torch-builds-gfx906-rocm7.2

Prebuilt wheel packages of torch and torchvision compiled for:

• ROCm 7.2
• AMD GPU architecture: gfx906 (Vega 20 – MI50 / MI60)

This repository provides reproducible HIP-enabled builds of PyTorch specifically targeting gfx906, which is not always covered by official upstream binaries.

---

Releases

Each GitHub Release contains:

• torch-<version>+rocm7.2-gfx906-*.whl
• torchvision-<version>+rocm7.2-gfx906-*.whl

Wheel filenames encode:

• Torch version
• ROCm version (7.2)
• Target architecture (gfx906)
• Python ABI tag (e.g. cp310, cp311, cp312)
• Linux x86_64

---

Prerequisites

Hardware

• AMD GPU with architecture gfx906

  • MI50
  • MI60

Verify architecture:

rocminfo | grep gfx

Expected:

Name: gfx906

---

Software Requirements

You must have:

• Linux (tested on Ubuntu 24)
• ROCm 7.2 runtime installed
• amdgpu kernel driver properly loaded
• Compatible Python version (matching the wheel tag)
• pip >= 23.x recommended

Check ROCm installation:

rocminfo
hipcc --version

---

Important Compatibility Notes

• ROCm version must be 7.2
• Your system glibc must be compatible with the build environment
• These wheels are built without CUDA
• Do NOT mix with CUDA PyTorch builds
• Ensure no conflicting torch installation exists:

pip uninstall torch torchvision -y

---

Installation

Download the required wheels from the Releases section and install:

pip install torch-<version>+rocm7.2-gfx906-*.whl
pip install torchvision-<version>+rocm7.2-gfx906-*.whl

Or:

pip install *.whl

---

Verification

import torch

print("Torch:", torch.__version__)
print("HIP:", torch.version.hip)
print("CUDA available:", torch.cuda.is_available())
print("Device:", torch.cuda.get_device_name(0))

Expected:

• torch.version.hip is not None
• torch.cuda.is_available() → True
• Device name corresponds to AMD Vega 20

---

Build Characteristics

• HIP backend enabled
• Targeted specifically for gfx906
• Built against ROCm 7.2
• No CUDA support
• Optimized for Vega 20

---

Scope & Limitations

These builds:

• Work only for gfx906

• Are not guaranteed to work on:

  • gfx90a
  • gfx1030
  • RDNA architectures

• Require matching ROCm runtime

If you need builds for other architectures or ROCm versions, open an Issue.

---

Upstream Projects

These binaries are built from:

• PyTorch
• AMD ROCm

All original licenses apply.

---

Reporting Issues

When reporting a problem, include:

• GPU model
• Output of rocminfo
• ROCm version
• Python version
• Full traceback/log

---

Project Status

Maintained on demand. New versions are added as needed.