[Refactor] Move unstreamed tool-arg flush from serving layer to parser (#44017)

Signed-off-by: sfeng33 <4florafeng@gmail.com>

tests/parser/test_streaming.py

@@ -235,3 +235,86 @@ def test_parse_delta_reasoning_only_thinking_disabled(tokenizer, request_obj):
     assert "Hello" in content
     assert "assist" in content
     assert len(tool_calls) == 0
+
+
+def test_parse_delta_finished_no_flush_without_tool_call_delta(tokenizer, request_obj):
+    """When finished=True but the final parse_delta produces no
+    tool-call delta, unstreamed args are not flushed."""
+    parser = make_parser(tokenizer, reasoning=False, tool=True)
+
+    results = stream_text(
+        parser, tokenizer, MODEL_OUTPUT, request_obj, prompt_token_ids=[]
+    )
+    _, _, tool_calls = collect_fields(results)
+    assert len(tool_calls) > 0
+
+    streamed = parser._tool_parser.streamed_args_for_tool[0]
+    assert len(streamed) > 5
+    parser._tool_parser.streamed_args_for_tool[0] = streamed[:-5]
+
+    # Prevent normal extraction from catching the gap — without a
+    # tool-call delta to merge into, the flush is skipped.
+    parser._tool_parser.extract_tool_calls_streaming = lambda *a, **kw: None
+
+    flush_result = parser.parse_delta("", [], request_obj, finished=True)
+    assert flush_result is None or flush_result.tool_calls is None
+
+
+def test_parse_delta_finished_no_extra_args_when_fully_streamed(tokenizer, request_obj):
+    """When all args have been streamed, finished=True must not
+    produce extra or duplicate arguments."""
+    parser = make_parser(tokenizer, reasoning=False, tool=True)
+    results = stream_text(
+        parser, tokenizer, MODEL_OUTPUT, request_obj, prompt_token_ids=[]
+    )
+    _, _, tool_calls = collect_fields(results)
+
+    assert len(tool_calls) > 0
+    assert tool_calls[0].function.name == "get_weather"
+    tool_args = "".join(
+        tc.function.arguments for tc in tool_calls if tc.function.arguments
+    )
+    assert json.loads(tool_args) == {"city": "Dallas"}
+
+    flush_result = parser.parse_delta("", [], request_obj, finished=True)
+    assert flush_result is None or flush_result.tool_calls is None
+
+
+def test_parse_delta_finished_appends_remaining_args(tokenizer, request_obj):
+    """When finished=True and the tool parser has unstreamed args,
+    parse_delta appends the remaining arguments to the tool-call delta."""
+    parser = make_parser(tokenizer, reasoning=False, tool=True)
+    token_ids = tokenizer.encode(MODEL_OUTPUT, add_special_tokens=False)
+
+    remainder = ',"unit":"celsius"}'
+    prompt_ids: list[int] | None = []
+    results: list[DeltaMessage | None] = []
+    for i, tid in enumerate(token_ids):
+        prev = results[-1] if results else None
+        prev_had_args = (
+            prev
+            and prev.tool_calls
+            and any(tc.function and tc.function.arguments for tc in prev.tool_calls)
+        )
+
+        if prev_had_args:
+            parser._tool_parser.get_remaining_unstreamed_args = lambda: remainder
+
+        result = parser.parse_delta(
+            tokenizer.decode([tid]),
+            [tid],
+            request_obj,
+            prompt_token_ids=prompt_ids,
+            finished=prev_had_args,
+        )
+        prompt_ids = None
+        results.append(result)
+
+        if prev_had_args:
+            break
+
+    _, _, tool_calls = collect_fields(results)
+    tool_args = "".join(
+        tc.function.arguments for tc in tool_calls if tc.function.arguments
+    )
+    assert tool_args.endswith(remainder)