[CI] Refactor CI build on GitHub

.github/workflows/build.yml

@@ -18,72 +18,78 @@ jobs:
       - name: 'Dependencies'
         run: |
           apt-get update
-          apt-get install -y git python3.9 pip cudnn9-cuda-12
+          apt-get install -y git python3.9 pip cudnn9-cuda-12 ccache
           pip install cmake==3.21.0 pybind11[global] ninja
       - name: 'Checkout'
         uses: actions/checkout@v3
         with:
           submodules: recursive
-      - name: ccache
-        uses: mozilla-actions/sccache-action@7d986dd989559c6ecdb630a3fd2557667be217ad
+      - uses: actions/cache@v4
+        with:
+          path: /root/.ccache
+          key: ccache-${{ runner.os }}-core-${{ github.ref }}
+          restore-keys: ccache-${{ runner.os }}-core-
       - name: 'Build'
-        run: NVTE_USE_CCACHE=1 NVTE_CCACHE_BIN=sccache pip install --no-build-isolation . -v
+        run: NVTE_USE_CCACHE=1 pip install --no-build-isolation . -v
         env:
           NVTE_FRAMEWORK: none
           MAX_JOBS: 1
-          SCCACHE_GHA_ENABLED: "true"
       - name: 'Sanity check'
         run: python3 -c "import transformer_engine"
         working-directory: /
   pytorch:
     name: 'PyTorch'
     runs-on: ubuntu-latest
+    container:
+      image: ubuntu:24.04
+      options: --user root
     steps:
-      - name: Move /var/lib/docker/
-        shell: bash -euxo pipefail {0}
-        run: sudo mv /var/lib/docker/ "${GITHUB_WORKSPACE}/docker"
-
-      - name: Maximize build space
-        uses: easimon/maximize-build-space@c28619d8999a147d5e09c1199f84ff6af6ad5794
-        with:
-          root-reserve-mb: 5120
-          temp-reserve-mb: 32
-          swap-size-mb: 10240
-          remove-dotnet: 'true'
-          remove-android: 'true'
-          remove-haskell: 'true'
-          remove-codeql: 'true'
-          build-mount-path: '/var/lib/docker/'
-
-      - name: Restore /var/lib/docker/
-        shell: bash -euxo pipefail {0}
-        run: sudo sh -c "mv ${GITHUB_WORKSPACE}/docker/* /var/lib/docker"
-
+      - name: 'Container dependencies'
+        run: apt-get update && apt-get install -y git python3 python3-pip ccache
       - name: 'Checkout'
         uses: actions/checkout@v3
         with:
           submodules: recursive
-
-      - name: Start named container
-        run: |
-          docker run -v $(pwd):$(pwd) -w $(pwd) --name builder -d nvcr.io/nvidia/cuda:12.8.0-devel-ubuntu22.04 sleep infinity
+      - name: "Disk space check before dependencies"
+        run: df -lh
 
       - name: 'Dependencies'
         run: |
-          docker exec builder bash -c '\
-            apt-get update && \
-            apt-get install -y git python3.9 pip cudnn9-cuda-12 && \
-            pip install cmake torch ninja pydantic importlib-metadata>=1.0 packaging pybind11 numpy einops onnxscript && \
-            apt-get clean \
-          '
+            export PIP_BREAK_SYSTEM_PACKAGES=1
+            pip install cmake ninja pybind11 numpy packaging
+            pip install torch --index-url https://download.pytorch.org/whl/cu130
+            pip install "nvidia-cuda-nvcc<13.1" "nvidia-cuda-cccl<13.1" "nvidia-cuda-crt<13.1" "nvidia-nvvm<13.1" "nvidia-cuda-profiler-api<13.1" "nvidia-nvml-dev<13.1"
+            export CUDA_PATH=$(python3 -c "import nvidia; print(list(nvidia.__path__)[0] + '/cu13')")
+            echo $CUDA_PATH
+            ln -s $CUDA_PATH/lib/libcudart.so.13 $CUDA_PATH/lib/libcudart.so
+            ln -s $CUDA_PATH/lib/libcublas.so.13 $CUDA_PATH/lib/libcublas.so
+            ln -s $CUDA_PATH/../nccl/lib/libnccl.so.2 $CUDA_PATH/../nccl/lib/libnccl.so
+            ln -s $CUDA_PATH/lib $CUDA_PATH/lib64
 
