[Tracking] SearchSpace distillation coverage and knowledge consolidation #91

Current state: Status checked 2026-11-09 21:41 Asia/Shanghai (live GitHub). The dated 2026-10-27 and 2026-10-30 sections below remain historical snapshots; their open/pending PR wording is superseded by this block and is not new scope or efficacy acceptance. Where a 2026-10-22 inventory row still lists a now-merged PR as pending, treat this Current state merge list as authoritative (addresses Nov 3 status-review findings formerly left only in comments).

Bounded source deliveries landed (checkbox = source distillation only; efficacy remains not_run unless separately measured): Batch 2 #263 merged as ac4c54bc; #138/#121/#122/#117 as f5bb4a10/cac287f6/25d6d5a6/ab03f3ff; Batch 3–5 #282/#283/#285 as 1fca842a/ebdbf5af/5d16e522; FlexGen reference #288; JustGoal #302 f58d0980; AutoTuneBench #306 f5a21bf0; NeMo-Speech.cpp #301 6c9f90da; Xe-Forge trial-identity #322 4715f957 (2026-11-09); Infera provenance #332 c38bb4e9 (2026-11-09); deepseek-v4-in-c byte/loader #324 e16a5c20 (2026-11-09). #133 and #156 stay closed unmerged.

Still open (source PRs / not checked): #325 NInfer contracts (open, head cbbee590); #339 faster-cosyvoice TTS playout guidance (open, head 881159057101c639504bd16d889c81cb5d514ee1); #342 Embodied.cpp / VLA benchmark guidance (open, head 0f4433b7c9bbc1512a4ef5526f087c2a9730d239). #339/#342 still amend the same inference-benchmark-design owner; exact heads and separate evidence stay explicit — combined pending changes were not tested together. Prior Ready CI / COMMENTED source reviews on those heads (see details below) are not APPROVE, merge authorization, SkillEval execution, or efficacy. SkillEval plans on #339/#342 remain unexecuted (zero comparisons; unsupported snapshots; budget/config blocked/unknown). Upstream rights outside inspected files remain UNKNOWN; device/audio/robot/performance unevaluated.

Still missing in the #177 queue (unchecked): NInfer (await #325 or equivalent), faster-cosyvoice (await #339 merge + licensing clarity), Embodied.cpp (await #342 merge), edge-dit.cpp, and the paper-led set (ReaDy, SoftCache, Continuous Depth Batching, LoopSpec, ReplaySSM). AutoProfiler / akas method intake is discussed only (not a checkbox yet). No new owner or experiment is assigned here.

Knowledge corpus catalog — discussion only (2026-11-09): Not approved; does not authorize tree moves, distillations, or a new child Issue this round. If a catalog-design Issue is opened later it would hang under architecture #362 after maintainer confirmation. #372 remains repository-tree governance only. Catalog would cover unit identity / classification / source pins + maturity / navigation index and Bundle publish — reusing skill/pack/Bundle IDs and hardware|frameworks|repositories|methods|workflows plus SearchSpace views; no third coverage authority. #177 dimensions may be navigation metadata only (no auto coverage or efficacy credit). User add-ons (not #177 daily rows): MindIE Agent pin Dong1017/mindie-agent@6a57cc8028d6dcf81117c207fcba2d668e6638bd + channel 0.1.12 is distinct from this Issue's existing mindie-agent @ 6a81b41 weight-loading row — do not overwrite; discussable patterns are Knowledge Admission gates, graded retrieval/budgets, local distill ≠ published corpus, thin Apps adapt. Knowledge CLI without valid Admission refuses knowledge ops; that does not mean remote_bash enforces coordinator leases. SpecPilot identity verified as arXiv:2611.02501 / StargazerX0/SpecPilot — do not confuse with SpecSpeech; public README says discovery loop/controller/memory are unreleased; author-reported speedups; no source-behavior audit here. Open discussion points: unit↔Bundle/skill relation; intake tags as navigation-only; future child-Issue scope under #362 vs keep on #91.

Integration note (historical): Earlier pairwise trial-merge conflict reports among #342/#339/#332/#325/#324/#322 and the #332-first sequencing proposal are partly superseded now that #332/#322/#324 have merged; remaining open heads (#325/#339/#342) still need current-head rebase/content reconciliation and regenerated projections by their existing authors/integrators. This status does not pick a merge order or grant merge authorization.

Decision needed: none for merges. Catalog axes remain discussion-only.
Next action: Keep #325/#339/#342 exact-head evidence separate from efficacy; check live claims before overlapping slices; leave empirical states not_run until measured. No new experiment, spend, or merge authorization from this refresh.

#339 / #342 exact-head evidence retained (unchanged claims; still open)

#339 head 881159057101c639504bd16d889c81cb5d514ee1: six Ready CI jobs passed; Codex summary completed with no findings (not formal approval/security/efficacy); SkillEval plan unexecuted; bot review no blocking defect; source review COMMENTED source-only LGTM (33 upstream files / 117-entry tree / 87 paths / 12 hashes) — not APPROVE or combined-amendment test.

#342 head 0f4433b7c9bbc1512a4ef5526f087c2a9730d239: Ready checks / declarations passed; Codex no findings; SkillEval unexecuted; source review COMMENTED LGTM. Owner remains candidate / not_run. Cross-PR conflict notes: 342, 339.

To-be-distilled: repo / paper tasks from research #177 — 2026-11-01

Queued for bounded source distillation. Checked items below mean bounded source delivery landed; they are not local validation or efficacy acceptance. Unchecked items remain undelivered or only proposed. Priorities rank knowledge value, not expected speedup. Pair each repository and its paper as one work unit. Before claiming a slice, check current canonical owners, source pins and overlapping PRs; reuse existing benchmark, search-policy and runtime owners. This addition does not assign workers, authorize experiments or promote efficacy.

First: evaluation and independent verification

• JustGoal repo + paper: same-model controlled comparisons, independent fixed correctness checks, single-/multi-shape coverage and workspace/stream/cache contracts. Preserve repaired results versus first-attempt success; kernels are not a complete campaign runner. Research. Bounded source delivery: #302 merged as f58d0980; source review, integration clearance and post-merge checks passed. The JustGoal reference and supporting source inventory are delivered; kernel correctness, performance and agent efficacy remain unverified. Wider backlog completion is not claimed.
• AutoTuneBench repo + paper v1: treatment-delivery checks, protocol/code gaps, saturation and weak-baseline counterexamples, anomaly quarantine. Extend existing evidence procedures rather than add a generic controller. Research. Bounded source delivery: #306 merged as f5a21bf0; source/integration review and post-merge checks passed. The reference and supporting source inventory are delivered; benchmark reproduction and agent efficacy remain unverified. Wider backlog completion is not claimed.
• Xe-Forge repo + paper v1: source-bound trial identities, external-evaluator verdicts and baseline-only oracle limitations. Separate kernel evidence from whole-model promotion; cited draft extensions are not delivered capabilities. Research. Bounded source delivery: #322 merged as 4715f957 (2026-11-09); source-only trial-origin/timing guidance on inference-benchmark-design. Approvals recorded on the PR; kernel/whole-model efficacy and Draft E2E remain unverified. Wider backlog completion is not claimed.
• deepseek-v4-in-c: independent byte/loader oracles, resident-state invariants during asynchronous misses, useful-miss accounting, and thermal/order-drift retractions. Tiny fixtures do not establish full-model quality. Research. Bounded source delivery: #324 merged as e16a5c20 (2026-11-09); source-only expert-cache byte-validity guidance on weight-loading-optimization. Approvals recorded on the PR; full-model quality and device efficacy remain unverified. Wider backlog completion is not claimed.

