ci: Fixing sccache behavior (#4866)

Signed-off-by: Dillon Cullinan <dcullinan@nvidia.com> Signed-off-by: Anant Sharma <anants@nvidia.com> Co-authored-by: Anant Sharma <anants@nvidia.com>

ci: Fixing sccache behavior (#4866)
Signed-off-by: Dillon Cullinan <dcullinan@nvidia.com> Signed-off-by: Anant Sharma <anants@nvidia.com> Co-authored-by: Anant Sharma <anants@nvidia.com>
242a4d5b · Dillon Cullinan · GitHub · 9b7152d7 · 242a4d5b · 242a4d5b
Unverified Commit 242a4d5b authored Dec 11, 2025 by Dillon Cullinan Committed by GitHub Dec 11, 2025
5 changed files
--- a/.github/actions/docker-build/action.yml
+++ b/.github/actions/docker-build/action.yml
@@ -143,7 +143,9 @@ runs:
          --vllm-max-jobs 10 \
          --framework ${{ inputs.framework }} \
          --platform ${{ inputs.platform }} \
-          $EXTRA_ARGS 2>&1 | tee "${BUILD_LOG_FILE}"
+          --use-sccache \
+          --sccache-bucket "$SCCACHE_S3_BUCKET" \
+          --sccache-region "$AWS_DEFAULT_REGION" $EXTRA_ARGS 2>&1 | tee "${BUILD_LOG_FILE}"
        BUILD_EXIT_CODE=${PIPESTATUS[0]}

--- a/container/Dockerfile
+++ b/container/Dockerfile
@@ -249,8 +249,11 @@ RUN --mount=type=secret,id=aws-key-id,env=AWS_ACCESS_KEY_ID \
 RUN --mount=type=secret,id=aws-key-id,env=AWS_ACCESS_KEY_ID \
    --mount=type=secret,id=aws-secret-id,env=AWS_SECRET_ACCESS_KEY \
    export SCCACHE_S3_KEY_PREFIX="${SCCACHE_S3_KEY_PREFIX:-${ARCH}}" && \
-    export CC=$(which gcc) && \
+    if [ "$USE_SCCACHE" = "true" ]; then \
-    export CXX=$(which g++) && \
+        export CMAKE_C_COMPILER_LAUNCHER="sccache" && \
+        export CMAKE_CXX_COMPILER_LAUNCHER="sccache" && \
+        export CMAKE_CUDA_COMPILER_LAUNCHER="sccache"; \
+    fi && \
    source ${VIRTUAL_ENV}/bin/activate && \
    git clone --depth 1 --branch ${NIXL_REF} "https://github.com/ai-dynamo/nixl.git" && \
    cd nixl && \
@@ -276,8 +279,11 @@ RUN echo "$NIXL_LIB_DIR" > /etc/ld.so.conf.d/nixl.conf && \
 RUN --mount=type=secret,id=aws-key-id,env=AWS_ACCESS_KEY_ID \
    --mount=type=secret,id=aws-secret-id,env=AWS_SECRET_ACCESS_KEY \
    export SCCACHE_S3_KEY_PREFIX="${SCCACHE_S3_KEY_PREFIX:-${ARCH}}" && \
-    export CC=$(which gcc) && \
+    if [ "$USE_SCCACHE" = "true" ]; then \
-    export CXX=$(which g++) && \
+        export CMAKE_C_COMPILER_LAUNCHER="sccache" && \
+        export CMAKE_CXX_COMPILER_LAUNCHER="sccache" && \
+        export CMAKE_CUDA_COMPILER_LAUNCHER="sccache"; \
+    fi && \
    cd /workspace/nixl && \
    uv build . --out-dir /opt/dynamo/dist/nixl --python $PYTHON_VERSION
@@ -292,8 +298,11 @@ ARG ENABLE_KVBM
 RUN --mount=type=secret,id=aws-key-id,env=AWS_ACCESS_KEY_ID \
    --mount=type=secret,id=aws-secret-id,env=AWS_SECRET_ACCESS_KEY \
    export SCCACHE_S3_KEY_PREFIX=${SCCACHE_S3_KEY_PREFIX:-${ARCH}} && \
-    export CC=$(which gcc) && \
+    if [ "$USE_SCCACHE" = "true" ]; then \
-    export CXX=$(which g++) && \
+        export CMAKE_C_COMPILER_LAUNCHER="sccache" && \
