Workstation Scripts: Environment and GPU

Saving environment queries to files makes it easier to compare the Windows and WSL setup later. This note covers those inventories and the GPU tests for PyTorch and TensorFlow. “Workstation Command Record” covers individual installation and inspection commands; “Workstation Scripts: Offline Recovery” covers preservation and recovery.

The scripts retain their original options and arguments. Each section explains the checks and includes execution results where I have them.

1. Saving an environment record with PowerShell

The following script collects Windows development information in one file. I have no saved result from that run, so I describe how the file is created and what needs care. It assumes that D:\Lab\OfflineLab\manifests already exists and overwrites the output file if present.

@(
  "=== Windows Research Development Baseline ==="
  "Recorded: $(Get-Date -Format 'yyyy-MM-dd HH:mm:ss K')"
  ""
  "=== Git ==="
  (git --version)
  ""
  "=== VS Code ==="
  (code --version)
  ""
  "=== uv ==="
  (uv --version)
  "uv cache: $(uv cache dir)"
  ""
  "=== Python ==="
  (python --version)
  "Python path:"
  (where.exe python)
  ""
  "=== JupyterLab ==="
  "JupyterLab $(jupyter-lab --version)"
  ""
  "=== WSL ==="
  (wsl --version)
  ""
  (wsl --status)
  ""
  (wsl -l -v)
) | Out-File `
  D:\Lab\OfflineLab\manifests\development-windows-baseline.txt `
  -Encoding utf8

The backtick at a line ending continues the PowerShell command. Trailing spaces after it can break that continuation when copying. This collection script does not automatically judge each command’s success or merge every error stream into the file. Creating the file alone is not a successful environment test.

See the Out-File reference and WSL basic commands for output behaviour and WSL options.

2. Saving an extension inventory with PowerShell

The following command can save the extension list to a separate file. A separate file makes it easier to compare extensions without searching the full environment record.

code --list-extensions --show-versions |
Out-File `
  D:\Lab\OfflineLab\manifests\development-vscode-extensions.txt `
  -Encoding utf8

This command is for Windows PowerShell; I have no saved result from that run. WSL remote extensions were checked separately inside WSL; the two lists should not be treated as one installation state. The options are documented in the VS Code CLI reference.

3. Checking Linux build tools in Bash

The following commands and their subsequent output were recorded in WSL Bash.

echo "=== Build Toolchain ==="

gcc --version | head -1
g++ --version | head -1
make --version | head -1
cmake --version | head -1
ninja --version
pkg-config --version

The recorded results were GCC and G++ 15.2.0, GNU Make 4.4.1, CMake 4.2.3, Ninja 1.13.2 and pkg-config 2.5.1. These queries showed that the tools ran and returned versions. Compilation and linking tests belong to separate records.

4. Selecting directory names in Bash

The following queries can inspect research projects and VS Code Server directories in the restored environment. They target different paths; I have no individual query results from that time.

find ~/research -maxdepth 1 -mindepth 1 -type d -printf '%f\n' | sort
find ~/.vscode-server/bin -maxdepth 1 -mindepth 1 -type d -printf '%f\n'

The first query was intended to list project names; the second inspected server directory names at that location. Directory presence and functional operation still require separate checks.

5. Exporting installation and storage inventories

I exported the installation inventory with WinGet. The later file and registry queries show how to inspect the result; I have no individual results for those queries. winget export -o selects a JSON file and --include-versions includes versions. Warnings reported packages or versions unavailable from configured sources. Get-Item reads file information, Get-Content reads contents, and Get-ItemProperty reads registry properties. The two HKLM paths query installation entries; Where-Object DisplayName keeps entries with a name. Format-Table -AutoSize adjusts display columns, not structured export. The Known Folder GUID is a folder identifier from the original command, not a personal identifier.

winget list | Out-File `
  D:\Lab\OfflineLab\manifests\winget-list-productivity-baseline.txt `
  -Encoding utf8
winget export `
  -o D:\Lab\OfflineLab\manifests\winget-productivity-baseline.json `
  --include-versions

Get-Item D:\Lab\OfflineLab\manifests\winget-productivity-baseline.json
Get-Content D:\Lab\OfflineLab\manifests\winget-productivity-baseline.json
Get-AppxPackage |
Select-Object Name, Version, PackageFullName |
Sort-Object Name |
Out-File `
  D:\Lab\OfflineLab\manifests\appx-productivity-baseline.txt `
  -Encoding utf8
Get-ItemProperty `
  HKLM:\Software\Microsoft\Windows\CurrentVersion\Uninstall\*, `
  HKLM:\Software\WOW6432Node\Microsoft\Windows\CurrentVersion\Uninstall\* `
  -ErrorAction SilentlyContinue |
Where-Object DisplayName |
Select-Object DisplayName, DisplayVersion, Publisher |
Sort-Object DisplayName |
Out-File `
  D:\Lab\OfflineLab\manifests\installed-programs-productivity-baseline.txt `
  -Encoding utf8

