# open-mmlab/mmrazor architecture

> Architecture diagram of the GitHub repository open-mmlab/mmrazor, generated by GitDiagram from its file tree, README and a sample of source files (last updated 2026-10-01). It is an AI-made overview: check details against the source.

- Interactive diagram: https://gitdiagram.com/open-mmlab/mmrazor
- Repository: https://github.com/open-mmlab/mmrazor

## Overview

MMRazor is an OpenMMLab model-compression toolkit for NAS, pruning, knowledge distillation, and quantization. Its principal workflow is to configure a compression method for a vision model, run the applicable search or training process, and obtain a compressed or quantized model. Shared model components, task modules, and registries support these methods. The README documents the major capabilities and plug-and-play design; the sampled source excerpts show selected search, delivery, recording, and graph behavior, not every algorithm’s runtime wiring.

Main entry points: [`mmrazor/models/algorithms/nas`](https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/models/algorithms/nas) (Architecture Search), [`mmrazor/models/distillers`](https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/models/distillers) (Distillers), [`mmrazor/models/algorithms/quantization`](https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/models/algorithms/quantization) (Quantization Algorithms), [`mmrazor/models/architectures`](https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/models/architectures) (Model Architectures)

## Diagram (Mermaid)

```mermaid
flowchart TD

subgraph group_workflow["Workflow"]
  node_configs["Method Configs"]
  node_train["Training Entry<br/>[train.py]"]
  node_training_loops["Training Loops"]
  node_compressed_model["Compressed Model"]
end

subgraph group_nas_pruning["Search and Pruning"]
  node_nas["Architecture Search"]
  node_search_loops["Search Loops"]
  node_predictors["Metric Predictors"]
  node_pruning["Pruning Algorithms"]
  node_group_fisher["Group Fisher"]
end

subgraph group_distillation["Distillation"]
  node_distill_algorithms["Distillation Algorithms"]
  node_distillers["Distillers"]
  node_losses["Distillation Losses"]
  node_deliveries["Distillation Delivery"]
  node_recorders["Model Recorders"]
end

subgraph group_quantization["Quantization"]
  node_quant_algorithms["Quantization Algorithms"]
  node_quantizers["Quantizers"]
  node_fake_quant["Fake Quantizers"]
  node_observers["Quantization Observers"]
  node_gptq["GPTQ Compression<br/>[gptq.py]"]
end

subgraph group_shared["Shared Model Services"]
  node_model_arch["Model Architectures"]
  node_dynamic_ops["Dynamic Operations"]
  node_mutables["Mutable Components"]
  node_mutators["Model Mutators"]
  node_module_graph["Module Graph<br/>[module_graph.py]"]
  node_crd_dataset["CRD Dataset"]
  node_registry["Component Registry<br/>[registry.py]"]
  node_setup["Module Registration<br/>[setup_env.py]"]
  node_calibrate["BN Calibration"]
end

node_engineer(("Model Engineer"))
node_vision_stack["Vision Project"]

node_engineer -->|"configures"| node_configs
node_configs -.->|"runs with"| node_train
node_train -.->|"launches"| node_training_loops
node_training_loops -.->|"produces"| node_compressed_model
node_vision_stack -->|"integrates with"| node_nas
node_vision_stack -->|"integrates with"| node_pruning
node_vision_stack -->|"integrates with"| node_distill_algorithms
node_vision_stack -->|"integrates with"| node_quant_algorithms
node_nas -->|"searches"| node_compressed_model
node_pruning -->|"slims"| node_compressed_model
node_distill_algorithms -->|"distills"| node_compressed_model
node_quant_algorithms -->|"quantizes"| node_compressed_model
node_search_loops -.->|"uses"| node_predictors
node_nas -.->|"uses"| node_mutators
node_pruning -->|"includes"| node_group_fisher
node_distill_algorithms -.->|"uses"| node_distillers
node_distillers -.->|"computes with"| node_losses
node_distillers -.->|"coordinates"| node_deliveries
node_distillers -.->|"records with"| node_recorders
node_quant_algorithms -.->|"uses"| node_quantizers
node_quantizers -.->|"uses"| node_fake_quant
node_quantizers -.->|"uses"| node_observers
node_quant_algorithms -.->|"includes"| node_gptq
node_nas -.->|"searches"| node_model_arch
node_model_arch -.->|"uses"| node_dynamic_ops
node_model_arch -.->|"contains"| node_mutables
node_mutators -.->|"controls"| node_mutables
node_module_graph -->|"traces"| node_model_arch
node_crd_dataset -.->|"provides samples"| node_vision_stack
node_setup -->|"registers modules"| node_registry
node_search_loops -.->|"calibrates with"| node_calibrate

click node_configs "https://github.com/open-mmlab/mmrazor/tree/main/configs"
click node_train "https://github.com/open-mmlab/mmrazor/blob/main/tools/train.py"
click node_training_loops "https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/engine/runner"
