PyTorch on Windows (ROCm on AMD Radeon 780M gfx1103)

Current Windows Support Status

Key Resources:

• GPU Compatibility: Check SUPPORTED_GPUS.md for the latest list of verified hardware.
• Component Support Status: Refer to windows_support.md for a detailed breakdown of ROCm components.
• GFX1103 Progress: Track the latest updates for Radeon 780M and other consumer GPUs in TheRock Issue #1337.

Support Summary:

• Supported: Core math libraries (rocBLAS, rocRAND, rocFFT, rocSOLVER, rocSPARSE), ML libraries (MIOpen, hipDNN), and the AMD-LLVM compiler toolchain are generally functional.
• Unsupported/Limited: Profiling tools (rocprofiler-sdk, aqlprofile), communication libraries (RCCL), and media decoding (rocDecode, rocJPEG) are currently unsupported or restricted on Windows. System-level tools like amdsmi and rocr-runtime are also pending full support.

Prerequisites

1. Install the latest Adrenaline driver.
2. Read TheRock release guidance (Windows and compatibility notes): https://github.com/ROCm/TheRock/blob/main/RELEASES.md.

ROCm installation paths

Support three ways to install ROCm:

Option 1 (recommended for most users): pip + nightlies

uv pip install --index-url https://rocm.nightlies.amd.com/v2/gfx110X-all/ "rocm[libraries,devel]"

Option 2 (recommended for reproducibility): artifacts download

1. Clone TheRock repository
2. Use the artifact installation helper:
   • TheRock\build_tools\install_rocm_from_artifacts.py
3. Supported channels:
   • dev (recommended)
   • nightly
4. Optional: target a specific GitHub Actions run ID for fixed behavior. You can install specific versions verified for certain device models. Example reference run:
   • https://github.com/ROCm/TheRock/actions/runs/23688149485/job/69056949871

Check more in Artifact install docs

Option 3 (source build, least recommended)

Source builds require ~100GB disk, tons of hours, should be equivalent to artifacts in code path. Use only if you need custom build control.

Prerequisites and Set up Environment

• Confirm prerequisites in RELEASES.md
• Reinstall Git with "Use Git and optional Unix tools from the Windows Command Prompt".
• Set these git options:
  • git config --global core.symlinks true
  • git config --global core.longpaths true
  • git config --global core.autocrlf true
• Clone https://github.com/ROCm/TheRock.
• Validate environment using:
  • .\TheRock\build_tools\validate_windows_install.ps1

Build flow

• uv pip install -r requirements.txt. If venv is not activated, activate it first.
• uv run python ./build_tools/fetch_sources.py. (Takes a long time, handles many files).
• TheRock\build_tools\setup_ccache.py (Optional).
• Build: cmake -B build -GNinja . -DTHEROCK_AMDGPU_FAMILIES=gfx110X-all

Additional docs:

• https://github.com/ROCm/TheRock/blob/main/docs/development/README.md
• https://github.com/ROCm/TheRock/blob/main/docs/development/windows_support.md

PyTorch installation

This step usually has no issues. Currently supports Python 3.10-3.12 for Torch 2.9 & 2.10.

Recommended:

uv pip install --index-url https://rocm.nightlies.amd.com/v2/gfx110X-all/ torch torchaudio torchvision

Fixing Known issue: HIP API error 0100 on iGPU mapping

checkHipErrors() HIP API error = 0100 "no ROCm-capable device is detected"

Observed GPU enumeration issue on Windows

TheRock issues confirm that Windows iGPU + discrete GPU enumeration is still under active work:

• AMD iGPU is cuda:0 when using ROCm wheels built by TheRock
  • AMD developer confirms Windows device enumeration priority is buggy and can skip the iGPU or conflict with dGPU.
  • Standard workaround: set HIP_VISIBLE_DEVICES=0 explicitly.
• Confused about current status of support for Radeon 780M (gfx1103) on Windows
  • Confirms gfx1103 support in TheRock nightlies is turned on, but some libraries are not yet fully bulletproof; first-load or op compilation hangs are reported.
• Hangs/fails on Radeon 780M (gfx1103)
  • Multiple iGPU users report silent lockups without explicit visible device, or memory scheduling issues.
  • See more on Issue #1184

Workaround (for current session)

To force the HIP runtime to recognize the iGPU, you must explicitly set the HIP_VISIBLE_DEVICES environment variable to 0.

In PowerShell (Current Session):

$env:HIP_VISIBLE_DEVICES = "0"

In cmd.exe (Current Session):

set HIP_VISIBLE_DEVICES=0

To make permanent, add HIP_VISIBLE_DEVICES user environment variable value 0 in Windows Settings.

More information can be found in GitHub issues.

Final Verification and Testing

import torch
import time

print("ROCm:", torch.version.hip)
print("GPU available:", torch.cuda.is_available())

if torch.cuda.is_available():
    print(torch.cuda.get_device_name(0))

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print(device)

# Create deterministic tensors
tensor_a_cpu = torch.full((1000, 1000), 2.0, device='cpu')
tensor_b_cpu = torch.full((1000, 1000), 3.0, device='cpu')

# CPU computation
start_time = time.time()
result_cpu = tensor_a_cpu + tensor_b_cpu
cpu_time = time.time() - start_time
print(f"CPU operation took: {cpu_time:.6f} seconds")

# GPU computation
if torch.cuda.is_available():
    tensor_a_gpu = tensor_a_cpu.to('cuda')
    tensor_b_gpu = tensor_b_cpu.to('cuda')

    torch.cuda.synchronize()  # ensure accurate timing
    start_time = time.time()

    result_gpu = tensor_a_gpu + tensor_b_gpu

    torch.cuda.synchronize()  # wait for GPU to finish
    gpu_time = time.time() - start_time
    print(f"GPU operation took: {gpu_time:.6f} seconds")

    # Move GPU result back to CPU for comparison
    result_gpu_cpu = result_gpu.to('cpu')

    # Verify correctness
    if torch.allclose(result_cpu, result_gpu_cpu):
        print("CPU and GPU results match!")
    else:
        print("Results differ!")

Sample output

CPU operation took: 0.004181 seconds
GPU operation took: 0.001731 seconds
CPU and GPU results match!