fix: keep mldev candidate parts separate

google/genai/models.py

@@ -189,6 +189,31 @@ def _Candidate_from_mldev(
   return to_object
 
 
+def _dedupe_mldev_candidate_parts(candidates: list[dict[str, Any]]) -> None:
+  if len(candidates) < 2:
+    return
+
+  first_parts = getv(candidates[0], ['content', 'parts'])
+  if not isinstance(first_parts, list):
+    return
+
+  duplicate_tail = []
+  for candidate in candidates[1:]:
+    parts = getv(candidate, ['content', 'parts'])
+    if not isinstance(parts, list) or not parts:
+      return
+    duplicate_tail.extend(parts)
+
+  if (
+      duplicate_tail
+      and len(first_parts) > len(duplicate_tail)
+      and first_parts[-len(duplicate_tail) :] == duplicate_tail
+  ):
+    content = getv(candidates[0], ['content'])
+    if isinstance(content, dict):
+      content['parts'] = first_parts[: -len(duplicate_tail)]
+
+
 def _CitationMetadata_from_mldev(
     from_object: Union[dict[str, Any], object],
     parent_object: Optional[dict[str, Any]] = None,
@@ -1697,13 +1722,15 @@ def _GenerateContentResponse_from_mldev(
     )
 
   if getv(from_object, ['candidates']) is not None:
+    candidates = [
+        _Candidate_from_mldev(item, to_object, root_object)
+        for item in getv(from_object, ['candidates'])
+    ]
+    _dedupe_mldev_candidate_parts(candidates)
     setv(
         to_object,
         ['candidates'],
-        [
-            _Candidate_from_mldev(item, to_object, root_object)
-            for item in getv(from_object, ['candidates'])
-        ],
+        candidates,
     )
 
   if getv(from_object, ['modelVersion']) is not None:

google/genai/tests/types/test_part_type.py

@@ -15,6 +15,7 @@
 
 
 import pytest
+from ... import models
 from ... import types
 
 
@@ -83,6 +84,40 @@ def test_two_candidates_text(caplog, generate_content_response):
   generate_content_response.candidates = None
 
 
+def test_mldev_candidate_count_keeps_candidate_parts_separate():
+  response_dict = models._GenerateContentResponse_from_mldev({
+      'candidates': [
+          {
+              'content': {
+                  'parts': [
+                      {'text': 'first'},
+                      {'text': 'second'},
+                      {'text': 'third'},
+                  ]
+              }
+          },
+          {'content': {'parts': [{'text': 'second'}]}},
+          {'content': {'parts': [{'text': 'third'}]}},
+      ]
+  })
+
+  response = types.GenerateContentResponse._from_response(
+      response=response_dict,
+      kwargs={},
+  )
+
+  assert [part.text for part in response.candidates[0].content.parts] == [
+      'first'
+  ]
+  assert [part.text for part in response.candidates[1].content.parts] == [
+      'second'
+  ]
+  assert [part.text for part in response.candidates[2].content.parts] == [
+      'third'
+  ]
+  assert response.text == 'first'
+
+
 def test_thought_signature_no_warning(caplog, generate_content_response):
   generate_content_response.candidates = [
       types.Candidate(