PyTorch: Tensor computation with GPU acceleration
Python framework for differentiable tensor computation and deep learning.
Learn more about pytorch
PyTorch is a Python machine learning library that provides tensor computation primitives and automatic differentiation capabilities. It implements reverse-mode automatic differentiation using a tape-based approach, which records operations during forward passes and replays them during backpropagation. The library supports both CPU and GPU computation, with particular emphasis on NVIDIA CUDA acceleration, while also supporting AMD ROCm and Intel GPU platforms. Common applications include deep learning research, scientific computing as a NumPy alternative, and production deployment of neural network models through its JIT compilation system.
Dynamic Computation Graphs
Tape-based autograd records operations at runtime, allowing network architectures to change between iterations without recompilation. Build conditional models with native Python control flow instead of framework-specific graph construction APIs.
Python-Native Execution Model
Imperative execution semantics mean tensors behave like NumPy arrays with standard Python debugging tools. Step through code with pdb, inspect tensors at any point, and iterate without separate compilation phases.
Unified Multi-GPU Backend
Provides seamless distributed training across multiple GPUs with automatic memory management and gradient synchronization. Scale from single GPU to multi-node clusters with minimal code changes using built-in parallelism primitives.
import torch
# Create tensors and perform operations
x = torch.tensor([[1.0, 2.0], [3.0, 4.0]])
y = torch.tensor([[5.0, 6.0], [7.0, 8.0]])
# Matrix multiplication and addition
result = torch.matmul(x, y) + 10
print(result)Patch release fixing memory regression in F.conv3d with bfloat16, torch.compile correctness bugs, and distributed crashes.
- –Install nvidia-cudnn 9.15+ from PyPI to work around significant memory regression in F.conv3d with bfloat16 inputs.
- –Fixes silent correctness error in torch.compile with error_on_graph_break and checkpoint, plus crashes in torch.bmm compilation.
Requires Python 3.10+ and macOS 14+ for MPS; ONNX exporter now defaults to dynamo=True, breaking TorchScript-based workflows.
- –Upgrade to Python 3.10 minimum and macOS 14+ for MPS backend; custom operators must not share storage between inputs and outputs.
- –Switch ONNX exports to use dynamo=True by default or explicitly set dynamo=False to preserve legacy TorchScript exporter behavior.
Drops CUDA support for Maxwell/Pascal GPUs (sm50–sm60) on CUDA 12.8/12.9 builds; Windows wheels exclude torch.segment_reduce due to CUDA 12.9.1 compiler bug.
- –Pin to CUDA 12.6 wheels if you require Maxwell or Pascal GPU support, or use torch.segment_reduce on Windows.
- –Update exception handling for unsupported dtypes from RuntimeError to NotImplementedError and replace mark_dynamic with maybe_mark_dynamic if specialization errors occur.
See how people are using pytorch
Related Repositories
Discover similar tools and frameworks used by developers
whisper.cpp
Lightweight, cross-platform speech recognition engine delivering real-time transcription with minimal dependencies and optimized inference performance.
servers
Reference implementations for LLM tool and data integration.
NAFNet
Efficient PyTorch architecture for image restoration tasks.
fish-speech
Transformer-based TTS with voice cloning from reference audio.
opencv
Cross-platform C++ library for real-time computer vision algorithms.