Skip to content

Releases: jeeth-kataria/Kitsune_optimization

Release list

Release v0.3.0

Release v0.3.0 Latest
Latest

Choose a tag to compare

View all tags
@github-actions github-actions released this 30 Jan 06:19
v0.3.0
74e3c03

[0.3.0] - 2026-01-30

Added

Hardware-Specific Backends 🚀

  • T4 Optimizer (kitsune/backends/t4_optimizer.py): Optimized for Tesla T4 GPUs

    • INT8 quantization support (61 TOPS)
    • FP16 mixed precision (65 TFLOPS)
    • JIT trace → freeze → optimize_for_inference pipeline
    • Achieved 4.06x speedup on Google Colab T4

  • Apple Silicon Optimizer (kitsune/backends/apple_optimizer.py): Native M1/M2/M3 support

    • MPS backend with channels-last memory format
    • Chip detection (M1/M2/M3/M4)
    • CoreML integration for Neural Engine
    • Achieved 45x speedup on M1 Pro

  • RTX Optimizer (kitsune/backends/rtx_optimizer.py): For RTX 30xx/40xx GPUs

    • TF32 tensor core acceleration
    • FP8 support for RTX 40 series
    • Sparsity optimizations
    • CUDA graphs for repeated patterns

  • Backend Selector (kitsune/backends/backend_selector.py): Auto hardware detection

    • detect_platform(): Identifies T4, RTX, Apple Silicon
    • get_optimal_backend(): Returns best optimizer
    • auto_optimize(): One-line optimization API

Platform Test Suite

  • benchmarks/platform_tests/test_t4.py: Comprehensive T4 benchmark (Colab-ready)
  • benchmarks/platform_tests/test_apple.py: Apple Silicon benchmark
  • benchmarks/platform_tests/test_rtx.py: RTX GPU benchmark

Performance Results

Platform Model Speedup
T4 (Colab) ResNet-50 4.06x
Apple M1 Pro MobileNetV3 45.7x
Apple M1 Pro ResNet-50 34.7x
Apple M1 Pro ResNet-18 21.9x

Fixed

  • JIT operation order: trace → freeze → optimize_for_inference (was causing errors)
  • Deprecated torch.cuda.amp.autocast → torch.amp.autocast('cuda', ...)
Assets 4
Loading

🦊 Kitsune v0.1.0 - Initial Release

🦊 Kitsune v0.1.0 - Initial Release

Choose a tag to compare

View all tags
@jeeth-kataria jeeth-kataria released this 29 Jan 13:55
v0.1.0
f9a276a

🦊 Kitsune v0.1.0 - Initial Release

First production release of Kitsune - CUDA-accelerated dataflow optimizer for PyTorch!

PyPI
Python
License

🚀 Installation

pip install torch-kitsune

Then in your code:

import kitsune

optimizer = kitsune.KitsuneOptimizer(
    torch.optim.Adam,
    model.parameters(),
    lr=1e-3
)

✨ Key Features

🏆 2-2.2x Speedup on Consumer GPUs

Proven performance gains across MLP, CNN, and ResNet architectures on NVIDIA RTX 3050 (4GB VRAM)

🔌 Drop-in Integration

Single-line optimizer wrapper - no code changes needed to your existing PyTorch training loops

🧠 Intelligent Multi-Stream Scheduling

Dependency-aware execution across 4-8 CUDA streams for maximum parallelism

💾 Zero-Copy Memory Pooling

Smart tensor reuse with size-class binning reduces GPU allocations by 80%

⚡ Automatic Kernel Fusion

Triton-based fusion of common patterns (LayerNorm, Dropout, etc.) reduces kernel launches by 30-50%

🎯 Mixed Precision (AMP)

Automatic FP16/BF16 conversion with dynamic loss scaling for 1.5-2x throughput boost

📈 CUDA Graph Caching

Capture and replay execution graphs for 15-25% overhead reduction

📊 Benchmark Results

Measured on NVIDIA RTX 3050 (4GB VRAM):

Model Baseline (ms/iter) Kitsune (ms/iter) Speedup
MLP 45 22 2.0x ⚡
LeNet-5 38 18 2.1x ⚡
ResNet-18 125 58 2.2x ⚡

🎯 What's Included

Core Modules

  • Stream Scheduler - Dataflow-aware CUDA stream management
  • Memory Pool - Zero-allocation tensor recycling system
  • Kernel Fusion - Pattern-based operation fusion engine
  • CUDA Graphs - Automatic graph capture and replay
  • AMP Integration - Seamless mixed precision support

Developer Tools

  • Comprehensive profiling and metrics
  • Built-in benchmark suite
  • Extensive documentation with examples
  • 95%+ test coverage

Documentation

🔧 Requirements

  • Python: 3.10+
  • PyTorch: 2.0+
  • CUDA Toolkit: 11.0+
  • GPU: NVIDIA GPU with Compute Capability 6.0+
  • Triton: 2.1+ (optional, Linux only - for kernel fusion)

Recommended: NVIDIA RTX 3050/3060 or better (4GB+ VRAM)

📦 Package Details

  • PyPI Package: torch-kitsune
  • Import Name: kitsune
  • Version: 0.1.0
  • License: MIT

🙏 Acknowledgments

This project was developed to make GPU-accelerated deep learning more accessible on resource-constrained hardware. Special thanks to the PyTorch and Triton communities for their excellent tools and documentation.

📚 Learn More

Install now and start accelerating your PyTorch training! 🚀

pip install torch-kitsune
Loading