Skip to content

Latest commit

 

History

History
22 lines (16 loc) · 4.65 KB

v0.2.0.md

File metadata and controls

22 lines (16 loc) · 4.65 KB

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