Skip to content

fix(flux): cast Q/K to value dtype before core attention #914

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

Draft
gphuang wants to merge 2 commits into
base: main
Choose a base branch
from
Draft

fix(flux): cast Q/K to value dtype before core attention #914

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
14cc0e7
fix(flux): cast Q/K to value dtype before core attention
gphuang Jul 23, 2026
ac7bc75
Merge branch 'main' into fix/flux/qk-attention-dtype-agi
gphuang Jul 24, 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
12 changes: 12 additions & 0 deletions primus/backends/megatron/core/models/diffusion/flux/attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -390,6 +390,12 @@ def forward(
query = apply_rotary_pos_emb(query, q_pos_emb, config=self.config, cu_seqlens=cu_seqlens_q)
key = apply_rotary_pos_emb(key, k_pos_emb, config=self.config, cu_seqlens=cu_seqlens_kv)

# Q/K RMSNorm (for_qk) can leave Q/K in fp32 while V stays bf16; AITER
# FlashAttention requires Q, K, V to share one dtype. Align Q/K to V at the
# attention-kernel boundary (covers both RMSNorm and RoPE promotion).
query = query.to(value.dtype)
key = key.to(value.dtype)

# Core attention computation
if self.checkpoint_core_attention and self.training:
core_attn_out = self._checkpointed_attention_forward(
Expand Down Expand Up @@ -566,6 +572,12 @@ def forward(
query = apply_rotary_pos_emb(query, q_pos_emb, config=self.config, cu_seqlens=cu_seqlens_q)
key = apply_rotary_pos_emb(key, k_pos_emb, config=self.config, cu_seqlens=cu_seqlens_kv)

# Q/K RMSNorm (for_qk) can leave Q/K in fp32 while V stays bf16; AITER
# FlashAttention requires Q, K, V to share one dtype. Align Q/K to V at the
# attention-kernel boundary (covers both RMSNorm and RoPE promotion).
query = query.to(value.dtype)
key = key.to(value.dtype)

# Core attention computation
if self.checkpoint_core_attention and self.training:
core_attn_out = self._checkpointed_attention_forward(
Expand Down