Next: runtime and workload contracts

• NInfer: joint KV/recurrent-state restoration, accepted-prefix replay, zero-accept/cancel negative controls and completion denominators. Coordinate existing recurrent-state knowledge. Research. Proposed, not merged: #325 (head cbbee590). Still missing: merge, current-head reconciliation, and any measured validation (not_run).
• AMD Infera: measured/projected provenance, seed-only comparator, failed-trial retention and trace-bound ranking calibration. Reuse serving-simulation/search-policy knowledge; predicted SLOs are not service guarantees. Research. Bounded source delivery: #332 merged as c38bb4e9 (2026-11-09); source-only serving-search provenance on calibrate-serving-simulation (+ search-policy route). Approvals recorded on the PR; projected SLO ≠ service SLO; ranking calibration and runtime efficacy remain unverified. Wider backlog completion is not claimed.
• faster-cosyvoice: PCM-playout/deadline scheduling, dependent-chunk batching, TTFP/underrun/quality and full-pipeline costs. Resolve licensing before copying code; retain voice-quality and repeatability gaps. Research. Proposed, not merged: #339 (head 88115905), extending inference-benchmark-design; licensing/rights remain UNKNOWN; efficacy not_run. Still missing: merge + current-head reconciliation after predecessors, SkillEval execution, device/audio quality evidence.
• NeMo-Speech.cpp: per-stream cache geometry, reset/end-of-utterance invariants and long-audio/load failures. Preserve precision and experiment-version caveats; skipped model tests are not passes. Research. Bounded source delivery: #301 merged as 6c9f90da; source review and post-merge checks passed. The cache-aware ASR reference is delivered; model/device, long-audio/load and service qualification remain unrun. Wider backlog completion is not claimed.
• Embodied.cpp repo + paper v3: CPU-observation→CPU-action timing, generated-action versus control-frequency distinctions, and per-task closed-loop quality failures. Preserve dirty-source/raw-episode gaps and unequal precision. Research. Proposed, not merged: #342 (head 0f4433b7), extending inference-benchmark-design. Still missing: merge, combined-amendment test with Add source-bound TTS playout benchmark guidance #339, SkillEval execution, robot/device/quality evidence.
• edge-dit.cpp: memory-budget/placement planning, double-buffer offload and cache fallback, retaining auto-fit quality failures. Fitting memory does not establish quality or E2E benefit. Research.

Additional paper-led tasks and deduplication boundaries

• ReaDy repo + paper v2: committed GDN-state conditioning and rejection rollback; distinguish fine-tuning benefit from the state-reader contribution. Research.
• SoftCache: cache staleness × sampler-step quality/cost tradeoff. Freeze exact paper version/readset before extraction. Research.
• Continuous Depth Batching: live exit and KV semantics; distinguish replay from online quality evidence. Freeze exact version/readset. Research.
• LoopSpec: intermediate-depth draft/final-depth verification, rejection and complete runtime cost. Freeze exact version/readset. Research.
• ReplaySSM repo + paper v1: input replay/checkpoint/accepted-prefix rollback, architecture boundaries, throughput versus single-request latency. Research.

Deduplication: existing Maya, Strata, stable-diffusion.cpp and other established source owners receive only uncovered increments, not duplicate whole-repo extraction. Preserve the active metadata slices #282/#283/#285. The former #263 partial delivery and foundation slices #277/#278/#279 are now merged; #288 is now merged, delivering its bounded FlexGen source reference; runtime validation and this backlog remain incomplete. Historical #133/#156 are closed and are not revived by this queue. The dated inventories below remain historical.

Watch, not promoted integration: retain SpecSpeech/AnySpeechRecognizer/Mach X completeness or numerical-drift counterexamples; PoSTCache/VLA-Rolling require source/configuration reconciliation first. Use the original #177 reports rather than treating these as proven speedups.

Completion evidence for each queued slice: pinned primary readset and license; overlap disposition and canonical output; applicable preconditions, invariants and counterexamples; source-linked PR plus deterministic integrity checks; evaluation plan/limits with empirical state left not_run until measured. A checked item means that bounded distillation was delivered, not that a repository, model or optimization was fully qualified.

Current decisions and bounded delivery — 2026-10-30

Superseded note (2026-11-09): #263 merged as ac4c54bc on 2026-10-31; the “remains open” wording below is historical only. See Current state.

Current outcome (historical text): #257's bounded source slice is merged; #263 remains open with the findings below. The tracker is not complete. Earlier source reconciliation was against 6e6ee245; the later delivery has its own linked evidence. The inventory and dated snapshots below remain historical; this is not a new full inventory or an efficacy assessment.

• Round 2 Batch 1 delivered: #243, merged as 401ee8a256c21e6c489e7d179e62c70ac08723d3, established the workflow sidecar shape and owner recommendations.
• Batch 2 decision is recorded, not a completed delivery: the updated scope/semantics record accepts projecting the four typed keys into the 17 frozen methods IDs, where the source body supports their stated meanings. Preserve the three excluded owners and source/section pointers; change only applicability/recommendations in the selected sidecars, regenerate projections from committed sources, and preserve maturity/evidence fields. The technical review is not blanket content approval. A closed key allowlist remains a later proposal, not an accepted Batch 2 requirement.
• Batch 2 partial PR remains open: #263, inspected head 1a52db0664d343628adcf5d8077aaa4ab65d91dc, proposes 15 of the 17 selected method sidecars; the two deferrals remain explicit, so this is not full Batch 2 delivery. The source review records a nonblocking count correction (no-applicability 41→26, not 42→27), current-main integration and missing current-head Ready evidence. The P2 provenance finding records unavailable generated-source snapshot 058fc76b656ea5a29bb4646a5e556e32d35bdbc9; the existing author must make the canonical snapshot reachable, preserve knowledge(hardware): audit sources and deepen memory, graph and collective guidance #262's source changes, regenerate projections and verify clean-checkout generator/integration checks. These are unresolved review records, not a new review or merge-ready conclusion.
• MindSpeed provenance slice delivered: #257 is merged. The final scoped review clears the source/integration findings with passing software checks. This completes only that bounded historical-source amendment—not Batch 2, runtime compatibility, model/agent benefit or the full tracker. docs: complete applicability metadata for 17 methods skills #263 remains the existing author's next unresolved partial delivery.

#257 delivery evidence and limits

Merge 436c950cf06ded742fef61ff4303493ac5175d4b; inspected head a1a511b0b2dcdde9d7cdc26a5d6fdca81ed5c65d. Integration review verifies the reachable joint source snapshot and preservation of #262. Ready checks tested integration into 6e6ee245; the later main delta was separately source-inspected, not presented as a rerun. Zero SkillEval comparisons/86 uncovered paths and the incomplete earlier bot review remain recorded limits; completed source/integration reviews support this source-only delivery. Candidate/not_run states and ownership are unchanged.

