hw-compat-matrix 0.2.0

Canonical CUDA/GPU compatibility metadata for hardware compatibility checking

hw-compat-matrix

Canonical CUDA/GPU compatibility metadata for hardware compatibility checking.

Overview

This package provides structured data and utilities for:

• CUDA toolkit version to minimum driver version mapping
• NVIDIA SM (Streaming Multiprocessor) version to GPU architecture mapping
• GPU name normalization across cloud providers (Vast.ai, AWS, Lambda, RunPod, Azure)

Installation

pip install hw-compat-matrix

Usage

from hw_compat_matrix import (
    get_min_driver_for_cuda,
    get_architecture_for_sm,
    normalize_gpu_name,
)

# Get minimum driver version for a CUDA toolkit version
min_driver = get_min_driver_for_cuda("12.8")  # Returns "570.26"

# Get GPU architecture for an SM version
arch = get_architecture_for_sm("89")  # Returns "Ada Lovelace"

# Normalize GPU names from different providers
canonical = normalize_gpu_name("RTX_4090", provider="vastai")  # Returns "RTX 4090"

Loading Raw Data

from hw_compat_matrix import (
    load_cuda_driver_matrix,
    load_nvidia_sm_map,
    load_gpu_aliases,
)

# Load raw data as Pydantic models
cuda_matrix = load_cuda_driver_matrix()
sm_map = load_nvidia_sm_map()
gpu_aliases = load_gpu_aliases()

Data Sources

• cuda_driver_matrix.yaml: From CUDA Toolkit Release Notes
• nvidia_sm_map.yaml: From CUDA GPUs
• gpu_name_aliases.yaml: Collected from provider APIs

Contributing

Adding New CUDA Versions

1. Update src/hw_compat_matrix/data/cuda_driver_matrix.yaml
2. Run validation: pytest tests/
3. Submit a PR

Adding New GPU Aliases

1. Update src/hw_compat_matrix/data/gpu_name_aliases.yaml
2. Run validation: pytest tests/
3. Submit a PR

License

MIT License - see LICENSE for details.