You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit efa9ba4
Browse filesBrowse the repository at this point in the historyBrowse files
Add a bounded source-derived procedure and offline declaration preflight for effective configuration, checkpoint transformations and VSA fallback. Preserve native and model validation gaps; integrate packaged discovery and assertion-backed cases.
Agent-assisted-by: Codex / gpt-6.1-sol [effort=xhigh] (author)
Signed-off-by: Codex <codex@openai.com>
repositories/<repo>/ Engine entry, sub-skills and focused complete skills
@@ -171,18 +171,18 @@ Complete procedures, sidecars, references and cases are packaged, with 13 Ascend
171
171
repositories, methods or workflows. The 21 engine and research-router skills
172
172
use the same ownership layout under `repositories/<engine>/` and
173
173
`workflows/inference-engine-research/`, including their sub-skills and references.
174
-
The library contains 115`SKILL.md` files, all in English.
174
+
The library contains 116`SKILL.md` files, all in English.
175
175
The collection overview, import notes and source manifest belong to the research
176
176
workflow's references; Strata design-space analysis belongs to Strata. The library
177
177
root contains only navigation and the registry.
178
178
179
179
The [generated catalog](../src/inferrsi/data/repo_packs/inferrsi/knowledge/README.md)
180
-
projects the 71 registered skills into the contribution catalog. Its records
180
+
projects the 72 registered skills into the contribution catalog. Its records
181
181
use `inferrsi.<backend>.skills.<name_with_underscores>` IDs, with `ascend` and
182
182
`cuda` for scopes that depend on those backends and `common` for the remaining skills.
183
183
Each `skill_documents.original` points to the full packaged procedure. The
184
184
`skills` CLI retains the original hyphenated skill IDs and defaults to this
185
-
71-ID registry; engine procedures remain available through library navigation
185
+
72-ID registry; engine procedures remain available through library navigation
186
186
and catalog document pointers.
187
187
188
188
The registered [h3.c runtime audit](../src/inferrsi/data/skill_library/repositories/h3-c/audit-h3-runtime-residency/SKILL.md)
@@ -271,3 +271,5 @@ The registered [mistral-rs candidate](../src/inferrsi/data/skill_library/reposit
271
271
The [mlx-swift-lm candidate](../src/inferrsi/data/skill_library/repositories/mlx-swift-lm/SKILL.md) owns selected Swift generation-lifetime and recurrent native/ops contracts, with source/observation limits preserved. No Swift compile, original native/host-control test, actual Metal/model/quality/service run is implied by its helpers or packaged discovery.
272
272
273
273
The [mlc-llm candidate](../src/inferrsi/data/skill_library/repositories/mlc-llm/SKILL.md) separates artifact metadata, supported consumers, Apple targets, pinned TVM runnable/runtime boundaries and compiled/runtime KV configuration. Canonical hashes, serializable fallback or estimated limits are not native/model/quality/fit/service evidence.
274
+
275
+
The registered [h3-metal FastH3 audit](../src/inferrsi/data/skill_library/repositories/h3-metal/audit-h3-metal-fasth3-semantics/SKILL.md) supplements earlier h3.c residency with fork-specific effective configuration, tensor/schedule semantics and VSA eligibility/fallback. Its offline declaration arithmetic and label probes do not qualify native conversion, sparse attention, media, memory fit or performance; required Apple/model evidence remains blocked/not_run.
