#!/usr/bin/env bash # GridShare Node Agent — AMD ROCm Installer # Supports: RX 400/500/Vega (GCN), RX 5000 (RDNA1), RX 6000/7000 (RDNA2/3) # GCN / RDNA1 / Ubuntu 20.04 → ROCm 5.7 (auto-detected) # RDNA2 / RDNA3 / Ubuntu 22.04 → ROCm 6.1 (best performance) # Run as root or with sudo: # curl -fsSL https://gridshare.in/dl/install-amd.sh | sudo bash # ───────────────────────────────────────────────────────────────────────────── set -euo pipefail CENTRAL_URL="${GRIDSHARE_CENTRAL_URL:-https://relay.gridshare.in}" GRIDSHARE_KEY="${GRIDSHARE_KEY:-}" GRIDSHARE_LABEL="${GRIDSHARE_LABEL:-$(hostname)}" GRIDSHARE_CITY="${GRIDSHARE_CITY:-}" SERVICE_NAME="gridshare-node" INSTALL_DIR="/opt/gridshare" VENV_DIR="$INSTALL_DIR/venv" CONFIG_DIR="/etc/gridshare" LOG_DIR="/var/log/gridshare" ROCM_VERSION="6.1" # may be overridden to 5.7.1 based on GPU arch / OS GPU_ARCH="rdna2" # gcn | rdna1 | rdna2 | rdna3 # ── Colour helpers ──────────────────────────────────────────────────────────── _bold=$(tput bold 2>/dev/null || true) _reset=$(tput sgr0 2>/dev/null || true) _green='\033[0;32m' _yellow='\033[1;33m' _red='\033[0;31m' _cyan='\033[0;36m' _nc='\033[0m' step() { echo -e "\n${_cyan}${_bold}── $1${_reset}"; } ok() { echo -e " ${_green}[✓]${_nc} $1"; } info() { echo -e " ${_yellow}$1${_nc}"; } warn() { echo -e " ${_yellow}[!]${_nc} $1"; } fail() { echo -e " ${_red}[✗]${_nc} $1"; exit 1; } ask() { local prompt="$1" default="${2:-}"; printf " %s" "$prompt"; [ -n "$default" ] && printf " [%s]" "$default"; printf ": "; read -r _ans; echo "${_ans:-$default}"; } # ── Banner ──────────────────────────────────────────────────────────────────── clear echo "" echo " ╔══════════════════════════════════════════════════════════════╗" echo " ║ GridShare — AMD ROCm Node Agent Installer ║" echo " ║ Earn INR from your AMD GPU • gridshare.in ║" echo " ║ Supports: RX 400/500/Vega · RX 5000–7000 · RDNA 1/2/3 ║" echo " ╚══════════════════════════════════════════════════════════════╝" echo "" # ── Root check ──────────────────────────────────────────────────────────────── if [[ $EUID -ne 0 ]]; then fail "Please run as root: sudo bash install-amd.sh" fi # ── OS check ───────────────────────────────────────────────────────────────── # ROCm 6.1 → requires Ubuntu 22.04+ (RDNA2/3 GPUs) # ROCm 5.7 → works on Ubuntu 20.04 and 22.04 (GCN/RDNA1 GPUs, legacy path) step "Checking OS compatibility" if [ -f /etc/os-release ]; then . /etc/os-release OS_NAME="$NAME" OS_VER="$VERSION_ID" else OS_NAME="Unknown" OS_VER="0" fi if [[ "$OS_NAME" != *"Ubuntu"* ]]; then fail "AMD ROCm requires Ubuntu. Detected: $OS_NAME. Use Ubuntu 20.04 or 22.04." fi OS_CODENAME="jammy" # Ubuntu 22.04 default if [[ "$OS_VER" == "20.04" ]]; then OS_CODENAME="focal" warn "Ubuntu 20.04 detected — will use ROCm 5.7 legacy path (Ubuntu 22.04 recommended for best performance)" elif [[ "$OS_VER" < "20.04" ]]; then fail "Ubuntu $OS_VER is too old. Minimum: Ubuntu 20.04 (ROCm 5.7) or 22.04 (ROCm 6.1)" fi ok "OS: Ubuntu $OS_VER ($OS_CODENAME) ✓" # ── Detect AMD GPU ──────────────────────────────────────────────────────────── step "Detecting AMD GPU" AMD_GPU_NAME="" AMD_VRAM_GB=0 # Method 1: lspci if command -v lspci &>/dev/null; then AMD_GPU_NAME=$(lspci | grep -i "VGA\|3D\|Display" | grep -i "AMD\|Radeon\|RX " | head -1 | sed 's/.