feat: add auditable workflow adapters.#11
Merged
Conversation
- select reconfigure, incremental, or targeted TileLang plans from repository and cache identity\n- emit deterministic environment, source, build, binary, and verdict artifacts\n- block unsafe benchmark inputs and cover lifecycle decisions with offline tests
Validate build, binary, service, workload, and artifact identities. Preserve artifact indexes and downgrade incomplete or overwritten runs. Keep smoke and quick results within their declared claim scope.
Record immutable launch, readiness, smoke, and cleanup artifacts per service attempt. Preserve process identities and require explicit NPU quiescence before formal claims. Keep legacy explicit log and PID paths compatible without treating them as formal evidence.
Validate formal run evidence, candidate identity, and normalized environment snapshots. Require campaign-defined thresholds instead of embedding machine-specific assumptions. Remove incident-specific values from the generic NPU precheck contract.
Capture normalized NPU health, utilization, process ownership, load, and swap facts. Preserve raw command outputs and block comparisons when collection evidence is incomplete. Support replay fixtures so parser behavior remains testable without live NPU access.
Block submodule commit mismatches and require complete binary provenance. Bind service cleanup to machine-collected NPU evidence and launched devices. Require model identity before readiness can pass.
Separate campaign and execution identities across repeatable optimization attempts. Enforce preflight, candidate comparison, retention, checksum, and archive gates. Generate resumable checkpoints and hash-chained run ledgers.
Require coarse loss budgets and executable Big-Rock gating before implementation. Connect the SOTA skill to the unified lifecycle and preserve validated replay lessons.
# Conflicts: # scripts/README.md
- derive evidence and fairness identities from the run manifest - aggregate build, service, evidence, fairness, and Big-Rock gates - add experimental vLLM-Ascend, SGLang, and NVIDIA adapters
Only declare environment artifacts and the active evidence type by default, so performance runs do not require accuracy and profiling outputs.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
变更概述
xllm-flow实验生命周期,覆盖工作区预检、run 创建、attempt ledger、checkpoint、finalize 和 archive。manifest.json作为身份信息的单一事实源,自动生成运行证据和公平性候选;gate all会阻断身份投影漂移。主要能力
workspace-tasks.json、experiment.yaml、preflight、run 生命周期、checkpoint、fingerprint 和 ledger。适配器支持范围
xllm_ops检查。解决的问题
此前 manifest、运行证据、公平性输入和各个独立 CLI 会重复维护身份字段,容易出现 commit、framework、设备顺序或 workload 漂移。同时,性能优化流程可能在没有证明最大瓶颈的情况下过早进入细节。
本 PR 将身份信息集中到 manifest,通过派生证据和统一门禁阻止漂移,并通过 Big-Rock 门禁要求先量化主要性能预算,再进入实现。
验证结果
pytest -q:138 项测试全部通过。git diff --check:通过。后续验证
在将 vLLM-Ascend 或 SGLang 从“实验性适配器”升级为“正式支持”之前,还需要分别完成真实仓库的构建、服务启动、benchmark、证据导出和
gate all端到端运行,并归档对应 artifacts。