Copy file name to clipboardExpand all lines: docs/reference-matrix.md
+1Lines changed: 1 addition & 0 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -44,6 +44,7 @@ Reading rules:
44
44
|---|---|---|---|---|---|---|
45
45
| NVlabs/kda @ `acdd79343c5e9d0c7cb96289f26fd3f1051ce785` → mit-han-lab/KernelWiki gitlink @ `76d27b56f804e7e7295d4c570e1e5d7eef4b0a75` | Direct bounded distillation, 2026-10-23; cutoff 2026-04-27. All 48 synthesis headers and 2,217 source headers screened; 171 SM90 source headers retained, selected bodies and independently matched DeepGEMM/CUTLASS code plus versioned NVIDIA sections inspected. [Readset/coverage/license qualifications](../src/inferrsi/data/skill_library/hardware/cuda/reason-hopper-kernels/references/sources.md); remaining PR bodies/algorithms unmined; KW pin has no root LICENSE, KDA MIT/sanitization assertions qualified | mined (conditional) — Hopper TMA/mbarrier/WGMMA ownership, resource/layout/pipeline/software scheduling and selected FP8 scaling → [reason-hopper-kernels](../src/inferrsi/data/skill_library/hardware/cuda/reason-hopper-kernels/SKILL.md); SM100 recipes and unsafe snippet generalizations excluded; native/precision/performance not_run | out-of-scope — device registers/SMEM/HBM are not host RAM | out-of-scope — no storage implementation extracted | out-of-scope — CTA/cluster synchronization is not network transport | out-of-scope — tile scheduling is not OS scheduling |
46
46
| antirez/h3.c @ `8974cc055ea9c02fcd14cc27dfda3e1027c05153` (MIT; retained ccv BSD-3-Clause notice) | Direct bounded source distillation, 2026-10-23; 34 selected file bytes matched to pinned Git blobs; [readset/architecture](../src/inferrsi/data/skill_library/repositories/h3-c/audit-h3-runtime-residency/references/source-notes.md). Loader, generation/cache, TE/vision/input-output VAEs, AdaLN, selected Metal/int8/SSD paths and native test assertions; full kernels/driver/media internals unread. Prior model-header observations reused, not rerun | mined — joint audio/video schedule and row-map dependencies, BF16 rounding/fusion versus numerical approximation, actual TensorOps dispatch → [audit-h3-runtime-residency](../src/inferrsi/data/skill_library/repositories/h3-c/audit-h3-runtime-residency/SKILL.md); native/quality/performance not_run | mined — shared/file-backed weights, Qwen prefetch ring, phase/cache/preview overlap and complete host output; tensor counters are not physical footprint or 64 GiB qualification | mined — exact raw BF16/F32 schema, direct/mapped reads and two-slot BF16 matrix streaming with completion/join boundaries; uncached-read request is not measured cache bypass or useful overlap | out-of-scope — local GPU/I/O ordering is not network transport | out-of-scope — no OS scheduling/control implementation extracted |
47
+
| zuiho-kai/h3-metal @ `4ba47a3eabdbebf801ac6664b31f8ec168bc57a8` (inspected MIT; retained BSD-3-Clause notice; machine resolver unavailable) | Bounded source distillation, 2026-10-28; 21 identity-checked files with selected ranges; [scope/readset](../src/inferrsi/data/skill_library/repositories/h3-metal/audit-h3-metal-fasth3-semantics/references/source-notes.md). Fork-specific FastH3 params, tensor/schedule/RoPE and VSA dispatch, test/script limits; full engine/external FastVideo/model surface not mined | mined — requested versus effective controls, ambient conflicts, head-interleaved QKV/FC1 swap/BF16 schedule, real partial VSA tiles/double top-k and dense-plus-coarse fallback → [audit-h3-metal-fasth3-semantics](../src/inferrsi/data/skill_library/repositories/h3-metal/audit-h3-metal-fasth3-semantics/SKILL.md); native/numerical/quality/performance not_run | out-of-scope — tile arithmetic is not measured host footprint or memory-fit qualification | out-of-scope — FastH3 disables requested SSD parameter; no new storage optimization or payload validation | out-of-scope — local attention/dispatch is not network transport | out-of-scope — runtime selection is not OS scheduling |
47
48
| Comfy-Org/ComfyUI @ `b1a247c9cf9dbfda4b2c0a0c1d50950667eeaa0c` (GNU GPL v3) | Source-reading archive, inspected 2026-10-21; bounded execution/cache/graph/model-management/static-dynamic-patcher/sampler paths and selected tests; [readset](../src/inferrsi/data/skill_library/repositories/comfy-ui/references/repo-provenance.md); not full-model/custom-node/provider-native coverage | mined — managed weight/patch identity, partial unload versus detach and dynamic-backend boundary → [comfy-ui](../src/inferrsi/data/skill_library/repositories/comfy-ui/SKILL.md); device behavior unverified | mined — output versus object identity, ancestor fingerprints, lazy consumer references and retention versus validity; host/pin ownership | out-of-scope — no storage implementation extracted | out-of-scope — no transport implementation extracted | out-of-scope — workflow scheduling is not OS scheduling |
48
49
| leejet/stable-diffusion.cpp @ `b13a48219d5cdf408edcb93da6cd69b612ed03d3` (MIT) | Supplied archive, inspected 2026-10-22: 18 bounded paths plus separately identity-checked README (19 blobs); historical receipt preserved. 2026-10-23 extension: 12 selected C++ blobs plus retained MIT text identity-checked, with line-scoped catalog/non-ZIP safetensors read/staging/mapping and immediate consumer cleanup; [readsets](../src/inferrsi/data/skill_library/repositories/stable-diffusion-cpp/audit-sd-cpp-residency/references/source-notes.md). File identity is not full-file review; archive/submodules, model kernels and other-format implementations remain outside the bounded audit | mined — native weight/runtime capacity, protected segment/prefetch lifetime → [audit-sd-cpp-residency](../src/inferrsi/data/skill_library/repositories/stable-diffusion-cpp/audit-sd-cpp-residency/SKILL.md); native/backend behavior unverified | mined — parameter storage versus compute staging and retained runtime ownership; loader host read/conversion buffers versus mapped owners; host peaks unmeasured | mined — bounded catalog stamps versus content identity, range admission versus actual reads, mapping/shared-owner and newly created block cleanup → [loader/storage boundaries](../src/inferrsi/data/skill_library/repositories/stable-diffusion-cpp/audit-sd-cpp-residency/references/loader-storage-boundaries.md); full other-format parsing, native failure/SSD/performance qualification remain open | out-of-scope — device copies are not network transport | out-of-scope — runner scheduling is not OS scheduling |
Copy file name to clipboardExpand all lines: docs/searchspace-matrix.md
+1Lines changed: 1 addition & 0 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -28,6 +28,7 @@ The primary compute and HBM resource. Most current work lives here.
