Skip to content

Drop the float32 copies in the Wan block residual adds #15225

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
klin2024 wants to merge 1 commit into
base: main
Choose a base branch
from
+24 −18
Open

Drop the float32 copies in the Wan block residual adds #15225

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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 src/diffusers/models/transformers/transformer_anyflow.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -487,7 +487,8 @@ def forward(
attn1_kwargs["kv_cache"] = kv_cache
attn1_kwargs["kv_cache_flag"] = kv_cache_flag
attn_output = self.attn1(**attn1_kwargs)
hidden_states = (hidden_states.float() + attn_output * gate_msa).type_as(hidden_states)
# gate_msa is float32, so the residual add is float32 through type promotion (no float32 copy of hidden_states)
hidden_states = (hidden_states + attn_output * gate_msa).type_as(hidden_states)

# 2. Cross-attention
norm_hidden_states = self.norm2(hidden_states.float()).type_as(hidden_states)
Expand All @@ -499,7 +500,7 @@ def forward(
hidden_states
)
ff_output = self.ffn(norm_hidden_states)
hidden_states = (hidden_states.float() + ff_output.float() * c_gate_msa).type_as(hidden_states)
hidden_states = (hidden_states + ff_output * c_gate_msa).type_as(hidden_states)

return hidden_states

Expand Down
5 changes: 3 additions & 2 deletions src/diffusers/models/transformers/transformer_anyflow_far.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -600,7 +600,8 @@ def forward(
attn1_kwargs["kv_cache"] = kv_cache
attn1_kwargs["kv_cache_flag"] = kv_cache_flag
attn_output = self.attn1(**attn1_kwargs)
hidden_states = (hidden_states.float() + attn_output * gate_msa).type_as(hidden_states)
# gate_msa is float32, so the residual add is float32 through type promotion (no float32 copy of hidden_states)
hidden_states = (hidden_states + attn_output * gate_msa).type_as(hidden_states)

# 2. Cross-attention
norm_hidden_states = self.norm2(hidden_states.float()).type_as(hidden_states)
Expand All @@ -612,7 +613,7 @@ def forward(
hidden_states
)
ff_output = self.ffn(norm_hidden_states)
hidden_states = (hidden_states.float() + ff_output.float() * c_gate_msa).type_as(hidden_states)
hidden_states = (hidden_states + ff_output * c_gate_msa).type_as(hidden_states)

return hidden_states

Expand Down
5 changes: 3 additions & 2 deletions src/diffusers/models/transformers/transformer_chronoedit.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -502,7 +502,8 @@ def forward(
# 1. Self-attention
norm_hidden_states = (self.norm1(hidden_states.float()) * (1 + scale_msa) + shift_msa).type_as(hidden_states)
attn_output = self.attn1(norm_hidden_states, None, None, rotary_emb)
hidden_states = (hidden_states.float() + attn_output * gate_msa).type_as(hidden_states)
# gate_msa is float32, so the residual add is float32 through type promotion (no float32 copy of hidden_states)
hidden_states = (hidden_states + attn_output * gate_msa).type_as(hidden_states)

# 2. Cross-attention
norm_hidden_states = self.norm2(hidden_states.float()).type_as(hidden_states)
Expand All @@ -514,7 +515,7 @@ def forward(
hidden_states
)
ff_output = self.ffn(norm_hidden_states)
hidden_states = (hidden_states.float() + ff_output.float() * c_gate_msa).type_as(hidden_states)
hidden_states = (hidden_states + ff_output * c_gate_msa).type_as(hidden_states)

return hidden_states

Expand Down
5 changes: 3 additions & 2 deletions src/diffusers/models/transformers/transformer_helios.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -458,7 +458,8 @@ def forward(
rotary_emb,
original_context_length,
)
hidden_states = (hidden_states.float() + attn_output * gate_msa).type_as(hidden_states)
# gate_msa is float32, so the residual add is float32 through type promotion (no float32 copy of hidden_states)
hidden_states = (hidden_states + attn_output * gate_msa).type_as(hidden_states)

