Upgrade pytorch-lightning to 1.5.9 (horovod#3480)

* Relax pytorch_lightning dependency to '>=1.3.8,<1.5.10'

Signed-off-by: Enrico Minack <[email protected]>

Author: EnricoMi

docker-compose.test.yml

@@ -12,7 +12,7 @@ services:
         TENSORFLOW_PACKAGE: tensorflow-cpu==2.8.0
         KERAS_PACKAGE: keras==2.8.0
         PYTORCH_PACKAGE: torch==1.11.0+cpu
-        PYTORCH_LIGHTNING_PACKAGE: pytorch-lightning==1.3.8
+        PYTORCH_LIGHTNING_PACKAGE: pytorch-lightning==1.5.9
         TORCHVISION_PACKAGE: torchvision==0.12.0+cpu
         MXNET_PACKAGE: mxnet==1.9.0
         PYSPARK_PACKAGE: pyspark==3.2.1
@@ -65,7 +65,6 @@ services:
         TENSORFLOW_PACKAGE: tensorflow==1.15.5
         KERAS_PACKAGE: keras==2.2.4
         PYTORCH_PACKAGE: torch==1.8.1+cpu
-        PYTORCH_LIGHTNING_PACKAGE: pytorch_lightning==1.3.8
         TORCHVISION_PACKAGE: torchvision==0.9.1+cpu
         MXNET_PACKAGE: mxnet==1.5.1.post0
   test-cpu-gloo-py3_8-tf2_6_3-keras2_6_0-torch1_9_1-mxnet1_7_0_p2-pyspark3_2_1:
@@ -95,7 +94,7 @@ services:
         KERAS_PACKAGE: None
         PYTORCH_PACKAGE: torch-nightly
         TORCHVISION_PACKAGE: torchvision
-        PYTORCH_LIGHTNING_PACKAGE: pytorch_lightning==1.3.8
+        PYTORCH_LIGHTNING_PACKAGE: pytorch-lightning==1.5.9
         MXNET_PACKAGE: mxnet-nightly
   # these are the lowest framework versions that Horovod compiles with, but they are not tested
   test-cpu-gloo-py3_7-tfmin-kerasmin-torchmin-mxnetmin-pysparkmin:
@@ -170,7 +169,7 @@ services:
         TENSORFLOW_PACKAGE: tensorflow-gpu==1.15.5
         KERAS_PACKAGE: keras==2.2.4
         PYTORCH_PACKAGE: torch==1.8.1+cu101
-        PYTORCH_LIGHTNING_PACKAGE: pytorch_lightning==1.3.8
+        PYTORCH_LIGHTNING_PACKAGE: pytorch-lightning==1.5.9
         TORCHVISION_PACKAGE: torchvision==0.9.1+cu101
         MXNET_PACKAGE: mxnet-cu100==1.5.1.post0
   # here we deviate from mxnet==1.7.0.post2 as there is none for cu101, only post1
@@ -185,7 +184,7 @@ services:
         TENSORFLOW_PACKAGE: tensorflow-gpu==2.6.3
         KERAS_PACKAGE: keras==2.6.0
         PYTORCH_PACKAGE: torch==1.9.1+cu102
-        PYTORCH_LIGHTNING_PACKAGE: pytorch_lightning==1.3.8
+        PYTORCH_LIGHTNING_PACKAGE: pytorch-lightning==1.5.9
         TORCHVISION_PACKAGE: torchvision==0.10.1+cu102
         MXNET_PACKAGE: mxnet-cu101==1.7.0.post1
   test-gpu-gloo-py3_8-tf2_7_1-keras2_7_0-torch1_10_2-mxnet1_8_0_p0-pyspark3_2_1:
@@ -198,7 +197,7 @@ services:
         TENSORFLOW_PACKAGE: tensorflow-gpu==2.7.1
         KERAS_PACKAGE: keras==2.7.0
         PYTORCH_PACKAGE: torch==1.10.2+cu111
-        PYTORCH_LIGHTNING_PACKAGE: pytorch_lightning==1.3.8
+        PYTORCH_LIGHTNING_PACKAGE: pytorch-lightning==1.5.9
         TORCHVISION_PACKAGE: torchvision==0.11.3+cu111
         MXNET_PACKAGE: mxnet-cu112==1.8.0.post0
   test-gpu-openmpi-gloo-py3_8-tf2_8_0-keras2_8_0-torch1_11_0-mxnet1_9_0-pyspark3_2_1:
@@ -212,7 +211,7 @@ services:
         TENSORFLOW_PACKAGE: tensorflow-gpu==2.8.0
         KERAS_PACKAGE: keras==2.8.0
         PYTORCH_PACKAGE: torch==1.11.0+cu113
-        PYTORCH_LIGHTNING_PACKAGE: pytorch-lightning==1.3.8
+        PYTORCH_LIGHTNING_PACKAGE: pytorch-lightning==1.5.9
         TORCHVISION_PACKAGE: torchvision==0.12.0+cu113
         MXNET_PACKAGE: mxnet-cu112==1.9.0
   test-gpu-gloo-py3_8-tfhead-keras_none-torchhead-mxnethead-pyspark3_2_1:
@@ -225,7 +224,7 @@ services:
         TENSORFLOW_PACKAGE: tf-nightly-gpu
         KERAS_PACKAGE: None
         PYTORCH_PACKAGE: torch-nightly-cu111
-        PYTORCH_LIGHTNING_PACKAGE: pytorch_lightning==1.3.8
+        PYTORCH_LIGHTNING_PACKAGE: pytorch-lightning==1.5.9
         TORCHVISION_PACKAGE: torchvision
         MXNET_PACKAGE: mxnet-nightly-cu112
   # these are the lowest framework versions that Horovod compiles with, but they are not tested
@@ -257,7 +256,7 @@ services:
         TENSORFLOW_PACKAGE: tensorflow-gpu==2.8.0
         KERAS_PACKAGE: keras==2.8.0
         PYTORCH_PACKAGE: torch==1.11.0+cu113
-        PYTORCH_LIGHTNING_PACKAGE: pytorch_lightning==1.3.8
+        PYTORCH_LIGHTNING_PACKAGE: pytorch-lightning==1.5.9
         TORCHVISION_PACKAGE: torchvision==0.12.0+cu113
         MXNET_PACKAGE: mxnet-cu112==1.9.0
         HOROVOD_BUILD_FLAGS: ""

