Skip to content

Commit 561c953

Browse files
q10q10
committed
[fbgemm_gpu] OSS build updates
- Upgrade gcc version to support newer libstdc++, which is required now that pytorch/pytorch#141035 has landed - Deprecate support for CUDA 12.1 and add support for 12.6, per changes in pytorch/pytorch#138899
1 parent e118f0c commit 561c953

19 files changed

+52
-45
lines changed

‎.github/scripts/fbgemm_gpu_build.bash

Lines changed: 3 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -192,7 +192,9 @@ __configure_fbgemm_gpu_build_cuda () {
192192
cuda_version_nvcc=$(conda run -n "${env_name}" nvcc --version)
193193
echo "[BUILD] Using the default architectures for CUDA $cuda_version_nvcc ..."
194194

195-
if [[ $cuda_version_nvcc == *"V12.1"* ]] || [[ $cuda_version_nvcc == *"V12.4"* ]]; then
195+
if [[ $cuda_version_nvcc == *"V12.1"* ]] ||
196+
[[ $cuda_version_nvcc == *"V12.4"* ]] ||
197+
[[ $cuda_version_nvcc == *"V12.6"* ]]; then
196198
# sm_90 and sm_90a are only available for CUDA 12.1+
197199
# NOTE: CUTLASS kernels for Hopper require sm_90a to be enabled
198200
# See:

‎.github/scripts/fbgemm_gpu_install.bash

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -182,7 +182,7 @@ install_fbgemm_gpu_pip () {
182182
echo " ${FUNCNAME[0]} build_env 0.8.0 cpu # Install the CPU variant, specific version from release channel"
183183
echo " ${FUNCNAME[0]} build_env release cuda/12.4.1 # Install the CUDA variant, latest version from release channel"
184184
echo " ${FUNCNAME[0]} build_env test/0.8.0 cuda/12.4.1 # Install the CUDA 12.4 variant, specific version from test channel"
185-
echo " ${FUNCNAME[0]} build_env nightly rocm/6.1 # Install the ROCM 6.1 variant, latest version from nightly channel"
185+
echo " ${FUNCNAME[0]} build_env nightly rocm/6.2 # Install the ROCM 6.2 variant, latest version from nightly channel"
186186
return 1
187187
else
188188
echo "################################################################################"

‎.github/scripts/fbgemm_gpu_test.bash

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -332,7 +332,7 @@ test_setup_conda_environment () {
332332
if [ "$pytorch_variant_type" == "" ]; then
333333
echo "Usage: ${FUNCNAME[0]} ENV_NAME COMPILER PYTHON_VERSION PYTORCH_INSTALLER PYTORCH_CHANNEL[/VERSION] PYTORCH_VARIANT_TYPE [PYTORCH_VARIANT_VERSION]"
334334
echo "Example(s):"
335-
echo " ${FUNCNAME[0]} build_env clang 3.12 pip test/0.8.0 cuda 12.1.0 # Setup environment with pytorch-test 0.8.0 for Clang + Python 3.12 + CUDA 12.1.0"
335+
echo " ${FUNCNAME[0]} build_env clang 3.12 pip test/0.8.0 cuda 12.4.1 # Setup environment with pytorch-test 0.8.0 for Clang + Python 3.12 + CUDA 12.4.1"
336336
return 1
337337
else
338338
echo "################################################################################"
@@ -494,8 +494,8 @@ test_fbgemm_gpu_setup_and_pip_install () {
494494
if [ "$variant_type" == "cuda" ] || [ "$variant_type" == "genai" ]; then
495495
local variant_versions=(
496496
11.8.0
497-
12.1.1
498497
12.4.1
498+
12.6.3
499499
)
500500
elif [ "$variant_type" == "rocm" ]; then
501501
local variant_versions=(

