fix: enhance device node checks for ollama-rocm.service and add GPU diagnostics script

README.md

@@ -170,13 +170,22 @@ sudo bash ./install-rootful.sh
 ```
 
 If logs show device mapping errors for `/dev/dri`, verify your runtime and GPU device nodes.
-The installers skip starting `ollama-rocm.service` when `/dev/kfd` or `/dev/dri` is missing.
+The installers skip starting `ollama-rocm.service` when `/dev/kfd` or `/dev/dri` is missing,
+or when `/dev/dri` exists but has no `renderD*` / `card*` nodes.
 
 Verify GPU device nodes:
 
 ```bash
 ls -l /dev/kfd
 ls -l /dev/dri
+ls -l /dev/dri/renderD* /dev/dri/card*
+```
+
+Run bundled diagnostics helper:
+
+```bash
+chmod +x ./diag-gpu.sh
+./diag-gpu.sh
 ```
 
 To force-refresh the Ollama image manually:
@@ -192,4 +201,4 @@ sudo systemctl restart ollama-rocm.service
 
 - `podman-mcp-server` is launched via `npx` inside a Node container because the upstream project is distributed as binary/npm package.
 - The Ollama unit mirrors your ROCm `docker run` flags.
-- If this host is not Linux with ROCm devices (`/dev/kfd`, `/dev/dri`), `ollama` will fail to start.
+- If this host is not Linux with ROCm devices (`/dev/kfd`, `/dev/dri/renderD*`), `ollama` will fail to start.

diag-gpu.sh

@@ -0,0 +1,82 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+ok() { echo "[OK]   $*"; }
+warn() { echo "[WARN] $*"; }
+fail() { echo "[FAIL] $*"; }
+info() { echo "[INFO] $*"; }
+
+FAILED=0
+
+echo "ROCm + Podman GPU diagnostics"
+echo
+
+if [[ -e /dev/kfd ]]; then
+  ok "/dev/kfd is present"
+else
+  fail "/dev/kfd is missing"
+  FAILED=1
+fi
+
+if [[ -d /dev/dri ]]; then
+  ok "/dev/dri directory is present"
+else
+  fail "/dev/dri directory is missing"
+  FAILED=1
+fi
+
+shopt -s nullglob
+DRI_NODES=(/dev/dri/renderD* /dev/dri/card*)
+shopt -u nullglob
+
+if (( ${#DRI_NODES[@]} > 0 )); then
+  ok "Detected DRM device nodes under /dev/dri"
+  ls -l /dev/dri/renderD* /dev/dri/card* 2>/dev/null || true
+else
+  fail "No /dev/dri/renderD* or /dev/dri/card* device nodes found"
+  FAILED=1
+fi
+
+echo
+info "Raw device listing"
+ls -l /dev/kfd 2>/dev/null || true
+ls -la /dev/dri 2>/dev/null || true
+
+echo
+if lsmod | grep -q '^amdgpu'; then
+  ok "Kernel module amdgpu is loaded"
+else
+  warn "Kernel module amdgpu is not loaded"
+fi
+
+if command -v podman >/dev/null 2>&1; then
+  ok "podman is installed"
+  if podman info >/dev/null 2>&1; then
+    ok "podman info succeeded"
+  else
+    warn "podman info failed (check podman setup/permissions)"
+  fi
+else
+  fail "podman is not installed"
+  FAILED=1
+fi
+
+if command -v rocminfo >/dev/null 2>&1; then
+  ok "rocminfo is installed"
+  if rocminfo >/dev/null 2>&1; then
+    ok "rocminfo executed successfully"
+  else
+    warn "rocminfo exists but failed to execute"
+  fi
+else
+  warn "rocminfo not found (install ROCm userspace tools for deeper checks)"
+fi
+
+echo
+if [[ "${FAILED}" -eq 0 ]]; then
+  ok "Diagnostic passed: host exposes required GPU device nodes for ollama-rocm"
+  exit 0
+fi
+
+fail "Diagnostic failed: fix missing device nodes/driver setup, then rerun"
+exit 1

install-rootful.sh

@@ -50,6 +50,9 @@ fi
 if [[ ! -d /dev/dri ]]; then
   echo "Skipping ollama-rocm.service: /dev/dri is missing on this host."
   OLLAMA_READY=false
+elif ! compgen -G "/dev/dri/renderD*" >/dev/null && ! compgen -G "/dev/dri/card*" >/dev/null; then
+  echo "Skipping ollama-rocm.service: /dev/dri has no render/card nodes on this host."
+  OLLAMA_READY=false
 fi
 
 if [[ "${OLLAMA_READY}" == "true" ]]; then
@@ -68,7 +71,7 @@ echo "  - podman.socket"
 if [[ "${OLLAMA_READY}" == "true" ]]; then
   echo "  - ollama-rocm.service"
 else
-  echo "  - ollama-rocm.service (skipped: missing /dev/kfd or /dev/dri)"
+  echo "  - ollama-rocm.service (skipped: missing /dev/kfd, /dev/dri, or /dev/dri nodes)"
 fi
 echo "  - open-webui.service"
 echo "  - podman-mcp-server.service"

install.sh

@@ -102,6 +102,9 @@ fi
 if [[ ! -d /dev/dri ]]; then
   echo "Skipping ollama-rocm.service: /dev/dri is missing on this host."
   OLLAMA_READY=false
+elif ! compgen -G "/dev/dri/renderD*" >/dev/null && ! compgen -G "/dev/dri/card*" >/dev/null; then
+  echo "Skipping ollama-rocm.service: /dev/dri has no render/card nodes on this host."
+  OLLAMA_READY=false
 fi
 
 if [[ "${OLLAMA_READY}" == "true" ]]; then
@@ -120,7 +123,7 @@ echo "  - podman.socket"
 if [[ "${OLLAMA_READY}" == "true" ]]; then
   echo "  - ollama-rocm.service"
 else
-  echo "  - ollama-rocm.service (skipped: missing /dev/kfd or /dev/dri)"
+  echo "  - ollama-rocm.service (skipped: missing /dev/kfd, /dev/dri, or /dev/dri nodes)"
 fi
 echo "  - open-webui.service"
 echo "  - podman-mcp-server.service"

preflight.sh

@@ -75,7 +75,11 @@ else
 fi
 
 if [[ -d /dev/dri ]]; then
-  ok "/dev/dri present"
+  if compgen -G "/dev/dri/renderD*" >/dev/null || compgen -G "/dev/dri/card*" >/dev/null; then
+    ok "/dev/dri has render/card device nodes"
+  else
+    warn "/dev/dri exists but has no render/card nodes (ROCm container will not start)"
+  fi
 else
   warn "/dev/dri missing (ROCm container will not start)"
 fi