kladml 0.10.1

KladML SDK - Enterprise-grade MLOps toolkit

Build ML pipelines with pluggable backends. Simple. Modular. Yours.

⭐ Star us on GitHub to support the project!

---

Why KladML?

Feature	KladML	MLflow	ClearML
Interface-based	✅ Pluggable	❌ Hardcoded	❌ Hardcoded
Server required	❌ No	⚠️ Optional	✅ Yes
Local-first	✅ Unified SQLite DB	✅ Yes	❌ No
Learning curve	🟢 Minutes	🟡 Days	🔴 Weeks
Hierarchy	✅ Workspace/Proj/Fam	❌ Exp/Run	❌ Project/task
User Interface	✅ TUI (Terminal)	⚠️ Web UI	✅ Web UI
Custom backends	✅ Easy	⚠️ Complex	❌ No
Data Engine	🚀 Polars (Fast)	🐢 Pandas	🐢 Pandas

---

---

Requirements

• Python: 3.10, 3.11, 3.12 (Native support for modern type hints)
• OS: Linux, macOS, Windows

Installation

# Core (lightweight, minimal dependencies)
pip install kladml

# Training + Data Engine (includes Polars, Torch)
pip install "kladml[train]"

# Full Suite (Tracking + TUI + Dev)
pip install "kladml[all]"

---

Workflow

1. Initialize Workspace

kladml init

Creates the standard folder structure (data/configs/, data/projects/, data/datasets/).

2. Interactive Management (TUI)

kladml ui

Explore projects, runs, and datasets visually in your terminal.

3. Training

# Train using a config file (auto-detects GPU/MPS)
kladml train --config data/configs/my_config.yaml

# Distributed Training (Multi-GPU)
kladml train --config ... --distributed --num-processes 2

✅ Universal Trainer: Supports Mixed Precision (FP16/BF16), Gradient Accumulation, and Multi-GPU without changing code.

---

Built-in Baselines

KladML is designed to work with any custom model (PyTorch, Scikit-learn, etc.). For convenience, we provide these reference implementations out-of-the-box:

Domain	Reference Model
Tabular	XGBoost (Coming Soon)
Time Series	Transformers
Computer Vision	ResNet / ViT (Coming Soon)
TEXT	BERT (Coming Soon)

---

Architecture

KladML uses dependency injection with abstract interfaces. Swap implementations without changing your code:

┌─────────────────────────────────────────────────────────────┐
│                      Your Code                              │
├─────────────────────────────────────────────────────────────┤
│                   ExperimentRunner                          │
├─────────────────────────────────────────────────────────────┤
│  StorageInterface  │  ConfigInterface  │  TrackerInterface  │
├─────────────────────────────────────────────────────────────┤
│  LocalStorage      │  YamlConfig       │  LocalTracker      │
│  S3Storage         │  EnvConfig        │  MLflowTracker     │
│  (your impl)       │  (your impl)      │  (your impl)       │
└─────────────────────────────────────────────────────────────┘

Implement Custom Backends

from kladml.interfaces import StorageInterface

class S3Storage(StorageInterface):
    """Custom S3 implementation."""
    
    def upload_file(self, local_path, bucket, key):
        # Your S3 logic
        ...

# Plug it in
runner = ExperimentRunner(storage=S3Storage())

---

Interfaces

Interface	Description	Default
StorageInterface	Object storage (files, artifacts)	LocalStorage
ConfigInterface	Configuration management	YamlConfig
PublisherInterface	Real-time metric publishing	ConsolePublisher
TrackerInterface	Experiment tracking	LocalTracker (MLflow + SQLite)

---

Configuration

Create kladml.yaml:

project:
  name: my-project
  version: 0.1.0

training:
  device: auto  # auto | cpu | cuda | mps

storage:
  artifacts_dir: ./data

Or use environment variables:

export KLADML_TRAINING_DEVICE=cuda
export KLADML_STORAGE_ARTIFACTS_DIR=/data/artifacts

---

CLI Commands

kladml --help                 # Show all commands
kladml init                   # Initialize workspace
kladml version                # Show version

# Training
kladml train quick ...        # Quick training (no DB setup)
kladml train single ...       # Full training with project/experiment

# Evaluation
kladml eval run ...           # Evaluate a model
kladml eval info              # Show available evaluators
kladml compare --runs r1,r2   # Compare runs side-by-side

# Data
kladml data inspect <path>    # Analyze a dataset (Parquet/PKL)
kladml data summary <dir>     # Summary of datasets
kladml data convert ...       # Convert PKL -> Parquet/HDF5

# Models
kladml export ...      # Export to ONNX

# Organization
kladml project list           # List all projects
kladml family list ...        # List families
kladml experiment list ...    # List experiments

---

Contributing

PRs welcome! See CONTRIBUTING.md for guidelines.

git clone https://github.com/kladml/kladml.git
cd kladml
pip install -e ".[dev]"
pytest

---

License

MIT License - see LICENSE for details.

---

Documentation · PyPI · GitHub

Made in 🇮🇹 by the KladML Team