Get-ComputerInfo | Out-File `
  D:\Lab\OfflineLab\manifests\system-baseline.txt `
  -Encoding utf8
Get-Volume | Format-Table DriveLetter,FileSystemLabel,FileSystem,Size,SizeRemaining -AutoSize |
Out-File D:\Lab\OfflineLab\manifests\storage-baseline.txt -Encoding utf8
Get-ItemProperty 'HKCU:\Software\Microsoft\Windows\CurrentVersion\Explorer\User Shell Folders' |
Select-Object Desktop, Personal, 'My Pictures', 'My Music', 'My Video', '{374DE290-123F-4565-9164-39C4925E467B}' |
Out-File D:\Lab\OfflineLab\manifests\known-folders-baseline.txt -Encoding utf8

6. Saving Windows and WSL baselines

These scripts save Windows and WSL baselines to files. The earlier development-baseline Note records saving the baseline files. I did not retain the immediate output of each script, so this section describes the saving method and what to check. Git configuration queries without a value read the setting. Output can contain personal email and user paths, so inspect files before publication. Bash { ...; } groups commands in the current shell; > saves standard output. date --iso-8601=seconds requests a timestamp to seconds, and ls -ld reports details of the directory itself.

@(
  "=== Windows Git Baseline ==="
  "Recorded: $(Get-Date -Format 'yyyy-MM-dd HH:mm:ss K')"
  ""
  "user.name=$(git config --global user.name)"
  "user.email=$(git config --global user.email)"
  "core.autocrlf=$(git config --global core.autocrlf)"
  "init.defaultBranch=$(git config --global init.defaultBranch)"
) | Out-File `
  D:\Lab\OfflineLab\manifests\development-git-config.txt `
  -Encoding utf8

{
  echo "=== WSL Research Development Baseline ==="
  echo "Recorded: $(date --iso-8601=seconds)"
  echo
  echo "=== OS ==="
  cat /etc/os-release
  echo
  echo "=== Kernel ==="
  uname -a
  echo
  echo "=== Python ==="
  python3 --version
  which python3
  echo
  echo "=== Git ==="
  git --version
  which git
  echo
  echo "user.name=$(git config --global user.name)"
  echo "user.email=$(git config --global user.email)"
  echo "core.autocrlf=$(git config --global core.autocrlf)"
  echo "init.defaultBranch=$(git config --global init.defaultBranch)"
  echo
  echo "=== Filesystem ==="
  echo "HOME=$HOME"
  echo "PWD=$PWD"
  echo
  echo "=== D: mount ==="
  ls -ld /mnt/d/Lab
} > /mnt/d/Lab/OfflineLab/manifests/development-wsl-baseline.txt

7. WSL GPU and post-reboot tests

I checked the output of the first PyTorch computation. I also recorded the post-Windows-reboot checks as passing at the time, but did not retain output for every item. python -c executes string code; python - reads code from standard input. <<'PY' passes a here-document without shell expansion. Matrices are allocated on CUDA and synchronise waits for completion. Model tests select float16, local files only and non-sampling token selection (do_sample=False); max_new_tokens=32 limits new tokens. grep -E uses extended regular expressions, uname -r reports the kernel release, test -e checks path existence, and && runs the next command after success. These are separate from the later physically disconnected tests.

python --version
which python
python -c "import torch; print('PyTorch:', torch.__version__); print('CUDA runtime:', torch.version.cuda); print('CUDA available:', torch.cuda.is_available()); print('GPU:', torch.cuda.get_device_name(0))"
python - <<'PY'
import torch

a = torch.randn((2048, 2048), device="cuda")
b = torch.randn((2048, 2048), device="cuda")
c = a @ b
torch.cuda.synchronize()

print("Shape:", c.shape)
print("Device:", c.device)
print("VS Code WSL PyTorch CUDA PASS")
PY

echo "=== OS / Kernel ==="
grep -E '^(PRETTY_NAME|VERSION_ID|VERSION_CODENAME)=' /etc/os-release
uname -r
echo
echo "=== Build Toolchain ==="
gcc --version | head -1
g++ --version | head -1
make --version | head -1
cmake --version | head -1
ninja --version
pkg-config --version
echo
echo "=== Python / uv ==="
python3 --version
uv --version
uv python list --only-installed
echo
echo "=== Jupyter Base ==="
cd ~/research/wsl-research-base
uv run jupyter-lab --version
uv run python -c "import IPython, ipykernel; print('IPython:', IPython.__version__); print('ipykernel:', ipykernel.__version__)"
echo
echo "=== GPU ==="
test -e /dev/dxg && echo "/dev/dxg: PASS"
/usr/lib/wsl/lib/nvidia-smi | head -15
echo
echo "=== VS Code Server ==="
find ~/.vscode-server/bin -maxdepth 1 -mindepth 1 -type d -printf '%f\n'
echo
echo "=== Remote Extensions ==="
code --list-extensions --show-versions | sort

