#!/usr/bin/env bash # GridShare Node Agent — AMD ROCm Installer # Supports: RX 400/500/Vega (GCN), RX 5000 (RDNA1), RX 6000/7000 (RDNA2/3), # Radeon PRO W6000/W7000, AMD Instinct MI100/MI200/MI300 (CDNA) # NOT yet: RX 9000 series (RDNA 4) — needs ROCm 6.4+; installer stops with a # clear message rather than installing a stack the card cannot use. # 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="${GRIDSHARE_CENTRAL_URL:-https://relay.gridshare.in}" NODE_VERSION="1.3.93" # Account-first install: set GRIDSHARE_TOKEN (a gsp_… provision token minted in the # portal) to install with ZERO prompts. Priority: TOKEN > KEY > interactive. GRIDSHARE_TOKEN="${GRIDSHARE_TOKEN:-}" 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} $1"; exit 1; } # Prompt goes to STDERR, not stdout — otherwise `x=$(ask "…")` captures the # prompt text itself into x. That is exactly how gpu_name once became the string # "Enter your GPU model manually…" on a detached install (found 2026-07-23). ask() { local prompt="$1" default="${2:-}"; { printf " %s" "$prompt"; [ -n "$default" ] && printf " [%s]" "$default"; printf ": "; } >&2; read -r _ans; echo "${_ans:-$default}"; } # ── Banner ──────────────────────────────────────────────────────────────────── clear echo "" echo " ╔══════════════════════════════════════════════════════════════╗" echo " ║ GridShare — AMD ROCm Node Agent Installer ║" echo " ║ Earn INR from your AMD GPU • gridshare.in ║" echo " ║ RX 400-7000 · Radeon PRO · Instinct MI100-MI300 ║" echo " ║ RX 9000 (RDNA 4) not yet supported ║" 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. NOTE the class list must include "Processing accelerators" — # AMD Instinct (MI100/MI200/MI300) reports as an accelerator, NOT as VGA/3D/ # Display, so a real MI300X went completely undetected here (found live # 2026-07-28 on a 129.212.x box: the card was present as # "Processing accelerators: [AMD/ATI] Device 74b5" and we saw nothing). if command -v lspci &>/dev/null; then AMD_GPU_NAME=$(lspci | grep -iE "VGA|3D|Display|Processing accelerator" \ | grep -iE "AMD|Radeon|RX |Instinct" \ | head -1 | sed 's/.*: //' | sed 's/ (rev .*//' || true) fi # Method 1b: Instinct cards often expose no marketing name before ROCm is # installed — lspci yields e.g. "Advanced Micro Devices, Inc. [AMD/ATI] Device # 74b5". Map the PCI device id to a real model so arch detection can work and # the provider sees a sane name instead of a hex id. if echo "$AMD_GPU_NAME" | grep -qiE "device [0-9a-f]{4}"; then _did=$(echo "$AMD_GPU_NAME" | grep -oiE "device [0-9a-f]{4}" | tail -1 | awk '{print tolower($2)}') case "$_did" in 74a0|74a1|74a5|74a9|74b5|74b9) AMD_GPU_NAME="AMD Instinct MI300X" ;; 7408|740c|740f|7410) AMD_GPU_NAME="AMD Instinct MI250X" ;; 738c|738e) AMD_GPU_NAME="AMD Instinct MI100" ;; esac 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 "$AMD_GPU_NAME" ]; then warn "No AMD GPU detected automatically." warn "Supported: RX 6700 XT, RX 6800 XT, RX 6900 XT, RX 7800 XT, RX 7900 XTX" # Only prompt when a human is actually at the keyboard (stdin is a TTY). On a # detached install (curl | bash, token/fleet mode) there is no one to answer — # prompting there is what let the prompt text get captured as gpu_name. Fail # CLOSED to an UNKNOWN GPU instead of guessing: register with no name, and let # the node agent's GPU-usability probe decide availability once ROCm is up # (a truly-dead GPU registers UNAVAILABLE; a working one is re-detected on the # agent's first heartbeat). Never a hard fail — that would abort a fleet roll. if [ -t 0 ]; then