# 2. Cross-attention
if self.guidance_cross_attn:
Expand Down Expand Up @@ -493,7 +494,7 @@ def forward(
hidden_states
)
ff_output = self.ffn(norm_hidden_states)
hidden_states = (hidden_states.float() + ff_output.float() * c_gate_msa).type_as(hidden_states)
hidden_states = (hidden_states + ff_output * c_gate_msa).type_as(hidden_states)

return hidden_states

Expand Down
5 changes: 3 additions & 2 deletions src/diffusers/models/transformers/transformer_skyreels_v2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -498,7 +498,8 @@ def forward(
# 1. Self-attention
norm_hidden_states = (self.norm1(hidden_states.float()) * (1 + scale_msa) + shift_msa).type_as(hidden_states)
attn_output = self.attn1(norm_hidden_states, None, attention_mask, rotary_emb)
hidden_states = (hidden_states.float() + attn_output * gate_msa).type_as(hidden_states)
# gate_msa is float32, so the residual add is float32 through type promotion (no float32 copy of hidden_states)
hidden_states = (hidden_states + attn_output * gate_msa).type_as(hidden_states)

# 2. Cross-attention
norm_hidden_states = self.norm2(hidden_states.float()).type_as(hidden_states)
Expand All @@ -510,7 +511,7 @@ def forward(
hidden_states
)
ff_output = self.ffn(norm_hidden_states)
hidden_states = (hidden_states.float() + ff_output.float() * c_gate_msa).type_as(hidden_states)
hidden_states = (hidden_states + ff_output * c_gate_msa).type_as(hidden_states)

return hidden_states

Expand Down
5 changes: 3 additions & 2 deletions src/diffusers/models/transformers/transformer_wan.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -487,7 +487,8 @@ def forward(
# 1. Self-attention
norm_hidden_states = (self.norm1(hidden_states.float()) * (1 + scale_msa) + shift_msa).type_as(hidden_states)
attn_output = self.attn1(norm_hidden_states, None, None, rotary_emb)
hidden_states = (hidden_states.float() + attn_output * gate_msa).type_as(hidden_states)
# gate_msa is float32, so the residual add is float32 through type promotion (no float32 copy of hidden_states)
hidden_states = (hidden_states + attn_output * gate_msa).type_as(hidden_states)

# 2. Cross-attention
norm_hidden_states = self.norm2(hidden_states.float()).type_as(hidden_states)
Expand All @@ -499,7 +500,7 @@ def forward(
hidden_states
)
ff_output = self.ffn(norm_hidden_states)
hidden_states = (hidden_states.float() + ff_output.float() * c_gate_msa).type_as(hidden_states)
hidden_states = (hidden_states + ff_output * c_gate_msa).type_as(hidden_states)

return hidden_states

Expand Down
5 changes: 3 additions & 2 deletions src/diffusers/models/transformers/transformer_wan_animate.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -966,7 +966,8 @@ def forward(
# 1. Self-attention
norm_hidden_states = (self.norm1(hidden_states.float()) * (1 + scale_msa) + shift_msa).type_as(hidden_states)
attn_output = self.attn1(norm_hidden_states, None, None, rotary_emb)
hidden_states = (hidden_states.float() + attn_output * gate_msa).type_as(hidden_states)
# gate_msa is float32, so the residual add is float32 through type promotion (no float32 copy of hidden_states)
hidden_states = (hidden_states + attn_output * gate_msa).type_as(hidden_states)

# 2. Cross-attention
norm_hidden_states = self.norm2(hidden_states.float()).type_as(hidden_states)
Expand All @@ -978,7 +979,7 @@ def forward(
hidden_states
)
ff_output = self.ffn(norm_hidden_states)
hidden_states = (hidden_states.float() + ff_output.float() * c_gate_msa).type_as(hidden_states)
hidden_states = (hidden_states + ff_output * c_gate_msa).type_as(hidden_states)

return hidden_states

Expand Down
7 changes: 3 additions & 4 deletions src/diffusers/models/transformers/transformer_wan_vace.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -112,7 +112,8 @@ def forward(
control_hidden_states
)
attn_output = self.attn1(norm_hidden_states, None, None, rotary_emb)
control_hidden_states = (control_hidden_states.float() + attn_output * gate_msa).type_as(control_hidden_states)
# gate_msa is float32, so the residual add is float32 through type promotion (no float32 copy of the input)
control_hidden_states = (control_hidden_states + attn_output * gate_msa).type_as(control_hidden_states)

# 2. Cross-attention
norm_hidden_states = self.norm2(control_hidden_states.float()).type_as(control_hidden_states)
Expand All @@ -124,9 +125,7 @@ def forward(
control_hidden_states
)
ff_output = self.ffn(norm_hidden_states)
control_hidden_states = (control_hidden_states.float() + ff_output.float() * c_gate_msa).type_as(
control_hidden_states
)
control_hidden_states = (control_hidden_states + ff_output * c_gate_msa).type_as(control_hidden_states)

conditioning_states = None
if self.proj_out is not None:
Expand Down
Loading