[Bugfix] Fix GLM tool-call finish chunk suffix alignment in streaming

[QwertyJack]

Summary

Fix the streaming finish-chunk logic for GLM tool calls when the parser's streamed JSON prefix does not exactly match json.dumps(...) formatting.

Current OpenAIServingChat computes the final tool-argument delta by serializing the parser state and calling replace() against the streamed prefix. That is brittle for GLM parsers because the streamed prefix is not always serialized in a single stable style:

• some runs are compact, e.g. {"a":1}
• some are default-spaced, e.g. {"a": 1}
• some GLM outputs are mixed, e.g. the outer object is compact but nested objects keep default whitespace

When replace() cannot align the prefix, the finish chunk either re-emits the full object (duplicated-value malformed JSON) or fails to backfill the missing suffix.

This PR fixes the suffix computation centrally in OpenAIServingChat by:

1. trying exact prefix matches against string-backed, default json.dumps, and compact json.dumps(..., separators=(",", ":")) candidates;
2. falling back to whitespace-minified prefix comparison for mixed-whitespace GLM fragments;
3. returning an empty suffix if nothing aligns, instead of re-emitting the whole object.

This keeps the existing GLM parser contract unchanged and fixes the finish-chunk backfill in the serving layer rather than mutating parser state.

Fixes #36857

Tests

Commands run:

python -m py_compile \
  vllm/entrypoints/openai/chat_completion/serving.py \
  tests/entrypoints/openai/chat_completion/test_serving_chat.py

PYTHONPATH=/tmp/vllm_pydeps:/tmp:$PYTHONPATH \
pytest -p skip_cleanup_plugin \
  tests/entrypoints/openai/chat_completion/test_serving_chat.py \
  -k "TestCreateRemainingArgsDelta or TestComputeRemainingToolArgs" -q

PYTHONPATH=/tmp:$PYTHONPATH \
pytest -p skip_cleanup_plugin \
  tests/tool_parsers/test_glm4_moe_tool_parser.py \
  -k "streaming_prev_tool_call_arr_updates or streaming_multiple_tool_calls_sequential" -q

Results:

• py_compile: passed
• serving tests: 9 passed, 31 deselected
• GLM parser contract tests: 2 passed, 20 deselected

[gemini-code-assist]

Code Review

This pull request addresses a bug in the streaming finish-chunk logic for GLM tool calls where JSON serialization format inconsistencies could lead to malformed JSON. The fix is implemented in OpenAIServingChat by introducing a more robust method, _compute_remaining_tool_args, to calculate the remaining tool arguments suffix. This new method correctly handles various JSON formatting styles, including compact, default, and mixed-whitespace, by trying multiple parsing strategies. The logic is well-encapsulated and supported by a comprehensive new set of unit tests. The changes appear correct and effectively resolve the issue without altering existing parser contracts.

[QwertyJack]

This PR is ready for review.

CI reported that it needs the ready label (or an author with at least 4 merged PRs) before the full workflow can proceed. I don't have permission to add labels on vllm-project/vllm from this fork.

Could a maintainer please add the ready label to trigger the gated CI path?