GB10 두 대를 연결해 K-EXAONE을 실행하는 한 페이지 가이드

ibdev2netdev
ip -br link
for IFACE in $(ibdev2netdev | awk '$NF == "(Up)" { print $(NF-1) }'); do
  echo "== $IFACE =="
  ethtool "$IFACE" | grep -E 'Speed|Lanes|Duplex|Link detected'
done

rocep1s0f0 port 1 ==> enp1s0f0np0 (Up)

roceP2p1s0f0 port 1 ==> enP2p1s0f0np0 (Up)

== enp1s0f0np0 ==

Speed: 200000Mb/s

Duplex: Full

Port: Direct Attach Copper

Link detected: yes

== enP2p1s0f0np0 ==

Speed: 200000Mb/s

Duplex: Full

Port: Direct Attach Copper

Link detected: yes

ibdev2netdev
for IFACE in $(ibdev2netdev | awk '$NF == "(Up)" { print $(NF-1) }'); do
  echo "== $IFACE =="
  ip -4 -br address show dev "$IFACE"
  nmcli -f GENERAL.STATE,IP4.ADDRESS device show "$IFACE" | sed 's/^/  /'
done
ip -4 route

== enp1s0f0np0 ==

enp1s0f0np0 UP 169.254.72.31/16

  GENERAL.STATE: 100 (connected)

  IP4.ADDRESS[1]: 169.254.72.31/16

== enP2p1s0f0np0 ==

enP2p1s0f0np0 UP 169.254.84.12/16

  GENERAL.STATE: 100 (connected)

  IP4.ADDRESS[1]: 169.254.84.12/16

169.254.0.0/16 dev enp1s0f0np0 proto kernel scope link src 169.254.72.31

169.254.0.0/16 dev enP2p1s0f0np0 proto kernel scope link src 169.254.84.12

ibdev2netdev
read -r -p "이 노드의 역할 (A/B): " NODE_ROLE
read -r -p "첫 번째 Up netdev (Path1): " QSFP_IF1
read -r -p "두 번째 Up netdev (Path2): " QSFP_IF2
case "$NODE_ROLE" in
  A) DEFAULT_LOCAL_IP1=10.42.0.1; DEFAULT_PEER_IP1=10.42.0.2; DEFAULT_LOCAL_IP2=10.42.1.1; DEFAULT_PEER_IP2=10.42.1.2 ;;
  B) DEFAULT_LOCAL_IP1=10.42.0.2; DEFAULT_PEER_IP1=10.42.0.1; DEFAULT_LOCAL_IP2=10.42.1.2; DEFAULT_PEER_IP2=10.42.1.1 ;;
  *) echo "A 또는 B만 입력하세요." >&2; exit 1 ;;
esac
echo "기본값: Path1 10.42.0.0/30, Path2 10.42.1.0/30"
read -r -p "Path1 local IP [$DEFAULT_LOCAL_IP1]: " LOCAL_IP1
read -r -p "Path1 peer IP  [$DEFAULT_PEER_IP1]: " PEER_IP1
read -r -p "Path2 local IP [$DEFAULT_LOCAL_IP2]: " LOCAL_IP2
read -r -p "Path2 peer IP  [$DEFAULT_PEER_IP2]: " PEER_IP2
: "${LOCAL_IP1:=$DEFAULT_LOCAL_IP1}"
: "${PEER_IP1:=$DEFAULT_PEER_IP1}"
: "${LOCAL_IP2:=$DEFAULT_LOCAL_IP2}"
: "${PEER_IP2:=$DEFAULT_PEER_IP2}"
mkdir -p "$HOME/.config"
cat > "$HOME/.config/qsfp-guide.env" <<EOF
export NODE_ROLE="$NODE_ROLE"
export QSFP_IF1="$QSFP_IF1"
export QSFP_IF2="$QSFP_IF2"
export LOCAL_IP1="$LOCAL_IP1"
export LOCAL_IP2="$LOCAL_IP2"
export PEER_IP1="$PEER_IP1"
export PEER_IP2="$PEER_IP2"
EOF
chmod 600 "$HOME/.config/qsfp-guide.env"
cat "$HOME/.config/qsfp-guide.env"

export NODE_ROLE="A"

export QSFP_IF1="enp1s0f0np0"

export QSFP_IF2="enP2p1s0f0np0"

export LOCAL_IP1="10.42.0.1"