+        export CMAKE_CXX_COMPILER_LAUNCHER="sccache" && \
+        export RUSTC_WRAPPER="sccache"; \
+    fi && \
    source ${VIRTUAL_ENV}/bin/activate && \
    cd /opt/dynamo && \
    uv build --wheel --out-dir /opt/dynamo/dist && \

--- a/container/Dockerfile.sglang
+++ b/container/Dockerfile.sglang
@@ -253,8 +253,11 @@ RUN --mount=type=secret,id=aws-key-id,env=AWS_ACCESS_KEY_ID \
 RUN --mount=type=secret,id=aws-key-id,env=AWS_ACCESS_KEY_ID \
    --mount=type=secret,id=aws-secret-id,env=AWS_SECRET_ACCESS_KEY \
    export SCCACHE_S3_KEY_PREFIX="${SCCACHE_S3_KEY_PREFIX:-${ARCH}}" && \
-    export CC=$(which gcc) && \
+    if [ "$USE_SCCACHE" = "true" ]; then \
-    export CXX=$(which g++) && \
+        export CMAKE_C_COMPILER_LAUNCHER="sccache" && \
+        export CMAKE_CXX_COMPILER_LAUNCHER="sccache" && \
+        export CMAKE_CUDA_COMPILER_LAUNCHER="sccache"; \
+    fi && \
    source ${VIRTUAL_ENV}/bin/activate && \
    git clone --depth 1 --branch ${NIXL_REF} "https://github.com/ai-dynamo/nixl.git" && \
    cd nixl && \
@@ -280,8 +283,11 @@ RUN echo "$NIXL_LIB_DIR" > /etc/ld.so.conf.d/nixl.conf && \
 RUN --mount=type=secret,id=aws-key-id,env=AWS_ACCESS_KEY_ID \
    --mount=type=secret,id=aws-secret-id,env=AWS_SECRET_ACCESS_KEY \
    export SCCACHE_S3_KEY_PREFIX="${SCCACHE_S3_KEY_PREFIX:-${ARCH}}" && \
-    export CC=$(which gcc) && \
+    if [ "$USE_SCCACHE" = "true" ]; then \
-    export CXX=$(which g++) && \
+        export CMAKE_C_COMPILER_LAUNCHER="sccache" && \
+        export CMAKE_CXX_COMPILER_LAUNCHER="sccache" && \
+        export CMAKE_CUDA_COMPILER_LAUNCHER="sccache"; \
+    fi && \
    cd /workspace/nixl && \
    uv build . --out-dir /opt/dynamo/dist/nixl --python $PYTHON_VERSION
@@ -301,8 +307,6 @@ RUN --mount=type=secret,id=aws-key-id,env=AWS_ACCESS_KEY_ID \
        export CMAKE_CXX_COMPILER_LAUNCHER="sccache" && \
        export RUSTC_WRAPPER="sccache"; \
    fi && \
-    export CC=$(which gcc) && \
-    export CXX=$(which g++) && \
    source ${VIRTUAL_ENV}/bin/activate && \
    cd /opt/dynamo && \
    uv build --wheel --out-dir /opt/dynamo/dist && \
@@ -358,9 +362,6 @@ ARG GRACE_BLACKWELL=false
 ARG ARCH
 ARG ARCH_ALT
 ARG PYTHON_VERSION
-ARG USE_SCCACHE
-ARG SCCACHE_BUCKET
-ARG SCCACHE_REGION
 ARG CARGO_BUILD_JOBS
 ARG CUDA_VERSION
@@ -463,17 +464,6 @@ RUN apt-get update \
    && rm -rf /var/lib/apt/lists/* \
    && apt-get clean
-# Install sccache if requested
-COPY container/use-sccache.sh /tmp/use-sccache.sh
-RUN if [ "$USE_SCCACHE" = "true" ]; then \
-    /tmp/use-sccache.sh install; \
-fi
-# Set environment variables - they'll be empty strings if USE_SCCACHE=false
-ENV SCCACHE_BUCKET=${USE_SCCACHE:+${SCCACHE_BUCKET}} \
-    SCCACHE_REGION=${USE_SCCACHE:+${SCCACHE_REGION}} \
-    SCCACHE_S3_KEY_PREFIX=${USE_SCCACHE:+${ARCH}}
 WORKDIR /sgl-workspace
 # GDRCopy installation
@@ -538,13 +528,7 @@ RUN --mount=type=cache,target=/var/cache/curl,uid=1000,gid=0 \