- **Separate hardware knowledge delivery:** [#262](https://github.com//pull/262) is merged as `6e6ee245d9e92ee9365976f2e4e262304999f147`; its governing [#260](https://github.com//issues/260) is closed. It adds bounded host-memory, TorchAir and HCCL implementation evidence and a 22-owner source-depth inventory. This does not complete Batch 2, promote candidate/not_run states, or establish target-device/model/performance/agent efficacy. - Existing owners remain unchanged. Coordinate against current source/provenance before editing overlapping owners; retain this living tracker open. No new extraction, execution budget, assignment or completion acceptance is created by this refresh.

---

Purpose and evidence boundaries

Track repository-derived knowledge by the five SearchSpace resource layers: what has been extracted, where it lives, what remains missing, and which versions need consolidation. A repository can appear in several sections because its mechanisms touch different resources; those rows are not independent source counts.

Snapshot: InferRSI main a1ee32664661, 2026-10-22. The bounded inventory is the matrices plus packaged source assets, not every potentially relevant upstream project.

Targeted update 2026-10-23: added verified historical delivery PR links and refreshed the affected CannBot, AKG, communication, TP-transfer, DLRover and intake rows. Other source pins/limits retain their recorded snapshot. Links below are ordinary references, not closing Development associations. A delivery link is not a new efficacy claim.

• Reference Matrix owns source pins, readsets and limits. Short pins below retain the recorded source identity; they do not mean current upstream HEAD.
• SearchSpace Matrix owns resource × optimization-stage coverage. This issue is a work/navigation view, not a third authoritative matrix.
• Extraction: scoped output / partial / located only / descriptor only / unknown. A scoped output is not whole-repository completion.
• Consolidation: not assessed / no overlap found / action needed / reconciled / unresolved conflict. Multiple pins alone do not prove a conflict. Unless specifically recorded below, consolidation is not assessed.
• Scoped extraction does not itself promote skills; retain each existing maturity label. Imported external cards remain locally unverified. Source reading, packaging checks and upstream reported results do not establish local device, model, service or performance validation.

Output directory convention

All paths below are relative to src/inferrsi/data/:

• skill_library/{hardware,frameworks,repositories,methods,workflows}/…/: complete procedures and, where applicable, metadata, references, cases and helpers.
• knowledge_packs/{dependency,peer}/<source>/…/*.json: imported cross-repository source cards.
• repo_packs/<repo>/knowledge/…: repository-owned knowledge or generated discovery projections; projections are not independent distillations.
• repo_packs/<repo>/pack.json: runtime descriptor only; its existence does not establish distilled knowledge.

Historical shared deliveries: #9 (engine guides), #11 (imported source cards), #12 (initial 48-skill library), #44 (canonical packaging), #46 (English/ownership cleanup), #80 (cross-repository synthesis), #102 (benchmark methodology). Packaging, cleanup and synthesis are not additional repository extractions.

Current delivery and consolidation delta — 2026-10-27

PR-status and provenance refresh against main 97f77a91373f. This is not a new full inventory: the 2026-10-22 inventory, output links and original source pins below retain their historical baseline unless explicitly updated. Open PRs are not main-branch coverage; merged navigation is not a new extraction or runtime acceptance.

Track	Verified delivery / pending work	Remaining boundary
Bounded source additions in main	CannBot #114, AFD #120, h3.c #110; earlier fastsafetensors #96, DLRover #95, AKG #107 and communication #108; Ascend host-offload #123, MLX cards #119 via #38, KDA KernelWiki #127 via #126	Candidate/source-specific limits remain. #123's bounded CPU-tier descriptor slice does not complete every host-KV gap deferred by #108 or establish native DMA/shutdown/ABI qualification
Multi-pin navigation in main	llama.cpp #124, vLLM-Ascend #130, SGLang #136, Diffusers #134, vLLM-Omni #139, TensorRT-LLM #142, PyTorch compiler/DCP #144	Separate pins/owners/evidence retained; no cross-pin/backend compatibility claim
Multi-pin navigation proposed, not merged	vLLM #138	Keep existing main owners; review/LGTM does not mark the contribution integrated
Bounded additions/extensions proposed, not merged	RLT #121; SD.cpp loader #122; Lithos #117 via #115	Do not count unmerged output as main coverage or infer native/agent efficacy
Knowledge access and evaluation	#129 whole-resource collect and #118 report-only PR evaluation planning/publishing are merged. #128 is closed with its separate evidence/limitations recorded in the acceptance comment	#88/#101 acceptance and authorized real addition/update evaluations remain separate; software delivery does not establish task benefit. #128 closure is not a general efficacy or context-cost improvement claim
Flare non-distillation delivery	#147 is closed for its bounded offline evidence-adapter scope, including corrective #155; see acceptance	Offline adapter acceptance does not qualify real service/model performance or complete #148 source distillation

Explicit distillation tasks added on 2026-10-25: #148 Flare (preparation contracts and complete session-state lifetimes) is delivered by #217, squash-merged as 509531c681479fbdd7ddd0e8b0032239e64a44c1. Fixed source: NVIDIA/local-inference (formerly NVIDIA/flare-local-inference) bce032c014bae34502ba22d3304fa084d5b40afe, tree f657cae770c66df9d6b47f9dbe8e40e14dc6fefa. Outputs: Flare session-lifecycle owner, reused weight-loading owner, native discovery/coverage and the bounded evaluation plan. Independent exact-head LGTM, Ready full/installed checks and static plan/publish passed. This accepts source-derived candidate knowledge / empirical not_run; model/device/service and efficacy remain unexecuted, with no automatic promotion. #149 LoopX (goal authority and quota decision boundaries) is delivered by #153, merged on 2026-10-28 as 840e2d868f136c6583c88c2ad75bc4c7d74b88bc. Bounded source readset: LoopX f094b2cfb2acde7a4a0d4d875bba52f64ac31ba9; canonical owner: workflows/rsi-evidence/references/loopx-goal-governance/, with thin optimization-search-policy navigation. Source/package review and Ready checks passed; candidate / empirical not_run and unexecuted evaluation-plan limits remain. This is one scoped source delivery, not #137 runtime implementation or agent/hardware efficacy. Both remain native sub-issues of this living tracker.

#133 and #156 are overlapping open proposals under #131. Coordinate with that existing owner rather than commission duplicate distillation; neither currently counts as main coverage.

Historical foundation extraction has positive records in the B-pack integration notes and A-pack prior coverage. They identify prior scheduling/KV artifacts, but do not recover the original archive's exact versions/readsets. #156 proposes new bounded PagedAttention/Sarathi-Serve slices; these remain distinct from original-archive recovery and are not integrated.

Keep this living tracker open. Existing task owners are unchanged; this refresh neither commissions new extraction nor reopens historical #48. Link future deliveries using ordinary references, not closing keywords.

