Skip to content

Include completion_mask in return statement #528

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
danielhanchen merged 6 commits into from
Mar 8, 2026
Merged

Include completion_mask in return statement #528

Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
74c0994
Include completion_mask in return statement
pluesclues Mar 2, 2026
a41bdbd
Refactor hidden_states handling in forward method
pluesclues Mar 2, 2026
f3936d2
Include mask in grpo_compute_loss return values
pluesclues Mar 3, 2026
bb459d2
Merge branch 'unslothai:main' into completion_mask_fix
pluesclues Mar 3, 2026
6d42b0c
Fix argument passing in compute_loss function
pluesclues Mar 3, 2026
cc0ffd3
Clarify memory management in forward method
pluesclues Mar 3, 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
10 changes: 5 additions & 5 deletions unsloth_zoo/rl_replacements.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -481,7 +481,7 @@ def masked_batch_mean(x):
mean_kl = mean_kl_per_reward.mean()
return completion_length, mean_kl
completion_length, mean_kl = masked_batch_mean(kl_i)
return loss, completion_length, mean_kl, delta, flat_is_ratio, coef_1
return loss, completion_length, mean_kl, delta, flat_is_ratio, coef_1, mask
Copy link
Copy Markdown

@chatgpt-codex-connector chatgpt-codex-connector Bot Mar 3, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P1 Badge Keep grpo_compute_loss return arity consistent

grpo_compute_loss now returns 7 values, but the UnslothEfficientGRPO.forward helper compute_loss still unpacks only 6 (loss, completion_length, mean_kl, delta, flat_is_ratio, coef_1 = grpo_compute_loss(...) at rl_replacements.py:504). In the GRPO training path this raises ValueError: too many values to unpack as soon as compute_loss is invoked, so training fails before loss/grad accumulation completes.

Useful? React with 👍 / 👎.

pass
RL_REPLACEMENTS["grpo_compute_loss"] = grpo_compute_loss
RL_REPLACEMENTS["grpo_compute_loss_slow"] = \
Expand All @@ -501,7 +501,7 @@ def forward(ctx, _new_logps, _old_logps, _ref_logps, _sampling_per_token_logps,
if extra_kwargs is None:
extra_kwargs = {}
def compute_loss(new_logps, old_logps, ref_logps, sampling_per_token_logps, input_ids, mask, advantages, scaling):
loss, completion_length, mean_kl, delta, flat_is_ratio, coef_1 = grpo_compute_loss(
loss, completion_length, mean_kl, delta, flat_is_ratio, coef_1, _mask = grpo_compute_loss(
ref_logps,
new_logps,
old_logps,
Expand Down Expand Up @@ -845,8 +845,8 @@ class Unsloth_Offloaded_Log_Softmax(torch.autograd.Function):
def forward(ctx, hidden_states, lm_head, index, chunks,
logit_scale_multiply, logit_scale_divide,
logit_softcapping, temperature):

ctx.saved_hidden_states = to_device(hidden_states, "cpu", non_blocking=True)
#Only the activations are needed so if we keep entire computational graph, keeps unnecessary memory on CPU so we detach it
ctx.saved_hidden_states = hidden_states.detach().contiguous().to("cpu", non_blocking=True)
Copy link
Copy Markdown
Collaborator

@Datta0 Datta0 Mar 3, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Please include a comment here describing the change and its implications (can ref the issue).

Copy link
Copy Markdown
Collaborator Author

@pluesclues pluesclues Mar 3, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Did that

ctx.device = hidden_states.device
ctx.dtype = hidden_states.dtype

Expand Down Expand Up @@ -990,7 +990,7 @@ def efficient_log_softmax(hidden_states, lm_head, index, chunks=32,
# Must force not returning hidden states but logits otherwise gibberish
os.environ["UNSLOTH_RETURN_HIDDEN_STATES"] = "0"

return loss, completion_length, mean_kl, delta, flat_is_ratio, coef_1
return loss, completion_length, mean_kl, delta, flat_is_ratio, coef_1, completion_mask
# Old non efficient code path
new_logits = torch.matmul(new_hidden_states, lm_head.t())
new_logits = new_logits[:, :-1, :] # exclude the last logit: it corresponds to the next token pred
Expand Down