librmm-cu12

What it is and what it does

librmm-cu12 is the CUDA 12 variant of the RAPIDS Memory Manager, a library that provides a common interface for customizing GPU and host memory allocation. It addresses the performance requirements of GPU-centric workflows by allowing developers to use optimizations like pinned host memory for asynchronous transfers or device memory pools to reduce allocation overhead. The package wraps the C++ librmm library and exposes it to Python applications, making it suitable for projects built on the RAPIDS ecosystem that need fine-grained control over memory behavior on NVIDIA GPUs.

The library is designed for developers building GPU-accelerated data science and machine learning applications. It depends on rapids-logger for logging and requires a modern NVIDIA GPU (Volta or newer) and CUDA 12.2 or later. Installation uses platform-specific wheels, so setup is straightforward on supported Linux architectures, though the GPU and CUDA toolkit must already be present on the system.

Use it for:

• Optimize asynchronous host-to-device memory transfers in GPU-accelerated data pipelines using pinned host memory.
• Reduce dynamic GPU memory allocation overhead in iterative machine learning training by using device memory pools.
• Customize memory allocation strategies in RAPIDS-based applications for specific workload patterns.
• Integrate custom memory management into C++ GPU applications via the RMM interface while maintaining Python compatibility.
• Profile and control memory usage patterns in multi-GPU workflows to avoid fragmentation and improve throughput.

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

librmm-cu12 provides GPU memory allocation and management for CUDA 12 workloads, offering customizable device and host memory pools to optimize GPU-centric applications.

Yes, if you are building GPU-accelerated applications on CUDA 12 and need fine-grained memory management. The package is actively maintained, has no known vulnerabilities, and carries a permissive license. Install friction is moderate due to platform-specific wheels, but setup is straightforward on supported systems. Not applicable for CPU-only or non-NVIDIA GPU workflows.

Install

librmm-cu12 on PyPI

pip install librmm-cu12

uv add librmm-cu12

poetry add librmm-cu12

Installing librmm-cu12

Before you install

Medium install friction due to platform-specific wheels (manylinux_2_24 and manylinux_2_28 for aarch64 and x86_64). Package is actively maintained with recent releases and depends only on rapids-logger. Last release 8 days old; repository shows active development.

License in practice

Apache-2.0 permissive license allows use in commercial and proprietary projects with minimal restrictions, requiring only license attribution.

Quickstart

pip install librmm-cu12
import librmm_cu12

Requires CUDA 12.2+ and Volta architecture GPU (Compute Capability 7.0+); NVIDIA GPU drivers and CUDA toolkit must be installed on the system.

Verify before relying

• Whether the package works on systems without NVIDIA GPUs or falls back gracefully.
• Exact Python version compatibility (requires_python field is unspecified in metadata).
• Whether rapids-logger dependency is optional or required at runtime.
• Specific API surface and memory allocation customization options available in version 26.8.0.

Package facts

Evidence: librmm_cu12-26.8.0-py3-none-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl; librmm_cu12-26.8.0-py3-none-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl

Tags

More Scientific/Engineering packages

Further reading