click node_nas "https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/models/algorithms/nas"
click node_search_loops "https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/engine/runner"
click node_predictors "https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/models/task_modules/predictor/handler"
click node_pruning "https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/models/algorithms/pruning"
click node_group_fisher "https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/implementations/pruning/group_fisher"
click node_distill_algorithms "https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/models/algorithms/distill/configurable"
click node_distillers "https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/models/distillers"
click node_losses "https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/models/losses"
click node_deliveries "https://github.com/open-mmlab/mmrazor/blob/main/mmrazor/models/task_modules/delivery/delivery_manager.py"
click node_recorders "https://github.com/open-mmlab/mmrazor/blob/main/mmrazor/models/task_modules/recorder/recorder_manager.py"
click node_quant_algorithms "https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/models/algorithms/quantization"
click node_quantizers "https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/models/quantizers"
click node_fake_quant "https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/models/fake_quants"
click node_observers "https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/models/observers"
click node_gptq "https://github.com/open-mmlab/mmrazor/blob/main/mmrazor/implementations/quantization/gptq/gptq.py"
click node_model_arch "https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/models/architectures"
click node_dynamic_ops "https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/models/architectures/dynamic_ops"
click node_mutables "https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/models/mutables"
click node_mutators "https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/models/mutators"
click node_module_graph "https://github.com/open-mmlab/mmrazor/blob/main/mmrazor/structures/graph/module_graph.py"
click node_crd_dataset "https://github.com/open-mmlab/mmrazor/blob/main/mmrazor/datasets/crd_dataset_wrapper.py"
click node_registry "https://github.com/open-mmlab/mmrazor/blob/main/mmrazor/registry/registry.py"
click node_setup "https://github.com/open-mmlab/mmrazor/blob/main/mmrazor/utils/setup_env.py"
click node_calibrate "https://github.com/open-mmlab/mmrazor/blob/main/mmrazor/engine/runner/utils/calibrate_bn_mixin.py"

classDef toneNeutral fill:#f8fafc,stroke:#334155,stroke-width:1.5px,color:#0f172a
classDef toneBlue fill:#dbeafe,stroke:#2563eb,stroke-width:1.5px,color:#172554
classDef toneAmber fill:#fef3c7,stroke:#d97706,stroke-width:1.5px,color:#78350f
classDef toneMint fill:#dcfce7,stroke:#16a34a,stroke-width:1.5px,color:#14532d
classDef toneRose fill:#ffe4e6,stroke:#e11d48,stroke-width:1.5px,color:#881337
classDef toneIndigo fill:#e0e7ff,stroke:#4f46e5,stroke-width:1.5px,color:#312e81
classDef toneTeal fill:#ccfbf1,stroke:#0f766e,stroke-width:1.5px,color:#134e4a
class node_configs,node_train,node_training_loops,node_compressed_model toneBlue
class node_nas,node_search_loops,node_predictors,node_pruning,node_group_fisher toneAmber
class node_distill_algorithms,node_distillers,node_losses,node_deliveries,node_recorders toneMint
class node_quant_algorithms,node_quantizers,node_fake_quant,node_observers,node_gptq toneRose
class node_model_arch,node_dynamic_ops,node_mutables,node_mutators,node_module_graph,node_crd_dataset,node_registry,node_setup,node_calibrate,node_engineer,node_vision_stack toneIndigo
```

## Components

### Workflow

- [Method Configs](https://github.com/open-mmlab/mmrazor/tree/main/configs): `configs`
- [Training Entry](https://github.com/open-mmlab/mmrazor/blob/main/tools/train.py): `tools/train.py`
- [Training Loops](https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/engine/runner): `mmrazor/engine/runner`
- **Compressed Model**

### Search and Pruning

- [Architecture Search](https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/models/algorithms/nas): `mmrazor/models/algorithms/nas`
- [Search Loops](https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/engine/runner): `mmrazor/engine/runner`
- [Metric Predictors](https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/models/task_modules/predictor/handler): `mmrazor/models/task_modules/predictor/handler`
- [Pruning Algorithms](https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/models/algorithms/pruning): `mmrazor/models/algorithms/pruning`
- [Group Fisher](https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/implementations/pruning/group_fisher): `mmrazor/implementations/pruning/group_fisher`

