Skip to content

chore: Upgrade to CUDA 12.1 #2020

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 6 commits into from
Jul 3, 2023
Merged

chore: Upgrade to CUDA 12.1 #2020

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
e9e50ad
chore: Upgrade to CUDA 12.1 and CuDNN 8.9
gs-olive Jun 13, 2023
27651e8
Downgrade to CuDNN 8.9.0
gs-olive Jun 15, 2023
9d33ee7
Revert "Downgrade to CuDNN 8.9.0"
gs-olive Jun 22, 2023
0e5e62e
with_pypi_cudnn_experiment
gs-olive Jun 27, 2023
16f5d0d
Revert "with_pypi_cudnn_experiment"
gs-olive Jun 27, 2023
e6cc509
Roll back to cuDNN 8.8
gs-olive Jun 28, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
94 changes: 40 additions & 54 deletions .circleci/config.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -92,17 +92,9 @@ commands:
key: cmake-<< parameters.cache-key >>-<< parameters.version >>
paths:
- << parameters.install-dir >>
uninstall-cuda:
description: "Uninstall CUDA-11.4"
steps:
- run:
name: "Uninstall CUDA-11.4"
command: |
sudo apt-get --purge remove "cuda*" "nsight*"
sudo apt-get --purge remove "*nvidia*"

install-cudnn:
description: "Install CUDNN 8.8.0"
description: "Install CUDNN 8.8.1"
Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Based on feedback and discussion with @HolyWu and @narendasan, the CuDNN was only upgraded in the patch version to avoid issues with PyTorch and CuDNN 8.9, while preserving the libtorch.so version installed by Torch. A subsequent upgrade to the newest CuDNN is to come when PyTorch support is more robust.

parameters:
os:
type: string
Expand All @@ -112,10 +104,10 @@ commands:
default: "x86_64"
cudnn-version:
type: string
default: "8.8.0.121"
default: "8.8.1.3"
cuda-version:
type: string
default: "cuda11.8"
default: "cuda12.0"
steps:
- run:
name: Install CUDNN
Expand Down Expand Up @@ -149,13 +141,13 @@ commands:
default: "amd64"
cuda-pkg-name:
type: string
default: "cuda-toolkit-11-8"
default: "cuda-toolkit-12-1"
cuda-pkg-version:
type: string
default: "11-8"
default: "12-1"
cuda-version:
type: string
default: "11.8.0"
default: "12.1.1"
steps:
- run:
name: Install CUDA
Expand Down Expand Up @@ -186,21 +178,12 @@ commands:
architecture:
type: string
default: "amd64"
cuda-pkg-name:
type: string
default: "cuda-toolkit-11-8"
cuda-pkg-version:
type: string
default: "11-8"
cuda-version:
type: string
default: "11.8.0"
cuda-string-version:
type: string
default: "cuda11.8"
default: "cuda12.0"
Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This version is cuda12.0 and not cuda12.1 like the others because the apt Ubuntu package directory does not have nvinfer libraries specifically for cuda12.1.

cudnn-version:
type: string
default: "8.8.0.121"
default: "8.8.1.3"
trt-version-short:
type: string
default: "8.6.1"
Expand Down Expand Up @@ -252,7 +235,7 @@ commands:
default: "8.6.1"
cudnn-version-long:
type: string
default: "8.8.0.121"
default: "8.8.1.3"
steps:
- run:
name: Set up python environment
Expand All @@ -261,27 +244,30 @@ commands:
pip3 install wheel setuptools
pip3 install nvidia-pyindex
pip3 install tabulate
pip3 install tensorrt==<< parameters.trt-version-long >> nvidia-cudnn-cu11==<< parameters.cudnn-version-long >>
pip3 install tensorrt==<< parameters.trt-version-long >> nvidia-cudnn-cu12==<< parameters.cudnn-version-long >>
pip3 install pytest parameterized expecttest nox

install-torch-from-index:
description: "Install Pytorch"
parameters:
torch-build:
type: string
default: "2.1.0.dev20230605+cu118"
default: "2.1.0.dev20230619+cu121"
torchvision-build:
type: string
default: "0.16.0.dev20230605+cu118"
default: "0.16.0.dev20230619+cu121"
torch-build-index:
type: string
default: "https://download.pytorch.org/whl/nightly/cu118"
default: "https://download.pytorch.org/whl/nightly/cu121"
torchvision-build-index:
type: string
default: "https://download.pytorch.org/whl/nightly/cu121"
steps:
- run:
name: Install Torch
command: |
pip3 install --upgrade pip
pip3 install torch==<< parameters.torch-build >> torchvision==<< parameters.torchvision-build >> --extra-index-url << parameters.torch-build-index >>
pip3 install torch==<< parameters.torch-build >> torchvision==<< parameters.torchvision-build >> --extra-index-url << parameters.torch-build-index >> --extra-index-url << parameters.torchvision-build-index >>

build-py:
description: "Build the torch-tensorrt python release (pre-cxx11-abi)"
Expand All @@ -293,7 +279,7 @@ commands:
- run:
name: Build torch-tensorrt python release (pre-cxx11-abi)
command: |
export CUDA_HOME=/usr/local/cuda-11.8/
export CUDA_HOME=/usr/local/cuda-12.1/
mv toolchains/ci_workspaces/WORKSPACE.<< parameters.platform >> WORKSPACE
cd py
python3 -m pip install wheel setuptools
Expand All @@ -313,7 +299,7 @@ commands:
- run:
name: Build torch-tensorrt python legacy release (pre-cxx11-abi)
command: |
export CUDA_HOME=/usr/local/cuda-11.8/
export CUDA_HOME=/usr/local/cuda-12.1/
mv toolchains/ci_workspaces/WORKSPACE.<< parameters.platform >> WORKSPACE
cd py
python3 -m pip install wheel setuptools
Expand Down Expand Up @@ -345,7 +331,7 @@ commands:
- run:
name: Build torch-tensorrt python release package
command: |
export CUDA_HOME=/usr/local/cuda-11.8/
export CUDA_HOME=/usr/local/cuda-12.1/
cd ~/project/py
python3 setup.py bdist_wheel --use-cxx11-abi --release
python3 setup.py install --use-cxx11-abi --release
Expand All @@ -357,7 +343,7 @@ commands:
- run:
name: Build torch-tensorrt python package
command: |
export CUDA_HOME=/usr/local/cuda-11.8/
export CUDA_HOME=/usr/local/cuda-12.1/
cd ~/project/py
python3 setup.py bdist_wheel --use-cxx11-abi
python3 setup.py install --use-cxx11-abi
Expand All @@ -375,7 +361,7 @@ commands:
- run:
name: Build torch-tensorrt python release with only the fx backend
command: |
export CUDA_HOME=/usr/local/cuda-11.8/
export CUDA_HOME=/usr/local/cuda-12.1/
mv toolchains/ci_workspaces/WORKSPACE.<< parameters.platform >> WORKSPACE
cd py
python3 -m pip install wheel setuptools
Expand Down Expand Up @@ -447,7 +433,7 @@ commands:
name: Build torch-tensorrt library with CMake
command: |
mkdir build
export PATH=$PATH:/usr/local/cuda-11.8/bin
export PATH=$PATH:/usr/local/cuda-12.1/bin
~/cmake/bin/cmake -S. -Bbuild \
-DCMAKE_MODULE_PATH=cmake/Module \
-DTorch_DIR=/opt/circleci/.pyenv/versions/3.9.4/lib/python3.9/site-packages/torch/share/cmake/Torch \
Expand Down Expand Up @@ -499,7 +485,7 @@ commands:
name: Run core / C++ tests
no_output_timeout: 15m
environment:
LD_LIBRARY_PATH: "/opt/circleci/.pyenv/versions/3.9.4/lib/python3.9/site-packages/torch_tensorrt.libs:/home/circleci/project/bazel-project/external/libtorch_pre_cxx11_abi/lib/:/home/circleci/project/bazel-project/external/tensorrt/lib/:/usr/local/cuda-11.8/lib64/:$LD_LIBRARY_PATH"
LD_LIBRARY_PATH: "/opt/circleci/.pyenv/versions/3.9.4/lib/python3.9/site-packages/torch_tensorrt.libs:/home/circleci/project/bazel-project/external/libtorch_pre_cxx11_abi/lib/:/home/circleci/project/bazel-project/external/tensorrt/lib/:/usr/local/cuda-12.1/lib64/:$LD_LIBRARY_PATH"
command: |
set -e
mv toolchains/ci_workspaces/WORKSPACE.<< parameters.platform >> WORKSPACE
Expand Down Expand Up @@ -529,7 +515,7 @@ commands:
environment:
USE_HOST_DEPS: "1"
PYT_PATH: "/opt/circleci/.pyenv/versions/3.9.4/lib/python3.9/site-packages/"
LD_LIBRARY_PATH: "/usr/lib/x86_64-linux-gnu/:/usr/local/cuda-11.8/lib64/:$LD_LIBRARY_PATH"
LD_LIBRARY_PATH: "/usr/lib/x86_64-linux-gnu/:/usr/local/cuda-12.1/lib64/:$LD_LIBRARY_PATH"
command: |
set -e
mkdir -p /tmp/artifacts/test_results
