Repository navigation
[Feat][Knowledge] Distill Ascend RC model and fusion performance history - #354
Conversation
Trace plugin discovery, mixed MiniMax cache storage and DSpark PD configuration through their consumers and test oracles. Agent-assisted-by: Codex / gpt-6-astra [effort=xhigh] (author) Signed-off-by: jiangkuaixue123 <jiangxiaozhou111@163.com>
Regenerate the discovery records against the committed skill sources and verify the new resources in an installed wheel. Agent-assisted-by: Codex / gpt-6-astra [effort=xhigh] (author) Signed-off-by: jiangkuaixue123 <jiangxiaozhou111@163.com>
Skill evaluationPLANNED: not executed; there are no task outcomes yet.
Author verification steps (nonpaid; not yet run)
python tools/build_skill_index.py --source-revision 2c766dbc95605e61b907ce71418ab941287b427a --checkThe declared source pin matches this head's canonical library bytes. The command has not run: retain its exit code and full drift/success output. It checks the generated projection, not claim truth or skill efficacy. Evidence and logs · Retention: 30 days. Plan bindings, budget, frozen exam and advisory detailsAdvisory: obtain missing evidenceRetained evidence availability: planned / unexecuted. No approval or merge authorization is issued.
Model/settings:
Automatic execution / cumulative PR budget: {"reason":"Trusted configuration or pricing missing, invalid or expired","remaining_usd":"unknown","reserved_or_unresolved_usd":"unknown","spent_or_legacy_usd":"unknown","state":"blocked"}Projection check selection (not an executed result): {"declared_in":"src/inferrsi/data/repo_packs/inferrsi/knowledge/README.md","source_revision":"2c766dbc95605e61b907ce71418ab941287b427a","status":"ready"}Provider/settings: Uncovered / unsupported paths (106)
|
|
Review at head Blocking
What I checked (no issues found) I spot-checked the new
I did not verify the DeepSeek-V4-Flash tutorial line references (P DP1/TP8 + DSA-CP, D DP8/TP1, the 360-token tier) or the one-card / four-card e2e test bodies, and I did not run anything. Happy to re-review once CI is green. |
|
Small correction to my comment above: the failing |
Agent-assisted-by: Codex / gpt-6-astra [effort=xhigh] (author) Signed-off-by: jiangkuaixue123 <jiangxiaozhou111@163.com>
Dong1017
left a comment
Dong1017
left a comment
There was a problem hiding this comment.
The catalog-count CI failure is cleared at 5360d88e. This is a scoped follow-up to Amy's review, not a whole-PR approval.
- I read both earlier failing logs: the failures were the Ascend count (
75versus74) and totalSKILL.mdcount (143versus142). The new two-file commit updates those expectations, including the second total assertion intest_skill_index.py; it changes no canonical knowledge or generated projection. - Both current-head full checks now pass: Python 3.10 and 3.11 each report 3,548 passed, 13 skipped. All six jobs in this Ready run succeeded.
- The separate request for a retained receipt of the pinned source/index command remains to be dispositioned. I did not execute that command or extend the upstream-source review, and these software checks establish no model/device efficacy.
Revision binding
Head 5360d88e65a9c75e7745e145662da197077e680d; logs identify merge 6e995c2 of this head into 9dded7adac0ac835f6a73493309399e328a43ab2. Inspected delta: 3311a100..5360d88e.
amy-why-3459
left a comment
amy-why-3459
left a comment
There was a problem hiding this comment.
LGTM
PR review verdict — PASS
Reviewed base 9dded7ad and head 5360d88e. Incremental review 3311a100..5360d88e (one commit, test: account for the new Ascend skill in catalog totals), checked against my previous review.
- Blocker 1 resolved: the earlier
full-check (3.10)failure was exactlytests/test_skill_index.py::test_hardware_filters_include_only_matching_and_shared_skills(75 == 74) andtests/test_skill_library.py::test_library_root_exposes_only_owner_categories_and_required_entrypoints(143 == 142). This commit bumps those totals (ascend 74→75, SKILL.md 142→143), which matches adding one Ascend skill. At this headfull-check (3.10),full-check (3.11),installed-package (3.10/3.11),declarations,pr-referencesandkv-task-qualificationare all green. - Item 2 resolved:
tests/test_skill_index.pyregenerates the full index projection withrender_index(...)from the recorded source revision and compares it with the committedDESTINATIONtree, so the index refresh in3311a100is covered by the now-green pytest run.
Still unverified by me (not blocking): the DeepSeek-V4-Flash tutorial line references and the one-card / four-card e2e test bodies noted last time. I did not run anything locally.
