add eval_accumulation_steps

train_od.py

@@ -192,6 +192,8 @@ def build_parser():
     p.add_argument("--disable-augmentations", action="store_true")
     p.add_argument("--eval-max-batches", type=int, default=None,
                    help="cap eval batches for profiling")
+    p.add_argument("--eval-accumulation-steps", type=int, default=4,
+                   help="offload eval predictions to CPU every N batches (avoids GPU OOM)")
     p.add_argument("--mini-train", action="store_true",
                    help="after profiling, also run a short real training+eval phase")
     return p
@@ -343,6 +345,7 @@ def main():
         remove_unused_columns=False,
         eval_strategy="steps",
         eval_steps=args.eval_steps,
+        eval_accumulation_steps=args.eval_accumulation_steps,
         logging_steps=20,
         save_strategy="steps",
         save_steps=args.save_steps,