The quickest way to get started with DeepSpeed is via pip, this will install the latest release of DeepSpeed which is not tied to specific PyTorch or CUDA versions. DeepSpeed includes several C++/CUDA extensions that we commonly refer to as our ‘ops’. By default, all of these extensions/ops will be built just-in-time (JIT) using torch’s JIT C++ extension loader that relies on ninja to build and dynamically link them at runtime.
pip install deepspeed
After installation, you can validate your installation and see which ops your machine
is compatible with via the DeepSpeed environment report with
python -m deepspeed.env_report. We’ve found this report useful when debugging
DeepSpeed install or compatibility issues.
Pre-install DeepSpeed Ops
Note: PyTorch must be installed before pre-compiling any DeepSpeed c++/cuda ops. However, this is not required if using the default mode of JIT compilation of ops.
Sometimes we have found it useful to pre-install either some or all DeepSpeed C++/CUDA ops instead of using the JIT compiled path. In order to support pre-installation we introduce build environment flags to turn on/off building specific ops.
You can indicate to our installer (either
pip install) that you
want to attempt to install all of our ops by setting the
environment variable to
1, for example:
DS_BUILD_OPS=1 pip install deepspeed
DeepSpeed will only install any ops that are compatible with your machine.
For more details on which ops are compatible with your system please try our
ds_report tool described above.
If you want to install only a specific op (e.g.,
FusedLamb), you can toggle
DS_BUILD environment variables at installation time. For example, to
install DeepSpeed with only the
FusedLamb op use:
DS_BUILD_FUSED_LAMB=1 pip install deepspeed
DS_BUILD options include:
DS_BUILD_OPStoggles all ops
DS_BUILD_CPU_ADAMbuilds the CPUAdam op
DS_BUILD_FUSED_ADAMbuilds the FusedAdam op (from apex)
DS_BUILD_FUSED_LAMBbuilds the FusedLamb op
DS_BUILD_SPARSE_ATTNbuilds the sparse attention op
DS_BUILD_TRANSFORMERbuilds the transformer op
DS_BUILD_TRANSFORMER_INFERENCEbuilds the transformer-inference op
DS_BUILD_STOCHASTIC_TRANSFORMERbuilds the stochastic transformer op
DS_BUILD_UTILSbuilds various optimized utilities
DS_BUILD_AIObuilds asynchronous (NVMe) I/O op
To speed up the build-all process, you can parallelize the compilation process with:
DS_BUILD_OPS=1 pip install deepspeed --global-option="build_ext" --global-option="-j8"
This should complete the full build 2-3 times faster. You can adjust
-j to specify how many cpu-cores are to be used during the build. In the example it is set to 8 cores.
You can also build a binary wheel and install it on multiple machines that have the same type of GPUs and the same software environment (CUDA toolkit, pytorch, python, etc.)
DS_BUILD_OPS=1 python setup.py build_ext -j8 bdist_wheel
This will create a pypi binary wheel under
dist/deepspeed-0.3.13+8cd046f-cp38-cp38-linux_x86_64.whl and then you can install it directly on multiple machines, in our example:
pip install dist/deepspeed-0.3.13+8cd046f-cp38-cp38-linux_x86_64.whl
Install DeepSpeed from source
After cloning the DeepSpeed repo from GitHub, you can install DeepSpeed in JIT mode via pip (see below). This installation should complete quickly since it is not compiling any C++/CUDA source files.
pip install .
For installs spanning multiple nodes we find it useful to install DeepSpeed
script in the repo. This will build a python wheel locally and copy it to all
the nodes listed in your hostfile (either given via
--hostfile, or defaults to
When the code using DeepSpeed is used for the first time it’ll automatically build only the CUDA
extensions, required for the run, and by default it’ll place them under
~/.cache/torch_extensions/. The next time the same program is executed these now precompiled
extensions will be loaded form that directory.
If you use multiple virtual environments this could be a problem, since by default there is only one
torch_extensions directory, but different virtual environments may use different setups (e.g., different
python or cuda versions) and then the loading of a CUDA extension built by another environment will
fail. Therefore, if you need to you can override the default location with the help of the
TORCH_EXTENSIONS_DIR environment variable. So in each virtual environment you can point it to a
unique directory and DeepSpeed will use it to save and load CUDA extensions.
You can also change it just for a specific run with:
TORCH_EXTENSIONS_DIR=./torch-extensions deepspeed ...
Building for the correct architectures
If you’re getting the following error:
RuntimeError: CUDA error: no kernel image is available for execution on the device
when running deepspeed, that means that the cuda extensions weren’t built for the card you’re trying to use it for.
When building from source deepspeed will try to support a wide range of architectures, but under jit-mode it’ll only support the architectures visible at the time of building.
You can build specifically for a desired range of architectures by setting a
TORCH_CUDA_ARCH_LIST env variable:
TORCH_CUDA_ARCH_LIST="6.1;7.5;8.6" pip install ...
It will also make the build faster when you only build for a few architectures.
This is also recommended to ensure your exact architecture is used. Due to a variety of technical reasons, a distributed pytorch binary isn’t built to fully support all architectures, skipping binary compatible ones, at a potential cost of underutilizing your full card’s compute capabilities. To see which architectures get included during the deepspeed build from source - save the log and grep for
The full list of nvidia GPUs and their compute capabilities can be found here.
Feature specific dependencies
Some DeepSpeed features require specific dependencies outside the general dependencies of DeepSpeed.
Python package dependencies per feature/op please see our requirements directory.
We attempt to keep the system level dependencies to a minimum, however some features do require special system-level packages. Please see our
ds_reporttool output to see if you are missing any system-level packages for a given feature.
Pre-compiled DeepSpeed builds from PyPI