diff --git a/docs/knowledge-base.md b/docs/knowledge-base.md
index 837652f24..d0b3eee6a 100644
--- a/docs/knowledge-base.md
+++ b/docs/knowledge-base.md
@@ -83,6 +83,36 @@ is not promoted by import, translation, a passing reader test or matrix placemen
External/internal origin and Skill/Knowledge/Case kind are attributes; contribution,
dependency and peer remain independent repository relationships.
+### Applicability keys and recommendations
+
+Sidecar `applicability` keys follow the contract settled in
+[issue #91](https://github.com/InferMatrixLab/InferRSI/issues/91#issuecomment-6512087181).
+
+| Key | States | Must not |
+|---|---|---|
+| `scope` | What the procedure applies to | Become a capability or efficacy claim |
+| `prerequisites` | What must hold first: access, tooling, authorization | Restate the input contract |
+| `input_contract` | What the caller supplies | Enumerate the procedure's own steps |
+| `produced_artifact` | What it emits, at the body's evidence level | Upgrade a proposal into a validated configuration or measured result |
+| `failure_modes` | How the procedure yields a wrong or unusable result | Restate `limitations` |
+| `refusal_point` | Where the body says it withholds a conclusion | Invent a refusal the body does not state |
+| `hardware` | A device dependency the body establishes | Be inferred from a source's hardware |
+
+`recommendations` holds owner-supplied procedure advice. `tools/build_skill_index.py`
+copies a declared list unchanged. A declared empty list states that the procedure
+has no specific advice and does not fall back; only an absent field produces the
+generic navigation defaults.
+
+`limitations` is not owner-only. The generator emits four sentences of its own and
+then appends the sidecar's `limitations` entries. `failure_modes` is therefore a
+separate field rather than more `limitations` entries.
+
+No mechanical allowlist restricts `applicability` keys. The packaged library uses
+nine distinct keys: the seven above plus `architectures` and `excluded_mechanisms`,
+which appear only in `hardware/cuda/reason-hopper-kernels`. A closed key set would
+reject that sidecar or have to grandfather it. Revisiting this is proposed for
+batch 4 of issue #91 and is not agreed.
+
## Sources and revisions
| Type | Identity and scope | Not interchangeable with |
diff --git a/docs/skill-coverage.md b/docs/skill-coverage.md
index ba3e1d7ad..81869415a 100644
--- a/docs/skill-coverage.md
+++ b/docs/skill-coverage.md
@@ -4,7 +4,7 @@ Generated from the registry, the skill sidecars and the two maintained
matrices. **Do not edit by hand**; regenerate after any registry, skill,
matrix or frozen-plan change.
-- Input digest: `sha256:827a48e08ad75f86c7f87c7b8ef2ec21f79fb8b897dc1f135f1d77723436e947`
+- Input digest: `sha256:b30daf60ce87156416fe29086c3c086f7b59d12a7d0f0b47f9540d796f4079a1`
- Regenerate: `python tools/build_skill_coverage.py`
- Verify without writing: `python tools/build_skill_coverage.py --check`
@@ -118,26 +118,26 @@ optimization is fully covered or that any procedure has measured benefit.
| Skill | Backend | Declared task | Applicability | Conditions | SearchSpace layer / stage | Reference Matrix | State | Gaps |
|---|---|---|---|---|---|---|---|---|
-| [assess-serving-preemption](../src/inferrsi/data/skill_library/methods/assess-serving-preemption/SKILL.md) | common | Assess whether LLM serving queue delays justify token-boundary priority preemption, accounting for suspended KV, host ... | **unknown** | — | 1 Accelerator · Diagnose / Configure / Optimize / Verify
2 RAM/Host · Diagnose / Configure / Optimize / Verify | 1 · Systems sources (Orca — OSDI 2022 proceedings) | candidate · not_run · 1 readset(s) | no-applicability, no-eval |
-| [assess-split-backward-schedule](../src/inferrsi/data/skill_library/methods/assess-split-backward-schedule/SKILL.md) | common | Evaluate memory-capped split-backward and V-shaped pipeline scheduling proposals, preserving synchronous update ... | **unknown** | — | 1 Accelerator · Diagnose / Configure / Optimize / Verify | 1 · Systems sources (Zero Bubble — [2401.10241v1](https://arxiv.org/abs/2401.10241v1)) | candidate · not_run · 1 readset(s) | no-applicability, no-eval |
-| [audit-fixed-workload-claim](../src/inferrsi/data/skill_library/methods/audit-fixed-workload-claim/SKILL.md) | common | Audit a performance claim from supplied fixed-workload reports and an explicit measurement contract. Check aggregation ... | **unknown** | — | — | — | candidate · not_run · 0 readset(s) | no-SearchSpace-layer, no-Reference-Matrix-row, no-applicability, no-eval |
-| [calibrate-serving-simulation](../src/inferrsi/data/skill_library/methods/calibrate-serving-simulation/SKILL.md) | common | Build an evidence-gated serving-configuration shortlist from a profiled simulator while separating empirical ... | **unknown** | — | 1 Accelerator · Diagnose / Verify | 1 · Systems sources (Vidur — [2405.05465v2](https://arxiv.org/abs/2405.05465v2)) | candidate · not_run · 1 readset(s) | no-applicability, no-eval |
-| [checkpoint-recovery-consistency](../src/inferrsi/data/skill_library/methods/checkpoint-recovery-consistency/SKILL.md) | common | Assess distributed checkpoint recovery using DLRover memory agreement, persistence completion and replica transport ... | **unknown** | — | 2 RAM/Host · Configure / Verify
3 SSD · Configure / Verify
4 Network · Diagnose / Configure / Verify
5 OS · Diagnose / Configure | 1 · Systems sources (C3 / DLRover FlashCheckpoint @ `e35b1ad90e2928ca6bacf5d22117f71356 ...)