    && sed -i 's/#define NUM_CPU_TIMEOUT_SECS 100/#define NUM_CPU_TIMEOUT_SECS 1000/' csrc/kernels/configs.cuh
 # Build and install NVSHMEM library only (without python library)
-RUN --mount=type=secret,id=aws-key-id,env=AWS_ACCESS_KEY_ID \
+RUN cd /sgl-workspace/nvshmem && \
-    --mount=type=secret,id=aws-secret-id,env=AWS_SECRET_ACCESS_KEY \
-    export SCCACHE_S3_KEY_PREFIX=${SCCACHE_S3_KEY_PREFIX:-${ARCH}} && \
-    export CC=$(which gcc) && \
-    export CXX=$(which g++) && \
-    export CUDA_CXX=$(which nvcc) && \
-    cd /sgl-workspace/nvshmem && \
    if [ "$GRACE_BLACKWELL" = true ]; then CUDA_ARCH="90;100;120"; else CUDA_ARCH="90"; fi && \
    NVSHMEM_SHMEM_SUPPORT=0 \
    NVSHMEM_UCX_SUPPORT=0 \
@@ -555,18 +539,11 @@ RUN --mount=type=secret,id=aws-key-id,env=AWS_ACCESS_KEY_ID \
    NVSHMEM_TIMEOUT_DEVICE_POLLING=0 \
    NVSHMEM_USE_GDRCOPY=1 \
    cmake -S . -B build/ -DCMAKE_INSTALL_PREFIX=${NVSHMEM_DIR} -DCMAKE_CUDA_ARCHITECTURES=${CUDA_ARCH} -DNVSHMEM_BUILD_PYTHON_LIB=OFF && \
-    cmake --build build --target install -j${CMAKE_BUILD_PARALLEL_LEVEL} && \
+    cmake --build build --target install -j${CMAKE_BUILD_PARALLEL_LEVEL}
-    /tmp/use-sccache.sh show-stats "NVSHMEM"
 # Build nvshmem4py wheels separately (Python 3.10, CUDA 12) to avoid building the python library twice for multiple python versions
 # Need to reconfigure with PYTHON_LIB=ON to add the nvshmem4py subdirectory
-RUN --mount=type=secret,id=aws-key-id,env=AWS_ACCESS_KEY_ID \
+RUN cd /sgl-workspace/nvshmem && \
-    --mount=type=secret,id=aws-secret-id,env=AWS_SECRET_ACCESS_KEY \
-    export SCCACHE_S3_KEY_PREFIX=${SCCACHE_S3_KEY_PREFIX:-${ARCH}} && \
-    export CC=$(which gcc) && \
-    export CXX=$(which g++) && \
-    export CUDA_CXX=$(which nvcc) && \
-    cd /sgl-workspace/nvshmem && \
    if [ "$GRACE_BLACKWELL" = true ]; then CUDA_ARCH="90;100;120"; else CUDA_ARCH="90"; fi && \
    NVSHMEM_SHMEM_SUPPORT=0 \
    NVSHMEM_UCX_SUPPORT=0 \
@@ -577,14 +554,10 @@ RUN --mount=type=secret,id=aws-key-id,env=AWS_ACCESS_KEY_ID \
    NVSHMEM_TIMEOUT_DEVICE_POLLING=0 \
    NVSHMEM_USE_GDRCOPY=1 \
    cmake -S . -B build/ -DCMAKE_INSTALL_PREFIX=${NVSHMEM_DIR} -DCMAKE_CUDA_ARCHITECTURES=${CUDA_ARCH} -DNVSHMEM_BUILD_PYTHON_LIB=ON && \
-    cmake --build build --target build_nvshmem4py_wheel_cu12_${PYTHON_VERSION} -j${CMAKE_BUILD_PARALLEL_LEVEL} && \
+    cmake --build build --target build_nvshmem4py_wheel_cu12_${PYTHON_VERSION} -j${CMAKE_BUILD_PARALLEL_LEVEL}
-    /tmp/use-sccache.sh show-stats "NVSHMEM4PY"
 # Install DeepEP
-RUN --mount=type=secret,id=aws-key-id,env=AWS_ACCESS_KEY_ID \
+RUN cd /sgl-workspace/DeepEP && \
-    --mount=type=secret,id=aws-secret-id,env=AWS_SECRET_ACCESS_KEY \
-    export SCCACHE_S3_KEY_PREFIX=${SCCACHE_S3_KEY_PREFIX:-${ARCH}} && \
-    cd /sgl-workspace/DeepEP && \
    NVSHMEM_DIR=${NVSHMEM_DIR} TORCH_CUDA_ARCH_LIST="9.0;10.0" pip install --no-build-isolation .