docker/horovod-cpu/Dockerfile

@@ -3,7 +3,7 @@ FROM ubuntu:20.04
 # TensorFlow version is tightly coupled to CUDA and cuDNN so it should be selected carefully
 ARG TENSORFLOW_VERSION=2.5.0
 ARG PYTORCH_VERSION=1.8.1
-ARG PYTORCH_LIGHTNING_VERSION=1.2.9
+ARG PYTORCH_LIGHTNING_VERSION=1.5.9
 ARG TORCHVISION_VERSION=0.9.1
 ARG MXNET_VERSION=1.8.0.post0
 

docker/horovod-ray/Dockerfile

@@ -7,7 +7,7 @@ FROM rayproject/ray:${RAY_DOCKER_VERSION}-gpu
 ARG RAY_DOCKER_VERSION=nightly
 ARG TENSORFLOW_VERSION=2.5.0
 ARG PYTORCH_VERSION=1.8.1+cu111
-ARG PYTORCH_LIGHTNING_VERSION=1.2.9
+ARG PYTORCH_LIGHTNING_VERSION=1.5.9
 ARG TORCHVISION_VERSION=0.9.1+cu111
 
 # to avoid interaction with apt-get

docker/horovod/Dockerfile

@@ -7,7 +7,7 @@ FROM nvidia/cuda:${CUDA_DOCKER_VERSION}
 ARG CUDA_DOCKER_VERSION=11.2.2-devel-ubuntu18.04
 ARG TENSORFLOW_VERSION=2.5.0
 ARG PYTORCH_VERSION=1.8.1+cu111
-ARG PYTORCH_LIGHTNING_VERSION=1.2.9
+ARG PYTORCH_LIGHTNING_VERSION=1.5.9
 ARG TORCHVISION_VERSION=0.9.1+cu111
 ARG CUDNN_VERSION=8.1.1.33-1+cuda11.2
 ARG NCCL_VERSION=2.8.4-1+cuda11.2

setup.py

@@ -162,7 +162,7 @@ def build_extensions(self):
 tensorflow_gpu_require_list = ['tensorflow-gpu']
 keras_require_list = ['keras>=2.0.8,!=2.0.9,!=2.1.0,!=2.1.1']
 # pytorch-lightning 1.3.8 is a stable version to work with horovod
-pytorch_require_list = ['torch', 'pytorch_lightning==1.3.8']
+pytorch_require_list = ['torch', 'pytorch_lightning>=1.3.8,<1.5.10']
 mxnet_require_list = ['mxnet>=1.4.1']
 pyspark_require_list = ['pyspark>=2.3.2;python_version<"3.8"',
                         'pyspark>=3.0.0;python_version>="3.8"']
@@ -189,7 +189,7 @@ def build_extensions(self):
                     'keras==2.3.1',
                     'torch==1.4.0',
                     'torchvision==0.5.0',
-                    'pytorch_lightning>=1.3.8',
+                    'pytorch_lightning>=1.3.8,<1.5.10',
                     'mxnet==1.5.0',
                     'pyspark==3.0.1'] + spark_require_list
 # torchvision 0.5.0 depends on torch==1.4.0