‎.github/scripts/test_torchrec.bash

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -29,10 +29,10 @@ usage () {
2929
echo " e.g., torchrec needs fbgemm-gpu while torchrec_nightly needs fbgemm-gpu-nightly"
3030
echo "PYTHON_VERSION : Python version (e.g., 3.10)"
3131
echo "PYTORCH_CHANNEL_NAME: PyTorch's channel name (e.g., pytorch-nightly, pytorch-test (=pre-release), pytorch (=stable release))"
32-
echo "CUDA_VERSION : PyTorch's CUDA version (e.g., 12.1)"
32+
echo "CUDA_VERSION : PyTorch's CUDA version (e.g., 12.4)"
3333
echo "FBGEMM_WHEEL_PATH : path to FBGEMM_GPU's wheel file"
3434
echo "MINICONDA_PREFIX : path to install Miniconda (default: \$HOME/miniconda)"
35-
echo "Example: Python 3.10 + PyTorch nightly (CUDA 12.1), install miniconda at \$HOME/miniconda, using dist/fbgemm_gpu_nightly.whl"
35+
echo "Example: Python 3.10 + PyTorch nightly (CUDA 12.4), install miniconda at \$HOME/miniconda, using dist/fbgemm_gpu_nightly.whl"
3636
# shellcheck disable=SC2086
3737
echo " bash $(basename ${BASH_SOURCE[0]}) -v -o torchrec_nightly -p 3.10 -P pytorch-nightly -c 11.7 -w dist/fbgemm_gpu_nightly.whl"
3838
}

‎.github/scripts/utils_build.bash

Lines changed: 10 additions & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -109,14 +109,18 @@ __conda_install_gcc () {
109109
# shellcheck disable=SC2155
110110
local env_prefix=$(env_name_or_prefix "${env_name}")
111111

112-
# NOTE: g++ 10.x is installed by default instead of 11.x+ becaue 11.x+ builds
113-
# binaries that reference GLIBCXX_3.4.29, which may not be available on
114-
# systems with older versions of libstdc++.so.6 such as CentOS Stream 8 and
115-
# Ubuntu 20.04. However, if libfolly is used, GLIBCXX_3.4.30+ will be
116-
# required, which will require 11.x+.
112+
# NOTE: Previously, g++ 10.x is installed by default instead of 11.x+ because
113+
# 11.x+ builds binaries that reference GLIBCXX_3.4.29, which may not be
114+
# available on systems with older versions of libstdc++.so.6 such as CentOS
115+
# Stream 8 and Ubuntu 20.04.
116+
#
117+
# However, since https://github.com/pytorch/pytorch/pull/141035 landed, g++
118+
# 11.1+ became a requirement, since std::bit_cast is only available with
119+
# libstdc++ 11.1+. See for details:
120+
# https://gcc.gnu.org/onlinedocs/libstdc++/manual/status.html#manual.intro.status.iso
117121
#
118122
# shellcheck disable=SC2155
119-
local gcc_version="${GCC_VERSION:-10.4.0}"
123+
local gcc_version="${GCC_VERSION:-11.4.0}"
120124

121125
echo "[INSTALL] Installing GCC (${gcc_version}, ${archname}) through Conda ..."
122126
# shellcheck disable=SC2086

‎.github/scripts/utils_cuda.bash

Lines changed: 2 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -59,7 +59,7 @@ install_cuda () {
5959

6060
# Ensure that the libraries are properly installed
6161
(test_filepath "${env_name}" libcuda.so) || return 1
62-
(test_filepath "${env_name}" libnvToolsExt.so) || return 1
62+
# (test_filepath "${env_name}" libnvToolsExt.so) || return 1
6363
(test_filepath "${env_name}" libnvidia-ml.so) || return 1
6464

6565
echo "[INSTALL] Appending libcuda.so path to LD_LIBRARY_PATH ..."
@@ -163,6 +163,7 @@ install_cudnn () {
163163
["118"]="https://developer.download.nvidia.com/compute/redist/cudnn/v8.7.0/local_installers/11.8/cudnn-${PLATFORM_NAME_LC}-8.7.0.84_cuda11-archive.tar.xz"
164164
["121"]="https://developer.download.nvidia.com/compute/cudnn/redist/cudnn/linux-x86_64/cudnn-linux-x86_64-8.9.2.26_cuda12-archive.tar.xz"
165165
["124"]="https://developer.download.nvidia.com/compute/cudnn/redist/cudnn/linux-x86_64/cudnn-linux-x86_64-8.9.2.26_cuda12-archive.tar.xz"
166+
["126"]="https://developer.download.nvidia.com/compute/cudnn/redist/cudnn/linux-x86_64/cudnn-linux-x86_64-9.5.1.17_cuda12-archive.tar.xz"
166167
)
167168

168169
# Split version string by dot into array, i.e. 11.7.1 => [11, 7, 1]