1 · Systems sources (C4 / NVIDIA Resiliency Extension @ ...) | candidate · not_run · 1 readset(s) | no-applicability, no-eval |
-| [choose-parallel-training-plan](../src/inferrsi/data/skill_library/methods/choose-parallel-training-plan/SKILL.md) | common | Compare stage, device-mesh and operator-sharding candidates for static distributed training while accounting for ... | **unknown** | — | 1 Accelerator · Diagnose / Configure / Optimize / Verify
4 Network · Diagnose / Configure / Optimize / Verify | 1 · Systems sources (Alpa — [2201.12023v3](https://arxiv.org/abs/2201.12023v3)) | candidate · not_run · 1 readset(s) | no-applicability, no-eval |
-| [dense-performance-evidence](../src/inferrsi/data/skill_library/methods/dense-performance-evidence/SKILL.md) | common | Taxonomy of performance evidence an optimization agent needs, ordered by collection cost and diagnostic power. Maps ... | **unknown** | — | 1 Accelerator · Observe / Profile / Verify
2 RAM/Host · Verify
4 Network · Verify | 1 · Systems sources (NVIDIA/TensorRT-LLM @ f33e0e5fd)
2 · Methodology and corpus references (JiusiServe/vllm-evolve @ 43cc82c)
2 · Methodology and corpus references (microsoft/SkillOpt @ 343db22) | candidate · not_run · 0 readset(s) | no-applicability |
-| [evaluate-elastic-serving-placement](../src/inferrsi/data/skill_library/methods/evaluate-elastic-serving-placement/SKILL.md) | common | Compare migrating an LLM request between model replicas with changing its sequence-parallel execution group, including ... | **unknown** | — | 1 Accelerator · Diagnose / Configure / Optimize / Verify
2 RAM/Host · Diagnose / Configure / Optimize / Verify
4 Network · Diagnose / Configure / Optimize / Verify | 1 · Systems sources (Llumnix — [2406.03243v1](https://arxiv.org/abs/2406.03243v1)) | candidate · not_run · 1 readset(s) | no-applicability, no-eval |
-| [inference-benchmark-design](../src/inferrsi/data/skill_library/methods/inference-benchmark-design/SKILL.md) | common | Select task quality gates, reference parity checks and performance workloads for text, agentic coding, work agents ... | **unknown** | — | — | 2 · Methodology and corpus references (JiusiServe/vllm-evolve @ 43cc82c)
Task-specific benchmark selection (issue #60) ([Issue #60 report](https://github.com/InferMatrixLab/InferRSI/issues/6 ...) | candidate · not_run · 2 readset(s) | no-SearchSpace-layer, no-applicability |
-| [inference-optimization-casebook](../src/inferrsi/data/skill_library/methods/inference-optimization-casebook/SKILL.md) | common | How to structure and use a casebook of past optimization decisions as a decision-support tool. Extracts the casebook ... | **unknown** | — | 1 Accelerator · Optimize | 1 · Systems sources (NVIDIA/TensorRT-LLM @ f33e0e5fd) | candidate · not_run · 0 readset(s) | no-applicability, no-eval |
-| [kernel-writing-guide](../src/inferrsi/data/skill_library/methods/kernel-writing-guide/SKILL.md) | common | Decision guide for writing custom inference kernels: when to write one, which framework to choose (Triton, CuTe DSL ... | **unknown** | — | 1 Accelerator · Optimize / Diagnose / Configure / Verify | 1 · Systems sources (Dao-AILab/causal-conv1d @ `cd81f0413cad2fc1e6f17e785ac39f59aae690cd` ...)
1 · Systems sources (NVIDIA/TensorRT-LLM @ f33e0e5fd)
1 · Systems sources (huggingface/transformers @ `0dad7b822255a0ae261ec45ae937371e859ffd1a` ...)
1 · Systems sources (mindspore-ai/akg @ `5aa15f3fe4d856eb60cd3b27447e04f5f1fb4dff` ...)
1 · Systems sources (sgl-project/sglang @ `16176278d0180f1a1a578e53e6eef432577876c2` ...)
1 · Systems sources (vllm-project/vllm @ `537a79f0fc5ba564d2bef43d54915b2f80f9f77d` (root ...) | candidate · not_run · 7 readset(s) | no-applicability, no-eval |
-| [optimization-search-policy](../src/inferrsi/data/skill_library/methods/optimization-search-policy/SKILL.md) | common | Policy for how an optimization agent should select, order, and stop optimization experiments. Covers candidate ... | **unknown** | — | 1 Accelerator · Optimize | 2 · Methodology and corpus references (JiusiServe/vllm-evolve @ 43cc82c)
2 · Methodology and corpus references (RRSI — arXiv 2609.24972v3)
2 · Methodology and corpus references (arXiv 2608.01867 (CRISP))
2 · Methodology and corpus references (microsoft/SkillOpt @ 343db22) | candidate · not_run · 9 readset(s) | no-applicability, no-eval |
+| [assess-serving-preemption](../src/inferrsi/data/skill_library/methods/assess-serving-preemption/SKILL.md) | common | Assess whether LLM serving queue delays justify token-boundary priority preemption, accounting for suspended KV, host ... | Diagnosing whether LLM serving queue delay justifies token-boundary priority preemption ... | failure_modes, input_contract, produced_artifact, refusal_point
4 advice | 1 Accelerator · Diagnose / Configure / Optimize / Verify
2 RAM/Host · Diagnose / Configure / Optimize / Verify | 1 · Systems sources (Orca — OSDI 2022 proceedings) | candidate · not_run · 1 readset(s) | no-eval |
+| [assess-split-backward-schedule](../src/inferrsi/data/skill_library/methods/assess-split-backward-schedule/SKILL.md) | common | Evaluate memory-capped split-backward and V-shaped pipeline scheduling proposals, preserving synchronous update ... | Assessing memory-capped split-backward and V-shaped pipeline schedule proposals when a ... | failure_modes, input_contract, produced_artifact, refusal_point
3 advice | 1 Accelerator · Diagnose / Configure / Optimize / Verify | 1 · Systems sources (Zero Bubble — [2401.10241v1](https://arxiv.org/abs/2401.10241v1)) | candidate · not_run · 1 readset(s) | no-eval |
+| [audit-fixed-workload-claim](../src/inferrsi/data/skill_library/methods/audit-fixed-workload-claim/SKILL.md) | common | Audit a performance claim from supplied fixed-workload reports and an explicit measurement contract. Check aggregation ... | Auditing a performance claim against supplied fixed-workload reports and an explicit ... | failure_modes, input_contract, produced_artifact, refusal_point
3 advice | — | — | candidate · not_run · 0 readset(s) | no-SearchSpace-layer, no-Reference-Matrix-row, no-eval |
+| [calibrate-serving-simulation](../src/inferrsi/data/skill_library/methods/calibrate-serving-simulation/SKILL.md) | common | Build an evidence-gated serving-configuration shortlist from a profiled simulator while separating empirical ... | Building an evidence-gated serving-configuration shortlist when a profiled simulator or ... | failure_modes, input_contract, produced_artifact, refusal_point
3 advice | 1 Accelerator · Diagnose / Verify | 1 · Systems sources (Vidur — [2405.05465v2](https://arxiv.org/abs/2405.05465v2)) | candidate · not_run · 1 readset(s) | no-eval |
+| [checkpoint-recovery-consistency](../src/inferrsi/data/skill_library/methods/checkpoint-recovery-consistency/SKILL.md) | common | Assess distributed checkpoint recovery using DLRover memory agreement, persistence completion and replica transport ... | Assessing distributed checkpoint recovery through DLRover memory agreement, persistence ... | failure_modes, input_contract, produced_artifact, refusal_point
4 advice | 2 RAM/Host · Configure / Verify
3 SSD · Configure / Verify
4 Network · Diagnose / Configure / Verify
5 OS · Diagnose / Configure | 1 · Systems sources (C3 / DLRover FlashCheckpoint @ `e35b1ad90e2928ca6bacf5d22117f71356 ...)
