[bugfix] Probabilities do not sum to one with Torch

doc/releases/changelog-dev.md

@@ -297,6 +297,9 @@
 
 <h3>Bug fixes 🐛</h3>
 
+* The probabilities now sum to one using the `torch` interface with `default_dtype` set to `torch.float32`. 
+  [(#5462)](https://github.com/PennyLaneAI/pennylane/pull/5462)
+
 * Avoid bounded value failures due to numerical noise with calls to `np.random.binomial`.
   [(#5447)](https://github.com/PennyLaneAI/pennylane/pull/5447)
 

pennylane/devices/qubit/sampling.py

@@ -448,10 +448,25 @@ def sample_state(
     with qml.queuing.QueuingManager.stop_recording():
         probs = qml.probs(wires=wires_to_sample).process_state(flat_state, state_wires)
 
+    # when using the torch interface with float32 as default dtype,
+    # probabilities must be renormalized as they may not sum to one
+    # see https://github.com/PennyLaneAI/pennylane/issues/5444
+    norm = qml.math.sum(probs, axis=-1)
+    abs_diff = np.abs(norm - 1.0)
+    cutoff = 1e-07
+
     if is_state_batched:
+
+        if np.any(abs_diff > 0) and np.all(abs_diff < cutoff):
+            probs = probs / norm[:, np.newaxis] if norm.shape else probs / norm
+
         # rng.choice doesn't support broadcasting
         samples = np.stack([rng.choice(basis_states, shots, p=p) for p in probs])
     else:
+
+        if 0 < abs_diff < cutoff:
+            probs /= norm
+
         samples = rng.choice(basis_states, shots, p=probs)
 
     powers_of_two = 1 << np.arange(num_wires, dtype=np.int64)[::-1]