Skip to content

AMD ROCm Setup

Jump to bottom
voidmap-bot edited this page Jun 5, 2026 · 1 revision

AMD GPU / ROCm Setup Guide

Voidmap supports AMD GPUs via PyTorch ROCm and ONNX Runtime ROCm. This guide covers installation, troubleshooting, and known limitations.

Quick Reference

GPU Generation Architecture ROCm Support Recommended Backend
RX 7900 XTX / 7900 XT RDNA 3 ✅ Full (5.7+) PyTorch ROCm 6.2
RX 7800 XT / 7700 XT RDNA 3 ✅ Full (5.7+) PyTorch ROCm 6.2
RX 6800 XT / 6900 XT RDNA 2 ✅ Full (5.4+) PyTorch ROCm 6.2
RX 6700 XT / 6800M RDNA 2 ✅ Full (5.4+) PyTorch ROCm 6.2
RX 6600 XT / 6600 RDNA 2 ✅ Full (5.4+) PyTorch ROCm 6.2
RX 5700 XT / 5700 RDNA 1 ⚠ Partial (5.4+) PyTorch ROCm 5.7
RX Vega 64 / Vega 56 GCN 5.0 ⚠ Partial PyTorch ROCm 5.4
RX 590 / 580 GCN 4.0 ❌ EOL Use ONNX CPU
Vega Cezanne iGPU GCN 5.1 ⚠ Limited Use ONNX CPU

Installation

HiveOS (Recommended for mining rigs)

ROCm is pre-installed on HiveOS. The install_hiveos.sh script auto-detects and configures it.

curl -fsSL https://raw.githubusercontent.com/thekillsquad007/voidmap/main/install_hiveos.sh | sudo bash

Ubuntu / Debian

# 1. Install ROCm
wget https://repo.radeon.com/amdgpu-install/6.2/ubuntu/jammy/amdgpu-install_6.2.60200-1_all.deb
sudo apt-get install ./amdgpu-install_6.2.60200-1_all.deb
sudo amdgpu-install --usecase=rocm

# 2. Add user to groups
sudo usermod -aG video,render $USER
newgrp video
newgrp render

# 3. Verify
rocm-smi

# 4. Install Voidmap
curl -fsSL https://raw.githubusercontent.com/thekillsquad007/voidmap/main/install_local.sh | bash

Fedora / RHEL

# 1. Install ROCm
sudo dnf install rocm ROCm

# 2. Add user to groups
sudo usermod -aG video,render $USER

# 3. Verify
rocm-smi

# 4. Install Voidmap
GPU_VENDOR=amd bash install_local.sh

Arch / Manjaro

# 1. Install ROCm
sudo pacman -S rocm-hip-sdk

# 2. Add user to groups
sudo usermod -aG video,render $USER

# 3. Verify
rocm-smi

# 4. Install Voidmap
GPU_VENDOR=amd bash install_local.sh

Critical Requirement: Python 3.10 or 3.11

PyTorch ROCm wheels are only built for Python 3.10 and 3.11.

If you have Python 3.12 or 3.13, you have two options:

Option A: Install Python 3.11 alongside

# Ubuntu
sudo apt-get install python3.11 python3.11-venv python3.11-dev

# Fedora
sudo dnf install python3.11

# Arch
sudo pacman -S python311

# Then verify
python3.11 --version

Then run the installer — it will auto-detect the right Python:

GPU_VENDOR=amd bash install_local.sh

Option B: Use ONNX CPU (slower, no Python restriction)

If you can't install Python 3.10/3.11, the miner falls back to ONNX CPU. This is slower but always works:

GPU_VENDOR=cpu bash install_local.sh

Performance: ~5-10x slower than ROCm, but still functional.

HSA_OVERRIDE_GFX_VERSION

For some AMD GPUs, you need to set HSA_OVERRIDE_GFX_VERSION because the GPU's reported GFX version isn't supported by the installed ROCm.

GPU GFX Version HSA_OVERRIDE_GFX_VERSION
RX 6800 / 6800 XT / 6900 XT 10.3.0 10.3.0
RX 6700 XT 10.3.0 10.3.0
RX 6600 / 6600 XT 10.3.0 10.3.0
RX 7900 XTX / 7900 XT 11.0.0 (default)
RX 7800 XT / 7700 XT 11.0.0 (default)
RX 5700 XT / 5700 10.1.0 10.1.0

Set the env var in your shell rc:

echo 'export HSA_OVERRIDE_GFX_VERSION=10.3.0' >> ~/.bashrc

Verifying Installation

After installation, run:

voidmap --detect

Expected output for AMD GPU:

  GPU: AMD Radeon RX 6800 XT (amd)
  Backend: pytorch-rocm
  PyTorch: 2.5.0+rocm6.2
  Platform: Linux-6.x.x

If you see "Backend: onnx-cpu" instead of "pytorch-rocm", check the troubleshooting below.

Troubleshooting

"ROCm not found" / PyTorch installed but not using GPU

Cause: PyTorch is using CPU version, not ROCm version.

Fix:

# Reinstall with ROCm
pip uninstall -y torch torchvision
pip install torch torchvision --index-url https://download.pytorch.org/whl/rocm6.2

"No HIP runtime" / "hipErrorNoBinaryForGpu"

Cause: Your GPU's GFX version isn't supported by installed ROCm.

Fix: Set HSA_OVERRIDE_GFX_VERSION:

export HSA_OVERRIDE_GFX_VERSION=10.3.0

"/dev/kfd not found"

Cause: KFD (Kernel Fusion Driver) not loaded.

Fix:

sudo modprobe amdgpu
sudo modprobe kfd

# Check
ls -la /dev/kfd

If /dev/kfd still doesn't exist, your kernel may not support KFD. Update to kernel 5.15+.

"Permission denied" on /dev/kfd

Cause: User not in video and render groups.

Fix:

sudo usermod -aG video,render $USER
newgrp video
newgrp render

# Log out and back in for changes to take effect

GPU detected but Backend = pytorch-cpu

Cause: PyTorch ROCm wheel not installed (Python version mismatch).

Fix: Check Python version:

python3 --version

If 3.12+, install Python 3.11 and reinstall:

sudo apt-get install python3.11
python3.11 -m pip install torch torchvision --index-url https://download.pytorch.org/whl/rocm6.2

"HSA status out of resources"

Cause: GPU out of memory.

Fix: Reduce batch size:

voidmap-mine --task exoplanet --batch 8

Mining crashes with "RuntimeError: HIP error"

Cause: Various ROCm issues. Often fixed by updating drivers.

Fix:

# Ubuntu
sudo amdgpu-install --usecase=rocm

# Or try the latest ROCm 6.3
wget https://repo.radeon.com/amdgpu-install/6.3/ubuntu/jammy/amdgpu-install_6.3.60300-1_all.deb

Old GPU (RX 500 series) not supported

Cause: GCN 4.0 and earlier are EOL in ROCm 5.4+.

Fix: Use ONNX CPU backend. The miner auto-falls back, but you can force it:

voidmap-mine --task exoplanet --backend onnx-cpu

Integrated GPU (Vega Cezanne) detected but not used

Cause: Integrated GPUs have limited ROCm support.

Fix: Use a discrete GPU. Or run ONNX CPU (integrated GPUs are slower than discrete even with ROCm).

Performance Comparison (RX 6800 XT)

Backend Round time Submissions/hr Notes
PyTorch ROCm 6.2 5.5s 182 Fastest
ONNX ROCm 7.2s 138 Good fallback
ONNX DirectML 10.1s 99 Windows only
ONNX CPU 35s 28 Slow but works

(12s on-chain cooldown not included in round time.)

HiveOS-Specific Notes

HiveOS uses /hive-config for configuration. To enable ROCm in HiveOS:

  1. Flight Sheet: Custom miner config:

    Miner: voidmap
URL: -
Wallet: 0xYOUR_ADDRESS
Extra args: --task exoplanet --submit --rpc https://mainnet.base.org --pk YOUR_PK

  2. HiveOS ROCm toggle: rocm on the rig config.

  3. Test: SSH into rig and run:

    rocm-smi
/opt/voidmap-miner/venv/bin/python -c "import torch; print(torch.cuda.is_available())"

    Both should return successfully.

Known Issues

Python 3.12+ no ROCm wheels

PyTorch team has not released ROCm wheels for Python 3.12+ as of mid-2026. Workarounds:

  • Use Python 3.11 (recommended)
  • Use ONNX CPU (slower)

ROCm 6.2 + RX 7900 series

ROCm 6.2 has full support for RDNA 3 (RX 7900 series). Use HSA_OVERRIDE_GFX_VERSION=11.0.0 if not auto-detected.

iGPU + dGPU hybrid systems

If your system has both an integrated GPU (Vega Cezanne) and a discrete GPU (RX 6800), the miner will use the discrete GPU by default. To force the integrated GPU (not recommended — slower):

CUDA_VISIBLE_DEVICES=0 voidmap --task exoplanet

See Also

Navigation

Quick Links

Protocol

Mining

Reference

Clone this wiki locally