ML Environment Setup & Troubleshooting

This skill creates isolated conda environments with PyTorch. It auto-detects your hardware (NVIDIA GPU, AMD GPU, or CPU) and installs the right PyTorch build, then validates it works. Designed for consistency across a small fleet of Linux boxes (NVIDIA, AMD Strix Halo, CPU).

Creating a New ML Project

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

1. Create the project directory.
2. Run hardware detection.
3. Run setup-universal.sh from the project dir — it creates a **named conda

env** (ml-<project> by default), installs PyTorch with the correct backend, and common ML libraries. It also drops a .gitignore if one is missing.

1. Run validate.sh to confirm the GPU/backend works.
2. Show you how to activate and use it.

# From the new project directory:
bash ~/.claude/skills/ml-env/scripts/setup-universal.sh
bash ~/.claude/skills/ml-env/scripts/validate.sh

The setup script (~/.claude/skills/ml-env/scripts/setup-universal.sh):

• Detects your GPU (NVIDIA via nvidia-smi, AMD via rocminfo, else CPU).
• Creates a conda env ml-<basename-of-cwd> (Python 3.13 by default).
• Installs PyTorch 2.13.0 with the right backend (CUDA cu132, ROCm 7.2,

or CPU); gfx1151 uses its own verified tracks.

• Installs: numpy, pandas, matplotlib, scikit-learn, jupyter, tensorboard,

accelerate.

• Writes a setup log into the project dir: setup-<project>-<timestamp>.log.

Using Your Environment

conda activate ml-<project>
python -c "import torch; print(torch.__version__); print('cuda:', torch.cuda.is_available())"

Env names default to ml-<project-dir>. Override at setup time:

ENV_NAME=myexp PYTHON_VERSION=3.12 bash ~/.claude/skills/ml-env/scripts/setup-universal.sh

Where envs live: ~/.condarc (linux-dotfiles) points new envs at ~/conda/envs. Existing envs under ~/miniforge/envs are unaffected.

Version Defaults (verified 2026-07-14; prefer latest)

Overridable via environment variables — see the header of setup-universal.sh.

Hardware-Specific Guidance

NVIDIA GPUs (Ampere / Ada / Blackwell)

• CUDA cu132 wheels (CUDA 13.2). Run on driver 580.x via minor-version compat.
• Blackwell (RTX 5090, GB10, sm_120+): natively supported by PyTorch 2.13 on

cu132 — no special handling. (If an older PyTorch is forced, fall back to nightly: --pre --index-url …/whl/nightly/cu132.)

• WSL2: use the Windows NVIDIA driver only — do not install a Linux

driver inside WSL.

AMD RDNA (RX 6000/7000) and other non-gfx1151

• ROCm rocm7.2 wheels. User must be in render and video groups:

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

AMD Strix Halo (gfx1151)

⚠️ Official PyTorch wheels fail with HIP error: invalid device function. The setup script offers two verified tracks:

1. TheRock multi-arch nightly (default) — latest PyTorch (~2.12), clean index:

pip install --index-url https://rocm.nightlies.amd.com/whl-multi-arch/ "torch[device-gfx1151]" "torchvision[device-gfx1151]" torchaudio

1. AMD supported stable (ROCm 7.2.1) — AMD-validated, torch 2.9.1; **cp312

only** (recreate the env with PYTHON_VERSION=3.12). Wheels from repo.radeon.com/rocm/manylinux/rocm-rel-7.2.1/.

Environment variables: do not set HSAENABLESDMA, PYTORCHHIPALLOC_CONF, etc. globally — they cause subtle bugs. Set them only per-process for a reproduced issue. (Matches the strix-halo-setup skill v2.0.0.)

For GTT memory (amd-ttm), flash-SDPA source builds, and kernel-backport checks, use the dedicated strix-halo-setup skill (~/Projects/amdtest). ml-env intentionally does not duplicate that. See [TROUBLESHOOTING.md](TROUBLESHOOTING.md#strix-halo-gfx1151) for the summary.

CPU-only

Works everywhere. Good for development/testing before scaling to GPU.

Validating an Existing Environment

cd ~/your-ml-project   # any dir; the env name defaults to ml-<this-dir>
ENV_NAME=ml-<project> bash ~/.claude/skills/ml-env/scripts/validate.sh

Checks Python, PyTorch, backend (CUDA/ROCm), device name + compute capability / gfx arch, and runs a real matmul.

Troubleshooting (quick)

GPU not detected:

nvidia-smi                     # NVIDIA driver present?
rocminfo | grep gfx            # AMD arch (expect gfx1151 on Strix Halo)
groups | grep -E "render|video"  # AMD requires these

PyTorch installed but cuda.is_available() is False:

• NVIDIA: driver too old for the CUDA wheel (cu132 needs driver 580+; the fleet's

580.x covers it via CUDA minor-version compat). Check nvidia-smi's "CUDA Version" line, or drop to cu130/cu128.

• AMD: check groups and that you installed a gfx-appropriate build (not vanilla).

CUDA out of memory / slow training: these are general PyTorch questions — see the PyTorch docs (mixed precision, gradient checkpointing, torch.cuda.empty_cache). This skill focuses on environment setup, not training tuning.

For hardware-specific issues, WSL2, and full Strix Halo setup, see [TROUBLESHOOTING.md](TROUBLESHOOTING.md). For updating packages, see [UPDATE.md](UPDATE.md).

Scripts in This Skill

Both live in ~/.claude/skills/ml-env/scripts/:

• setup-universal.sh — detect hardware, create conda env, install PyTorch + libs.
• validate.sh — verify an existing env and exercise the GPU/CPU.

When to Use This Skill

• Creating a new ML project / PyTorch environment.
• Setting up or repairing PyTorch with GPU support.
• Troubleshooting GPU/CUDA/ROCm detection.
• Updating or maintaining an ML environment.
• Hardware-specific questions (NVIDIA, AMD, Strix Halo).