AMD_GPU_NAME=$(ask "Enter your GPU model manually (or leave blank to auto-detect later)" "") else warn "Non-interactive install — leaving GPU model UNKNOWN; the node agent will re-detect and gate on real GPU health." fi fi if [ -n "$AMD_GPU_NAME" ]; then ok "GPU detected: $AMD_GPU_NAME" else warn "GPU model unknown — registering as UNKNOWN; the agent verifies GPU health before earning." fi # ── Detect GPU architecture → pick ROCm version ─────────────────────────────── # GCN (Polaris) RX 4xx, RX 5xx (3-digit), Vega 56/64, Radeon VII → ROCm 5.7 # RDNA1 (Navi 10) RX 5500/5600/5700 (4-digit) → ROCm 5.7 # RDNA2 (Navi 2x) RX 6xxx → ROCm 6.1 # RDNA3 (Navi 3x) RX 7xxx → ROCm 6.1 GPU_LOWER_ARCH=$(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]|$)"; then # RX 470/480/570/580/590 (3-digit). The trailing ([^0-9]|$) alternation is # load-bearing: the old pattern demanded a non-digit AFTER the digits, so a # name ENDING in the model — "AMD Radeon RX 580", which is exactly how lspci # reports it — matched nothing and fell through to the rdna2 default, i.e. the # wrong ROCm on one of the most common older AMD cards. Caught 2026-07-28 by # test_amd_installer_detection.py on its first run. GPU_ARCH="gcn" 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 elif echo "$GPU_LOWER_ARCH" | grep -qE "rx[[:space:]_-]*9[0-9]{3}|rx9[0-9]{3}"; then GPU_ARCH="rdna4" # RX 9000 series (Navi 4x — 9070 XT/9070/9060 XT/9060) elif echo "$GPU_LOWER_ARCH" | grep -qE "rx[[:space:]_-]*6[0-9]{3}|rx6[0-9]{3}"; then GPU_ARCH="rdna2" # RX 6000 series — EXPLICIT, no longer a silent default elif echo "$GPU_LOWER_ARCH" | grep -qE "instinct|mi[[:space:]_-]?(100|210|250|300|325|355)"; then GPU_ARCH="cdna" # Instinct MI-series (CDNA) — ROCm 6.1 covers MI100→MI300X elif echo "$GPU_LOWER_ARCH" | grep -qE "w7[0-9]{3}|pro[[:space:]_-]*w7"; then GPU_ARCH="rdna3" # Radeon PRO W7900/W7800 (RDNA3 workstation) elif echo "$GPU_LOWER_ARCH" | grep -qE "w6[0-9]{3}|pro[[:space:]_-]*w6"; then GPU_ARCH="rdna2" # Radeon PRO W6800 etc (RDNA2 workstation) else GPU_ARCH="unknown" # nothing matched — do NOT guess (see the guard below) fi # ── Unsupported / unrecognised architecture → FAIL LOUDLY ──────────────────── # Until 2026-07-27 anything unmatched silently fell through to the rdna2 default # and installed ROCm 6.1. An RX 9060 XT (RDNA 4, needs ROCm 6.4+) therefore # "installed successfully" onto a stack its GPU cannot use — the card was dead # but the installer said OK, and the provider only found out when jobs failed. # A wrong install is worse than a refused one: fail here with a real reason. # ── UNSUPPORTED / UNKNOWN AMD ARCH — INSTALL ANYWAY, BUT NEVER SERVE BUYERS ── # Policy (Vamsi, 2026-07-28): "don't block install on any card... let them be # listed but make sure it won't launch for the buyer... once the AMD node is # there mail me and we start building." Turning a provider away loses them # permanently AND loses us the only hardware we could develop against. So we # install, register, and let the agent's GPU-usability probe hold the node OUT # of the rental pool (fail-closed) until we ship real support. The provider # keeps their setup; we get a real card to build against. if [[ "$GPU_ARCH" == "rdna4" ]] || [[ "$GPU_ARCH" == "unknown" ]]; then echo "" warn "This GPU (${AMD_GPU_NAME:-unknown model}) is NOT fully supported yet." warn "Installing anyway so your machine is registered and we can build support" warn "for it — but it will NOT be rented to buyers until that work lands, so" warn "you will not earn from it in the meantime. We are notified