Refactor: Extract Scoring API into Dedicated Mixin Files + CODEOWNERS

Motivation

The Scoring API (/v1/score, Engine.score(), Engine.async_score()) is a self-contained feature spanning entrypoints, managers, and tests. This PR extracts scoring-specific logic into dedicated files and adds CODEOWNERS entries so that scoring contributors are automatically requested for review on scoring-related changes.

What Changed

1. New file: python/sglang/srt/entrypoints/engine_score_mixin.py

Extracted score() and async_score() from the Engine class into an EngineScoreMixin. The Engine class now inherits from this mixin — no behavior change, just cleaner separation.

2. Renamed: tokenizer_manager_multiitem_mixin.py → tokenizer_manager_score_mixin.py

Renamed the file and class (TokenizerManagerMultiItemMixin → TokenizerManagerScoreMixin) to better reflect its purpose. This file contains the core scoring logic: score_request(), score_prompts(), multi-item scoring helpers, and the ScoreResult dataclass.

3. Updated: python/sglang/srt/entrypoints/engine.py

• Engine now inherits from EngineScoreMixin (MRO: Engine → EngineScoreMixin → EngineBase → ABC)
• Removed the inline score() / async_score() methods (now provided by the mixin)

4. Updated: .github/CODEOWNERS

Added per-file ownership for all scoring-specific files:

/python/sglang/srt/entrypoints/engine_score_mixin.py @sundar24295s @chanh @fortunecookiee
/python/sglang/srt/entrypoints/openai/serving_score.py @sundar24295s @chanh @fortunecookiee
/python/sglang/srt/managers/tokenizer_manager_score_mixin.py @sundar24295s @chanh @fortunecookiee
/test/registered/core/test_score_api.py @sundar24295s @chanh @fortunecookiee
/benchmark/prefill_only/bench_score.py @sundar24295s @chanh @fortunecookiee

5. Updated: test/registered/openai_server/basic/test_serving_rerank.py

Updated import path to reflect the renamed module.

Files NOT Touched (and why)

These files contain scoring-related infrastructure (GPU kernels, scheduler logprob processing, attention masking) that is shared with other features. They were intentionally left in place:

• layers/logits_processor.py — compute_logprobs_for_multi_item_scoring() operates on GPU tensors using logits-processor-internal APIs
• layers/attention/flashinfer_backend.py — MultiItemScoringParams and _process_multi_item_scoring() are flashinfer attention-level code
• managers/scheduler_output_processor_mixin.py — multi-item scoring conditionals are interleaved with regular logprob processing
• server_args.py — multi_item_scoring_delimiter is a server config flag

Validation

Server startup

python -m sglang.launch_server \
  --model-path /shared/public/sharing/job-rank/kbehdin/f389cde308efd4dbb8d9-2025-06-06-18-31-30/best_model/epoch=0-step=498-HF \
  --port 30000 --host 0.0.0.0 \
  --chunked-prefill-size -1 \
  --enable-torch-compile \
  --dtype float16 \
  --max-prefill-tokens 100000 \
  --mem-fraction-static 0.3 \
  --enable-tokenizer-batch-encode \
  --disable-radix-cache \
  --disable-cuda-graph \
  --multi-item-scoring-delimiter 128255

Server starts successfully with no import errors.

Single-item scoring

curl -X POST "http://localhost:30000/v1/score" \
  -H "Content-Type: application/json" \
  -d '{
    "query": "What is the capital of California? Answer Yes or No for each of the following options:",
    "items": ["Scaramento"],
    "label_token_ids": [9454, 2753],
    "model": "..."
  }'

Response:

{
    "scores": [[6.398421076647679e-06, 3.3389641633503636e-06]],
    "model": "...",
    "usage": {"prompt_tokens": 23, "total_tokens": 23},
    "object": "scoring"
}

Multi-item scoring (3 items, softmax)

curl -X POST "http://localhost:30000/v1/score" \
  -H "Content-Type: application/json" \
  -d '{
    "query": "What is the capital of California? Answer Yes or No for each of the following options:",
    "items": ["Sacramento", "San Jose", "San Francisco"],
    "label_token_ids": [9454, 2753],
    "apply_softmax": true,
    "model": "..."
  }'

Response:

{
    "scores": [
        [0.4857216775417328, 0.5142783522605896],
        [0.6157812476158142, 0.3842187821865082],
        [0.5241511464118958, 0.475848913192749]
    ],
    "model": "...",
    "usage": {"prompt_tokens": 28, "total_tokens": 28},
    "object": "scoring"
}

Import verification

from sglang.srt.entrypoints.engine import Engine
assert hasattr(Engine, 'score')
assert hasattr(Engine, 'async_score')
# MRO: Engine → EngineScoreMixin → EngineBase → ABC → object

Test Plan

• Server starts without import errors
• /v1/score endpoint returns correct results for single-item scoring
• /v1/score endpoint returns correct results for multi-item scoring with softmax
• Engine.score() and Engine.async_score() are accessible via mixin inheritance
• All Python imports resolve correctly (no circular imports)
• No stale references to old file/class names remain in the codebase
• Existing CI test test/registered/core/test_score_api.py passes
• Existing CI test test/registered/openai_server/basic/test_serving_rerank.py passes