Fix broken streaming response with --incremental-streaming-output

[kpham-sgl]

Motivation

Fixes #22510

--incremental-streaming-output + stream=True produces garbled text for all models (not just Gemma 4).

Root cause: #21037 (c37200f) changed tokenizer_manager.py to emit incremental deltas (state.text[last_text_offset:]) instead of cumulative text (state.text) when incremental_streaming_output is enabled. serving_chat.py was not updated and still slices the already-incremental text by the accumulated buffer length, producing garbage fragments.

Repro on any model:

# Server
python3 -m sglang.launch_server --model-path Qwen/Qwen2.5-0.5B-Instruct --incremental-streaming-output --port 51000

# Client
curl -s http://127.0.0.1:51000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"default","messages":[{"role":"user","content":"Hello, who are you?"}],"stream":true}'

# Before fix: chunks = ['Hello', 'e', 'age', 's', 'g'] -> "Helloeagesg"
# After fix:  chunks = ['I', ' am', ' Q', 'wen', ',', ' a', ' large', ...] -> coherent text

Modifications

• serving_chat.py: When incremental_streaming_output is enabled, use content["text"] directly as the delta instead of slicing by accumulated buffer length.
• test_serving_chat.py: Add regression test test_incremental_streaming_output_delta.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[kpham-sgl]

/tag-run-ci-label

[kpham-sgl]

/rerun-test test/registered/openai_server/basic/test_serving_chat.py

[github-actions]

✅ 1-gpu-5090 (1 test): View workflow run

cd test/ && python3 registered/openai_server/basic/test_serving_chat.py

[kpham-sgl]

/rerun-failed-ci