automatically" warn "when a card like yours joins, and we'll email you the moment it's live." warn "Questions: support@gridshare.in" GRIDSHARE_UNSUPPORTED_GPU=1 # RDNA 4 needs ROCm 6.4+; our 6.1 packages cannot drive it. Install the newest # stack we have so the box is otherwise ready, and let the probe gate it. [[ "$GPU_ARCH" == "rdna4" ]] && ROCM_VERSION="6.1" fi # Force ROCm 5.7 for GCN / RDNA1 — ROCm 6.x dropped support for these if [[ "$GPU_ARCH" == "rdna4" ]]; then # Only reachable with GRIDSHARE_AMD_EXPERIMENTAL=1 (the guard above stops # everyone else). RDNA 4 = gfx1200/gfx1201, first supported in ROCm 6.4. # UNVERIFIED on real hardware — we have no RDNA 4 card in-house. ROCM_VERSION="6.4" warn "EXPERIMENTAL: RDNA 4 path selected (ROCm 6.4) — UNVERIFIED on real hardware." warn "If the GPU probe fails after install, the node self-delists rather than" warn "accepting jobs it can't run. Please report what you see to support@gridshare.in." if [[ "$OS_CODENAME" == "focal" ]]; then warn "RDNA 4 wants Ubuntu 22.04+; on 20.04 the GPU will stay unavailable." fi elif [[ "$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 [[ "$OS_CODENAME" == "focal" ]]; then # Ubuntu 20.04 + RDNA2/3: ROCm 6.1 doesn't support focal; drop to 5.7 ROCM_VERSION="5.7.1" warn "Ubuntu 20.04 + RDNA2 detected — ROCm 6.1 requires Ubuntu 22.04" warn "Using ROCm 5.7 fallback. Upgrade to Ubuntu 22.04 for best RDNA2/3 performance." else ok "$GPU_ARCH 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 / unknown AMD card — don't guess a rate (a high-VRAM Instinct # would be under-floored) OR a VRAM size. A fabricated 8 GB is exactly what # once registered as a MI300X's VRAM (real: 192 GB). Leave VRAM UNKNOWN (0); # the agent re-detects the real size and central assigns the authoritative # rate on the first heartbeat. PRICE_PER_HOUR=""; TIER_NAME="set on registration"; AMD_VRAM_GB=0 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, free monthly bank payout)" else echo " │ Rate: set by GridShare on registration — see your dashboard" echo " │ Yours: 70% of every rented hour it is booked (free monthly bank 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} ${OS_CODENAME} main" \ > /etc/apt/sources.list.d/rocm.list # amdgpu driver repo — needed by every arch that may require the DKMS kernel # driver. GCN/RDNA1 ship an in-tree amdgpu module, so they are the ONLY ones # that can skip it. cdna (Instinct) and rdna4 were missing here, which is why # a real MI300X could not get amdgpu-dkms at all: the repo was never added, so # `apt-get install amdgpu-dkms` reported "Unable to locate package" and the # card stayed permanently unusable (found live 2026-07-28). if [[ "$GPU_ARCH" == "rdna2" ]] || [[ "$GPU_ARCH" == "rdna3" ]] \ || [[ "$GPU_ARCH" == "rdna4" ]] || [[ "$GPU_ARCH" == "cdna" ]]; 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 # APT PIN — without this ROCm CANNOT install on Ubuntu (found live 2026-07-28 # on a real MI300X). Ubuntu's own universe repo ships ancient ROCm 5.0.0-1 # packages (rocminfo, rocm-device-libs, rocm-cmake). apt prefers those over # repo.radeon.com's 6.1 builds, so the dependency set is unsatisfiable: # rocm-hip-runtime : Depends: rocminfo (= 1.0.0.60100-82~22.04) # but 5.0.0-1 is to be installed # …and the whole install dies with "held broken packages". Because every # apt-get here ends in `|| true`, that failure used to be SILENT: the script # reported success while /opt/rocm did not exist and the GPU was unusable. # Pinning repo.radeon.com above the distro fixes resolution outright. cat > /etc/apt/preferences.d/rocm-pin-600 <<'PIN' Package: * Pin: release o=repo.radeon.com Pin-Priority: 600 PIN ok "ROCm $ROCM_VERSION repository added (repo pinned above distro packages)" fi info "Installing ROCm packages (this may take 5–10 minutes)..." apt-get update -qq # ── apt helpers: never swallow the error that explains the failure ─────────── # `apt-get ... 