1 · Systems sources (C4 / NVIDIA Resiliency Extension @ ...) | candidate · not_run · 1 readset(s) | no-eval |
+| [choose-parallel-training-plan](../src/inferrsi/data/skill_library/methods/choose-parallel-training-plan/SKILL.md) | common | Compare stage, device-mesh and operator-sharding candidates for static distributed training while accounting for ... | Comparing stage, device-mesh and operator-sharding candidates for static distributed ... | failure_modes, input_contract, produced_artifact, refusal_point
3 advice | 1 Accelerator · Diagnose / Configure / Optimize / Verify
4 Network · Diagnose / Configure / Optimize / Verify | 1 · Systems sources (Alpa — [2201.12023v3](https://arxiv.org/abs/2201.12023v3)) | candidate · not_run · 1 readset(s) | no-eval |
+| [dense-performance-evidence](../src/inferrsi/data/skill_library/methods/dense-performance-evidence/SKILL.md) | common | Taxonomy of performance evidence an optimization agent needs, ordered by collection cost and diagnostic power. Maps ... | Deciding which performance evidence to collect next and which claim each evidence level ... | failure_modes, input_contract, produced_artifact, refusal_point
4 advice | 1 Accelerator · Observe / Profile / Verify
2 RAM/Host · Verify
4 Network · Verify | 1 · Systems sources (NVIDIA/TensorRT-LLM @ f33e0e5fd)
2 · Methodology and corpus references (JiusiServe/vllm-evolve @ 43cc82c)
2 · Methodology and corpus references (microsoft/SkillOpt @ 343db22) | candidate · not_run · 0 readset(s) | — |
+| [evaluate-elastic-serving-placement](../src/inferrsi/data/skill_library/methods/evaluate-elastic-serving-placement/SKILL.md) | common | Compare migrating an LLM request between model replicas with changing its sequence-parallel execution group, including ... | Comparing migration of a running LLM request between compatible model replicas with ... | failure_modes, input_contract, produced_artifact, refusal_point
3 advice | 1 Accelerator · Diagnose / Configure / Optimize / Verify
2 RAM/Host · Diagnose / Configure / Optimize / Verify
4 Network · Diagnose / Configure / Optimize / Verify | 1 · Systems sources (Llumnix — [2406.03243v1](https://arxiv.org/abs/2406.03243v1)) | candidate · not_run · 1 readset(s) | no-eval |
+| [inference-benchmark-design](../src/inferrsi/data/skill_library/methods/inference-benchmark-design/SKILL.md) | common | Select task quality gates, reference parity checks and performance workloads for text, agentic coding, work agents ... | Selecting task quality gates, reference parity checks and performance workloads for text ... | failure_modes, input_contract, produced_artifact, refusal_point
4 advice | — | 2 · Methodology and corpus references (JiusiServe/vllm-evolve @ 43cc82c)
Task-specific benchmark selection (issue #60) ([Issue #60 report](https://github.com/InferMatrixLab/InferRSI/issues/6 ...) | candidate · not_run · 2 readset(s) | no-SearchSpace-layer |
+| [inference-optimization-casebook](../src/inferrsi/data/skill_library/methods/inference-optimization-casebook/SKILL.md) | common | How to structure and use a casebook of past optimization decisions as a decision-support tool. Extracts the casebook ... | Structuring and querying a casebook of past optimization decisions as decision support ... | failure_modes, input_contract, produced_artifact, refusal_point
4 advice | 1 Accelerator · Optimize | 1 · Systems sources (NVIDIA/TensorRT-LLM @ f33e0e5fd) | candidate · not_run · 0 readset(s) | no-eval |
+| [kernel-writing-guide](../src/inferrsi/data/skill_library/methods/kernel-writing-guide/SKILL.md) | common | Decision guide for writing custom inference kernels: when to write one, which framework to choose (Triton, CuTe DSL ... | Deciding whether to write a custom inference kernel, routing a kernel pattern to Triton ... | failure_modes, input_contract, produced_artifact, refusal_point
6 advice | 1 Accelerator · Optimize / Diagnose / Configure / Verify | 1 · Systems sources (Dao-AILab/causal-conv1d @ `cd81f0413cad2fc1e6f17e785ac39f59aae690cd` ...)
1 · Systems sources (NVIDIA/TensorRT-LLM @ f33e0e5fd)
1 · Systems sources (huggingface/transformers @ `0dad7b822255a0ae261ec45ae937371e859ffd1a` ...)
1 · Systems sources (mindspore-ai/akg @ `5aa15f3fe4d856eb60cd3b27447e04f5f1fb4dff` ...)
1 · Systems sources (sgl-project/sglang @ `16176278d0180f1a1a578e53e6eef432577876c2` ...)