*: //' | sed 's/ (rev .*//' || true) fi # Method 2: /sys/class/drm if [ -z "$AMD_GPU_NAME" ]; then for card in /sys/class/drm/card*/device; do vendor=$(cat "$card/vendor" 2>/dev/null || echo "") if [ "$vendor" = "0x1002" ]; then # AMD vendor ID AMD_GPU_NAME=$(cat "$card/product" 2>/dev/null || echo "AMD Radeon GPU") break fi done fi # Method 3: rocminfo (if already installed) if [ -z "$AMD_GPU_NAME" ] && command -v rocminfo &>/dev/null; then AMD_GPU_NAME=$(rocminfo 2>/dev/null | grep "Marketing Name:" | head -1 | sed 's/.*Marketing Name: *//' || true) fi if [ -z "$AMD_GPU_NAME" ]; then warn "No AMD GPU detected automatically." warn "Supported: RX 6700 XT, RX 6800 XT, RX 6900 XT, RX 7800 XT, RX 7900 XTX" AMD_GPU_NAME=$(ask "Enter your GPU model manually (or press Ctrl+C to exit)" "") [ -z "$AMD_GPU_NAME" ] && fail "No AMD GPU found. This installer is for AMD ROCm only." fi ok "GPU detected: $AMD_GPU_NAME" # ── Detect GPU architecture → pick ROCm version ─────────────────────────────── # GCN (Polaris) RX 4xx, RX 5xx (3-digit), Vega 56/64, Radeon VII → ROCm 5.7 # RDNA1 (Navi 10) RX 5500/5600/5700 (4-digit) → ROCm 5.7 # RDNA2 (Navi 2x) RX 6xxx → ROCm 6.1 # RDNA3 (Navi 3x) RX 7xxx → ROCm 6.1 GPU_LOWER_ARCH=$(echo "$AMD_GPU_NAME" | tr '[:upper:]' '[:lower:]') if echo "$GPU_LOWER_ARCH" | grep -qE "radeon vii|radeon-vii"; then GPU_ARCH="gcn" # Radeon VII = Vega 20 GCN elif echo "$GPU_LOWER_ARCH" | grep -qE "vega.?(5[0-9]|6[0-9]|64|56)|rx vega"; then GPU_ARCH="gcn" # Vega 56 / Vega 64 elif echo "$GPU_LOWER_ARCH" | grep -qE "rx[[:space:]_-]*[45][0-9]{2}[^0-9]"; then GPU_ARCH="gcn" # RX 470/480/570/580/590 (3-digit models) elif echo "$GPU_LOWER_ARCH" | grep -qE "rx[[:space:]_-]*5[0-9]{3}|rx5[0-9]{3}"; then GPU_ARCH="rdna1" # RX 5500/5600/5700 XT (4-digit starting with 5) elif echo "$GPU_LOWER_ARCH" | grep -qE "rx[[:space:]_-]*7[0-9]{3}|rx7[0-9]{3}"; then GPU_ARCH="rdna3" # RX 7000 series fi # default stays rdna2 for RX 6xxx and anything unrecognised # Force ROCm 5.7 for GCN / RDNA1 — ROCm 6.x dropped support for these if [[ "$GPU_ARCH" == "gcn" ]] || [[ "$GPU_ARCH" == "rdna1" ]]; then ROCM_VERSION="5.7.1" warn "Legacy GPU architecture detected ($GPU_ARCH) — using ROCm 5.7 compatibility path" info "Supported: RX 470/480/570/580/590, Vega 56/64, Radeon VII, RX 5000 series" info "PyTorch, TensorFlow and inference workloads work on ROCm 5.7" elif [[ "$OS_CODENAME" == "focal" ]]; then # Ubuntu 20.04 + RDNA2/3: ROCm 6.1 doesn't support focal; drop to 5.7 ROCM_VERSION="5.7.1" warn "Ubuntu 20.04 + RDNA2 detected — ROCm 6.1 requires Ubuntu 22.04" warn "Using ROCm 5.7 fallback. Upgrade to Ubuntu 22.04 for best RDNA2/3 performance." else ok "RDNA2/3 architecture on Ubuntu 