InferRSI v0.1.0

Source: https://github.com/InferMatrixLab/InferRSI/tree/6c9f90daae2d404915378b910f581b4dcff7e7da

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.

What's Changed

Other changes

• docs: OmniRSI repo template (RSI Team chapters) by @hsliuustc0106 in #4
• feat: portable RSI repo scaffold (copyable layout + four-lane CI) by @hsliuustc0106 in #5
• docs: design draft → scaffold test design map (RSI + ProductResearch locks) by @hsliuustc0106 in #6
• docs: portable root AGENTS.md for coding agents by @hsliuustc0106 in #7
• feat: bootstrap CLI experiment workbench and diffusion knowledge by @david6666666 in #2
• feat: add three-tier knowledge catalog and daily learning prompt by @david6666666 in #11
• feat: add bilingual knowledge for five inference engines by @david6666666 in #9
• Record mixed audio scheduling and codec workspace isolation lessons by @david6666666 in #13
• Record native C64 reference accounting and feasibility evidence by @david6666666 in #14
• Record PDL feasibility and native execution audit lessons by @david6666666 in #16
• Record metadata readiness and native Lt diagnostic outcomes by @david6666666 in #18
• Record measured occupancy and staged native prefix reuse lessons by @david6666666 in #19
• Record exact prefix reuse and compressed allocation lessons by @david6666666 in #20
• Record production prefix service outcome and paused results by @david6666666 in #22
• docs(collaboration): define issue-led agent coordination by @Dong1017 in #23
• RFC 17: knowledge content and contract foundation (K0/K1) by @Dong1017 in #12
• feat: add normalized knowledge reads and bounded Collect by @Dong1017 in #24
• feat: publish store-owned candidate bundles in an authorized local store by @Dong1017 in #25
• fix: keep declared knowledge dependencies in bundle revisions by @Dong1017 in #26
• Drama3 NPU readiness: stage/state map, correctness + measurement contracts, noninterference fixture by @Dong1017 in #28
• docs: add privacy and security rules for shared artifacts by @zengchuang-hw in #34
• Add Qwen3-TTS Ascend optimization knowledge and evidence index by @amy-why-3459 in #36
• docs: onboarding paths, operation guide and implemented Collect/Store status (#32 PR 1) by @Dong1017 in #37
• feat: read-only store path and knowledge CLI bridge (#32 PR 2) by @Dong1017 in #39
• Organize complete skill library and add bilingual architecture/audit docs by @david6666666 in #44
• Remove obsolete root skills directory by @david6666666 in #45
• Streamline README and organize the English skill library by @david6666666 in #46
• docs: standardize generated engine naming and attribution by @zengchuang-hw in #47
• fix: collect text output contract, clean-install skills path and operation guide accuracy (#32 corrective) by @Dong1017 in #42
• fix: harden process cleanup, evidence integrity and CI coverage by @david6666666 in #52
• chore: remove unused root recipe and benchmark scaffolding by @david6666666 in #59
• feat: add report-only knowledge PR admission by @Dong1017 in #62
• fix: prevent shared-head knowledge admission overwrite by @Dong1017 in #64
• Batch committed Git blob reads in the skill index generator by @Dong1017 in #70
• Distill ComfyUI graph validity and model residency by @Dong1017 in #73
• Distill MLX lazy evaluation and stream lifetimes by @Dong1017 in #72
• feat: add bounded skill evaluation and retain neutral pilot results by @Dong1017 in #69
• Distill MLX-LM generation and cache lifecycles by @Dong1017 in #74
• Distill llama.cpp runtime state and cancellation contracts by @Dong1017 in #71
• Distill stable-diffusion.cpp residency and segment lifetimes by @Dong1017 in #77
• Distill SGLang diffusion request and component lifetimes by @Dong1017 in #79
• Trace Diffusers invocation and scheduler state by @Dong1017 in #76
• knowledge: trace vLLM-Omni abort acknowledgment and reclamation by @Dong1017 in #78
• knowledge: import DeepInfer mechanisms as candidates by @Dong1017 in #68
• knowledge: synthesize repository ownership and source scope by @Dong1017 in #80
• fix: balance skill-eval group order and correct timing claims by @Dong1017 in #82
