Skip to content

[Perf] [NPU]Fix rotary embedding dimension mismatch & add auto-adaptation for full-dim cos/sin in mindiesd RoP #2369

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
Li-8916 wants to merge 8 commits into
base: main
Choose a base branch
from
Open

[Perf] [NPU]Fix rotary embedding dimension mismatch & add auto-adaptation for full-dim cos/sin in mindiesd RoP #2369

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
58d78f7
cos/sin shape already is [s,d], just use rotary_position_embedding di…
Li-8916 Mar 28, 2026
eb375af
Update rope.py
Li-8916 Mar 31, 2026
492294a
Merge branch 'vllm-project:main' into main
Li-8916 Mar 31, 2026
a25f131
optimize rope on NPU
Li-8916 Mar 31, 2026
38a7941
update rotary
Li-8916 Mar 31, 2026
b91cbf2
Reorganize import statements in wan2_2_transformer.py
Li-8916 Mar 31, 2026
e09cdb0
Remove unnecessary blank line in wan2_2_transformer.py
Li-8916 Mar 31, 2026
bc28e90
Remove unnecessary line in rope.py
Li-8916 Mar 31, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
5 changes: 3 additions & 2 deletions vllm_omni/diffusion/layers/rope.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -49,7 +49,9 @@ def apply_rotary_emb_mindiesd(
# (B, S, D/2) -> (S, D/2)
cos = cos[0]
sin = sin[0]

if cos.shape[-1] == x.shape[-1]:
half_head_dim = False
# already expanded to (S, D), just use directly
if interleaved:
# if last dim of sin and cos is D/2, expand to (S, D) to adapt to mindiesd operators
if half_head_dim:
Expand Down Expand Up @@ -97,7 +99,6 @@ def forward_cuda(
# (B, S, D/2) -> (S, D/2)
cos = cos[0]
sin = sin[0]

return apply_rotary_emb(
x,
cos,
Expand Down
12 changes: 10 additions & 2 deletions vllm_omni/diffusion/models/wan2_2/wan2_2_transformer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3,6 +3,7 @@

import math
from collections.abc import Iterable
from importlib.util import find_spec
from typing import Any

import torch
Expand All @@ -29,6 +30,8 @@
SequenceParallelOutput,
)
from vllm_omni.diffusion.forward_context import get_forward_context
from vllm_omni.diffusion.layers.rope import RotaryEmbedding
from vllm_omni.platforms import current_omni_platform

logger = init_logger(__name__)

Expand Down Expand Up @@ -392,6 +395,7 @@ def __init__(
softmax_scale=1.0 / (head_dim**0.5),
causal=False,
)
self.rope = RotaryEmbedding(is_neox_style=False)

def forward(
self,
Expand All @@ -418,8 +422,12 @@ def forward(
# Apply rotary embeddings
if rotary_emb is not None:
freqs_cos, freqs_sin = rotary_emb
query = apply_rotary_emb_wan(query, freqs_cos, freqs_sin)
key = apply_rotary_emb_wan(key, freqs_cos, freqs_sin)
if find_spec("mindiesd") is not None and current_omni_platform.is_npu():
Copy link
Copy Markdown
Collaborator

@hsliuustc0106 hsliuustc0106 Mar 31, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

any better ways to add supports from mindiesd? cc @gcanlin

Copy link
Copy Markdown
Collaborator

@hsliuustc0106 hsliuustc0106 Apr 1, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

???? @Li-8916 @gcanlin

Copy link
Copy Markdown
Collaborator

@gcanlin gcanlin Apr 1, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I prefer not to introduce it here. Will investigate how mindie-sd implement it.

Copy link
Copy Markdown
Collaborator

@gcanlin gcanlin Apr 1, 2026 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This PR #2393 probably has the same effect and provide the general optimization for all platforms.

query = self.rope(query, freqs_cos, freqs_sin)
key = self.rope(key, freqs_cos, freqs_sin)
else:
query = apply_rotary_emb_wan(query, freqs_cos, freqs_sin)
key = apply_rotary_emb_wan(key, freqs_cos, freqs_sin)

# Create attention metadata if mask is provided
attn_metadata = None
Expand Down
Loading