1. Accelerator

Type / source and recorded pin	Scope / extraction state	Output / next action
Repo · Comfy-Org/ComfyUI @ b1a247c9cf9d	SCOPED · managed weight/patch identity, partial unload versus detach and dynamic-backend boundary	skill_library/repositories/comfy-ui/ Delivery: #73
Repo · leejet/stable-diffusion.cpp @ b13a48219d5c	SCOPED · native weight/runtime capacity, protected segment/prefetch lifetime	skill_library/repositories/stable-diffusion-cpp/audit-sd-cpp-residency/ Delivery: #77
Repo · NVIDIA/TensorRT-LLM @ f33e0e5fd	SCOPED · evidence taxonomy, kernel writing, casebook, roofline refs	skill_library/methods/dense-performance-evidence/; skill_library/methods/inference-optimization-casebook/; skill_library/methods/kernel-writing-guide/; skill_library/methods/performance-ceiling-analysis/ Consolidation: navigation delivered in merged #142; methodology and PD/KV source evidence stay distinct. No validated shared assumptions are established. Delivery: #12 (initial library import)
Repo · Ascend/MindIE-SD @ 8481d39	SCOPED · 5-layer analysis, bottleneck casebook	skill_library/hardware/ascend/ascend-optimization-casebook/; skill_library/hardware/ascend/ascend-performance-analysis/ Delivery: #12 (initial library import)
Repo · ascend-ai-coding/awesome-ascend-skills @ 75ec89c	SCOPED · graph mode, operator profiling/tuning, tiling design; includes historical cannbot-skills mirror material, not direct cann/cannbot application intake	skill_library/hardware/ascend/ascend-graph-mode/; skill_library/hardware/ascend/ascend-operator-profiling/; skill_library/hardware/ascend/ascend-operator-tuning/; skill_library/hardware/ascend/ascendc-tiling-design/ Delivery: #12 (initial library import)
Repo · cann/cannbot @ c0e2ad23ee1a73e3c0f84c8df1cc459e2aae3110	SCOPED · fixed-stage evidence acceptance; instructions versus lexical callback predicates; candidate only, host/NPU/agent effects untested	skill_library/hardware/ascend/ascend-performance-analysis/; direct canonical application source, separate from historical cannbot-skills Delivery: #114
Repo · vllm-project/vllm-ascend @ bc8115ed	SCOPED · FP8/KV-replication/multimodal adaptation	skill_library/hardware/ascend/ascend-model-adaptation/ Consolidation: version/responsibility navigation delivered in merged #130. Historical adapter documentation and ACLGraph runtime remain separate; no new OS-control or compatibility evidence. Delivery: #12 (initial library import)
Repo · vllm-project/vllm @ 7f730b4cea9a	SCOPED · independent admission budgets and async-KV readiness	skill_library/repositories/vllm/audit-vllm-kv-lifetimes/; skill_library/repositories/vllm/diagnose-vllm-admission-pressure/ Consolidation: navigation proposed in open #138; engine and scheduler/KV scopes stay separate. Not yet main delivery or backend qualification. Delivery: #12 (initial library import)
Repo · vllm-project/vllm-ascend @ 8a8e1e05ecb6	SCOPED · effective dispatch, caller-owned stable inputs, branch-specific replay/update order and capture-resource failures	skill_library/hardware/ascend/ascend-aclgraph-capture-replay/ Delivery: #12 (initial library import)
Repo · Ascend/pytorch (TorchNPU) @ 67511c53cd4c	SCOPED · stream-use lifetime versus ordering, HBM counters/fragmentation and feature-gated controls	skill_library/hardware/ascend/torchnpu-device-memory-lifecycle/ Delivery: #12 (initial library import)
Repo · hahnyuan/LLM-Viewer @ 1893e4b	SCOPED · roofline model + hardware params	skill_library/methods/performance-ceiling-analysis/ Delivery: #12 (initial library import)
Repo · ebugger/Empirical-Roofline-Toolkit @ 0771fa9	SCOPED · empirical roofline method	skill_library/methods/performance-ceiling-analysis/ Delivery: #12 (initial library import)
Repo · sgl-project/sglang @ df13d5dce31a	SCOPED · stage/use readiness, retention and executor rank boundaries	skill_library/repositories/sglang/audit-sglang-diffusion-lifetimes/ Consolidation: three-scope navigation delivered in merged #136; engine, HiCache and diffusion preserve separate pins/evidence. No cross-pin/backend compatibility or new Network qualification. Delivery: #79
Repo · ggml-org/llama.cpp @ 707616feb5f3	SCOPED · partial-ubatch progress, topology/input graph-reuse gates and backend copy/event obligations	skill_library/repositories/llama-cpp/audit-llama-runtime-state/ Consolidation: navigation reconciled in merged #124; loading/host placement and runtime keep separate pins, prerequisites and owners. No cross-pin/backend qualification. Delivery: #71
Repo · cann/oam-tools @ 8440935f (!302 import; later master f3cf4115 labeled [V2], never merged into !302 claims)	SCOPED · representative-step selection, code_ref source ownership, attribution-aware diagnosis, dependency DAG (data/state/event/collective_order), bounded overlap, correctness-first acceptance, metric separation, device-memory domain qualification	skill_library/hardware/ascend/cann-attribution-overlap/ Delivery: #12 (initial library import)
Repo · pytorch/pytorch (DCP async-save only) @ 7d239835e773; live async recipe retrieved 2026-10-15	SCOPED · training-visible staging pause/mutation boundary	skill_library/frameworks/pytorch/dcp-async-save-performance/ Consolidation: scope map delivered in merged #144; compiler and DCP pins/lifetimes remain distinct. No tested combined stack is established. Delivery: #12 (initial library import)
