fix the crash in Wan-AI/Wan2.2-TI2V-5B-Diffusers if CP is enabled #12562
+19
−0
Conversation
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
Signed-off-by: Wang, Yi <yi.a.wang@intel.com>
|
torchrun --nproc-per-node 2 test.py crash stack: |
|
@yiyixuxu @sayakpaul please help review |
|
Could you also supplement an output with the fix? |
sayakpaul
reviewed
Oct 30, 2025
| ) | ||
| if ts_seq_len is not None: | ||
| # Check if running under context parallel and split along seq_len dimension | ||
| if hasattr(self, '_parallel_config') and self._parallel_config is not None: |
There was a problem hiding this comment.
Could you elaborate why this is needed?
There was a problem hiding this comment.
when cp is enabled, seq_len is split, timestep_shape is [batch_size, seq_len, 6, inner_dim], so should be split in dim_1 as well since hidden state is split in seq_len dim as well. or else shape miss match will occur
There was a problem hiding this comment.
do you think we can just change _cp_plan? https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/transformers/transformer_wan.py#L546
There was a problem hiding this comment.
you mean split timestep in forward? adding
"": {
"timestep": ContextParallelInput(split_dim=1, split_output=False)
}, to _cp_plan will make 5B work, but 14B fail since 5B timestep dims is 2. 14 timestep dims is 1.
There was a problem hiding this comment.
Hmm, this is an interesting situation. To tackle these, I think we might have to revisit the ContextParallelInput and ContextParallelOutput definitions a bit more.
If we had a way to tell the partitioner that the input might have "dynamic" dimensions depending on the model configs (like in this case), and what it should do if that's the case, it might be more flexible as a solution.
@DN6 curious to know what you think.
There was a problem hiding this comment.
another solution that come to my mind is a fix in pipeline wan. clean and more generic change like
from .pipeline_output import WanPipelineOutput
-
+from ...models._modeling_parallel import ContextParallelInput, ContextParallelOutput
if is_torch_xla_available():
import torch_xla.core.xla_model as xm
@@ -150,6 +150,11 @@ class WanPipeline(DiffusionPipeline, WanLoraLoaderMixin):
)
self.register_to_config(boundary_ratio=boundary_ratio)
self.register_to_config(expand_timesteps=expand_timesteps)
+ if expand_timesteps:
+ transformer._cp_plan.update({"": {
+ "timestep": ContextParallelInput(split_dim=1, expected_dims=2, split_output=False)
+ }})
WDYT ?
There was a problem hiding this comment.
change in src/diffusers/pipelines/wan/pipeline_wan.py
There was a problem hiding this comment.
That is probably a bit more intrusive. I think what you suggested earlier would work as @DN6 suggested.
There was a problem hiding this comment.
Actually, I like this the best 👍🏽
There was a problem hiding this comment.
The only downside to setting it via the pipeline is that it would still error out if you were trying to run inference outside the WanPipeline. So perhaps this is the best way to handle it for now.
def _prepare_cp_input(self, x: torch.Tensor, cp_input: ContextParallelInput) -> torch.Tensor:
if cp_input.expected_dims is not None and x.dim() != cp_input.expected_dims:
logger.warning(
f"Expected input tensor to have {cp_input.expected_dims} dimensions, but got {x.dim()} dimensions."
)
return x
else:
return EquipartitionSharder.shard(x, cp_input.split_dim, self.parallel_config._flattened_mesh)
sayakpaul
reviewed
Oct 30, 2025
|
seems I can not attach the video here, blocked may be .... 
|
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 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.
fix the crash when testing CP for wan2.2-TI2V-5B
test script: