Skip to content

Commit 27b64c6

Browse files
committed
Integrate the reviewed AFD intake with current main
Preserve the AFD and RLT matrix additions alongside the existing benchmark and H3 knowledge. Keep the AFD report and descriptor unchanged. Agent-assisted-by: Codex / gpt-6.1-sol [effort=xhigh] (author) Signed-off-by: GUOGUO <xwdong1998@163.com>
2 parents 64f2efc + d600992 commit 27b64c6

95 files changed

Lines changed: 3247 additions & 235 deletions

File tree

Some content is hidden

Large Commits have some content hidden by default. Use the searchbox below for content that may be hidden.

‎docs/knowledge-base.md‎

Lines changed: 10 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -14,7 +14,7 @@ separate work, not an implemented capability of this knowledge base.
1414
```text
1515
src/inferrsi/data/skill_library/
1616
README.md Library navigation
17-
skill-index.json Stable IDs → categorized paths/backend (57 skills)
17+
skill-index.json Stable IDs → categorized paths/backend (58 skills)
1818
hardware/{ascend,cuda,host}/<id>/ Hardware/platform-specific complete skills
1919
frameworks/<framework>/<id>/ Framework-owned complete skills
2020
repositories/<repo>/ Engine entry, sub-skills and focused complete skills
@@ -168,20 +168,26 @@ Complete procedures, sidecars, references and cases are packaged, with 13 Ascend
168168
repositories, methods or workflows. The 21 engine and research-router skills
169169
use the same ownership layout under `repositories/<engine>/` and
170170
`workflows/inference-engine-research/`, including their sub-skills and references.
171-
The library contains 78 `SKILL.md` files, all in English.
171+
The library contains 79 `SKILL.md` files, all in English.
172172
The collection overview, import notes and source manifest belong to the research
173173
workflow's references; Strata design-space analysis belongs to Strata. The library
174174
root contains only navigation and the registry.
175175

176176
The [generated catalog](../src/inferrsi/data/repo_packs/inferrsi/knowledge/README.md)
177-
projects the 57 registered skills into the contribution catalog. Its records
177+
projects the 58 registered skills into the contribution catalog. Its records
178178
use `inferrsi.<backend>.skills.<name_with_underscores>` IDs, with `ascend` and
179179
`cuda` for those hardware-specific scopes and `common` for the remaining skills.
180180
Each `skill_documents.original` points to the full packaged procedure. The
181181
`skills` CLI retains the original hyphenated skill IDs and defaults to this
182-
57-ID registry; engine procedures remain available through library navigation
182+
58-ID registry; engine procedures remain available through library navigation
183183
and catalog document pointers.
184184

185+
The registered [h3.c runtime audit](../src/inferrsi/data/skill_library/repositories/h3-c/audit-h3-runtime-residency/SKILL.md)
186+
owns MiniMax-H3 component/lifetime, AdaLN, precision/SSD mechanisms and the
187+
retained Apple feasibility/header-inventory research. Its source-derived
188+
candidate state and imported header observations remain separate from native
189+
execution, media quality, target memory and performance qualification.
190+
185191
Use `inferrsi knowledge search --repo-type inferrsi --limit 100` in an ordinary
186192
installation. Catalog JSON and full package bundle reads need no YAML parser:
187193
`inferrsi read inferrsi.ascend.skills.ascend_graph_mode --material` reads the

‎docs/operations.md‎

Lines changed: 23 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -17,6 +17,7 @@ Effect classes used below:
1717
| Intent | Entry | Effect class | Boundary |
1818
| --- | --- | --- | --- |
1919
| Search packaged records | CLI `knowledge search` / `knowledge show`; catalog API | Read-only | Installed package catalog with the existing tier/repo/backend/scenario filters and IDs; these package-only filters exclude skills and cases from every other authority |
20+
| Prepare model-aware engine research | CLI `knowledge research`; Python `research_knowledge` | Read-only | Independent model/framework/task/modality/mechanism queries over the same package catalog; source-grounded known profiles, deduplicated matches and explicit target annotations; no engine/hardware qualification |
2021
| Discover/read packaged or custom skills | CLI `skills list` / `show` / `resource` / `reference` | Read-only | Metadata-first listing; `show` loads one named instruction body; `resource`/`reference` resolve paths for your file reader without loading or executing them |
2122
| Collect normalized knowledge | CLI `collect`; Python `KnowledgeReader.collect` | Read-only | Explicit sources, authority-qualified identities, exact revisions, bounded output and explicit omissions |
2223
| Read exact material/resources | CLI `read` / `resource`; Python `KnowledgeReader.read` / `KnowledgeReader.resource` | Read-only | Returns the requested revision or an explicit unavailable/stale error; never substitutes newer bytes |
@@ -46,6 +47,28 @@ The tier/repo/backend/scenario filters are package-only: an empty filtered
4647
result does not mean that no checkout or store knowledge exists. See the
