slangpy-torch 0.4.0

Fast PyTorch tensor access for slangpy

slangpy-torch

Minimal PyTorch native extension providing fast (~28ns) tensor metadata access from native code without Python API overhead (~350ns).

Prerequisites

• Python 3.9+
• PyTorch 2.0+ installed
• C++ compiler

Windows

Install Visual Studio 2019 or 2022 with the "Desktop development with C++" workload.

Linux

# Ubuntu/Debian
sudo apt-get install build-essential

Install

This extension must be installed with --no-build-isolation to ensure ABI compatibility with your installed PyTorch version:

pip install slangpy-torch --no-build-isolation

Verify Installation

import slangpy_torch
print(slangpy_torch.get_api_ptr())  # Should print a non-zero integer

Troubleshooting

• "torch not found": Ensure PyTorch is installed first (pip install torch)
• Windows linker errors: Run from a "Developer Command Prompt for VS" or ensure MSVC is in PATH

Usage from native code (slangpy_ext)

#include "tensor_bridge_api.h"

// At init time:
auto bridge = nb::module_::import_("slangpy_torch");
auto api = reinterpret_cast<const TensorBridgeAPI*>(
    nb::cast<uintptr_t>(bridge.attr("get_api_ptr")())
);

// In hot path (~28ns):
TensorBridgeInfo info;
api->extract(handle.ptr(), &info);
// Use info.data_ptr, info.shape, info.strides, etc.