Skip to content

Fix training stuck on all-gather cp #2123

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

Merged
zhuzilin merged 1 commit into from
Jul 1, 2026
Merged

Fix training stuck on all-gather cp #2123

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
11 changes: 6 additions & 5 deletions slime/backends/megatron_utils/loss.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -202,7 +202,7 @@ def _allgather_cp_redistribute(
response_length,
dtype=ref_dtype,
device=ref_device,
requires_grad=True,
requires_grad=ref_value.requires_grad,
)
else:
resp_start = s - logit_global_start
Expand All @@ -213,7 +213,7 @@ def _allgather_cp_redistribute(
full_resps.append(full_resp)
seq_start += total_length

# Single differentiable all-reduce to gather full response from all CP ranks
# Single differentiable all-reduce to gather full response from all CP ranks.
all_cat = torch.cat(full_resps, dim=0)
all_cat = dist.nn.all_reduce(all_cat, group=cp_group)

Expand Down Expand Up @@ -444,9 +444,9 @@ def _extract_per_sample(
s = max(logit_global_start, chunk_start)
e = min(logit_global_end, chunk_end)
if e <= s:
log_probs_list.append(torch.zeros((0,), dtype=log_prob_full.dtype, device=log_prob_full.device))
log_probs_list.append(log_prob_full[:0])
if entropy_full is not None:
entropy_list.append(torch.zeros((0,), dtype=entropy_full.dtype, device=entropy_full.device))
entropy_list.append(entropy_full[:0])
else:
log_probs_list.append(log_prob_full[s - chunk_start : e - chunk_start])
if entropy_full is not None:
Expand Down Expand Up @@ -1073,7 +1073,8 @@ def policy_loss_function(
train_rollout_logprob_abs_diff = None
if "rollout_log_probs" in batch and batch["rollout_log_probs"]:
rollout_log_probs = torch.cat(batch["rollout_log_probs"], dim=0)
train_rollout_logprob_abs_diff = sum_of_sample_mean((old_log_probs - rollout_log_probs).abs())
log_probs_to_compare = log_probs if args.use_rollout_logprobs else old_log_probs
train_rollout_logprob_abs_diff = sum_of_sample_mean((log_probs_to_compare - rollout_log_probs).abs())

reported_loss = {
"loss": loss.clone().detach(),
Expand Down
Loading