diff --git a/python/sglang/srt/model_executor/piecewise_cuda_graph_runner.py b/python/sglang/srt/model_executor/piecewise_cuda_graph_runner.py
index c567a6d6f..6f8a102fa 100644
--- a/python/sglang/srt/model_executor/piecewise_cuda_graph_runner.py
+++ b/python/sglang/srt/model_executor/piecewise_cuda_graph_runner.py
@@ -217,8 +217,12 @@ class PiecewiseCudaGraphRunner:
         self.capture_forward_mode = ForwardMode.EXTEND
         self.capture_hidden_mode = CaptureHiddenMode.NULL
 
-        # If returning hidden states is enabled, set initial capture hidden mode to full to avoid double-capture on startup
-        if model_runner.server_args.enable_return_hidden_states:
+        # If returning hidden states is enabled, or if speculative prefill needs
+        # aux hidden states (DFLASH), capture the FULL variant up front.
+        if (
+            model_runner.server_args.enable_return_hidden_states
+            or model_runner.spec_algorithm.is_dflash()
+        ):
             self.capture_hidden_mode = CaptureHiddenMode.FULL
 
         self.max_num_tokens = (
@@ -416,7 +420,7 @@ class PiecewiseCudaGraphRunner:
                 mrope_positions=mrope_positions,
                 spec_algorithm=None,
                 spec_info=None,
-                capture_hidden_mode=CaptureHiddenMode.NULL,
+                capture_hidden_mode=self.capture_hidden_mode,
                 num_token_non_padded=None,
                 num_token_non_padded_cpu=num_tokens,
                 global_forward_mode=ForwardMode.EXTEND,
@@ -587,7 +591,7 @@ class PiecewiseCudaGraphRunner:
                 mrope_positions=mrope_positions,
                 spec_algorithm=None,
                 spec_info=None,
-                capture_hidden_mode=CaptureHiddenMode.NULL,
+                capture_hidden_mode=self.capture_hidden_mode,
                 num_token_non_padded=None,
                 num_token_non_padded_cpu=num_tokens,
                 global_forward_mode=ForwardMode.EXTEND,
diff --git a/test/registered/piecewise_cuda_graph/test_pcg_with_speculative_decoding_dflash.py b/test/registered/piecewise_cuda_graph/test_pcg_with_speculative_decoding_dflash.py
new file mode 100644
index 000000000..3095aacca
--- /dev/null
+++ b/test/registered/piecewise_cuda_graph/test_pcg_with_speculative_decoding_dflash.py
@@ -0,0 +1,47 @@
+"""Test piecewise CUDA graph coexisting with speculative decoding (DFLASH).
+
+PCG handles prefill/extend path while DFlash needs target aux hidden states
+from prefill to materialize draft KV cache. This verifies PCG captures that
+path with the DFlash hidden-state variant enabled.
+"""
+
+import unittest
+
+from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.server_fixtures.pcg_spec_fixture import PCGSpecBase
+from sglang.test.test_utils import (
+    DEFAULT_DRAFT_MODEL_DFLASH,
+    DEFAULT_TARGET_MODEL_DFLASH,
+    CustomTestCase,
+)
+
+register_cuda_ci(est_time=531, stage="base-b", runner_config="1-gpu-small")
+
+
+class TestPCGWithDFlash(PCGSpecBase, CustomTestCase):
+    """PCG + DFLASH on Llama-3.1-8B-Instruct."""
+
+    model = DEFAULT_TARGET_MODEL_DFLASH
+    server_args = [
+        "--trust-remote-code",
+        "--attention-backend",
+        "flashinfer",
+        "--enforce-piecewise-cuda-graph",
+        "--speculative-algorithm",
+        "DFLASH",
+        "--speculative-draft-model-path",
+        DEFAULT_DRAFT_MODEL_DFLASH,
+        "--page-size",
+        "1",
+        "--max-running-requests",
+        "64",
+        "--cuda-graph-bs",
+        *[str(i) for i in range(1, 65)],
+    ]
+    server_env = {"SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN": "1"}
+    accuracy_threshold = 0.75
+    speedup_threshold = 2.8
+
+
+if __name__ == "__main__":
+    unittest.main()