Skip to content

Add elementwise fusions #9495

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

Merged
comfyanonymous merged 2 commits into from
Aug 22, 2025
Merged

Add elementwise fusions #9495

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
714950f
Add elementwise fusions
contentis Aug 22, 2025
17038d5
Add addcmul pattern to Qwen
contentis Aug 22, 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
12 changes: 7 additions & 5 deletions comfy/ldm/modules/diffusionmodules/mmdit.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -109,7 +109,7 @@ def forward(self, x):
def modulate(x, shift, scale):
if shift is None:
shift = torch.zeros_like(scale)
return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
return torch.addcmul(shift.unsqueeze(1), x, 1+ scale.unsqueeze(1))


#################################################################################
Expand Down Expand Up @@ -564,10 +564,7 @@ def post_attention_x(self, attn, attn2, x, gate_msa, shift_mlp, scale_mlp, gate_
assert not self.pre_only
attn1 = self.attn.post_attention(attn)
attn2 = self.attn2.post_attention(attn2)
out1 = gate_msa.unsqueeze(1) * attn1
out2 = gate_msa2.unsqueeze(1) * attn2
x = x + out1
x = x + out2
x = gate_cat(x, gate_msa, gate_msa2, attn1, attn2)
x = x + gate_mlp.unsqueeze(1) * self.mlp(
modulate(self.norm2(x), shift_mlp, scale_mlp)
)
Expand All @@ -594,6 +591,11 @@ def forward(self, x: torch.Tensor, c: torch.Tensor) -> torch.Tensor:
)
return self.post_attention(attn, *intermediates)

def gate_cat(x, gate_msa, gate_msa2, attn1, attn2):
out1 = gate_msa.unsqueeze(1) * attn1
out2 = gate_msa2.unsqueeze(1) * attn2
x = torch.stack([x, out1, out2], dim=0).sum(dim=0)
return x

def block_mixing(*args, use_checkpoint=True, **kwargs):
if use_checkpoint:
Expand Down
12 changes: 6 additions & 6 deletions comfy/ldm/qwen_image/model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -214,9 +214,9 @@ def __init__(
operations=operations,
)

def _modulate(self, x, mod_params):
shift, scale, gate = mod_params.chunk(3, dim=-1)
return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1), gate.unsqueeze(1)
def _modulate(self, x: torch.Tensor, mod_params: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
shift, scale, gate = torch.chunk(mod_params, 3, dim=-1)
return torch.addcmul(shift.unsqueeze(1), x, 1 + scale.unsqueeze(1)), gate.unsqueeze(1)

def forward(
self,
Expand Down Expand Up @@ -248,11 +248,11 @@ def forward(

img_normed2 = self.img_norm2(hidden_states)
img_modulated2, img_gate2 = self._modulate(img_normed2, img_mod2)
hidden_states = hidden_states + img_gate2 * self.img_mlp(img_modulated2)
hidden_states = torch.addcmul(hidden_states, img_gate2, self.img_mlp(img_modulated2))

txt_normed2 = self.txt_norm2(encoder_hidden_states)
txt_modulated2, txt_gate2 = self._modulate(txt_normed2, txt_mod2)
encoder_hidden_states = encoder_hidden_states + txt_gate2 * self.txt_mlp(txt_modulated2)
encoder_hidden_states = torch.addcmul(encoder_hidden_states, txt_gate2, self.txt_mlp(txt_modulated2))

return encoder_hidden_states, hidden_states

Expand All @@ -275,7 +275,7 @@ def __init__(
def forward(self, x: torch.Tensor, conditioning_embedding: torch.Tensor) -> torch.Tensor:
emb = self.linear(self.silu(conditioning_embedding))
scale, shift = torch.chunk(emb, 2, dim=1)
x = self.norm(x) * (1 + scale)[:, None, :] + shift[:, None, :]
x = torch.addcmul(shift[:, None, :], self.norm(x), (1 + scale)[:, None, :])
return x


Expand Down
14 changes: 7 additions & 7 deletions comfy/ldm/wan/model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -148,8 +148,8 @@ def forward(self, x, context, context_img_len):

def repeat_e(e, x):
repeats = 1
if e.shape[1] > 1:
repeats = x.shape[1] // e.shape[1]
if e.size(1) > 1:
repeats = x.size(1) // e.size(1)
if repeats == 1:
return e
return torch.repeat_interleave(e, repeats, dim=1)
Expand Down Expand Up @@ -219,15 +219,15 @@ def forward(

# self-attention
y = self.self_attn(
self.norm1(x) * (1 + repeat_e(e[1], x)) + repeat_e(e[0], x),
torch.addcmul(repeat_e(e[0], x), self.norm1(x), 1 + repeat_e(e[1], x)),
freqs)

x = x + y * repeat_e(e[2], x)
x = torch.addcmul(x, y, repeat_e(e[2], x))

# cross-attention & ffn
x = x + self.cross_attn(self.norm3(x), context, context_img_len=context_img_len)
y = self.ffn(self.norm2(x) * (1 + repeat_e(e[4], x)) + repeat_e(e[3], x))
x = x + y * repeat_e(e[5], x)
y = self.ffn(torch.addcmul(repeat_e(e[3], x), self.norm2(x), 1 + repeat_e(e[4], x)))
x = torch.addcmul(x, y, repeat_e(e[5], x))
return x


Expand Down Expand Up @@ -342,7 +342,7 @@ def forward(self, x, e):
else:
e = (comfy.model_management.cast_to(self.modulation, dtype=x.dtype, device=x.device).unsqueeze(0) + e.unsqueeze(2)).unbind(2)

x = (self.head(self.norm(x) * (1 + repeat_e(e[1], x)) + repeat_e(e[0], x)))
x = (self.head(torch.addcmul(repeat_e(e[0], x), self.norm(x), 1 + repeat_e(e[1], x))))
return x


Expand Down
Loading