Upstream sync 2026-05-25 (158 commits)

[Clarit-AI]

Upstream sync — 2026-05-25

Merges sgl-project/sglang upstream/main @ 2b1e53c98 into the Engram fork.

• Base: origin/main @ 9d85d453b
• Upstream: upstream/main @ 2b1e53c98
• Commits merged: 158

⚠️ DRAFT — not ready. Awaiting GPU validation (scripts/validate-sync.sh on the H100) and a version tag before this is marked ready. Do not merge yet.

📌 Deviation from brief: the brief was authored against origin/main @ 46a9f41e1 / upstream/main @ 2aa699530 (121 commits, 2 conflict files). Both refs advanced before execution — origin/main via merged hygiene PRs #99/#100, upstream/main via normal movement — yielding 158 commits and 4 conflict files. Proceeded with reviewer approval; both conflict guards stayed armed.

Conflict resolution (4 files)

File	Resolution
python/sglang/srt/managers/schedule_batch.py	Keep ENGRAM conversation_id block and adopt upstream array("q", origin_input_ids) storage.
python/sglang/srt/managers/tokenizer_manager.py	Keep ENGRAM_MODIFIED header and add upstream from __future__ import annotations.
python/sglang/srt/managers/scheduler.py	Forward-port — see note below.
sgl-kernel/python/sgl_kernel/flash_mla.py	Keep ENGRAM_MODIFIED header and add upstream import dataclasses.

scheduler.py forward-port note

Upstream refactored the inline construction of output_streamer / batch_result_processor out of __init__ into two new helper methods, init_output_streamer() and init_batch_result_processor(). This collided with two ENGRAM kwarg injections. Resolution: adopt upstream's extracted-method structure, and relocate the two ENGRAM kwargs into the new method bodies:

• send_to_tokenizer=self.ipc_channels.send_to_tokenizer (M3 stateful-generate output routing) → now at scheduler.py:1755
• snapshot_hook_manager=self.snapshot_hook_manager (M2 hook) → now at scheduler.py:1782

Runtime-ordering verified safe statically: the relocated reads execute when init_output_streamer()/init_batch_result_processor() are called (in __init__), and both dependencies are assigned earlier — init_ipc_channels() at L421 and init_snapshot_system() at L507, both before the L630/632 call site. ⚠️ Validate at Phase 3b test #1 (server startup) and #5 — an AttributeError here would only surface at runtime, not in py_compile.

Marker accounting (for reviewers / CodeRabbit / Codex — do not re-flag)

• Broad BEGIN ENGRAM / END ENGRAM = 336 / 338. The −2 is entirely pre-existing example markers in .engram/SYNC_PLAYBOOK.md (2 BEGIN / 4 END) — documentation prose describing the marker convention, not code markers. Identical on origin/main; untouched by this merge. Independently flagged as benign in the 2026-05-25 sync report.
• Code markers balanced: python/ = 73 / 73; every individually merged file balances.
• Headers unchanged: 99 broad / 30 narrow.
• Preflight: the dangling KHA-390 marker in schedule_batch.py was converted to the inline ENGRAM_CHANGED form in a separate commit before the merge (KHA-390 is closed).

Phase 3b validation checklist (post-GPU, before ready)

• scripts/validate-sync.sh — all 6 tests pass (incl. test Implement startup snapshot warm restore #6 contract-conformance)
• Test feat: Add Mamba state snapshot persistence system (Phase 2) #1 server startup — confirms scheduler.py kwarg-ordering forward-port
• Test fix: implement restore_snapshots_on_startup (Gap 3) #5 — exercises snapshot hook (snapshot_hook_manager wiring)
• Snapshot save/restore smoke (Mamba/SSM stateful path)
• Cut version tag
• Mark PR ready

ℹ️ The ~9 hardware CI checks always fail on the fork (no GPU runners) and are not merge blockers.

Local quality gates (CPU-only, this checkout)

• ✅ Conflict markers: 0 remaining
• ✅ py_compile: all 4 modules pass (syntax-only; no CUDA/torch init)
• ✅ Code-marker conservation: python/ 73/73, headers unchanged
• ⏸️ Lint / types / tests: deferred to Phase 3b (GPU)

🤖 Generated with Claude Code

---

CI States

Latest PR Test (Base): ❌ Run #26477063450
Latest PR Test (Extra): ❌ Run #26477086339