‎.github/scripts/utils_pip.bash

Lines changed: 5 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -41,7 +41,7 @@ __export_package_channel_info () {
4141
__export_package_variant_info () {
4242
local package_variant_type_version="$1"
4343

44-
local FALLBACK_VERSION_CUDA="12.1.1"
44+
local FALLBACK_VERSION_CUDA="12.4.1"
4545
local FALLBACK_VERSION_ROCM="6.0.2"
4646

4747
if [ "$package_variant_type_version" == "cuda" ]; then
@@ -60,7 +60,7 @@ __export_package_variant_info () {
6060
local variant_version=""
6161

6262
else
63-
# Split along '/', e.g. cuda/12.1.0
63+
# Split along '/', e.g. cuda/12.4.0
6464
# shellcheck disable=SC2207
6565
local package_variant_type_version_arr=($(echo "${package_variant_type_version}" | tr '/' '\n'))
6666
local variant_type="${package_variant_type_version_arr[0]}"
@@ -71,7 +71,7 @@ __export_package_variant_info () {
7171
local cuda_version="${variant_version:-${FALLBACK_VERSION_CUDA}}"
7272
# shellcheck disable=SC2206
7373
local cuda_version_arr=(${cuda_version//./ })
74-
# Convert, i.e. cuda 12.1.0 => cu121
74+
# Convert, i.e. cuda 12.4.0 => cu124
7575
local variant_type="cu"
7676
local variant_version="${cuda_version_arr[0]}${cuda_version_arr[1]}"
7777

@@ -204,7 +204,7 @@ install_from_pytorch_pip () {
204204
echo "Example(s):"
205205
echo " ${FUNCNAME[0]} build_env torch 1.11.0 cpu # Install the CPU variant, specific version from release channel"
206206
echo " ${FUNCNAME[0]} build_env torch release cpu # Install the CPU variant, latest version from release channel"
207-
echo " ${FUNCNAME[0]} build_env fbgemm_gpu test/0.8.0 cuda/12.1.0 # Install the CUDA 12.1 variant, specific version from test channel"
207+
echo " ${FUNCNAME[0]} build_env fbgemm_gpu test/0.8.0 cuda/12.4.0 # Install the CUDA 12.4 variant, specific version from test channel"
208208
echo " ${FUNCNAME[0]} build_env fbgemm_gpu nightly rocm/6.1 # Install the ROCM 6.1 variant, latest version from nightly channel"
209209
echo " ${FUNCNAME[0]} build_env pytorch_triton 1.11.0 # Install specific version from release channel"
210210
echo " ${FUNCNAME[0]} build_env pytorch_triton release # Install latest version from release channel"
@@ -249,7 +249,7 @@ download_from_pytorch_pip () {
249249
echo "Example(s):"
250250
echo " ${FUNCNAME[0]} build_env torch 1.11.0 cpu # Download the CPU variant, specific version from release channel"
251251
echo " ${FUNCNAME[0]} build_env torch release cpu # Download the CPU variant, latest version from release channel"
252-
echo " ${FUNCNAME[0]} build_env fbgemm_gpu test/0.8.0 cuda/12.1.0 # Download the CUDA 12.1 variant, specific version from test channel"
252+
echo " ${FUNCNAME[0]} build_env fbgemm_gpu test/0.8.0 cuda/12.4.0 # Download the CUDA 12.4 variant, specific version from test channel"
253253
echo " ${FUNCNAME[0]} build_env fbgemm_gpu nightly rocm/6.1 # Download the ROCM 6.1 variant, latest version from nightly channel"
254254
return 1
255255
else

‎.github/scripts/utils_pytorch.bash

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -112,7 +112,7 @@ install_pytorch_pip () {
112112
echo "Example(s):"
113113
echo " ${FUNCNAME[0]} build_env test/2.1.0 cpu # Install the CPU variant for a specific version"
114114
echo " ${FUNCNAME[0]} build_env release cpu # Install the CPU variant, latest release version"
115-
echo " ${FUNCNAME[0]} build_env test cuda/12.1.0 # Install the CUDA 12.1 variant, latest test version"
115+
echo " ${FUNCNAME[0]} build_env test cuda/12.4.0 # Install the CUDA 12.4 variant, latest test version"
116116
echo " ${FUNCNAME[0]} build_env nightly rocm/6.1 # Install the ROCM 6.1 variant, latest nightly version"
117117
return 1
118118
else