 # Copy rust installation from dynamo_base to avoid duplication efforts

--- a/container/Dockerfile.trtllm
+++ b/container/Dockerfile.trtllm
@@ -232,8 +232,7 @@ RUN if [ "$USE_SCCACHE" = "true" ]; then \
 # Set SCCACHE environment variables
 ENV SCCACHE_BUCKET=${USE_SCCACHE:+${SCCACHE_BUCKET}} \
-    SCCACHE_REGION=${USE_SCCACHE:+${SCCACHE_REGION}} \
+    SCCACHE_REGION=${USE_SCCACHE:+${SCCACHE_REGION}}
-    RUSTC_WRAPPER=${USE_SCCACHE:+sccache}
 # Build and install UCX
 RUN --mount=type=secret,id=aws-key-id,env=AWS_ACCESS_KEY_ID \
@@ -274,8 +273,11 @@ RUN --mount=type=secret,id=aws-key-id,env=AWS_ACCESS_KEY_ID \
 RUN --mount=type=secret,id=aws-key-id,env=AWS_ACCESS_KEY_ID \
    --mount=type=secret,id=aws-secret-id,env=AWS_SECRET_ACCESS_KEY \
    export SCCACHE_S3_KEY_PREFIX="${SCCACHE_S3_KEY_PREFIX:-${ARCH}}" && \
-    export CC=$(which gcc) && \
+    if [ "$USE_SCCACHE" = "true" ]; then \
-    export CXX=$(which g++) && \
+        export CMAKE_C_COMPILER_LAUNCHER="sccache" && \
+        export CMAKE_CXX_COMPILER_LAUNCHER="sccache" && \
+        export CMAKE_CUDA_COMPILER_LAUNCHER="sccache"; \
+    fi && \
    source ${VIRTUAL_ENV}/bin/activate && \
    git clone --depth 1 --branch ${NIXL_REF} "https://github.com/ai-dynamo/nixl.git" && \
    cd nixl && \
@@ -300,9 +302,12 @@ RUN echo "$NIXL_LIB_DIR" > /etc/ld.so.conf.d/nixl.conf && \
 RUN --mount=type=secret,id=aws-key-id,env=AWS_ACCESS_KEY_ID \
    --mount=type=secret,id=aws-secret-id,env=AWS_SECRET_ACCESS_KEY \
-    export CC=$(which gcc) && \
-    export CXX=$(which g++) && \
    export SCCACHE_S3_KEY_PREFIX="${SCCACHE_S3_KEY_PREFIX:-${ARCH}}" && \
+    if [ "$USE_SCCACHE" = "true" ]; then \
+        export CMAKE_C_COMPILER_LAUNCHER="sccache" && \
+        export CMAKE_CXX_COMPILER_LAUNCHER="sccache" && \
+        export CMAKE_CUDA_COMPILER_LAUNCHER="sccache"; \
+    fi && \
    cd /workspace/nixl && \
    uv build . --out-dir /opt/dynamo/dist/nixl --python $PYTHON_VERSION
@@ -323,8 +328,6 @@ RUN --mount=type=secret,id=aws-key-id,env=AWS_ACCESS_KEY_ID \
        export CMAKE_CXX_COMPILER_LAUNCHER="sccache" && \
        export RUSTC_WRAPPER="sccache"; \
    fi && \
-    export CC=$(which gcc) && \
-    export CXX=$(which g++) && \
    source ${VIRTUAL_ENV}/bin/activate && \
    cd /opt/dynamo && \
    uv build --wheel --out-dir /opt/dynamo/dist && \

--- a/container/Dockerfile.vllm
+++ b/container/Dockerfile.vllm
@@ -276,8 +276,11 @@ RUN --mount=type=secret,id=aws-key-id,env=AWS_ACCESS_KEY_ID \
 RUN --mount=type=secret,id=aws-key-id,env=AWS_ACCESS_KEY_ID \
    --mount=type=secret,id=aws-secret-id,env=AWS_SECRET_ACCESS_KEY \
    export SCCACHE_S3_KEY_PREFIX="${SCCACHE_S3_KEY_PREFIX:-${ARCH}}" && \
-    export CC=$(which gcc) && \
+    if [ "$USE_SCCACHE" = "true" ]; then \
-    export CXX=$(which g++) && \
+        export CMAKE_C_COMPILER_LAUNCHER="sccache" && \
+        export CMAKE_CXX_COMPILER_LAUNCHER="sccache" && \
+        export CMAKE_CUDA_COMPILER_LAUNCHER="sccache"; \
+    fi && \
    source ${VIRTUAL_ENV}/bin/activate && \
    git clone --depth 1 --branch ${NIXL_REF} "https://github.com/ai-dynamo/nixl.git" && \