22.04 — using ROCm 6.1 (optimal)" fi # ── AMD GPU tier pricing (72% of equivalent NVIDIA tier) ───────────────────── # AMD ROCm nodes are priced at 72% of equivalent NVIDIA VRAM because ROCm has # a smaller buyer pool than CUDA. Rates will increase as ROCm adoption grows. # # Tier mapping: # RX 7900 XTX / XTX (24GB) → ₹32/hr (vs RTX 4090 ₹44/hr) # RX 7900 GRE / XT (16GB) → ₹27/hr (vs RTX 3090 ₹34/hr) # RX 7800 XT (16GB) → ₹20/hr (vs RTX 3080 ₹24/hr) # RX 6900 XT (16GB) → ₹27/hr (vs RTX 3090 ₹34/hr) # RX 6800 XT (16GB) → ₹18/hr (vs RTX 3080 Ti ₹28/hr) # RX 6800 (16GB) → ₹16/hr (vs RTX 3080 ₹24/hr) # RX 6750 XT / 6700 XT (12GB) → ₹14/hr (vs RTX 3080 ₹24/hr, less VRAM) # Other / Unknown AMD → ₹10/hr (Starter equivalent) GPU_LOWER=$(echo "$AMD_GPU_NAME" | tr '[:upper:]' '[:lower:]') PRICE_PER_HOUR=10 TIER_NAME="AMD Starter" AMD_VRAM_GB=0 if echo "$GPU_LOWER" | grep -qE "7900 xtx|rx 7900xtx"; then PRICE_PER_HOUR=32; TIER_NAME="AMD Ultra (ROCm)"; AMD_VRAM_GB=24 elif echo "$GPU_LOWER" | grep -qE "7900 gre|7900 xt|6900 xt"; then PRICE_PER_HOUR=27; TIER_NAME="AMD Pro Max (ROCm)"; AMD_VRAM_GB=16 elif echo "$GPU_LOWER" | grep -qE "7800 xt|6800 xt"; then PRICE_PER_HOUR=20; TIER_NAME="AMD Pro (ROCm)"; AMD_VRAM_GB=16 elif echo "$GPU_LOWER" | grep -qE "rx 6800[^x]|6800$"; then PRICE_PER_HOUR=16; TIER_NAME="AMD Standard+ (ROCm)"; AMD_VRAM_GB=16 elif echo "$GPU_LOWER" | grep -qE "6750 xt|6700 xt"; then PRICE_PER_HOUR=14; TIER_NAME="AMD Standard (ROCm)"; AMD_VRAM_GB=12 elif echo "$GPU_LOWER" | grep -qE "7700 xt|7600 xt|6650 xt|6600 xt"; then PRICE_PER_HOUR=12; TIER_NAME="AMD Plus (ROCm 6.1)"; AMD_VRAM_GB=8 # ── GCN / RDNA1 tiers (ROCm 5.7 path) ──────────────────────────────────────── elif echo "$GPU_LOWER" | grep -qE "radeon vii|radeon-vii"; then PRICE_PER_HOUR=14; TIER_NAME="AMD Radeon VII (ROCm 5.7)"; AMD_VRAM_GB=16 elif echo "$GPU_LOWER" | grep -qE "vega.?64|rx vega 64|vega64"; then PRICE_PER_HOUR=8; TIER_NAME="AMD Vega 64 (ROCm 5.7)"; AMD_VRAM_GB=8 elif echo "$GPU_LOWER" | grep -qE "vega.?56|rx vega 56|vega56"; then PRICE_PER_HOUR=7; TIER_NAME="AMD Vega 56 (ROCm 5.7)"; AMD_VRAM_GB=8 elif echo "$GPU_LOWER" | grep -qE "rx.?5700.?xt|rx5700xt"; then PRICE_PER_HOUR=8; TIER_NAME="AMD RX 5700 XT (ROCm 5.7)"; AMD_VRAM_GB=8 elif echo "$GPU_LOWER" | grep -qE "rx.?5700[^x ]|rx.?5700$"; then PRICE_PER_HOUR=7; TIER_NAME="AMD RX 5700 (ROCm 5.7)"; AMD_VRAM_GB=8 elif echo "$GPU_LOWER" | grep -qE "rx.?56[0-9]{2}|rx.?55[0-9]{2}"; then PRICE_PER_HOUR=5; TIER_NAME="AMD RDNA1 (ROCm 5.7)"; AMD_VRAM_GB=6 elif echo "$GPU_LOWER" | grep -qE "rx.?5[89][0-9]|rx.?48[0-9]|rx.?47[0-9]"; then PRICE_PER_HOUR=5; TIER_NAME="AMD GCN Polaris (ROCm 5.7)"; AMD_VRAM_GB=8 else # Unmatched AMD card — don't guess a rate (a high-VRAM Instinct would be # under-floored). Central assigns the authoritative rate on first heartbeat. PRICE_PER_HOUR=""; TIER_NAME="set on registration"; AMD_VRAM_GB=8 fi echo "" echo " ┌─────────────────────────────────────────────────────────────┐" echo " │ GPU: $AMD_GPU_NAME" echo " │ Tier: $TIER_NAME" if [ -n "$PRICE_PER_HOUR" ]; then NET_EARNINGS=$(echo "scale=2; $PRICE_PER_HOUR * 0.7" | bc 2>/dev/null || echo "?") echo " │ Rate: ~₹${PRICE_PER_HOUR}/hr (buyers pay; final set on registration)" echo " │ Yours: ~₹${NET_EARNINGS}/hr (70% share, daily UPI payout)" else echo " │ Rate: set by GridShare on registration — see your dashboard" echo " │ Yours: 70% of every rented hour (daily UPI payout)" fi echo " └─────────────────────────────────────────────────────────────┘" echo "" # ── Install ROCm 6.1 ────────────────────────────────────────────────────────── step "Installing AMD ROCm $ROCM_VERSION" # Add AMD ROCm APT repository (version-specific URL) if [ ! -f /etc/apt/sources.list.d/rocm.list ]; then info "Adding AMD ROCm $ROCM_VERSION repository..." apt-get update -qq apt-get install -y -qq curl gnupg wget # AMD ROCm signing key (same for all versions) curl -fsSL https://repo.radeon.com/rocm/rocm.gpg.key | \ gpg --dearmor -o /usr/share/keyrings/rocm.gpg # ROCm repo — version-specific URL, OS-specific codename echo "deb [arch=amd64 signed-by=/usr/share/keyrings/rocm.gpg] \ https://repo.radeon.com/rocm/apt/${ROCM_VERSION} ${OS_CODENAME} main" \ > /etc/apt/sources.list.d/rocm.list # amdgpu driver repo — only for RDNA2/3 (GCN uses built-in kernel driver) if [[ "$GPU_ARCH" == "rdna2" ]] || [[ "$GPU_ARCH" == "rdna3" ]]; then echo "deb [arch=amd64 signed-by=/usr/share/keyrings/rocm.gpg] \ https://repo.radeon.com/amdgpu/latest/ubuntu ${OS_CODENAME} main" \ > /etc/apt/sources.list.d/amdgpu.list info "RDNA2/3: amdgpu driver repo added" else info "GCN/RDNA1: using built-in kernel amdgpu driver (no proprietary driver needed)" fi ok "ROCm $ROCM_VERSION repository added" fi info "Installing ROCm packages (this may take 5–10 minutes)..." apt-get update -qq if [[ "$ROCM_VERSION" == "5.7.1" ]]; then # ── ROCm 5.7 packages — GCN / RDNA1 / Ubuntu 20.04 legacy path ────────────── # hipblaslt not in 5.7 → use hipblas # python3-rocm not packaged in 5.7 → install via pip later apt-get install -y -qq \ rocm-hip-sdk \ rocm-developer-tools \ hipblas \ miopen-hip 2>/dev/null || true else # ── ROCm 6.1 packages — RDNA2/3 path ───────────────────────────────────────── apt-get install -y -qq \ rocm-hip-sdk \ rocm-developer-tools \ python3-rocm \ hipblaslt \ miopen-hip 2>/dev/null || true fi # Add user to video/render groups (required for GPU access) usermod -aG video,render root 2>/dev/null || true ok "ROCm $ROCM_VERSION installed" # Verify ROCm if command -v rocminfo &>/dev/null; then ROCM_DETECTED=$(rocminfo 2>/dev/null | grep "Agent " | grep -c "gfx" || echo "0") if [ "$ROCM_DETECTED" -gt 0 ]; then ok "ROCm GPU agents detected: $ROCM_DETECTED" else warn "ROCm installed but no GPU agents detected — check driver compatibility" fi fi # ── Docker with ROCm support ────────────────────────────────────────────────── step "Setting up Docker with ROCm" if ! command -v docker &>/dev/null; then info "Installing Docker..." curl -fsSL https://get.docker.com | sh systemctl enable docker systemctl start docker fi # Add ROCm device access to Docker (no special runtime needed for ROCm — uses --device) ok "Docker ready for ROCm (uses --device=/dev/kfd --device=/dev/dri)" # ── Python & PyTorch ROCm ──────────────────────────────────────────────────── step "Checking Python 3.10+" if ! python3 --version 2>/dev/null | grep -qE "3\.