Skip to content

Commit da06369

Browse files
committed
docs: streamline README and retain English skills
Replace long speech-engine sections with sourced C1 comparisons, remove the requested WIP and packaged-library blocks, and remove Chinese skill editions. Update references, package checks and generated source snapshots for the retained English library. Agent-assisted-by: Codex (implementation) Signed-off-by: David Chen <530634352@qq.com>
1 parent ddeba86 commit da06369

155 files changed

Lines changed: 330 additions & 4427 deletions

File tree

Some content is hidden

Large Commits have some content hidden by default. Use the searchbox below for content that may be hidden.

‎AGENTS.md‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -97,7 +97,7 @@ Agents and human reviewers **clear** work only with artifacts—not vibes. Green
9797

9898
## Skills
9999

100-
Portable operator guidance and candidate procedures live in [src/inferrsi/data/skill_library/](src/inferrsi/data/skill_library/README.md), with hardware-specific guidance under `hardware/<backend-or-host>/`, framework/repository guidance under its owner, and generic `methods/` and `workflows/`. The library owns complete procedures, metadata, references and cases. Engine translations live under their owner's `translations/zh-CN/`; supporting documents stay in owner-local resource directories. The [knowledge-base contract](docs/knowledge-base.md) and two matrices define ownership and coverage. Keep **AGENTS.md** short; follow skill bodies for procedures—skills must remain free of campaign SHAs.
100+
Portable operator guidance and candidate procedures live in [src/inferrsi/data/skill_library/](src/inferrsi/data/skill_library/README.md), with hardware-specific guidance under `hardware/<backend-or-host>/`, framework/repository guidance under its owner, and generic `methods/` and `workflows/`. The library owns complete procedures, metadata, references and cases. Keep library content in English; supporting documents stay in owner-local resource directories. The [knowledge-base contract](docs/knowledge-base.md) and two matrices define ownership and coverage. Keep **AGENTS.md** short; follow skill bodies for procedures—skills must remain free of campaign SHAs.
101101

102102
## Dev smoke (local)
103103

‎PHASE1_REPORT.md‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -2,7 +2,7 @@
22

33
> Historical snapshot from 2026-10-14. Its inventory, proposed next steps and
44
> sibling-workstream reports are not current implementation or validation status.
5-
> Use the [three-objective overview](README.md#three-objectives-of-the-current-wip),
5+
> Use the [three-objective overview](docs/architecture.md#three-objective-delivery-framing),
66
> [Reference Matrix](docs/reference-matrix.md) and
77
> [SearchSpace Matrix](docs/searchspace-matrix.md) for current framing and evidence.
88
> The original report below is preserved; new Magpie coordination evidence remains

‎README.md‎

Lines changed: 9 additions & 112 deletions
Original file line numberDiff line numberDiff line change
@@ -34,34 +34,6 @@ You drive profiling and code changes; InferRSI runs and evaluates the benchmark
3434

3535
For every CLI/Python operation — discovery, normalized reads, store candidates, experiments — see the [operation guide](docs/operations.md).
3636

