Multiple updates to Tensorflow easyblock#1453
Conversation
| # enable mkl-dnn by default, but only if cuDNN is not listed as dependency | ||
| if self.cfg['with_mkl_dnn'] is None and get_software_root('cuDNN') is None: | ||
| self.log.info("Enabling use of mkl-dnn since cuDNN is not listed as dependency") | ||
| self.cfg['with_mkl_dnn'] = True |
There was a problem hiding this comment.
@akesandgren Can you clarify this? Can both be used together? There must have been a reason why did this?
There was a problem hiding this comment.
We did this because i originally thought there would be a problem, but I later tried and they can be compiled in at the same time.
So it's basically removing a bit of code that shouldn't have been there in the first place.
|
Verified to work with old TensorFlow-1.5.0-foss-2017b-Python-3.6.3.eb |
This might really be a bug in protobuf-python, protocolbuffers/protobuf#1296
|
Tested with TensorFlow-1.5.0-foss-2017b-Python-3.6.3.eb and easybuilders/easybuild-easyconfigs#6677 (TensorFlow-1.10.0-fosscuda-2018b-Python-2.7.15.eb) |
| 'NCCL_INSTALL_PATH': nccl_root, | ||
| }) | ||
| else: | ||
| nccl_version = '1.3' # Use simple downloadable version |
There was a problem hiding this comment.
@akesandgren Where did you get this 1.3 from? Isn't this something that will vary across TF versions?
There was a problem hiding this comment.
1.3 is the fallback version if there is no external NCCL v2.x installed (requires manual download).
1.3 can be auto-downloaded and is the version they mention in all TF versions from 1.4, i think, as the default if no external NCCL is available.
But, from TF 1.10 they have changed the Q&A stuff so one now must specify it.
|
@akesandgren I've tested this with a whole bunch of existing
Didn't see any problems, so unless you're planning further changes, this looks good to go? |
|
No more changes in sight, gfood to go from my side. |
Solves issue #1445