    cd nixl && \
@@ -302,9 +305,12 @@ RUN echo "$NIXL_LIB_DIR" > /etc/ld.so.conf.d/nixl.conf && \
 RUN --mount=type=secret,id=aws-key-id,env=AWS_ACCESS_KEY_ID \
    --mount=type=secret,id=aws-secret-id,env=AWS_SECRET_ACCESS_KEY \
-    export CC=$(which gcc) && \
-    export CXX=$(which g++) && \
    export SCCACHE_S3_KEY_PREFIX="${SCCACHE_S3_KEY_PREFIX:-${ARCH}}" && \
+    if [ "$USE_SCCACHE" = "true" ]; then \
+        export CMAKE_C_COMPILER_LAUNCHER="sccache" && \
+        export CMAKE_CXX_COMPILER_LAUNCHER="sccache" && \
+        export CMAKE_CUDA_COMPILER_LAUNCHER="sccache"; \
+    fi && \
    cd /workspace/nixl && \
    uv build . --out-dir /opt/dynamo/dist/nixl --python $PYTHON_VERSION
@@ -324,8 +330,6 @@ RUN --mount=type=secret,id=aws-key-id,env=AWS_ACCESS_KEY_ID \
        export CMAKE_CXX_COMPILER_LAUNCHER="sccache" && \
        export RUSTC_WRAPPER="sccache"; \
    fi && \
-    export CC=$(which gcc) && \
-    export CXX=$(which g++) && \
    source ${VIRTUAL_ENV}/bin/activate && \
    cd /opt/dynamo && \
    uv build --wheel --out-dir /opt/dynamo/dist && \
@@ -416,38 +420,12 @@ ARG MAX_JOBS=16
 ENV MAX_JOBS=$MAX_JOBS
 ENV CUDA_HOME=/usr/local/cuda
-# Install sccache if requested
-COPY container/use-sccache.sh /tmp/use-sccache.sh
-# Install sccache if requested
-ARG USE_SCCACHE
-ARG ARCH_ALT
-ARG SCCACHE_BUCKET
-ARG SCCACHE_REGION
-ENV ARCH_ALT=${ARCH_ALT}
-RUN if [ "$USE_SCCACHE" = "true" ]; then \
-        /tmp/use-sccache.sh install; \
-    fi
-# Set environment variables - they'll be empty strings if USE_SCCACHE=false
-ENV SCCACHE_BUCKET=${USE_SCCACHE:+${SCCACHE_BUCKET}} \
-    SCCACHE_REGION=${USE_SCCACHE:+${SCCACHE_REGION}}
 # Install VLLM and related dependencies
 RUN --mount=type=bind,source=./container/deps/,target=/tmp/deps \
    --mount=type=cache,target=/root/.cache/uv \
-    --mount=type=secret,id=aws-key-id,env=AWS_ACCESS_KEY_ID \
-    --mount=type=secret,id=aws-secret-id,env=AWS_SECRET_ACCESS_KEY \
-    export SCCACHE_S3_KEY_PREFIX=${SCCACHE_S3_KEY_PREFIX:-${ARCH}} && \
-    if [ "$USE_SCCACHE" = "true" ]; then \
-        export CMAKE_C_COMPILER_LAUNCHER="sccache" && \
-        export CMAKE_CXX_COMPILER_LAUNCHER="sccache" && \
-        export CMAKE_CUDA_COMPILER_LAUNCHER="sccache"; \
-    fi && \
    cp /tmp/deps/vllm/install_vllm.sh /tmp/install_vllm.sh && \
    chmod +x /tmp/install_vllm.sh && \
-        /tmp/install_vllm.sh --vllm-ref $VLLM_REF --max-jobs $MAX_JOBS --arch $ARCH --installation-dir /opt ${DEEPGEMM_REF:+--deepgemm-ref "$DEEPGEMM_REF"} ${FLASHINF_REF:+--flashinf-ref "$FLASHINF_REF"} ${LMCACHE_REF:+--lmcache-ref "$LMCACHE_REF"} --cuda-version $CUDA_VERSION && \
+    /tmp/install_vllm.sh --vllm-ref $VLLM_REF --max-jobs $MAX_JOBS --arch $ARCH --installation-dir /opt ${DEEPGEMM_REF:+--deepgemm-ref "$DEEPGEMM_REF"} ${FLASHINF_REF:+--flashinf-ref "$FLASHINF_REF"} ${LMCACHE_REF:+--lmcache-ref "$LMCACHE_REF"} --cuda-version $CUDA_VERSION
-        /tmp/use-sccache.sh show-stats "vLLM";
 ENV LD_LIBRARY_PATH=\
 /opt/vllm/tools/ep_kernels/ep_kernels_workspace/nvshmem_install/lib:\