2>/dev/null || true` cost us two full install runs on an MI300X # (2026-07-28). apt printed the answer immediately — `E: Unable to locate # package python3-rocm` — and both halves of that idiom threw it away: the # redirect binned the message, `|| true` binned the exit code. Worse, # `apt-get install` is all-or-nothing at the RESOLUTION stage, so ONE bogus # package name made the whole transaction a no-op. dpkg was never invoked, # nothing appeared in /var/log/apt/history.log, no ROCm userspace landed, and # the installer printed a green tick and registered the node into a PAID # marketplace with a GPU that could not run a single buyer workload. # # Rules now: keep stderr, wait for locks instead of racing them, retry only # genuinely transient failures, and make REQUIRED packages fatal. GRIDSHARE_APT_LOG=/var/log/gridshare-apt.log export NEEDRESTART_MODE=a # don't let needrestart prompt/defer on us _apt_run() { local tries=0 rc=0 out while :; do tries=$((tries+1)) out=$(apt-get -o DPkg::Lock::Timeout=300 -o Acquire::Retries=3 "$@" 2>&1) && rc=0 || rc=$? printf '%s\n' "$out" >> "$GRIDSHARE_APT_LOG" 2>/dev/null || true [ $rc -eq 0 ] && return 0 # Resolution errors are permanent — fail fast and SHOW them. if printf '%s' "$out" | grep -qiE 'Unable to locate package|has no installation candidate|held broken packages|Unable to correct problems'; then printf '%s' "$out" | grep -iE '^E:' >&2 return "$rc" fi if [ $tries -ge 4 ]; then printf '%s' "$out" | grep -iE '^E:' >&2 return "$rc" fi sleep $((tries * 15)) done } # Optional packages: warn, keep going. _apt_try() { _apt_run install -y --no-install-recommends "$@" \ || warn "optional packages failed: $* (see $GRIDSHARE_APT_LOG)"; } # Required packages: the GPU cannot work without these. _apt_need() { _apt_run install -y --no-install-recommends "$@" \ || fail "REQUIRED ROCm packages failed to install: $* The apt errors are above and in $GRIDSHARE_APT_LOG. NOT registering this machine as rentable supply."; } # Catch a bogus package NAME before it silently no-ops the whole transaction. _apt_assert_exists() { local missing=() for p in "$@"; do apt-cache show "$p" >/dev/null 2>&1 || missing+=("$p"); done [ ${#missing[@]} -eq 0 ] || fail "package(s) not in the configured repos: ${missing[*]} This is a bug in the installer's package list, not your machine." } if [[ "$ROCM_VERSION" == "5.7.1" ]]; then # ── ROCm 5.7 packages — GCN / RDNA1 / Ubuntu 20.04 legacy path ────────────── # hipblaslt not in 5.7 → use hipblas _apt_assert_exists rocm-hip-sdk rocm-smi-lib _apt_need rocm-hip-sdk rocm-smi-lib _apt_try rocm-developer-tools hipblas miopen-hip else # ── ROCm 6.1 packages — RDNA2/3/4 + CDNA path ────────────────────────────── # NO python3-rocm: it does not exist in ANY AMD repo (verified live on an # MI300X, 2026-07-28 — `apt-cache search python3-rocm` is empty). The python # bindings ship INSIDE amd-smi-lib at /opt/rocm/share/amd_smi and are # pip-installed into the agent venv below. _apt_assert_exists rocm-hip-sdk rocm-smi-lib amd-smi-lib rocminfo _apt_need rocm-hip-sdk rocm-smi-lib amd-smi-lib rocminfo _apt_try rocm-developer-tools hipblaslt miopen-hip fi # ── amdsmi python bindings — REQUIRED by PyTorch-ROCm ──────────────────────── # torch>=2.4 +rocm does `import amdsmi` inside torch.cuda.device_count() and # get_device_name(). Without it the GPU still COMPUTES fine (a raw matmul works) # but every standard call blows up with: # NameError: name 'amdsmi' is not defined # device_count() appears in almost every real ML script (DDP init, device # selection, HF accelerate), so the node looks healthy and rentable while buyer # workloads crash on line one. Found live 2026-07-28 on the MI300X: matmul OK, # device_count FAILED. `amd-smi-lib` (added above) ships the python package # under /opt/rocm*/share/amd_smi — install it into the agent venv. _AMDSMI_SRC=$(ls -d /opt/rocm*/share/amd_smi 2>/dev/null | head -1 || true) if [ -n "$_AMDSMI_SRC" ] && [ -x "$VENV_DIR/bin/pip" ]; then if "$VENV_DIR/bin/pip" install -q "$_AMDSMI_SRC" 2>/dev/null; then ok "amdsmi python bindings installed (PyTorch device queries will work)" else warn "Could not install amdsmi from $_AMDSMI_SRC — torch.cuda.device_count()" warn "may fail for buyers on this node. Report to support@gridshare.in." fi elif [ -z "$_AMDSMI_SRC" ]; then warn "amd-smi python bindings not found under /opt/rocm*/share/amd_smi —" warn "PyTorch device queries (device_count/get_device_name) may fail." fi # ── Did ROCm USERSPACE actually install? ───────────────────────────────────── # Every apt-get above ends in `2>/dev/null || true`, which keeps a fleet roll # alive but also SWALLOWS a total failure: on the 2026-07-28 MI300X the ROCm # packages never installed at all (no /opt/rocm), yet the installer printed a # clean success and the provider had no idea. Say it out loud instead. We do NOT # abort — the node is still useful for CPU work and the agent's GPU-health probe # already refuses to rent out a card it cannot verify. ROCM_USERSPACE_OK=0 if [ ! -d /opt/rocm ] && ! ls -d /opt/rocm-* >/dev/null 2>&1; then warn "ROCm userspace did NOT install (no /opt/rocm) — the GPU will not be usable." warn "This node will stay OUT of the GPU rental pool until ROCm is working." warn "Re-run this installer, or contact support@gridshare.in with the log above." else ROCM_USERSPACE_OK=1 ok "ROCm userspace present: $(ls -d /opt/rocm* 2>/dev/null | head -1)" fi # ── Kernel driver (amdgpu) — the piece ROCm userspace CANNOT work without ──── # We install rocm-hip-sdk etc. above, but that is all USERSPACE. On a normal # desktop Radeon the in-tree `amdgpu` module is already loaded, so nothing more # is needed. On a bare cloud VM with a passed-through card (e.g. an Instinct # MI300X) there is NO driver: /dev/kfd never appears and every ROCm call fails, # while the install still "succeeds" (found live 2026-07-28 — the MI300X node # came up and was correctly held out of supply as gpu_unusable, but it could # never have recovered on its own). Install the DKMS driver in that case and be # explicit that a reboot is required to load it. if [ ! -e /dev/kfd ] && ! lsmod 2>/dev/null | grep -q '^amdgpu'; then warn "No amdgpu kernel driver loaded — installing amdgpu-dkms (kernel module)." info "This compiles against your kernel and can take several minutes..." # amdgpu-dkms is built WITH RDMA peer-memory support, so the module imports # ib_register_peer_memory_client / ib_unregister_peer_memory_client from # ib_core. On a stock cloud/VM kernel ib_core is in linux-modules-extra and is # NOT loaded, so `modprobe amdgpu` dies with: # amdgpu: Unknown symbol ib_register_peer_memory_client (err -2) # …even after a reboot. Found live 2026-07-28 on an MI300X: the driver built # fine, the card stayed dark, and nothing in