[Router] Standardize route_request signature across all routers (rebase of #1023)

[chrikrah]

Mohith26's #1023 on main 7b1164d: every router's route_request becomes async with request_json: Optional[Dict] = None, and the three dispatch sites in request.py stop type-sniffing. Fixes #1022.

Since the original base, PriorityRouter and LoadAwareRouter arrived and get the same signature here, and the three-way isinstance chain #1021 put on the multipart call collapses to one awaited call passing {}. The DisaggregatedPrefill* handler dispatch stays, since those routers use different handlers.

Cost: an out-of-tree router with a sync route_request breaks on the first request, with TypeError: object str can't be used in 'await' expression. The fake router in test_request_stats.py from #1115 hit that and is now async. Once this lands I would close my #1095, whose isinstance addition becomes dead code. #929 by nejch and #950 by guba-elleschr make the same async change, both before #1023, and #1020 patches the multipart call alone. All three are open, and this series covers each.

On #1095 I said I would keep Mohith Gajjela as author, but the new routers and the #1021 chain meant rewriting his commits, so he is credited in a Co-authored-by: trailer on both commits. The second extends his signature tests to the two new routers and checks that request_json defaults to None.

# tree: git archive 7b1164d, vllm_router.__file__ printed from the exported tree's src/; Python 3.12, ~/.venvs/production-stack
$ PYTHONPATH=$PWD/src python -m pytest -q -p no:cacheprovider src/tests   # base 7b1164d
272 passed, 1 warning in 16.25s
$ PYTHONPATH=$PWD/src python -m pytest -q -p no:cacheprovider src/tests   # this series
292 passed, 1 warning in 17.56s
$ # falsification: routing_logic.py and request.py restored to 7b1164d, every test file kept
18 failed, 274 passed, 2 warnings in 15.49s   rc=1
# 14 in test_route_request_signature.py, 3 in test_roundrobin_router.py, 1 in test_request_stats.py
$ uvx black@25.1.0 --check <six changed files>
6 files would be left unchanged.
$ uvx isort@6.0.0 --check-only --profile black <six changed files>   # rc=0, no output
$ uvx ruff@0.12.0 check <six changed files>
All checks passed!
$ uvx codespell@2.4.1 <six changed files>   # rc=0, no output
# not run: the helm, actionlint, hadolint and shellcheck hooks, because no file they cover changes; the router end to end

• Make sure the code changes pass the pre-commit checks.
  • black, isort, ruff and codespell ran at the pinned versions on all six changed files.
• Sign-off your commit by using -s when doing git commit
• Try to classify PRs for easy understanding of the type of changes, such as [Bugfix], [Feat], and [CI].
  • [Router], the prefix [Router] Standardize route_request signature across all routers #1023 used.

@Shaoting-Feng, you asked on #1095 for someone to pick #1023 up, and this is that rebase as its own pull request. If Mohith26 comes back, should it fold into his branch instead?

[gemini-code-assist]

Code Review

This pull request standardizes the route_request signature across all routers to be asynchronous and accept an optional request_json parameter, simplifying call sites by removing complex isinstance dispatch guards. Feedback was provided regarding a potential KeyError in PrefixAwareRouter when request_json is empty or missing the prompt key.

[gemini-code-assist]

If request_json is None or an empty dictionary (such as the {} passed during multipart/form-data requests), PrefixAwareRouter will raise a KeyError: 'prompt' at line 934 because it directly accesses request_json["prompt"] when "messages" is not present. \n\nTo prevent this runtime crash, we should ensure that request_json always contains at least an empty string for "prompt" if neither "prompt" nor "messages" is specified.

        if request_json is None:\n            request_json = {\"prompt\": \"\"}\n        elif \"prompt\" not in request_json and \"messages\" not in request_json:\n            request_json = {**request_json, \"prompt\": \"\"}