Repository navigation
knowledge: complete AREX distillation of PyTorch and TorchNPU #309
Description
Clean resubmission completed (2026-11-02): The owner-requested revert #311 merged as 977289d290092dcb4251cadb9593950771b6178c; replacement #312 merged as 2baebf6ebb815ebf07d21c41e3b3754858c0798e. One replacement commit retains all skill runtime content while reducing additions by about 96%. Acceptance and validation records the 122 passing local tests, passing dual-version CI/install checks and independent review. The original process reports remain available only as historical evidence.
Current state: Completed and accepted after the owner-requested revert and clean resubmission.
Decision needed: None for implementation, PR creation, integration or closure; these were explicitly delegated by the owner. Any unmet repository admission or required backend gate remains visible.
Next action: No remaining work for this requested revision. Native/hardware qualification remains explicitly unverified and follows the retained historical plans when infrastructure is available.
Goal and acceptance
Use VectorSpaceLab/AREX-Skill's actual portable Creator meta-skills to distill Ascend/pytorch and pytorch/pytorch into InferRSI, based on latest main at implementation start: 919ed42c0d01dc0aa3b613085b9041d735d2d4db. Refresh against current main before integration.
Acceptance:
- Immutable source snapshots and reproducible readsets; complete repository capability/directory inventory with inspected, reused, excluded and unverified areas explicit.
- Self-contained operating skill graphs covering major user workflows (installation/build, tensor/dispatcher/autograd, training/data, compiler/export, memory/streams/graphs, distributed/checkpoint, profiling/debugging and extension development where applicable), with substantive source-grounded references and troubleshooting.
- No claim of every line or device being verified. Existing canonical skills retain provenance; cross-link them rather than duplicating ownership.
- AREX scope, grounding, environment preparation, construction, assertion-backed verification, source-script inventory, license checks, native candidate/backend plan, integration report and explicit remaining gaps.
- InferRSI registry, discovery/catalog and coverage documentation; focused and feasible full/package checks; independent review of actual head/base.
- Signed-off commits, linked PR, passing applicable admission/checks, squash integration and a closure summary before closing this issue.
Scope and lineage
This is a fresh full-source distillation, not a relabeling or cherry-pick of the closed unmerged PR #305. Issue #304 remains the prior contributor's separate bounded delivery record. No historical authorship is rewritten.
AREX method snapshot: ac3fe1afa80fb9a09775ecfb2b6cc3ba850a2db6 (confirmed against remote main).
Pinned sources: PyTorch 9a90a4fefcaafa63ac984621963a6fcc66aee68c; TorchNPU 046934140b96836c3fcef14d0f39ae54dc0a06bf; parent-pinned op-plugin ad946817034b3270e309c81c54f0ee808d0b62cb.
No campaign performance PASS, accuracy claim, paid model evaluation, or hardware qualification is implied by knowledge delivery. No policy weakening or branch-protection bypass.
Authority and responsibility
The user explicitly requested both repository distillations, creation of an Issue and PR, PR merge and Issue closure. This authorizes the accountable parent to implement and integrate the requested contribution after checks. Source authors and independent reviewer identities will be truthfully declared with exact configured models/efforts. Native hardware limitations will be reported independently from source and CPU checks.
Activity
david6666666 commentedon Nov 2, 2026
Implementation is underway from latest main
919ed42c0d01dc0aa3b613085b9041d735d2d4dbin an isolated branch. AREX portable workflow is executed fromac3fe1afa80fb9a09775ecfb2b6cc3ba850a2db6; no DisCo CLI or managed-library import is claimed.Two independent source authors and one integration author are configured as
gpt-6.1-sol, effortxhigh, per current knowledge admission. The two proposed registered roots arepytorch(8 workflow subskills) andtorch-npu(8 workflow subskills). Existing canonical source owners retain their revisions.Baseline verification before changes:
ruff check src tests toolspassed; 95 focused skill/library/index/knowledge-tier tests passed. The local environment is Windows without detected CUDA/NPU tools. The user's workspace instructions explicitly authorize implementation with unavailable local tests recorded and remote runtime validation left to the user. Backend blocks will remain visible; source and packaging checks will not be represented as native hardware qualification.Current InferRSI local-skill-v1 requires frontmatter to contain only name and description. AREX license/role/visibility fields will be retained in sidecars, with repository-format checks plus an explicitly labeled verification-only portable projection for the official AREX license gate. No repository policy or runtime type changes are planned.
david6666666 commentedon Nov 2, 2026
Accepted: PyTorch and TorchNPU AREX operating graphs
flowchart LR A["Pinned PyTorch and TorchNPU sources"] --> B["Eight operating routes per framework"] B --> C["Discoverable skills and packaged resources"] C --> D["Independent review and passing CPU CI"] B -.-> E["Device and source-build qualification remains open"]e668cff3b12e46a0dae1a007099c521afa848152, verified on currentmain. The PyTorch graph and TorchNPU graph provide installation, tensors/dispatch, training, compilation, memory/streams/graphs, distributed/checkpoint, profiling and operator-development procedures. Two roots, sixteen children and twelve helpers are available through the existing catalog and wheel/plugin resource readers.Full acceptance record
Authority and scope. The owner explicitly requested both distillations, an Issue and PR, PR merge and Issue closure. The user-provided workflow permits implementation without unavailable local backends, provided unverified scope is explicit. The accepted destination is InferRSI's candidate library, with no managed DisCo import claimed. Prior Issue #304 and closed unmerged PR #305 are separate historical work and were not reused or modified.
