[CPU] Skip CPU support unimplemented error#3633
[CPU] Skip CPU support unimplemented error#3633tjruwase merged 21 commits intodeepspeedai:masterfrom
Conversation
Yejing-Lai
requested review from
jeffra,
loadams,
mrwyattii and
tjruwase
as code owners
Yejing-Lai
force-pushed
the
lyj/cpu_infer_workflow
branch
2 times, most recently
from
5838577 to
7e534ab
Compare
|
I found that if no tests were collected, pytest will return exit code 5 (the correct exit code should be 0). I updated the cmd of running ut to ensure that pytest can exit correctly. |
Yejing-Lai
force-pushed
the
lyj/cpu_infer_workflow
branch
from
7e534ab to
7ea1c8b
Compare
accelerator/cuda_accelerator.py
Outdated
| return False | ||
|
|
||
| def supported_dtypes(self): | ||
| return [torch.float, torch.half] |
There was a problem hiding this comment.
Cuda support torch.bfloat16 as well, right?
There was a problem hiding this comment.
Is there any reason we cannot use the data types defined here
https://pytorch.org/docs/stable/tensors.html#data-types
There was a problem hiding this comment.
Hi tjruwase, I am sorry for the late reply. Cuda support torch.bfloat16 as well. I will update my PR later, Thanks~
|
@Yejing-Lai, is this PR ready for review? Thanks! |
It's ready. The ut can work correctly on my local env. Please review. Thanks~ |
|
@Yejing-Lai, PR looks good. Please resolve conflict to enable merge. Thanks. |
Hi @tjruwase , can I run the CI again? I think the reason for the failure of both checks is a network issue. |
github-merge-queue
bot
removed this pull request from the merge queue due to failed status checks
Yejing-Lai
requested review from
RezaYazdaniAminabadi,
arashb,
awan-10,
cmikeh2 and
samyam
as code owners
Revert "remove skip FusedAdamBuilder; add suported_dtypes" Revert "remove unused parameters" Revert "enable zero stage 1 and 2 for synchronized accelerator (a.k.a. CPU)" Revert "use cpu adam to implement fused adam" Revert "fused adam can build"
|
Hi @tjruwase, I am sorry to revert the commit about fusedadam. Our fusedadam needs to be further modified, and we will submit another PR about fusedadam after this PR merge. Thanks for your review/merge~ |
|
@Yejing-Lai, no worries. Thanks for the update. Please ping me when you are ready to move forward with this PR. Thanks! |
Hi @tjruwase. this PR is ready for merge. Please review. Thanks~ |
4 participants
This PR aims to add CPU inference UT support. We skip CPU support unimplemented errors and update cpu inference workflow.
Skip logic: