fix(repo): main 복구 -- #230 이 되돌린 다른 세션의 작업 33 파일 복원 + KDA 2라운드 재적용

[choiceoh]

무엇이 잘못됐나

워크트리들이 .git 을 공유하므로 다른 세션의 git fetch 가 이 세션의 origin/main 참조를 새 main(0bd06b7)으로 올려놓았고, 옛 base(83c2896) 기준 작업 트리를 git reset --soft origin/main 으로 스쿼시한 커밋(7a287e0)이 #226~#229 의 추가분을 통째로 되돌린 채 #230 으로 머지됐습니다(36 파일, 11,145 줄 삭제). 감사 결과 #224(MHC)도 같은 방식으로 prep_fused 작업을 한 번 지웠고, 그쪽이 #226/#228 로 다시 올렸던 것입니다.

이 PR

0bd06b7(머지 직전 main)의 전체 트리로 되돌린 뒤, KDA 2라운드의 진짜 차이(83c2896..7a287e0 의 megakernel .cu / README / tests / MEASUREMENTS)만 3-way 로 재적용(충돌 없음)하고 복구된 프로필로 build/glm53 을 다시 합성했습니다(33 overlays from 22 modules).

• git diff --stat 0bd06b7 = KDA 2라운드의 5개 파일뿐(MEASUREMENTS +6, README, overlay/build glm53_megakernel.cu, tests +27).
• tests/test_logic.py 통과(그쪽 테스트 + 내 계약).

재발 방지(메모리에 기록): 스쿼시 base 는 git merge-base origin/main HEAD; PR 전 git fetch origin && git diff --stat origin/main...HEAD 로 내 파일만인지 확인; 머지 후 first-parent diff 로 삭제 여부 확인.

🤖 Generated with Claude Code

---

Note

High Risk
Changes decode input preparation, speculative scheduling, and fp32 indexer numerics on serving-critical paths; mitigated by env opt-in, preimage DISARM, and shadow/self-check for prep-fused, but armed modes can still affect KV/attention if guards fail.

Overview
Restores GLM-5.3 kernel-campaign work that had been dropped from main, alongside documentation and build/glm53 artifacts for three opt-in decode optimizations.

glm53_prep_fused (EXP-7) adds a Triton fused prep path that replaces per-step eager prepare_inputs / prepare_attn / KV metadata builders with one H2D copy, one _glm53_prep_fused_kernel launch, and deep_gemm schedule setup. It patches the V2 GPU runner (preimage checks, shadow/self-check, live UVA handle reads, metadata cache for DFlash-only boots).

glm53_async_dflash (EXP-8) vendors a forked config/vllm.py with _deneb_dflash_async_ok so VLLM_GLM53_ASYNC_DFLASH=1 whitelists dflash for async scheduling (stock vLLM disables async for dflash by method name only).

glm53_indexer_gate_splitk (EXP-9) adds a small-M Triton split-K head-gate and wires glm53_prefill_fastpath through _glm53_head_gate when VLLM_GLM53_INDEXER_GATE_SPLITK=1.

MEASUREMENTS.md, STEP_KERNEL_MAP.md, and RUNBOOK_KERNEL_CAMPAIGN2.md document trace-driven ceilings, v2 offline gates, and experiment ladders; the runbook also fixes megakernel launcher examples to use caller env instead of invalid EXTRA_ENV= profile keys.

^{Reviewed by Cursor Bugbot for commit 0e01d48. Bugbot is set up for automated code reviews on this repo. Configure here.}

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-02T13:24:27.956168Z	0e01d48	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[Copilot]

🟡 Changes recommended

The newly added probe/tools scripts have confirmed runtime errors on valid inputs (CUDA/CPU mask mismatch and unguarded indexing on short/no-gate traces) that should be fixed before merging.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

Restores and re-applies previously reverted GLM53 optimization work by introducing three opt-in overlay modules (prep-fused, async-dflash allowlist, indexer head-gate split-K), plus accompanying probes, documentation, and contract-style logic tests to prevent silent drift/regressions.

