ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

modified the num_tokens logic

Browse source
This commit is contained in:
sshah229 2026-02-16 00:38:40 -07:00
commit 63b34660ed
1 changed files with 4 additions and 14 deletions
Download patch file Download diff file Expand all files Collapse all files

18
studio/backend/core/training/trainer.py
View file

@ -567,6 +567,7 @@ class UnslothTrainer:
"seed": training_args.get('random_seed', 3407),
"output_dir": output_dir,
"report_to": ["wandb"] if training_args.get('enable_wandb', False) else "none",
"include_num_input_tokens_seen": True, # Enable token counting
}
# Add warmup parameter - use warmup_ratio if provided, otherwise warmup_steps
@ -739,20 +740,9 @@ class UnslothTrainer:
time_per_step = elapsed_seconds / current_step
eta_seconds = time_per_step * steps_remaining
# Calculate num_tokens
num_tokens = None
if (self.trainer_instance.batch_size is not None and
self.trainer_instance.max_seq_length is not None and
self.trainer_instance.gradient_accumulation_steps is not None):
# Approximate: batch_size * gradient_accumulation_steps * max_seq_length * steps
# Note: This is an approximation as actual sequences may be shorter
# Gradient accumulation means each step processes batch_size * gradient_accumulation_steps samples
num_tokens = (
self.trainer_instance.batch_size *
self.trainer_instance.gradient_accumulation_steps *
self.trainer_instance.max_seq_length *
current_step
)
# Extract num_tokens from TRL SFTTrainer state (real counter)
# Requires include_num_input_tokens_seen=True in SFTConfig
num_tokens = getattr(state, "num_input_tokens_seen", None)
self.trainer_instance._update_progress(
step=current_step,