[surajm20061998]

Overview

--enable-log-outputs is silently ignored by /v1/completions. The handler never accepted the flag or logged generated text, while chat, Responses, Anthropic (#60982), Cohere (#61089) and /inference/v1/generate all do. This adds output logging to completions, mirroring chat.

Claims

• With --enable-log-requests --enable-log-outputs, /v1/completions logs generated text: one line per output when not streaming; when streaming, one line per delta (respects --enable-log-deltas) plus the full text when the stream ends.
• Output lines reuse the per-prompt request id from the input log (cmpl-<id>-<i>), so the two can be correlated.
• Nothing changes when the flag is off (the default).

Validation

End-to-end, vllm serve Qwen/Qwen2.5-0.5B-Instruct --enable-log-requests --enable-log-outputs (CPU backend). For each request, the counts of Received request / Generated response log lines:

Request	main	This PR
/v1/completions, non-streaming	1 / 0	1 / 1
/v1/completions, streaming, 8 tokens	1 / 0	1 / 9 (8 deltas + complete)
/v1/chat/completions (control)	1 / 1	1 / 1

Log lines with this PR

Generated response cmpl-e2e-cmpl-nonstream-0: output: ' Paris. It was founded in 7', finish_reason: length
Generated response cmpl-e2e-cmpl-stream-0 (streaming delta): output: ' Paris', finish_reason: None
Generated response cmpl-e2e-cmpl-stream-0 (streaming delta): output: '.', finish_reason: None
...
Generated response cmpl-e2e-cmpl-stream-0 (streaming delta): output: '7', finish_reason: length
Generated response cmpl-e2e-cmpl-stream-0 (streaming complete): output: ' Paris. It was founded in 7', finish_reason: streaming_complete
Generated response chatcmpl-e2e-chat-nonstream: output: 'Hello! How can I assist you today', finish_reason: length

Tests: test_completion_logs_outputs_when_enabled[stream=False|True] drives create_completion with a mock engine and checks the logged text, token ids, finish reason, delta/complete lines, and that every line uses the input log's request id.

$ pytest tests/entrypoints/openai/completion/test_completion_error.py tests/entrypoints/openai/completion/test_completion_text_offset.py
main:    2 failed, 51 passed  (TypeError: ... unexpected keyword argument 'enable_log_outputs')
this PR: 53 passed

Regression: pytest tests/entrypoints/openai/completion/{test_completion_error,test_completion_text_offset,test_lora_resolvers}.py tests/entrypoints/openai/chat_completion/test_serving_chat.py tests/entrypoints/serve/utils: main 279 passed, this PR 281 passed. Both have the same 19 environment-only failures (GPT-OSS/Harmony tokenizers unavailable offline). pre-commit ruff and mypy are clean.

Validated on macOS (Apple Silicon, CPU backend), Python 3.12, torch 2.14.0, rebased on main @ e744eacf6. No model evaluation: logging only.

Details

The original --enable-log-outputs PR (#20707) implemented logging for chat and Responses only. Later endpoints that inherit from the chat handler only needed the flag forwarded (#60982, #61089). Completions has its own handler, so it needs the logging itself:

• OpenAIServingCompletion takes enable_log_outputs / enable_log_deltas, and init_generate_state passes them, as for chat.
• Non-streaming: after the response is built, log each output's generated text, token ids and finish reason.
• Streaming: log each non-empty delta, accumulate text only when logging is on, and log each choice's full text after the stream, with finish_reason="streaming_complete" as chat does.
• The generated text is logged, not the echoed prompt (echo=true); log_inputs already logs the prompt.

The batch chat endpoint and speech-to-text have the same gap. I'll follow up separately to keep this one small.

Duplicate check: searched issues/PRs for enable-log-outputs, enable_log_outputs, and log outputs + completions, and inspected every open PR touching vllm/entrypoints/openai/completion/serving.py and vllm/entrypoints/generate/api_router.py. #45634 adds request arrival/completion lifecycle logs, a different feature. #40155 is about prompt log levels. None add output logging to completions.

AI assistance: Claude (Anthropic) helped find, reproduce and write this change. I reviewed every changed line and ran the tests and the end-to-end check above myself.

---

Pull Request Checklist

• I used vLLM's /pr-checklist skill. (Mandatory for agents, optional for humans).

• AI assistance was used during the creation of this PR.

• Design Fit: Minimizes impact on core components, reuses existing functionality, and justifies added complexity.

• Testing and Validation: Validates the change and ensures any added tests are meaningful and reliable, with CI coverage or documented CI resource constraints and validation performed outside CI.

• Code Quality and Style: Keeps code and comments clear and concise, and updates relevant documentation and examples.

• Pull Request Contents: Includes a brief summary and relevant links, supports claims with evidence, explains root causes and implementation trade-offs, and follows the contributing guide.

[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]

@claude review

[claude]

Code review found no issues

No high-confidence issues detected in this change.

[DarkLight1337]

Thx

[DarkLight1337]

/ci run

[github-actions]

✅ Triggered Buildkite CI #95761 for commit 38ae9fd214e8.

[surajm20061998]

/ci retry

[github-actions]

✅ Queued 1 failed job(s) for retry in Buildkite CI #95761.

[surajm20061998]

The /ci retry hit the same single failure: H200 LM Eval KV-Offload, simple-nemotron-h-8b GSM8K run 2 = 0.395, i.e. 79/200 vs the 80/200 bar (0.45 − 0.05). The other 3 checks in that job and the remaining 124 jobs passed.

I don't think it's related to this PR:

• The eval server doesn't set --enable-log-requests / --enable-log-outputs (no request/output log lines in the job log), so the new code only reads two False flags.
• The same job passed on [CI][AMD] Split LM Eval Small by CDNA generation, unify AMD step titles, clone MoRIIO steps into test-amd.yaml #62363 (base 8 commits newer than ours, run 2 = 83/200) and [KV Offload] Load small CPU-tier hits synchronously within the step #61961 (older base, 82/200), and within each build run 1 vs run 2 already swings by up to 6 questions, the variance tracked in [ROCm]: run-to-run GSM8K accuracy variance for simple-nemotron-h-8b in KV-Offload test on MI300 #59818 / [CI] Stabilize KV-offload baseline accuracy check #56632.

Happy to retry again or rebase if you'd prefer. @DarkLight1337, let me know what works best for merging.

[DarkLight1337]

Retrying