diff --git a/.github/workflows/config/gitlab_commits.txt b/.github/workflows/config/gitlab_commits.txt
index 40b8604660b..d0a266f8bf5 100644
--- a/.github/workflows/config/gitlab_commits.txt
+++ b/.github/workflows/config/gitlab_commits.txt
@@ -1,2 +1,2 @@
 nvidia-mgpu-repo: cuda-quantum/cuquantum-mgpu.git
-nvidia-mgpu-commit: 17279375c13290a202ae5c5267f1f571227f1ef4
+nvidia-mgpu-commit: de32bf2abae42f118a0da36be0066296d76b0dbc
diff --git a/python/tests/visualization/test_draw.py b/python/tests/visualization/test_draw.py
index 733d97a24a6..b8c1756bbb5 100644
--- a/python/tests/visualization/test_draw.py
+++ b/python/tests/visualization/test_draw.py
@@ -105,6 +105,28 @@ def kernel():
     assert expected_str == produced_string
 
 
+# This test will run on the default simulator. For machines with GPUs, that
+# will be a GPU-accelerated simulator, but for machines without GPUs, it
+# will run on a CPU simulator.
+def test_draw_with_exp_pauli():
+    @cudaq.kernel
+    def kernel_exp_pauli():
+        q = cudaq.qvector(2)
+        exp_pauli(0.2, q, "ZZ")
+
+
+    expected_str = R"""
+                           
+q0 : ──●────────────────●──
+     ╭─┴─╮╭──────────╮╭─┴─╮
+q1 : ┤ x ├┤ rz(-0.4) ├┤ x ├
+     ╰───╯╰──────────╯╰───╯
+"""
+    expected_str = expected_str[1:]
+    produced_string = cudaq.draw(kernel_exp_pauli)
+    assert expected_str == produced_string
+
+
 def test_draw_hw_target():
 
     @cudaq.kernel
diff --git a/runtime/nvqir/CircuitSimulator.h b/runtime/nvqir/CircuitSimulator.h
index 759cc971d58..f1dd2dc7387 100644
--- a/runtime/nvqir/CircuitSimulator.h
+++ b/runtime/nvqir/CircuitSimulator.h
@@ -386,13 +386,12 @@ class CircuitSimulatorBase : public CircuitSimulator {
   /// @brief Reference to the current circuit name.
   std::string currentCircuitName = "";
 
-private:
+protected:
   /// @brief Return true if the simulator is in the tracer mode.
   bool isInTracerMode() const {
     return executionContext && executionContext->name == "tracer";
   }
 
-protected:
   /// @brief The current Execution Context (typically this is null,
   /// sampling, or spin_op observation.
   cudaq::ExecutionContext *executionContext = nullptr;
diff --git a/runtime/nvqir/custatevec/CuStateVecCircuitSimulator.cpp b/runtime/nvqir/custatevec/CuStateVecCircuitSimulator.cpp
index e5ee20cd184..580c7d602ed 100644
--- a/runtime/nvqir/custatevec/CuStateVecCircuitSimulator.cpp
+++ b/runtime/nvqir/custatevec/CuStateVecCircuitSimulator.cpp
@@ -456,6 +456,10 @@ class CuStateVecCircuitSimulator
   void applyExpPauli(double theta, const std::vector<std::size_t> &controlIds,
                      const std::vector<std::size_t> &qubits,
                      const cudaq::spin_op &op) override {
+    if (this->isInTracerMode()) {
+      nvqir::CircuitSimulator::applyExpPauli(theta, controlIds, qubits, op);
+      return;
+    }
     flushGateQueue();
     cudaq::info(" [cusv decomposing] exp_pauli({}, {})", theta,
                 op.to_string(false));
diff --git a/runtime/nvqir/cutensornet/simulator_mps_register.cpp b/runtime/nvqir/cutensornet/simulator_mps_register.cpp
index 57b7f090bef..e9246a49949 100644
--- a/runtime/nvqir/cutensornet/simulator_mps_register.cpp
+++ b/runtime/nvqir/cutensornet/simulator_mps_register.cpp
@@ -112,6 +112,10 @@ class SimulatorMPS : public SimulatorTensorNetBase {
                              const std::vector<std::size_t> &controls,
                              const std::vector<std::size_t> &qubitIds,
                              const cudaq::spin_op &op) override {
+    if (this->isInTracerMode()) {
+      nvqir::CircuitSimulator::applyExpPauli(theta, controls, qubitIds, op);
+      return;
+    }
     // Special handling for equivalence of Rxx(theta), Ryy(theta), Rzz(theta)
     // expressed as exp_pauli.
     //  Note: for MPS, the runtime is ~ linear with the number of 2-body gates