Skip to content

Partially enabled FP16 reasoning in the Qwen Image Models #12522

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
Icbears wants to merge 2 commits into
base: master
Choose a base branch
from
+60 −22
Open

Partially enabled FP16 reasoning in the Qwen Image Models #12522

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
c5549de
Add files via upload
Icbears Feb 19, 2026
526cf1b
Remove unnecessary blank line in model.py
Icbears Feb 19, 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
82 changes: 60 additions & 22 deletions comfy/ldm/qwen_image/model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -150,35 +150,73 @@ def forward(
seq_txt = encoder_hidden_states.shape[1]

# Project and reshape to BHND format (batch, heads, seq, dim)
img_query = self.to_q(hidden_states).view(batch_size, seq_img, self.heads, -1).transpose(1, 2).contiguous()
img_key = self.to_k(hidden_states).view(batch_size, seq_img, self.heads, -1).transpose(1, 2).contiguous()
img_value = self.to_v(hidden_states).view(batch_size, seq_img, self.heads, -1).transpose(1, 2)
if hidden_states.dtype not in (torch.bfloat16,):
# fp16 optimization: IMG_K, IMG_V, TXT_Q, TXT_V use fp16; IMG_Q and TXT_K stay original dtype
img_h = hidden_states.half()
txt_h = encoder_hidden_states.half()

txt_query = self.add_q_proj(encoder_hidden_states).view(batch_size, seq_txt, self.heads, -1).transpose(1, 2).contiguous()
txt_key = self.add_k_proj(encoder_hidden_states).view(batch_size, seq_txt, self.heads, -1).transpose(1, 2).contiguous()
txt_value = self.add_v_proj(encoder_hidden_states).view(batch_size, seq_txt, self.heads, -1).transpose(1, 2)
img_query = self.to_q(hidden_states).view(batch_size, seq_img, self.heads, -1).transpose(1, 2).contiguous()
img_key = self.to_k(img_h).view(batch_size, seq_img, self.heads, -1).transpose(1, 2).contiguous()
img_value = self.to_v(img_h).view(batch_size, seq_img, self.heads, -1).transpose(1, 2)

img_query = self.norm_q(img_query)
img_key = self.norm_k(img_key)
txt_query = self.norm_added_q(txt_query)
txt_key = self.norm_added_k(txt_key)
txt_query = self.add_q_proj(txt_h).view(batch_size, seq_txt, self.heads, -1).transpose(1, 2).contiguous()
txt_key = self.add_k_proj(encoder_hidden_states).view(batch_size, seq_txt, self.heads, -1).transpose(1, 2).contiguous()
txt_value = self.add_v_proj(txt_h).view(batch_size, seq_txt, self.heads, -1).transpose(1, 2)

joint_query = torch.cat([txt_query, img_query], dim=2)
joint_key = torch.cat([txt_key, img_key], dim=2)
joint_value = torch.cat([txt_value, img_value], dim=2)
del img_h, txt_h

joint_query = apply_rope1(joint_query, image_rotary_emb)
joint_key = apply_rope1(joint_key, image_rotary_emb)
img_query = self.norm_q(img_query)
img_key = self.norm_k(img_key)
txt_query = self.norm_added_q(txt_query)
txt_key = self.norm_added_k(txt_key)

if encoder_hidden_states_mask is not None:
attn_mask = torch.zeros((batch_size, 1, seq_txt + seq_img), dtype=hidden_states.dtype, device=hidden_states.device)
attn_mask[:, 0, :seq_txt] = encoder_hidden_states_mask
joint_query = torch.cat([txt_query, img_query], dim=2)
joint_key = torch.cat([txt_key, img_key], dim=2)
joint_value = torch.cat([txt_value, img_value], dim=2)

joint_query = apply_rope1(joint_query, image_rotary_emb)
joint_key = apply_rope1(joint_key, image_rotary_emb)

if encoder_hidden_states_mask is not None:
attn_mask = torch.zeros((batch_size, 1, seq_txt + seq_img), dtype=hidden_states.dtype, device=hidden_states.device)
attn_mask[:, 0, :seq_txt] = encoder_hidden_states_mask
else:
attn_mask = None

joint_hidden_states = optimized_attention_masked(joint_query.half(), joint_key.half(), joint_value.half(), self.heads,
attn_mask.half() if attn_mask is not None else None, transformer_options=transformer_options,
skip_reshape=True).float()
else:
attn_mask = None
# bfloat16: use original dtype throughout
img_query = self.to_q(hidden_states).view(batch_size, seq_img, self.heads, -1).transpose(1, 2).contiguous()
img_key = self.to_k(hidden_states).view(batch_size, seq_img, self.heads, -1).transpose(1, 2).contiguous()
img_value = self.to_v(hidden_states).view(batch_size, seq_img, self.heads, -1).transpose(1, 2)

txt_query = self.add_q_proj(encoder_hidden_states).view(batch_size, seq_txt, self.heads, -1).transpose(1, 2).contiguous()
txt_key = self.add_k_proj(encoder_hidden_states).view(batch_size, seq_txt, self.heads, -1).transpose(1, 2).contiguous()
txt_value = self.add_v_proj(encoder_hidden_states).view(batch_size, seq_txt, self.heads, -1).transpose(1, 2)

img_query = self.norm_q(img_query)
img_key = self.norm_k(img_key)
txt_query = self.norm_added_q(txt_query)
txt_key = self.norm_added_k(txt_key)

joint_query = torch.cat([txt_query, img_query], dim=2)
joint_key = torch.cat([txt_key, img_key], dim=2)
joint_value = torch.cat([txt_value, img_value], dim=2)

joint_query = apply_rope1(joint_query, image_rotary_emb)
joint_key = apply_rope1(joint_key, image_rotary_emb)

if encoder_hidden_states_mask is not None:
attn_mask = torch.zeros((batch_size, 1, seq_txt + seq_img), dtype=hidden_states.dtype, device=hidden_states.device)
attn_mask[:, 0, :seq_txt] = encoder_hidden_states_mask
else:
attn_mask = None

joint_hidden_states = optimized_attention_masked(joint_query, joint_key, joint_value, self.heads,
attn_mask, transformer_options=transformer_options,
skip_reshape=True)
joint_hidden_states = optimized_attention_masked(joint_query, joint_key, joint_value, self.heads,
attn_mask, transformer_options=transformer_options,
skip_reshape=True)

txt_attn_output = joint_hidden_states[:, :seq_txt, :]
img_attn_output = joint_hidden_states[:, seq_txt:, :]
Expand Down