Building the GPU
Build Farm for AI

WheelForge exists because compiling CUDA kernels is a nightmare. Someone needed to build a dedicated farm to handle the heavy lifting—so you don't have to.

Our Mission

To make GPU-native software build, install, and run as easily as pip install.

AI developers shouldn’t fight CUDA versions, compiler flags, SM architectures, broken builds, or hardware drift.
WheelForge is the infrastructure layer that removes that friction forever.

The Problem

Package Managers Weren’t Built for GPUs

CUDA, PyTorch wheels, arch-specific kernels — pip never signed up for this.

Hardware Variation Has Exploded

V100, A100, H100, L40s, Jetson, consumer GPUs — everything requires different SM-native builds.

Teams Waste Days on Builds

Rebuilding kernels, debugging environments, CI failures, mismatched versions — it’s all busywork.

Our Mission

To curate a trusted catalogue of high-pain GPU libraries and become the place engineers go to get binary-clean, reproducible wheels.

We are starting by solving the immediate pain: building FlashAttention, xFormers, and Triton for the most common hardware targets. From there, we will expand to cover the entire long-tail of GPU software.

  • Automated GPU builds (wheels & source ops)
  • Global binary cache
  • Environment intelligence
⚙️ ⚡ 📦
Build → Cache → Deploy

Our Philosophy

1. Build Once. Install Anywhere.

If one developer builds it, everyone benefits via the cache.

2. Performance is a Feature.

SM-native kernels give huge wins over generic wheels. We’re not leaving performance on the table.

3. Infrastructure should be invisible.

If we do our job right, nobody thinks about WheelForge — it “just works.”

The Road Ahead

Phase 1 — Early Access (Now)
  • Working builder core
  • Runpod-powered GPU builds
  • Flash-attn prototype support
  • Cache v1
Phase 2 — Platform
  • Builder dashboard
  • Team accounts
  • Logs + environments
  • Cached builds explorer
  • Notifications
Phase 3 — Engine
  • PGO-optimized kernels
  • Custom CUDA builds
  • On-prem / VPC offerings
  • Binary provenance + SBOM

A note from the founder

Hardware fragmentation is here to stay, and traditional tooling simply hasn’t kept up.
Building AI shouldn’t require surgery on your toolchain.

We’re fixing that – one kernel at a time.

– Daniel G.

Want to follow WheelForge’s journey?

Join the waitlist to get updates and early access.