{2023.06}[system] cuDNN/8.9.2.26-CUDA-12.1.1 #581
Conversation
|
Instance
|
|
Instance
|
|
bot: build inst:aws repo:eessi.io-2023.06-software arch:x86_64/amd/zen2 |
Updates by the bot instance
|
Updates by the bot instance
|
|
New job on instance
|
eb_hooks.py
Outdated
| # iterate over all files in the CUDA installation directory | ||
| for dir_path, _, files in os.walk(self.installdir): | ||
| for filename in files: | ||
| full_path = os.path.join(dir_path, filename) | ||
| # we only really care about real files, i.e. not symlinks | ||
| if not os.path.islink(full_path): | ||
| # check if the current file is part of the allowlist | ||
| basename = filename.split('.')[0] | ||
| if '.' in filename: | ||
| extension = '.' + filename.split('.')[1] | ||
| if basename in allowlist: | ||
| self.log.debug("%s is found in allowlist, so keeping it: %s", basename, full_path) | ||
| elif '.' in filename and extension in allowlist: | ||
| self.log.debug("%s is found in allowlist, so keeping it: %s", extension, full_path) | ||
| else: | ||
| self.log.debug("%s is not found in allowlist, so replacing it with symlink: %s", | ||
| filename, full_path) | ||
| # if it is not in the allowlist, delete the file and create a symlink to host_injections | ||
| host_inj_path = full_path.replace('versions', 'host_injections') | ||
| # make sure source and target of symlink are not the same | ||
| if full_path == host_inj_path: | ||
| raise EasyBuildError("Source (%s) and target (%s) are the same location, are you sure you " | ||
| "are using this hook for a NESSI installation?", | ||
| full_path, host_inj_path) | ||
| remove_file(full_path) | ||
| symlink(host_inj_path, full_path) |
There was a problem hiding this comment.
Isn't this identical to what is done for CUDA? We should probably just create a function that takes the installdir and allowlist as arguments and does this.
There was a problem hiding this comment.
Actually there are subtle differences. For CUDA, the EULA/README lists files you can distribute. For cuDNN the LICENSE lists what type of files you can distribute. These differences require small modifications. For example, in the hook for CUDA we have:
basename = filename.split('.')[0]
if basename in allowlist:
self.log.debug("%s is found in allowlist, so keeping it: %s", basename, full_path)
else:
self.log.debug("%s is not found in allowlist, so replacing it with symlink: %s",
basename, full_path)For cuDNN, we have
basename = filename.split('.')[0]
if '.' in filename:
extension = '.' + filename.split('.')[1]
if basename in allowlist:
self.log.debug("%s is found in allowlist, so keeping it: %s", basename, full_path)
elif '.' in filename and extension in allowlist:
self.log.debug("%s is found in allowlist, so keeping it: %s", extension, full_path)
else:
self.log.debug("%s is not found in allowlist, so replacing it with symlink: %s",
filename, full_path)Anyhow, the differences are relatively small, so a function would require a parameter that allows it to distinguish between CUDA and cuDNN (and in the future maybe other packages such as cuTENSOR).
There was a problem hiding this comment.
For the extension part, perhaps we should split on all . and look for the last non-numeric entry (which should be the extension)? I can imagine there could be files like libcuda.so.520.12.1
There was a problem hiding this comment.
Scratch that, you already have a good solution, you are taking the second entry which is virtually guaranteed to the the extension
There was a problem hiding this comment.
Added a function that implements the suggestion in 74a9a55
eb_hooks.py
Outdated
| ec_dict['builddependencies'].append(dep) | ||
| value = '\n'.join([value, 'setenv("EESSICUDNNVERSION","%s")' % cudnn_version]) | ||
| if key in ec_dict: | ||
| if not value in ec_dict[key]: |
There was a problem hiding this comment.
This check is probably no longer good enough, we're looking for the exact string, but that is not likely to exist (even though the add_property("arch","gpu") most likely does exist since the applications also should have a CUDA dep). What we really need to do is
- Grab what is there already
- Split it on
\n - Add any missing elements
- Put it back together again and replace it
Either this, or only the modify/add the modluafooter once in the entire function
There was a problem hiding this comment.
You mean the if not value in ec_dict[key] is not good enough?
There was a problem hiding this comment.
No, because the value is a composite string of property and the setenv, and the property will already (very likely) exist from the CUDA part of this hook
There was a problem hiding this comment.
