Skip to content

fix: align correct-sample rewards with DP-local lengths #1900

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
miamia0 wants to merge 2 commits into
base: main
Choose a base branch
from
+9 −2
Open

fix: align correct-sample rewards with DP-local lengths #1900

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
84db77d
fix: align correct-sample rewards with DP-local lengths
miamia0 May 10, 2026
9f6be82
Merge branch 'main' into fix/correct-sample-raw-reward-alignment
miamia0 May 11, 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
7 changes: 5 additions & 2 deletions slime/backends/megatron_utils/data.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -416,6 +416,7 @@ def log_rollout_data(
"rollout_routed_experts",
"max_seq_lens",
"dynamic_global_batch_size",
"raw_rewards_partitioned",
]:
continue
# Upload per sample mean for each rollout value
Expand Down Expand Up @@ -505,13 +506,15 @@ def quantile(total_value, n_quantiles, data) -> dict:
percentile = {f"p{min(math.ceil(q*100),100)}": p for q, p in zip(quantiles, percentile, strict=True)}
return percentile

raw_rewards = rollout_data["raw_reward"]
# raw_reward may stay global for passrate; correct-sample metrics
# must use rewards aligned with local response_lengths/total_lengths.
raw_rewards_partitioned = rollout_data["raw_rewards_partitioned"]
# Additional metrics for correct cases are calculated separately below.
correct_response_lengths = []
correct_total_lengths = []
correct_loss_masks = []
correct_entropy = []
for i, raw_reward in enumerate(raw_rewards):
for i, raw_reward in enumerate(raw_rewards_partitioned):
if raw_reward == 1:
correct_response_lengths.append(response_lengths[i])
correct_total_lengths.append(total_lengths[i])
Expand Down
4 changes: 4 additions & 0 deletions slime/utils/data.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -306,5 +306,9 @@ def process_rollout_data(args, rollout_data_ref, dp_rank, dp_size):
# save the seqlen of the whole rollout batch
Timer().seq_lens = total_lengths
rollout_data["total_lengths"] = [total_lengths[i] for i in partition]
if "raw_reward" in rollout_data:
# Keep raw_reward global for passrate metrics, but provide a DP-local view
# for correct-sample metrics that index local response_lengths/total_lengths.
rollout_data["raw_rewards_partitioned"] = [rollout_data["raw_reward"][i] for i in partition]

return rollout_data