feat: v3.10 GPU stress and NCCL burn updates

This commit is contained in:
Mikhail Chusavitin
2026-03-31 11:22:27 +03:00
parent 6dee8f3509
commit c850b39b01
12 changed files with 423 additions and 88 deletions
+24 -1
View File
@@ -9,6 +9,7 @@
#
# Output layout:
# $CACHE_DIR/bin/all_reduce_perf
# $CACHE_DIR/lib/libcudart.so* copied from the nvcc toolchain used to build nccl-tests
set -e
@@ -30,7 +31,7 @@ CACHE_DIR="${DIST_DIR}/nccl-tests-${NCCL_TESTS_VERSION}"
CACHE_ROOT="${BEE_CACHE_DIR:-${DIST_DIR}/cache}"
DOWNLOAD_CACHE_DIR="${CACHE_ROOT}/nccl-tests-downloads"
if [ -f "${CACHE_DIR}/bin/all_reduce_perf" ]; then
if [ -f "${CACHE_DIR}/bin/all_reduce_perf" ] && [ "$(find "${CACHE_DIR}/lib" -maxdepth 1 -name 'libcudart.so*' 2>/dev/null | wc -l)" -gt 0 ]; then
echo "=== nccl-tests cached, skipping build ==="
echo "binary: ${CACHE_DIR}/bin/all_reduce_perf"
exit 0
@@ -52,6 +53,23 @@ echo "nvcc: $NVCC"
CUDA_HOME="$(dirname "$(dirname "$NVCC")")"
echo "CUDA_HOME: $CUDA_HOME"
find_cudart_dir() {
for dir in \
"${CUDA_HOME}/targets/x86_64-linux/lib" \
"${CUDA_HOME}/targets/x86_64-linux/lib/stubs" \
"${CUDA_HOME}/lib64" \
"${CUDA_HOME}/lib"; do
if [ -d "$dir" ] && find "$dir" -maxdepth 1 -name 'libcudart.so*' -type f | grep -q .; then
printf '%s\n' "$dir"
return 0
fi
done
return 1
}
CUDART_DIR="$(find_cudart_dir)" || { echo "ERROR: libcudart.so* not found under ${CUDA_HOME}"; exit 1; }
echo "cudart dir: $CUDART_DIR"
# Download libnccl-dev for nccl.h
REPO_BASE="https://developer.download.nvidia.com/compute/cuda/repos/debian${DEBIAN_VERSION}/x86_64"
DEV_PKG="libnccl-dev_${NCCL_VERSION}+cuda${NCCL_CUDA_VERSION}_amd64.deb"
@@ -136,6 +154,11 @@ mkdir -p "${CACHE_DIR}/bin"
cp "./build/all_reduce_perf" "${CACHE_DIR}/bin/all_reduce_perf"
chmod +x "${CACHE_DIR}/bin/all_reduce_perf"
mkdir -p "${CACHE_DIR}/lib"
find "${CUDART_DIR}" -maxdepth 1 -name 'libcudart.so*' -type f -exec cp -a {} "${CACHE_DIR}/lib/" \;
[ "$(find "${CACHE_DIR}/lib" -maxdepth 1 -name 'libcudart.so*' -type f | wc -l)" -gt 0 ] || { echo "ERROR: libcudart runtime copy failed"; exit 1; }
echo "=== nccl-tests build complete ==="
echo "binary: ${CACHE_DIR}/bin/all_reduce_perf"
ls -lh "${CACHE_DIR}/bin/all_reduce_perf"
ls -lh "${CACHE_DIR}/lib/"libcudart.so* 2>/dev/null || true