Changes:

• Add glm53_prep_fused overlay to fuse decode-step host-side input prep into a single Triton launch (with preimage pinning + shadow/self-check modes).
• Add glm53_async_dflash overlay to allow async scheduling for dflash behind an explicit env knob.
• Add glm53_indexer_gate_splitk overlay to optionally replace the indexer fp32 head-gate torch.mm with a split-K Triton kernel for small-M decode shapes, and wire it into both attention + prefill fastpath.

File summaries

File	Description
tools/trace_step_timeline.py	Adds a trace timeline helper (needs a couple robustness guards).
tools/trace_step_composition.py	Adds per-step kernel-category composition analysis (needs a guard for short traces).
tests/test_logic.py	Expands overlay contract tests and makes symbol-contract checking profile-aware.
STEP_KERNEL_MAP.md	Adds supplemental trace decomposition/interpretation write-up.
RUNBOOK_KERNEL_CAMPAIGN2.md	Updates runbook and adds EXP-7/8/9 arming ladders and caller-env guidance.
profiles/glm53.env	Mounts the new overlay modules and introduces new opt-in knobs (default off).
probes/run_prep_fused_check.sh	Adds a containerized probe wrapper with base-preimage verification.
probes/prep_fused_check.py	Adds an offline bit-exactness + timing probe for prep-fused (has a device bug).
probes/moe_gate_tile_sweep.py	Adds an offline Triton tile sweep script for MoE gate kernel configs.
probes/indexer_gate_check.py	Adds offline numerics + timing probe for indexer head-gate split-K.
overlay/modules/glm53_prep_fused/requires	Declares module dependencies for prep-fused overlay.
overlay/modules/glm53_prep_fused/README.md	Documents rationale, contract, guards, and arming for prep-fused.
overlay/modules/glm53_prep_fused/manifest.tsv	Adds the new glm53_prep_fused.py file to the overlay manifest.
overlay/modules/glm53_model_wiring/glm5next_model.py	Installs prep-fused at model init with safe ImportError handling.
overlay/modules/glm53_model_wiring/glm53_prefill_fastpath.py	Routes indexer head-gate through a lazily-resolved helper.
overlay/modules/glm53_indexer_gate_splitk/requires	Declares dependency on model wiring for the split-K module.
overlay/modules/glm53_indexer_gate_splitk/README.md	Documents split-K behavior, numerics caveat, and arming.
overlay/modules/glm53_indexer_gate_splitk/manifest.tsv	Overlays attention.py and adds glm53_indexer_gate.py into image.
overlay/modules/glm53_indexer_gate_splitk/glm5next_attention.py	Attention overlay uses _glm53_head_gate helper for head-gate projection.
overlay/modules/glm53_indexer_gate_splitk/glm53_indexer_gate.py	Implements split-K Triton kernel and opt-in routing logic.
overlay/modules/glm53_async_dflash/README.md	Documents async-dflash allowlisting mechanism and operational checks.
overlay/modules/glm53_async_dflash/manifest.tsv	Overlays vllm/config/vllm.py with a pinned preimage.
MEASUREMENTS.md	Records trace findings and offline measurements for EXP-7/8/9 and lever analysis.
build/glm53/manifest.tsv	Updates composed build manifest to include new overlay artifacts.
build/glm53/glm5next_model.py	Composed build output mirroring wiring changes for prep-fused install.
build/glm53/glm5next_attention.py	Composed build output mirroring attention overlay changes.
build/glm53/glm53_prep_fused.py	Composed build output for prep-fused implementation.
build/glm53/glm53_prefill_fastpath.py	Composed build output mirroring fastpath head-gate helper routing.
build/glm53/glm53_indexer_gate.py	Composed build output for split-K head-gate kernel helper.

Review details

Suppressed comments (1)

tools/trace_step_timeline.py:27

• gates[0] and statistics.median(...) will crash when there are no _deneb_gate_partial* kernels in the selected step segment. Handle the empty case so the timeline tool still works on traces without gate kernels.

• Files reviewed: 30/33 changed files
• Comments generated: 3
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[cursor]

Left a non-blocking comment: Cursor Bugbot was present but still in progress after the 8-minute wait, so this run is not approving. Human review is needed; assigning a reviewer from the repo’s other contributor on the affected paths.

  

_{Sent by Cursor Approval Agent: Pull Request Router and Approver}