Repo · NVIDIA DALI @ 6fe47c48d3b0	SCOPED · CPU/GPU stage overlap and executor-mode gates	skill_library/methods/training-input-pipeline/ Delivery: #12 (initial library import)
Repo · NVIDIA Resiliency Extension @ 49af0b03f145	SCOPED · CUDA D2H synchronized before async request returns; no NPU qualification	skill_library/methods/checkpoint-recovery-consistency/ Delivery: #12 (initial library import)
Repo · DeepSpeed ZeRO-Infinity @ 5f266e26d216	SCOPED · parameter stage-3 and optimizer stage-1/2/3 boundaries	skill_library/methods/training-resource-budget/ Delivery: #12 (initial library import)
Repo · Ascend MindSpeed adaptive memory — core_r0.8.0, page history 53193db8 (full pin and license unresolved)	SCOPED · spare-NPU-memory versus recompute/swap decision	skill_library/methods/training-resource-budget/ Delivery: #12 (initial library import)
Repo · ml-explore/mlx @ bff10220c742	SCOPED · Lazy evaluation, dependency ordering and backend dispatch; native execution not run	skill_library/frameworks/mlx/lazy-evaluation-stream-lifetimes/ Delivery: #72
Repo · ml-explore/mlx-lm @ ab7f7048dc34	SCOPED · Prefill evaluation and async decode submission/read boundary; no backend qualification	skill_library/repositories/mlx-lm/generation-cache-lifecycle/ Delivery: #74
Repo · huggingface/diffusers @ aea7b70be0f5	SCOPED · Callback/interrupt and normal cleanup ownership; submitted kernels are not cancelled by loop skipping	skill_library/frameworks/diffusers/diffusers-invocation-state/ Consolidation: navigation delivered in merged #134; reviewed scope remains separate offload/invocation owners. No runtime qualification is established. Delivery: #76
Repo · deepseek-ai/DeepInfer @ 8b16149741a4	SCOPED · DFlash verification budget and bounded query replay; sampling, quality and native performance unrun	skill_library/repositories/deepinfer/ Delivery: #68
Repo · vllm-project/vllm-omni @ 18c3bd5d2d43	SCOPED · Abort acknowledgment versus actual device/transfer reclamation; incremental source guidance	skill_library/repositories/vllm-omni/sub-skills/omni-validation/ Consolidation: navigation delivered in merged #139; engine and abort/reclamation guidance retain separate scope. No runtime qualification is established. Delivery: #78
Repo · Dao-AILab/flash-attention @ 1fc0221183e7	SCOPED · Generation-specific GPU/dtype/API dispatch contracts; imported	knowledge_packs/dependency/flashattention/cuda/attention/ Delivery: #11 (source-card import)
Repo · flashinfer-ai/flashinfer @ 3637083a3a3d	SCOPED · Paged decode planning/dispatch and graph-buffer ownership; imported	knowledge_packs/dependency/flashinfer/cuda/llm/ Delivery: #11 (source-card import)
Repo · pytorch/pytorch @ f15b476814eb	SCOPED · Inductor CUDA graph shapes, guards and lifetimes; imported	knowledge_packs/dependency/pytorch/cuda/compiler/ Delivery: #11 (source-card import)
Repo · huggingface/transformers @ 2126f65b0ab2	SCOPED · Dynamic/static/offloaded/quantized KV contracts; imported	knowledge_packs/dependency/transformers/common/llm/ Delivery: #11 (source-card import)
Repo · triton-lang/triton @ 4eda6cc91e86	SCOPED · Fusion geometry versus register/shared-memory footprint; imported	knowledge_packs/dependency/triton/common/kernel/ Delivery: #11 (source-card import)
Repo · vipshop/cache-dit @ a57c826d3527	SCOPED · Residual approximation, boundary blocks and warmup; imported	knowledge_packs/peer/cache_dit/common/diffusion/ Delivery: #11 (source-card import)
Repo · InternLM/lmdeploy @ 822dddb8b3bc	SCOPED · KV quantization capacity/cost tradeoff and latency counterexamples; imported	knowledge_packs/peer/lmdeploy/cuda/llm/ Delivery: #11 (source-card import)
Repo · InferMatrixLab/qwen3-tts-engine @ 9f190577aa2a	SCOPED · Codec constants, graph batching/request state, fixed CustomVoice prompt; three imported cards	knowledge_packs/peer/qwen3_tts_engine/ascend/audio/ Delivery: #36
Repo · mindspore-ai/akg @ 5aa15f3f	SCOPED · CUDA tiling boundaries, full-point loops and tail guards; source-derived candidate only	skill_library/methods/kernel-writing-guide/; not whole-compiler coverage Delivery: #107
Repo · InferMatrixLab/vllm-rlt (formerly ThinkFlowLab)	TODO · bounded intake in main; recurrent KV candidate submitted in open #121, not yet merged	repo_packs/vllm_rlt/pack.json; retain the pinned intake; open #121 covers finalization/retirement at 4a350344027960a48384626aa646f1765fda6d4f, with candidate / empirical not_run boundaries Delivery: #104 (intake only)
Paper · PagedAttention / Sarathi-Serve	HISTORICAL EXTRACTION RECORDED · original foundation archive unavailable in the integrated corpus; exact original versions/readsets unrecovered	Main integration notes identify prior scheduling/KV artifacts. New bounded slices in open #156 under #131 are not merged. Recover provenance and coordinate existing deliveries before new extraction.
Paper · ZeRO-Infinity / FlexGen	TODO · Recorded foundation input missing; proposed capacity/offload tradeoff slices	Recover/check evidence; existing DeepSpeed configuration guidance is not the missing paper extraction.