Screen all tag-reachable history and retain focused source procedures for DeepSeek V4, Kimi K3, MiniMax M3, GLM5 families and communication/configuration gates. Agent-assisted-by: Codex / gpt-6-astra [effort=xhigh] (author) Signed-off-by: jiangkuaixue123 <jiangxiaozhou111@163.com>
Agent-assisted-by: Codex / gpt-6-astra [effort=xhigh] (author) Signed-off-by: jiangkuaixue123 <jiangxiaozhou111@163.com>
jiangkuaixue123
changed the title
|
Re-review at head Needs a fix
Questions / please confirm
What I checked (no issues found)
Not verified: NPU numerics/performance (as the PR states), and the remaining commit/line anchors beyond the ones listed above. |
Preserve the full discovery ledger outside runtime package data. Agent-assisted-by: Codex / gpt-6-astra [effort=xhigh] (author) Signed-off-by: jiangkuaixue123 <jiangxiaozhou111@163.com>
Correct field anchors and validate pinned definitions without replacement objects or lazy fetches. Archive the full history outside runtime resources, restore source backlinks, and verify compact JSON and script resources in installed wheels. Agent-assisted-by: Codex / gpt-6-astra [effort=xhigh] (author) Signed-off-by: jiangkuaixue123 <jiangxiaozhou111@163.com>
Regenerate every skill projection from the committed review fixes and remove the archived inventory from runtime source paths. Agent-assisted-by: Codex / gpt-6-astra [effort=xhigh] (author) Signed-off-by: jiangkuaixue123 <jiangxiaozhou111@163.com>
amy-why-3459
left a comment
amy-why-3459
left a comment
There was a problem hiding this comment.
LGTM
PR review verdict — PASS
Reviewed base 9dded7ad and head aeb267fb. Incremental review f426ae47..aeb267fb (3 commits: 9fd6fd1e archive inventory, 2c766dbc review fixes, aeb267fb index refresh), checked against my previous review at f426ae47.
All four earlier findings are resolved:
- Wrong anchors:
references/communication-config.mdnow linksascend_config.py#L413(enable_fused_mc2: int = 0) and#L404(mega_moe_max_tokens: int = 65536) at3b318862. The newtools/check_vllm_ascend_source_anchors.pychecks that a 1–5 line span at the metadata pin contains the exact definition, reading only Git objects with--no-replace-objects --no-lazy-fetch.tests/test_vllm_ascend_source_anchors.pycovers blank or unrelated lines, wrong pins,git replacesubstitution, mutable or abbreviated revisions, broad spans and the CLI path. - 1.4 MB inventory in the wheel:
history-inventory.jsonmoved todocs/reviews/vllm-ascend-0271rc1/and was removed frommetadata.jsonprovenance_readsetsand from thevllm-ascend.jsonrecord'ssources[0].pathsandsource_files. The newtest_runtime_reader_keeps_full_history_outside_the_bundleasserts it is not inKnowledgeReader().read(...).content. - Readset back-link: the route test now requires a
](sources.md)or](history-scope.md)link, andtest_every_instruction_route_retains_a_readset_linkapplies that to every routed reference. The four performance references gained thesources.mdback-link. - Installed-package coverage:
tests/check_installed.pynow byte-compareshistory-reviewed.jsonandscripts/screen_history.pyand asserts the inventory is absent from the wheel.
The aeb267fb refresh only moves the source pin 5004c0ef → 2c766dbc across the generated records, plus the vllm-ascend.json path and digest changes above. CI at aeb267fb is green: full-check (3.10/3.11), installed-package (3.10/3.11), kv-task-qualification, pr-references, declarations.
Note (non-blocking): the archived inventory link in history-scope.md pins PR-branch commit 9fd6fd1e, which won't be on main after a squash or rebase merge. It follows the same pinning convention as the generated records, so it should keep resolving through the PR ref.
Not verified: NPU numerics or performance (as the PR states), and I did not run the anchor check against a real vllm-ascend clone.
|
@Dong1017 @amy-why-3459 Thanks for the reviews. All four findings are addressed at current head The exact-head source/index command, environment, exit status and output are retained in the PR description. Ready CI passed all six jobs; Python 3.10 and 3.11 each report 3,578 passed and 13 skipped. @amy-why-3459 Thank you for approving this head. @Dong1017 Could you re-review the updated evidence and fixes, and approve if there are no remaining blockers? Please let us know if anything else needs addressing. Native NPU numerics/performance and agent efficacy remain unverified, as documented. |
|
@jiangkuaixue123 Re-checked the four review fixes and updated evidence at head
I found no remaining blocker in these fixes and the evidence update. This is a scoped review comment, not a formal approval. I read the source, tests and recorded CI results; I did not rerun the commands locally or validate the full upstream history, native NPU numerics/performance, or agent efficacy. The documented candidate status and broader #350 acceptance boundary remain unchanged. |
Dong1017
left a comment
Dong1017
left a comment
There was a problem hiding this comment.