• docs: explain experiment and knowledge loops in How It Works by @zengchuang-hw in #56
• docs: require architecture and end-to-end optimization figures in engine READMEs by @david6666666 in #85
• fix: stop stale knowledge sweeps from failing merged PRs by @Dong1017 in #86
• docs: add closure summary card rule and template by @Dong1017 in #87
• docs: distill fastsafetensors TP transfer and completion contracts by @Dong1017 in #96
• fix: harden deterministic skill grading and offline failure handling by @Dong1017 in #94
• docs: add task-specific benchmark selection knowledge by @david6666666 in #102
• docs: record bounded vLLM-RLT intake by @Dong1017 in #104
• docs: distill h3.c runtime residency into repository knowledge by @david6666666 in #110
• feat: add model-aware inference research preparation by @david6666666 in #111
• docs: record bounded AFD plugin intake by @Dong1017 in #105
• docs: distill AKG affine tail-guard contracts by @Dong1017 in #107
• docs: distill Ascend communication benchmark contracts by @Dong1017 in #108
• docs: distill DLRover checkpoint persistence and replica transport by @Dong1017 in #95
• knowledge: ground CannBot stage evidence in Ascend analysis by @Dong1017 in #114
• feat: add configurable offline skill addition and update comparisons by @Dong1017 in #109
• feat: distill AFD rank and P2P payload ownership by @Dong1017 in #120
• docs: clarify vLLM-Ascend guidance versions and responsibilities by @Dong1017 in #130
• docs: Apache-2.0 LICENSE and v0.1.0 Developer Preview notes by @hsliuustc0106 in #135
• fix: invalidate an old admission verdict before evaluating by @Dong1017 in #98
• feat: add opt-in skill selection and context to collect by @Dong1017 in #129
• docs: connect llama.cpp knowledge owners and source scopes by @Dong1017 in #124
• docs: route SGLang engine, HiCache and diffusion knowledge by @Dong1017 in #136
• fix: give each admission target a bounded share of the sweep by @Dong1017 in #146
• docs: route PyTorch compiler and checkpoint knowledge by scope by @Dong1017 in #144
• docs: add a campaign plan template for the plan evidence segment by @zengchuang-hw in #31
• Add campaign gain summary chart, rendered when a run ends by @amy-why-3459 in #92
• fix: isolate local command input and clarify audit aggregation by @Dong1017 in #145
• docs: route vLLM-Omni engine and abort guidance by @Dong1017 in #139
• feat: distill Ascend offload descriptor ownership by @Dong1017 in #123
• feat: add MLX discovery and execution knowledge cards by @Dong1017 in #119
• feat: adapt offline Flare delivery evidence by @Dong1017 in #151
• feat: derive offline goal checkpoints and next actions by @Dong1017 in #150
• knowledge: distill qualified Hopper guidance from KDA KernelWiki by @jiangkuaixue123 in #127
• fix: isolate trigger PR admission from sweep health by @Dong1017 in #154
• feat: plan and report skill PR evaluations with explicit authorization by @Dong1017 in #118
• docs: connect TensorRT-LLM methodology and disaggregation guidance by @Dong1017 in #142
• docs: distinguish Diffusers offload and invocation guidance by @Dong1017 in #134
• fix: accept pinned Flare derived analysis after sidecar save by @Dong1017 in #155
• fix: retain safe Skill Eval publisher HTTP diagnostics by @Dong1017 in #163
• feat: add bounded offline harness improvement with quality and cost gates by @david6666666 in #158
• ci: add model-free Skill Eval plan dispatch by @Dong1017 in #165
• fix: retain safe publisher permission requirement diagnostics by @Dong1017 in #168
• fix: grant Skill Eval publisher PR comment write access by @Dong1017 in #170
• feat: add a pure conditional Roofline helper by @Dong1017 in #167
• fix: compact Skill Eval comments while retaining evidence boundaries by @Dong1017 in #173
• docs: define bounded Yibuapi agent preflight by @Dong1017 in #172
• feat: add offline bounded Messages agent contract by @Dong1017 in #174
• feat: enable bounded Yibuapi Skill Eval execution by @Dong1017 in #181
• knowledge: drop default sidecar fields, keep case digests in records by @david6666666 in #176