(10|11|12)"; then apt-get install -y -qq python3.11 python3.11-venv python3-pip update-alternatives --install /usr/bin/python3 python3 /usr/bin/python3.11 1 fi ok "Python: $(python3 --version)" # ── System deps ─────────────────────────────────────────────────────────────── step "Installing system dependencies" # python3-venv/pip REQUIRED for the agent venv — generic pkg matches the system python3 # (minimal Ubuntu 24.04 has python3 3.12 but not python3.12-venv → "ensurepip" failure). apt-get install -y -qq bc curl jq lm-sensors net-tools pciutils fuse3 python3-venv python3-pip ok "System deps ready" # ── rclone — required for Network Volume mounting ──────────────────────────── step "Installing rclone (Network Volume support)" if ! command -v rclone &>/dev/null; then info "Downloading rclone..." RCLONE_ARCH="amd64" case "$(uname -m)" in arm64|aarch64) RCLONE_ARCH="arm64" ;; esac curl -fsSL "https://downloads.rclone.org/rclone-current-linux-${RCLONE_ARCH}.zip" \ -o /tmp/rclone.zip 2>/dev/null && \ unzip -q /tmp/rclone.zip -d /tmp/rclone-extract 2>/dev/null && \ cp /tmp/rclone-extract/rclone-*/rclone /usr/local/bin/rclone && \ chmod +x /usr/local/bin/rclone && \ rm -rf /tmp/rclone.zip /tmp/rclone-extract && \ ok "rclone installed: $(rclone --version 2>/dev/null | head -1)" || \ warn "rclone install failed — Network Volumes won't mount. Install manually." else ok "rclone already installed: $(rclone --version 2>/dev/null | head -1)" fi # ── Open DDP/NCCL inter-node port ──────────────────────────────────────────── # Gaming cafe cluster mode: when a buyer requests multiple GPUs across LAN nodes, # PyTorch DistributedDataParallel uses NCCL for gradient sync. # NCCL's rendezvous listens on port 29500 (configurable via master_port). # We open this port from the LAN subnet so rank-0 (master) can accept connections # from worker nodes without firewall blocks. # Ref: https://pytorch.org/docs/stable/distributed.html#initialization step "Opening NCCL inter-node port (29500) for cluster mode" if command -v ufw &>/dev/null; then UFW_STATUS=$(ufw status 2>/dev/null | grep "Status:" | awk '{print $2}') if [ "$UFW_STATUS" = "active" ]; then ufw allow 29500/tcp comment "GridShare NCCL cluster DDP" 2>/dev/null && \ ok "ufw: port 29500/tcp opened for NCCL" || \ warn "ufw rule failed — open port 29500/tcp manually if using cluster mode" else info "ufw is inactive — no firewall rule needed for NCCL" fi else info "ufw not found — skipping NCCL firewall rule (iptables-based systems: open port 29500/tcp manually)" fi ok "Cluster mode: NCCL port configured" # ── Provider Registration ───────────────────────────────────────────────────── step "Provider Registration" OWNER_NAME="" UPI_ID="" CITY="$GRIDSHARE_CITY" EMAIL_OPT="" MACHINE_LABEL="$GRIDSHARE_LABEL" API_KEY="" if [ -n "$GRIDSHARE_KEY" ]; then echo "" echo " ── Fleet Mode ─────────────────────────────────────────────────────────" echo " GRIDSHARE_KEY found — skipping registration." API_KEY="$GRIDSHARE_KEY" ok "Fleet mode: API key loaded" else echo "" echo " GridShare will pay your earnings to your UPI ID." echo "" OWNER_NAME=$(ask "Your name" "") UPI_ID=$(ask "UPI ID for payouts (GPay/PhonePe/Paytm)" "") CITY=$(ask "City" "") EMAIL_OPT=$(ask "Email (optional, for payout receipts)" "") [ -z "$OWNER_NAME" ] && fail "Name is required" [ -z "$UPI_ID" ] && fail "UPI ID is required" echo "" echo " Registering with GridShare..." REG_RESPONSE=$(curl -s -X POST "$CENTRAL_URL/nodes/issue-key" \ -H "Content-Type: application/json" \ -d "{\"owner_name\":\"$OWNER_NAME\",\"upi_id\":\"$UPI_ID\",\"email\":\"$EMAIL_OPT\",\"city\":\"$CITY\",\"gpu_backend\":\"rocm\"}" \ 2>/dev/null || echo '{"error":"connection failed"}') API_KEY=$(echo "$REG_RESPONSE" | python3 -c "import sys,json; d=json.load(sys.stdin); print(d.get('api_key',''))" 2>/dev/null || true) if [ -z "$API_KEY" ]; then warn "Could not auto-register." API_KEY=$(ask "Paste your API key from gridshare.in/portal" "") [ -z "$API_KEY" ] && fail "API key required." else ok "API key issued" echo "" echo " ┌──────────────────────────────────────────────────────────────┐" echo " │ YOUR API KEY — SAVE THIS, IT WON'T BE SHOWN AGAIN: │" echo " │ $API_KEY" echo " └──────────────────────────────────────────────────────────────┘" echo "" fi fi # ── Install gridshare-node ──────────────────────────────────────────────────── step "Installing GridShare Node Agent" mkdir -p "$INSTALL_DIR" "$CONFIG_DIR" "$LOG_DIR" python3 -m venv "$VENV_DIR" "$VENV_DIR/bin/pip" install -q --upgrade pip WHEEL_URL="$CENTRAL_URL/dl/gridshare_node-1.3.64-py3-none-any.whl" if curl -sfI "$WHEEL_URL" &>/dev/null; then "$VENV_DIR/bin/pip" install -q "$WHEEL_URL" else "$VENV_DIR/bin/pip" install -q gridshare-node 2>/dev/null || \ fail "Could not install gridshare-node. Contact support@gridshare.in" fi ok "Node agent installed" # Install PyTorch with ROCm support — version-matched to installed ROCm if [[ "$ROCM_VERSION" == "5.7.1" ]]; then TORCH_ROCM_TAG="rocm5.7" else TORCH_ROCM_TAG="rocm6.0" fi info "Installing PyTorch ($TORCH_ROCM_TAG)..." "$VENV_DIR/bin/pip" install -q \ torch torchvision torchaudio \ --index-url "https://download.pytorch.org/whl/${TORCH_ROCM_TAG}" \ 2>/dev/null && ok "PyTorch ${TORCH_ROCM_TAG} installed" || warn "PyTorch install failed — will retry on first run" # ── Write config ────────────────────────────────────────────────────────────── step "Writing configuration" LOCAL_ID="amd-$(hostname | tr '[:upper:]' '[:lower:]')-$(cat /proc/sys/kernel/random/uuid 2>/dev/null | head -c 8 || date +%s | tail -c 8)" python3 - < /etc/systemd/system/${SERVICE_NAME}.service <