API Reference¶

Welcome to the FusionBench API reference. This documentation covers all public classes, functions, and modules available in the FusionBench package.

Quick Links¶

Algorithms - Fusion algorithms (ensemble, merging, mixing)
Model Pool - Model management and loading
Task Pool - Evaluation tasks and benchmarks
Mixins - Reusable functionality
Programs - Execution programs
Utilities - Helper functions

Module Overview¶

Core Components¶

Module	Description
`fusion_bench.method`	Fusion algorithms including ensemble, merging, mixing, and compression methods
`fusion_bench.modelpool`	Model pool implementations for managing and loading models
`fusion_bench.taskpool`	Task pool implementations for evaluation
`fusion_bench.mixins`	Reusable mixins for extending functionality
`fusion_bench.programs`	Program execution frameworks

Supporting Modules¶

Module	Description
`fusion_bench.utils`	Utility functions for logging, caching, data processing
`fusion_bench.models`	Model wrapper definitions
`fusion_bench.dataset`	Dataset definitions and loaders
`fusion_bench.metrics`	Evaluation metrics
`fusion_bench.tasks`	Task base classes
`fusion_bench.constants`	Constant definitions
`fusion_bench.optim`	Optimizers and learning rate schedulers

Deprecated¶

Module	Status
`fusion_bench.compat`	Legacy v0.1.x compatibility layer (deprecated)

Base Classes¶

Algorithm Base¶

The base class for all fusion algorithms in FusionBench.

fusion_bench.BaseAlgorithm - Abstract base class defining the algorithm interface
fusion_bench.BaseModelFusionAlgorithm - Base class for model fusion algorithms

from fusion_bench import BaseAlgorithm

class MyAlgorithm(BaseAlgorithm):
    def run(self, modelpool):
        # implement the fusion logic
        pass

Pool Base Classes¶

fusion_bench.BaseModelPool - Base class for managing and loading models
fusion_bench.BaseTaskPool - Base class for evaluation tasks

Program Base¶

fusion_bench.BaseHydraProgram - Base class for Hydra-based execution programs
fusion_bench.FabricModelFusionProgram - Program for fusing models with Lightning Fabric

Usage Examples¶

Using an Algorithm¶

from fusion_bench import SimpleAverageAlgorithm

# Create algorithm instance
algorithm = SimpleAverageAlgorithm(
    weight_key="task_weight",
    scale_factor=1.0
)

# Run fusion
merged_model = algorithm.run(modelpool)

Using Model Pool¶

from fusion_bench import CLIPVisionModelPool

# Load models from pool
modelpool = CLIPVisionModelPool.from_config("config/modelpool/CLIPVisionModelPool/clip-vit-base-patch32_TA8.yaml")
models = modelpool.load_models()

Using Task Pool¶

from fusion_bench import CLIPVisionModelTaskPool

# Evaluate model on tasks
taskpool = CLIPVisionModelTaskPool()
results = taskpool.evaluate(model)

Browse the API documentation by category:

Model Ensemble Algorithms - Simple and weighted ensembles
Model Merging Algorithms - Linear interpolation and optimization-based methods
Model Mixing Algorithms - Layer-level and MoE-based mixing
Model Compression Algorithms - Pruning and compression techniques
Training Algorithms - Fine-tuning and training methods