fix: Evaluations no longer associate incorrect inputs with predictions

tests/trace/test_evaluations.py

@@ -148,6 +148,19 @@ def is_op_ref_with_name(val: Any, name: str):
     expected_predict_ref = model_obj.val["predict"]
     assert is_op_ref_with_name(expected_predict_ref, "MyModel.predict")
 
+    predict_and_score_calls = [
+        c
+        for (c, d) in flattened_calls
+        if op_name_from_ref(c.op_name) == "Evaluation.predict_and_score"
+    ]
+    assert len(predict_and_score_calls) == 3
+
+    # Assert that all the inputs are unique
+    inputs = set()
+    for call in predict_and_score_calls:
+        inputs.add(call.inputs["example"])
+    assert len(inputs) == 3
+
 
 @weave.op
 def gpt_mocker(question: str):

weave/trace/vals.py

@@ -352,7 +352,7 @@ def _local_iter_with_remote_fallback(self) -> Generator[dict, None, None]:
 
         for ndx, row in enumerate(self._prefetched_rows):
             next_id_future = wc.future_executor.defer(
-                lambda: cached_table_ref.row_digests[ndx]
+                lambda ndx_closure=ndx: cached_table_ref.row_digests[ndx_closure]
             )
             new_ref = self.ref.with_item(next_id_future)
             val = self._prefetched_rows[ndx]