the install said why. Pull the # extras and load ib_core FIRST. apt-get install -y -qq "linux-modules-extra-$(uname -r)" 2>/dev/null || true if apt-get install -y -qq amdgpu-dkms 2>/dev/null; then ok "amdgpu-dkms installed" modprobe ib_core 2>/dev/null || true modprobe amdgpu 2>/dev/null || true # Persist the load order so the GPU survives a reboot without hand-holding. printf 'ib_core\namdgpu\n' > /etc/modules-load.d/gridshare-amd.conf 2>/dev/null || true # The reboot warning and the flag BOTH belong inside the else. Previously # `GRIDSHARE_NEEDS_REBOOT=1` sat outside the `if`, so it overwrote the 0 set # two lines above and every install demanded a reboot — including the ones # where /dev/kfd was already live and the GPU worked immediately (verified # on an MI300X 2026-07-28: kfd appeared the moment modprobe ran, and no # reboot was ever needed). Crying wolf here trains providers to reboot # rented cloud boxes for nothing and buries the times it genuinely matters. if [ -e /dev/kfd ]; then ok "GPU driver live now (/dev/kfd present) — no reboot needed" GRIDSHARE_NEEDS_REBOOT=0 else warn "A REBOOT IS REQUIRED before the GPU can be used: sudo reboot" warn "Until then this node stays listed as unavailable (gpu_unusable) — it" warn "will never be rented out with a non-working GPU, and re-lists itself" warn "automatically once the driver is live." GRIDSHARE_NEEDS_REBOOT=1 fi else warn "amdgpu-dkms install failed — the GPU will remain unusable." warn "The node stays out of the rental pool until the driver works." fi fi # Add user to video/render groups (required for GPU access) usermod -aG video,render root 2>/dev/null || true # Report the TRUTH, not the fact that we reached the end of the section. This # said `ok "ROCm 6.1 installed"` unconditionally, printing four lines after # "ROCm userspace did NOT install — the GPU will not be usable" (seen live on # the 2026-07-28 MI300X). A provider reading the tail of the log saw a green # tick and reasonably assumed it worked. if [ "${ROCM_USERSPACE_OK:-0}" = "1" ]; then ok "ROCm $ROCM_VERSION installed" else warn "ROCm $ROCM_VERSION step finished, but ROCm userspace is NOT present — GPU unusable (see above)." fi # Verify ROCm if command -v rocminfo &>/dev/null; then ROCM_DETECTED=$(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="" NODE_ID="" LOCAL_ID="" if [ -n "$GRIDSHARE_TOKEN" ]; then # ── Account-first mode: single tokenized exchange, ZERO prompts ───────────── # Name/UPI/city were saved at portal signup; central reads them off the # provider record and burns the single-use token. A stale token sends the # provider back to the portal — we do NOT fall back to prompting for PII. echo "" echo " ── Account-First Install ──────────────────────────────────────────────" echo " Provision token found — no questions asked." echo "" LOCAL_ID="amd-$(hostname | tr '[:upper:]' '[:lower:]')-$(cat /proc/sys/kernel/random/uuid 2>/dev/null | head -c 8 || date +%s | tail -c 8)" CPU_MODEL=$(grep -m1 "model name" /proc/cpuinfo 2>/dev/null | sed 's/.*: //' | xargs || echo "unknown") CPU_CORES=$(nproc 2>/dev/null || echo 1) RAM_GB=$(free -g 2>/dev/null | awk '/^Mem:/{print $2}' || echo 0) GPU_JSON="[{\"name\":\"$AMD_GPU_NAME\",\"vram_total_mb\":$(( ${AMD_VRAM_GB:-0} * 1024 ))}]" HW_JSON="{\"cpu_model\":\"$CPU_MODEL\",\"cpu_cores\":$CPU_CORES,\"ram_total_gb\":$RAM_GB,\"gpu_count\":1,\"gpus\":$GPU_JSON,\"gpu_backend\":\"rocm\",\"platform\":\"linux\"}" info "Provisioning with GridShare ..." # NOTE: -sSL, NOT -fsSL. `-f` makes curl exit non-zero on any HTTP >= 400, # which triggers