Skip to content

policy: record exact review for vllm-ascend-quant-hust#8 - #59

Open
jxd1111 wants to merge 1 commit into
mainfrom
quant-hust-8-review
Open

jxd1111 wants to merge 1 commit into
mainfrom
quant-hust-8-review

Conversation

@jxd1111

@jxd1111 jxd1111 commented Oct 28, 2026

Copy link
Copy Markdown

Adds a central exact-diff review for vLLM-HUST/vllm-ascend-quant-hust#8.

contracts/ascend-quant-evidence-v1.schema.json is flagged by MEASURED because the file is JSON and its published contract identifiers (mean_ttft_ms, p99_ttft_ms, mean_tpot_ms, p99_tpot_ms) sit next to JSON Schema bounds ("minimum": 0). The file defines the shape of the BF16/W8A8/W4A4/W4A8 evaluation record validated by ascend-quant-toolkit validate-evidence; it contains no measured value, dataset, request log, timing or run result.

It cannot be reworded without changing contract 1.1.0: the property names are the contract, and dropping their numeric bounds would let the schema accept records the toolkit rejects. Reformatting the JSON only to move digits out of the heuristic's 60-character window would evade the check rather than answer it, so the bytes are left exactly as published.

The review note records the reasoning. The ledger entry binds PR #8 head 1d36d36e7e7cececf8ffb6ae152aa05dd7307fd1: blob 4918c81b324f3e4fbdfef1929f0474cdb17b318e, fingerprint fb6cdf33ba0121529a4803e634c72432a71635439eefa646f7ae9ed9d9f9a2d0. Verified with the central script on the live PR file list: this entry resolves the single remaining finding and nothing else.

Rollout: after this merges, the toolkit repository pins the merge commit and passes --repository/--pull-request in a separate base-branch PR; PR #8 is then re-evaluated. The heuristics, DATA_PARTS, DATA_SUFFIXES, complete-list validation, removal semantics and the debt baseline are unchanged.

The toolkit's evidence JSON Schema is flagged by MEASURED on its own field
names (mean_ttft_ms, p99_tpot_ms) and schema bounds. It is a published
contract with no measured values, dataset, log or run result, and cannot be
rewritten without changing contract 1.1.0.

The review note explains the decision; the ledger entry binds the exact diff
of PR #8 head 1d36d36 (blob 4918c81b..., fingerprint fb6cdf33...).
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant