Building PyTorch from Source with CUDA 13.3 + cuDNN 9.22 on Arch Linux (Artix) (RTX 3050)

A battle-tested, step-by-step manual for compiling PyTorch 2.12.0 from source against CUDA 13.3 and cuDNN 9.22 on Arch Linux (and Artix / other derivatives), targeting an NVIDIA RTX 3050 (compute capability 8.6).

This repo exists because the "happy path" does not work on a bleeding-edge Arch system. The stock Arch cuda package (13.2 at the time of writing) triggers a double free detected in tcache 2 crash even on a trivial nvcc-compiled C program, and glibc 2.43 cannot be downgraded without breaking the whole system. The route that actually works is documented here end-to-end.

⚠️ Read this first. The exact version numbers below (driver 610.43.02, CUDA 13.3.0, cuDNN 9.22.0.52, glibc 2.43, gcc 15, PyTorch v2.12.0) are the combination that worked on one specific machine at one point in time. Arch moves fast. Always re-check current version compatibility before you start — see docs/01-prerequisites-and-compatibility.md.

---

TL;DR

# 1. System build dependencies
./scripts/00-system-deps.sh

# 2. NVIDIA open beta driver (fixes glibc 2.43 incompatibility)
./scripts/01-beta-drivers.sh

# 3. CUDA 13.3 via the official runfile (NOT the Arch package)
./scripts/02-install-cuda-13.3.sh

# 4. cuDNN 9.22 (manual copy into the CUDA tree)
./scripts/03-install-cudnn.sh

# 5. Python 3.12.3 via pyenv + a clean virtualenv
./scripts/04-pyenv-setup.sh

# 6. Load the build environment and compile PyTorch 2.12.0
source ./scripts/exports.sh
./scripts/05-build-pytorch.sh

# 7. Verify
./scripts/test-pytorch.sh

Each step is explained in detail in docs/. Do not blindly run the scripts — several steps involve downloading multi-GB installers, replacing your GPU driver, and a compile that can take an hour and eat all your RAM. Read the matching doc page for each script first.

---

Target system

Component	Value
OS	Arch Linux / Artix (rolling)
GPU	NVIDIA RTX 3050 (Ampere, compute capability 8.6)
Driver	nvidia-open-beta 610.43.02
CUDA Toolkit	13.3.0 (runfile install to /usr/local/cuda-13.3)
cuDNN	9.22.0.52 for CUDA 13
glibc	2.43
Host compiler	gcc-15 / g++-15
Python	3.12.3 (via pyenv)
PyTorch	v2.12.0 (built from source)

---

Why build from source at all?

For this exact CUDA 13.3 + RTX 3050 + Arch stack there is no matching prebuilt PyTorch wheel. Official wheels lag behind the newest CUDA toolkits, and mixing a prebuilt wheel's bundled CUDA with a hand-installed system CUDA is a common source of subtle runtime crashes. Building from source guarantees PyTorch links against your toolkit, your cuDNN, and is compiled with a host compiler CUDA 13.3 actually accepts.

---

Repository layout

.
├── README.md                     ← you are here
├── docs/
│   ├── 00-overview.md            The full story: what broke and why
│   ├── 01-prerequisites-and-compatibility.md
│   ├── 02-nvidia-beta-drivers.md
│   ├── 03-install-cuda-13.3.md
│   ├── 04-install-cudnn.md
│   ├── 05-pyenv-python-setup.md
│   ├── 06-build-pytorch.md
│   ├── 07-verification.md
│   ├── 08-troubleshooting.md
│   └── 09-environment-variables.md
├── scripts/
│   ├── 00-system-deps.sh
│   ├── 01-beta-drivers.sh
│   ├── 02-install-cuda-13.3.sh
│   ├── 03-install-cudnn.sh
│   ├── 04-pyenv-setup.sh
│   ├── 05-build-pytorch.sh
│   ├── exports.sh                Source this before building
│   ├── test-cuda.sh             Standalone nvcc smoke test
│   └── test-pytorch.sh          PyTorch/CUDA/cuDNN verification
├── test/
│   └── test.cu                   Minimal CUDA "does the GPU exist" program
├── .gitignore
└── LICENSE

---

Recommended reading order

1. Overview — what broke and why
2. Prerequisites & compatibility
3. NVIDIA beta drivers
4. Install CUDA 13.3
5. Install cuDNN 9.22
6. pyenv & Python setup
7. Build PyTorch 2.12.0
8. Verification
9. Troubleshooting
10. Environment variables explained

---

Adapting this for a different GPU

The only GPU-specific knob is the compute capability. This guide uses 8.6 for the RTX 3050. If your card differs, change TORCH_CUDA_ARCH_LIST in scripts/exports.sh:

GPU family	Example cards	TORCH_CUDA_ARCH_LIST
Ampere	RTX 3050 / 3060 / 3090	8.6
Ada Lovelace	RTX 4070 / 4090	8.9
Hopper	H100	9.0
Blackwell	RTX 5090	12.0

Look up your card's compute capability at https://developer.nvidia.com/cuda-gpus before changing this.

---

Disclaimer

This is a community field-guide, not official NVIDIA or PyTorch documentation. Replacing GPU drivers and installing CUDA via runfile carries real risk of leaving your system without working graphics. Have a way to reach a TTY / recover before you begin. Nothing here is guaranteed to work on a future Arch snapshot.