Enable Deepspeed Ulysses for Wan#353
Merged
Wei-Lin-Intel merged 9 commits intoNov 6, 2025
Merged
Conversation
ikurtchen
reviewed
Oct 31, 2025
| start = sp_seq_len * parallel_state.get_sequence_parallel_rank() | ||
| end = sp_seq_len * (parallel_state.get_sequence_parallel_rank() + 1) | ||
|
|
||
| hidden_states = hidden_states[:, start:end, :] |
There was a problem hiding this comment.
Padded sequence length in SDPA without a mask would trigger the accuracy issue. It is necessary to add mask here if seq_len is padded. Use bool tensor for the mask can save some memory.
mengker33
force-pushed
the
oh_fork_wan_enable_cp
branch
4 times, most recently
from
556f5ef to
ef2b847
Compare
yingjie-han
reviewed
Nov 4, 2025
mengker33
force-pushed
the
oh_fork_wan_enable_cp
branch
from
ef2b847 to
9e27dc2
Compare
|
LGTM, if update some performance data compared with 1/2/4 card would be good. |
yingjie-han
reviewed
Nov 5, 2025
Author
Linwei suggested to use classic SP to reduce memory consumption. I will include SP in this PR too and provide perf data for both deepspeed ulysses and SP soon. |
mengker33
force-pushed
the
oh_fork_wan_enable_cp
branch
from
9e27dc2 to
d1c12df
Compare
This bug has been fixed in upstream deepspeed but not included in habana deepspeed 1.22. See upstream fix commit: deepspeedai/DeepSpeed@ecb4bf3
When seq_len cannot be divided by cp_size, there is padding in seq_len. and attn_mask is helpful to reduce the padding impact on accuracy. Considering padding is usually very small compared to seq_len in the diffuser case, we don't apply the attn_mask by default. The attn_mask can be activated by setting env var CP_USE_MASK.
Co-authored-by: Wei Lin <wei2.lin@intel.com>
mengker33
force-pushed
the
oh_fork_wan_enable_cp
branch
from
d1c12df to
c5aeffc
Compare
mengker33
changed the title
Author
@Wei-Lin-Intel @huijuanzh Please help to review again, thank you:) |
huijuanzh
reviewed
Nov 6, 2025
| is_casual=False, | ||
| scale=None, | ||
| softmax_mode=softmax_mode, | ||
| query.transpose(1, 2).contiguous(), |
There was a problem hiding this comment.
why CogVideoXAttnProcessorGaudi need to change? Since you put qkv transpose in ModuleFusedSDPA, need to check other function which not use sp or ulysses.
|
LGTM |
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 participants
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.
No description provided.