1 · Systems sources (vllm-project/vllm @ `537a79f0fc5ba564d2bef43d54915b2f80f9f77d` (root ...) | candidate · not_run · 7 readset(s) | no-eval |
+| [optimization-search-policy](../src/inferrsi/data/skill_library/methods/optimization-search-policy/SKILL.md) | common | Policy for how an optimization agent should select, order, and stop optimization experiments. Covers candidate ... | Selecting, ordering and stopping optimization experiments: candidate generation ... | failure_modes, input_contract, produced_artifact, refusal_point
4 advice | 1 Accelerator · Optimize | 2 · Methodology and corpus references (JiusiServe/vllm-evolve @ 43cc82c)
2 · Methodology and corpus references (RRSI — arXiv 2609.24972v3)
2 · Methodology and corpus references (arXiv 2608.01867 (CRISP))
2 · Methodology and corpus references (microsoft/SkillOpt @ 343db22) | candidate · not_run · 9 readset(s) | no-eval |
| [performance-ceiling-analysis](../src/inferrsi/data/skill_library/methods/performance-ceiling-analysis/SKILL.md) | common | Estimate optimization headroom before modifying an inference system. Produces a three-level ceiling model (analytical ... | inference serving systems (LLM, diffusion, audio) | — | 1 Accelerator · Diagnose / Verify
5 OS · Configure | 1 · Systems sources (Mooncake — FAST '25 proceedings, pp. 155-170)
1 · Systems sources (NVIDIA/TensorRT-LLM @ f33e0e5fd)
1 · Systems sources (ebugger/Empirical-Roofline-Toolkit @ 0771fa9)
1 · Systems sources (hahnyuan/LLM-Viewer @ 1893e4b) | candidate · not_run · 17 readset(s) | no-eval |
-| [plan-activation-residency](../src/inferrsi/data/skill_library/methods/plan-activation-residency/SKILL.md) | common | Propose training activation and gradient memory reductions by distinguishing lifetime reordering, address packing and ... | **unknown** | — | 1 Accelerator · Diagnose / Optimize / Verify | 1 · Systems sources (Checkmate — [1910.02653v3](https://arxiv.org/abs/1910.02653v3)) | candidate · not_run · 1 readset(s) | no-applicability, no-eval |
+| [plan-activation-residency](../src/inferrsi/data/skill_library/methods/plan-activation-residency/SKILL.md) | common | Propose training activation and gradient memory reductions by distinguishing lifetime reordering, address packing and ... | Proposing training activation and gradient memory reductions when pressure appears to ... | failure_modes, input_contract, produced_artifact, refusal_point
3 advice | 1 Accelerator · Diagnose / Optimize / Verify | 1 · Systems sources (Checkmate — [1910.02653v3](https://arxiv.org/abs/1910.02653v3)) | candidate · not_run · 1 readset(s) | no-eval |
| [repository-evolution-analysis](../src/inferrsi/data/skill_library/methods/repository-evolution-analysis/SKILL.md) | common | Analyze a repository's commit history to build a capability-evolution model: what the system can do, how it got there ... | any software repository with a recoverable git history | — | — | 2 · Methodology and corpus references (LearningToCommit/LearningToCommit @ 7e5a1d7)
2 · Methodology and corpus references (yetone/magpie @ 77800032) | candidate · not_run · 0 readset(s) | no-SearchSpace-layer, no-eval |
-| [skill-evolution-playbook](../src/inferrsi/data/skill_library/methods/skill-evolution-playbook/SKILL.md) | common | Practical design choices for evolving agent skills from execution traces: when one trajectory should update a skill ... | **unknown** | — | — | 2 · Methodology and corpus references (Qwen-Applications/Trace2Skill @ 3d0b52a)
2 · Methodology and corpus references (microsoft/SkillOpt @ 343db22) | candidate · not_run · 0 readset(s) | no-SearchSpace-layer, no-applicability, no-eval |
-| [training-input-pipeline](../src/inferrsi/data/skill_library/methods/training-input-pipeline/SKILL.md) | common | Diagnose training input stalls and choose bounded queue, shard-cache and cache-ownership changes from DALI, Mosaic ... | **unknown** | — | 1 Accelerator · Observe / Configure / Optimize
2 RAM/Host · Observe / Diagnose / Configure / Optimize
3 SSD · Diagnose / Configure / Optimize
4 Network · Configure / Optimize
5 OS · Configure / Optimize | 1 · Systems sources (B1 / NVIDIA DALI @ `6fe47c48d3b0ca335f5e1afd3c77c750b2c7e041` ...)
1 · Systems sources (B2 / Mosaic Streaming @ `d99bf9c7cdf2dd4ed62a4960ed35270b93337a5f` ...)
1 · Systems sources (B3 / FFCV @ `3a12966b3afe3a81733a732e633317d747bfaac7` ...) | candidate · not_run · 1 readset(s) | no-applicability, no-eval |
-| [training-resource-budget](../src/inferrsi/data/skill_library/methods/training-resource-budget/SKILL.md) | common | Choose training offload and resource-budget changes using pinned DeepSpeed configuration, attributed expert diagnosis ... | **unknown** | — | 1 Accelerator · Diagnose / Configure / Optimize
2 RAM/Host · Observe / Diagnose / Configure / Optimize / Verify
3 SSD · Configure / Optimize
4 Network · Observe / Diagnose / Verify
5 OS · Diagnose / Configure | 1 · Systems sources (CR1 / DeepSpeed ZeRO-Infinity @ `5f266e26d216a705d9c64cac88fb813ce ...)
1 · Systems sources (CR5 / Ascend MindSpeed adaptive memory — historical `core_r0.8.0` ...)
1 · Systems sources (ZeRO-Offload — [2101.06840v1](https://arxiv.org/abs/2101.06840v1))
2 · Methodology and corpus references (CR6 / Stas Bekman ML Engineering Open Book @ ...) | candidate · not_run · 9 readset(s) | no-applicability, no-eval |
+| [skill-evolution-playbook](../src/inferrsi/data/skill_library/methods/skill-evolution-playbook/SKILL.md) | common | Practical design choices for evolving agent skills from execution traces: when one trajectory should update a skill ... | Design choices for evolving agent skills from execution traces, drawn from Trace2Skill ... | failure_modes, input_contract, produced_artifact, refusal_point
4 advice | — | 2 · Methodology and corpus references (Qwen-Applications/Trace2Skill @ 3d0b52a)
2 · Methodology and corpus references (microsoft/SkillOpt @ 343db22) | candidate · not_run · 0 readset(s) | no-SearchSpace-layer, no-eval |
+| [training-input-pipeline](../src/inferrsi/data/skill_library/methods/training-input-pipeline/SKILL.md) | common | Diagnose training input stalls and choose bounded queue, shard-cache and cache-ownership changes from DALI, Mosaic ... | Diagnosing training input stalls and choosing bounded queue, shard-cache and ... | failure_modes, input_contract, produced_artifact, refusal_point
4 advice | 1 Accelerator · Observe / Configure / Optimize
2 RAM/Host · Observe / Diagnose / Configure / Optimize
3 SSD · Diagnose / Configure / Optimize
4 Network · Configure / Optimize
5 OS · Configure / Optimize | 1 · Systems sources (B1 / NVIDIA DALI @ `6fe47c48d3b0ca335f5e1afd3c77c750b2c7e041` ...)
1 · Systems sources (B2 / Mosaic Streaming @ `d99bf9c7cdf2dd4ed62a4960ed35270b93337a5f` ...)