Approved at head aeb267fb5cefa1139efb2ebac2dcc2220ded5bf5, following the current-head review and evidence disposition. I rechecked the head, discussion, review threads and CI status before submitting; no material change or new blocker was found.
The four review findings and missing source/index receipt are resolved within the documented source-grounded candidate slice. Native NPU numerics/performance and agent efficacy remain unverified. This approval does not complete the broader #350 acceptance or promote the candidate.
…ith #354 Keep the installed-record formula of 49 topic records plus the skill index, and record main's added skill so the skill count is 80. Regenerate docs/skill-coverage.md from the combined registry, matrices and plans. Agent-assisted-by: Grok 4.7 (conflict resolution) Signed-off-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: amy-why-3459 <amy-why-3459@users.noreply.github.com>
…licts with #354 Conflicts were limited to generated skill projections and the skill coverage view. Regenerated docs/skill-coverage.md from the merged registry and matrices. Knowledge abstracts are refreshed from this merge commit in the following change, because the generator pins a committed source revision. Agent-assisted-by: Cursor / grok-4.7 [effort=unknown] (conflict resolution) Signed-off-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: amy-why-3459 <amy-why-3459@users.noreply.github.com>
… conflicts with #354 Keep both appended matrix sections: main's vLLM Ascend v0.27.1rc1 notes and this branch's Issue #49 conditioning-transport notes. The installed catalog counts 129 knowledge records because each side added one distinct record. Generated skill abstracts follow main in this merge and are refreshed from this commit next. Skill coverage is regenerated from the merged matrices and registry. Agent-assisted-by: Cursor / grok-4.7 (conflict resolution) Signed-off-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: amy-why-3459 <amy-why-3459@users.noreply.github.com>
Governing Issue and delivery slice
Refs #350
The requester authorized v0.27.1rc1 Repo-to-Skill distillation, PR submission and the expanded performance scope below. This slice complements the separately assigned historical ACLGraph and checkpoint-adaptation work.
Knowledge authorship declaration
Changed artifacts and evidence
The registered
vllm-ascendskill now covers performance mechanisms, source evolution, current eligibility, reversals and the closest test oracles.references/deepseek-v4-evolution.mdtraces original support, HcPre reversal, cache plans, async metadata, IndexCache, DSpark projection/input work and boundary fixesreferences/fused-operators.mdcovers graph norm/quant/RoPE, GMM-SiTU/SwiGLU, one-collective SFA DCP, MLA prolog, GDN paths, sampling and removed fusionsreferences/communication-config.mdand the DSpark resources trace requested versus derived state, SP ownership, model/device conflicts, environment migration, MegaMoe rollback and capacityreferences/model-performance.mdcovers KDA precision/overlap/padding, MSA TP/FP8/top-k and GLM5/5.1/5.2/5.3 index reuse, rotation, loader and fusion constraintsAll resources live under
src/inferrsi/data/skill_library/repositories/vllm-ascend/. Registry, both matrices, generated coverage and package discovery records are updated. The index generator pins a single committed source snapshot across all records.History boundary
The official tag resolves to 3b31886237ed65c435e0965b10e8002ff5769766.
docs/reviews/vllm-ascend-0271rc1/history-inventory.json, with an immutable archive link and a packaged reproducible read-onlyscripts/screen_history.py. The full inventory is excluded from the runtime bundle and wheel.Candidate screening is broader than semantic extraction. Historical benchmarks remain attributed upstream reports. The procedures keep Kimi K3/GLM source coverage separate from the RC's formal MiniMax-M3/DeepSeek-V4 validation table.
Findings that change optimization decisions
Review corrections
Addresses the latest source/package review.
Checks and dependencies
2feac43of headaeb267fb5cefa1139efb2ebac2dcc2220ded5bf5into base9dded7adac0ac835f6a73493309399e328a43ab2. The knowledge-model-policy declaration also passed.Retained source/index receipt
The common generated source pin now identifies the committed canonical review fixes. The following read-only check ran on the current PR head and returned the recorded output:
The separate local upstream-source command
python tools/check_vllm_ascend_source_anchors.py --repo <local-vllm-ascend-checkout>exited 0 and returned (the local checkout path is redacted):This verifies the two selected source definitions and generated index consistency. It does not establish NPU numerical accuracy, service performance or agent efficacy.
Remaining work and Issue closure
This PR delivers source-grounded performance procedures and the auditable history screen for review. Broader parent-issue acceptance and native/efficacy validation remain open. No issue closure or empirical promotion is requested.