‎.github/scripts/utils_torchrec.bash

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -26,7 +26,7 @@ install_torchrec_pip () {
2626
echo "Example(s):"
2727
echo " ${FUNCNAME[0]} build_env test/2.1.0rc0 cpu # Install the CPU variant for a specific version"
2828
echo " ${FUNCNAME[0]} build_env release cpu # Install the CPU variant, latest release version"
29-
echo " ${FUNCNAME[0]} build_env nightly/0.9.0.dev20240716 cuda/12.1.0 # Install the CUDA 12.1 variant, nightly version"
29+
echo " ${FUNCNAME[0]} build_env nightly/0.9.0.dev20240716 cuda/12.4.0 # Install the CUDA 12.4 variant, nightly version"
3030
return 1
3131
else
3232
echo "################################################################################"

‎.github/workflows/fbgemm_ci.yml

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -144,7 +144,7 @@ jobs:
144144
run: . $PRELUDE; create_conda_environment $BUILD_ENV 3.12
145145

146146
- name: Install C/C++ Compilers
147-
run: . $PRELUDE; install_cxx_compiler $BUILD_ENV ${{ matrix.compiler }}
147+
run: . $PRELUDE; GCC_VERSION=10.4.0 install_cxx_compiler $BUILD_ENV ${{ matrix.compiler }}
148148

149149
- name: Install Build Tools
150150
run: . $PRELUDE; install_build_tools $BUILD_ENV

‎.github/workflows/fbgemm_gpu_ci_cuda.yml

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -67,7 +67,7 @@ jobs:
6767
{ arch: x86, instance: "linux.24xlarge" },
6868
]
6969
python-version: [ "3.9", "3.10", "3.11", "3.12" ]
70-
cuda-version: [ "11.8.0", "12.1.1", "12.4.1" ]
70+
cuda-version: [ "11.8.0", "12.4.1", "12.6.3" ]
7171
compiler: [ "gcc", "clang" ]
7272

7373
steps:
@@ -150,7 +150,7 @@ jobs:
150150
# { arch: x86, instance: "linux.gcp.a100" },
151151
]
152152
python-version: [ "3.9", "3.10", "3.11", "3.12" ]
153-
cuda-version: [ "11.8.0", "12.1.1", "12.4.1" ]
153+
cuda-version: [ "11.8.0", "12.4.1", "12.6.3" ]
154154
# Specify exactly ONE CUDA version for artifact publish
155155
cuda-version-publish: [ "12.4.1" ]
156156
compiler: [ "gcc", "clang" ]

‎.github/workflows/fbgemm_gpu_ci_genai.yml

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -67,7 +67,7 @@ jobs:
6767
{ arch: x86, instance: "linux.24xlarge" },
6868
]
6969
python-version: [ "3.9", "3.10", "3.11", "3.12" ]
70-
cuda-version: [ "11.8.0", "12.1.1", "12.4.1" ]
70+
cuda-version: [ "11.8.0", "12.4.1", "12.6.3" ]
7171
compiler: [ "gcc", "clang" ]
7272

7373
steps:
@@ -149,7 +149,7 @@ jobs:
149149
# { arch: x86, instance: "linux.gcp.a100" },
150150
]
151151
python-version: [ "3.9", "3.10", "3.11", "3.12" ]
152-
cuda-version: [ "11.8.0", "12.1.1", "12.4.1" ]
152+
cuda-version: [ "11.8.0", "12.4.1", "12.6.3" ]
153153
# Specify exactly ONE CUDA version for artifact publish
154154
cuda-version-publish: [ "12.4.1" ]
155155
compiler: [ "gcc", "clang" ]

‎.github/workflows/fbgemm_gpu_ci_genai_generic_infra.yml

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -54,7 +54,7 @@ jobs:
5454
{ arch: x86, instance: "ubuntu-latest" },
5555
]
5656
python-version: [ "3.9", "3.10", "3.11", "3.12" ]
57-
cuda-version: [ "11.8.0", "12.1.1", "12.4.1" ]
57+
cuda-version: [ "11.8.0", "12.4.1", "12.6.3" ]
5858
compiler: [ "gcc", "clang" ]
5959

