InferRSI v0.2.0

Source: https://github.com/InferMatrixLab/InferRSI/tree/223ee523115d8d59eebf4568fd6ac34f4708b1d4

These notes describe software changes. Release builds and scaffold checks do not establish campaign PASS, model quality or service performance. Any such claim requires its acceptance issue and complete frozen evidence.

Reviewed v0.2.0 highlights

InferRSI v0.2.0

Highlights compared to v0.1.1. The published Release identifies the exact source revision and validated package assets.

Highlights since v0.1.1

• Three packaged optimization tasks. Run two-NPU MoE forward, single-NPU Qwen-Image 2.1 denoising, and Qwen3-TTS 1.7B CustomVoice workloads from their task-owned baselines and evaluation contracts. Qwen denoising also supports multi-round sessions, frozen candidates and budgets, independent feedback, and a final retest. MoE and TTS currently expose single-candidate entries. Task guide · Session workflow
• A unified optimization lifecycle. Ordinary project campaigns and benchmark sessions share the Harness controller, with retained attempts, interrupted-run recovery, explicit acceptance criteria, usage observations, and independent final evaluation. User criteria are preserved; ceiling analysis fills missing criteria instead of silently replacing them. #338 · #400
• Inspect the experiment, not just the verdict. Offline and live dashboards now show performance curves, event tracks, stage timing comparisons, retrieval explanations, per-round searches, and candidate/evidence drill-down. Unknown observations remain unknown rather than becoming invented zeroes. #319 · #401 · #405 · #428
• Stronger records and replay boundaries. A local SQLite RunStore, standardized Apps output, complete preparation archives, and saved requests bound to frozen worker specifications make execution and evidence provenance more explicit. The module migration is incremental, not a claim that all proposed architecture work is complete. #398 · #397 · #424 · #427
• More practical optimization examples and evidence. The README now maps image, video, and speech workloads across B300, RTX 3090, L20X and Ascend devices to their source experiments and upstream changes: H3 Ref2VA/H3-Image, MiniCPM-o, Qwen3-Omni, LFM3-Omni and LingBot-World alongside Qwen-Image. The merged four-device 910C H3 evidence remains INCONCLUSIVE overall. Some links point to upstream reports or unmerged evidence PRs, not new executable recipes or independently reproduced results. Results and limitations · #317
• Broader integration and reusable knowledge. Native WorkSwarm packaging and a DeepSeek Harness integration join the existing Codex path. Knowledge additions include MindIE and audited Ascend, TTS and serving techniques, while per-round search/adoption records show what context was delivered. A delivered record is not proof that a model consumed it. MindIE #393 · Search/adoption #384

Compatibility and scope

• Python 3.10+; no PyPI upload is part of this release process.
• The synthetic CPU KV slice tooling was retired during the v0.1.x series. Use the current documented workloads instead. Retirement
• Knowledge and single-CPU execution ownership moved incrementally into their canonical packages; some compatibility shells remain. #418 · #421
• Packaging/CPU checks establish software behavior only. They do not certify model quality, GPU/NPU support or service-level gains. Preserve each campaign's PASS/FAIL/INCONCLUSIVE verdict and exact workload. Automatic model-backed search and automatic knowledge promotion remain planned.
• Existing h3-metal and standalone Drama 3/Qwen3-TTS/MiniCPM engine comparisons were already in v0.1.1; they are not presented as new v0.2.0 implementations.