4748
[three-tier guide](../src/inferrsi/data/knowledge_packs/README.md) for all sources.
4849

50+
### Prepare model-aware research — `knowledge research`
51+
52+
- **Inputs:** model identity; repeatable `--framework`, `--task`, `--modality`, `--mechanism`; existing `--tier`, `--repo-type`, `--backend`, `--scenario` hard filters and `--limit` output cap; optional `--target-backend`, `--target-device`, positive finite `--target-memory-gib` annotations.
53+
- **Scope:** installed package records only. A packaged MiniMax-H3 profile supplies its source-grounded diffusion/video/audio identity and component roles. Unknown models use only caller-supplied dimensions and need source inspection; no architecture is guessed.
54+
- **Output:** JSON query groups and match counts, a bounded deduplicated union with `matched_queries`, unchanged full records/provenance, active filters, component roles, preparation/example resource paths and an unverified target state. Queries take turns contributing results; this ordering does not rank engine quality.
55+
- **Side effects:** none; no network requests, hardware probes, model download or engine execution. Target annotations do not add runtime backends.
56+
- **Failure behavior:** no match returns an explicit empty envelope and exits 0; invalid limits/memory or unsupported catalog filters exit 2. A zero-match dimension stays visible.
57+
- **Example:** `inferrsi knowledge research MiniMax-H3 --framework mlx --target-backend metal --target-device "M5 Mac Studio" --target-memory-gib 64 --limit 12`
58+
- **Python:** `from inferrsi.research import research_knowledge`; `research_knowledge("MiniMax-H3", frameworks=["mlx"], modalities=["video"], limit=12)`.
59+
60+
Exact `knowledge search` and `context` still AND query tokens. Independent
61+
research queries broaden discovery without relaxing the active package filters.
62+
Keep those filters explicit when transferring mechanisms from other backends;
63+
target intent does not qualify their applicability. Read the canonical
64+
[preparation procedure](../src/inferrsi/data/skill_library/workflows/inference-engine-research/references/model-aware-preparation.md)
65+
and the repository-owned [h3.c residency audit](../src/inferrsi/data/skill_library/repositories/h3-c/audit-h3-runtime-residency/SKILL.md)
66+
for native-engine reference selection, full encoder/DiT/VAE accounting,
67+
quantized-package compatibility and target hardware decisions. Returned resource
68+
paths are relative to `importlib.resources.files("inferrsi")` and ship in the wheel.
69+
The small model profile links that skill's full inventory and worked example;
70+
those resources retain their repository owner and source attribution.
71+
4972
### Discover/read packaged or custom skills — `skills list` / `show` / `resource` / `reference`
5073

5174
- **Inputs:** `list` reads frontmatter and sidecar metadata; `show <skill>` selects one instruction body; `resource <skill> <relative-path>` and `reference <skill> <relative-path>` resolve one path. The default is the installed library's registered skills; an explicit `--root` selects a custom flat tree or a categorized library with `skill-index.json`. The optional `skills` extra supplies the YAML parser.

‎docs/reference-matrix.md‎

Lines changed: 8 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -42,6 +42,7 @@ Reading rules:
4242

4343
| Reference (origin @ revision) | Inspected scope | Accelerator | RAM / Host | SSD / Storage | Communication | OS / System |
4444
|---|---|---|---|---|---|---|
45+
| 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 |
4546
| 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 |
4647
| leejet/stable-diffusion.cpp @ `b13a48219d5cdf408edcb93da6cd69b612ed03d3` (MIT) | Supplied fixed-revision archive, inspected 2026-10-22; 18 bounded readset paths plus one separately identity-checked README matched to pinned commit/tree (19 blobs total); header coverage is selected owner/state declarations; full archive/submodules not byte-authenticated; bounded image API, native runner, capacity, segment/prefetch and cleanup paths; [readset](../src/inferrsi/data/skill_library/repositories/stable-diffusion-cpp/audit-sd-cpp-residency/references/source-notes.md); GGML submodules/model kernels unread | 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; host peaks unmeasured | out-of-scope — demotion eligibility observed, complete loader/I/O not extracted | out-of-scope — device copies are not network transport | out-of-scope — runner scheduling is not OS scheduling |
4748
| NVIDIA/TensorRT-LLM @ f33e0e5fd | skill-doc-only (`.claude/skills/`) | mined — evidence taxonomy, kernel writing, casebook, roofline refs → `dense-performance-evidence`, `kernel-writing-guide`, `inference-optimization-casebook`, `performance-ceiling-analysis` | mined — host overhead verdict + line-profiling loop → `host-overhead-profiling`, `host-overhead-optimization` | mined — weight-movement/overlap + KV host-cache patterns → `weight-loading-optimization`; disk-level: negative (see below) | out-of-scope — host/device weight-movement overlap remains in RAM/Storage, not network coverage | mined — CPU affinity rule + env preflight → `os-process-placement` |
