CPUTorch

CPUTorch is a lightweight tensor and neural-network framework built from scratch, with a C++ numerical core, Python API, CPU execution, and Vulkan compute support.

The project is designed as a general-purpose computational foundation for machine learning and artificial intelligence systems.

Author

Frandika Imam Arifin

Repository

https://github.com/Mangono1/minitorch

Architecture

• C++ numerical core
• Python API
• CPU backend
• Vulkan compute backend
• Automatic differentiation
• Tensor operations
• Neural-network modules
• Optimizers
• Hardware profiling
• Hardware-aware execution
• Adaptive execution architecture

Current Version

CPUTorch V0.5.5

V0.5.5

Current development includes:

• Tensor engine
• requires_grad
• Computational graph infrastructure
• Automatic differentiation infrastructure
• Neural-network modules
• Linear layers
• ReLU
• Sequential models
• MSE loss
• SGD optimizer
• Hardware profiler
• CPU backend
• Vulkan compute infrastructure
• Vulkan buffer management
• Vulkan compute pipeline
• Vector arithmetic backend operations
• Scalar multiplication operation
• Python package
• PyPI distribution
• Multi-platform wheel build automation

Installation

Install the released package directly from PyPI:

python -m pip install cputorch

Or install a specific version:

python -m pip install cputorch==0.5.5

Quick Example

import cputorch

x = cputorch.Tensor(
    [1.0, 2.0, 3.0, 4.0],
    [2, 2],
    requires_grad=True
)

y = x.multiply_scalar(2.0)

print(y.data)

Design Philosophy

CPUTorch is intended to remain a general computational framework.

It is not tied to a specific domain or dataset.

It can serve as a foundation for:

• Machine learning
• Deep learning
• Computer vision
• Scientific computing
• Robotics
• Simulation
• Reinforcement learning
• Custom AI systems
• Domain-specific AI frameworks

Higher-level frameworks and models can use CPUTorch as their computational backend.

Roadmap

• V0.1 Tensor core
• V0.2 Automatic differentiation
• V0.3 Neural-network modules
• V0.4 Optimizers
• V0.5 Hardware profiler
• V0.5.x Vulkan compute development
• V0.6 Expanded Vulkan backend
• V0.7 GPU watchdog and adaptive scheduler
• V0.8 Transformer architecture
• V0.9 Tokenizer and language-model infrastructure
• V1.0 General training engine

License

MIT License.