Add embedding backward workaround #1756
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
pmarkovicTT
requested review from
sdjordjevicTT,
svuckovicTT,
mtopalovicTT,
nobradovictt,
jserbedzijaTT,
jnie-TT,
azecevicTT,
nsmithtt and
mrakitaTT
as code owners
There was a problem hiding this comment.
Great change, thanks @pmarkovicTT! Can you change a silicon test to use f32 instead of bf16, so we can test the workaround on silicon as well?
pmarkovicTT
force-pushed
the
pmarkovic/embedding_backward_workaround
branch
from
9a639d9 to
ddd511b
Compare
|
Changed the test and it passed tests ( |
This was referenced Jan 27, 2025
Merged
pmarkovicTT
added a commit
to tenstorrent/tt-tvm
that referenced
this pull request
Jan 27, 2025
We don't need explicit embedding dataformat cast in tvm (from float32 to bf16) as dataformat workaround for this case is implemented in mlir. PRs for reference: - [TVM cast workaround](#55) - [Embedding Op workaround](tenstorrent/tt-mlir#1583) - [EmbeddingBackward Op workaround](tenstorrent/tt-mlir#1756) Related to this issue tenstorrent/tt-forge-fe#1112
pmarkovicTT
added a commit
to tenstorrent/tt-forge-fe
that referenced
this pull request
Feb 7, 2025
#1111) ### Ticket Close #1112 ### Problem description We don't need explicit embedding dataformat cast in tvm (from float32 to bf16) as dataformat workaround for this case is implemented in mlir. PRs for reference: - [TVM change](tenstorrent/tt-tvm#59) - [Embedding Op workaround](tenstorrent/tt-mlir#1583) - [EmbeddingBackward Op workaround](tenstorrent/tt-mlir#1756) ### What's changed Removed explicit cast to bfloat16 if dataformat for embedding weights is float32. Updated llama backward test to reflect new forge api for training (setting training argument). ### Checklist - [x] Remove explicit cast in third_party/tvm/python/tvm/relay/frontend/pytorch.py - [x] Update test_llama_backward.py --------- Co-authored-by: Vladimir Milosevic <157983820+vmilosevic@users.noreply.github.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR introduces workaround for embedding backward op. Due to TTNN constraint to accept only BF16 and BF8 data types, this workaround casts F32 to BF16 and back. This is follow up to #1583
Closes #1503