+      - name: "Disk space check after dependencies"
+        run: df -lh
+      - uses: actions/cache@v4
+        with:
+          path: /root/.ccache
+          key: ccache-${{ runner.os }}-pytorch-${{ github.ref }}
+          restore-keys: ccache-${{ runner.os }}-pytorch-
+      - name: "Disk space check after dependencies and ccache"
+        run: df -lh
       - name: 'Build'
-        run: docker exec builder bash -c 'pip install --no-build-isolation . -v --no-deps'
+        run: |
+          export CUDA_PATH=$(python3 -c "import nvidia; print(list(nvidia.__path__)[0] + '/cu13')")
+          export CUDNN_PATH=$(python3 -c "import nvidia; print(list(nvidia.__path__)[0] + '/cudnn')")
+          export PATH=$CUDA_PATH/bin:$PATH
+          export NVTE_BUILD_USE_NVIDIA_WHEELS=1
+          export PIP_BREAK_SYSTEM_PACKAGES=1
+          NVTE_USE_CCACHE=1 pip install --no-build-isolation . -v
         env:
           NVTE_FRAMEWORK: pytorch
-          MAX_JOBS: 1
+          MAX_JOBS: 2
+          CCACHE_DIR: /root/.ccache
+          CCACHE_MAXSIZE: 2G
       - name: 'Sanity check'
-        run: docker exec builder bash -c 'python3 tests/pytorch/test_sanity_import.py'
+        run: python3 tests/pytorch/test_sanity_import.py
   jax:
     name: 'JAX'
     runs-on: ubuntu-latest
@@ -92,65 +98,79 @@ jobs:
       options: --user root
     steps:
       - name: 'Dependencies'
-        run: pip install cmake==3.21.0 pybind11[global]
+        run: |
+          pip install cmake==3.21.0 pybind11[global]
+          apt-get update && apt-get install -y ccache
       - name: 'Checkout'
         uses: actions/checkout@v3
         with:
           submodules: recursive
-      - name: ccache
-        uses: mozilla-actions/sccache-action@7d986dd989559c6ecdb630a3fd2557667be217ad
+      - uses: actions/cache@v4
+        with:
+          path: /root/.ccache
+          key: ccache-${{ runner.os }}-jax-${{ github.ref }}
+          restore-keys: ccache-${{ runner.os }}-jax-
       - name: 'Build'
         run: |
-          NVTE_CCACHE_BIN=sccache NVTE_USE_CCACHE=1 pip install --no-build-isolation . -v
+          NVTE_USE_CCACHE=1 pip install --no-build-isolation . -v
         env:
           NVTE_FRAMEWORK: jax
           MAX_JOBS: 1
-          SCCACHE_GHA_ENABLED: "true"
       - name: 'Sanity check'
         run: python3 tests/jax/test_sanity_import.py
   all:
     name: 'All'
     runs-on: ubuntu-latest
+    container:
+      image: ubuntu:24.04
+      options: --user root
     steps:
-      - name: Move /var/lib/docker/
-        shell: bash -euxo pipefail {0}
-        run: sudo mv /var/lib/docker/ "${GITHUB_WORKSPACE}/docker"
-
-      - name: Maximize build space
-        uses: easimon/maximize-build-space@c28619d8999a147d5e09c1199f84ff6af6ad5794
-        with:
-          root-reserve-mb: 5120
-          temp-reserve-mb: 32
-          swap-size-mb: 10240
-          remove-dotnet: 'true'
-          remove-android: 'true'
-          remove-haskell: 'true'
-          remove-codeql: 'true'
-          build-mount-path: '/var/lib/docker/'
-
-      - name: Restore /var/lib/docker/
-        shell: bash -euxo pipefail {0}
-        run: sudo sh -c "mv ${GITHUB_WORKSPACE}/docker/* /var/lib/docker"
-
+      - name: 'Container dependencies'
+        run: apt-get update && apt-get install -y git python3 python3-pip ccache
       - name: 'Checkout'
         uses: actions/checkout@v3
         with:
           submodules: recursive
-
-      - name: Start named container
-        run: |
-          docker run -v $(pwd):$(pwd) -w $(pwd) --name builder -d ghcr.io/nvidia/jax:jax sleep infinity
+      - name: "Disk space check before dependencies"
+        run: df -lh
 
       - name: 'Dependencies'
         run: |
-          docker exec builder bash -c '\
-            pip install cmake==3.21.0 pybind11[global] einops onnxscript && \
-            pip install torch --no-cache-dir --index-url https://download.pytorch.org/whl/cu130
-          '
+            export PIP_BREAK_SYSTEM_PACKAGES=1
+            pip install cmake ninja pybind11 numpy packaging
+            pip install torch --index-url https://download.pytorch.org/whl/cu130
+            pip install jax[cuda13]
+            pip install "nvidia-cuda-nvcc<13.1" "nvidia-cuda-cccl<13.1" "nvidia-cuda-crt<13.1" "nvidia-nvvm<13.1" "nvidia-cuda-profiler-api<13.1" "nvidia-nvml-dev<13.1"
+            export CUDA_PATH=$(python3 -c "import nvidia; print(list(nvidia.__path__)[0] + '/cu13')")
+            echo $CUDA_PATH
+            ln -s $CUDA_PATH/lib/libcudart.so.13 $CUDA_PATH/lib/libcudart.so
+            ln -s $CUDA_PATH/lib/libcublas.so.13 $CUDA_PATH/lib/libcublas.so
+            ln -s $CUDA_PATH/../nccl/lib/libnccl.so.2 $CUDA_PATH/../nccl/lib/libnccl.so
+            ln -s $CUDA_PATH/lib $CUDA_PATH/lib64
+
+      - name: "Disk space check after dependencies"
+        run: df -lh
+      - uses: actions/cache@v4
+        with:
+          path: /root/.ccache
+          key: ccache-${{ runner.os }}-all-${{ github.ref }}
+          restore-keys: ccache-${{ runner.os }}-all-
+      - name: "Disk space check after dependencies and ccache"
+        run: df -lh
       - name: 'Build'
-        run: docker exec builder bash -c 'pip install --no-cache-dir --no-build-isolation . -v --no-deps'
+        run: |
+          export CUDA_PATH=$(python3 -c "import nvidia; print(list(nvidia.__path__)[0] + '/cu13')")
+          export CUDNN_PATH=$(python3 -c "import nvidia; print(list(nvidia.__path__)[0] + '/cudnn')")
+          export PATH=$CUDA_PATH/bin:$PATH
+          export NVTE_BUILD_USE_NVIDIA_WHEELS=1
+          export PIP_BREAK_SYSTEM_PACKAGES=1
+          NVTE_USE_CCACHE=1 pip install --no-build-isolation . -v
         env:
           NVTE_FRAMEWORK: all
-          MAX_JOBS: 1
-      - name: 'Sanity check'
-        run: docker exec builder bash -c 'python3 tests/pytorch/test_sanity_import.py && python3 tests/jax/test_sanity_import.py'
+          MAX_JOBS: 2
+          CCACHE_DIR: /root/.ccache
+          CCACHE_MAXSIZE: 2G
+      - name: 'Sanity check (pytorch)'
+        run: python3 tests/pytorch/test_sanity_import.py
+      - name: 'Sanity check (jax)'
+        run: python3 tests/jax/test_sanity_import.py

