[rllib] PPO doesn't work with fractional num gpus (#3396)

* frac ppo

* gpu test

python/ray/rllib/optimizers/multi_gpu_optimizer.py

@@ -3,6 +3,7 @@
 from __future__ import print_function
 
 import logging
+import math
 import numpy as np
 from collections import defaultdict
 import tensorflow as tf
@@ -44,7 +45,9 @@ def _init(self,
         if not num_gpus:
             self.devices = ["/cpu:0"]
         else:
-            self.devices = ["/gpu:{}".format(i) for i in range(num_gpus)]
+            self.devices = [
+                "/gpu:{}".format(i) for i in range(int(math.ceil(num_gpus)))
+            ]
         self.batch_size = int(sgd_batch_size / len(self.devices)) * len(
             self.devices)
         assert self.batch_size % len(self.devices) == 0

test/jenkins_tests/run_multi_node_tests.sh

@@ -53,6 +53,14 @@ docker run --rm --shm-size=10G --memory=10G $DOCKER_SHA \
     --stop '{"training_iteration": 2}' \
     --config '{"simple_optimizer": true, "num_sgd_iter": 2, "model": {"use_lstm": true}}'
 
+docker run --rm --shm-size=10G --memory=10G $DOCKER_SHA \
+    python /ray/python/ray/rllib/train.py \
+    --env CartPole-v1 \
+    --run PPO \
+    --stop '{"training_iteration": 2}' \
+    --config '{"num_gpus": 0.1}' \
+    --ray-num-gpus 1
+
 docker run --rm --shm-size=10G --memory=10G $DOCKER_SHA \
     python /ray/python/ray/rllib/train.py \
     --env CartPole-v1 \