diff --git a/src/prime_rl/trainer/sft/train.py b/src/prime_rl/trainer/sft/train.py
index 1da7fe60af..3917f414c7 100644
--- a/src/prime_rl/trainer/sft/train.py
+++ b/src/prime_rl/trainer/sft/train.py
@@ -287,8 +287,19 @@ def run_eval_loop(data_iter):
         total_token_count = torch.tensor(0, dtype=torch.int64, device="cuda")
         nan_count = torch.tensor(0, device="cuda")
 
+        # Variable-length packing yields different per-rank batch counts. Under FSDP
+        # every forward is a collective, so all ranks must agree on when to stop —
+        # otherwise the first rank to exit deadlocks the rest in the next all-gather.
+        # Sync per batch and exit together as soon as any rank exhausts its iterator.
+        data_iter = iter(data_iter)
+
         with torch.no_grad():
-            for micro_batch in data_iter:
+            while True:
+                micro_batch = next(data_iter, None)
+                has_data = torch.tensor(micro_batch is not None, dtype=torch.int32, device="cuda")
+                dist.all_reduce(has_data, op=dist.ReduceOp.MIN)
+                if has_data.item() == 0:
+                    break
                 loss_sum, token_count = compute_loss(micro_batch)
                 if not torch.isnan(loss_sum.detach()):
                     total_loss_sum += loss_sum.detach()