• fix: bind explicit token-stop headroom for paid pairs by @Dong1017 in #192
• knowledge: distill DwarfStar with portable AREX workflow by @david6666666 in #184
• docs: correct knowledge-policy enforcement status and architecture map by @Dong1017 in #182
• perf: reduce knowledge retrieval manifest work with measured evidence by @david6666666 in #193
• knowledge: distill Splash with portable AREX workflow by @david6666666 in #186
• knowledge: distill Lily with portable AREX workflow by @david6666666 in #188
• chore: remove retrieval audit and evidence artifacts by @david6666666 in #195
• knowledge: add Husky document-only research candidate by @david6666666 in #191
• docs: distill LoopX goal authority and quota boundaries by @Dong1017 in #153
• evidence: freeze approved 8K search and retain FP8 optimization results by @jiangkuaixue123 in #97
• skills: add goal-driven Ascend video diffusion optimization workflow by @BLANKETusers in #199
• fix: bound admission writer queue and rejected status retries by @Dong1017 in #204
• fix: diagnose unknown status rejections without raw response data by @Dong1017 in #205
• docs: add a generated skill coverage reverse view by @Dong1017 in #194
• fix: recognize exact status-capacity messages in string responses by @Dong1017 in #206
• feat: distill recurrent KV finalization and retirement by @Dong1017 in #121
• feat: distill diffusion loader and storage boundaries by @Dong1017 in #122
• fix: refresh only affected admissions after merges by @Dong1017 in #208
• knowledge: distill Project Maya with AREX source and CPU verification by @david6666666 in #211
• docs: add h3-metal comparison table with Device column by @david6666666 in #210
• diagnostic: capture one encrypted admission rejection by @Dong1017 in #212
• fix: contain skill resource walks and harden bundle identity inputs by @Dong1017 in #196
• fix: make run state and comparison verdicts match their evidence by @Dong1017 in #202
• fix: isolate PR admission writes and remove temporary capture by @Dong1017 in #213
• fix: make source and device identity reject unsupported evidence by @Dong1017 in #197
• fix: separate push reconciliation completion from unresolved admissions by @Dong1017 in #214
• feat: derive revision-bound Skill Eval review guidance by @Dong1017 in #216
• knowledge: distill vllm-metal with portable AREX workflow by @david6666666 in #219
• knowledge: distill mistral.rs Metal paged-attention contracts by @david6666666 in #224
• feat: label-gated Skill Eval with cumulative PR budget and live authorization by @Dong1017 in #223
• docs: Apple silicon playbook and h3-metal lessons (measured on M5 Max) by @zuiho-kai in #221
• knowledge: distill MLX Swift LM lifecycle and Metal recurrence by @david6666666 in #226
• knowledge: MLC LLM compiled Metal source contracts by @david6666666 in #228
• refactor: remove the unread platform_packs descriptors and fix the generated-path guidance by @Dong1017 in #201
• fix(skill-eval): separate trusted policy and PR run identities by @Dong1017 in #229
• fix(skill-eval): preserve same-plan results after disabling execution by @Dong1017 in #230
• knowledge: audit h3-metal FastH3 execution and fallback contracts by @david6666666 in #232
• Distill Flare preparation and session-state contracts by @Dong1017 in #217
• docs: route vLLM engine, admission and KV knowledge by @Dong1017 in #138
• skill: add fixed-workload claim audit procedure by @Dong1017 in #159
• skill: clarify fixed-workload report aggregation by @Dong1017 in #160
• knowledge: qualify causal-conv1d forward and state contracts by @Dong1017 in #164
• knowledge: trace vLLM GDN convolution phases and state by @Dong1017 in #169
• docs: correct LearningToCommit implementation attribution by @Dong1017 in #233
• knowledge: add Local Dream Android diffusion boundary candidate by @david6666666 in #236
• feat: add bounded diagnostic evaluation and daily admission by @Dong1017 in #237
• config: bind audited diagnostic pilot admission by @Dong1017 in #240
• feat: add bounded one-shot full-context evaluation by @Dong1017 in #241
