[Refactor] Move unstreamed tool-arg flush from serving layer to parser

[sfeng33]

Purpose

• Relocate the "flush unstreamed tool-call arguments" logic from OpenAIServingChat into DelegatingParser._append_unstreamed_tool_args and ToolParser.get_remaining_unstreamed_args, where the streaming state naturally lives
• Pass finished=True from the serving layer when output.finish_reason is set, letting the parser emit any remaining arguments in its final parse_delta call
• Simplify the finish_reason condition by removing the auto_tools_called variable

Test Plan

E2E test

vllm serve meta-llama/Llama-3.2-3B-Instruct \
    --enable-auto-tool-choice \
    --tool-call-parser llama3_json 

curl -sN http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "meta-llama/Llama-3.2-3B-Instruct",
    "stream": true, 
    "messages": [{"role":"user","content":"What is the weather in Tokyo in celsius? Use the tool."}],
    "tools": [{"type":"function","function":{
      "name":"get_weather",
      "description":"Get current weather for a city",
      "parameters":{"type":"object",
        "properties":{"city":{"type":"string"},
                      "unit":{"type":"string","enum":["celsius","fahrenheit"]}},
        "required":["city","unit"]}}}]
    }'

Unit test

pytest tests/entrypoints/openai/chat_completion/test_serving_chat.py -v
pytest tests/parser/test_streaming.py -v

[hclsys]

moving the flush into the parser where the streaming state lives is the right call, and the merge-into-last_tc branch is clean.

but the standalone branch worries me: when there's no concurrent tool-call delta to append onto, _append_unstreamed_tool_args emits DeltaToolCall(index=index, function=DeltaFunctionCall(arguments=remaining)) — no id, no type, no function.name. that's exactly what the deleted _create_remaining_args_delta existed to preserve ("id/type/name ... which would otherwise be lost"), and TestCreateRemainingArgsDelta is gone with it. so a finish chunk that flushes remaining args on its own now drops id/type/name — isn't that the regression the old helper guarded? if the merge branch always fires in practice it's moot, but worth a test that asserts id/type/name survive a standalone final flush.

[sfeng33]

but the standalone branch worries me: when there's no concurrent tool-call delta to append onto, _append_unstreamed_tool_args emits DeltaToolCall(index=index, function=DeltaFunctionCall(arguments=remaining)) — no id, no type, no function.name. that's exactly what the deleted _create_remaining_args_delta existed to preserve ("id/type/name ... which would otherwise be lost"), and TestCreateRemainingArgsDelta is gone with it. so a finish chunk that flushes remaining args on its own now drops id/type/name — isn't that the regression the old helper guarded? if the merge branch always fires in practice it's moot, but worth a test that asserts id/type/name survive a standalone final flush.

Thanks for the review, your concern about the standalone branch dropping id/type/name makes sense.

I think the part is actually unnecessary though. If there's no existing tool-call delta to merge into, we shouldn't be emitting a bare one without id/type/name — better to just skip the flush entirely, which matches the old serving-layer behavior where the flush only happened when there was already a delta to attach to.

I've simplified _append_unstreamed_tool_args to remove the extra logic: it now only appends when delta_message already carries a tool-call delta with a .function, so id/type/name are naturally preserved.

[yzong-rh]

One small question, LGTM otherwise

[AndreasKaratzas]

Tests look good, but the rest of the changes probably someone else should review.

[hclsys]

yeah merge-only is cleaner — last_tc already carries id/type/name, so appending args to it just works and there's no bare-delta path that could drop them. lgtm.

[chaunceyjiang]

LGTM