#0: untilize tensor on device #12001
Conversation
sraizada-tt
requested review from
eyonland,
arakhmati,
patrickroberts,
yan-zaretskiy,
cfjchu,
xanderchin,
TT-BrianLiu,
ayerofieiev-tt and
dmakoviichuk-tt
as code owners
|
It would be good to run at least 1 T3000 and 1 TG pipeline. |
TG Llama fails, issue open: #11509 @tarafdarTT |
sraizada-tt
had a problem deploying
to
production
GitHub Actions
Failure
Could you run.either https://github.com/tenstorrent/tt-metal/actions/workflows/tg-nightly-tests.yaml or https://github.com/tenstorrent/tt-metal/actions/workflows/tg-unit-tests.yaml ? |
| @@ -309,12 +309,12 @@ def to_torch( | |||
| tensor([[-0.3008, -0.8438, 0.3242], | |||
| [ 0.9023, -0.5820, 0.5312]], dtype=torch.bfloat16) | |||
| """ | |||
| if tensor.layout != ttnn.ROW_MAJOR_LAYOUT: | |||
There was a problem hiding this comment.
we convert to row-major on-device before moving it from device because it activates untilize on device by default?
There was a problem hiding this comment.
are there differences in the output between untilize on host vs. device or is it negligible?
There was a problem hiding this comment.
if there are changes, we might want to consider adding a toggleable option for switching between host vs. device untilizer
There was a problem hiding this comment.
there is a perf hit for utilizing on host vs device. Adding this as a toggleable option is a good idea
|
Hey folks, just wanted to check if this is still happening? |
Ticket
Link to Github Issue
Problem description
Unitize tensor on device in the ttnn.to_torch function
What's changed
Unitize tensor on device in the ttnn.to_torch function
Checklist