• skills: extract backend-agnostic video diffusion optimization skeleton by @BLANKETusers in #207
• fix: preserve trial deadline across live authorization by @Dong1017 in #242
• fix: keep evaluation reporting free of optional parser imports by @Dong1017 in #244
• knowledge: add a source-pinned RTK commit-to-lesson example by @Dong1017 in #245
• knowledge: qualify HF Qwen3.5 conv/GDN bindings by @Dong1017 in #250
• feat: explain retained SkillEval answers and component results by @Dong1017 in #247
• feat: add an installable Codex plugin preview by @zengchuang-hw in #162
• docs: qualify SGLang GDN and causal-conv caller contracts by @Dong1017 in #251
• knowledge: add source-grounded Lithos Metal candidate skills by @jiangkuaixue123 in #117
• docs: settle the workflow sidecar shape and let sidecars supply recommendations by @Dong1017 in #243
• test(indexer): retain retrieval efficiency and accuracy evidence for #255 by @Dong1017 in #256
• feat(skill-eval): add manual local artifact evaluation for #253 by @Dong1017 in #258
• test(indexer): retain failed whole-token accuracy experiment for #255 by @Dong1017 in #259
• fix(skill-eval): write frozen local inputs as UTF-8 bytes by @Dong1017 in #261
• knowledge(hardware): audit sources and deepen memory, graph and collective guidance by @david6666666 in #262
• evidence(indexer): consolidate negative accuracy and cue-study results for #255 by @Dong1017 in #265
• feat(skilleval): add manual strict verifier qualification by @Dong1017 in #266
• feat(skilleval): bounded cloud generation and isolated artifact replay by @Dong1017 in #267
• feat(skilleval): fixed manual historical cloud generation job by @Dong1017 in #269
• Recover historical MindSpeed provenance and allocator prerequisites by @Dong1017 in #257
• feat(skilleval): explain artifact deltas and scoped next steps by @Dong1017 in #271
• feat(indexer): add opt-in raw-string substring ranking by @Dong1017 in #272
• perf(indexer): avoid redundant view sizing after item limit by @Dong1017 in #270
• feat: qualify a bounded synthetic CPU KV reclamation task by @Dong1017 in #274
• feat: add bounded KV editing and scripted actor qualification by @Dong1017 in #275
• feat: add inactive bounded KV pilot launcher by @Dong1017 in #276
• Activate one owner-frozen CPU KV editing pair by @Dong1017 in #280
• Retain closed KV failure diagnostics without retry by @Dong1017 in #281
• docs: qualify Sarathi chunked-prefill budget case by @Dong1017 in #277
• docs: retain DistServe topology-constrained phase planning by @Dong1017 in #278
• docs: add Mooncake prefix reuse feasibility reference by @Dong1017 in #279
• docs: complete applicability metadata for 17 methods skills by @Dong1017 in #263
• Bound one additional KV pair to the PR275 budget exception by @Dong1017 in #287
• Bind the single additional KV pair to the refreshed public freeze by @Dong1017 in #290
• Test the manual SkillEval CLI handoff lifecycle by @Dong1017 in #291
• Make nonpaid SkillEval plans actionable for PR authors by @Dong1017 in #292
• ci: report closing syntax inside PR body text by @Dong1017 in #296
• policy: admit GPT-6.1 Terra at xhigh and max by @Dong1017 in #298
• docs: add FlexGen offload feasibility reference by @Dong1017 in #288
• ci: automate daily releases with Python and Codex plugin packages by @zengchuang-hw in #239
• docs: define the label list, rules and nightly drift report by @Dong1017 in #299
• fix: preserve unknown command counts in author imports by @Dong1017 in #300
• docs: add cache-aware ASR stream-boundary reference by @Dong1017 in #301

New Contributors

• @hsliuustc0106 made their first contribution in #4
• @david6666666 made their first contribution in #2
• @Dong1017 made their first contribution in #23
• @zengchuang-hw made their first contribution in #34
• @amy-why-3459 made their first contribution in #36
• @jiangkuaixue123 made their first contribution in #127
• @BLANKETusers made their first contribution in #199
• @zuiho-kai made their first contribution in #221

Full Changelog: https://github.com/InferMatrixLab/InferRSI/commits/v0.1.0