the `|| echo 000` fallback below and REPLACES the real status # code with 000 — so every 401/409/410 below was unreachable and a used or # expired install token reported as "Could not reach GridShare server", sending # the provider off to debug their network. Hit live on 2026-07-28. PROV_HTTP=$(curl -sSL -w "\n%{http_code}" -X POST "$CENTRAL_URL/nodes/provision" \ -H "Content-Type: application/json" \ -d "{\"token\":\"$GRIDSHARE_TOKEN\",\"local_id\":\"$LOCAL_ID\",\"machine_label\":\"$MACHINE_LABEL\",\"hardware\":$HW_JSON,\"version\":\"$NODE_VERSION\"}" \ 2>/dev/null || echo $'\n000') PROV_CODE=$(echo "$PROV_HTTP" | tail -1) PROV_RESP=$(echo "$PROV_HTTP" | sed '$d') case "$PROV_CODE" in 200|201) : ;; 401) fail "That install link is invalid. Copy a fresh one from your dashboard." ;; 409) fail "This install link was already used. Mint a new one per machine." ;; 410) fail "This install link expired. Generate a new one from your dashboard." ;; 000) fail "Could not reach GridShare server. Check your connection and re-run." ;; *) fail "Provisioning failed (HTTP $PROV_CODE). Mint a fresh install link and re-run." ;; esac NODE_ID=$(echo "$PROV_RESP" | python3 -c "import sys,json; d=json.load(sys.stdin); print(d.get('node_id') or '')" 2>/dev/null || echo "") API_KEY=$(echo "$PROV_RESP" | python3 -c "import sys,json; d=json.load(sys.stdin); print(d.get('api_key') or '')" 2>/dev/null || echo "") OWNER_NAME=$(echo "$PROV_RESP" | python3 -c "import sys,json; d=json.load(sys.stdin); print(d.get('owner_name') or '')" 2>/dev/null || echo "") UPI_ID=$(echo "$PROV_RESP" | python3 -c "import sys,json; d=json.load(sys.stdin); print(d.get('upi_id') or '')" 2>/dev/null || echo "") CITY=$(echo "$PROV_RESP" | python3 -c "import sys,json; d=json.load(sys.stdin); print(d.get('city') or '')" 2>/dev/null || echo "") if [ -z "$API_KEY" ] || [ -z "$NODE_ID" ]; then fail "Provisioning response was incomplete. Mint a fresh install link and re-run." fi ok "Node registered: $NODE_ID" ok "Installed for ${OWNER_NAME:-your account} · payouts to ${UPI_ID:-UPI on file}" elif [ -n "$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 "$REG_RESPONSE" | python3 -c "import sys,json; d=json.load(sys.stdin); print(d.get('api_key') or '')" 2>/dev/null || true) if [ -z "$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="$CENTRAL_URL/dl/gridshare_node-1.3.93-py3-none-any.whl" if curl -sfI "$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 ${TORCH_ROCM_TAG} installed" || warn "PyTorch install failed — will retry on first run" # ── Write config ────────────────────────────────────────────────────────────── step "Writing configuration" # Token path already minted its local_id (it went into /nodes/provision). [ -z "$LOCAL_ID" ] && 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 - < HIP error; without it -> OK. # The override is a COMPATIBILITY HACK for consumer RDNA cards that ROCm has no # native kernels for. Datacenter CDNA (Instinct) is natively supported and must # never be overridden. RDNA3 is already gfx11xx, so it needs nothing either. HSA_ENV_LINE="" case "$GPU_ARCH" in rdna2) HSA_ENV_LINE="Environment=HSA_OVERRIDE_GFX_VERSION=10.3.0"$'\n' ;; rdna4) HSA_ENV_LINE="Environment=HSA_OVERRIDE_GFX_VERSION=11.0.0"$'\n' ;; *) HSA_ENV_LINE="" ;; # cdna / rdna3 / gcn / rdna1 -> native, no override esac [ -n "$HSA_ENV_LINE" ] && info "HSA override for $GPU_ARCH: ${HSA_ENV_LINE%$'\n'}" \ || info "No HSA override needed for $GPU_ARCH (natively supported)" step "Installing systemd service" cat > /etc/systemd/system/${SERVICE_NAME}.service <