easywheels

Install GPU Python packages without the headache.

easywheels auto-detects your CUDA version, GPU architecture, PyTorch version, and Python version, then installs the exact right pre-built wheel. No more hunting through compatibility matrices or building from source.

Install

pip install easywheels

Quick Start

# Log in with your GitHub account
easywheels login

# Install any GPU package
easywheels install flash-attn

easywheels detects your environment automatically:

Detected: Python 3.12, CUDA 12.8, RTX 4090 (sm_89), torch 2.9.0

Resolving flash-attn...
  Found: flash_attn-2.8.3+cu128torch2.9-cp312-cp312-linux_x86_64.whl
  CUDA: cu128, Torch: 2.9

Running: pip install flash_attn-2.8.3+cu128torch2.9-cp312-cp312-linux_x86_64.whl
Done in 9 seconds.

Why?

Installing GPU Python packages on CUDA is painful:

• Some packages ship source-only on PyPI. No pre-built wheels at all.
• Pre-built wheels that do exist are scattered across GitHub repos, custom indices, and community forks.
• Many CUDA/Python/platform combos simply don't have a wheel anywhere.
• Building from source takes 30-120 minutes and frequently fails.

easywheels solves this. It mirrors pre-built wheels from every upstream source and builds the gaps on GPU infrastructure. 2,200+ wheels across 10 packages, served through a single registry.

Commands

easywheels install <package>

Detects your environment and installs the best matching wheel.

easywheels install flash-attn
easywheels install mamba-ssm causal-conv1d
easywheels install flash-attn==2.8.3    # pin a version
easywheels install flash-attn -U        # upgrade
easywheels install flash-attn --dry-run # show what would install

easywheels detect

Shows your detected environment without installing anything.

easywheels detect

easywheels login

Authenticates via GitHub device OAuth. Opens your browser, you authorize, done. Your API key is stored in ~/.easywheels/config.toml.

easywheels login

easywheels search <package>

Shows all available wheels for a package that match your environment.

easywheels search flash-attn

easywheels config

Manage configuration.

easywheels config --show
easywheels config --set-key ew_xxx

What's in the Registry?

2,200+ pre-built wheels across 10 packages:

• flash-attn, flash-attn-3, deepspeed, mamba-ssm, causal-conv1d, exllamav2, llama-cpp-python, gptqmodel, sageattention, flashinfer-jit-cache
• CUDA 12.4 through 13.0
• Python 3.10-3.13
• Linux fully covered. Windows build-out in progress.

GPU architectures: Turing (sm_75) through Hopper (sm_90) with PTX forward compatibility.

How It Works

1. Detection. The CLI checks nvidia-smi, nvcc, CUDA_PATH/CUDA_HOME, and torch to determine your CUDA version, GPU compute capability, and PyTorch version.

2. Resolution. Your environment is sent to the EasyWheels API, which finds the best compatible wheel considering CUDA version, torch ABI, platform, and architecture.

3. Installation. The exact right wheel is downloaded and handed to pip. No guessing, no source builds.

Configuration

Config lives in ~/.easywheels/config.toml:

api_key = "ew_abc123..."
api_url = "https://easywheels.io"

You can also set EASYWHEELS_API_KEY as an environment variable.

Pricing

The CLI is free and open source. The registry requires a subscription because building and hosting GPU wheels costs real money.

Plan	Price	Downloads
Trial	Free 14 days	3 downloads
Lite	$9/mo	10/mo
Pro	$19/mo	Unlimited
Team	$49/mo	Unlimited, 5 seats

Sign up at easywheels.io

Requirements

• Python 3.9+
• pip
• NVIDIA GPU with CUDA drivers (for GPU packages)
• PyTorch (optional, improves detection accuracy)

Links

• Registry: easywheels.io
• Packages: easywheels.io/packages

License

MIT