Expand Down Expand Up @@ -819,7 +805,7 @@ jobs:
type: boolean
default: false
machine:
image: linux-cuda-11:2023.02.1
image: linux-cuda-12:2023.05.1
resource_class: gpu.nvidia.small
steps:
- checkout
Expand Down Expand Up @@ -881,7 +867,7 @@ jobs:
cudnn-version:
type: string
machine:
image: linux-cuda-11:2023.02.1
image: linux-cuda-12:2023.05.1
resource_class: gpu.nvidia.large
parallelism: 4
steps:
Expand Down Expand Up @@ -922,7 +908,7 @@ jobs:
python-version:
type: string
machine:
image: linux-cuda-11:2023.02.1
image: linux-cuda-12:2023.05.1
resource_class: gpu.nvidia.large
steps:
- checkout
Expand Down Expand Up @@ -956,7 +942,7 @@ jobs:
type: string
parallelism: 8
machine:
image: linux-cuda-11:2023.02.1
image: linux-cuda-12:2023.05.1
resource_class: gpu.nvidia.large
steps:
- checkout
Expand Down Expand Up @@ -992,7 +978,7 @@ jobs:
type: string
parallelism: 8
machine:
image: linux-cuda-11:2023.02.1
image: linux-cuda-12:2023.05.1
resource_class: gpu.nvidia.large
steps:
- checkout
Expand Down Expand Up @@ -1030,7 +1016,7 @@ jobs:
python-version:
type: string
machine:
image: linux-cuda-11:2023.02.1
image: linux-cuda-12:2023.05.1
resource_class: gpu.nvidia.large
steps:
- checkout
Expand Down Expand Up @@ -1068,7 +1054,7 @@ jobs:
type: string
parallelism: 4
machine:
image: linux-cuda-11:2023.02.1
image: linux-cuda-12:2023.05.1
resource_class: gpu.nvidia.small
steps:
- when:
Expand Down Expand Up @@ -1116,7 +1102,7 @@ jobs:
torch-build-index:
type: string
machine:
image: linux-cuda-11:2023.02.1
image: linux-cuda-12:2023.05.1
resource_class: gpu.nvidia.small
steps:
- when:
Expand Down Expand Up @@ -1252,7 +1238,7 @@ jobs:
python-version:
type: string
machine:
image: linux-cuda-11:2023.02.1
image: linux-cuda-12:2023.05.1
resource_class: gpu.nvidia.small
steps:
- checkout
Expand Down Expand Up @@ -1288,7 +1274,7 @@ jobs:
torch-base-image:
type: string
machine:
image: linux-cuda-11:2023.02.1
image: linux-cuda-12:2023.05.1
resource_class: gpu.nvidia.small
steps:
- checkout
Expand Down Expand Up @@ -1316,7 +1302,7 @@ jobs:
torch-base-image:
type: string
machine:
image: linux-cuda-11:2023.02.1
image: linux-cuda-12:2023.05.1
resource_class: gpu.nvidia.small
steps:
- when:
Expand Down Expand Up @@ -1352,13 +1338,13 @@ parameters:
# Nightly platform config
torch-build:
type: string
default: "2.1.0.dev20230605+cu118"
default: "2.1.0.dev20230619+cu121"
torchvision-build:
type: string
default: "0.16.0.dev20230605+cu118"
default: "0.16.0.dev20230619+cu121"
torch-build-index:
type: string
default: "https://download.pytorch.org/whl/nightly/cu118"
default: "https://download.pytorch.org/whl/nightly/cu121"
torch-build-legacy:
type: string
default: "1.13.1+cu117"
Expand All @@ -1373,7 +1359,7 @@ parameters:
default: true
cudnn-version:
type: string
default: "8.8.0.121"
default: "8.8.1.3"
trt-version-short:
type: string
default: "8.6.1"
Expand Down
6 changes: 3 additions & 3 deletions README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -116,9 +116,9 @@ torch.jit.save(trt_ts_module, "trt_torchscript_module.ts") # save the TRT embedd
These are the following dependencies used to verify the testcases. Torch-TensorRT can work with other versions, but the tests are not guaranteed to pass.

