diff --git a/python/sglang/srt/model_executor/runner/prefill_cuda_graph_runner.py b/python/sglang/srt/model_executor/runner/prefill_cuda_graph_runner.py
index d87e54353..96b211ad0 100644
--- a/python/sglang/srt/model_executor/runner/prefill_cuda_graph_runner.py
+++ b/python/sglang/srt/model_executor/runner/prefill_cuda_graph_runner.py
@@ -1608,6 +1608,13 @@ class PrefillCudaGraphRunner(BaseCudaGraphRunner):
                 s["extend_start_loc"][bs:r].fill_(self.raw_num_tokens)
                 s["req_pool_indices"][bs:r].zero_()
                 s["orig_seq_lens"][bs:r].zero_()
+                # The captured track scatter reads these rows too, and a stale
+                # mask row still carries a live destination slot: it would land
+                # this replay's window in an earlier request's checkpoint.
+                registry = self.buffer_registry
+                for name in ("mamba_track_mask", "mamba_track_indices"):
+                    if registry.has_slot(name):
+                        registry.get_slot(name).buffer[bs:r].zero_()
 
         # Refresh the static buffer the captured graph reads from.
         if (
diff --git a/test/registered/models_e2e/test_inkling_small_nvfp4.py b/test/registered/models_e2e/test_inkling_small_nvfp4.py
index e11f0f550..40998b8a6 100644
--- a/test/registered/models_e2e/test_inkling_small_nvfp4.py
+++ b/test/registered/models_e2e/test_inkling_small_nvfp4.py
@@ -163,7 +163,6 @@ class TestInklingSmallNvfp4Deterministic(CustomTestCase):
                 "--mem-fraction-static",
                 "0.85",
                 "--enable-deterministic-inference",
-                "--disable-prefill-cuda-graph",
             ],
             env={**os.environ, "SGLANG_ENABLE_UNIFIED_RADIX_TREE": "1"},
         )