37-
## Three objectives of the current WIP
38-
39-
These objectives frame the candidate-intelligence work; they are not another
40-
coverage matrix or claims that an autonomous system has been implemented.
41-
42-
| Objective | Question | Existing artifacts | Supported findings | Remaining work |
43-
|---|---|---|---|---|
44-
| **1. Roofline / Optimization Headroom** | What limits performance, how much headroom may remain, and which experiments are worth pursuing? | [Performance ceiling analysis](src/inferrsi/data/skill_library/methods/performance-ceiling-analysis/SKILL.md), measurement and search guidance | Source-backed analytical/empirical/Amdahl reasoning; estimates depend on workload, algorithm and measurement assumptions | Calibrate predictions and qualify application outcomes; no automatic ceiling calculator |
45-
| **2. Dense Commit Coordination** | How can ownership, concurrent changes, review, testing, integration, release and feedback sustain agent-driven development? | [Repository evolution analysis](src/inferrsi/data/skill_library/methods/repository-evolution-analysis/SKILL.md), [historical Magpie analysis](PHASE1_REPORT.md#workstream-2--repository-evolution-and-fast-iteration) | Historical commit/capability analysis and temporal-split methodology; statistics are not causal evidence of development speed | Dot's first-hand Magpie coordination case, then owner review; no implemented Magpie-style autonomous pipeline |
46-
| **3. Skill System** | How do we preserve, organize and expose reusable optimization intelligence? | [Reference Matrix](docs/reference-matrix.md), [SearchSpace Matrix](docs/searchspace-matrix.md), [skill library](src/inferrsi/data/skill_library/), explicit read-only CLI | Source-backed candidate procedures, knowledge and cases with scoped evidence; metadata listing and explicit instruction/resource access | Content review and separately authorized efficacy/compatibility validation; matrices do not automatically route tasks |
47-
48-
Dot owns first-hand extraction; the owner approves sources and reviews content
49-
and mapping; the integration workstream reconciles accepted material with the
50-
actual local system. Exactly two primary matrices remain. Green coverage means
51-
candidate intelligence with source support, not measured gains or completed A/B
52-
validation. The earlier [Phase 1 report](PHASE1_REPORT.md) is historical.
53-
54-
Reviewed paper intelligence extends [simulation calibration](src/inferrsi/data/skill_library/methods/calibrate-serving-simulation/SKILL.md),
55-
[serving-state decisions](src/inferrsi/data/skill_library/methods/assess-serving-preemption/SKILL.md),
56-
[training memory](src/inferrsi/data/skill_library/methods/plan-activation-residency/SKILL.md),
57-
[attention diagnosis](src/inferrsi/data/skill_library/hardware/cuda/diagnose-attention-bottlenecks/SKILL.md) and
58-
[skill/search methodology](src/inferrsi/data/skill_library/workflows/npu-skill-factory/SKILL.md). Exact paper
59-
versions, readsets, mechanism comparisons and cases remain linked through the two
60-
matrices. On 2026-10-16 the inventory is **43 branch-added candidates + 5 inherited
61-
skills = 48 directories**. This is a content-integration count, not an efficacy
62-
result. The earlier six-paper foundation archive remains an unavailable input;
63-
new Magpie coordination evidence is still pending.
64-
6537
<a id="results"></a>
6638

6739
## 📊 From profiling to a 30.60% latency reduction
@@ -89,38 +61,17 @@ Author-reported means from three requests after two warmups; not independently v
8961

9062
[Ascend measurements and rejected rounds →](docs/qwen-image-2.1-turbo-ascend.md) · [Upstream NPU patch](https://github.com/vllm-project/vllm-omni/pull/8729)
9163

92-
### Drama 3: a dedicated streaming TTS engine on one B300
93-
94-
[drama-3-single-b300-inference](https://github.com/david6666666/drama-3-single-b300-inference)
95-
is a standalone Python runtime built around native FP8 MoE, persistent GPU state,
96-
CUDA Graphs and incremental DAC audio decoding. Its recorded comparison with
97-
vLLM-Omni and SGLang-Omni covers C1–C64 throughput, TTFA and RTF, alongside WER/SIM
98-
quality scores, actual repeat counts and failed-run records.
99-
100-
Campaign status is **INCONCLUSIVE**: the original all-concurrency throughput target
101-
remains unmet and optimization is paused.
102-
103-
[Architecture & quickstart →](https://github.com/david6666666/drama-3-single-b300-inference#architecture) · [Recorded comparison →](https://github.com/david6666666/drama-3-single-b300-inference/blob/58a1882a4d941f8d21bfce6be37d3d5e7c9fde80/README.md#performance) · [Optimization lessons](evidence/drama3-b300-scale150/results/diagnostic-lessons.md) · [Acceptance issue](https://github.com/InferMatrixLab/InferRSI/issues/10)
104-
105-
### Qwen3-TTS CustomVoice on Ascend: imported optimization techniques
64+
### Standalone speech engines
10665

