Windows 11 + AMD Radeon RX 9070 XT (gfx1201) + ROCm 7.14 + PyTorch + Voicebox

[raveformovie]

AMD ROCm 7.14 + PyTorch + Qwen-TTS + Voicebox — Windows 11 Installation Guide

Tested on: Windows 11 + RX 9070 XT (gfx1201)

This guide includes the pitfalls we actually hit during setup — not just the happy path.

---

Part 1 — Clean System Preparation

Step 1. Install the latest AMD Adrenalin driver

Download the latest WHQL/Recommended driver from AMD.

After installation, confirm in:

Device Manager → Display adapters → AMD Radeon RX 9070 XT is visible.

Step 2. Remove any old HIP SDK

If previously installed, go to:

Control Panel → Programs → Uninstall a program

Remove:

• HIP SDK
• Old ROCm SDK
• Any obsolete AMD AI SDK

Reboot.

Step 3. Disable conflicting Windows security features

Disable Windows Defender Application Guard:

Turn Windows features on/off → uncheck Microsoft Defender Application Guard

Disable Smart App Control:

Settings → Privacy & Security → Windows Security → App & Browser control → Smart App Control → Off

Reboot.

---

Part 2 — Python

Install Python 3.12.x (64-bit).

Verify:

python --version

---

Part 3 — Create a Virtual Environment

py -3.12 -m venv venv
venv\Scripts\activate
python -m pip install --upgrade pip

---

Part 4 — Install ROCm

Exactly as AMD recommends:

python -m pip install --index-url https://repo.amd.com/rocm/whl-multi-arch/ "rocm[libraries,device-gfx1201]==7.14.0"

Verify:

pip freeze

Should include:

rocm==7.14.0
rocm-sdk-core
rocm-sdk-device-gfx1201
rocm-sdk-libraries

---

Part 5 — Install PyTorch

Install AMD's architecture-specific packages:

python -m pip install --index-url https://repo.amd.com/rocm/whl-multi-arch/ ^
"torch[device-gfx1201]==2.12.0+rocm7.14.0" ^
"torchvision[device-gfx1201]==0.27.0+rocm7.14.0" ^
"torchaudio==2.11.0+rocm7.14.0"

This additionally installs:

• amd-torch-device-gfx1201
• amd-torch-device-gfx12-0
• amd-torchvision-device-gfx1201

These packages proved essential in our testing.

---

Part 6 — Verify ROCm

hipinfo

Should report:

AMD Radeon RX 9070 XT
gfx1201

Also check:

python -c "import torch; print(torch.cuda.is_available())"

Expected: True

---

Part 7 — Verify Device Detection

python -c "import torch; print(torch.cuda.device_count()); print(torch.cuda.get_device_name(0))"

• If you have only one GPU, this should show AMD Radeon RX 9070 XT.
• If you have multiple GPUs, continue to Part 8.

---

Part 8 — Multi-GPU Systems (IMPORTANT)

Many Ryzen systems expose:

• Device 0 → AMD Radeon(TM) Graphics (iGPU)
• Device 1 → AMD Radeon RX 9070 XT (dGPU)

Without intervention, PyTorch may choose the iGPU.

Create backend/.env with only this content:

HIP_VISIBLE_DEVICES=1

⚠️ Do NOT add HSA_OVERRIDE_GFX_VERSION unless ROCm fails to detect gfx1201.

⚠️ Do NOT keep AMD_SERIALIZE_KERNEL set except while actively debugging.

---

Part 9 — Load .env Before Torch Initializes

Install:

pip install python-dotenv

In backend/main.py, immediately after:

from dotenv import load_dotenv

add:

load_dotenv(
    Path(__file__).with_name(".env"),
    override=True
)

This must occur before from .app import app — otherwise Torch may initialize before the environment variables are loaded.

---

Part 10 — Verify Environment Is Actually Loaded

Temporarily add:

print(os.getenv("HIP_VISIBLE_DEVICES"))

Expected output: 1

Then remove the debug print once confirmed.

---

Part 11 — Minimal GPU Test

python -c "import os; os.environ['HIP_VISIBLE_DEVICES']='1'; import torch; print(torch.cuda.get_device_name(0))"

Expected: AMD Radeon RX 9070 XT

Kernel test:

python -c "import os; os.environ['HIP_VISIBLE_DEVICES']='1'; import torch; print(torch.randn(2,2,device='cuda'))"

Should execute without crashing.

---

Part 12 — Voicebox Test

just dev

Generate speech in the app.

Expected:

• GPU usage: 20–40%
• CPU usage: 20–40%
• VRAM increases during generation

---

Common Errors

hipErrorInvalidImage

Usually indicates: wrong PyTorch wheel, or missing amd-torch-device packages.
Solution: reinstall the architecture-specific PyTorch packages (Part 5).

"Device kernel image invalid" on torch.randn(...)

Usually indicates: wrong architecture wheel installed.

Torch detects the iGPU (AMD Radeon(TM) Graphics) instead of the dGPU

Solution: set HIP_VISIBLE_DEVICES=1 (Part 8). Only add HSA_OVERRIDE_GFX_VERSION if ROCm still fails to detect gfx1201 afterward.

AMD_SERIALIZE_KERNEL

Debug-only. Do not leave this enabled in normal use.

---

Final Working Configuration (Validated)

Component	Version
OS	Windows 11 Pro (24H2)
Python	3.12.x
GPU	AMD Radeon RX 9070 XT (gfx1201)
ROCm	7.14.0
PyTorch	2.12.0+rocm7.14.0

Installed AMD packages:

rocm
rocm-sdk-core
rocm-sdk-device-gfx1201
rocm-sdk-libraries
amd-torch-device-gfx1201
amd-torch-device-gfx12-0
amd-torchvision-device-gfx1201

Voicebox-specific:

• python-dotenv installed
• .env loaded via:

  load_dotenv(Path(__file__).with_name(".env"), override=True)

  before importing the application

.env contents:

HIP_VISIBLE_DEVICES=1