PT 1.5 & TF 1.15.0 w/cuda10.1 + ORT 1.2.0 (#347)

* PyTorch 1.5 & Redis 5.0.8

* fixes 1

* ONNXRuntime 1.2.0

* GPU-related changes

* fixes 1

Co-authored-by: filipe oliveira <[email protected]>

Author: rafie

Dockerfile

@@ -1,6 +1,6 @@
-# BUILD redisfab/redisai-cpu-${OSNICK}:${VERSION}-${ARCH}
+# BUILD redisfab/redisai:${VERSION}-cpu-${ARCH}-${OSNICK}
 
-ARG REDIS_VER=5.0.7
+ARG REDIS_VER=5.0.8
 
 # OSNICK=bionic|stretch|buster
 ARG OSNICK=buster
@@ -39,7 +39,6 @@ RUN if [ "$PACK" = "1" ]; then make -C opt pack; fi
 RUN if [ "$TEST" = "1" ]; then make -C opt test $BUILD_ARGS NO_LFS=1; fi
 
 #----------------------------------------------------------------------------------------------
-# FROM redis:latest
 FROM redisfab/redis:${REDIS_VER}-${ARCH}-${OSNICK}
 
 RUN set -e; apt-get -qq update; apt-get -q install -y libgomp1

Dockerfile.gpu

@@ -1,6 +1,6 @@
-# BUILD redisfab/redisai-gpu:${VERSION}-${ARCH}-${OSNICK}
+# BUILD redisfab/redisai:${VERSION}-gpu-${ARCH}-${OSNICK}
 
-ARG REDIS_VER=5.0.7
+ARG REDIS_VER=5.0.8
 
 # OSNICK=bionic|centos7|centos6
 ARG OSNICK=bionic
@@ -10,7 +10,7 @@ ARG ARCH=x64
 
 #----------------------------------------------------------------------------------------------
 FROM redisfab/redis:${REDIS_VER}-${ARCH}-${OSNICK} AS redis
-FROM nvidia/cuda:10.0-cudnn7-devel-ubuntu18.04 AS builder
+FROM nvidia/cuda:10.1-cudnn7-devel-ubuntu18.04 AS builder
 
 # ENV DEPS "git ca-certificates wget unzip cmake libgomp1 patchelf coreutils tcl libjemalloc-dev"
 ENV NVIDIA_VISIBLE_DEVICES all
@@ -40,7 +40,7 @@ RUN if [ "$PACK" = "1" ]; then make -C opt pack GPU=1; fi
 RUN if [ "$TEST" = "1" ]; then TEST="" make -C opt test GPU=1 $BUILD_ARGS NO_LFS=1; fi
 
 #----------------------------------------------------------------------------------------------
-FROM nvidia/cuda:10.0-cudnn7-runtime-ubuntu18.04
+FROM nvidia/cuda:10.1-cudnn7-runtime-ubuntu18.04
 
 RUN set -e; apt-get -qq update; apt-get -q install -y libgomp1
 

Dockerfile.gpu-test

@@ -1,6 +1,6 @@
-# BUILD redisfab/redisai-gpu:${VERSION}-${ARCH}-${OSNICK}-test
+# BUILD redisfab/redisai:${VERSION}-gpu-${ARCH}-${OSNICK}-test
 
-ARG REDIS_VER=5.0.7
+ARG REDIS_VER=5.0.8
 
 # OSNICK=bionic|centos7|centos6
 ARG OSNICK=bionic
@@ -10,7 +10,7 @@ ARG ARCH=x64
 
 #----------------------------------------------------------------------------------------------
 FROM redisfab/redis:${REDIS_VER}-${ARCH}-${OSNICK} AS redis
-FROM nvidia/cuda:10.0-cudnn7-devel-ubuntu18.04 AS builder
+FROM nvidia/cuda:10.1-cudnn7-devel-ubuntu18.04 AS builder
 
 SHELL ["/bin/bash", "-c"]
 
