Skip to content

fix(trainer): fix train/eval metric deflation at eval steps #302

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
liusong1222 merged 1 commit into from
May 14, 2026
Merged

fix(trainer): fix train/eval metric deflation at eval steps #302

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
43 changes: 27 additions & 16 deletions angelslim/compressor/speculative/train/trainer/eagle3_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -47,10 +47,10 @@ def __init__(self, draft_model: nn.Module, length: int, **kwargs):
super().__init__(model=draft_model, **kwargs)
self.length = length
self._train_start_time = None
self._pending_log: dict = (
{}
) # cache acc/ploss log for merging with base Trainer's loss log
self._pending_log_count: int = 0 # accumulated batch count for averaging the cached log
self._train_pending_log: dict = {}
self._train_pending_log_count: int = 0
self._eval_pending_log: dict = {}
self._eval_pending_log_count: int = 0

def train(self, *args, **kwargs):
"""Override train method to record training start time for estimating remaining time."""
Expand All @@ -59,12 +59,11 @@ def train(self, *args, **kwargs):

def log(self, logs: dict, start_time: Optional[float] = None) -> None:
"""
rewrite log method to merge acc/ploss log with base Trainer's loss log.
Merge acc/ploss accumulators with the base Trainer's loss log.
"""
if "loss" in logs and self._pending_log:
# merge cached acc/ploss data (average)
count = max(self._pending_log_count, 1)
acc_ploss = {k: v / count for k, v in self._pending_log.items()}
if "loss" in logs and self._train_pending_log:
train_count = max(self._train_pending_log_count, 1)
acc_ploss = {k: v / train_count for k, v in self._train_pending_log.items()}
merged = {}

# step
Expand All @@ -85,9 +84,16 @@ def log(self, logs: dict, start_time: Optional[float] = None) -> None:
if "learning_rate" in logs:
merged["lr"] = logs["learning_rate"]

# acc/ploss
# train acc/ploss
merged.update(acc_ploss)

# eval acc/ploss — merged when a training log fires
if self._eval_pending_log:
eval_count = max(self._eval_pending_log_count, 1)
merged.update({k: v / eval_count for k, v in self._eval_pending_log.items()})
self._eval_pending_log.clear()
self._eval_pending_log_count = 0

# remaining_time
if (
self.state is not None
Expand All @@ -102,8 +108,8 @@ def log(self, logs: dict, start_time: Optional[float] = None) -> None:
minutes, seconds = divmod(remainder, 60)
merged["remaining_time"] = f"{hours:02d}h:{minutes:02d}m:{seconds:02d}s"

self._pending_log.clear()
self._pending_log_count = 0
self._train_pending_log.clear()
self._train_pending_log_count = 0
super().log(merged, start_time)
else:
super().log(logs, start_time)
Expand Down Expand Up @@ -294,10 +300,15 @@ def draft_model_training_time_test(

log = {f"{log_prefix}/acc_{i}": acces[i] for i in range(len(acces))}
log.update({f"{log_prefix}/ploss_{i}": plosses[i].item() for i in range(len(plosses))})
# Cache log for merging with base Trainer's loss log
for k, v in log.items():
self._pending_log[k] = self._pending_log.get(k, 0.0) + v
self._pending_log_count += 1
# Route into the appropriate accumulator.
if log_prefix == "eval":
for k, v in log.items():
self._eval_pending_log[k] = self._eval_pending_log.get(k, 0.0) + v
self._eval_pending_log_count += 1
else:
for k, v in log.items():
self._train_pending_log[k] = self._train_pending_log.get(k, 0.0) + v
self._train_pending_log_count += 1
# Step 9: Return loss
return ploss

Expand Down
Loading