Skip to content

fix grpo example #9352

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
hjh0119 wants to merge 2 commits into
base: main
Choose a base branch
from
Open

fix grpo example #9352

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
8073a0a
fix example
hjh0119 May 15, 2026
5c2374e
update doc
hjh0119 May 15, 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: 2 additions & 1 deletion docs/source/Instruction/GRPO/AdvancedResearch/GSPO.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -56,7 +56,8 @@ importance_weights = torch.exp(log_importance_weights)
```bash
--epsilon 3e-4 # from paper section 5.1
--epsilon_high 4e-4 # from paper section 5.1
--steps_per_generation 4 # from paper section 5.1 (each batch of rollout data is partitioned into four minibatches for gradient updates)
--gradient_accumulation_steps 8
--steps_per_generation 32 # from paper section 5.1 (each batch of rollout data is partitioned into four minibatches for gradient updates)
--beta 0 # zero kl regularization https://github.com/volcengine/verl/pull/2775#issuecomment-3131807306
```

Expand Down
3 changes: 2 additions & 1 deletion docs/source_en/Instruction/GRPO/AdvancedResearch/GSPO.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -58,7 +58,8 @@ Other hyperparameters in the paper
```bash
--epsilon 3e-4 # from paper section 5.1
--epsilon_high 4e-4 # from paper section 5.1
--steps_per_generation 4 # from paper section 5.1 (each batch of rollout data is partitioned into four minibatches for gradient updates)
--gradient_accumulation_steps 8
--steps_per_generation 32 # from paper section 5.1 (each batch of rollout data is partitioned into four minibatches for gradient updates)
--beta 0 # zero kl regularization https://github.com/volcengine/verl/pull/2775#issuecomment-3131807306
```

Expand Down
1 change: 0 additions & 1 deletion examples/train/grpo/internal/chord.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,7 +19,6 @@ swift rlhf \
--load_from_cache_file true \
--torch_dtype bfloat16 \
--beta 0.0 \
--steps_per_generation 4 \
--num_train_epochs 1 \
--per_device_train_batch_size 4 \
--gradient_accumulation_steps 8 \
Expand Down
4 changes: 2 additions & 2 deletions examples/train/grpo/internal/gspo.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3,7 +3,7 @@
# hyperparameter
# - epsilon = 3e-4 from paper serction 5.1
# - epsilon_high = 4e-4 from paper serction 5.1
# - steps_per_generation = 4 from paper serction 5.1 (each batch of rollout data is partitioned into four minibatches for gradient updates)
# - steps_per_generation = 32 (= 4 * gradient_accumulation_steps): paper section 5.1 partitions each batch of rollout data into four minibatches for gradient updates; in swift HF GRPO, steps_per_generation counts micro-batches, so it must be multiplied by gradient_accumulation_steps
# - beta = 0: zero kl regularization https://github.com/volcengine/verl/pull/2775#issuecomment-3131807306

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
Expand All @@ -17,7 +17,7 @@ swift rlhf \
--beta 0.0 \
--epsilon 3e-4 \
--epsilon_high 4e-4 \
--steps_per_generation 4 \
--steps_per_generation 32 \
--importance_sampling_level sequence \
--num_train_epochs 1 \
--per_device_train_batch_size 2 \
Expand Down
Loading