1 · Systems sources (B3 / FFCV @ `3a12966b3afe3a81733a732e633317d747bfaac7` ...) | candidate · not_run · 1 readset(s) | no-eval |
+| [training-resource-budget](../src/inferrsi/data/skill_library/methods/training-resource-budget/SKILL.md) | common | Choose training offload and resource-budget changes using pinned DeepSpeed configuration, attributed expert diagnosis ... | Choosing training offload and resource-budget changes from pinned DeepSpeed ... | failure_modes, input_contract, produced_artifact, refusal_point
4 advice | 1 Accelerator · Diagnose / Configure / Optimize
2 RAM/Host · Observe / Diagnose / Configure / Optimize / Verify
3 SSD · Configure / Optimize
4 Network · Observe / Diagnose / Verify
5 OS · Diagnose / Configure | 1 · Systems sources (CR1 / DeepSpeed ZeRO-Infinity @ `5f266e26d216a705d9c64cac88fb813ce ...)
1 · Systems sources (CR5 / Ascend MindSpeed adaptive memory — historical `core_r0.8.0` ...)
1 · Systems sources (ZeRO-Offload — [2101.06840v1](https://arxiv.org/abs/2101.06840v1))
2 · Methodology and corpus references (CR6 / Stas Bekman ML Engineering Open Book @ ...) | candidate · not_run · 9 readset(s) | no-eval |
| [video-diffusion-optimization](../src/inferrsi/data/skill_library/methods/video-diffusion-optimization/SKILL.md) | common | Run and verify goal-driven video diffusion optimization, backend-agnostic. Use when the user asks for latency, RTF ... | Backend-agnostic video diffusion inference and serving optimization | — | — | — | candidate · not_run · 0 readset(s) | no-SearchSpace-layer, no-Reference-Matrix-row, no-eval |
-| [weight-loading-optimization](../src/inferrsi/data/skill_library/methods/weight-loading-optimization/SKILL.md) | common | Optimize the serving-level weight load path from storage to device: judge the load path on four dimensions (source ... | **unknown** | — | 1 Accelerator † guidance: Flare preparation and complete-session source contracts
2 RAM/Host † guidance: Flare preparation and complete-session source contracts
3 SSD · Observe / Profile / Diagnose / Optimize / Verify † guidance: Flare preparation and complete-session source contracts
4 Network · Configure / Optimize / Verify | 1 · Systems sources (Ascend/MindIE-SD @ 8481d39)
1 · Systems sources (NVIDIA/TensorRT-LLM @ f33e0e5fd)
1 · Systems sources (ascend-ai-coding/awesome-ascend-skills @ 75ec89c)
1 · Systems sources (foundation-model-stack/fastsafetensors @ 219e256)
1 · Systems sources (ggml-org/llama.cpp @ `3e390af18de1131dda27913c00511725b4274c8a` (MIT ...)
1 · Systems sources (mindie-agent/mindie-agent @ 6a81b41) | candidate · not_run · 3 readset(s) | no-applicability, no-eval |
+| [weight-loading-optimization](../src/inferrsi/data/skill_library/methods/weight-loading-optimization/SKILL.md) | common | Optimize the serving-level weight load path from storage to device: judge the load path on four dimensions (source ... | Diagnosing the serving-level weight load path from storage to device and choosing changes ... | failure_modes, input_contract, produced_artifact, refusal_point
5 advice | 1 Accelerator † guidance: Flare preparation and complete-session source contracts
2 RAM/Host † guidance: Flare preparation and complete-session source contracts
3 SSD · Observe / Profile / Diagnose / Optimize / Verify † guidance: Flare preparation and complete-session source contracts
4 Network · Configure / Optimize / Verify | 1 · Systems sources (Ascend/MindIE-SD @ 8481d39)
1 · Systems sources (NVIDIA/TensorRT-LLM @ f33e0e5fd)
1 · Systems sources (ascend-ai-coding/awesome-ascend-skills @ 75ec89c)
1 · Systems sources (foundation-model-stack/fastsafetensors @ 219e256)
1 · Systems sources (ggml-org/llama.cpp @ `3e390af18de1131dda27913c00511725b4274c8a` (MIT ...)
1 · Systems sources (mindie-agent/mindie-agent @ 6a81b41) | candidate · not_run · 3 readset(s) | no-eval |
## Registered: Workflows (6)
@@ -221,7 +221,7 @@ an entrypoint routes to procedures below it rather than adding one.
| `no-SearchSpace-layer` | 15: ascend-video-diffusion-optimization, audit-fixed-workload-claim, audit-recurrent-kv-finalization, cite-check-downstream-pr, competitor-scan, draft-acceptance-issue, dwarfstar, husky, inference-benchmark-design, lily, npu-skill-factory, repository-evolution-analysis, skill-evolution-playbook, splash, video-diffusion-optimization |
| `no-Reference-Matrix-row` | 8: ascend-video-diffusion-optimization, audit-fixed-workload-claim, cite-check-downstream-pr, claim-contract, competitor-scan, draft-acceptance-issue, rsi-evidence, video-diffusion-optimization |
| `no-sidecar` | 0: — |
-| `no-applicability` | 41: as listed in the Gaps column above |
+| `no-applicability` | 24: as listed in the Gaps column above |
| `no-eval` | 75: as listed in the Gaps column above |
These are inventory gaps, not defects. A workflow skill may legitimately
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/README.md b/src/inferrsi/data/repo_packs/inferrsi/knowledge/README.md
index ae13524c5..55f201a2b 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/README.md
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/README.md
@@ -15,7 +15,7 @@ These are knowledge abstracts (`--kind knowledge`), with namespaced package
IDs distinct from explicitly selected `checkout_ref` skills. Ascend/CUDA
procedures use their corresponding backend; shared methods use `common`.
-Source snapshot: [ad322aab86d9a148bc4b378a15b2762e41b86908](https://github.com/InferMatrixLab/InferRSI/tree/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library).
+Source snapshot: [54d4c89d4db82cbfc10d6096fe49bb077136e8b1](https://github.com/InferMatrixLab/InferRSI/tree/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library).
```bash
inferrsi knowledge search 'profiling' --repo-type inferrsi --limit 50
@@ -29,7 +29,7 @@ contain the current input paths and bytes (Git checkout filters are respected).
```bash
python tools/build_skill_index.py --source-revision
-python tools/build_skill_index.py --source-revision ad322aab86d9a148bc4b378a15b2762e41b86908 --check
+python tools/build_skill_index.py --source-revision 54d4c89d4db82cbfc10d6096fe49bb077136e8b1 --check
```
`--check` detects missing, stale and unexpected generated records without writing.
