task: distill PyTorch and TorchNPU into inference skills #304

Current state: Bounded readsets already exist for PyTorch host copy (v2.8.0), DCP async save, Inductor CUDA-graph documents, and TorchNPU device/pinned allocators. Dispatcher, SDPA, device allocator configuration, profiler, AOTInductor, NPU graph capture, PrivateUse1 registration, streams, HCCL control, and compile hooks were not covered.
Decision needed: None. The repository owner requested this distillation and authorized the issue, pull request, merge, and issue closure.
Next action: Land the distillation pull request, then close this issue after the closure summary.

Delivery update (2026-11-02): PR #305 closed without merging at 37b01ae0. The landing step above is therefore pending a disposition from the existing contributor; no merged delivery or completion is established by that closure. The original scope, ownership and acceptance text below are preserved.

Goal and acceptance

• Problem / desired outcome: Distill pytorch/pytorch and Ascend/pytorch into InferRSI skills and dependency knowledge using the AREX repo-to-skill sequence (scope, ground, construct, verify), without duplicating existing owners and without fabricating measurements.
• Completion criteria: New candidate procedures and knowledge cards are pinned to immutable commits, coverage docs are updated, CI other than an unsatisfiable knowledge-authorship admission is understood, and the delivery pull request is merged or the exact merge blocker is recorded.

Scope and constraints

• In scope / out of scope: Inference control-plane contracts at the pins below. Out of scope: rewriting the v2.8 host-copy, DCP, CUDA-graph, or 67511c53 allocator owners; running devices; claiming PASS or a speedup; mining c10d, Inductor codegen, ACL runtime, HCCL transport, or kernel sources.
• Relevant dependencies, shared interfaces or limits (if any): Skill registry, knowledge scopes, reference matrix, and searchspace matrix. torch_npu is a knowledge scope only, not a runtime target.

Ownership and authorization

• Proposed by / actual requester: Repository owner (chenweiqing), via the distillation task assigned to this agent.
• Status: Accepted
• Accountable parent (Issue-level owner, if assigned): Cursor Grok 4.7 acting for the owner-requested distillation
• Slice implementer(s) / acting contributor (if assigned): Cursor Grok 4.7
• Goal-acceptance owner: Repository owner
• Accepted decision or scoped delegation: Owner task text authorizing the issue, implementation, merge, and closure.

Inputs and deliverables

• Relevant sources, revisions or linked issues:
  • pytorch/pytorch 31a77fc9c0a7c6fff15a1ff58c743f53b1a9d457 (v2.15.0)
  • Ascend/pytorch 54e365d3e3614299bccd17f06916544c3880c9fc (tag v26.2.0-pytorch2.14.0)
  • AREX-Skill method: scope/ground/construct/verify. DisCo CLI was not executed. Prior method commit ac3fe1afa80fb9a09775ecfb2b6cc3ba850a2db6 was not re-fetched.
• Expected artifacts: Registered skills, dependency cards, readsets, coverage updates, and a review note. Candidate / not_run only.

Completion and closure

• Who accepts completion and may close: Repository owner authorized closure with this task. The accountable agent will post the closure summary and close the issue after merge, or record the blocker if merge is refused.
• Remaining uncertainty / next decision: Knowledge-model admission may reject this agent identity. That is reported rather than bypassed. Large unread areas remain listed in the pull request.