@@ -72,6 +73,7 @@ Reading rules:
7273
| **C4 / NVIDIA Resiliency Extension** @ `49af0b03f14550d36a60dcfebd573848b5fd8435` (Apache-2.0 source headers; Megatron-Core lineage) | local `base_manager.py` complete-iteration selection/save/finalizer and `local_manager.py` dirty/final file lifecycle | mined (contract only) — CUDA D2H synchronized before async request returns; no NPU qualification | mined — snapshot/replica lifetime constraints → `checkpoint-recovery-consistency` | mined — clean/dirty publication, complete-rank generation selection and incomplete-save rejection; no power-loss proof | mined — all-rank availability/finalization and replica failure domains, not transport throughput | out-of-scope |
7374
| **CR1 / DeepSpeed ZeRO-Infinity** @ `5f266e26d216a705d9c64cac88fb813cefdbac4b` (Apache-2.0) | complete short `offload_config.py`, official ZeRO-3 live offload-field docs; no runtime/checkpoint audit | mined — parameter stage-3 and optimizer stage-1/2/3 boundaries → `training-resource-budget` | mined — CPU optimizer budget, pinned RAM, element-count residency and gated core share | mined — NVMe buffer pools and tiled read/compute/write overlap intent; buffer-size units unresolved in approved excerpt | out-of-scope — CPU/device transfer is not network coverage | mined — CPU resource-budget qualification; no privileged changes |
7475
| **CR5 / Ascend MindSpeed adaptive memory** — `core_r0.8.0`, page history `53193db8` (full pin and license unresolved) | original GitCode historical `docs/features/adaptive-memory.md` body directly read; implementation unavailable | mined (docs-only case) — spare-NPU-memory versus recompute/swap decision → `training-resource-budget` historical case; changed-layer comparison rejected as controlled gain | mined (docs-only) — CPU swap and allocator/affinity prerequisites | out-of-scope | out-of-scope — NPU/CPU copies not network evidence | mined (historical conditions) — incompatible allocator features and CPU-affinity prerequisites; no current compatibility claim |
76+
| InferMatrixLab/vllm-rlt (legacy ThinkFlowLab/vllm-rlt) @ `4a350344027960a48384626aa646f1765fda6d4f` (root Apache-2.0; model attribution in NOTICE) | Fixed-pin bounded depth-aware KV manager/runner/docs/test reads; [intake](reviews/issue91-vllm-rlt-intake.md); descriptor-only / intake-plan | located-not-mined — synchronous early-exit LAST_EXITED propagation versus SHARED and bookkeeping-only marking; no operating skill or runtime qualification | out-of-scope — device KV does not establish host-RAM coverage | out-of-scope — no storage mechanism extracted | out-of-scope — inspected KV copies are local tensor copies; no transport mechanism extracted | out-of-scope — no OS mechanism extracted |
7577
| vllm-project/afd-plugin @ `df8d6a0c403c3ca1b4ed9cf1687e7204908bebec` (Apache-2.0 selected Python readset; retained MIT CAM subset notice) | Fixed-pin bounded connector/config/topology/test reads; [intake](reviews/issue91-afd-plugin-intake.md); descriptor-only / intake-plan | located-not-mined — CUDA P2pNccl role/world/subgroup mapping candidate; no operating skill or backend qualification | out-of-scope — no host-memory mechanism extracted | out-of-scope — no storage mechanism extracted | located-not-mined — actual NCCL send/recv anchors found; transport/completion/performance unverified, no Network coverage credit | out-of-scope — no OS mechanism extracted |
7678

7779
Wave-8 implementation provenance (Dot bundles integrated 2026-10-15):
@@ -349,3 +351,9 @@ those dependencies, a build-complete checkout or a complete serving system.
349351
The three peer cards summarize related source observations; generated abstracts
350352
point to the complete procedures. Neither independently corroborates or promotes
351353
the candidate mechanisms.
354+
355+
## Task-specific benchmark selection (issue #60)
356+
357+
| Source / revision | Inspected scope | Reuse, extraction and limits |
358+
| --- | --- | --- |
359+
| [Issue #60 report](https://github.com/InferMatrixLab/InferRSI/issues/60#issuecomment-6310920198), 2026-10-21; eight Git source inventories in the [owner metadata](../src/inferrsi/data/skill_library/methods/inference-benchmark-design/metadata.json) | Task/quality/workload separation across text, agentic coding, work agents, TTS, full-duplex interaction and image/video; selected pinned AIPerf/realtime/FDB/video sources rechecked and the current InferRSI metric contract inspected | Extend [inference-benchmark-design](../src/inferrsi/data/skill_library/methods/inference-benchmark-design/SKILL.md) with [task selection](../src/inferrsi/data/skill_library/methods/inference-benchmark-design/references/task-benchmark-selection.md) and [command templates](../src/inferrsi/data/skill_library/methods/inference-benchmark-design/references/benchmark-commands.md). Candidate guidance; source review and bounded smoke do not qualify all tasks or establish a performance gain. Replay scores, reference parity and task quality retain separate meanings. |

0 commit comments

Comments
 (0)