Repository navigation
[Task] 0 Day Model Support: Reusable Integration Experience from Transformers to GPU and NPU Inference Engines #315
Description
Current state: Scope is defined at roadmap level only. The body defers the initial models, GPU/NPU engines and environments to the owner, so acceptance criteria 1-5 cannot yet be judged, and no delivery has started. The 2026-11-02 review additionally asks which existing assets this work extends, since workflows/inference-engine-research, hardware/ascend/ascend-model-adaptation and methods/weight-loading-optimization already cover parts of this path.
Decision needed: The initial model, GPU/NPU engine and environment list with versions, the incremental gap relative to those existing skills, and who accepts completion, from the owner @amy-why-3459.
Next action: The owner records that list and the incremental gap in this Issue; knowledge delivery then follows the protected-path admission and index-regeneration rules. No scope, assignment or acceptance is changed by adding this status block.
Parent issue: #316 · InferRSI Roadmap
Goal
Expand reusable integration and optimization experience along the Transformers → GPU Inference Engine → NPU Inference Engine path to improve InferRSI's efficiency in supporting new models.
Owner: @amy-why-3459. Corresponds to original roadmap item 8.
Scope
- Map the integration path: Starting from a Transformers reference implementation, document how to adapt model architecture, weight mappings, operator implementations, and inference behavior to GPU and NPU engines.
- Collect engineering cases: Record key changes, compatibility problems, correctness checks, performance bottlenecks, and both successful and unsuccessful attempts.
- Consolidate reusable experience: Turn cases into searchable, reusable InferRSI knowledge / skills. Specify the model, engine, backend, versions, applicability conditions, and limitations.
- Validate reuse: Use concrete model cases to assess whether the workflow helps identify gaps, complete integration, and reduce repeated trial and error.
The owner will list the initial models, GPU / NPU engines, and environments in this issue. Record each stage separately as completed, requiring further work, or unverified.
Acceptance Criteria
- Define the initial model and engine scope, and document the three-stage integration path and its gaps.
- Produce reusable integration instructions and case records covering key implementation differences, validation methods, and lessons from resolving problems.
- Include traceable correctness / performance evidence for each case; explicitly mark stages that have not been executed as unverified.
- Add the experience to the existing knowledge system and make it discoverable through relevant model integration queries.
- Demonstrate how the knowledge supports new model integration using a concrete case, and record remaining limitations.
Activity
[-][Task] 0 Day Model Support：完善 Transformers → GPU → NPU 推理引擎适配经验[/-][+][Task] 0 Day Model Support: Reusable Integration Experience from Transformers to GPU and NPU Inference Engines[/+]agent-ci-infra-bot commentedon Nov 2, 2026
该 Issue 要求在 Transformers → GPU → NPU 引擎路径上沉淀可复用的新模型接入经验、案例与可检索知识，并以具体案例验证复用价值（Roadmap #316 工作项 8，负责人 @amy-why-3459）。结论：需要补充信息并由人决策，暂不适合无人值守开发——首批模型/引擎/环境与版本尚未列出，交付物与现有技能库的增量缺口未界定，Issue 也缺少仓库约定的状态块与完成验收人；正文没有明确要求修改代码的句子，且 Issue 已被指派。Labels: kind:task, area:knowledge（属知识线交付任务；因无 status:* 标签且 repository 规定 priority 仅随 status:proposed/accepted 使用，故不返回 priority）。
缺少首批模型 / 引擎 / 环境清单，范围与验收无法判定
AGENTS.md:117），并把版本与适用范围作为知识记录的必备字段（docs/knowledge-base.md:13-20）。交付物与现有知识资产高度重叠，未说明真正的增量缺口
src/inferrsi/data/skill_library/workflows/inference-engine-research/SKILL.md:15、references/model-aware-preparation.md:1-5）、NPU 侧 checkpoint 接入（src/inferrsi/data/skill_library/hardware/ascend/ascend-model-adaptation/SKILL.md:1-12）、权重加载路径（src/inferrsi/data/skill_library/methods/weight-loading-optimization/SKILL.md:1-6），以及 GPU→NPU 缺口清单先例（docs/drama3-npu-readiness.md:1-9）。SearchSpace 矩阵已把「Model adapter / FP8 / KV 复制」标为已覆盖（docs/searchspace-matrix.md:42）。Issue 未说明要新增或修订哪一条、哪个矩阵单元格仍是缺口。ascend-model-adaptation只覆盖 vllm-ascend 侧），并说明是修订既有 skill 还是新建，遵循npu-skill-factory的 artifact 路由。相关且已核对：父 Roadmap [Roadmap] InferRSI: Optimization, Evaluation, and Engineering #316（工作项 8 → [Task] 0 Day Model Support: Reusable Integration Experience from Transformers to GPU and NPU Inference Engines #315）、总体 Roadmap [Roadmap] InferRSI: reusable knowledge, verifiable inference optimization #61、GPU→NPU 先例 [RFC] Drama3 NPU task readiness: execution contracts, validation and knowledge supplementation #27。缺少仓库约定的三行状态块
kind:taskIssue 以三行状态块开头并由 acting agent 保持更新（docs/agent-collaboration.md:212-219），scoped-task模板也把它放在正文首位（.github/ISSUE_TEMPLATE/scoped-task.md:17-19）。本 Issue 是kind:task，但三项均缺失。**Current state:**、**Decision needed:**、**Next action:**三行，并保持更新。未记录完成验收/关闭责任人，也未记录接受范围或授权
docs/agent-collaboration.md:15、docs/agent-collaboration.md:17）；模板相应字段为.github/ISSUE_TEMPLATE/scoped-task.md:36、:40、:55。Status、Accepted decision or scoped delegation（链接）与Who accepts completion and may close三项，明确分片的独立评审与最终验收人。知识交付落在受保护路径，未说明知识准入声明与索引再生成要求
src/inferrsi/data/skill_library/**等（.github/knowledge-model-policy.json:4-5），main上knowledge-model-policy是必需检查，需要在 PR 正文给出唯一 JSON 声明，且模型 ID/effort 仅限策略清单（.github/knowledge-model-policy.json:24、docs/knowledge-model-policy.md:31、:46）；不合规需 @Dong1017 的例外审批。同时skill-index.json是生成物（docs/knowledge-base.md:17），新增/修订 skill 后需按tools/build_skill_index.py:1-6生成并刷新记录（docs/knowledge-base.md:271），catalog.py/skill_catalog.py同为保护路径。inferrsi knowledge search能命中。验收项 3 的可追溯证据缺少冻结的门槛与修订版本
evidence/<campaign-id>/，判定仅 PASS/FAIL/INCONCLUSIVE（docs/contributing.md:94-112、:175-184）。本 Issue 既未链接相应 acceptance Issue，也未给出所需 L0–L4 级别，无法产出可追溯证据。Reviewed commit: 2baebf6
Coverage: Snapshot: 2637 files, 42149780 bytes; skipped 0 entries; Initial model context: 28 files; other accepted files available through tools; Git workspace contains a fixed snapshot, not full upstream history