Fix CI build and improve parallelism (#47)

Summary:
This PR introduces two main improvements to the CI workflow:

1.  **Run Jobs in Parallel:** The `test` job no longer waits for the `format-check` job to complete. By removing the `needs` dependency in `.github/workflows/test.yml`, both jobs can run concurrently, which helps to speed up the overall CI pipeline.

2.  **Fix Broken cuDNN Installation:** The CI was failing during the environment setup because the `install_cudnn.sh` script from the PyTorch repository was removed. This has been fixed by:
    *   Using the newer `install_cuda.sh` script which contains the necessary `install_cudnn` function.
    *   Correctly calling the `install_cudnn` function with the required `CUDA_VERSION` and `CUDNN_VERSION` arguments.

These changes make the CI process both faster and more reliable.

Pull Request resolved: #47

Reviewed By: adamomainz

Differential Revision: D79178763

Pulled By: FindHao

fbshipit-source-id: cdd4aaf4c0cc1ea7715f99d27b532d18202950af

Author: FindHao

.ci/setup.sh

@@ -87,14 +87,13 @@ else
     HAS_CORRECT_CUDA=false
 fi
 
-if [ "$HAS_CORRECT_CUDA" = "true" ]; then
-    echo "🔧 Installing development libraries only..."
-    # Install other dev libraries but skip CUDA toolkit
-    sudo apt-get install -y libstdc++6 libstdc++-12-dev libffi-dev libncurses-dev zlib1g-dev libxml2-dev git build-essential
-else
-    echo "📦 Installing CUDA $CUDA_VERSION_REQUIRED and development libraries..."
+echo "🔧 Installing development libraries"
+sudo apt-get install -y libstdc++6 libstdc++-12-dev libffi-dev libncurses-dev zlib1g-dev libxml2-dev git build-essential cmake bc gdb curl wget
+
+if [ "$HAS_CORRECT_CUDA" != "true" ]; then
+    echo "📦 Installing CUDA $CUDA_VERSION_REQUIRED"
     # Install all packages including CUDA toolkit (this is the big download)
-    sudo apt-get install -y cuda-toolkit-12.8 libstdc++6 libstdc++-12-dev libffi-dev libncurses-dev zlib1g-dev libxml2-dev git build-essential
+    sudo apt-get install -y cuda-toolkit-12.8
 fi
 
 # Verify clang installation
@@ -170,7 +169,14 @@ echo "Using cuDNN version: $CUDNN_VERSION"
 
 # Install cuDNN using PyTorch's script
 echo "Installing cuDNN using PyTorch's script..."
-curl -s https://raw.githubusercontent.com/pytorch/pytorch/main/.ci/docker/common/install_cudnn.sh | sudo bash
+curl -s https://raw.githubusercontent.com/pytorch/pytorch/main/.ci/docker/common/install_cuda.sh -o /tmp/install_cuda.sh
+chmod +x /tmp/install_cuda.sh
+# The install_cudnn function is defined in install_cuda.sh.
+# We source the script and call the function with sudo to install cuDNN.
+# The -E flag preserves the environment variables. The function expects
+# CUDA major version (e.g., "12") and CUDNN version as arguments.
+CUDA_MAJOR_VERSION="${CUDA_VERSION%%.*}"
+sudo -E bash -c "source /tmp/install_cuda.sh && install_cudnn \"${CUDA_MAJOR_VERSION}\" \"${CUDNN_VERSION}\""
 
 # Install PyTorch nightly
 echo "Installing PyTorch nightly..."

.github/workflows/test.yml

@@ -59,7 +59,6 @@ jobs:
     test:
         runs-on: 4-core-ubuntu-gpu-t4
         timeout-minutes: 120
-        needs: format-check
         steps:
             - uses: actions/checkout@v4
 

Makefile

@@ -55,5 +55,4 @@ clean:
 
 install-dev:
 	@echo "Installing development dependencies..."
-	pip install -e ".[test]"
-	pip install black usort ruff
+	pip install black usort ruff coverage

README.md

@@ -71,6 +71,8 @@ pip install -e .
 
 **Prerequisites:** Python ≥ 3.10, Triton > 3.3.1 ([install from source](https://github.com/triton-lang/triton)), GPU required (NVIDIA/AMD)
 
+TritonParse relies on new features in Triton > 3.3.1. Please install Triton from source for now.
+
 ## 📚 Complete Documentation
 
 | 📖 Guide | Description |

pyproject.toml

@@ -15,6 +15,7 @@ test = [
     "coverage>=7.0.0",
 ]
 
+
 [tool.setuptools.packages.find]
 include = ["tritonparse*"]