6060
steps:
@@ -141,7 +141,7 @@ jobs:
141141
{ arch: x86, instance: "ubuntu-latest" },
142142
]
143143
python-version: [ "3.9", "3.10", "3.11", "3.12" ]
144-
cuda-version: [ "11.8.0", "12.1.1", "12.4.1" ]
144+
cuda-version: [ "11.8.0", "12.4.1", "12.6.3" ]
145145
# Specify exactly ONE CUDA version for artifact publish
146146
cuda-version-publish: [ "12.4.1" ]
147147
compiler: [ "gcc", "clang" ]

‎.github/workflows/fbgemm_gpu_ci_rocm.yml

Lines changed: 4 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -64,9 +64,9 @@ jobs:
6464
host-machine: [
6565
{ arch: x86, instance: "linux.24xlarge" },
6666
]
67-
container-image: [ "ubuntu:20.04" ]
67+
container-image: [ "ubuntu:22.04" ]
6868
python-version: [ "3.9", "3.10", "3.11", "3.12" ]
69-
rocm-version: [ "6.1" ]
69+
rocm-version: [ "6.1", "6.2" ]
7070
compiler: [ "gcc", "clang" ]
7171

7272
steps:
@@ -129,7 +129,7 @@ jobs:
129129
if: ${{ github.repository_owner == 'pytorch' }}
130130
runs-on: ${{ matrix.host-machine.instance }}
131131
container:
132-
image: "rocm/dev-ubuntu-20.04:${{ matrix.rocm-version }}-complete"
132+
image: "rocm/dev-ubuntu-22.04:${{ matrix.rocm-version }}-complete"
133133
options: --user root --device=/dev/kfd --device=/dev/dri --ipc=host --shm-size 16G --group-add video --cap-add=SYS_PTRACE --security-opt seccomp=unconfined
134134
defaults:
135135
run:
@@ -147,7 +147,7 @@ jobs:
147147
]
148148
# ROCm machines are limited, so we only test a subset of Python versions
149149
python-version: [ "3.12" ]
150-
rocm-version: [ "6.1" ]
150+
rocm-version: [ "6.2" ]
151151
compiler: [ "gcc", "clang" ]
152152
needs: build_artifact
153153

‎.github/workflows/fbgemm_gpu_pip.yml

Lines changed: 3 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -121,7 +121,7 @@ jobs:
121121
{ instance: "linux.g5.4xlarge.nvidia.gpu" },
122122
]
123123
python-version: [ "3.9", "3.10", "3.11", "3.12" ]
124-
cuda-version: [ "11.8.0", "12.1.1", "12.4.1" ]
124+
cuda-version: [ "11.8.0", "12.4.1", "12.6.3" ]
125125

126126
steps:
127127
# Cannot upgrade to actions/checkout@v4 yet because GLIBC on the instance is too old
@@ -168,7 +168,7 @@ jobs:
168168
if: ${{ github.repository_owner == 'pytorch' && (github.event_name == 'schedule' || (github.event_name == 'workflow_dispatch' && github.event.inputs.fbgemm_gpu_variant_type == 'rocm')) }}
169169
runs-on: ${{ matrix.host-machine.instance }}
170170
container:
171-
image: "rocm/dev-ubuntu-20.04:${{ matrix.rocm-version }}-complete"
171+
image: "rocm/dev-ubuntu-22.04:${{ matrix.rocm-version }}-complete"
172172
options: --user root --device=/dev/kfd --device=/dev/dri --ipc=host --shm-size 16G --group-add video --cap-add=SYS_PTRACE --security-opt seccomp=unconfined
173173
defaults:
174174
run:
@@ -186,7 +186,7 @@ jobs:
186186
]
187187
# ROCm machines are limited, so we only test a subset of Python versions
188188
python-version: [ "3.11", "3.12" ]
189-
rocm-version: [ "6.1" ]
189+
rocm-version: [ "6.2" ]
190190

191191
steps:
192192
- name: Setup Build Container

‎.github/workflows/fbgemm_gpu_release_cuda.yml

Lines changed: 3 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -34,7 +34,7 @@ on:
3434
description: CUDA Version to Use for Building Artifact
3535
type: choice
3636
required: false
37-
options: [ "11.8.0", "12.1.1", "12.4.1" ]
37+
options: [ "11.8.0", "12.4.1", "12.6.3" ]
3838
default: "12.4.1"
3939
publish_to_pypi:
4040
description: Publish Artifact to PyPI
@@ -71,7 +71,7 @@ jobs:
7171
{ arch: x86, instance: "linux.24xlarge" },
7272
]
7373
python-version: [ "3.9", "3.10", "3.11", "3.12" ]
74-
cuda-version: [ "11.8.0", "12.1.1", "12.4.1" ]
74+
cuda-version: [ "11.8.0", "12.4.1", "12.6.3" ]
7575