107-
The [standalone Qwen3-TTS engine](https://github.com/InferMatrixLab/qwen3-tts-engine)
108-
contributes three searchable peer-technique records: codec constant/state reuse,
109-
graph batching with request-owned state, and fixed CustomVoice prompt preparation.
110-
These are source-backed transfer candidates, not a built-in InferRSI runtime.
66+
Concurrency 1 (C1) results reported by each project. Values are **engine / vLLM-Omni / SGLang-Omni**; compare within a row, since workloads and audio-frame definitions differ.
11167

112-
The import's performance verdict is **INCONCLUSIVE**; accelerator inference and
113-
audio-quality validation were not rerun. Historical combined-engine reports do
114-
not establish per-feature or downstream framework gains.
68+
| Project | Device | Throughput (frames/s) ↑ | TTFA p50 (ms) ↓ | Report and status |
69+
| --- | --- | --- | --- | --- |
70+
| [Drama 3](https://github.com/InferMatrixLab/drama-3-single-b300-engine) | 1× B300 | **309.17** / 195.89 / 169.80 | **32.83** / 47.29 / 79.38 | [3-run medians](https://github.com/InferMatrixLab/drama-3-single-b300-engine/blob/58a1882a4d941f8d21bfce6be37d3d5e7c9fde80/docs/current-results.md); overall target unmet |
71+
| [Qwen3-TTS](https://github.com/InferMatrixLab/qwen3-tts-rsi-npu-engine) | 1× Ascend 910B1 | **61.12** / 12.71 / 4.31 | **50.41** / 263.51 / 439.72 | [3-run means](https://github.com/InferMatrixLab/qwen3-tts-rsi-npu-engine/blob/a29f8193b557c415175d70919bdfb9bd3dcd51fc/evidence/campaign-20261017/same-card-performance.md); source campaign accepted |
72+
| [MiniCPM-o 4.5](https://github.com/InferMatrixLab/minicpm-o-4.5-single-gpu-engine) | 1× L20X | **397.43** / 187.52 / 278.72 | **126.2** / 295.3 / 381.2 | [5-run means](https://github.com/InferMatrixLab/minicpm-o-4.5-single-gpu-engine/blob/a4f4778f312791c2d190b48e521141887ed160c8/docs/current-results.md); target unmet, quality unqualified |
11573

116-
[Evidence and reproduction index](evidence/qwen3-tts-ascend-import/results/README.md) ·
117-
[Knowledge records](src/inferrsi/data/knowledge_packs/README.md#qwen3-tts-ascend-source-import) ·
118-
[Scope issue](https://github.com/InferMatrixLab/InferRSI/issues/35)
119-
120-
```bash
121-
inferrsi knowledge search 'Qwen3-TTS' --tier peer \
122-
--repo-type qwen3_tts_engine --backend ascend --scenario audio.tts
123-
```
74+
MiniCPM's SGLang baseline returns a complete waveform, so its TTFA equals end-to-end latency. These external results were not rerun in InferRSI.
12475

12576
<a id="quickstart"></a>
12677

@@ -147,60 +98,6 @@ inferrsi context --repo-type vllm_omni --backend ascend \
14798

14899
Ready to measure your own model? [Connect your benchmark and quality checks →](docs/quickstart.md#real-remote-experiment)
149100

