diff --git a/vllm/v1/worker/gpu/cudagraph_utils.py b/vllm/v1/worker/gpu/cudagraph_utils.py
index 55462283db..0ea45ccf0e 100644
--- a/vllm/v1/worker/gpu/cudagraph_utils.py
+++ b/vllm/v1/worker/gpu/cudagraph_utils.py
@@ -1133,6 +1133,11 @@ def _teardown_profiling_state(runner: "GPUModelRunner") -> None:
     torch.accelerator.synchronize()
     if hasattr(runner.model_state, "_mamba_ctx"):
         runner.model_state._mamba_ctx = None
+    # A PIECEWISE profiling capture records its RecoverSSM step (only FULL
+    # captures skip it), and that step's commit contexts reference the
+    # profiling KV cache.
+    if (recoverssm := getattr(runner.model_state, "recoverssm", None)) is not None:
+        recoverssm.reset()
     # Invalidate the align-mode Mamba group metadata cached from the
     # profiling KVCacheConfig: the real (e.g. PP-projected) config may
     # place Mamba layers into a different group layout, so it must be
diff --git a/vllm/v1/worker/gpu/model_states/recoverssm.py b/vllm/v1/worker/gpu/model_states/recoverssm.py
index b02a6c19aa..3cfbecea26 100644
--- a/vllm/v1/worker/gpu/model_states/recoverssm.py
+++ b/vllm/v1/worker/gpu/model_states/recoverssm.py
@@ -16,6 +16,11 @@ class RecoverSSMState:
     def __init__(self) -> None:
         self._step: tuple[RecoverSSMMetadata, ...] | None = None
 
+    def reset(self) -> None:
+        """Drop the recorded step. Its metadata holds the commit contexts, which
+        reference the KV cache tensors they were built for."""
+        self._step = None
+
     def record_step(
         self,
         attn_metadata: dict[str, Any],