Repository navigation
[Tracking] SearchSpace distillation coverage and knowledge consolidation #91
Description
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 — 2026-10-25
Checked against main 65147a03e465 and the current PR metadata. The 2026-10-22 inventory and original source pins below remain historical baseline; only the listed rows/statuses are refreshed. 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 remain delivered | Candidate/source-specific limits remain; host-KV scope deferred by #108 is still not extracted |
| Multi-pin navigation in main | llama.cpp #124, vLLM-Ascend #130, SGLang #136 | Separate pins/owners/evidence retained; no cross-pin/backend compatibility claim |
| Multi-pin navigation proposed, not merged | Diffusers #134, vLLM #138, vLLM-Omni #139, TensorRT-LLM #142, PyTorch compiler/DCP #144 | Keep existing main owners; review/LGTM does not mark these contributions integrated |
| Bounded additions/extensions proposed, not merged | RLT #121; SD.cpp loader #122; Ascend host-offload #123; Lithos #117 via #115; MLX cards #119 via #38; KDA KernelWiki #127 via #126 | Do not count unmerged output as main coverage or infer native/agent efficacy |
| Knowledge access and evaluation | #129 whole-resource collect is merged; #88/#101 and open #118 own evaluation automation; #128 owns its separate ongoing exposure comparison | Software delivery and owner-reported ongoing trials do not establish task benefit or final acceptance |
| New source / non-distillation proposal | NVIDIA/flare-local-inference — #147 | Intake/proposal only: no imported skill, backend adapter, runtime qualification or resource coverage credited |
Explicit distillation tasks added on 2026-10-25: #148 Flare (preparation contracts and complete session-state lifetimes; related RFC #147) and #149 LoopX (goal authority and quota decision boundaries; related RFC #137). Both are native sub-issues of this tracker, pending source delivery; no new extraction or efficacy credit yet.
The new Flare proposal is a separate engineering opportunity, not reopening the accepted historical #48 repository batch. Keep this tracker open. 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 proposed in open #142; methodology and PD/KV source evidence stay distinct. Not yet main delivery or validated shared assumptions. 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-skillsDelivery: #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 proposed in open #144; compiler and DCP pins/lifetimes remain distinct. Not yet main delivery; no tested combined stack. 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 proposed in open #134; reviewed scope remains separate offload/invocation owners. Not yet main delivery or runtime qualification. 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 proposed in open #139; engine and abort/reclamation guidance retain separate scope. Not yet main delivery or runtime qualification. 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 coverageDelivery: #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 boundariesDelivery: #104 (intake only) |
| Paper · PagedAttention / Sarathi-Serve | TODO · Recorded foundation input missing; proposed KV allocation/admission and prefill/decode scheduling slices | Recover/check existing readsets and exact paper versions 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 proposed in open #142; methodology and PD/KV source evidence stay distinct. Not yet main delivery or validated shared assumptions. 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 proposed in open #144; compiler and DCP pins/lifetimes remain distinct. Not yet main delivery; no tested combined stack. 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 proposed in open #134; reviewed scope remains separate offload/invocation owners. Not yet main delivery or runtime qualification. 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 proposed in open #139; engine and abort/reclamation guidance retain separate scope. Not yet main delivery or runtime qualification. 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 / proposed 8a8e1e05ecb64a30804c99c8f362909f9a0cb5bf |
TODO · historical doc-only gap; bounded runtime host-offload candidate submitted, not yet main coverage | Open #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 proposed in open #142; methodology and PD/KV source evidence stay distinct. Not yet main delivery or validated shared assumptions. 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 proposed in open #144; compiler and DCP pins/lifetimes remain distinct. Not yet main delivery; no tested combined stack. 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 limitsDelivery: #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 proposed in open #142; methodology and PD/KV source evidence stay distinct. Not yet main delivery or validated shared assumptions. 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 candidateDelivery: #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 proposed in open #142; methodology and PD/KV source evidence stay distinct. Not yet main delivery or validated shared assumptions. 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 proposed in open #144; compiler and DCP pins/lifetimes remain distinct. Not yet main delivery; no tested combined stack. 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. This does not prove nobody ever read or distilled them. Recover/check existing extraction before re-distilling. 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.
Activity
Dong1017 commentedon Oct 22, 2026
Suggested execution order
Prioritize bounded gaps with existing source evidence and canonical owners before descriptor-only intake. These are recommendations, not blanket execution or paid-evaluation authorization.
Link with #88 without another registry
flowchart LR A["Tracker gap or consolidation"] --> B["Bounded delivery PR"] B --> C["Skill Eval plan"] C --> D["Authorized run and evidence"] D --> E["Update tracker row"]not evaluated/uncovered. Offline Fake-client tests validate the evaluation tool, not skill efficacy.Dong1017 commentedon Oct 23, 2026
Reopened at the owner's request. This is a living SearchSpace coverage/consolidation tracker, not the acceptance issue for a single distillation PR. A merged scoped delivery updates its relevant row; it does not complete the whole tracker. Please reference this issue with
Refs #91/Related to #91, rather than closing keywords. Pinning is for visibility, not protection against closure.docs: clarify vLLM-Ascend guidance versions and responsibilities (#130)