### Distillation

- [Distillation Algorithms](https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/models/algorithms/distill/configurable): `mmrazor/models/algorithms/distill/configurable`
- [Distillers](https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/models/distillers): `mmrazor/models/distillers`
- [Distillation Losses](https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/models/losses): `mmrazor/models/losses`
- [Distillation Delivery](https://github.com/open-mmlab/mmrazor/blob/main/mmrazor/models/task_modules/delivery/delivery_manager.py): `mmrazor/models/task_modules/delivery/delivery_manager.py`
- [Model Recorders](https://github.com/open-mmlab/mmrazor/blob/main/mmrazor/models/task_modules/recorder/recorder_manager.py): `mmrazor/models/task_modules/recorder/recorder_manager.py`

### Quantization

- [Quantization Algorithms](https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/models/algorithms/quantization): `mmrazor/models/algorithms/quantization`
- [Quantizers](https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/models/quantizers): `mmrazor/models/quantizers`
- [Fake Quantizers](https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/models/fake_quants): `mmrazor/models/fake_quants`
- [Quantization Observers](https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/models/observers): `mmrazor/models/observers`
- [GPTQ Compression](https://github.com/open-mmlab/mmrazor/blob/main/mmrazor/implementations/quantization/gptq/gptq.py): `mmrazor/implementations/quantization/gptq/gptq.py`

### Shared Model Services

- [Model Architectures](https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/models/architectures): `mmrazor/models/architectures`
- [Dynamic Operations](https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/models/architectures/dynamic_ops): `mmrazor/models/architectures/dynamic_ops`
- [Mutable Components](https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/models/mutables): `mmrazor/models/mutables`
- [Model Mutators](https://github.com/open-mmlab/mmrazor/tree/main/mmrazor/models/mutators): `mmrazor/models/mutators`
- [Module Graph](https://github.com/open-mmlab/mmrazor/blob/main/mmrazor/structures/graph/module_graph.py): `mmrazor/structures/graph/module_graph.py`
- [CRD Dataset](https://github.com/open-mmlab/mmrazor/blob/main/mmrazor/datasets/crd_dataset_wrapper.py): `mmrazor/datasets/crd_dataset_wrapper.py`
- [Component Registry](https://github.com/open-mmlab/mmrazor/blob/main/mmrazor/registry/registry.py): `mmrazor/registry/registry.py`
- [Module Registration](https://github.com/open-mmlab/mmrazor/blob/main/mmrazor/utils/setup_env.py): `mmrazor/utils/setup_env.py`
- [BN Calibration](https://github.com/open-mmlab/mmrazor/blob/main/mmrazor/engine/runner/utils/calibrate_bn_mixin.py): `mmrazor/engine/runner/utils/calibrate_bn_mixin.py`

### Other components

- **Model Engineer**
- **Vision Project**

## Connections

- Model Engineer → Method Configs: configures (evidence: `README.md`)
- Method Configs → Training Entry: runs with
- Training Entry → Training Loops: launches
- Training Loops → Compressed Model: produces
- Vision Project → Architecture Search: integrates with (evidence: `README.md`)
- Vision Project → Pruning Algorithms: integrates with (evidence: `README.md`)
- Vision Project → Distillation Algorithms: integrates with (evidence: `README.md`)
- Vision Project → Quantization Algorithms: integrates with (evidence: `README.md`)
- Architecture Search → Compressed Model: searches (evidence: `README.md`)
- Pruning Algorithms → Compressed Model: slims (evidence: `README.md`)
- Distillation Algorithms → Compressed Model: distills (evidence: `README.md`)
- Quantization Algorithms → Compressed Model: quantizes (evidence: `README.md`)
- Search Loops → Metric Predictors: uses
- Architecture Search → Model Mutators: uses
- Pruning Algorithms → Group Fisher: includes (evidence: `README.md`)
- Distillation Algorithms → Distillers: uses
- Distillers → Distillation Losses: computes with
- Distillers → Distillation Delivery: coordinates
- Distillers → Model Recorders: records with
- Quantization Algorithms → Quantizers: uses
- Quantizers → Fake Quantizers: uses
- Quantizers → Quantization Observers: uses
- Quantization Algorithms → GPTQ Compression: includes
- Architecture Search → Model Architectures: searches
- Model Architectures → Dynamic Operations: uses (evidence: `mmrazor/models/architectures/dynamic_ops/__init__.py`)
- Model Architectures → Mutable Components: contains
- Model Mutators → Mutable Components: controls
- Module Graph → Model Architectures: traces (evidence: `mmrazor/structures/graph/module_graph.py`)
- CRD Dataset → Vision Project: provides samples
- Module Registration → Component Registry: registers modules (evidence: `mmrazor/utils/setup_env.py`)
- Search Loops → BN Calibration: calibrates with (evidence: `mmrazor/engine/runner/utils/__init__.py`)