2. RAM / Host

Type / source and recorded pin	Scope / extraction state	Output / next action
Repo · Comfy-Org/ComfyUI @ b1a247c9cf9d	SCOPED · output versus object identity, ancestor fingerprints, lazy consumer references and retention versus validity; host/pin ownership	skill_library/repositories/comfy-ui/ Delivery: #73
Repo · leejet/stable-diffusion.cpp @ b13a48219d5c	SCOPED · parameter storage versus compute staging and retained runtime ownership; host peaks unmeasured	skill_library/repositories/stable-diffusion-cpp/audit-sd-cpp-residency/ Delivery: #77
Repo · NVIDIA/TensorRT-LLM @ f33e0e5fd	SCOPED · host overhead verdict + line-profiling loop	skill_library/hardware/cuda/host-overhead-profiling/; skill_library/hardware/host/host-overhead-optimization/ Consolidation: navigation delivered in merged #142; methodology and PD/KV source evidence stay distinct. No validated shared assumptions are established. Delivery: #12 (initial library import)
Repo · Ascend/MindIE-SD @ 8481d39	SCOPED · host-opt async/pattern sections	skill_library/methods/weight-loading-optimization/ Delivery: #12 (initial library import)
Repo · Ascend/pytorch (TorchNPU) @ 67511c53cd4c	SCOPED · owned/external pinned buffers, host-specific counters, default/expandable paths and unresolved registration-version warning	skill_library/hardware/ascend/torchnpu-pinned-host-memory-lifecycle/ Delivery: #12 (initial library import)
Repo · foundation-model-stack/fastsafetensors @ 219e256	SCOPED · pinned bounce-buffer staging, reader NUMA binding	skill_library/methods/weight-loading-optimization/ Delivery: #12 (initial library); #96 (TP-transfer increment)
Repo · LMCache/LMCache @ c3b875a	SCOPED · L1 DRAM tier, admission dial, watermark eviction, multi-L1 overflow	skill_library/repositories/lmcache/kv-tier-management/ Delivery: #12 (initial library import)
Repo · sgl-project/sglang @ df13d5dce31a	SCOPED · host request/job/scheduler-object and finalizer lifetimes only; device memory is not RAM	skill_library/repositories/sglang/audit-sglang-diffusion-lifetimes/ Consolidation: three-scope navigation delivered in merged #136; engine, HiCache and diffusion preserve separate pins/evidence. No cross-pin/backend compatibility or new Network qualification. Delivery: #79
Repo · ggml-org/llama.cpp @ 3e390af	SCOPED · mmap/mlock residency, lazy tensors, host weight placement (--cpu-moe family), KV host offload	skill_library/methods/weight-loading-optimization/ Consolidation: navigation reconciled in merged #124; loading/host placement and runtime keep separate pins, prerequisites and owners. No cross-pin/backend qualification. Delivery: #12 (initial library import)
Repo · ggml-org/llama.cpp @ 707616feb5f3	SCOPED · host-owned task/result waiters and slot prompt-token bookkeeping; stateful retention versus cancellation; device KV is not host RAM	skill_library/repositories/llama-cpp/audit-llama-runtime-state/ Delivery: #71
Repo · sysstat/sysstat @ efd0c445	SCOPED · per-process I/O/fault attribution caveats	skill_library/hardware/host/storage-io-measurement/ Delivery: #12 (initial library import)
Docs · Linux cgroup v2 docs — kernel.org live 2026-10-15 (revision unidentified) + v6.12 cgroup-v2.rst (reproducible baseline)	SCOPED · memory pressure/events triage, high/low/max semantics	skill_library/hardware/host/cgroup-resource-controls/ Delivery: #12 (initial library import)
Repo · pytorch/pytorch (DCP async-save only) @ 7d239835e773; live async recipe retrieved 2026-10-15	SCOPED · separate staging/upload futures, cached pinned CPU buffers, reuse only after prior upload, CPU/RAM contention	skill_library/frameworks/pytorch/dcp-async-save-performance/ Consolidation: scope map delivered in merged #144; compiler and DCP pins/lifetimes remain distinct. No tested combined stack is established. Delivery: #12 (initial library import)
Repo · NVIDIA DALI @ 6fe47c48d3b0	SCOPED · queue/buffer sizing and per-output real/reserved memory statistics; legacy stats require exec_dynamic=False	skill_library/methods/training-input-pipeline/ Delivery: #12 (initial library import)
Repo · Mosaic Streaming @ d99bf9c7cdf2	SCOPED · per-worker sample look-ahead versus memory/cache budget	skill_library/methods/training-input-pipeline/ Delivery: #12 (initial library import)
Repo · FFCV @ 3a12966b3afe	SCOPED · dataset-fits-RAM versus process cache, batches-ahead buffers	skill_library/methods/training-input-pipeline/ Delivery: #12 (initial library import)
Repo · DLRover FlashCheckpoint @ e35b1ad90e29	SCOPED · shared-memory readiness and consistent positive-step read gate	skill_library/methods/checkpoint-recovery-consistency/ Delivery: #12 (initial library); #95 (persistence/replica increment)
Repo · NVIDIA Resiliency Extension @ 49af0b03f145	SCOPED · snapshot/replica lifetime constraints	skill_library/methods/checkpoint-recovery-consistency/ Delivery: #12 (initial library import)
Repo · DeepSpeed ZeRO-Infinity @ 5f266e26d216	SCOPED · CPU optimizer budget, pinned RAM, element-count residency and gated core share	skill_library/methods/training-resource-budget/ Delivery: #12 (initial library import)
Repo · Ascend MindSpeed adaptive memory — core_r0.8.0, page history 53193db8 (full pin and license unresolved)	SCOPED · CPU swap and allocator/affinity prerequisites	skill_library/methods/training-resource-budget/ Delivery: #12 (initial library import)
Repo · ml-explore/mlx @ bff10220c742	SCOPED · Graph detachment versus in-flight data retention; not allocator eviction	skill_library/frameworks/mlx/lazy-evaluation-stream-lifetimes/ Delivery: #72
Repo · ml-explore/mlx-lm @ ab7f7048dc34	SCOPED · Request/cache row alignment, prefix identity, eviction and cooperative stop	skill_library/repositories/mlx-lm/generation-cache-lifecycle/ Delivery: #74
Repo · huggingface/diffusers @ efe46a5b177a; aea7b70be0f5	SCOPED · Offload staging plus separately pinned invocation/scheduler ownership	knowledge_packs/dependency/diffusers/cuda/diffusion/; skill_library/frameworks/diffusers/diffusers-invocation-state/ Consolidation: navigation delivered in merged #134; reviewed scope remains separate offload/invocation owners. No runtime qualification is established. Delivery: #11 (offload); #76 (invocation)
Repo · deepseek-ai/DeepInfer @ 8b16149741a4	SCOPED · Host scheduling overlap, retained state/repair/drain; retained query versus KV ownership	skill_library/hardware/host/host-overhead-optimization/; skill_library/repositories/deepinfer/ Delivery: #68
Repo · Niko1221/Strata @ 296b59d2f73c	SCOPED · CPU expert work/residency and CPU/PCIe balance; scoped engine procedures	skill_library/repositories/strata/; repo_packs/strata/knowledge/common/llm/ Delivery: #9 (initial engine guides)
Repo · sgl-project/sglang @ f7ad4345b86d	SCOPED · Host prefix residency and transfer-versus-recompute tradeoff; imported	knowledge_packs/peer/sglang/cuda/llm/ Delivery: #11 (HiCache source-card import)
Repo · vllm-project/vllm-omni @ 18c3bd5d2d43	SCOPED · Correlated waiter, wait-only timeout, partial-output and queued-release lifetimes	skill_library/repositories/vllm-omni/sub-skills/omni-validation/ Consolidation: navigation delivered in merged #139; engine and abort/reclamation guidance retain separate scope. No runtime qualification is established. Delivery: #78
Repo · awesome-ascend-skills @ 75ec89c	TODO · Partial: model-infer-kvcache located, not extracted	Retain existing procedures; distill only missing host-KV scope. Delivery: #108 (assessment; host-KV remains deferred)
Repo · vllm-ascend @ bc8115ed / 8a8e1e05ecb64a30804c99c8f362909f9a0cb5bf	SCOPED · bounded runtime host-offload candidate delivered in #123; broader historical gap not automatically completed	Merged #123 uses the existing host-memory owner for single CPU-tier KV-offload descriptor lifetime. Native NPU DMA, shutdown safety and external vLLM ABI remain unqualified.
Paper · FlexGen / ZeRO-Infinity	TODO · Recorded foundation input missing; proposed host staging/residency budgets	Recover exact paper/readsets; distinguish CPU RAM from HBM and reuse canonical paper owners.
Repo · SGLang-Omni @ 2cfb3b67becb	SCOPED · stage DAG, host process/state ownership, preparation/abort lifecycle; process transport is not Network coverage	skill_library/repositories/sglang-omni/; repo_packs/sglang_omni/knowledge/common/; consolidation not assessed. Delivery: #9 (initial engine guides)