build_tools/jax.py

@@ -9,7 +9,12 @@
 
 import setuptools
 
-from .utils import get_cuda_include_dirs, all_files_in_dir, debug_build_enabled
+from .utils import (
+    get_cuda_include_dirs,
+    all_files_in_dir,
+    debug_build_enabled,
+    get_cuda_library_dirs,
+)
 from typing import List
 
 
@@ -92,6 +97,9 @@ def setup_jax_extension(
         ]
     )
 
+    # Library dirs
+    library_dirs = get_cuda_library_dirs()
+
     # Compile flags
     cxx_flags = ["-O3"]
     if debug_build_enabled():
@@ -109,4 +117,5 @@ def setup_jax_extension(
         include_dirs=[str(path) for path in include_dirs],
         extra_compile_args=cxx_flags,
         libraries=["nccl"],
+        library_dirs=[str(path) for path in library_dirs],
     )

build_tools/utils.py

@@ -251,6 +251,33 @@ def get_cuda_include_dirs() -> Tuple[str, str]:
     ]
 
 
+@functools.lru_cache(maxsize=None)
+def get_cuda_library_dirs() -> Tuple[str, str]:
+    """Returns the CUDA library directory."""
+
+    force_wheels = bool(int(os.getenv("NVTE_BUILD_USE_NVIDIA_WHEELS", "0")))
+    # If cuda is installed via toolkit, all necessary headers
+    # are bundled inside the top level cuda directory.
+    if not force_wheels and cuda_toolkit_include_path() is not None:
+        return []
+
+    # Use pip wheels to include all headers.
+    try:
+        import nvidia
+    except ModuleNotFoundError as e:
+        raise RuntimeError("CUDA not found.")
+
+    if nvidia.__file__ is not None:
+        cuda_root = Path(nvidia.__file__).parent
+    else:
+        cuda_root = Path(nvidia.__path__[0])  # namespace
+    return [
+        subdir / "lib"
+        for subdir in cuda_root.iterdir()
+        if subdir.is_dir() and (subdir / "lib").is_dir()
+    ]
+
+
 @functools.lru_cache(maxsize=None)
 def cuda_archs() -> str:
     archs = os.getenv("NVTE_CUDA_ARCHS")
@@ -292,6 +319,13 @@ def cuda_version() -> Tuple[int, ...]:
         version_str = get_version("nvidia-cuda-runtime-cu12")
         version_tuple = tuple(int(part) for part in version_str.split(".") if part.isdigit())
         return version_tuple
+    except:
+        pass
+
+    try:
+        version_str = get_version("nvidia-cuda-runtime")
+        version_tuple = tuple(int(part) for part in version_str.split(".") if part.isdigit())
+        return version_tuple
     except importlib.metadata.PackageNotFoundError:
         raise RuntimeError("Could neither find NVCC executable nor CUDA runtime Python package.")
 

transformer_engine/common/util/logging.h

@@ -12,10 +12,10 @@
 #include <cudnn.h>
 #include <nvrtc.h>
 
-#include "nccl.h"
-
 #ifdef NVTE_WITH_CUBLASMP
 #include <cublasmp.h>
+
+#include "nccl.h"
 #endif  // NVTE_WITH_CUBLASMP
 
 #include <iostream>

transformer_engine/jax/csrc/extensions/cgemm_helper.h

@@ -7,6 +7,7 @@
 #ifndef TRANSFORMER_ENGINE_JAX_CGEMM_HELPER_H_
 #define TRANSFORMER_ENGINE_JAX_CGEMM_HELPER_H_
 
+#include <nccl.h>
 #include <unistd.h>
 
 #include <chrono>