XPU and MPS take 3 #276
XPU and MPS take 3 #276
Conversation
jwallwork23
added
enhancement
jwallwork23
force-pushed
the
127_xpu-take3
branch
from
4aa9a4a to
55b929f
Compare
|
Offline testing for CUDA version of MultiGPU example with 2 devices passed on Ampere. In the queue for XPU testing on PVC. |
There was a problem hiding this comment.
Thanks @jwallwork23 I have only done a quick pass of this so far and will need to schedule time for a closer look, but I suspect you know my first comment - can you update the docs at utils/README etc. to reflect the new args and usage of pt2ts?
I would also like to see it documented somewhere how the device enums are managed from CMake. Perhaps under the developer docs. As, whilst a very nifty solution, it's slightly abstract if you are not the one who came up with it 😉
|
Also looks like you may want a rebase after #268 |
|
Re-tested on Dawn with latest version of branch - all good |
There was a problem hiding this comment.
Thanks for driving all of this forward @jwallwork23 the devices make a fine addition to the collection.
I think we are on the edge of glory, but have a couple of suggestions/clarifications before approval.
More general ruminations on reviewing this:
- I wonder if there is a better way to keep the pt2ts files in sync... (not to be resolved in this PR)
conda/README.md
Outdated
There was a problem hiding this comment.
I'll need to update the mac conda to include this in #284
Co-authored-by: Jack Atkinson <[email protected]>
Yeah agreed. Perhaps the script could be set up such that it doesn't need to be modified, although that might be a tall order. |
|
Latest version now passes tests on Ampere. Awaiting XPU job on PVC. |
Passed on PVC, too! |
|
Tests run as |
There was a problem hiding this comment.
I think I am now happy with everything here @jwallwork23 .
Ran all OK on Mac (conda, with MPS to run MultiGPU example) and you checked PVC, CUDA, and CPU.
Fantastic stuff.
You may merge when ready (I would lean towards a squash based on the commit history. The PR touches a lot of files, but generally fairly specific lines in each).
|
Amazing! Thanks for the thorough review @jatkinson1000 - will merge now. |
* Add MacOS GPU device option * Add XPU device option * Update C++ XPU interface to handle multiple devices indices. * Update ftorch.F90 for XPU support * Make device enums consistent with PyTorch * Accept command line arguments in MultiGPU example * Introduce GPU_DEVICE preprocessor option * Update pt2ts scripts; use argparse over sys.argv * Update GPU docs * Update READMEs * Add explanation of GPU device codes in dev docs --------- Co-authored-by: ElliottKasoar <[email protected]> Co-authored-by: Jack Atkinson <[email protected]> Co-authored-by: Matt Archer <[email protected]> Co-authored-by: Jack Atkinson <[email protected]>
Closes #127.
Builds upon #125 and #209.
(Contains changes from #268 so that will need to be merged first.)
This PR adds support XPU and MPS. Unfortunately, it ended up requiring an overhaul of the
pt2tsscripts, too.Notable changes:
ENABLE_CUDAto the more general and extensibleGPU_DEVICE=<NONE/CUDA/XPU/MPS>.argparsefor reading command line arguments into Python scripts, rather thansys.argv.Checklist