150-
### Read the packaged skill library
151-
152-
All 48 former root skills now live in
153-
[the skill library](src/inferrsi/data/skill_library/README.md), including their
154-
full procedures, metadata, references and cases. Hardware-specific guidance is
155-
grouped under `hardware/ascend`, `hardware/cuda` and `hardware/host`; the remaining
156-
skills use `frameworks`, `repositories`, `methods` and `workflows`. The 42 bilingual
157-
engine skill documents follow the same ownership layout: English entries live
158-
under `repositories/<engine>` and `workflows/inference-engine-research`, with
159-
complete Chinese bundles in each owner's `translations/zh-CN/`. Supporting
160-
Markdown and JSON live with their owner; the library root keeps only `README.md`
161-
and `skill-index.json`.
162-
163-
Install the optional parser with `python -m pip install 'inferrsi[skills]'` (or
164-
`python -m pip install '.[skills]'` from this checkout). The `skills` CLI defaults
165-
to the installed library's 48 registered skills; IDs stay unchanged. An explicit
166-
`--root /path/to/skills` before the subcommand still selects a custom flat tree or
167-
a categorized library with `skill-index.json`:
168-
169-
```bash
170-
inferrsi skills list --json
171-
inferrsi skills show weight-loading-optimization
172-
inferrsi skills resource <skill-name> <relative-resource-path>
173-
```
174-
175-
The [generated catalog](src/inferrsi/data/repo_packs/inferrsi/knowledge/README.md)
176-
also exposes the 48 skills through ordinary package knowledge commands, without
177-
the optional YAML parser:
178-
179-
```bash
180-
inferrsi knowledge search 'weight-loading' --repo-type inferrsi
181-
inferrsi knowledge show inferrsi.common.skills.weight_loading_optimization
182-
inferrsi read inferrsi.ascend.skills.ascend_graph_mode --material
183-
```
184-
185-
Each catalog record's `skill_documents.original` points to its full packaged
186-
procedure. Normalized reads retain the library resources and dependency bundle.
187-
Ascend and CUDA records have matching backend scopes; `common` means generic
188-
guidance, not demonstrated compatibility with every backend. Candidate state,
189-
source provenance and `UNKNOWN` licenses are preserved.
190-
191-
Listing reads frontmatter and sidecar metadata only. `show` selects one instruction
192-
body and lists resources; `resource` resolves one path for your file reader.
193-
The package ships the content, but reading it does not execute instructions,
194-
install agent configuration, select an optimization or promote a candidate.
195-
The Reference and SearchSpace matrices remain the two authoritative matrices.
196-
See the [knowledge-base contract](docs/knowledge-base.md) for ownership,
197-
compatibility and declared cross-skill references.
198-
Normalized reads and bounded Collect ([`KnowledgeReader`](docs/knowledge-operations.md))
199-
and authorized local Create/Update ([`KnowledgeStore`](docs/knowledge-store.md))
200-
are implemented Python interfaces, with a thin CLI route (`collect` / `read` /
201-
`resource`) for the read side — see the [operation guide](docs/operations.md).
202-
203-
204101
<a id="architecture"></a>
205102

206103
## 🧩 Architecture
@@ -252,7 +149,7 @@ for the 34 review points, qualifications and suggested acceptance checks.
252149
| [Skill catalog](src/inferrsi/data/repo_packs/inferrsi/knowledge/README.md) | Searchable records for 48 categorized skills, linking full packaged procedures and resources |
253150
| [Source audit](docs/reviews/source-audit-ac70272.md) | 34 review points with proposed remedies, qualifications and validation criteria |
254151
| [Daily learning prompt](docs/daily-learning-prompt.md) | 06:00 Asia/Shanghai L2C + performance workflow and prerequisites |
255-
| [Complete skill library (English / 中文)](src/inferrsi/data/skill_library/README.md) | Hardware, frameworks, repositories, methods and workflows, plus 42 bilingual five-engine skill nodes |
152+
| [Complete skill library](src/inferrsi/data/skill_library/README.md) | 48 registered skills and 21 English engine/workflow documents, organized by owner |
256153
| [vLLM-Omni experience records](src/inferrsi/data/repo_packs/vllm_omni/knowledge/README.md) | Backend-scoped tuning methods, sources, and applicability |
257154
| [Contributor guide](docs/contributing.md) | Development workflow, evidence contracts, and CI |
258155
| [Engine delivery template](docs/template/README.md) | Copyable engine scaffold, branding, evidence and schema-v1 catalog recipe |

‎docs/architecture.md‎

Lines changed: 5 additions & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -7,8 +7,7 @@ boundary called out separately.
77

88
## Three-objective delivery framing
99

10-
The current WIP develops three connected objectives, detailed in the
11-
[README overview](../README.md#three-objectives-of-the-current-wip):
10+
The current WIP develops three connected objectives:
1211

