[Bugfix][Frontend] Report Cohere stop sequences correctly

[taking-lying-flat]

Summary

• report Cohere v2 stop-sequence termination as STOP_SEQUENCE instead of COMPLETE
• preserve the OpenAI-compatible stop_reason through the Cohere streaming state machine
• distinguish matched string stop sequences from EOS and integer stop-token IDs
• add non-streaming and streaming regression coverage

Root cause

vLLM represents EOS, stop-token, and stop-string termination with finish_reason="stop". A matched stop string is carried separately in stop_reason, but the Cohere v2 conversion only inspected finish_reason. As a result, its existing STOP_SEQUENCE mapping was unreachable in both response paths.

User impact

Cohere v2 clients can now distinguish a configured stop-sequence match from ordinary completion in both regular responses and SSE message-end events.

Why this is not duplicating an existing PR

I searched open pull requests using cohere stop_reason, cohere stop sequence, and Cohere STOP_SEQUENCE, and found no open PR covering this behavior. PR #45807 addresses the analogous Anthropic Messages API conversion, not the Cohere v2 endpoint or its streaming state.

Testing

.venv/bin/python -m pytest -q \
  tests/entrypoints/cohere/test_serving_conversion.py \
  tests/entrypoints/cohere/test_serving_streaming.py \
  tests/entrypoints/cohere/test_protocol.py

Result: 150 passed.

uvx ruff check \
  vllm/entrypoints/cohere/serving.py \
  tests/entrypoints/cohere/test_serving_conversion.py \
  tests/entrypoints/cohere/test_serving_streaming.py

uvx ruff format --check \
  vllm/entrypoints/cohere/serving.py \
  tests/entrypoints/cohere/test_serving_conversion.py \
  tests/entrypoints/cohere/test_serving_streaming.py

Result: all checks passed; all three files were already formatted.

Model evaluation

Not run (not applicable). This change only maps already-produced termination metadata to the Cohere wire enum; it does not change model execution, token generation, or accuracy.

AI Assistance

AI assistance was used to investigate the lifecycle mismatch, construct regression cases, draft the implementation, and review counterexamples. I reviewed the final diff and the test results below.

The submitting author reviewed every changed line in the final diff.

[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.

[DarkLight1337]

/ci run

[github-actions]

✅ Triggered Buildkite CI #83232 for commit 757a50264d76.

[taking-lying-flat]

The previous full CI run (Buildkite #83232) was triggered for commit 757a502. I subsequently merged main into this branch, which changed the current HEAD to 3a3cc93. The checks currently attached to the new HEAD are passing, but the full Buildkite CI run is still associated with the old commit.

Sorry for invalidating the previous run. Could a maintainer please trigger full CI again for the current HEAD?

[DarkLight1337]

/ci run

[github-actions]

✅ Triggered Buildkite CI #83283 for commit 3a3cc930bb21.