|
| 1 | +# InferRSI v0.2.0 |
| 2 | + |
| 3 | +Highlights compared to v0.1.1. The published Release identifies the exact source revision and validated package assets. |
| 4 | + |
| 5 | +## Highlights since v0.1.1 |
| 6 | + |
| 7 | +- **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](https://github.com/InferMatrixLab/InferRSI/blob/57f8e8a4915b38201d9fa58d4d05acbe3c4199ee/docs/engineering-benchmarks.md) · [Session workflow](https://github.com/InferMatrixLab/InferRSI/blob/57f8e8a4915b38201d9fa58d4d05acbe3c4199ee/docs/optimization-benchmarks.md) |
| 8 | +- **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](https://github.com/InferMatrixLab/InferRSI/pull/338) · [#400](https://github.com/InferMatrixLab/InferRSI/pull/400) |
| 9 | +- **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](https://github.com/InferMatrixLab/InferRSI/pull/319) · [#401](https://github.com/InferMatrixLab/InferRSI/pull/401) · [#405](https://github.com/InferMatrixLab/InferRSI/pull/405) · [#428](https://github.com/InferMatrixLab/InferRSI/pull/428) |
| 10 | +- **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](https://github.com/InferMatrixLab/InferRSI/pull/398) · [#397](https://github.com/InferMatrixLab/InferRSI/pull/397) · [#424](https://github.com/InferMatrixLab/InferRSI/pull/424) · [#427](https://github.com/InferMatrixLab/InferRSI/pull/427) |
| 11 | +- **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](https://github.com/InferMatrixLab/InferRSI/blob/57f8e8a4915b38201d9fa58d4d05acbe3c4199ee/README.md#results) · [#317](https://github.com/InferMatrixLab/InferRSI/pull/317) |
| 12 | +- **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](https://github.com/InferMatrixLab/InferRSI/pull/393) · [Search/adoption #384](https://github.com/InferMatrixLab/InferRSI/pull/384) |
| 13 | + |
| 14 | +## Compatibility and scope |
| 15 | + |
| 16 | +- Python 3.10+; no PyPI upload is part of this release process. |
| 17 | +- The synthetic CPU KV slice tooling was retired during the v0.1.x series. Use the current documented workloads instead. [Retirement](https://github.com/InferMatrixLab/InferRSI/commit/1246348f252dbc7e473efb1752c2d81da15e23fd) |
| 18 | +- Knowledge and single-CPU execution ownership moved incrementally into their canonical packages; some compatibility shells remain. [#418](https://github.com/InferMatrixLab/InferRSI/pull/418) · [#421](https://github.com/InferMatrixLab/InferRSI/pull/421) |
| 19 | +- 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. |
| 20 | +- 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. |
| 21 | + |
| 22 | +[Full comparison from v0.1.1](https://github.com/InferMatrixLab/InferRSI/compare/v0.1.1...v0.2.0) |
0 commit comments