1312
1. **Roofline / Optimization Headroom:** source-grounded bounds, observed
1413
attainable limits, calibrated predictions and experiment selection. The
@@ -103,7 +102,7 @@ src/inferrsi/
103102
data/
104103
repo_packs/ Contribution descriptors and versioned tuning knowledge
105104
knowledge_packs/ Dependency and peer knowledge, preserving existing tiers
106-
skill_library/ Complete categorized skills + bilingual procedures/resources
105+
skill_library/ Complete categorized English skills and resources
107106
platform_packs/ Backend metadata and qualification context
108107
execution.py Compatibility entrypoint -> harness.runner
109108
evaluation.py Compatibility exports -> harness.evaluation
@@ -252,9 +251,9 @@ lives under `hardware/ascend`, `hardware/cuda` or `hardware/host`; other skills
252251
are grouped by framework, repository, method or workflow. Full procedures,
253252
sidecars, references and cases ship with the package. Engine entry skills live
254253
under `repositories/<engine>/`, the collection router under
255-
`workflows/inference-engine-research/`, and complete Chinese editions under
256-
each owner's `translations/zh-CN/`. Supporting documents are owner-local; only
257-
`README.md` and `skill-index.json` remain at the library root. An explicit
254+
`workflows/inference-engine-research/`. All library content is in English.
255+
Supporting documents are owner-local; only `README.md` and `skill-index.json`
256+
remain at the library root. An explicit
258257
`--root` can still select a custom flat skills tree or a registry-based library. The CLI uses PyYAML from the `skills` extra,
259258
lists metadata, reads a named body or resolves a named resource for an external
260259
file reader. It never executes, installs or promotes skills. Core knowledge

‎docs/contributing.md‎

Lines changed: 5 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -52,13 +52,13 @@ their generated abstracts with `tools/build_skill_index.py`. Keep all skill
5252
content in the canonical library. Preserve source state and
5353
`UNKNOWN` license declarations; moving content does not resolve licensing.
5454

55-
Keep English engine procedures under `repositories/<engine>/` and the collection
56-
router under `workflows/inference-engine-research/`, with each complete Chinese
57-
bundle in its owner's `translations/zh-CN/`. Preserve both editions' sub-skills
58-
and references. Keep supporting Markdown and JSON with their owner; the library
55+
Keep all library content in English. Engine procedures belong under
56+
`repositories/<engine>/` and the collection router under
57+
`workflows/inference-engine-research/`, with their sub-skills and references.
58+
Keep supporting Markdown and JSON with their owner; the library
5959
root contains only `README.md` and `skill-index.json`.
6060
Catalog abstracts link to packaged bodies via `skill_documents.original` for
61-
migrated skills or the existing language fields for bilingual imports. Update
61+
migrated skills or `skill_documents.en` for engine procedures. Update
6262
affected source content and catalog records together. Source snapshot revisions
6363
identify research evidence, not campaign defaults. Check linked resources,
6464
licenses, local links and reads from a noneditable wheel installation.

‎docs/daily-learning-prompt.md‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -121,7 +121,7 @@ E. 检查、候选 PR 与晋升
121121
17. 学习者只写 staging 目录;独立校验/发布步骤核对改动 allowlist 后复制:
122122
- src/inferrsi/data/skill_library/ 下按主要归属放置的候选技能:
123123
hardware/<backend>/、frameworks/<framework>/、repositories/<repo>/、methods/ 或 workflows/;
124-
翻译放所属技能的 translations/<language>/,参考资料放该技能的 references/;
124+
技能库内容统一使用英文,参考资料放所属技能的 references/;
125125
- src/inferrsi/data/repo_packs/<repo>/knowledge/ 的知识记录和导航;
126126
- 本次 acquisition/learning manifest、评测摘要与性能候选队列的可追溯归档。
127127
新增 vllm 或 vllm_ascend 记录时创建该 pack 的知识导航;知识身份不表示运行支持。

0 commit comments

Comments
 (0)