3. SSD / Storage

Type / source and recorded pin	Scope / extraction state	Output / next action
Repo · NVIDIA/TensorRT-LLM @ f33e0e5fd	SCOPED · weight-movement/overlap + KV host-cache patterns	skill_library/methods/weight-loading-optimization/ Consolidation: navigation delivered in merged #142; methodology and PD/KV source evidence stay distinct. No validated shared assumptions are established. Delivery: #12 (initial library import)
Repo · Ascend/MindIE-SD @ 8481d39	SCOPED · load-path judgment, fixed-overhead accounting, offload tiers, weights-prep	skill_library/methods/weight-loading-optimization/ Delivery: #12 (initial library import)
Repo · ascend-ai-coding/awesome-ascend-skills @ 75ec89c	SCOPED · device-side prefetch + diffusers CPU-offload playbook	skill_library/methods/weight-loading-optimization/ Delivery: #12 (initial library import)
Repo · vllm-project/vllm-ascend @ bc8115ed	SCOPED · FP8 paired-tensor loading pattern	skill_library/hardware/ascend/ascend-model-adaptation/ Consolidation: version/responsibility navigation delivered in merged #130. Historical adapter documentation and ACLGraph runtime remain separate; no new OS-control or compatibility evidence. Delivery: #12 (initial library import)
Repo · mindie-agent/mindie-agent @ 6a81b41	SCOPED · logical-vs-physical I/O + cache-flush measurement caveat	skill_library/methods/weight-loading-optimization/ Delivery: #12 (initial library import)
Repo · foundation-model-stack/fastsafetensors @ 219e256	SCOPED · loading I/O mechanics (no-mmap design, bounce reader, GDS, readiness overlap, partial reads, static planner) + measurement protocol	skill_library/methods/weight-loading-optimization/ Delivery: #12 (initial library); #96 (TP-transfer increment)
Repo · LMCache/LMCache @ c3b875a	SCOPED · L2 local adapters (fs/nixl/raw_block/dax), serde capacity lever, persist/recover	skill_library/repositories/lmcache/kv-tier-management/ Delivery: #12 (initial library import)
Repo · ggml-org/llama.cpp @ 3e390af	SCOPED · loading modes (mmap/mlock/dio/lazy), page-cache interactions, fit-probe bypass	skill_library/methods/weight-loading-optimization/ Consolidation: navigation reconciled in merged #124; loading/host placement and runtime keep separate pins, prerequisites and owners. No cross-pin/backend qualification. Delivery: #12 (initial library import)
Repo · axboe/fio @ acb7a302	SCOPED · workload matching, requested-vs-achieved depth, cache-state ledger, run-shape classification, output schema/units	skill_library/hardware/host/storage-io-measurement/ Delivery: #12 (initial library import)
Repo · sysstat/sysstat @ efd0c445	SCOPED · per-device attribution, saturation guards (%util/%iowait)	skill_library/hardware/host/storage-io-measurement/ Delivery: #12 (initial library import)
Docs · Linux cgroup v2 docs — kernel.org live 2026-10-15 (revision unidentified) + v6.12 cgroup-v2.rst (reproducible baseline)	SCOPED · io.max caps gated on confirmed block-backed traffic	skill_library/hardware/host/cgroup-resource-controls/ Delivery: #12 (initial library import)
Repo · pytorch/pytorch (DCP async-save only) @ 7d239835e773; live async recipe retrieved 2026-10-15	SCOPED · bounded in-flight saves and upload timing, writer-specific completion/durability distinction; no distributed restore audit	skill_library/frameworks/pytorch/dcp-async-save-performance/ Consolidation: scope map delivered in merged #144; compiler and DCP pins/lifetimes remain distinct. No tested combined stack is established. Delivery: #12 (initial library import)
Repo · Mosaic Streaming @ d99bf9c7cdf2	SCOPED · raw+compressed shard admission, cache minimum and LRU/redownload trade-off	skill_library/methods/training-input-pipeline/ Delivery: #12 (initial library import)
Repo · FFCV @ 3a12966b3afe	SCOPED · read-only mmap and documented quasi-random read-locality choice; .beton required	skill_library/methods/training-input-pipeline/ Delivery: #12 (initial library import)
Repo · NVIDIA Resiliency Extension @ 49af0b03f145	SCOPED · clean/dirty publication, complete-rank generation selection and incomplete-save rejection; no power-loss proof	skill_library/methods/checkpoint-recovery-consistency/ Delivery: #12 (initial library import)
Repo · DeepSpeed ZeRO-Infinity @ 5f266e26d216	SCOPED · NVMe buffer pools and tiled read/compute/write overlap intent; buffer-size units unresolved in approved excerpt	skill_library/methods/training-resource-budget/ Delivery: #12 (initial library import)
Repo · sgl-project/sglang @ f7ad4345b86d	SCOPED · HiCache L3 sharing, prefetch, layout and eviction; storage backend qualification pending	knowledge_packs/peer/sglang/cuda/llm/ Consolidation: three-scope navigation delivered in merged #136; engine, HiCache and diffusion preserve separate pins/evidence. No cross-pin/backend compatibility or new Network qualification. Delivery: #11 (HiCache source-card import)
Repo · DLRover FlashCheckpoint @ e35b1ad90e29	SCOPED · persistence/recovery boundaries; save completion is not power-loss durability	skill_library/methods/checkpoint-recovery-consistency/; retain source and failure-domain limits Delivery: #12 (initial library); #95 (persistence/replica increment)
Repo · stable-diffusion.cpp @ b13a48219d5c	TODO · loader/storage extension submitted in open #122; not yet main coverage	Residency skill exists. Open #122 adds bounded non-ZIP safetensors loader/storage guidance in the existing owner; native parser, storage and image results remain unrun. Delivery: #77 (main residency slice)
Paper · ZeRO-Infinity / FlexGen	TODO · Recorded foundation input missing; proposed SSD/offload slice	Recover/check exact versions/readsets; existing offload-config guidance is not a substitute.
Paper · Mooncake	TODO · Recorded foundation input missing; storage-tier applicability unassessed	Recover readset first; qualify any storage mechanism, do not assume KV pools use SSDs.

4. Communication / Network

Only actual network/collective/remote-transfer evidence belongs here. H2D/D2H, local IPC, stream events and engine queues do not create Network coverage.

Type / source and recorded pin	Scope / extraction state	Output / next action
Repo · Ascend/MindIE-SD @ 8481d39	SCOPED · USP strategy, masking formula, topology rules	skill_library/hardware/ascend/ascend-distributed-parallel/ Delivery: #12 (initial library import)
Repo · cann/oam-tools @ 8440935f (!302 import; later master f3cf4115 labeled [V2], never merged into !302 claims)	SCOPED · exposed-communication diagnosis, wait pollution, evidence-gated compute/comm overlap	skill_library/hardware/ascend/cann-attribution-overlap/ Delivery: #12 (initial library import)
Repo · Mosaic Streaming @ d99bf9c7cdf2	SCOPED · remote dataset download look-ahead/cache-churn decisions only; no collective optimization	skill_library/methods/training-input-pipeline/ Delivery: #12 (initial library import)
Repo · DLRover FlashCheckpoint @ e35b1ad90e29	SCOPED · replica transport/rank mapping and availability limits; no distributed runtime qualification	skill_library/methods/checkpoint-recovery-consistency/ Delivery: #12 (initial library); #95 (persistence/replica increment)
Repo · NVIDIA Resiliency Extension @ 49af0b03f145	SCOPED · all-rank availability/finalization and replica failure domains, not transport throughput	skill_library/methods/checkpoint-recovery-consistency/ Delivery: #12 (initial library import)
Repo · NVIDIA/TensorRT-LLM @ c9b18eefa731	SCOPED · Disaggregated prefill/decode KV transfer; interconnect/layout/queue costs; imported	knowledge_packs/peer/tensorrt_llm/cuda/llm/ Consolidation: navigation delivered in merged #142; methodology and PD/KV source evidence stay distinct. No validated shared assumptions are established. Delivery: #11 (source-card import)
Repo · xdit-project/xDiT @ f7c16aa95837	SCOPED · Topology-dependent USP/CFG/PipeFusion and freshness tradeoffs; imported	knowledge_packs/peer/xdit/cuda/diffusion/ Delivery: #11 (source-card import)
Repo · awesome-ascend-skills @ 75ec89c	SCOPED · communication-test wrapper selection, byte/timing/oracle meaning and rank-map boundaries; no native transport audit	skill_library/hardware/ascend/ascend-distributed-parallel/; source-derived candidate Delivery: #108
Repo · fastsafetensors @ 219e256	SCOPED · TP broadcast/scatter/coalesced-transfer ownership and completion boundaries; source-only, no measured transport gain	skill_library/methods/weight-loading-optimization/ Delivery: #12 (initial library); #96 (TP-transfer increment)
Repo · vllm-project/afd-plugin @ df8d6a0c403c3ca1b4ed9cf1687e7204908bebec	SCOPED · CUDA P2pNccl eager dense-FFN rank namespaces and per-stage/per-peer payload ownership; Attention-side gating disabled	skill_library/repositories/afd-plugin/audit-afd-p2p-exchanges/ Delivery: #105 (historical intake); merged #120 (bounded source candidate). Native communication/performance and NPU CAM remain unqualified.
Paper · DistServe / Mooncake	TODO · Recorded foundation input missing; proposed disaggregation/KV-transfer slices	Recover/check versions and readsets; transport discovery is not paper mining.

5. OS / System

Process placement, resource controls and documented host contention are distinguished from engine scheduling; a runtime scheduler is not OS-priority control.

