[sglang-miles] Extract Anthropic conversion into standalone utils

[guapisolo]

Summary

Expose SGLang's Anthropic/OpenAI conversion for reuse by external frontends.

Motivation

External OpenAI-compatible frontends that must interpret Anthropic Messages exactly as this server does (e.g. Miles PR #2358) previously had to vendor protocol.py/serving.py and maintain a derived codec, plus a sync ledger to keep the copies honest. This gives the server and external callers one conversion implementation, maintained here.

Usage

An external handler validates AnthropicMessagesRequest, then wraps its own OpenAI path with the shared converters:

from sglang.srt.entrypoints.anthropic.serving import (
    convert_response,
    convert_to_chat_completion_request,
)

chat_request = convert_to_chat_completion_request(
    anthropic_request,
    merge_inline_system=merge_inline_system,
)
chat_response = await openai_backend(chat_request)
anthropic_response = convert_response(chat_response)

The result uses the same conversion as SGLang's /v1/messages; endpoint validation, routing, and transport remain caller-owned.

Design Notes

• Mental model: An external /v1/messages handler parses AnthropicMessagesRequest; passes its deployment's merge_inline_system choice plus optional reasoning callbacks into convert_to_chat_completion_request; runs the resulting ChatCompletionRequest through its OpenAI-compatible backend; then converts the completed response with convert_response or to_anthropic_fake_sse_events. SGLang owns format conversion; the caller owns endpoint policy, session state, and transport.
• Runtime boundary: convert_to_chat_completion_request receives request data plus explicit inline-system policy and optional reasoning hooks; convert_response is data-only. AnthropicServing invokes both while supplying hooks from OpenAIServingChat, preserving server routing plus live streaming.
• Utility boundary: utils.py contains Anthropic error-envelope mapping plus eager fake-SSE construction from a completed OpenAI response. Raw-body parsing, feature policy, routing, and state remain caller-owned; no fake serving instance, AnthropicRequestContext, or allow_* matrix is exposed.
• Streaming boundary: Eager fake SSE materializes one delta per block after completion; SGLang's token-by-token streaming path is unchanged.

Verification

• PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=python python -m pytest -p no:cacheprovider -q test/registered/unit/entrypoints/anthropic/test_utils.py test/registered/unit/entrypoints/anthropic/test_serving.py: 84 passed, 55 subtests passed; covers request/response goldens, existing serving behavior, error mapping, fake SSE, and import hygiene.
• Miles tests/fast/router/test_session_anthropic.py plus tests/fast/utils/test_utils/test_mock_sglang_server.py: 58 passed, 1 skipped; verifies the paired session/TITO integration.
• pre-commit run --files python/sglang/srt/entrypoints/anthropic/utils.py test/registered/unit/entrypoints/anthropic/test_utils.py: passed.

Review Focus

• convert_to_chat_completion_request / AnthropicServing: reasoning-callback injection and preservation of server conversion behavior.
• utils.py: error mapping, message scrubbing, and the eager fake-SSE boundary.
• Cross-repo ABI with Miles [Bug] runtime_endpoint encounter a JSONDecodeError #2358: merge_inline_system, request/response types, and completed-response streaming.

---

CI States

Latest PR Test (Base): ❌ Run #33150255041
Latest PR Test (Extra): ❌ Run #33150254812
Latest PR Test (AMD ROCm 7.2): ❌ Run #33150255067