[Tracking] Upstream remaining sglang-miles changes to main

Track the remaining work to land on main. Source PRs are already merged into miles branches; they are references, not pending PRs.

Status guide

Main PR / next action	Meaning
A PR link only	Existing main PR, open at the last check. Advance it.
Create/link main PR	No upstream PR is linked in this tracker. Link an existing PR or create one.
Decide / link PR or Decide: …	First confirm whether main still needs the change; then port, revise, or drop it.
Partial; … or Related PR; …	Existing PR coverage is partial or unconfirmed. Identify and complete the remaining work.
Reconcile …	Overlapping proposals or later fixes need to be aligned before proceeding.
Include in on-policy port	This fix depends on the on-policy feature and should be delivered with it.
Assign an internal coordinator	An internal maintenance owner still needs to be assigned.

Each work item has one checkbox. The handle beside each PR is its GitHub PR author. Ownership: internal PR author first; otherwise, the maintainer who merged that PR (mergedBy). For an unmerged external main PR, the corresponding Source PR owner coordinates upstreaming. Reviewer and commit identity do not determine ownership. If there is no Source PR or verified merger, leave ownership unassigned. External authors are not pinged.

Maintenance follow-up: @yueming-yuan

• Shared-indexer checker exemption; include None guard

  • Source: #29339 (yueming-yuan)
  • Main PR / next action: #29340 (yueming-yuan)

• Refresh attention-sink caches

  • Source: #30421 (yueming-yuan)
  • Main PR / next action: #29916 (yueming-yuan)

• Positional-buffer checker exemptions

  • Source: #37649 (yueming-yuan)
  • Main PR / next action: #37651 (yueming-yuan)

• True on-policy support

  • Source: #18639 (yueming-yuan)
  • Main PR / next action: Create/link main PR

• R3 capture, padding and raw-indexer replay

  • Source: #18642 (yueming-yuan), #36404 (yueming-yuan)
  • Main PR / next action: Create/link main PR

• Remaining PD retract behavior

  • Source: #18646 (yueming-yuan)
  • Main PR / next action: Create/link main PR

• MTP graph switch; revalidate clamping

  • Source: #18647 (yueming-yuan)
  • Main PR / next action: Create/link main PR

• Per-process DeepGEMM cache

  • Source: #27604 (yueming-yuan)
  • Main PR / next action: Create/link main PR

• DSv4 packed-scale weight updates

  • Source: #27728 (yueming-yuan)
  • Main PR / next action: Create/link main PR

• Decide whether to retain reload logging changes

  • Source: #27603 (yueming-yuan)
  • Main PR / next action: Decide / link PR

• Remaining INT4 QAT weight-processing/checker changes

  • Source: #18565 (yueming-yuan)
  • Main PR / next action: Create/link main PR

• P2P mappings and transfer, including NextN

  • Source: #21278 (JD-ETH), #22663 (JD-ETH), #33478 (yueming-yuan)
  • Main PR / next action: #20907 (JD-ETH), #32565 (JD-ETH); Partial; complete remaining port

• Pausable DSA metadata

  • Source: #33479 (yueming-yuan)
  • Main PR / next action: Create/link main PR

• Multimem override

  • Source: #36110 (yueming-yuan)
  • Main PR / next action: Create/link main PR

• PD R3 capture and mini_lb payload

  • Source: #22916 (zyzshishui)
  • Main PR / next action: Create/link main PR

• Host-local checkpoint pull

  • Source: #30366 (nanjiangwill)
  • Main PR / next action: #30367 (nanjiangwill)

• Revalidate VLM legacy fallback

  • Source: #18781 (yueming-yuan)
  • Main PR / next action: Decide: port, rewrite or drop

• Revalidate GLM shared-expert metadata

  • Source: —
  • Main PR / next action: #26041 (yueming-yuan); Decide: update or close

• Streamed LoRA registration and sessions

  • Source: #36384 (yueming-yuan)
  • Main PR / next action: Create/link main PR

• Weight-processing/session integration

  • Source: #28001 (yueming-yuan)
  • Main PR / next action: Create/link main PR

• Weight-checker snapshot release

  • Source: #36561 (yueming-yuan)
  • Main PR / next action: #37284 (yueming-yuan)

Maintenance follow-up: @yushengsu-thu

• Distributed LoRA loading

  • Source: #27268 (gongyisheng), #31759 (yushengsu-thu)
  • Main PR / next action: #31521 (yushengsu-thu)

• LoRA upsert

  • Source: #30913 (mathewjhan)
  • Main PR / next action: #32238 (yushengsu-thu)

• LoRA post-load processing

  • Source: —
  • Main PR / next action: #31524 (yushengsu-thu)

• MoE-LoRA, DP/graphs and reload behavior

  • Source: #29874 (yushengsu-thu), #31251 (yushengsu-thu), #25141 (nanjiangwill)
  • Main PR / next action: #31525 (yushengsu-thu), #32241 (yushengsu-thu), #32584 (yushengsu-thu), #32704 (yushengsu-thu), #31066 (kailash109), #36389 (erikwijmans); Reconcile alternatives and later fixes