Full comparison from v0.1.1

What's Changed

Other changes

• docs: describe conditions and advice for 22 hardware skills by @Dong1017 in #283
• docs: describe conditions and advice for 29 repository and framework skills by @Dong1017 in #285
• feat: distinguish resource-layer exclusions and unresolved coverage by @Dong1017 in #282
• ci: summarize release notes with a configurable AI API by @zengchuang-hw in #313
• docs: require explicit pull request review verdicts by @zengchuang-hw in #347
• [Feat][Knowledge] Distill Ascend RC model and fusion performance history by @jiangkuaixue123 in #354
• Add offline and live optimization workflow dashboard by @david6666666 in #319
• fix(labels): exempt outside-app labels from the drift report by @Dong1017 in #343
• Qualify RSI trials with explicit completeness and lifecycle checks by @amy-why-3459 in #341
• [BugFix][Dashboard] Render external Ascend optimization records by @jiangkuaixue123 in #358
• feat: add native WorkSwarm plugin and release artifact by @zengchuang-hw in #327
• knowledge: Drama-3 decode graphs on Ascend 910B1 by @amy-why-3459 in #132
• chore: retire the synthetic CPU KV slice tooling by @Dong1017 in #357
• feat(dsh): add structured iteration form and SSH execution by @zengchuang-hw in #361
• knowledge: audit Infera serving-search result provenance by @Dong1017 in #332
• docs: anchor the KV retirement page to the v0.1.4 release by @Dong1017 in #369
• knowledge: distinguish Xe-Forge trial origin and timing by @Dong1017 in #322
• fix: read shared provenance in trusted SkillEval planning by @Dong1017 in #360
• knowledge: audit DeepSeek C expert-cache byte validity by @Dong1017 in #324
• docs: add an English Codex plugin installation guide by @zengchuang-hw in #330
• fix(labels): align labels.json with live area:harness by @Dong1017 in #373
• docs: record the target module layout and import gate by @Dong1017 in #376
• test: lock the diffusion reference-PSNR ranking regression by @Dong1017 in #375
• knowledge: validate NInfer state and completion contracts by @Dong1017 in #325
• test: add a fake prepare-to-assess handoff loop by @Dong1017 in #377
• test: connect one CPU offline read, record, and assessment by @Dong1017 in #378
• chore: move skill-eval fixtures out of docs/examples by @Dong1017 in #381
• docs: archive stale planning notes under docs/plans by @Dong1017 in #383
• prototype: reduce knowledge projection and declaration churn by @Dong1017 in #359
• Add source-bound TTS playout benchmark guidance by @Dong1017 in #339
• evidence: track H3 Ref2VA on four logical Ascend 910C devices by @Dong1017 in #317
• docs: P0a contracts, schema, fixtures, and D023 errata by @Dong1017 in #391
• docs: propose the first-wave Knowledge compatibility gate (P0b) by @Dong1017 in #395
• harness: add local SQLite RunStore core (H01) by @Dong1017 in #398
• Review measured request profiles and decisions across optimization rounds by @david6666666 in #389
• Integrate project-scoped MindIE knowledge into InferRSI by @jiaran-king in #393
• docs: InferRSI × vLLM-Omni Optimization Results by @zengchuang-hw in #415
• Compare stage timings and inspect knowledge searches for each profile by @david6666666 in #401
• feat: share RSI optimization management across project and benchmark tasks by @Bounty-hunter in #338
• knowledge: move seven bodies into inferrsi.knowledge and revise the manifest (K01) by @Dong1017 in #418
• feat: record per-round knowledge searches and adoption evidence by @jiaran-king in #384
• harness: move single-CPU execute and the process domain into H (H02) by @Dong1017 in #421
• Publish standardized Apps run output by @zengchuang-hw in #397
• Use user acceptance criteria, with ceiling-analysis fallback by @amy-why-3459 in #400
• feat: Dashboard 实验性能曲线、事件轨道与证据下钻 by @amy-why-3459 in #405
• Archive complete Apps preparations by @zengchuang-hw in #424
• Bind saved Apps requests to frozen worker specs by @zengchuang-hw in #427
• policy: admit Claude Fable 5.2 at xhigh and max by @Dong1017 in #439
• feat: 补齐知识检索解释与逐轮候选证据导航 by @amy-why-3459 in #428
• knowledge: move packaged resources into knowledge/data and harness/data (K02) by @Dong1017 in #426
• Route fixed artifacts through Apps storage mounts by @zengchuang-hw in #432
• Project fixed Harness records into assessment evidence by @zengchuang-hw in #433
• release: prepare one-off formal v0.2.0 with reviewed highlights by @hsliuustc0106 in #442
• Compose standalone knowledge reads in Apps by @zengchuang-hw in #436

New Contributors

• @jiaran-king made their first contribution in #393
• @Bounty-hunter made their first contribution in #338

Full Changelog: v0.1.1...v0.2.0