diff --git a/deepinfer/neuron_ops/nki_attention.py b/deepinfer/neuron_ops/nki_attention.py
index 98221af..1467008 100644
--- a/deepinfer/neuron_ops/nki_attention.py
+++ b/deepinfer/neuron_ops/nki_attention.py
@@ -255,6 +255,8 @@ def flash_mla_decode(
         if mask is not None:
             invalid = invalid | mask
         scopes.append((cache, torch.where(invalid, torch.zeros_like(idx), idx), invalid))
+    # Prompt chunks only: DSpark verification steps (6 rows) on the table path lost acceptance (3.85 -> 1.27 tokens
+    # a step) and ran slower (bench21), so they stay on the per-row path.
     if rows == PREFILL_BUCKET and sum(c.shape[0] * c.shape[1] for c, _, _ in scopes) <= _TABLE_MAX_SLOTS:
         out = _table_attention(q.reshape(rows, h_q, d), scopes, softmax_scale, attn_sink)
     elif rows == PREFILL_BUCKET and any(idx.shape[1] >= _SHARED_SCOPE_MIN for _, idx, _ in scopes):
diff --git a/deepinfer/neuron_ops/regions.py b/deepinfer/neuron_ops/regions.py
index d71e424..72c226c 100644
--- a/deepinfer/neuron_ops/regions.py
+++ b/deepinfer/neuron_ops/regions.py
@@ -119,6 +119,9 @@ def _boundary() -> None:
 
 #: Rows of the step being run (set by NeuronGraphRunner.replay): a prefill chunk has more than one.
 _STEP_ROWS = [1]
+#: Steps up to this many rows keep whole-layer graphs (DEEPINFER_NEURON_SPLIT_MIN_ROWS). Whole 6-row layers (DSpark
+#: verification) crash neuronx-cc on the full model, so they are split like prompt chunks by default.
+_SPLIT_MIN_ROWS = int(os.environ.get("DEEPINFER_NEURON_SPLIT_MIN_ROWS", "1"))
 
 
 def set_step_rows(rows: int) -> None:
@@ -131,5 +134,5 @@ def split() -> None:
     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):
+    if envs.DEEPINFER_NEURON_REGION_SPLIT or (envs.DEEPINFER_NEURON_REGION_SPLIT_PREFILL and _STEP_ROWS[0] > _SPLIT_MIN_ROWS):
         _boundary()
diff --git a/examples/neuron/bench_neuron.py b/examples/neuron/bench_neuron.py
index 1dff11a..8a750c8 100644
--- a/examples/neuron/bench_neuron.py
+++ b/examples/neuron/bench_neuron.py
@@ -82,7 +82,7 @@ def _run_one(llm, prompt: list[int] | None, params) -> tuple[float, float, int]
         unfinished = engine.has_unfinished_requests()
         outputs = engine.step() if unfinished else engine.flush_inflight_reqs()
         now = time.perf_counter()
-        steps += bool(outputs)
+        steps += len(outputs) > 0
         for out in outputs:
             tokens += 1
             first = first if first is not None else now