[Bugfix][Frontend] Keep the trailing parameter when qwen3 closes the function without it

[he-yufeng]

Purpose

Fixes #57699. _qwen3_arg_converter ended a parameter value only at </parameter> or the next <parameter=. Models sometimes close </function> without closing the last parameter, and then:

• non-streaming: the final conversion (partial=False) dropped the trailing value — arguments came back without it (with several parameters, the call silently lost the last one);
• streaming: partial conversions had already streamed the value, but the flush at TOOL_CALL_END produced {} — shorter than what was already sent, so nothing more was emitted and the client kept the clipped version.

Not a duplicate: the two earlier qwen3 parser fixes I checked (#43243's JSON-first parsing, and the null-handling change) cover different shapes; nothing open touches this one.

The fix is to run the trailing-parameter capture in both modes instead of only in partial mode. The accumulated raw body never contains the </function> marker (the engine consumes it as a terminal event), so the trailing value runs cleanly to the end of the buffer. Empty trailing values behave exactly like explicitly-closed empty parameters.

Test Plan

python -m pytest tests/parser/engine/test_qwen3.py -q
python -m pytest tests/parser/ -q --ignore=tests/parser/cohere -p no:randomly
ruff check / ruff format --check on the two touched files

Three new tests pin the shape: the BFCL example from the issue (non-streaming), earlier params kept with an unclosed tail (non-streaming), and no truncation in the streamed flush.

Test Result

• Red-green: all three new tests fail on the unpatched converter (non-streaming returns {}/drops the tail; the streamed flush truncates), pass with the fix.
• tests/parser/engine/test_qwen3.py: 54 passed (51 existing + 3 new).
• tests/parser/: 4608 passed with -p no:randomly (the default random-order run surfaces a pre-existing ordering-sensitive flake in tests/parser/mistral/test_reasoning.py, present with and without this change; that file passes standalone in both states). tests/parser/cohere needs the optional cohere_melody package, not installed here.
• ruff check / ruff format: clean.
• No GPU model eval: parser-level change; the issue's raw model text is pinned as a unit test.

AI assistance: prepared with AI assistance (Kimi Code). I reviewed every changed line and ran the tests above.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[he-yufeng]

Rebased onto current main (54020c3) after the Streaming Parser Engine rewrite landed.

Worth being explicit about what changed around this PR: #45701 moved qwen3 parsing into the engine, but the bug this PR fixes still reproduces on main. The engine's flush converts with partial=False, and the one-shot answer comes from that conversion, so the unclosed trailing parameter is still dropped there. Measured on current main before this rebase:

extract_tool_calls('<tool_call><function=get_weather>'
  '<parameter=city>Seattle</parameter>'
  '<parameter=unit>celsius</function></tool_call>')
-> {"city": "Seattle"}            # main: trailing unit dropped
-> {"city": "Seattle", "unit": "celsius"}   # this head

(The engine never puts the marker into the accumulated raw body, so the trailing capture runs cleanly to end of buffer. The sibling tail-class PRs for minimax_m2/glm47/deepseek_v4 were covered by the rewrite and closed; qwen3's was not.)

Verification on the rebased head: regression input and the fully-closed control both convert identically (no double counting), and the engine qwen3 suite passes 54/54 including the new TestUnclosedLastParameter cases (.venv/bin/python -m pytest tests/parser/engine/test_qwen3.py -q).

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @he-yufeng.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[he-yufeng]

Rebased onto current main (8ee4069). The conflict mergify flagged no longer exists against the current tree: nothing under vllm/parser/ has moved since the flagged base, so the rebase applied cleanly with no manual resolution.

pytest tests/parser/engine/test_qwen3.py -q on the rebased head: 69 passed.