From 4e1394d6d54266f3aa263de78feec38d468d8090 Mon Sep 17 00:00:00 2001
From: MickLesk <47820557+MickLesk@users.noreply.github.com>
Date: Tue, 20 Oct 2026 11:03:02 +0200
Subject: [PATCH] Only treat an NVIDIA GPU as present when its device nodes are

lspci shows the host's PCI bus inside the container, so every NVIDIA
card on the host counted as "passthrough detected" and pulled the CUDA
userspace before the device check at the end said otherwise (ProxmoxVE
#18020). The unpinned fallback keeps libcuda1 on the chosen version, and
the final check refuses a libcuda from another driver series.
---
 lib/hwaccel.func | 27 +++++++++++++++++++--------
 1 file changed, 19 insertions(+), 8 deletions(-)

diff --git a/lib/hwaccel.func b/lib/hwaccel.func
index 71783aa..3793a8b 100644
--- a/lib/hwaccel.func
+++ b/lib/hwaccel.func
@@ -122,7 +122,12 @@ setup_hwaccel() {
     elif echo "$gpu_name" | grep -qiwE 'AMD|ATI|Radeon|Advanced Micro Devices'; then
       gpu_type="AMD"
     elif echo "$gpu_name" | grep -qi 'NVIDIA'; then
-      gpu_type="NVIDIA"
+      # lspci shows the host's bus; the card is only ours with its device nodes.
+      if compgen -G '/dev/nvidia[0-9]*' >/dev/null; then
+        gpu_type="NVIDIA"
+      else
+        msg_custom "ℹ️" "${YW}" "NVIDIA GPU on the host, but no /dev/nvidia* device in this container - skipping it (add the devices to the container config to use it)"
+      fi
     fi
 
     if [[ -n "$gpu_type" ]]; then
@@ -810,9 +815,9 @@ NVIDIA_PIN
           msg_warn "Same driver series, different release. CUDA workloads may still fail on this."
         fi
       else
-        msg_warn "Version-pinned install failed - trying unpinned"
+        msg_warn "Version-pinned install failed - keeping libcuda1 at ${available_version}, the rest unpinned"
         $STD apt-get -y -o Dpkg::Options::="--force-confold" install --no-install-recommends \
-          libcuda1 libnvcuvid1 libnvidia-encode1 libnvidia-ml1 2>/dev/null || true
+          "libcuda1=${available_version}" libnvcuvid1 libnvidia-encode1 libnvidia-ml1 2>/dev/null || true
       fi
     else
       # Deliberately install nothing. The previous code fell back to whatever
@@ -927,16 +932,22 @@ NVIDIA_PIN
   local have_libcuda=0
   ldconfig -p 2>/dev/null | grep -q 'libcuda\.so\.1' && have_libcuda=1
 
-  if [[ $have_libcuda -eq 1 && ${#nvidia_nodes[@]} -gt 0 ]]; then
-    _cs_nvidia_usable=1
-    msg_ok "NVIDIA GPU configured (host driver ${nvidia_host_version}, ${#nvidia_nodes[@]} device nodes)"
-  elif [[ ${#nvidia_nodes[@]} -eq 0 ]]; then
+  local libcuda_version=""
+  libcuda_version="$(dpkg-query -W -f='${Version}' libcuda1 2>/dev/null || true)"
+
+  if [[ ${#nvidia_nodes[@]} -eq 0 ]]; then
     msg_error "NVIDIA GPU not usable: no /dev/nvidia* device nodes in this container"
     msg_warn "The userspace driver is in place, but the devices were never passed through."
     msg_warn "On the host, check the container config lists /dev/nvidia0, /dev/nvidiactl and /dev/nvidia-uvm."
-  else
+  elif [[ $have_libcuda -eq 0 ]]; then
     msg_error "NVIDIA GPU not usable: libcuda.so.1 is not available in this container"
     msg_warn "The host driver is ${nvidia_host_version}. CUDA workloads will fail until a matching userspace library is present."
+  elif [[ -n "$libcuda_version" && "${libcuda_version%%.*}" != "${nvidia_host_version%%.*}" ]]; then
+    msg_error "NVIDIA GPU not usable: libcuda ${libcuda_version} is not the host driver's ${nvidia_host_version%%.*} series"
+    msg_warn "The userspace library must match the host module. Install the ${nvidia_host_version%%.*} series on the host, or run NVIDIA's ${nvidia_host_version} .run installer with --no-kernel-module in the container."
+  else
+    _cs_nvidia_usable=1
+    msg_ok "NVIDIA GPU configured (host driver ${nvidia_host_version}, ${#nvidia_nodes[@]} device nodes)"
   fi
 
   # nvidia-smi separately: in NVIDIA's CUDA repos the package is an empty