Dev/zirui/flux mlperf#876

Draft

zirui wants to merge 31 commits into

mainfrom

dev/zirui/flux-mlperf

Commits on Jul 8, 2026

first commit for diffusion training

zirui
and
cursoragent
committed
docs(diffusion): unify single/multi-node torchrun launch in READMEs

zirui
and
cursoragent
committed
docs(diffusion): simplify unified run instructions
zirui
committed
fix(diffusion): broadcast Wan timesteps per sample

zirui
and
cursoragent
committed
Remove unused pip packages and clean up README.md
zirui
committed
fix launch command
zirui
committed
set aiter as default attention backend
zirui
committed
fix: add missing newline at end of examples/diffusion/README.md

Copilot
authored and
zirui
committed
set aiter as default attention backend & add set seperated_timestep=True for wan2.2-5B
zirui
committed
fix(diffusion): address Wan code quality findings

zirui
and
cursoragent
committed
fix(diffusion): resolve remaining code quality findings

zirui
and
cursoragent
committed
fix(diffusion): address Copilot PR review findings

zirui
and
cursoragent
committed
style(diffusion): apply pre-commit formatting

zirui
and
cursoragent
committed
fix(diffusion): address remaining PR review comments

zirui
and
cursoragent
committed
fix(diffusion): align Wan timestep config naming

zirui
and
cursoragent
committed
fix(diffusion): resolve Wan training review issues

zirui
and
cursoragent
committed
Fix black formatting in train_pipeline.py to pass code-lint CI

Copilot
authored and
zirui
committed
fix(diffusion): remove Wan attention debug leftovers

zirui
and
cursoragent
committed
fix(diffusion): resolve Wan review follow-ups

zirui
and
cursoragent
committed
fix(diffusion): address latest Wan review comments

zirui
and
cursoragent
committed
feat(diffusion): add FLUX training support

zirui
and
cursoragent
committed
fix(diffusion): make FLUX examples portable

zirui
and
cursoragent
committed
style(diffusion): apply pre-commit formatting

zirui
and
cursoragent
committed
fix(diffusion): use AITER attention for FLUX

zirui
and
cursoragent
committed
fix(tests): add __init__.py to test dirs to prevent module name conflicts

Copilot
authored and
zirui
committed
fix(diffusion): address PR review comments

zirui
and
cursoragent
committed
docs(diffusion): generalize utils docstring

zirui
and
cursoragent
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Dev/zirui/flux mlperf#876

Dev/zirui/flux mlperf#876
zirui wants to merge 31 commits into
mainfrom
dev/zirui/flux-mlperf

Commits on Jul 8, 2026

first commit for diffusion training

docs(diffusion): unify single/multi-node torchrun launch in READMEs

docs(diffusion): simplify unified run instructions

fix(diffusion): broadcast Wan timesteps per sample

Remove unused pip packages and clean up README.md

fix launch command

set aiter as default attention backend

fix: add missing newline at end of examples/diffusion/README.md

set aiter as default attention backend & add set seperated_timestep=True for wan2.2-5B

fix(diffusion): address Wan code quality findings

fix(diffusion): resolve remaining code quality findings

fix(diffusion): address Copilot PR review findings

style(diffusion): apply pre-commit formatting

fix(diffusion): address remaining PR review comments

fix(diffusion): align Wan timestep config naming

fix(diffusion): resolve Wan training review issues

Fix black formatting in train_pipeline.py to pass code-lint CI

fix(diffusion): remove Wan attention debug leftovers

fix(diffusion): resolve Wan review follow-ups

fix(diffusion): address latest Wan review comments

feat(diffusion): add FLUX training support

fix(diffusion): make FLUX examples portable

style(diffusion): apply pre-commit formatting

fix(diffusion): use AITER attention for FLUX

fix(tests): add init.py to test dirs to prevent module name conflicts

fix(diffusion): address PR review comments

docs(diffusion): generalize utils docstring

Commits on Jul 14, 2026

fix(diffusion): harden FLUX training numerics and edge cases

merge main into FLUX diffusion PR

fix(diffusion): update module utils import after main merge

docs(diffusion): clarify FLUX pretraining launch