perf(vertex/gemini): eliminate O(n²) json.loads on accumulated streaming chunks (#26181)

Replace the per-chunk `accumulated_json += chunk` + full json.loads with a
list-based buffer plus a cheap brace/bracket completeness heuristic so the
buffer is only joined and parsed when the latest shard looks like a closed
envelope. The empty-string flush from __next__/__anext__ at StopIteration
still forces a final attempt so partial-completion semantics are unchanged.

Re-PR of #26187 against the new litellm_internal_staging base (the original
PR auto-closed when the daily-dated litellm_oss_staging_04_25_2026 base
branch was retired). Scope intentionally narrowed to the Gemini hot path;
the parallel fixes for Anthropic chat handler and SageMaker common_utils
will follow once this lands so each can be reviewed independently.

Signed-off-by: Tai An <antai12232931@outlook.com>

litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py

@@ -3263,7 +3263,12 @@ class ModelResponseIterator:
 
         self.streaming_response = streaming_response
         self.chunk_type: Literal["valid_json", "accumulated_json"] = "valid_json"
-        self.accumulated_json = ""
+        # Buffer SSE shards as a list to avoid O(n²) string concatenation when
+        # Gemini emits multi-MB tool-call payloads as many small chunks. The
+        # parser only joins+json.loads when the trailing brace/bracket suggests
+        # a complete envelope, instead of re-parsing the growing buffer on every
+        # chunk and freezing the event loop.
+        self.accumulated_json_chunks: list = []
         self.sent_first_chunk = False
         self.logging_obj = logging_obj
         self.response_headers = response_headers or {}
@@ -3486,16 +3491,22 @@ class ModelResponseIterator:
         chunk = litellm.CustomStreamWrapper._strip_sse_data_from_chunk(chunk) or ""
         message = chunk.replace("\n\n", "")
 
-        # Accumulate JSON data
-        self.accumulated_json += message
-
-        # Try to parse the accumulated JSON
+        self.accumulated_json_chunks.append(message)
+        # Cheap completeness heuristic: only join + json.loads when the latest
+        # shard looks like it might have closed the envelope. Avoids parsing
+        # the growing buffer on every chunk -- the empty-string flush from
+        # __next__/__anext__ at StopIteration still forces a final attempt.
+        _stripped = message.rstrip()
+        if _stripped and _stripped[-1] not in ('}', ']'):
+            return None
+        _full_json = "".join(self.accumulated_json_chunks)
+        if not _full_json:
+            return None
         try:
-            _data = json.loads(self.accumulated_json)
-            self.accumulated_json = ""  # reset after successful parsing
+            _data = json.loads(_full_json)
+            self.accumulated_json_chunks = []
             return self.chunk_parser(chunk=_data)
         except json.JSONDecodeError:
-            # If it's not valid JSON yet, continue to the next event
             return None
 
     def _common_chunk_parsing_logic(
@@ -3522,7 +3533,7 @@ class ModelResponseIterator:
         try:
             chunk = self.response_iterator.__next__()
         except StopIteration:
-            if self.chunk_type == "accumulated_json" and self.accumulated_json:
+            if self.chunk_type == "accumulated_json" and self.accumulated_json_chunks:
                 return self.handle_accumulated_json_chunk(chunk="")
             raise StopIteration
         except ValueError as e:
@@ -3544,7 +3555,7 @@ class ModelResponseIterator:
         try:
             chunk = await self.async_response_iterator.__anext__()
         except StopAsyncIteration:
-            if self.chunk_type == "accumulated_json" and self.accumulated_json:
+            if self.chunk_type == "accumulated_json" and self.accumulated_json_chunks:
                 return self.handle_accumulated_json_chunk(chunk="")
             raise StopAsyncIteration
         except ValueError as e: