diff --git a/mstar/model/qwen3_tts/submodules.py b/mstar/model/qwen3_tts/submodules.py
index ad3b99bbc..2dde9468e 100644
--- a/mstar/model/qwen3_tts/submodules.py
+++ b/mstar/model/qwen3_tts/submodules.py
@@ -708,7 +708,12 @@ def get_cuda_graph_configs(
                 },
             ),
             capture_batch_sizes=self.DECODE_CAPTURE_BATCH_SIZES,
-            compile=True,
+            # Capture the eager kernels. A torch.compile'd capture picks its
+            # kernels by benchmarking in each process, and sixty sampling
+            # steps turn the resulting rounding differences into a different
+            # token sequence per server instance for the same seed. The eager
+            # kernels are the same everywhere, and the capture is faster.
+            compile=False,
         )]
 
     def get_piecewise_cuda_graph_configs(
diff --git a/test/modular/test_qwen3_tts_model.py b/test/modular/test_qwen3_tts_model.py
index eed6b957d..dfb87ead5 100644
--- a/test/modular/test_qwen3_tts_model.py
+++ b/test/modular/test_qwen3_tts_model.py
@@ -702,6 +702,10 @@ def test_qwen3_tts_talker_batches_and_captures_decode():
     graph_config = submodule.get_cuda_graph_configs(torch.device("cpu"))[0]
     assert graph_config.capture_graph_walk == "talker_decode"
     assert graph_config.capture_batch_sizes == [1, 2, 4, 8, 16, 32]
+    # The decode graph captures the eager kernels: a compiled capture picks
+    # kernels by benchmarking per process, and the same seed then gives a
+    # different token sequence on every server instance.
+    assert graph_config.compile is False
     assert graph_config.single_request_inputs.tensor_inputs[
         "suppress_eos"
     ].item() is True