Type / source and recorded pin	Scope / extraction state	Output / next action
Repo · NVIDIA/TensorRT-LLM @ f33e0e5fd	SCOPED · CPU affinity rule + env preflight	skill_library/hardware/host/os-process-placement/ Consolidation: navigation delivered in merged #142; methodology and PD/KV source evidence stay distinct. No validated shared assumptions are established. Delivery: #12 (initial library import)
Repo · Ascend/MindIE-SD @ 8481d39	SCOPED · device health, card-group contracts, measurement discipline	skill_library/hardware/ascend/ascend-device-health/ Delivery: #12 (initial library import)
Repo · ascend-ai-coding/awesome-ascend-skills @ 75ec89c	SCOPED · env checks, device-busy preflight, thread/env tables	skill_library/hardware/host/os-process-placement/ Delivery: #12 (initial library import)
Repo · vllm-project/vllm-ascend @ bc8115ed	SCOPED · Dockerfile thread/env ship defaults	skill_library/hardware/host/os-process-placement/ Consolidation: version/responsibility navigation delivered in merged #130. Historical adapter documentation and ACLGraph runtime remain separate; no new OS-control or compatibility evidence. Delivery: #12 (initial library import)
Repo · foundation-model-stack/fastsafetensors @ 219e256	SCOPED · reader worker NUMA binding + O_DIRECT per-filesystem policy	skill_library/methods/weight-loading-optimization/ Delivery: #12 (initial library); #96 (TP-transfer increment)
Repo · LMCache/LMCache @ c3b875a	SCOPED · MP process/thread concurrency model (isolation, client-affinity pools, TTL locks)	skill_library/repositories/lmcache/kv-tier-management/ Delivery: #12 (initial library import)
Repo · ggml-org/llama.cpp @ 3e390af	SCOPED · NUMA modes (distribute/isolate/numactl), gen-vs-batch thread/affinity splits, numa_balancing rule	skill_library/hardware/host/os-process-placement/ Consolidation: navigation reconciled in merged #124; loading/host placement and runtime keep separate pins, prerequisites and owners. No cross-pin/backend qualification. Delivery: #12 (initial library import)
Repo · sysstat/sysstat @ efd0c445	SCOPED · CPU-side scheduling-pressure vs disk-hypothesis rules (%wait, context switches)	skill_library/hardware/host/storage-io-measurement/ Delivery: #12 (initial library import)
Docs · Linux cgroup v2 docs — kernel.org live 2026-10-15 (revision unidentified) + v6.12 cgroup-v2.rst (reproducible baseline)	SCOPED · cpu.max/weight quota-vs-contention (cpu.stat vs .local), requested/effective cpuset, hierarchy/delegation preflight	skill_library/hardware/host/cgroup-resource-controls/ Delivery: #12 (initial library import)
Repo · cann/oam-tools @ 8440935f (!302 import; later master f3cf4115 labeled [V2], never merged into !302 claims)	SCOPED · host-event coverage of device bubbles, uncertainty-aware	skill_library/hardware/ascend/cann-attribution-overlap/ Delivery: #12 (initial library import)
Repo · pytorch/pytorch (DCP async-save only) @ 7d239835e773; live async recipe retrieved 2026-10-15	SCOPED · background-worker contention and caller-owned stager cleanup; no OS scheduling changes	skill_library/frameworks/pytorch/dcp-async-save-performance/ Consolidation: scope map delivered in merged #144; compiler and DCP pins/lifetimes remain distinct. No tested combined stack is established. Delivery: #12 (initial library import)
Repo · FFCV @ 3a12966b3afe	SCOPED · cache ownership and traversal configuration; distributed QUASI_RANDOM excluded, current compatibility unverified	skill_library/methods/training-input-pipeline/ Delivery: #12 (initial library import)
Repo · DLRover FlashCheckpoint @ e35b1ad90e29	SCOPED · worker versus helper/host lifetime; fallback daemon cannot persist after parent exits	skill_library/methods/checkpoint-recovery-consistency/ Delivery: #12 (initial library); #95 (persistence/replica increment)
Repo · DeepSpeed ZeRO-Infinity @ 5f266e26d216	SCOPED · CPU resource-budget qualification; no privileged changes	skill_library/methods/training-resource-budget/ Delivery: #12 (initial library import)
Repo · Ascend MindSpeed adaptive memory — core_r0.8.0, page history 53193db8 (full pin and license unresolved)	SCOPED · incompatible allocator features and CPU-affinity prerequisites; no current compatibility claim	skill_library/methods/training-resource-budget/ Delivery: #12 (initial library import)
Repo · vllm-ascend @ bc8115ed / 8a8e1e05ecb6	SCOPED · version/responsibility navigation reconciled	Merged #130 preserves older Docker/thread-env scope; ACLGraph runtime does not establish new OS controls.
Repo · MindSpeed adaptive memory @ core_r0.8.0 / 53193db8	TODO · Historical full pin/license and current applicability unresolved	Resolve provenance and CPU-affinity/allocator assumptions before updating the historical case.

Additional post-snapshot source delivery: #110 adds skill_library/repositories/h3-c/audit-h3-runtime-residency/, with Accelerator, RAM and SSD scopes recorded in the current matrices. This is public repository-source knowledge, not runtime qualification.

Link to Skill Eval

#88 owns evaluation tooling; this tracker owns source coverage and consolidation work. A delivery that adds a skill links its absent/new evaluation plan; a skill update or behavior-changing consolidation links its old/new plan. Until those capabilities ship, record manual evidence or not evaluated honestly. Pure navigation/provenance edits need deterministic integrity checks, not automatic paid agent runs. Missing tests remain uncovered, not a pass; source-grounded delivery and measured efficacy stay separate.

Maintenance and acceptance

• SCOPED means a bounded output exists; TODO records missing, unknown, deferred or consolidation work. Proposed TODO resource mappings do not earn coverage. A multi-version repo is not automatically conflicting.
• Update affected rows with pinned readset, canonical output, evidence/PR link, remaining limits and consolidation state. Do not count navigation or generated abstracts as additional extraction.
• Consolidate by mechanism: merge duplicates, preserve original pins/evidence, record semantic changes/applicability, explicitly deprecate superseded advice and retain unresolved conflicts. Different topics keep modular owners and unified navigation, not one monolithic file.
• Check affected indexes/readers and links before marking reconciliation complete. Consolidation never promotes a candidate or inherits runtime validation.
• Paper-input caveat: the Reference Matrix records an earlier foundation archive containing ZeRO-Infinity, FlexGen, Sarathi-Serve, DistServe, PagedAttention and Mooncake as missing from the supplied corpus; these extractions were not imported/countable there. Main integration notes positively record historical extraction and prior scheduling/KV artifacts, but the original archive identity, exact versions/readsets and recoverability remain unverified. Recover/check existing extraction before re-distilling; docs: guide stalled optimization from whole-system bottlenecks #156's proposed new slices are separate, not a recovered original archive. Rows above are proposed gap mappings, not new coverage.
• Already extracted Vidur, Orca, FastServe, Llumnix, LoongServe, Checkmate, OLLA, Alpa, Zero Bubble and FlashAttention 1–3 remain governed by the Reference Matrix; pending runtime validation does not make them unextracted.
• Methodology-only sources remain in Reference Matrix §2, rather than being forced into hardware-resource coverage. The tracker stays open while maintained; selected intake and experiments need agreed scope.