diff --git a/deepinfer/environ.py b/deepinfer/environ.py
index ccd6bb5..6df4fd9 100644
--- a/deepinfer/environ.py
+++ b/deepinfer/environ.py
@@ -113,7 +113,9 @@ class envs:
     # graphs (neuron_ops/regions.py, the vllm-neuron#71 layout); "model" compiles the whole step as one graph.
     DEEPINFER_NEURON_GRAPH_MODE = EnvVar.str(default="regions")
     # Neuron only: cut the region graphs further at the stage boundaries marked by neuron_ops.regions.split().
-    DEEPINFER_NEURON_REGION_SPLIT = EnvVar.bool(default=True)
+    DEEPINFER_NEURON_REGION_SPLIT = EnvVar.bool(default=False)
+    # Neuron only: split the region graphs of prefill chunks (steps of more than one row) regardless.
+    DEEPINFER_NEURON_REGION_SPLIT_PREFILL = EnvVar.bool(default=True)
     # Neuron only: synchronize around every region graph call and log per-kind times each decode step (profiling).
     DEEPINFER_NEURON_REGION_TIMING = EnvVar.bool()
     # Keep the token embedding (and the text model's unused vision tower) in host memory on Neuron.
diff --git a/deepinfer/neuron_ops/regions.py b/deepinfer/neuron_ops/regions.py
index cd5bc6d..d71e424 100644
--- a/deepinfer/neuron_ops/regions.py
+++ b/deepinfer/neuron_ops/regions.py
@@ -117,7 +117,19 @@ def _boundary() -> None:
         _MARK[0], _MARK[1] = now, _MARK[1] + 1
 
 
+#: Rows of the step being run (set by NeuronGraphRunner.replay): a prefill chunk has more than one.
+_STEP_ROWS = [1]
+
+
+def set_step_rows(rows: int) -> None:
+    _STEP_ROWS[0] = rows
+
+
 def split() -> None:
-    """Cut the current region graph here (only while compiling; free in eager)."""
-    if torch.compiler.is_compiling() and envs.DEEPINFER_NEURON_REGION_SPLIT:
+    """Cut the current region graph here (only while compiling; free in eager). Always for prefill chunks under
+    DEEPINFER_NEURON_REGION_SPLIT_PREFILL: neuronx-cc runs the split 128-row attention ~2x faster (62 vs 122 ms a
+    layer, issue #26), while decode's 1-row graphs are faster whole."""
+    if not torch.compiler.is_compiling():
+        return
+    if envs.DEEPINFER_NEURON_REGION_SPLIT or (envs.DEEPINFER_NEURON_REGION_SPLIT_PREFILL and _STEP_ROWS[0] > 1):
         _boundary()
diff --git a/deepinfer/worker/neuron_graph_runner.py b/deepinfer/worker/neuron_graph_runner.py
index 3ca70a0..702abbd 100644
--- a/deepinfer/worker/neuron_graph_runner.py
+++ b/deepinfer/worker/neuron_graph_runner.py
@@ -170,6 +170,7 @@ class NeuronGraphRunner(CudaGraphRunner):
             else:
                 compiled = torch.compile(forward, backend=backend, fullgraph=fullgraph, dynamic=False)
             # First call compiles (and runs) this batch size's graph.
+            regions.set_step_rows(batch_size)
             input_metadata_device.recursive_copy_(input_metadata_source)
             with hide_cuda():
                 compiled(input_metadata_device)
@@ -187,6 +188,7 @@ class NeuronGraphRunner(CudaGraphRunner):
         if not self.input_on_device:
             input_metadata_device.recursive_copy_(cuda_graph_data.input_metadata_source)
         batch_size = next(bs for bs, data in self.graphs.items() if data is cuda_graph_data)
+        regions.set_step_rows(batch_size)
         if not envs.DEEPINFER_NEURON_REGION_TIMING:
             with hide_cuda():
                 return self.compiled[batch_size](input_metadata_device)