[Spec] Stage Inkling MTP draft metadata before verify

[paulzhang-tm]

Motivation

Inkling MTP draft extension reads shared request/KV and conv-slot mappings after target verification, keeping the scheduler from updating those mappings for the next batch. Stage those reads into draft-owned buffers before verify so the target metadata-completion event can release the scheduler earlier.

Runtime-only mirror, following the staging approach in #35126 and #35289.

Modifications

Exactly five files under python/sglang/srt/:

File	Change
layers/attention/base_attn_backend.py	Add an opt-in staging capability, disabled by default.
layers/attention/flashattention_backend.py	Own the staging capability checks and fill draft-owned metadata using the existing metadata kernel.
layers/attention/linear/inkling_sconv_backend.py	Stage attention metadata and conv slots outside draft graphs; retain result-dependent conv metadata inside them. Commit target conv state through prepared slots.
speculative/multi_layer_eagle_draft_extend_cuda_graph_runner.py	Stage mapping reads before verify; populate result-dependent inputs afterward. Clear padded request slots on batch reuse.
speculative/multi_layer_eagle_worker_v2.py	Stage before verify and select the target fence only when staging and both graph paths succeed.

Preserves #36752: unstaged/eager fallback still uses the final draft runner. Combined draft graphs remain supported. Staging requires compatible top-k-1 attention/conv backends, a target graph marker, and a supported batch bucket. The Inkling wrapper queries the full-attention capability without importing a concrete backend. Standalone FlashAttention keeps its existing in-graph metadata path and final draft fence.

Accuracy Tests

• Existing public unit tests: 9 passed (test_multi_layer_eagle_shared_read_event.py and test_decode_cuda_graph_shared_read_fence.py).
• 14 targeted tests passed on GB300 against this public checkout, including those nine unit tests, four GPU integration cases, and a regression test preserving the standalone FlashAttention draft fence. The GPU cases cover stage → prepare → replay after shared mappings change (page sizes 1/16/128, boundary windows, batch reuse and fallback), plus real target conv-state writes after slot reassignment.
• Ruff 0.15.1 format/lint, isort 7.0.0, Python syntax, and git diff --check passed.

The GPU validation harness is separate from this runtime-only mirror. No test, private-extension, or benchmark files are added.

Speed Tests and Profiling

No end-to-end speed benchmark was run against this upstream checkout. The change moves shared mapping reads before verify; model computation and KV allocation are unchanged.

Checklist

• Follow upstream formatting and import conventions.
• Keep the diff limited to the five relevant srt/ files.
• Preserve upstream's final-draft-runner fallback.

-Robot

---

CI States

Latest PR Test (Base): 🚫 Run #34536403208
Latest PR Test (Extra): ❌ Run #34536403316
Latest PR Test (AMD ROCm 10): ⏳ Run #34536403290

[Qiaolin-Yu]

/tag-and-rerun-ci

[Qiaolin-Yu]

/rerun-test test/registered/e2e/models/test_mimo_v2.py test/registered/e2e/models/test_mimo_v2_flash.py

[github-actions]

⚠️ Rebase Required Before Re-run

A major update has landed on main. Your PR is diverged relative to required base commit 3700c4ee26a1.

Re-run was not dispatched. What to do:

• Rebase your branch onto the latest main and push again
• Follow issue #21065 for context
• CI-fix PRs may request the bypass-maintenance label to skip this check

[Qiaolin-Yu]

/rerun-test test/registered/e2e/models/test_mimo_v2.py test/registered/e2e/models/test_mimo_v2_flash.py

[github-actions]

Results for /rerun-test test/registered/e2e/models/test_mimo_v2.py test/registered/e2e/models/test_mimo_v2_flash.py:

🚀 8-gpu-h200 (2 tests): ✅ View workflow run

cd test/ && python3 registered/e2e/models/test_mimo_v2.py
cd test/ && python3 registered/e2e/models/test_mimo_v2_flash.py