7676
steps:
7777
- name: Setup Build Container
@@ -144,7 +144,7 @@ jobs:
144144
{ arch: x86, instance: "linux.g5.4xlarge.nvidia.gpu" },
145145
]
146146
python-version: [ "3.9", "3.10", "3.11", "3.12" ]
147-
cuda-version: [ "11.8.0", "12.1.1", "12.4.1" ]
147+
cuda-version: [ "11.8.0", "12.4.1", "12.6.3" ]
148148
needs: build_artifact
149149

150150
steps:

‎.github/workflows/fbgemm_gpu_release_genai.yml

Lines changed: 3 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -34,7 +34,7 @@ on:
3434
description: CUDA Version to Use for Building Artifact
3535
type: choice
3636
required: false
37-
options: [ "11.8.0", "12.1.1", "12.4.1" ]
37+
options: [ "11.8.0", "12.4.1", "12.6.3" ]
3838
default: "12.4.1"
3939
publish_to_pypi:
4040
description: Publish Artifact to PyPI
@@ -71,7 +71,7 @@ jobs:
7171
{ arch: x86, instance: "linux.24xlarge" },
7272
]
7373
python-version: [ "3.9", "3.10", "3.11", "3.12" ]
74-
cuda-version: [ "11.8.0", "12.1.1", "12.4.1" ]
74+
cuda-version: [ "11.8.0", "12.4.1", "12.6.3" ]
7575

7676
steps:
7777
- name: Setup Build Container
@@ -144,7 +144,7 @@ jobs:
144144
{ arch: x86, instance: "linux.g5.4xlarge.nvidia.gpu" },
145145
]
146146
python-version: [ "3.9", "3.10", "3.11", "3.12" ]
147-
cuda-version: [ "11.8.0", "12.1.1", "12.4.1" ]
147+
cuda-version: [ "11.8.0", "12.4.1", "12.6.3" ]
148148
needs: build_artifact
149149

150150
steps:

‎fbgemm_gpu/README.md

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -9,7 +9,7 @@ PyTorch GPU operator libraries for training and inference. The library provides
99
efficient table batched embedding bag, data layout transformation, and
1010
quantization supports.
1111

12-
FBGEMM_GPU is currently tested with CUDA 12.1 and 11.8 in CI, and with PyTorch
12+
FBGEMM_GPU is currently tested with CUDA 12.4 and 11.8 in CI, and with PyTorch
1313
packages (2.1+) that are built against those CUDA versions.
1414

1515
See the full [Documentation](https://pytorch.org/FBGEMM) for more information

‎fbgemm_gpu/docs/src/fbgemm_gpu-development/BuildInstructions.rst

Lines changed: 4 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -117,7 +117,7 @@ Install the full CUDA package through Conda, which includes
117117
.. code:: sh
118118
119119
# See https://anaconda.org/nvidia/cuda for all available versions of CUDA
120-
cuda_version=12.1.0
120+
cuda_version=12.4.1
121121
122122
# Install the full CUDA package
123123
conda install -n ${env_name} -y cuda -c "nvidia/label/cuda-${cuda_version}"
@@ -177,10 +177,10 @@ desired ROCm version:
177177

178178
.. code:: sh
179179
180-
# Run for ROCm 6.1.2
181-
docker run -it --entrypoint "/bin/bash" rocm/rocm-terminal:6.1.2
180+
# Run for ROCm 6.2.0
181+
docker run -it --entrypoint "/bin/bash" rocm/rocm-terminal:6.2.0
182182
183-
While the `full ROCm Docker image <https://hub.docker.com/r/rocm/dev-ubuntu-20.04>`__
183+
While the `full ROCm Docker image <https://hub.docker.com/r/rocm/dev-ubuntu-22.04>`__
184184
comes with all ROCm packages pre-installed, it results in a very large Docker
185185
container, and so for this reason, the minimal image is recommended for building
186186
and running FBGEMM_GPU.

0 commit comments

Comments
 (0)