igamenovoer/pixi-make-cu-build-env

Pixi Make CUDA Build Env

Trigger

Use this skill when the user asks to:

• "Setup a CUDA build environment in this project"
• "Prepare the current Pixi project for compiling .cu files"
• "Add CUDA compilation support to <project_path>"

Workflow

1. Requirements Gathering

Mandatory: You MUST identify:

1. Project Context: The existing Pixi project path. (Default to current working directory if valid).

2. Manifest File: Identify if the project uses pixi.toml or pyproject.toml.

3. Target Environment Name: "Which Pixi environment should this be set up in?"

   • Default: If not specified, assume the default environment.
   • Named: If the user provides a name (e.g., cuda-12), use that feature.

4. CUDA Version: "Which CUDA version do you need?" (e.g., 11.8, 12.1).

5. Host Tools: Check if cmake, ninja, and a C++ compiler are available and suitable.

   • Action: Run cmake --version, ninja --version, and check for gcc/clang (Linux) or cl (Windows).
   • Decision: If missing or old, include them in the Pixi environment.

6. Tools: "Do you need profiling/debugging tools like nsight-compute?"

2. Conflict Detection & Resolution

Before modifying the project, verify the current state of <MANIFEST_FILE>.

• Check: Are there existing CUDA libs (cuda-toolkit, cudnn)? Is the nvidia channel prioritized? Are versions conflicting?
• Conflict Logic: If conflicts are detected (e.g., conda-forge prioritized, incompatible pins):
  • Autonomous Check: Did the user say "adjust settings", "proceed without asking", or "do what you have to"?
    • Yes: Automatically proceed with Option 1 (New Env) for safety, or Option 2 (Force) if the user explicitly asked to fix this environment.
  • Manual Resolution: If no authorization, STOP and offer:
    1. Create New Environment (Recommended): Create a separate environment (e.g., cu<VERSION>-build) with its own solve-group to isolate dependencies.
    2. Force Adjust (Risky): Reorder channels and overwrite pins in the target environment. Warn: This may break existing code.
    3. Cancel: Abort the operation.
    4. Custom: Ask the user for specific instructions on how to handle the conflict.

3. Adding Dependencies (Based on Selection)

Execute the workflow corresponding to the user's choice.

Option 1: New Isolated Environment (Recommended)

Create a new feature and environment with a dedicated solve-group.

# 1. Add Feature with Dependencies (Pin channel to nvidia)
pixi add --manifest-path <MANIFEST_FILE> --feature <NEW_ENV_NAME> --channel nvidia \
    cmake ninja cxx-compiler make pkg-config \
    cuda-toolkit=<VERSION> cuda-nvcc=<VERSION> \
    [cudnn nccl nsight-compute]

# 2. Create Environment with Dedicated Solve-Group
# (Use 'pixi workspace environment add' or edit manifest manually if CLI lacks support)
# TOML equivalent:
# [tool.pixi.environments]
# <NEW_ENV_NAME> = { features = ["<NEW_ENV_NAME>"], solve-group = "<NEW_ENV_NAME>" }

Option 2: Force Adjust Existing

Best Practice: Always add the nvidia channel to the project configuration before installing packages.

Command Logic:

• Default Environment: Do NOT use --feature.
• Named Environment: Use --feature <ENV_NAME>.

# 1. Setup Channels (Prioritize NVIDIA)
pixi project channel add nvidia --prepend --manifest-path <PROJECT_PATH>/<MANIFEST_FILE>

# 2. Core Build Tools
pixi add --manifest-path <PROJECT_PATH>/<MANIFEST_FILE> [--feature <ENV_NAME>] cmake ninja cxx-compiler make pkg-config

# 3. CUDA Toolchain
pixi add --manifest-path <PROJECT_PATH>/<MANIFEST_FILE> [--feature <ENV_NAME>] cuda-toolkit=<VERSION> cuda-nvcc=<VERSION>

Optional Extras/Tools (Only if requested):

pixi add --manifest-path <PROJECT_PATH>/<MANIFEST_FILE> [--feature <ENV_NAME>] cudnn=<VERSION> nccl=<VERSION> nsight-compute

4. Configuring Build Tasks

Add a standard CMake build task to <MANIFEST_FILE>.

• Note: Tasks typically run in the activated environment. Ensure CMAKE_CUDA_COMPILER points to the Pixi-managed nvcc ($CONDA_PREFIX/bin/nvcc).

Task Definition:

[tool.pixi.tasks]
configure = { cmd = "cmake -G Ninja -S . -B build -DCMAKE_CUDA_COMPILER=$CONDA_PREFIX/bin/nvcc -DCUDAToolkit_ROOT=$CONDA_PREFIX", env = { CUDACXX = "$CONDA_PREFIX/bin/nvcc" } }
build = "cmake --build build"
test = "ctest --test-dir build"

5. Verification (Automated)

To verify the setup, deploy the self-contained test suite from the skill's resource directory.

1. Deploy Test Suite: Check if the project has a tmp/ directory. If so, place the verification project there to keep the root clean.

# Determine verification path
TARGET_DIR="build-check"
if [ -d "tmp" ]; then TARGET_DIR="tmp/build-check"; fi

# Deploy
mkdir -p "$TARGET_DIR"
cp -r magic-context/skills/tools/pixi/pixi-make-cu-build-env/demo-code/* "$TARGET_DIR"/
chmod +x "$TARGET_DIR/build-and-run.sh"

2. Execute: Run the script using the configured environment. The script automatically handles source path detection.

pixi run --manifest-path <MANIFEST_FILE> [--feature <ENV_NAME>] "$TARGET_DIR/build-and-run.sh"

• Note: If the host has no GPU or an incompatible driver, the build step should still succeed, but the binary execution will fail. In this case, verify that $TARGET_DIR/build/check_app exists to confirm the compilation setup.

Troubleshooting

CMake Errors

• Source Directory: The provided build-and-run.sh handles directory context automatically. If you see errors about missing CMakeLists.txt, ensure you haven't moved the script out of the verification directory.
• Pthread Failure: Performing Test CMAKE_HAVE_LIBC_PTHREAD - Failed is often normal; CMake usually finds pthread in the next step.
• Architecture Warnings: nvcc warning : Support for offline compilation... indicates the default architecture might be older. You can safely ignore this warning for verification purposes.

Runtime Errors

• Driver Mismatch: If the kernel fails to launch, check nvidia-smi. The host driver must support the installed CUDA Toolkit version.

Best Practices

• Never mix channels for CUDA: Stick to nvidia for cuda-* packages.
• Explicit Compilation: Always prefer passing -DCMAKE_CUDA_COMPILER over relying on implicit path resolution, as CMake might pick up /usr/bin/nvcc.
• User Space: Remind the user that this setup requires NO sudo and works on any Linux machine with a driver.