If you don't see 7.5 make sure to upgrade your brew formulas: brew update ModuleNotFoundError: No module named 'apex'Ĭlearly, deepspeed itself was installed successfully but it still says failed because the basic install test failed due to unavailability of apex on my Mac.Īlso, I used the -d and -k flags when running install.sh, so why was it trying to uninstall torch 1.4 and install torch 1.2 in the first place? I thought I shouldn't have needed to touch any file in the /requirements directory if I used -d and -k, but in fact I had to in order to get this to install fine.Make sure that the installed cuda version is 7.5 you can check the version with brew cask info cuda runtime.engine import DeepSpeedEngineįile "/path/to/my/venv/lib/python3.6/site-packages/deepspeed/runtime/engine.py", line 10, in Removed build tracker: '/private/var/folders/yf/dvfk4pps6ndd5xx1q5fw5mm9ssqdt8/T/pip-req-tracker-jp7hzynz'įile "basic_install_test.py", line 19, in įile "basic_install_test.py", line 16, in įile "/path/to/my/venv/lib/python3.6/site-packages/deepspeed/_init_.py", line 9, in įrom. Successfully installed deepspeed-0.3.0+4fef478 tensorboardX-1.9 torchvision-0.5.0 Given no hashes to check 0 links for project 'clang-format': discarding no candidatesĮRROR: Could not find a version that satisfies the requirement clang-format (from deepspeed)ĮRROR: No matching distribution found for clang-format
Skipping link: none of the wheel's tags match: py2-none-manylinux1_x86_64, p圓-none-manylinux1_x86_64: (from ) Skipping link: none of the wheel's tags match: py2-none-manylinux1_x86_64: (from ) Request header has "max_age" as 0, cache bypassed Reply to this email directly, view it on GitHubġ location(s) to search for versions of clang-format:įetching project page and analyzing links: You are receiving this because you were mentioned. For PyTorch 1.2 that should beĬudatoolkit=10.0, and for PyTorch >=1.3 that should be cudatoolkit=10.1 Whatever CUDA toolkit that environment has. install.sh should build you the correct Apex commit with Ensure the build directories are clean underģ. Uninstall Apex (you might need sudo pip -H uninstall apex if it'sĢ. I believe the error that you're seeing is due to it linking with a versionĬompiled against CUDA toolkit 10.2, which is newer than what you areġ.
#Should i install cuda on mac code#
Your training will crash on certain code paths. Just installing DeepSpeed will not test the Apex interfaces, only theĪbility to import it. On Thu, Mar 26, 2020, 12:34 PM Shaden Smith wrote:
Assuming those are installed and working, you could try to install the remaining prerequisites and see if there are any issues? Do you have a CUDA-enabled PyTorch already installed on that system? PyTorch's website suggests that CUDA is only available on Macs if installed from source. I don't have access to a Mac, so investigation is hard on my end. install.sh to make sure you have the right Apex version. I'd recommend uninstalling the latest Apex you have, clearing deepspeed/third_party/apex/, and installing DeepSpeed with a fresh. Were you working from a source directory that you had previously used to build+install? Apex could have reused some build files from that. I agree that error suggests that Apex was compiled with 10.2. Catching up with Apex HEAD is a work in progress. Apex does not use versioning, so we use a submodule pointing to specific commits to enforce that. Hi the latest commit of Apex is not API-compatible with the one we are pointing to.
Is there a particular reason why exactly that version of apex is being pointed to in install.sh? Clearly it seems to be giving some issues above. The above import error seems to have gone away now! install.sh -k -d to ensure that I don't install the apex pointed to by DeepSpeed. This libcudart.so.10.2 thing seems to suggest that the version of apex pointed to by the install.sh is only compatible with 10.2 - do you have any ideas on what's going on?ĮDIT: I just fresh-installed apex separately (outside of install.sh) directly by cloning the apex repo and running pip install -v -no-cache-dir -global-option="-cpp_ext" -global-option="-cuda_ext". ImportError: libcudart.so.10.2: cannot open shared object file: No such file or directory Return _bootstrap._gcd_import(name, package, level)įile "", line 955, in _find_and_load_unlockedįile "", line 219, in _call_with_frames_removed File "/home/ec2-user/anaconda3/envs/pytorch_p36/lib/python3.6/site-packages/apex/optimizers/fused_adam.py", line 40, in _init_įused_adam_cuda = importlib.import_module("fused_adam_cuda")įile "/home/ec2-user/anaconda3/envs/pytorch_p36/lib/python3.6/importlib/_init_.py", line 126, in import_module
0 kommentar(er)
