A script I use to quickly build llama.cpp for nvidia spark dgx- NCCL enabled. can run multiple times to update server

rebuild_llama.sh

#!/bin/bash

set -e

LLAMA_DIR="$HOME/llama.cpp"

# 1. Clone or update
cd "$HOME"
if [ ! -d "$LLAMA_DIR/.git" ]; then
    echo "=== Cloning llama.cpp (first time) ==="
    git clone --depth 1 https://github.com/ggml-org/llama.cpp.git "$LLAMA_DIR"
else
    echo "=== Updating llama.cpp ==="
    cd "$LLAMA_DIR"
    git pull
fi
cd "$LLAMA_DIR"

# 2. Clean previous builds
echo "Cleaning up old build cache..."
rm -rf build build-cuda CMakeCache.txt CMakeFiles cmake_install.cmake

# 3. Locate NCCL — find the real header + lib, not a guessed path.
#    NCCL's Makefile build lands headers in build/include and libs in build/lib.
#    Prefer that; fall back to a system install (e.g. /usr/local) if present.
NCCL_H=""
NCCL_LIB=""
for root in "$HOME/nccl" /usr/local /usr; do
  if [[ -z "$NCCL_H" ]]; then
    NCCL_H=$(find "$root" -name 'nccl.h' -path '*/include/*' 2>/dev/null | head -1 || true)
  fi
  if [[ -z "$NCCL_LIB" ]]; then
    NCCL_LIB=$(find "$root" -name 'libnccl.so' 2>/dev/null | head -1 || true)
  fi
  [[ -n "$NCCL_H" && -n "$NCCL_LIB" ]] && break
done

if [[ -n "$NCCL_H" && -n "$NCCL_LIB" ]]; then
  NCCL_INC_DIR=$(dirname "$NCCL_H")
  NCCL_LIB_DIR=$(dirname "$NCCL_LIB")
  echo "NCCL header: $NCCL_H"
  echo "NCCL lib:    $NCCL_LIB"
  NCCL_FLAGS="-DNCCL_LIBRARY=$NCCL_LIB -DNCCL_INCLUDE_DIR=$NCCL_INC_DIR"
else
  NCCL_FLAGS=""
  echo "Warning: NCCL not found — building without NCCL (set up ~/nccl first)"
fi

# 4. Configure for Blackwell + NCCL
echo "Configuring CMake (CUDA, FA, MoE, NCCL)..."
cmake -B build \
  -DCMAKE_BUILD_TYPE=Release \
  -DGGML_CUDA=ON \
  -DGGML_CUDA_FA=ON \
  -DGGML_MOE=ON \
  -DGGML_CUDA_NCCL=ON \
  $NCCL_FLAGS

# 5. Compile
echo "Compiling llama-server + llama-bench..."
cmake --build build -j --target llama-server llama-bench

# 6. Verify
echo "=== Build complete ==="
./build/bin/llama-server --version | head -1

# 7. Set up runtime env for NCCL (only if we found it)
if [[ -n "$NCCL_LIB_DIR" ]]; then
cat > "$HOME/.nccl_env" <<EOF
export LD_LIBRARY_PATH="$NCCL_LIB_DIR:/usr/local/cuda/lib64:${LD_LIBRARY_PATH}"
export NCCL_DEBUG=INFO
EOF
echo "Runtime env written to ~/.nccl_env (source it before running multi-GPU)"
fi