Handle "gpus" not in facts

README.md

@@ -1,18 +1,15 @@
-
 # nvidia_docker_runtime
 
 This is a PDK based module https://puppet.com/pdk/latest/pdk_generating_modules.html .
 
 #### Table of Contents
 
 1. [Description](#description)
-3. [Usage - Configuration options and additional functionality](#usage)
+2. [Usage - Configuration options and additional functionality](#usage)
 
 ## Description
 
-Modifies Docker for use with NVIDIA GPU based containers, including in swarm mode. 
-
-## Setup
+Modifies Docker for use with NVIDIA GPU based containers, including in swarm mode.
 
 ## Usage
 
@@ -22,3 +19,10 @@ class { 'nvidia_docker_runtime':
   nvidia_docker2_version => '2.2.2-1',
 }
 ```
+
+## Limitations
+
+Since GPU UUIDs are needed to specify `node-generic-resources`, and these are not available through facts until
+the NVIDIA driver has been installed, it takes two applies to fully setup. Until then swarm cannot use gpu resources.
+
+Due to Docker limitations you can only specify gpus in compose files for compose format 2.3. As `docker stack` requires format 3.0 this means you can't use gpus in stacks.

manifests/init.pp

@@ -66,7 +66,11 @@
   }
   ~> Service['docker']
 
-  $gpu_ids = $facts['gpus'].map |$gpu| { $gpu['gpu_uuid'][0,11] }
+  $gpu_ids = if 'gpus' in $facts {
+    $facts['gpus'].map |$gpu| { $gpu['gpu_uuid'][0,11] }
+  } else {
+    []
+  }
   Package['nvidia-docker2']
   -> augeas { 'daemon.json':
     lens    => 'Json.lns',