28
28
|---|---|---|---|---|---|
29
29
|**Diagnose / Configure / Verify**| CUDA eager dense-FFN Attention/FFN rank namespaces and per-stage/per-peer hidden-state ownership |[audit-afd-p2p-exchanges](../src/inferrsi/data/skill_library/repositories/afd-plugin/audit-afd-p2p-exchanges/SKILL.md)| A=3,F=2 control sender differs from data peer; matching total tokens does not preserve peer rows | afd-plugin @ df8d6a0c403c3ca1b4ed9cf1687e7204908bebec; [source ledger](../src/inferrsi/data/skill_library/repositories/afd-plugin/audit-afd-p2p-exchanges/references/source-notes.md), [proposed cases](reviews/issue91-afd-p2p.md)| 🟢 (source-derived candidate; native/model/efficacy not_run; no Network qualification) |
30
30
|**Observe / Diagnose / Verify**| MiniMax-H3 joint audio/video AdaLN schedule/row-map validity, Metal BF16 rounding and actual int8 dispatch versus approximate controls |[audit-h3-runtime-residency](../src/inferrsi/data/skill_library/repositories/h3-c/audit-h3-runtime-residency/SKILL.md)| Fifty layers/reuse 1 does not disable M5 int8 or BF16 head; test-source parity is not full media quality | h3.c @ 8974cc055ea9c02fcd14cc27dfda3e1027c05153; [precision contracts](../src/inferrsi/data/skill_library/repositories/h3-c/audit-h3-runtime-residency/references/precision-and-streaming.md)| 🟢 (source-derived candidate; native/quality/performance not_run) |
31
+
|**Observe / Diagnose / Verify**| FastH3 effective parameters/ambient override, semantic checkpoint transformations, partial/prefix VSA membership and sparse eligibility/fallback |[audit-h3-metal-fasth3-semantics](../src/inferrsi/data/skill_library/repositories/h3-metal/audit-h3-metal-fasth3-semantics/SKILL.md)| Shape agreement is not conversion; sparse-disabled remains dense fine+coarse; parameter normalization does not erase ambient token enable | h3-metal @ 4ba47a3eabdbebf801ac6664b31f8ec168bc57a8; [semantic counterchecks](../src/inferrsi/data/skill_library/repositories/h3-metal/audit-h3-metal-fasth3-semantics/references/semantic-contracts.md)| 🟢 (source-derived candidate; native/sparse/schedule/media/performance not_run) |
31
32
|**Diagnose / Configure / Verify**| Workflow model-weight residency, partial unload versus wrapper removal, patch identity and static/dynamic backend dispatch |[comfy-ui](../src/inferrsi/data/skill_library/repositories/comfy-ui/SKILL.md)| CPU dynamic construction reroutes to the static patcher; wrapper/free-byte bookkeeping is not physical allocator release | ComfyUI @ b1a247c9cf9dbfda4b2c0a0c1d50950667eeaa0c; [model ownership](../src/inferrsi/data/skill_library/repositories/comfy-ui/references/model-residency.md)| 🟢 (source-derived audit; device/native allocator unverified) |
32
33
|**Observe / Diagnose / Verify**| Native diffusion managed budget versus device headroom, graph-segment weight protection and optional prefetch completion |[audit-sd-cpp-residency](../src/inferrsi/data/skill_library/repositories/stable-diffusion-cpp/audit-sd-cpp-residency/SKILL.md)| CPU bypasses accelerator capacity and segmentation gates; pin release is not physical free; backend capability does not prove overlap | stable-diffusion.cpp @ b13a48219d5cdf408edcb93da6cd69b612ed03d3; [phase residency](../src/inferrsi/data/skill_library/repositories/stable-diffusion-cpp/audit-sd-cpp-residency/references/phase-residency.md)| 🟢 (source-derived audit; native/device/model unverified) |
0 commit comments