cd ~/research/pytorch-smoke-test
uv run python - <<'PY'
import torch

print("PyTorch:", torch.__version__)
print("CUDA runtime:", torch.version.cuda)
print("CUDA available:", torch.cuda.is_available())
print("GPU:", torch.cuda.get_device_name(0))

a = torch.randn((2048, 2048), device="cuda")
b = torch.randn((2048, 2048), device="cuda")
c = a @ b
torch.cuda.synchronize()

print("Device:", c.device)
print("WSL PyTorch persistence PASS")
PY

cd ~/research/transformers-smoke-test
HF_HUB_CACHE=/mnt/d/Lab/Models/HuggingFace \
HF_HUB_OFFLINE=1 \
TRANSFORMERS_OFFLINE=1 \
uv run python - <<'PY'
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "Qwen/Qwen3-0.6B"

tokenizer = AutoTokenizer.from_pretrained(
    model_id,
    local_files_only=True,
)

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    dtype=torch.float16,
    local_files_only=True,
).to("cuda")

inputs = tokenizer(
    "Explain CUDA in one short sentence.",
    return_tensors="pt",
).to("cuda")

with torch.inference_mode():
    outputs = model.generate(
        **inputs,
        max_new_tokens=32,
        do_sample=False,
    )

print("Device:", next(model.parameters()).device)
print("WSL Transformers persistence PASS")
PY

8. TensorFlow execution, memory and XLA

The first block checks a basic matrix computation; I have no result for it. For the subsequent memory-growth and XLA tests, I kept both the commands and their output. Memory growth is configured before GPU initialisation; jit_compile=True requests XLA compilation. Reading c.numpy() materialises the result. PTX JIT and AutoGraph source-discovery warnings also occurred beyond these excerpts. Nightly success remains separate from stable adoption.

uv run python - <<'PY'
import tensorflow as tf

print("TensorFlow:", tf.__version__)
print("GPUs:", tf.config.list_physical_devices("GPU"))

with tf.device("/GPU:0"):
    a = tf.random.normal((2048, 2048))
    b = tf.random.normal((2048, 2048))
    c = tf.matmul(a, b)
    _ = c.numpy()

print("Shape:", c.shape)
print("Device:", c.device)
print("TensorFlow Nightly GPU execution PASS")
PY

uv run python - <<'PY'
import tensorflow as tf

gpus = tf.config.list_physical_devices("GPU")
print("TensorFlow:", tf.__version__)
print("GPUs:", gpus)

for gpu in gpus:
    tf.config.experimental.set_memory_growth(gpu, True)

print("Memory growth:",
      tf.config.experimental.get_memory_growth(gpus[0]))

with tf.device("/GPU:0"):
    a = tf.random.normal((2048, 2048))
    b = tf.random.normal((2048, 2048))
    c = tf.matmul(a, b)
    _ = c.numpy()

print("Shape:", c.shape)
print("Device:", c.device)
print("TensorFlow Nightly memory-growth GPU PASS")
PY

TensorFlow: 2.22.0-dev20260923
GPUs: [PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]
Memory growth: True
Shape: (2048, 2048)
Device: /job:localhost/replica:0/task:0/device:GPU:0
TensorFlow Nightly memory-growth GPU PASS

uv run python - <<'PY'
import tensorflow as tf

gpus = tf.config.list_physical_devices("GPU")
for gpu in gpus:
    tf.config.experimental.set_memory_growth(gpu, True)

@tf.function(jit_compile=True)
def gpu_matmul(a, b):
    return tf.matmul(a, b)

with tf.device("/GPU:0"):
    a = tf.random.normal((1024, 1024))
    b = tf.random.normal((1024, 1024))
    c = gpu_matmul(a, b)
    _ = c.numpy()

print("TensorFlow:", tf.__version__)
print("Shape:", c.shape)
print("Device:", c.device)
print("TensorFlow Nightly XLA GPU PASS")
PY

TensorFlow: 2.22.0-dev20260923
Shape: (1024, 1024)
Device: /job:localhost/replica:0/task:0/device:GPU:0
TensorFlow Nightly XLA GPU PASS

Environment inventories provide a comparison point; GPU tests record what ran under the conditions stated above. The nightly TensorFlow result remains separate from the stable path awaiting a retest.