Skip to content

feat: enable attention dispatch for huanyuan video #12591

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
wants to merge 5 commits into
base: main
Choose a base branch
from
+39 −20
Open

feat: enable attention dispatch for huanyuan video #12591

Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
af9814b
feat: enable attention dispatch for huanyuan video
DefTruth Nov 4, 2025
cf296d0
feat: enable attention dispatch for huanyuan video
DefTruth Nov 4, 2025
73f44c0
Merge branch 'main' into attn-dispatch-hyvideo
DefTruth Nov 4, 2025
02753af
Merge branch 'main' into attn-dispatch-hyvideo
DefTruth Nov 5, 2025
653d87b
Merge branch 'main' into attn-dispatch-hyvideo
DefTruth Nov 6, 2025
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
59 changes: 39 additions & 20 deletions src/diffusers/models/transformers/transformer_hunyuan_video.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -24,6 +24,7 @@
from ...loaders import PeftAdapterMixin
from ...utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers
from ..attention import FeedForward
from ..attention_dispatch import dispatch_attention_fn
from ..attention_processor import Attention, AttentionProcessor
from ..cache_utils import CacheMixin
from ..embeddings import (
Expand All @@ -42,6 +43,9 @@


class HunyuanVideoAttnProcessor2_0:
_attention_backend = None
_parallel_config = None

def __init__(self):
if not hasattr(F, "scaled_dot_product_attention"):
raise ImportError(
Expand All @@ -64,9 +68,9 @@ def __call__(
key = attn.to_k(hidden_states)
value = attn.to_v(hidden_states)

query = query.unflatten(2, (attn.heads, -1)).transpose(1, 2)
key = key.unflatten(2, (attn.heads, -1)).transpose(1, 2)
value = value.unflatten(2, (attn.heads, -1)).transpose(1, 2)
query = query.unflatten(2, (attn.heads, -1))
key = key.unflatten(2, (attn.heads, -1))
value = value.unflatten(2, (attn.heads, -1))

# 2. QK normalization
if attn.norm_q is not None:
Expand All @@ -81,46 +85,61 @@ def __call__(
if attn.add_q_proj is None and encoder_hidden_states is not None:
query = torch.cat(
[
apply_rotary_emb(query[:, :, : -encoder_hidden_states.shape[1]], image_rotary_emb),
query[:, :, -encoder_hidden_states.shape[1] :],
apply_rotary_emb(
query[:, : -encoder_hidden_states.shape[1]],
image_rotary_emb,
sequence_dim=1,
),
query[:, -encoder_hidden_states.shape[1] :],
],
dim=2,
dim=1,
)
key = torch.cat(
[
apply_rotary_emb(key[:, :, : -encoder_hidden_states.shape[1]], image_rotary_emb),
key[:, :, -encoder_hidden_states.shape[1] :],
apply_rotary_emb(
key[:, : -encoder_hidden_states.shape[1]],
image_rotary_emb,
sequence_dim=1,
),
key[:, -encoder_hidden_states.shape[1] :],
],
dim=2,
dim=1,
)
else:
query = apply_rotary_emb(query, image_rotary_emb)
key = apply_rotary_emb(key, image_rotary_emb)
query = apply_rotary_emb(query, image_rotary_emb, sequence_dim=1)
key = apply_rotary_emb(key, image_rotary_emb, sequence_dim=1)

# 4. Encoder condition QKV projection and normalization
if attn.add_q_proj is not None and encoder_hidden_states is not None:
encoder_query = attn.add_q_proj(encoder_hidden_states)
encoder_key = attn.add_k_proj(encoder_hidden_states)
encoder_value = attn.add_v_proj(encoder_hidden_states)

encoder_query = encoder_query.unflatten(2, (attn.heads, -1)).transpose(1, 2)
encoder_key = encoder_key.unflatten(2, (attn.heads, -1)).transpose(1, 2)
encoder_value = encoder_value.unflatten(2, (attn.heads, -1)).transpose(1, 2)
encoder_query = encoder_query.unflatten(2, (attn.heads, -1))
encoder_key = encoder_key.unflatten(2, (attn.heads, -1))
encoder_value = encoder_value.unflatten(2, (attn.heads, -1))

if attn.norm_added_q is not None:
encoder_query = attn.norm_added_q(encoder_query)
if attn.norm_added_k is not None:
encoder_key = attn.norm_added_k(encoder_key)

query = torch.cat([query, encoder_query], dim=2)
key = torch.cat([key, encoder_key], dim=2)
value = torch.cat([value, encoder_value], dim=2)
query = torch.cat([query, encoder_query], dim=1)
key = torch.cat([key, encoder_key], dim=1)
value = torch.cat([value, encoder_value], dim=1)

# 5. Attention
hidden_states = F.scaled_dot_product_attention(
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
hidden_states = dispatch_attention_fn(
query,
key,
value,
attn_mask=attention_mask,
dropout_p=0.0,
is_causal=False,
backend=self._attention_backend,
parallel_config=self._parallel_config,
)
hidden_states = hidden_states.transpose(1, 2).flatten(2, 3)
hidden_states = hidden_states.flatten(2, 3)
hidden_states = hidden_states.to(query.dtype)

# 6. Output projection
Expand Down