- Bazel 5.2.0
- Libtorch 2.1.0.dev20230605 (built with CUDA 11.8)
- CUDA 11.8
- cuDNN 8.8.0
- Libtorch 2.1.0.dev20230619 (built with CUDA 12.1)
- CUDA 12.1
- cuDNN 8.8.1
- TensorRT 8.6.1

## Prebuilt Binaries and Wheel files
Expand Down
20 changes: 10 additions & 10 deletions WORKSPACE
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -41,7 +41,7 @@ local_repository(
new_local_repository(
name = "cuda",
build_file = "@//third_party/cuda:BUILD",
path = "/usr/local/cuda-11.8/",
path = "/usr/local/cuda-12.1/",
)

#############################################################################################################
Expand All @@ -51,17 +51,17 @@ new_local_repository(
http_archive(
name = "libtorch",
build_file = "@//third_party/libtorch:BUILD",
sha256 = "999becce82b73e566d0ffe010cd21fea8cf3a33f90f09dcc6b01150b820ae063",
sha256 = "5ba55259b65e071346a2b547b8d1378595f1467a39aaa923fecb09f134f1bcba",
strip_prefix = "libtorch",
urls = ["https://download.pytorch.org/libtorch/nightly/cu118/libtorch-cxx11-abi-shared-with-deps-2.1.0.dev20230605%2Bcu118.zip"],
urls = ["https://download.pytorch.org/libtorch/nightly/cu121/libtorch-cxx11-abi-shared-with-deps-2.1.0.dev20230619%2Bcu121.zip"],
)

http_archive(
name = "libtorch_pre_cxx11_abi",
build_file = "@//third_party/libtorch:BUILD",
sha256 = "786cc728c63ea69c40bd8fb535cf8e5e1dfff1d43eaad3eb5256b9ed89c1b268",
sha256 = "8f6661bfc11597e77400e9e36cc8dd8e5e385ba82361d630d9ccede8518d7c7e",
strip_prefix = "libtorch",
urls = ["https://download.pytorch.org/libtorch/nightly/cu118/libtorch-shared-with-deps-2.1.0.dev20230605%2Bcu118.zip"],
urls = ["https://download.pytorch.org/libtorch/nightly/cu121/libtorch-shared-with-deps-2.1.0.dev20230619%2Bcu121.zip"],
)

# Download these tarballs manually from the NVIDIA website
Expand All @@ -71,20 +71,20 @@ http_archive(
http_archive(
name = "cudnn",
build_file = "@//third_party/cudnn/archive:BUILD",
sha256 = "36fff137153ef73e6ee10bfb07f4381240a86fb9fb78ce372414b528cbab2293",
strip_prefix = "cudnn-linux-x86_64-8.8.0.121_cuda11-archive",
sha256 = "79d77a769c7e7175abc7b5c2ed5c494148c0618a864138722c887f95c623777c",
strip_prefix = "cudnn-linux-x86_64-8.8.1.3_cuda12-archive",
urls = [
"https://developer.download.nvidia.com/compute/cudnn/secure/8.8.0/local_installers/11.8/cudnn-linux-x86_64-8.8.0.121_cuda11-archive.tar.xz",
"https://developer.nvidia.com/downloads/compute/cudnn/secure/8.8.1/local_installers/12.0/cudnn-linux-x86_64-8.8.1.3_cuda12-archive.tar.xz",
],
)

http_archive(
name = "tensorrt",
build_file = "@//third_party/tensorrt/archive:BUILD",
sha256 = "15bfe6053d45feec45ecc7123a9106076b0b43fa0435f242d89dca0778337759",
sha256 = "0f8157a5fc5329943b338b893591373350afa90ca81239cdadd7580cd1eba254",
strip_prefix = "TensorRT-8.6.1.6",
urls = [
"https://developer.nvidia.com/downloads/compute/machine-learning/tensorrt/secure/8.6.1/tars/TensorRT-8.6.1.6.Linux.x86_64-gnu.cuda-11.8.tar.gz",
"https://developer.nvidia.com/downloads/compute/machine-learning/tensorrt/secure/8.6.1/tars/TensorRT-8.6.1.6.Linux.x86_64-gnu.cuda-12.0.tar.gz",
],
)

Expand Down
4 changes: 2 additions & 2 deletions docker/Dockerfile
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,7 +1,7 @@
# Base image starts with CUDA
ARG BASE_IMG=nvidia/cuda:11.8.0-devel-ubuntu22.04
ARG BASE_IMG=nvidia/cuda:12.1.1-devel-ubuntu22.04
FROM ${BASE_IMG} as base
ENV BASE_IMG=nvidia/cuda:11.8.0-devel-ubuntu22.04
ENV BASE_IMG=nvidia/cuda:12.1.1-devel-ubuntu22.04

ARG TENSORRT_VERSION
ENV TENSORRT_VERSION=${TENSORRT_VERSION}
Expand Down
Loading