Add Scoring API support for SequenceClassification models

Summary

• Extends the /v1/score endpoint to support SequenceClassification models (e.g., Qwen3ForSequenceClassification, Qwen2ForSequenceClassification, LlamaForClassification) in addition to the existing CausalLM path. For classification models, the scoring API returns pooled class logits from the model's classification head — no label_token_ids required.
• Adds multi-item scoring (MIS) support for classification models via --multi-item-scoring-delimiter. All items are packed into a single sequence separated by a delimiter token and scored in one forward pass, with score_and_pool() extracting per-item scores at delimiter positions.
• Adds comprehensive unit and E2E tests covering both single-item and multi-item classification scoring paths.

Changes

Design

• Single-item scoring: Each query + item pair is sent as a separate EmbeddingReqInput. The model's pooler extracts the last-token hidden state, the classification head produces logits, and those logits are returned as scores.
• Multi-item scoring: When --multi-item-scoring-delimiter <token_id> is configured, all items are packed into one sequence: query<delim>item1<delim>item2<delim>.... The score_and_pool() function applies the classification head to ALL hidden states, then extracts logits at positions just before each delimiter — the same pattern CausalLM uses in LogitsProcessor.
• Delimiter positions are found dynamically via (input_ids == delimiter_token).nonzero() at forward time, avoiding the need to thread delimiter indices through the scheduling pipeline.

Test plan

Unit tests (CPU, no model needed)

source /workspace/venvs/sglang-repos/bin/activate
python test/registered/unit/test_pooler_score_and_pool.py -v

All 6 tests pass:

test_mis_extracts_positions_before_delimiter ... ok
test_mis_falls_back_when_no_delimiters_in_input ... ok
test_mis_falls_back_when_not_prefill_only ... ok
test_mis_returns_list_of_scores ... ok
test_single_item_returns_scores ... ok
test_single_item_scores_match_manual_computation ... ok
----------------------------------------------------------------------
Ran 6 tests in 0.007s
OK

E2E tests (GPU required)

Uses json_model_override_args to override Qwen/Qwen3-0.6B to Qwen3ForSequenceClassification with a random classification head.

python test/registered/core/test_score_classification.py -v

All 14 tests pass:

test_basic_single_item ... ok
test_deterministic ... ok
test_label_token_ids_ignored ... ok
test_raw_logits_without_softmax ... ok
test_single_item_edge_case ... ok
test_tokenized_inputs ... ok
test_deterministic (MIS) ... ok
test_items_produce_distinct_scores ... ok
test_mis_basic ... ok
test_mis_many_items ... ok
test_mis_single_item ... ok
test_softmax_valid ... ok
test_many_items_distinct (12 labels) ... ok
test_many_labels_shape (12 labels) ... ok
----------------------------------------------------------------------
Ran 14 tests in 45.538s
OK

Manual server test — single-item scoring

Launch server with a SequenceClassification model:

python -m sglang.launch_server \
  --model-path <path-to-seq-cls-model> \
  --port 30000 --host 0.0.0.0 \
  --chunked-prefill-size -1 \
  --dtype float16 \
  --mem-fraction-static 0.5 \
  --disable-radix-cache \
  --disable-cuda-graph

Single-item curl:

curl -s -X POST "http://localhost:30000/v1/score" \
  -H "Content-Type: application/json" \
  -d '{
    "query": "What is the capital of California?",
    "items": ["Sacramento"],
    "model": "default"
  }' | python3 -m json.tool

Returns class logits directly (12 labels for this model):

{
    "scores": [
        [-0.44, 0.27, -0.85, -0.41, 0.69, 0.32, 0.59, 0.38, 0.17, 0.70, -0.07, -0.07]
    ],
    "prompt_tokens": 14
}

Manual server test — multi-item scoring

Launch server with MIS delimiter:

python -m sglang.launch_server \
  --model-path <path-to-seq-cls-model> \
  --port 30000 --host 0.0.0.0 \
  --chunked-prefill-size -1 \
  --dtype float16 \
  --mem-fraction-static 0.5 \
  --disable-radix-cache \
  --disable-cuda-graph \
  --multi-item-scoring-delimiter 151643

Multi-item curl:

curl -s -X POST "http://localhost:30000/v1/score" \
  -H "Content-Type: application/json" \
  -d '{
    "query": "What is the capital of California?",
    "items": ["Sacramento", "Los Angeles", "San Francisco"],
    "model": "default",
    "apply_softmax": true
  }' | python3 -m json.tool

Returns one probability distribution per item (all items scored in a single forward pass):

{
    "scores": [
        [0.05, 0.10, 0.03, 0.05, 0.15, 0.11, 0.14, 0.11, 0.09, 0.15, 0.07, 0.07],
        [0.05, 0.10, 0.03, 0.05, 0.16, 0.10, 0.14, 0.11, 0.09, 0.15, 0.07, 0.07],
        [0.05, 0.10, 0.03, 0.05, 0.15, 0.10, 0.14, 0.11, 0.09, 0.15, 0.07, 0.07]
    ],
    "prompt_tokens": 26
}

Pre-commit

SKIP=clang-format pre-commit run  # all hooks pass

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.

Review and Merge Process

1. Ping Merge Oncalls to start the process. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • Common commands include /tag-and-rerun-ci, /tag-run-ci-label, /rerun-failed-ci
4. After green CI and required approvals, ask Merge Oncalls or people with Write permission to merge the PR.