export LOCAL_IP2="10.42.1.1"

export PEER_IP1="10.42.0.2"

export PEER_IP2="10.42.1.2"

source "$HOME/.config/qsfp-guide.env"
ensure_profile() {
  name="$1"; iface="$2"; address="$3"
  if nmcli -t -f NAME connection show | grep -Fxq "$name"; then
    sudo nmcli connection modify "$name" connection.interface-name "$iface" ipv4.method manual ipv4.addresses "$address/30" ipv4.gateway "" ipv4.dns "" ipv4.never-default yes ipv6.method link-local 802-3-ethernet.mtu 9000 connection.autoconnect yes
  else
    sudo nmcli connection add type ethernet ifname "$iface" con-name "$name" ipv4.method manual ipv4.addresses "$address/30" ipv4.gateway "" ipv4.dns "" ipv4.never-default yes ipv6.method link-local 802-3-ethernet.mtu 9000 connection.autoconnect yes
  fi
  sudo nmcli connection up "$name"
}
ensure_profile qsfp-primary "$QSFP_IF1" "$LOCAL_IP1"
ensure_profile qsfp-secondary "$QSFP_IF2" "$LOCAL_IP2"

Connection successfully activated (D-Bus active path: /org/freedesktop/NetworkManager/ActiveConnection/12)

Connection successfully activated (D-Bus active path: /org/freedesktop/NetworkManager/ActiveConnection/13)

source "$HOME/.config/qsfp-guide.env"
ip -4 -br address show dev "$QSFP_IF1"
ip -4 -br address show dev "$QSFP_IF2"
ip route get "$PEER_IP1"
ip route get "$PEER_IP2"
ping -I "$QSFP_IF1" -M do -s 8972 -c 4 "$PEER_IP1"
ping -I "$QSFP_IF2" -M do -s 8972 -c 4 "$PEER_IP2"

enp1s0f0np0 UP 10.42.0.1/30

enP2p1s0f0np0 UP 10.42.1.1/30

10.42.0.2 dev enp1s0f0np0 src 10.42.0.1 uid 1000

10.42.1.2 dev enP2p1s0f0np0 src 10.42.1.1 uid 1000

4 packets transmitted, 4 received, 0% packet loss, time 3001ms

4 packets transmitted, 4 received, 0% packet loss, time 3001ms

source "$HOME/.config/qsfp-guide.env"
STAMP="$(date +%F-%H%M%S)"
OUT="$HOME/qsfp-evidence/$(hostname)-$STAMP"
mkdir -p "$OUT"
{
  date -Is
  uname -a
  ip -br link
  ip -4 -br address
  ip -4 route
  ibdev2netdev
  rdma link show
  nvidia-smi
} > "$OUT/inventory.txt" 2>&1
for IFACE in "$QSFP_IF1" "$QSFP_IF2"; do
  ethtool "$IFACE" > "$OUT/ethtool-$IFACE.txt" 2>&1
  ethtool -S "$IFACE" > "$OUT/ethtool-stats-$IFACE.txt" 2>&1
done
sudo lspci -tv > "$OUT/lspci-tree.txt" 2>&1
sudo lspci -vv > "$OUT/lspci-vv.txt" 2>&1
sudo journalctl -k -b --no-pager > "$OUT/kernel-journal.txt" 2>&1
printf 'Evidence saved: %s
' "$OUT"
find "$OUT" -maxdepth 1 -type f -printf '%f
' | LC_ALL=C sort

Evidence saved: /home/donghyeon/qsfp-evidence/spark-a-2026-07-23-153000

ethtool-enP2p1s0f0np0.txt

ethtool-enp1s0f0np0.txt

ethtool-stats-enP2p1s0f0np0.txt

ethtool-stats-enp1s0f0np0.txt

inventory.txt

kernel-journal.txt

lspci-tree.txt

lspci-vv.txt

git clone https://github.com/ggml-org/llama.cpp.git "$HOME/llama.cpp-b10088"
cd "$HOME/llama.cpp-b10088"
git checkout b10088
cmake -B build -DGGML_CUDA=ON -DGGML_RPC=ON -DCMAKE_BUILD_TYPE=Release
cmake --build build -j"$(nproc)"
sha256sum build/bin/ggml-rpc-server
build/bin/llama-server --version

<sha256>  build/bin/ggml-rpc-server

version: b10088