• Prefix abort

  • Source: #30912 (mathewjhan)
  • Main PR / next action: #32237 (yushengsu-thu)

• One-sided LoRA target warnings

  • Source: #32376 (yushengsu-thu)
  • Main PR / next action: #32406 (yushengsu-thu)

• Complete LoRA lease/reloadable lifecycle

  • Source: #33299 (yushengsu-thu)
  • Main PR / next action: #31808 (shafeeqibraheem), #34215 (souvikDevloper), #37824 (Swigler); Partial; complete remaining port

• Staged LoRA publication

  • Source: #38165 (yushengsu-thu)
  • Main PR / next action: Create/link main PR

Maintenance follow-up: @guapisolo

• Draft/all-runner weight updates and checksum wire

  • Source: #28575 (guapisolo), #27750 (guapisolo)
  • Main PR / next action: Create/link main PR

• Kimi raw tool-call IDs

  • Source: #25196 (guapisolo)
  • Main PR / next action: Create/link main PR

• Anthropic conversion plus remaining error/fake-SSE helpers

  • Source: #35127 (guapisolo)
  • Main PR / next action: #34312 (Shi-Dong), #34313 (Shi-Dong); Partial; port remaining helpers

• FP8 tuple handling for the on-policy path

  • Source: #30742 (XinyuJiangCMU)
  • Main PR / next action: Include in on-policy port

• Quantized checker exemptions

  • Source: #32809 (guapisolo)
  • Main PR / next action: Create/link main PR

• TMS graph-pool backup

  • Source: #32718 (zyzshishui)
  • Main PR / next action: Create/link main PR

• Checker reset exemptions

  • Source: #27750 (guapisolo; reset exemption)
  • Main PR / next action: #31621 (yuchenwang3); Advance the reset-side fix

Maintenance follow-up: @maocheng23

• Ray RDT/NIXL weight synchronization

  • Source: #27723 (xyuzh)
  • Main PR / next action: Create/link main PR

• Remaining IdleSleeper pause guards

  • Source: #22623 (maocheng23)
  • Main PR / next action: #25319 (maocheng23); Related PR; cover guard remainder

• Pause-aware weight-update and post-process locking

  • Source: #22754 (maocheng23), #29675 (maocheng23)
  • Main PR / next action: #25319 (maocheng23); Partial; reconcile remaining locking

Maintenance follow-up: @Zhichenzzz

• Flush/release after retract

  • Source: #31962 (Zhichenzzz)
  • Main PR / next action: #32153 (Zhichenzzz)

• GDN norm device handling

  • Source: #37954 (Zhichenzzz)
  • Main PR / next action: Create/link main PR

• Remaining Nemotron/Mamba metadata

  • Source: #27110 (Zhichenzzz)
  • Main PR / next action: Create/link main PR

• Nemotron-H replay layer IDs

  • Source: #38718 (Shi-Dong)
  • Main PR / next action: Create/link main PR

Shared follow-up: @yushengsu-thu, @yueming-yuan; fzyzcjy

• GDN mapping, CLI targets and TP/DP sharding
  • Source: #30499 (yushengsu-thu), #37466 (yueming-yuan), #37887 (fzyzcjy)
  • Main PR / next action: #31522 (yushengsu-thu), #31523 (yushengsu-thu), #30189 (buffett0323), #30963 (huangtingwei9988); Reconcile overlapping PRs
  • Owners: yushengsu-thu for his PRs and upstream coordination; yueming-yuan for Source #37466; fzyzcjy for Source #37887 (self-merged maintainer).

Needs an internal coordinator

• Revalidate default DSv4 fusion disabling

  • Source: Follow-up commits (PR author unconfirmed)
  • Main PR / next action: Decide: port, narrow or drop
  • Owner: Unassigned; identify the originating PR before routing.

• MRoPE axis_map checker exemption

  • Source: Follow-up commits (PR author unconfirmed)
  • Main PR / next action: Create/link main PR
  • Owner: Unassigned; identify the originating PR before routing.

Resolved / no separate action

• #39142 (zianglih) → #39141 (zianglih): already merged into main.
• #27749 (guapisolo) → #28575 (guapisolo): old implementation replaced; the replacement's upstream work remains above.
• #23672 (ByronHsu), #23887 (ByronHsu), #39119 (guapisolo): obsolete PD guard removed; do not port it.

Scope and verification

• PR states checked on September 14, 2026; functional scope comes from the September 11 audit plus the September 14 branch additions. PR authors and merge actors verified against GitHub on September 17, 2026.
• Partial/overlapping PRs require a coverage check. Include later branch fixes when porting.
• Dependencies: LoRA load/lifecycle → streamed sessions → staged publication; P2P mappings → RDT; R3 capture → indexer/PD payloads.
• Report the main PR link, dependencies, ETA, and validation status. Internal PR authors own their work; external contributions route to the maintainer who merged them.