Motivation

This PR tightens the visibility of structs and functions across the source files in sgl-model-gateway, mainly data_connector and grpc module.

As a result, some dead code are appearing (as clippy masks dead code when it is pub), the unused methods are removed.

Modifications

• Tighten visibility across the repo.

• Add allow(dead_code) to the code that are unused.

• Removed dead code in grpc/context.rs
    1. StreamingState struct - Removed entirely. The streaming implementation uses local variables within functions instead of this struct.
    2. Removed fields from ResponseState:
      - streaming: StreamingState
      - collected: Option<Vec<ProtoGenerateComplete>>
      - collected_embeddings: Option<Vec<ProtoEmbedComplete>>
      - harmony_parser: Option<HarmonyParserAdapter>
      - harmony_parser_per_index: Option<HashMap<...>>
    3. Removed unused accessor methods from impl RequestContext:
      - request()
      - responses_request()
      - embedding_request()
      - embedding_request_arc()
      - classify_request()
      - classify_request_arc()
    4. Removed unused imports: HashMap, Value, ProtoGenerateComplete

• Remove unused methods in impl ProtoRequest
    - Removed as_generate() and as_embed() methods since they were truly unused (code uses pattern matching instead)

• Remove unused methods in impl HarmonyResponsesContext

• Fix inconsistent patterns in Embedding and Classify
    1. embedding/response_processing.rs:
      - Removed .clone() and Ok(Some(Json(...)))
      - Now returns Ok(None) like Chat/Generate
    2. classify/response_processing.rs:
      - Same change - returns Ok(None)
    3. pipeline.rs execute_embeddings:
      - Changed from error case to extract response from FinalResponse::Embedding(response)
      - Returns axum::Json(response).into_response()
    4. pipeline.rs execute_classify:
      - Same pattern - extracts and returns the response
    5. context.rs FinalResponse:
      - Removed #[allow(dead_code)] from Embedding and Classify variants

  All four endpoint types (Chat, Generate, Embedding, Classify) now follow the same pipeline pattern:
  - Stage stores response in FinalResponse::Variant(response) and returns Ok(None)
  - Pipeline extracts response from FinalResponse and converts to HTTP response

• Fix inconsistencies in model_id during metrics recording in pipeline.rs
  • Fix execute_generate to use hardcoded UNKNOWN_MODEL_ID constant
  • Remove unnecessary model_for_metrics clone

Accuracy Tests

Benchmarking and Profiling

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 Process

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