FAIL, quarantine
Artifact Validation	Corrupt output	Write error or truncation	FAIL
Benchmark	Timeout	Benchmark exceeds wall clock	INCOMPLETE

Cross-Links

• Merge Strategies for available strategies.
• Tensor Operations for low-level ops.
• Merge Validation for post-merge checks.
• Model Compatibility for the compatibility decision.