p

training/train.py

@@ -35,6 +35,23 @@ torch.backends.cudnn.allow_tf32       = True
 
 import transformers
 from transformers import TrainingArguments, Trainer, TrainerCallback, PrinterCallback
+from transformers.trainer_callback import ProgressCallback
+
+
+class _NoEvalTqdmCallback(ProgressCallback):
+    """Same as HF's ProgressCallback but with the per-batch eval bar disabled.
+
+    In a Colab `!python -m ...` subprocess HF Trainer's `is_in_notebook()`
+    returns False (no IPython kernel in the child) so it falls back to plain
+    tqdm. Colab's text renderer mishandles `\\r` for fast updates, so the
+    eval bar (~1 batch/sec × 1250 batches) prints a fresh line every step
+    and lags the browser tab. Training tqdm updates slowly enough (one bar
+    line per ~9s at 24M params + LoRA + bf16) that it stays clean, so we
+    only kill the prediction bar. eval_loss is still logged at the end of
+    each eval pass via the standard log_history mechanism."""
+
+    def on_prediction_step(self, args, state, control, **kwargs):  # noqa: D401
+        return
 
 # Add project root to path
 sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
@@ -377,6 +394,10 @@ def get_trainer(
         data_collator   = collator,
     )
     trainer.remove_callback(PrinterCallback)
+    # Replace default ProgressCallback with one that skips the eval per-batch
+    # bar — see _NoEvalTqdmCallback docstring for the Colab-subprocess rationale.
+    trainer.remove_callback(ProgressCallback)
+    trainer.add_callback(_NoEvalTqdmCallback())
     return trainer