@@ -35,7 +35,7 @@ COPY ./get_deps.sh .
 RUN set -e ;\
 	. venv/bin/activate ;\
 	if [[ -z $DEPS_ARGS ]]; then \
-		./get_deps.sh gpu ;\
+		VERBOSE=1 ./get_deps.sh gpu ;\
 	else \
 		env $DEPS_ARGS ./get_deps.sh gpu ;\
 	fi

get_deps.sh

@@ -14,13 +14,15 @@ if [[ $1 == --help || $1 == help ]]; then
 		[ARGVARS...] get_deps.sh [cpu|gpu] [--help|help]
 		
 		Argument variables:
-		VERBOSE=1       Print commands
-		FORCE=1         Download even if present
-		WITH_DLPACK=0   Skip dlpack
-		WITH_TF=0       Skip Tensorflow
-		WITH_TFLITE=0   Skip Tensorflow
-		WITH_PT=0       Skip PyTorch
-		WITH_ORT=0      Skip OnnxRuntime
+		CPU=1              Get CPU dependencies
+		GPU=1              Get GPU dependencies
+		VERBOSE=1          Print commands
+		FORCE=1            Download even if present
+		WITH_DLPACK=0      Skip dlpack
+		WITH_TF=0|S3       Skip Tensorflow or download from S3 repo
+		WITH_TFLITE=0|S3   Skip TensorflowLite  or download from S3 repo
+		WITH_PT=0|S3       Skip PyTorch or download from S3 repo
+		WITH_ORT=0|S3      Skip OnnxRuntime or download from S3 repo
 
 	END
 	exit 0
@@ -29,10 +31,10 @@ fi
 set -e
 [[ $VERBOSE == 1 ]] && set -x
 
-if [[ "$1" == "cpu" ]]; then
+if [[ "$1" == "cpu" || $CPU == 1 ]]; then
 	GPU=0
 	DEVICE=cpu
-elif [[ "$1" == "gpu" ]]; then
+elif [[ "$1" == "gpu" || $GPU == 1 ]]; then
 	GPU=1
 	DEVICE=gpu
 else
@@ -100,7 +102,13 @@ if [[ $WITH_TF != 0 ]]; then
 			if [[ $ARCH == x64 ]]; then
 				TF_VERSION=1.15.0
 				TF_ARCH=x86_64
-				LIBTF_URL_BASE=https://storage.googleapis.com/tensorflow/libtensorflow
+				
+				# special case for 1.15.0 as official version does not suport CUDA 10.1
+				if [[ $WITH_TF == S3 || $GPU == 1 ]]; then
+					LIBTF_URL_BASE=https://s3.amazonaws.com/redismodules/tensorflow
+				else
+					LIBTF_URL_BASE=https://storage.googleapis.com/tensorflow/libtensorflow
+				fi
 			elif [[ $ARCH == arm64v8 ]]; then
 				TF_VERSION=1.15.0
 				TF_ARCH=arm64
@@ -115,7 +123,11 @@ if [[ $WITH_TF != 0 ]]; then
 			TF_OS=darwin
 			TF_BUILD=cpu
 			TF_ARCH=x86_64
-			LIBTF_URL_BASE=https://storage.googleapis.com/tensorflow/libtensorflow
+			if [[ $WITH_TF == S3 ]]; then
+				LIBTF_URL_BASE=https://s3.amazonaws.com/redismodules/tensorflow
+			else
+				LIBTF_URL_BASE=https://storage.googleapis.com/tensorflow/libtensorflow
+			fi
 		fi
 
 		LIBTF_ARCHIVE=libtensorflow-${TF_BUILD}-${TF_OS}-${TF_ARCH}-${TF_VERSION}.tar.gz
@@ -186,7 +198,7 @@ fi # WITH_TFLITE
 
 ###################################################################################### LIBTORCH
 
-PT_VERSION="1.4.0"
+PT_VERSION="1.5.0"
 
 if [[ $WITH_PT != 0 ]]; then
 	[[ $FORCE == 1 ]] && rm -rf $LIBTORCH
@@ -270,7 +282,7 @@ fi
 
 ################################################################################### ONNXRUNTIME
 
-ORT_VERSION="1.0.0"
+ORT_VERSION="1.2.0"
 
 if [[ $WITH_ORT != 0 ]]; then
 	[[ $FORCE == 1 ]] && rm -rf $ONNXRUNTIME

opt/Makefile

@@ -44,13 +44,14 @@ make test         # run tests
   CALLGRIND|CGD=1   # test with Callgrind (implies DEBUG=1)
 make cov-upload   # upload coverage data to codecov.io (requires CODECOV_TOKEN)
 
