Skip to content

[fsdp] fix: preserve Hugging Face fp32-keep modules in FSDP2 #7165

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
chengcuiping wants to merge 4 commits into
base: main
Choose a base branch
from
Open

[fsdp] fix: preserve Hugging Face fp32-keep modules in FSDP2 #7165

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
dee36eb
[fsdp] fix: preserve fp32-keep modules in FSDP2
chengcuiping Jul 27, 2026
f695687
[fsdp] fix: preserve fp32 activation boundaries
MagellaX Jul 28, 2026
53e9899
[fsdp] test: cover fp16 activation-boundary parity
chengcuiping Jul 29, 2026
f621a30
[fsdp] test: cover fp32 keep with forward prefetch
chengcuiping Sep 22, 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
3 changes: 3 additions & 0 deletions tests/special_distributed/run_all.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,3 +25,6 @@ torchrun --nproc-per-node=2 --standalone tests/special_distributed/test_fsdp2_cp
# Regression for colocated FSDP2 full-parameter model transfers. It locks in
# PyTorch's implicit pinned allocation for non-blocking D2H Module.to().
torchrun --nproc-per-node=2 --standalone tests/special_distributed/test_fsdp2_pinned_model_transfer.py
# Regression for verl#7092 (FSDP2 must honour HF _keep_in_fp32_modules*). Two
# ranks are enough to exercise sharded fp32 units and the full-state-dict load.
torchrun --nproc-per-node=2 --standalone tests/special_distributed/test_fsdp2_keep_in_fp32.py
Loading