---

name: ml-env description: Set up ML environments with PyTorch and auto-detect hardware. Use this when creating new ML projects, setting up PyTorch, or troubleshooting GPU/environment issues. Guides you through creating isolated project environments with hardware-specific PyTorch builds (NVIDIA/AMD/CPU). allowed-tools: Read, Bash, WebFetch activation-precedence: high

ML Environment Setup & Troubleshooting

This skill helps you create and manage isolated ML environments with PyTorch. It auto-detects your hardware (NVIDIA GPU, AMD GPU, or CPU) and installs the appropriate PyTorch build.

Creating a New ML Project

I can help you set up a complete ML project with PyTorch in seconds. Here's what I'll do:

1. Create your project directory
2. Create a .gitignore for ML files
3. Run hardware detection
4. Install PyTorch with the right backend
5. Install ML libraries
6. Validate everything works

To get started, tell me:

• Project name/path where you want it created
• I'll handle the rest!

Interactive Setup Process

When you ask me to set up a new ML project, I will:

# 1. Create the project directory
mkdir -p ~/projects/my-ml-project

# 2. Create .gitignore
# (ignores ml-env/, data/, models/, logs/, etc.)

# 3. Run the setup script
bash ~/.claude/skills/ml-env/scripts/setup-universal.sh

# 4. Show you the results

The setup script will:

• Detect your GPU (NVIDIA with nvidia-smi, AMD with rocminfo, or fallback to CPU)
• Ask questions for special hardware (Blackwell GPUs, Strix Halo, etc.)
• Install Python 3.13 virtual environment with uv
• Install PyTorch 2.10.0 with correct backend
• Install ML libraries: numpy, pandas, scikit-learn, jupyter, accelerate, etc.
• Create ml-env/ directory in your project
• Optionally initialize git

After Setup: Using Your Environment

Once created, activating is simple:

cd ~/projects/my-ml-project
source ml-env/bin/activate  # Regular environments
# OR
source ml-env/activate-safe.sh  # If you use conda (ignores conda settings)

Check it works:

python -c "import torch; print(f'PyTorch: {torch.__version__}'); print(f'CUDA: {torch.cuda.is_available()}')"

Hardware-Specific Guidance

NVIDIA GPUs

• Supported: RTX 3090, 4090, 5090, and most Ampere/Ada/Blackwell
• Installation: CUDA 12.8 (stable) or CUDA 13.0 (for RTX 5090)
• Driver requirement: 520+ for CUDA 12.8, 550+ for CUDA 13.0
• WSL2 users: Use Windows NVIDIA driver only (do NOT install Linux driver)

AMD RDNA (RX 6000/7000 series)

• Installation: ROCm 6.2
• Requirements: User in render and video groups
• Setup: sudo usermod -aG render,video $USER && newgrp render

AMD Strix Halo (gfx1151)

⚠️ This requires special handling - official PyTorch wheels do NOT work!

• GPU: Ryzen AI MAX+ 395 with gfx1151
• Critical issue: Official PyTorch wheels fail with "HIP error: invalid device function"
• Solution: Use AMD gfx1151-specific builds
• ROCm 7 (Recommended): https://repo.amd.com/rocm/whl/gfx1151/ - ~31 TFLOPS BF16
• ROCm 6.4.4+ (Fallback): https://rocm.nightlies.amd.com/v2/gfx1151/ - ~12 TFLOPS BF16
• Memory limits: Default ~33GB; configure GTT for larger models (30B+)
• Setup requires: User in render and video groups, Linux kernel 6.14+ (6.16.9+ recommended for automatic UMA/GTT behavior)

Reference project: See ~/Projects/amdtest for a working gfx1151 setup example.

See TROUBLESHOOTING.md for complete Strix Halo setup and GTT memory configuration.

CPU-Only Systems

• Works everywhere
• Good for development/testing
• Use for learning before scaling to GPU

Current Versions (2026)

• PyTorch: 2.10.0
• Python: 3.13 (or 3.12 if needed)
• CUDA: 12.8 (main), 13.0 (Blackwell experimental)
• ROCm: 6.2 (RDNA), 7.x preferred for Strix Halo (6.4.4+ as fallback)
• Key ML libs: numpy, pandas, matplotlib, scikit-learn, jupyter, accelerate, tensorboard

Validating an Existing Environment

If you already have a project and want to verify it works:

cd ~/your-ml-project
bash ~/.claude/skills/ml-env/scripts/validate.sh

This will check:

• Python and PyTorch versions
• GPU/CPU backend detection
• GPU memory and specifications
• Computation tests

Troubleshooting

GPU Not Detected

NVIDIA:

nvidia-smi  # Check driver is installed
python -c "import torch; print(torch.cuda.is_available())"

AMD:

rocm-smi  # Check ROCm installation
rocminfo | grep gfx  # Check GPU architecture

CUDA Out of Memory

1. Reduce batch size
2. Enable mixed precision training
3. Use torch.cuda.empty_cache() between batches
4. Try gradient accumulation

PyTorch Not Finding GPU After Install

1. Activate the environment: source ml-env/bin/activate
2. Check driver version
3. Reinstall PyTorch with correct index URL:

   uv pip install --upgrade torch --index-url https://download.pytorch.org/whl/cu128
   

Strix Halo Specific Issues

See TROUBLESHOOTING.md for detailed Strix Halo troubleshooting, GTT memory setup, and performance optimization.

Best Practices

1. Always activate first: Before running any Python/ML code
2. Use virtual environments: Never install to system Python
3. Move models to device: Explicitly move tensors to GPU
4. Monitor memory: Keep an eye on GPU memory usage
5. Test on CPU first: Develop with small data on CPU, scale to GPU
6. Save checkpoints: Don't train for hours without saving progress

Common Workflows

Training a Model

import torch

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = YourModel().to(device)

for batch in dataloader:
    x, y = batch
    x, y = x.to(device), y.to(device)
    loss = model(x, y)
    loss.backward()

Mixed Precision Training (Faster, Less Memory)

from torch.cuda.amp import autocast, GradScaler

scaler = GradScaler()

for batch in dataloader:
    with autocast():
        loss = model(x, y)
    scaler.scale(loss).backward()
    scaler.step(optimizer)
    scaler.update()

Checking GPU Memory

import torch
if torch.cuda.is_available():
    print(f"Allocated: {torch.cuda.memory_allocated()/1e9:.2f}GB")
    print(f"Reserved: {torch.cuda.memory_reserved()/1e9:.2f}GB")
    print(f"Total: {torch.cuda.get_device_properties(0).total_memory/1e9:.2f}GB")

Reference Documentation

• TROUBLESHOOTING.md - Common issues, hardware-specific setup (especially Strix Halo)
• UPDATE.md - Updating PyTorch and dependencies

Scripts in This Skill

All scripts are in ~/.claude/skills/ml-env/scripts/:

• setup-universal.sh - Hardware detection and PyTorch installation (used during initial setup)
• validate.sh - Validate an existing environment and test GPU/CPU

When to Use This Skill

Use me when you:

• Want to create a new ML project
• Need to set up PyTorch with GPU support
• Are troubleshooting GPU/CUDA/ROCm issues
• Want to update or maintain your ML environment
• Have hardware-specific questions (NVIDIA, AMD, Strix Halo)
• Need guidance on ML best practices

Questions?

Ask me anything about:

• Creating new ML projects
• Hardware setup and troubleshooting
• PyTorch installation
• GPU/CPU configuration
• ML best practices
• Updating packages