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-aclgraph-capture-replay.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-aclgraph-capture-replay.json
index d8066efbf..d83a6f417 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-aclgraph-capture-replay.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-aclgraph-capture-replay.json
@@ -21,8 +21,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/hardware/ascend/ascend-aclgraph-capture-replay/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/hardware/ascend/ascend-aclgraph-capture-replay/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/hardware/ascend/ascend-aclgraph-capture-replay/SKILL.md",
"src/inferrsi/data/skill_library/hardware/ascend/ascend-aclgraph-capture-replay/metadata.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-device-health.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-device-health.json
index 96894d30e..cee368c5d 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-device-health.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-device-health.json
@@ -23,8 +23,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/hardware/ascend/ascend-device-health/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/hardware/ascend/ascend-device-health/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/hardware/ascend/ascend-device-health/SKILL.md",
"src/inferrsi/data/skill_library/hardware/ascend/ascend-device-health/metadata.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-distributed-parallel.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-distributed-parallel.json
index f617341b2..2ae1692ca 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-distributed-parallel.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-distributed-parallel.json
@@ -27,8 +27,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/hardware/ascend/ascend-distributed-parallel/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/hardware/ascend/ascend-distributed-parallel/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/hardware/ascend/ascend-distributed-parallel/SKILL.md",
"src/inferrsi/data/skill_library/hardware/ascend/ascend-distributed-parallel/metadata.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-graph-mode.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-graph-mode.json
index 14202b727..0681960df 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-graph-mode.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-graph-mode.json
@@ -26,8 +26,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/hardware/ascend/ascend-graph-mode/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/hardware/ascend/ascend-graph-mode/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/hardware/ascend/ascend-graph-mode/SKILL.md",
"src/inferrsi/data/skill_library/hardware/ascend/ascend-graph-mode/metadata.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-model-adaptation.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-model-adaptation.json
index 1b2c84c1b..1beec5a9b 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-model-adaptation.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-model-adaptation.json
@@ -22,8 +22,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/hardware/ascend/ascend-model-adaptation/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/hardware/ascend/ascend-model-adaptation/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/hardware/ascend/ascend-model-adaptation/SKILL.md",
"src/inferrsi/data/skill_library/hardware/ascend/ascend-model-adaptation/metadata.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-operator-profiling.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-operator-profiling.json
index ecfae544f..7f18eacad 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-operator-profiling.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-operator-profiling.json
@@ -23,8 +23,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/hardware/ascend/ascend-operator-profiling/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/hardware/ascend/ascend-operator-profiling/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/hardware/ascend/ascend-operator-profiling/SKILL.md",
"src/inferrsi/data/skill_library/hardware/ascend/ascend-operator-profiling/metadata.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-operator-tuning.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-operator-tuning.json
index 1684c6f25..2485f5d1c 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-operator-tuning.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-operator-tuning.json
@@ -24,8 +24,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/hardware/ascend/ascend-operator-tuning/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/hardware/ascend/ascend-operator-tuning/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/hardware/ascend/ascend-operator-tuning/SKILL.md",
"src/inferrsi/data/skill_library/hardware/ascend/ascend-operator-tuning/metadata.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-optimization-casebook.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-optimization-casebook.json
index 26127cc3f..e150ddfbf 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-optimization-casebook.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-optimization-casebook.json
@@ -22,8 +22,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/hardware/ascend/ascend-optimization-casebook/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/hardware/ascend/ascend-optimization-casebook/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/hardware/ascend/ascend-optimization-casebook/SKILL.md",
"src/inferrsi/data/skill_library/hardware/ascend/ascend-optimization-casebook/metadata.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-performance-analysis.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-performance-analysis.json
index 139174fa6..f98d36f12 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-performance-analysis.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-performance-analysis.json
@@ -22,8 +22,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/hardware/ascend/ascend-performance-analysis/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/hardware/ascend/ascend-performance-analysis/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/hardware/ascend/ascend-performance-analysis/SKILL.md",
"src/inferrsi/data/skill_library/hardware/ascend/ascend-performance-analysis/metadata.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-video-diffusion-optimization.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-video-diffusion-optimization.json
index b499a1e27..892275b5c 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-video-diffusion-optimization.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascend-video-diffusion-optimization.json
@@ -26,8 +26,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/hardware/ascend/ascend-video-diffusion-optimization/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/hardware/ascend/ascend-video-diffusion-optimization/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/hardware/ascend/ascend-video-diffusion-optimization/SKILL.md",
"src/inferrsi/data/skill_library/hardware/ascend/ascend-video-diffusion-optimization/metadata.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascendc-tiling-design.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascendc-tiling-design.json
index a1a7a1ce9..4088bb290 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascendc-tiling-design.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/ascendc-tiling-design.json
@@ -24,8 +24,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/hardware/ascend/ascendc-tiling-design/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/hardware/ascend/ascendc-tiling-design/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/hardware/ascend/ascendc-tiling-design/SKILL.md",
"src/inferrsi/data/skill_library/hardware/ascend/ascendc-tiling-design/metadata.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/cann-attribution-overlap.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/cann-attribution-overlap.json
index 1c57df82e..3f1af08e1 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/cann-attribution-overlap.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/cann-attribution-overlap.json
@@ -24,8 +24,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/hardware/ascend/cann-attribution-overlap/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/hardware/ascend/cann-attribution-overlap/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/hardware/ascend/cann-attribution-overlap/SKILL.md",
"src/inferrsi/data/skill_library/hardware/ascend/cann-attribution-overlap/metadata.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/torchnpu-device-memory-lifecycle.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/torchnpu-device-memory-lifecycle.json