Revisions.
ac3fe1afa80fb9a09775ecfb2b6cc3ba850a2db6.9a90a4fefcaafa63ac984621963a6fcc66aee68c(source 2.16.0a0).046934140b96836c3fcef14d0f39ae54dc0a06bf; parent-pinned op-pluginad946817034b3270e309c81c54f0ee808d0b62cb.f5a21bf0bd684fab70bfae02562076eb478bd627.8115da15b9ef34978de500b08aa9f14ea790c35b.01c64d053b1f8e1bc85695cbb707c6c5f540ed87.e668cff3b12e46a0dae1a007099c521afa848152.Breadth and grounding. Full tree inventories cover 22,762 PyTorch entries and 4,453 TorchNPU entries, with script/dependency dispositions and unread boundaries. Independent identity checks matched 122 selected files / 172 inspected spans and seven retained LICENSE/NOTICE resources. Tree discovery is not all-lines semantic reading. Existing focused allocator, DCP and host-memory owners retain their independent pins.
AREX and format adaptation. Scope, source/environment grounding, construction, script dispositions, assertion-backed cases, self-refinement, verification and integration evidence are retained under
tests/fixtures/repo_to_skill/{pytorch,torch-npu}. All 18 documents use InferRSI local-skill-v1 frontmatter; license, role and invocation fields remain in sidecars. The official pinned AREX license checker passed both nine-document verification-only projections. Its literal frontmatter check on the final normalized trees is recorded as an expected format mismatch, not disguised as a pass. No license grant is inferred from NOASSERTION.Review and runtime evidence. A fresh independent reviewer answered 22 prompts before reading their grading assertions; all selected semantic cases passed. Fifteen other fixture cases were not independently replayed. Structural phrase assertions are separate from response grading. Seven adapted diagnostics passed under installed torch 2.15.0+cpu, distinct from the 2.16 source pin; the pinned-source setup deprecation negative also passed. No missing accelerator was substituted with CPU success.
Repairs and failed attempts preserved. The export helper's expected shape-rejection exception handling was repaired and rerun while retaining its initial failure. Initial full CI found omitted plugin ZIP notices and an outdated normalized-root count; both were fixed and the entire Ubuntu matrix passed at the final head. The local Windows full run was interrupted at 14%; isolated reruns identified three existing platform-specific failures (CRLF fixture assumptions, unavailable FIFO creation and locale-default decoding) in addition to the fixed plugin omission. Initial local build TLS and path-length failures were resolved using cached build dependencies and a shorter extraction directory. The separate native execute workflow was intentionally skipped and provides no backend evidence.
Attribution and integration. Signed-off commits and the final squash retain the three actual material authors, each configured as gpt-6.1-sol with xhigh effort. Independent review is recorded separately. The final plugin-only repair left protected source/catalog bytes unchanged, so earlier source/wheel evidence remains applicable, with new-head CI providing fresh packaging coverage. No policy bypass, candidate promotion or new runtime repository type was introduced.
The requested repository contribution is accepted and complete; target-device qualification remains a clearly stated downstream validation task.
david6666666 commentedon Nov 2, 2026
Owner requested reverting merge commit
e668cff3b12e46a0dae1a007099c521afa848152because the delivery diff is too large, then submitting a clean replacement. This reopens the portable delivery scope. Plan: (1) an explicit revert PR, preserving shared history; (2) a focused replacement retaining the two operating graphs, necessary immutable provenance/notices and integration tests, while removing bulk source-tree inventories and process reports from the maintained tree. Historical evidence remains available in PR #310 and its commits. I will report the actual retained scope and validation before integration. Prior hardware limitations remain unchanged.david6666666 commentedon Nov 2, 2026
The explicit revert PR #311 merged as
977289d290092dcb4251cadb9593950771b6178c. Independent exact-inverse review and both Python 3.10/3.11 full CI/install jobs passed atb5dce653842efcec268a67275c357912246b3ade(CI). The tracked tree is restored to the pre-delivery baseline. The reduced replacement is being prepared; this Issue remains open.david6666666 commentedon Nov 2, 2026
Accepted: clean PyTorch and TorchNPU resubmission
flowchart LR A["Oversized delivery"] --> B["Explicit revert"] B --> C["One focused replacement commit"] C --> D["Skills and packaging verified"] C -.-> E["Historical process evidence linked externally"]977289d290092dcb4251cadb9593950771b6178c; clean replacement #312 merged as2baebf6ebb815ebf07d21c41e3b3754858c0798e. The replacement is one signed-off commit: 206 files, +9,352 / −277, about 96% fewer additions than the reverted delivery. All 111 runtime files remain intact; 160 process-fixture files are omitted. Immutable historical report links retain their original evidence.The owner's requested revert and clean resubmission are complete. This supersedes the former delivery-size acceptance while preserving its historical evidence.