[guapisolo]

Summary

Port #42349 to sglang-miles: chat/session return_indexer_topk, plus its layer-count tests.

Symptom & Reproduction

• Symptom: on sglang-miles, /v1/chat/completions with return_indexer_topk: true plus return_meta_info: true returns no meta_info.indexer_topk; a session turn drops the flag the same way, with no error. Miles session rollouts set this flag on chat requests (miles/rollout/session/request_args.py), so indexer replay through sessions silently gets no data.
• Reproduction: python -m pytest test/registered/unit/entrypoints/openai/test_serving_chat.py -k "internal_request_single or indexer_topk_without_meta_info" expects the flag on GenerateReqInput plus a 400 without return_meta_info; on sglang-miles at 258bc1a8d9 both fail, and with this change both pass. The session drop is at Session.create_req (Root Cause step 3).

Root Cause

1. ChatCompletionRequest declares no return_indexer_topk, so pydantic drops it.
2. _convert_to_internal_request therefore never sets GenerateReqInput.return_indexer_topk.
3. Session.create_req copies return_routed_experts onto the turn's Req, not return_indexer_topk.

Fix

The first commit cherry-picks #42349's two chat/session commits; the hunks are identical. Declare the field on ChatCompletionRequest, pass it into GenerateReqInput, and copy it in Session.create_req. Chat rejects return_indexer_topk without return_meta_info, matching return_sampling_mask, because meta_info is the only place a chat response carries it. /generate, /v1/completions, and requests without the flag are unchanged.

The second commit ports only the tests of #42349's layer-count commit (a4a57b3db1). sglang-miles already returns indexer_topk_num_layers through the scheduler payload, the multi-tokenizer split, and meta_info, and already has #42349's sequence-relative DSA capture, but nothing tested the layer count. The output-streamer test patches get_global_indexer_capturer at sglang.srt.state_capturer.indexer_topk, because sglang-miles imports it lazily in to_payload. The RL test reshapes by the returned meta_info["indexer_topk_num_layers"] instead of a hardcoded count. No production code changes in this commit.

Verification

• The reproduction above: 2 failed on 258bc1a8d9, 2 passed with this change.
• At 26336e6f98 (the first commit, on sglang-miles 9d86e9157a), on an H200 devbox whose image tree is 9d86e9157a with only that change applied: test_serving_chat.py and test_session_token_share_unit.py 161 passed, 113 subtests passed.
• At this head (afd6d7e3eb), CPU on macOS with PYTHONPATH=python: test_serving_chat.py, test_session_token_share_unit.py, test_multi_tokenizer_mixin.py, test_output_streamer_customized_info.py, and test_tokenizer_manager_rid_cleanup.py 246 passed, 163 subtests passed, including the 3 new layer-count tests. With to_payload forced to send indexer_topk_num_layers=None, the output-streamer test fails in both IPC codecs.
• Not run: the GPU e2e test/registered/rl/test_return_indexer_topk.py (DeepSeek-V3.2, 8 GPUs); only py_compile at this head.

Review Focus

• OpenAIServingChat._validate_request: new 400 for return_indexer_topk without return_meta_info.
• Session.create_req: its replay flags must stay in sync with the non-session Req built in scheduler.py.
• TestOutputStreamerIndexerTopk: its capturer patch target must follow where output_streamer imports get_global_indexer_capturer from.

---

CI States

Latest PR Test (Base): ❌ Run #39675164268
Latest PR Test (Extra): ❌ Run #39675163815
Latest PR Test (AMD ROCm 10): ❌ Run #39675163936

[github-actions]

Required checks failed on afd6d7e; this PR cannot be merged until they pass.

PR title (run)

PR title must use "[Type] scope: title" or "[Type] [Hardware] scope: title". Type must be one of: Feat, Fix, Refactor, Perf, Doc, Test, CI, Chore. Scope is required and may be any module or model name, with unrestricted letter case (letters, numbers, dots, underscores, slashes, or hyphens). Hardware is optional; use a single tag without spaces. Use uppercase hardware tags (AMD, NVIDIA, NPU, MI300X), except for exactly these four spellings: Intel, Apple, Huawei, Ascend. Reverts must use GitHub's default format: Revert "original title". Use one space between fields and after the colon, with a non-empty title and no leading or trailing whitespace. Examples: "[Fix] attention: Fix KV cache indexing", "[Feat] qwen3: Add model support", "[Perf] [AMD] flash-attn: Optimize decode performance".
Edit the title; the check reruns on its own.