index b5eec59d3..5a15fa487 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/torchnpu-device-memory-lifecycle.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/torchnpu-device-memory-lifecycle.json
@@ -22,8 +22,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/hardware/ascend/torchnpu-device-memory-lifecycle/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/hardware/ascend/torchnpu-device-memory-lifecycle/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/hardware/ascend/torchnpu-device-memory-lifecycle/SKILL.md",
"src/inferrsi/data/skill_library/hardware/ascend/torchnpu-device-memory-lifecycle/metadata.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/torchnpu-pinned-host-memory-lifecycle.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/torchnpu-pinned-host-memory-lifecycle.json
index b9a213af2..dd95945b8 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/torchnpu-pinned-host-memory-lifecycle.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/ascend/skills/torchnpu-pinned-host-memory-lifecycle.json
@@ -22,8 +22,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/hardware/ascend/torchnpu-pinned-host-memory-lifecycle/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/hardware/ascend/torchnpu-pinned-host-memory-lifecycle/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/hardware/ascend/torchnpu-pinned-host-memory-lifecycle/SKILL.md",
"src/inferrsi/data/skill_library/hardware/ascend/torchnpu-pinned-host-memory-lifecycle/metadata.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/assess-serving-preemption.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/assess-serving-preemption.json
index 0c4d08720..217d654b4 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/assess-serving-preemption.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/assess-serving-preemption.json
@@ -16,8 +16,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/methods/assess-serving-preemption/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/methods/assess-serving-preemption/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/methods/assess-serving-preemption/SKILL.md",
"src/inferrsi/data/skill_library/methods/assess-serving-preemption/metadata.json",
@@ -34,7 +34,7 @@
"checkout_ref": "checkout:assess-serving-preemption",
"source_files": {
"src/inferrsi/data/skill_library/methods/assess-serving-preemption/SKILL.md": "0a5f48d9352c5bf460775ef47f9136829978811f4a16e84983e299782cae305d",
- "src/inferrsi/data/skill_library/methods/assess-serving-preemption/metadata.json": "77ce5ba16245cbd80d46d0abdf80aaf26fc0924fa2d2a5624ff430de68a1daf0",
+ "src/inferrsi/data/skill_library/methods/assess-serving-preemption/metadata.json": "3e2cea37f5669e7c60f7c01c390a48ea960b15cbd6de8f5340cf8fe2afb9e1a1",
"src/inferrsi/data/skill_library/methods/assess-serving-preemption/references/source-notes.md": "f29fcecfba0490008a354c11b42403983fe165e56ff907ad61cc6040d2b6407f",
"src/inferrsi/data/skill_library/skill-index.json": "b5a9f05bd8179644f3fea79d7eb04fe4a8baad9e8d68aa8d0f1db213cf633317"
},
@@ -50,6 +50,27 @@
"origin": [
"external"
],
+ "applicability": {
+ "scope": "Diagnosing whether LLM serving queue delay justifies token-boundary priority preemption, including suspended KV, host swap or recomputation and long-request service guarantees; admission-versus-service-order diagnosis, not current scheduler implementation or tenant-fairness certification",
+ "prerequisites": "The packaged source notes are read before quoting a metric or comparing a baseline; installing, benchmarking, changing scheduling or admission, migrating state, editing a repository or running a reproduction needs a separately authorized task",
+ "input_contract": "The late boundary (first token, token gap, completion or a stated cohort's service) with its exact percentile or attainment target and denominator; model and KV format, engine revision, safe interruption boundary, request batch limit, input/output/context distributions, arrival bursts, device live/reserved/allocated bytes, active and suspended requests, host budget, transfer path, recovery time and queue/phase/service timings; per-tenant or priority-class obligations kept apart from aggregate latency",
+ "produced_artifact": "A read-only decision naming the observed failure boundary, the smallest candidate that addresses it, unchanged assumptions, missing evidence, resource cost, affected cohorts, counter-signal and next authorized gate, plus a proposed verification plan; not a validated configuration",
+ "failure_modes": [
+ "An ordinary iteration boundary is taken to permit interrupting a prefill halfway through",
+ "Unknown output length is treated as known from the prompt",
+ "KV is counted for the running batch only, omitting started requests that are suspended",
+ "Reservations are removed because some reserved capacity is not live",
+ "Recompute-based eviction is proposed without a progress invariant, so two jobs can repeatedly displace each other",
+ "Host-device copies or same-host IPC are charged to Network instead of their local resource"
+ ],
+ "refusal_point": "Withholds a preemption recommendation when the apparent gain can be explained by different kernels, unmatched parallelism or an incomparable latency metric; unknown safe boundaries, unsupported state recovery, uncalibrated timing or unclear service obligations are blockers"
+ },
+ "recommendations": [
+ "Name the late boundary and its exact target and denominator before choosing among batching, admission, preemption and state recovery.",
+ "Send prefill interference to choose-prefill-decode-schedule and physical KV slack to diagnose-kv-block-efficiency when those candidates are available; otherwise report the diagnosed boundary and the missing evidence.",
+ "In a later authorized test, isolate service ordering first and state recovery second, and evaluate their interaction only after both are understood.",
+ "Weigh FastServe's same-engine FCFS comparison above its historical-vLLM headline, and do not read its P95 aggregate gain as a starvation or tenant-fairness guarantee."
+ ],
"limitations": [
"Source-derived candidate; no local efficacy validation.",
"Original typed source inventory is not normalized; retain source identities and limits in the procedure/readsets."
@@ -59,10 +80,26 @@
],
"consumption_mode": "packaged"
},
- "applicability": {},
+ "applicability": {
+ "scope": "Diagnosing whether LLM serving queue delay justifies token-boundary priority preemption, including suspended KV, host swap or recomputation and long-request service guarantees; admission-versus-service-order diagnosis, not current scheduler implementation or tenant-fairness certification",
+ "prerequisites": "The packaged source notes are read before quoting a metric or comparing a baseline; installing, benchmarking, changing scheduling or admission, migrating state, editing a repository or running a reproduction needs a separately authorized task",
+ "input_contract": "The late boundary (first token, token gap, completion or a stated cohort's service) with its exact percentile or attainment target and denominator; model and KV format, engine revision, safe interruption boundary, request batch limit, input/output/context distributions, arrival bursts, device live/reserved/allocated bytes, active and suspended requests, host budget, transfer path, recovery time and queue/phase/service timings; per-tenant or priority-class obligations kept apart from aggregate latency",
+ "produced_artifact": "A read-only decision naming the observed failure boundary, the smallest candidate that addresses it, unchanged assumptions, missing evidence, resource cost, affected cohorts, counter-signal and next authorized gate, plus a proposed verification plan; not a validated configuration",
+ "failure_modes": [
+ "An ordinary iteration boundary is taken to permit interrupting a prefill halfway through",
+ "Unknown output length is treated as known from the prompt",
+ "KV is counted for the running batch only, omitting started requests that are suspended",
+ "Reservations are removed because some reserved capacity is not live",
+ "Recompute-based eviction is proposed without a progress invariant, so two jobs can repeatedly displace each other",
+ "Host-device copies or same-host IPC are charged to Network instead of their local resource"
+ ],
+ "refusal_point": "Withholds a preemption recommendation when the apparent gain can be explained by different kernels, unmatched parallelism or an incomparable latency metric; unknown safe boundaries, unsupported state recovery, uncalibrated timing or unclear service obligations are blockers"
+ },
"recommendations": [
- "Read the pinned SKILL.md and its prerequisites before choosing this procedure.",
- "Read the packaged references and owner-local cases alongside the procedure."