-make pack       # create installation packages
-  PACK_DEPS=0     # do not pack dependencies
-  INTO=dir        # place artifacts in `dir`
-  BRANCH=name     # use `name` as branch name
-make deploy     # copy packages to S3
-make release    # release a version
-make docker     # build docker image
+make pack             # create installation packages
+  PACK_DEPS=0           # do not pack dependencies
+  INTO=dir              # place artifacts in `dir`
+  BRANCH=name           # use `name` as branch name
+make deploy           # copy packages to S3
+make release          # release a version
+make docker           # build docker image
+make docker-gpu-test  # run GPU tests
 
 fetch and build options:
   WITH_TF=0     # SKip TensofFlow
@@ -204,3 +205,9 @@ callgrind:
 
 docker:
 	$(SHOW)docker build -t redisai --build-arg TEST=1 --build-arg PACK=1 ..
+
+docker-gpu-test:
+	$(SHOW)set -e ;\
+		cd .. ;\
+		docker build -f Dockerfile.gpu-test -t redisai-test:latest-$(DEVICE)-$(ARCH)-$(OS) .
+		docker run --gpus all -it redisai-test:latest-$(DEVICE)-$(ARCH)-$(OS)

opt/build/libtorch/Makefile

@@ -1,7 +1,7 @@
 
 ROOT=../../..
 
-VERSION ?= 1.4.0
+VERSION ?= 1.5.0
 OSNICK ?= buster
 
 #----------------------------------------------------------------------------------------------

opt/build/libtorch/repack.sh

@@ -11,15 +11,24 @@ ROOT=$(realpath $ROOT)
 
 if [[ "$1" == "cpu" ]]; then
 	GPU=0
+	DEVICE=cpu
 elif [[ "$1" == "gpu" ]]; then
 	GPU=1
+	DEVICE=gpu
 else
 	GPU=${GPU:-0}
+	if [[ $GPU == 1 ]]; then
+		DEVICE=gpu
+	else
+		DEVICE=cpu
+	fi
 fi
 
 OS=$(python3 $ROOT/opt/readies/bin/platform --os)
 ARCH=$(python3 $ROOT/opt/readies/bin/platform --arch)
 
+TARGET_DIR=$ROOT/deps//$OS-$ARCH-$DEVICE
+
 # avoid wget warnings on macOS
 [[ $OS == macosx ]] && export LC_ALL=en_US.UTF-8
 
@@ -74,5 +83,5 @@ if [[ $OS == linux ]]; then
 fi
 
 unzip -q -o $LIBTORCH_ZIP
-tar czf libtorch-${PT_BUILD}-${PT_OS}-${PT_ARCH}-${PT_VERSION}.tar.gz libtorch/
+tar czf $TARGET_DIR/libtorch-${PT_BUILD}-${PT_OS}-${PT_ARCH}-${PT_VERSION}.tar.gz libtorch/
 rm -rf libtorch/ $LIBTORCH_ZIP

src/libtorch_c/CMakeLists.txt

@@ -1,3 +1,3 @@
 add_library(torch_c STATIC torch_c.cpp)
 target_link_libraries(torch_c "${TORCH_LIBRARIES}")
-set_property(TARGET torch_c PROPERTY CXX_STANDARD 11)
+set_property(TARGET torch_c PROPERTY CXX_STANDARD 14)

src/libtorch_c/torch_c.cpp

@@ -1,11 +1,12 @@
 #include "torch_c.h"
-#include <torch/torch.h>
-#include <torch/csrc/jit/import.h>
-#include <torch/csrc/jit/script/compilation_unit.h>
+#include "torch/torch.h"
+#include "torch/csrc/jit/serialization/import.h"
+#include "torch/csrc/jit/api/compilation_unit.h"
+#include "ATen/Functions.h"
+
 #include <iostream>
 #include <sstream>
 
-#include <ATen/Functions.h>
 
 namespace {