Repository navigation
knowledge(hardware): audit sources and deepen memory, graph and collective guidance - #262
Conversation
Replace unsafe cross-stream and buffer-reuse examples with separate ordering, host mutation and device allocator lifetime contracts. Retain bounded PyTorch and CPython readsets, source notices and unverified hardware recipes. Agent-assisted-by: Codex / gpt-6-astra [effort=max] (author) Signed-off-by: David Chen <530634352@qq.com>
Correct configuration ownership and separate the registered backend documentation from the selected TorchAir implementation. Retain a bounded immutable readset for graph keys, persistent eager fallback, input staging and output lifetime, with native validation explicitly unrun. Agent-assisted-by: Codex / gpt-6-astra [effort=max] (author) Signed-off-by: David Chen <530634352@qq.com>
Map the existing hardware owners to their retained source channels, inspection scopes, identities and limitations. Preserve candidate states and distinguish historical imports from direct implementation readsets. Agent-assisted-by: Codex / gpt-6-astra [effort=max] (author) Signed-off-by: David Chen <530634352@qq.com>
Trace a selected multi-rank AICPU mesh AllReduce through selection, scratch-bounded orchestration, peer transfers and reduction. Retain pinned source identities and simulated-test boundaries without claiming native execution or gains. Agent-assisted-by: Codex / gpt-6-astra [effort=max] (author) Signed-off-by: David Chen <530634352@qq.com>
Agent-assisted-by: Codex / gpt-6-astra [effort=max] (author) Signed-off-by: David Chen <530634352@qq.com>
Agent-assisted-by: Codex / gpt-6-astra [effort=max] (author) Signed-off-by: David Chen <530634352@qq.com>
Agent-assisted-by: Codex / gpt-6-astra [effort=max] (author) Signed-off-by: David Chen <530634352@qq.com>
Agent-assisted-by: Codex / gpt-6-astra [effort=max] (author) Signed-off-by: David Chen <530634352@qq.com>
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
Skill evaluationAutomatic execution / cumulative PR budget: Advisory: obtain missing evidenceRetained evidence availability: planned / unexecuted. No approval or merge authorization is issued.
PLANNED: execution is not authorized by knowing this hash. Selected comparisons: 0 · Estimated task calls: 0. Applicable only while these revisions and inputs match the current PR. Plan bindings, budget, frozen exam and advisory detailsUnchanged-correct is valid neutral evidence, not incremental benefit. Per-case outcomes (discovery excluded; regressions and gaps first): []Showing at most 20 of 0 pairs; the evidence link contains all outcomes, including omitted regressions/gaps. Plan SHA-256: Model/settings:
Provider/settings: Uncovered / unsupported paths (105)
|
|
Independent content review by the coordinating Codex reviewer (separate from the three material author agents; no protected-content edits by this reviewer). Reviewed head: Reviewed the host slice at Checks covered source channel/depth and retained unknowns; immutable file identities/ranges and license attribution; asynchronous copy/consumer/buffer lifetime guidance; TorchAir configuration, persistent fallback, staging and output aliases; HCCL selection, scratch/chunking, peer synchronization, local reduction and simulator boundaries; cross-owner navigation; stable IDs/state; and derived-content scope. Two integration concerns were resolved before publication: a shared matrix row could grant an unrelated KV-tier skill an extra stage, and author-worktree CRLF bytes caused a coverage digest inconsistent with the committed LF source. The row was separated by responsibility, and the digest was regenerated from committed bytes. The final delta contains no source or package-input change after the wheel validation. The original 125-test selection produced 124 passes and the digest failure; all 17 affected coverage tests passed after correction. Ruff, index/coverage freshness, source/navigation checks and the isolated noneditable-wheel checks described in the PR passed. The unchanged checked consumers support reuse across main's unrelated local SkillEval newline fix. Residual limits: source definitions were inspected, not executed on CUDA/NPU; no native/simulator numerical, transport, performance, model or service result is claimed. Candidate/not_run states and later NCCL/dispatch/UCX/ROCm gaps remain explicit. |
Governing Issue and delivery slice
Refs #260
The hardware library mixed imported skill guidance with direct implementation evidence, and the host-memory example allowed computation to race an asynchronous copy. This change records the source depth of all 22 hardware owners and adds bounded PyTorch, TorchAir and native HCCL implementation paths, with corrected ordering and lifetime guidance.
The requesting human authorized implementation and Issue/PR publication on 2026-10-30, as recorded in #260. Goal acceptance, merge and closure remain with that human.
Changed artifacts and evidence
ba56102387ef21a3b04b357e5b183d48f0afefc7and two CPython documentation/license files at60403a5409ff2c3f3b07dd2ca91a7a3e096839c7.e1118e8b0fc1b95190ab53b732c0872989550e20, stopping at TorchNPU/CANN calls.96fab0a2acb761e7f8ab0f94aeb5138f200a345b; distinguish simulator semantics from native numerical evidence and stop at HCOMM/ACL boundaries.Canonical source revision:
f05ab7b583c66a9c34cb9f9e8a2c15b2ccfc0db3. All 77 generated abstracts point to that reachable commit; 74 change only source revision/URL, and three change substantive content. The coverage view changes the three target owners and its digest. No runtime, registry schema or directory migration is included.Checks and dependencies
ruff check src tests tools: passed.7b292c1: 124 passed, 1 coverage-digest failure, 2 subtests passed. The failure exposed two author-worktree CRLF inputs differing from committed LF bytes. Commit205b2f6changes only the derived coverage digest; all 17 coverage tests then passed on the current head.--check: passed.git diff --check: passed.7b292c1archive and installed it in a fresh environment outside the checkout: all 22 hardware bundles read, 18 modified-owner reference declarations resolved, and all 720 packaged JSON/Markdown files matched source bytes. The final one-line documentation-only correction does not change any package input.b3549ce1aea1556f76de1d2e14a4240474947b6b: its change since implementation basedabd184is limited to local SkillEval frozen-input newline handling and its test, outside this change and its checked consumers.CUDA/NPU execution, upstream stub/simulator/native tests, numerical quality, performance, service behavior and agent efficacy were not run. Source pins do not prove a compatible combined stack. Native NCCL, broad kernel dispatch, workload-dependent UCX and AMD/ROCm remain explicit follow-up gaps.
Knowledge authorship declaration
Remaining work and Issue closure
This PR proposes the bounded knowledge-maintenance delivery for review. Target-device validation is a separate activity; no candidate promotion, campaign PASS, merge or Issue closure is claimed. The requesting human retains acceptance and closure authority.