[hnyls2002]

Req.decoded_text is set to "" at construction and never assigned again, so every batch the scheduler sends to the detokenizer carries a column of empty strings.

• Remove Req.decoded_text, the decoded_texts field of BatchTokenIDOutput, and the code that fills, copies, and reads it. The detokenizer starts each request's text at "" directly.
• In _check_str_based_finish, stop_str in self.decoded_text was always False, so drop it together with the inner condition that was then always True.

No behavior change: the removed values were constant, and the rewritten stop-string loop was checked against the old one on every stop list of up to two strings and every tail drawn from a small alphabet (including empty stop strings). Tests that built or read the field are updated; the test that set decoded_text by hand to reach the unreachable branch is removed.

Verification

• CPU unit tests under test/registered/unit/managers/ plus beam_search/test_output_decode.py and sampling/test_sampling_params.py, one process per file: 1050 passed before, 1049 after (the removed test); the one failure is the same on both heads and unrelated.

---

CI States

Latest PR Test (Base): 🚫 Run #40955057028
Latest PR Test (Extra): ❌ Run #40955056883
Latest PR Test (AMD ROCm 10): 🚫 Run #40955056894

[claude]

Claude Code Review

This repository is configured for manual code reviews. Comment @claude review for a one-time review, or @claude review always to subscribe this PR to a review on every future push.

_{Tip: disable this comment in your organization's Code Review settings.}

[hnyls2002]

/tag-and-rerun-ci

[hnyls2002]

/rerun-test test_flat_raw_top_logprobs.py test_multi_tokenizer_mixin.py test_output_streamer_customized_info.py test_output_streamer_logprobs.py test_stop_str_speculative.py test_trim_matched_stop.py test_skip_tokenizer_init.py test_multi_tokenizer.py test_basic_sanity.py test_srt_endpoint.py registered/openai_server/validation/test_matched_stop.py

[github-actions]

🚀 ubuntu-latest (6 tests): ✅ View workflow run

cd test/ && python3 registered/unit/managers/test_flat_raw_top_logprobs.py
cd test/ && python3 registered/unit/managers/test_multi_tokenizer_mixin.py
cd test/ && python3 registered/unit/managers/test_output_streamer_customized_info.py
cd test/ && python3 registered/unit/managers/test_output_streamer_logprobs.py
cd test/ && python3 registered/unit/managers/test_stop_str_speculative.py
cd test/ && python3 registered/unit/managers/test_trim_matched_stop.py

🚀 1-gpu-5090 (4 tests): ✅ View workflow run

cd test/ && python3 registered/tokenizer/test_skip_tokenizer_init.py
cd test/ && python3 registered/core/test_basic_sanity.py
cd test/ && python3 registered/core/test_srt_endpoint.py
cd test/ && python3 registered/openai_server/validation/test_matched_stop.py

🚀 1-gpu-h100 (1 test): ✅ View workflow run

cd test/ && python3 registered/tokenizer/test_multi_tokenizer.py