The module file for cuDNN contains the following
-- Built with EasyBuild version 4.9.1
add_property("arch","gpu")
setenv("EESSICUDAVERSION","12.1.1")For something that builds on top of cuDNN, we would the above and something like
setenv("EESSICUDNNVERSION","8.9.2.26")There was a problem hiding this comment.
As currently implemented, for something that builds on top of cuDNN I believe you will have
-- Built with EasyBuild version 4.9.1
add_property("arch","gpu")
setenv("EESSICUDAVERSION","12.1.1")
add_property("arch","gpu")
setenv("EESSICUDNNVERSION","8.9.2.26")
as it will see if the entire string add_property("arch","gpu")\nsetenv("EESSICUDNNVERSION","8.9.2.26") is in the footer
There was a problem hiding this comment.
Ack. Working on something to implement the desired footer (and avoiding duplication of code).
There was a problem hiding this comment.
Updated the function. It still produces the same footer for cuDNN. I guess a real test would be a build that uses cuDNN. @ocaisa can you check if the function looks better now?
|
Retry after fixing args to bot: build inst:aws repo:eessi.io-2023.06-software arch:x86_64/amd/zen2 |
Updates by the bot instance
|
Updates by the bot instance
|
|
New job on instance
|
|
@trz42 The installation looks suspiciously large at 700MB, are you sure your hook is cleaning out the files it should? |
Full package is 1.4 GB. |
|
Rebuild after changing hook function that handles dependencies and creates modluafooter entries... bot: build inst:aws repo:eessi.io-2023.06-software arch:x86_64/amd/zen2 |
Updates by the bot instance
|
Updates by the bot instance
|
|
New job on instance
|
|
One more time... bot: build inst:aws repo:eessi.io-2023.06-software arch:x86_64/amd/zen2 |
Updates by the bot instance
|
Updates by the bot instance
|
|
New job on instance
|
|
@trz42 I will take your updated |
|
I also get the feeling that if we are going to move to easystack files (a good idea) then we should probably ship the ones we expect people to use |
Just updated the script with some improvements/fixes after my own testing. |
…-layer into 2023.06-software.eessi.io-cuDNN-8.9.2.26-system
- `EESSI-install-software.sh`
- use `scripts/gpu_support/nvidia/install_cuda_and_libraries.sh` with
`scripts/gpu_support/nvidia/eessi-2023.06-cuda-and-libraries.yml`
- `create_lmodsitepackage.py`
- consolidate `eessi_{cuda,cudnn}_enabled_load_hook` functions in a single one
(`eessi_cuda_and_libraries_enabled_load_hook`)
- the remaining hook is prepared to easily add new modules, e.g., cuTENSOR
- `eb_hooks.py`
- put code that iterates over all files replacing non-distributable ones with
symlinks into `host_injections` with a common function
(`replace_non_distributable_files_with_symlinks`)
- `install_scripts.sh`
- add files to copy to CVMFS (see `nvidia_files`)
- `scripts/gpu_support/nvidia/install_cuda_and_libraries.sh`
- improved creation of tmp directory
|
Run another build after several changes... bot: build inst:aws repo:eessi.io-2023.06-software arch:x86_64/amd/zen2 |
Updates by the bot instance
|
Updates by the bot instance
|
|
New job on instance
|
Attempt to add cuDNN which is a dependency of other packages such as TensorFlow and PyTorch.
Major additions/changes:
scripts/gpu_support/nvidia/install_cuda_and_libraries.shwithscripts/gpu_support/nvidia/eessi-2023.06-cuda-and-libraries.ymlCUDAandcuDNNpackages under.../host_injectionsEESSI-install-software.shscripts/gpu_support/nvidia/install_cuda_and_libraries.shwithscripts/gpu_support/nvidia/eessi-2023.06-cuda-and-libraries.ymlto installCUDA,cuDNNunder.../host_injectionseb_hooks.pysymlinks into
host_injectionswith a common function(
replace_non_distributable_files_with_symlinks)post_sanitycheck_hookwhich replaces files with symlinks into corresponding paths under.../host_injectionsfor all files that cannot be redistributedcuDNNto a build dependency (seeinject_gpu_property)create_lmodsitepackage.pyeessi_{cuda,cudnn}_enabled_load_hookfunctions in a single one(
eessi_cuda_and_libraries_enabled_load_hook)install_scripts.shnvidia_files)