How eml-eda/plinio works
Project type and purpose
• PLiNIO is an open-source Python library (built on PyTorch) for gradient-based, hardware-aware DNN inference optimization.
• Its main goal is to let users plug lightweight optimization (DNAS, mixed-precision search, pruning) into their standard training loop with minimal code changes.
File structure
• Top level:
• setup.py, requirements.txt – packaging and installation scripts
• README.md, LICENSE – documentation and license
• examples/ – Jupyter notebooks & helper scripts demonstrating usage
• unit_test/ – model definitions and test suites
• Core package plinio/:
• cost/ – hardware cost models (generic and platform-specific)
• graph/ – representation, annotation, transformation, feature extraction of DNN computation graph
• methods/ – optimization strategies subdivided into:
– dnas_base (core DNAS framework)
– supernet (coarse layer selection)
– pit (fine-grained pruning)
– mps (mixed-precision search) with multiple backends (diana, onnx, match, maupiti)
– odimo_mps (multi-accelerator mapping)
• regularizers/ – loss regularizers for cost terms and constraints
• __init__.py – public API entry points
• Supporting directories: .assets/ (logo), .vscode/ (editor settings), .gitignore
README for additional insights
• Confirms four main optimization methods (SuperNet, PIT, MPS, ODiMO MPS)
• Highlights cost modeling sub-package and regularizers for flexible loss terms
• Usage pattern: user imports a Method class, invokes .search(...) or similar, and plugs model.cost or regularizer(model) into the loss
1) User Training Loop (PyTorch)
2) PLiNIO Public API
3) Graph Module (plinio.graph)
4) Cost Models (plinio.cost)
5) Optimization Methods (plinio.methods)
– DNAS Base
– SuperNet
– PIT
– MPS (+ backends)
– ODiMO MPS
6) Regularizers (plinio.regularizers)
7) Examples & Notebooks
8) Unit Tests & CI
• Training Loop → calls → PLiNIO Public API
• Public API → requests → Graph Module to build/annotate model graph
• Public API → queries → Cost Models to compute cost metrics
• Public API → dispatches to → Selected Optimization Method
• Optimization Methods → use → Graph Module & Cost Models internally
• MPS Method → loads → specific backends (Diana, ONNX, Match, Maupiti)
• Public API → returns → modified model and cost attributes
• Training Loop ← consumes ← model.cost or regularizer(model) from Regularizers
• Examples & Notebooks illustrate → usage connecting Training Loop and PLiNIO API
• Unit Tests validate → each module and end-to-end flows
c. Architectural patterns & principles:
• Modular “plugin” design: new methods, cost models, backends, regularizers drop into their respective subpackages
• Strategy pattern: Method classes encapsulate different optimization algorithms
• Factory or registry for quantization backends inside MPS
• Clear separation of concerns: graph construction, cost calculation, optimization logic, and loss regularization
d. Key technologies & frameworks:
• Python 3.x, PyTorch tensors and modules
• Jupyter notebooks for examples
• pytest (or unittest) for unit_test
• setup.py & pip for distribution
Essential architectural elements
• Plugin architecture and extension points
• Flow from user loop → PLiNIO API → internal modules → back to loop is obvious