Skip to content

flux: reduce VRAM usage #10737

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
Kosinkadink merged 1 commit into from
Nov 14, 2025
+26 −7
Merged

flux: reduce VRAM usage #10737

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
33 changes: 26 additions & 7 deletions comfy/ldm/flux/layers.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -167,39 +167,55 @@ def forward(self, img: Tensor, txt: Tensor, vec: Tensor, pe: Tensor, attn_mask=N
img_modulated = self.img_norm1(img)
img_modulated = apply_mod(img_modulated, (1 + img_mod1.scale), img_mod1.shift, modulation_dims_img)
img_qkv = self.img_attn.qkv(img_modulated)
del img_modulated
img_q, img_k, img_v = img_qkv.view(img_qkv.shape[0], img_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
del img_qkv
img_q, img_k = self.img_attn.norm(img_q, img_k, img_v)

# prepare txt for attention
txt_modulated = self.txt_norm1(txt)
txt_modulated = apply_mod(txt_modulated, (1 + txt_mod1.scale), txt_mod1.shift, modulation_dims_txt)
txt_qkv = self.txt_attn.qkv(txt_modulated)
del txt_modulated
txt_q, txt_k, txt_v = txt_qkv.view(txt_qkv.shape[0], txt_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
del txt_qkv
txt_q, txt_k = self.txt_attn.norm(txt_q, txt_k, txt_v)

if self.flipped_img_txt:
q = torch.cat((img_q, txt_q), dim=2)
del img_q, txt_q
k = torch.cat((img_k, txt_k), dim=2)
del img_k, txt_k
v = torch.cat((img_v, txt_v), dim=2)
del img_v, txt_v
# run actual attention
attn = attention(torch.cat((img_q, txt_q), dim=2),
torch.cat((img_k, txt_k), dim=2),
torch.cat((img_v, txt_v), dim=2),
attn = attention(q, k, v,
pe=pe, mask=attn_mask, transformer_options=transformer_options)
del q, k, v

img_attn, txt_attn = attn[:, : img.shape[1]], attn[:, img.shape[1]:]
else:
q = torch.cat((txt_q, img_q), dim=2)
del txt_q, img_q
k = torch.cat((txt_k, img_k), dim=2)
del txt_k, img_k
v = torch.cat((txt_v, img_v), dim=2)
del txt_v, img_v
# run actual attention
attn = attention(torch.cat((txt_q, img_q), dim=2),
torch.cat((txt_k, img_k), dim=2),
torch.cat((txt_v, img_v), dim=2),
attn = attention(q, k, v,
pe=pe, mask=attn_mask, transformer_options=transformer_options)
del q, k, v

txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1]:]

# calculate the img bloks
img += apply_mod(self.img_attn.proj(img_attn), img_mod1.gate, None, modulation_dims_img)
del img_attn
img += apply_mod(self.img_mlp(apply_mod(self.img_norm2(img), (1 + img_mod2.scale), img_mod2.shift, modulation_dims_img)), img_mod2.gate, None, modulation_dims_img)

# calculate the txt bloks
txt += apply_mod(self.txt_attn.proj(txt_attn), txt_mod1.gate, None, modulation_dims_txt)
del txt_attn
txt += apply_mod(self.txt_mlp(apply_mod(self.txt_norm2(txt), (1 + txt_mod2.scale), txt_mod2.shift, modulation_dims_txt)), txt_mod2.gate, None, modulation_dims_txt)

if txt.dtype == torch.float16:
Expand Down Expand Up @@ -249,12 +265,15 @@ def forward(self, x: Tensor, vec: Tensor, pe: Tensor, attn_mask=None, modulation
qkv, mlp = torch.split(self.linear1(apply_mod(self.pre_norm(x), (1 + mod.scale), mod.shift, modulation_dims)), [3 * self.hidden_size, self.mlp_hidden_dim], dim=-1)

q, k, v = qkv.view(qkv.shape[0], qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
del qkv
q, k = self.norm(q, k, v)

# compute attention
attn = attention(q, k, v, pe=pe, mask=attn_mask, transformer_options=transformer_options)
del q, k, v
# compute activation in mlp stream, cat again and run second linear layer
output = self.linear2(torch.cat((attn, self.mlp_act(mlp)), 2))
mlp = self.mlp_act(mlp)
output = self.linear2(torch.cat((attn, mlp), 2))
x += apply_mod(output, mod.gate, None, modulation_dims)
if x.dtype == torch.float16:
x = torch.nan_to_num(x, nan=0.0, posinf=65504, neginf=-65504)
Expand Down