aind-torch-utils

Generic, queue-based, multi-GPU PyTorch inference pipeline.

Features

• Block/patch based volume processing with optional overlap
• Multi-threaded preparation & writing stages
• Multi-GPU inference workers
• Flexible blending/trim seam handling
• Queue & system monitoring hooks

Installation

To use the software, in the root directory, run

pip install -e .

To develop the code, run

pip install -e . --group dev

Note: --group flag is available only in pip versions >=25.1

Alternatively, if using uv, run

uv sync