Skip to content

Add skiplist and more robust calculation on val #965

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鈥檒l occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
yueming-yuan merged 1 commit into from
Apr 9, 2026
Merged

Add skiplist and more robust calculation on val #965

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: 10 additions & 1 deletion miles/backends/training_utils/log_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -121,6 +121,8 @@ def log_rollout_data(rollout_id: int, args: Namespace, rollout_data: RolloutBatc
"rollout_routed_experts",
"max_seq_lens",
"dynamic_global_batch_size",
"weight_versions",
"metadata",
]:
continue
# Upload per sample mean for each rollout value
Expand Down Expand Up @@ -151,7 +153,14 @@ def log_rollout_data(rollout_id: int, args: Namespace, rollout_data: RolloutBatc
else:
val = val.mean() * cp_size
else:
val = sum(val) / len(val)
# Flatten nested lists (e.g. list of lists from async rollout)
flat = val
if isinstance(val[0], (list, tuple)):
flat = [x for sublist in val for x in sublist]
# Skip non-numeric values (e.g. strings from async rollout metadata)
if flat and not isinstance(flat[0], (int, float)):
continue
val = sum(flat) / len(flat)
Comment on lines +157 to +163

@gemini-code-assist gemini-code-assist Bot Apr 9, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

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

high

The current logic for processing list values has a few potential issues that could lead to runtime errors:

  • IndexError: isinstance(val[0], ...) on line 158 will fail if val is an empty list.
  • TypeError: sum(flat) on line 163 will fail if flat contains a mix of numeric and non-numeric types (e.g., [1, 2, 'a']), as the type check on line 161 only inspects the first element.
  • ZeroDivisionError: len(flat) on line 163 can be zero if val is an empty list or a list of empty lists, leading to a division by zero.

To make this more robust, I suggest checking if the list is empty before accessing elements, filtering all items for numeric types, and then calculating the average only if there are numeric items.

flat = val
if val and isinstance(val[0], (list, tuple)):
    flat = [x for sublist in val for x in sublist]
numeric_items = [item for item in flat if isinstance(item, (int, float))]
if not numeric_items:
    continue
val = sum(numeric_items) / len(numeric_items)

elif isinstance(val, torch.Tensor):
val = val.float().mean()
else:
Expand Down
Loading