diff --git a/src/prime_rl/trainer/rl/train.py b/src/prime_rl/trainer/rl/train.py
index 568ca6e1d8..fe63de8d47 100644
--- a/src/prime_rl/trainer/rl/train.py
+++ b/src/prime_rl/trainer/rl/train.py
@@ -604,7 +604,12 @@ def train(config: TrainerConfig):
                 broadcast_weights_start_time = time.perf_counter()
                 # The per-layer gather + fp8 conversion peaks ~50 GiB above the
                 # resident weights; release cached blocks (incl. offload-stream
-                # pools) so the broadcast gets the full headroom.
+                # pools) so the broadcast gets the full headroom. Drain all
+                # pending work first: empty_cache returns blocks to the driver,
+                # so a still-running kernel holding a cached block (e.g. the
+                # optimizer step's tail) faults with an illegal memory access
+                # once its block is freed under it.
+                torch.cuda.synchronize()
                 torch.cuda.empty_cache()
                 weight_sender.broadcast(model, step=progress.step)
                 broadcast_weights_time = time.perf_counter() - broadcast_weights_start_time