+ "Name the late boundary and its exact target and denominator before choosing among batching, admission, preemption and state recovery.",
+ "Send prefill interference to choose-prefill-decode-schedule and physical KV slack to diagnose-kv-block-efficiency when those candidates are available; otherwise report the diagnosed boundary and the missing evidence.",
+ "In a later authorized test, isolate service ordering first and state recovery second, and evaluate their interaction only after both are understood.",
+ "Weigh FastServe's same-engine FCFS comparison above its historical-vLLM headline, and do not read its P95 aggregate gain as a starvation or tenant-fairness guarantee."
],
"limitations": [
"This generated abstract points to the complete packaged skill; external evidence and upstream repositories remain source links.",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/assess-split-backward-schedule.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/assess-split-backward-schedule.json
index f6eb0e8ce..9e58f9d99 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/assess-split-backward-schedule.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/assess-split-backward-schedule.json
@@ -16,8 +16,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/methods/assess-split-backward-schedule/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/methods/assess-split-backward-schedule/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/methods/assess-split-backward-schedule/SKILL.md",
"src/inferrsi/data/skill_library/methods/assess-split-backward-schedule/metadata.json",
@@ -34,7 +34,7 @@
"checkout_ref": "checkout:assess-split-backward-schedule",
"source_files": {
"src/inferrsi/data/skill_library/methods/assess-split-backward-schedule/SKILL.md": "1e64f28ea71672fb53c56f6108b87864738c6d03d6dcd32eedaef9e0625e96e3",
- "src/inferrsi/data/skill_library/methods/assess-split-backward-schedule/metadata.json": "c6acae57d70190cea0e18c71c79127962181e58169e011df75d8635b3db4daa7",
+ "src/inferrsi/data/skill_library/methods/assess-split-backward-schedule/metadata.json": "a089dd2c2e5a4b41e970e426c63b5b8f65acc77d2e5a40a44b38543b5d730a00",
"src/inferrsi/data/skill_library/methods/assess-split-backward-schedule/references/source-notes.md": "61b1609f71c37b61d136b8c4d39f793de87cffb1cf31c8ec737a1c3fa6fb1b4e",
"src/inferrsi/data/skill_library/skill-index.json": "b5a9f05bd8179644f3fea79d7eb04fe4a8baad9e8d68aa8d0f1db213cf633317"
},
@@ -49,6 +49,26 @@
"origin": [
"external"
],
+ "applicability": {
+ "scope": "Assessing memory-capped split-backward and V-shaped pipeline schedule proposals when a pipeline trace suggests input-gradient propagation waits behind deferrable weight-gradient work or an optimizer barrier, while preserving synchronous update, global-batch and optimizer-validation semantics",
+ "prerequisites": "The packaged evidence notes are read before citing the paper; installing or running frameworks, benchmarking schedules, changing training semantics, configuring hardware or modifying repositories needs separate authorization",
+ "input_contract": "Stage and microbatch counts, microbatch and global batch, precision, optimizer and clipping/NaN behavior; F, input-gradient B, weight-gradient W and transfer timings; per-pass retained and released bytes; total device peak and margin; collective ordering; kernel-splitting support; calculated bubble rate kept apart from measured idle time and useful samples/s",
+ "produced_artifact": "A candidate schedule and fallback with per-pass order rationale, peak-memory and transfer costs, unchanged semantic requirements, counter-signals and a future verification proposal; a reviewable plan, not an enabled schedule",
+ "failure_modes": [
+ "Completion of B is equated with releasing all activations, so W inputs retained by deferral go uncounted",
+ "The equal-time zero-bubble construction replaces observed F/B/W and transfer times, or kernel-splitting and synchronization overhead is left out",
+ "The schedule is chosen by minimum modeled bubble rate instead of whole-step throughput under the memory and quality constraint",
+ "pMB activation capacity is read as total memory, hiding HBM, workspace, buffers or per-rank imbalance",
+ "Removing the optimizer barrier drops global norm, clipping or NaN decisions, rollback and re-update, or the correctness of work that depends on a tentative update",
+ "Optimizer post-validation is conflated with accepting stale-gradient delayed updates"
+ ],
+ "refusal_point": "Withholds an enablement recommendation under frequent clipping/NaNs, small available memory, strongly unequal passes, unsupported split kernels, hidden collectives or dynamic shapes, and retains the synchronized fallback while optimizer reversibility and rollback coverage are unverified"
+ },
+ "recommendations": [
+ "Compare ordinary 1F1B with a memory-capped split schedule first; add V-shaped two-chunk placement only when memory permits, and at matched global batch also try smaller microbatches with more microbatches.",
+ "Decide optimizer post-validation separately from the schedule change.",
+ "In a later authorized comparison, hold model, global batch and precision fixed and compare real with calculated idle intervals."
+ ],
"limitations": [
"Source-derived candidate; no local efficacy validation.",
"Original typed source inventory is not normalized; retain source identities and limits in the procedure/readsets."
@@ -58,10 +78,25 @@
],
"consumption_mode": "packaged"
},
- "applicability": {},
+ "applicability": {
+ "scope": "Assessing memory-capped split-backward and V-shaped pipeline schedule proposals when a pipeline trace suggests input-gradient propagation waits behind deferrable weight-gradient work or an optimizer barrier, while preserving synchronous update, global-batch and optimizer-validation semantics",
+ "prerequisites": "The packaged evidence notes are read before citing the paper; installing or running frameworks, benchmarking schedules, changing training semantics, configuring hardware or modifying repositories needs separate authorization",
+ "input_contract": "Stage and microbatch counts, microbatch and global batch, precision, optimizer and clipping/NaN behavior; F, input-gradient B, weight-gradient W and transfer timings; per-pass retained and released bytes; total device peak and margin; collective ordering; kernel-splitting support; calculated bubble rate kept apart from measured idle time and useful samples/s",
+ "produced_artifact": "A candidate schedule and fallback with per-pass order rationale, peak-memory and transfer costs, unchanged semantic requirements, counter-signals and a future verification proposal; a reviewable plan, not an enabled schedule",
+ "failure_modes": [
+ "Completion of B is equated with releasing all activations, so W inputs retained by deferral go uncounted",
+ "The equal-time zero-bubble construction replaces observed F/B/W and transfer times, or kernel-splitting and synchronization overhead is left out",
+ "The schedule is chosen by minimum modeled bubble rate instead of whole-step throughput under the memory and quality constraint",
+ "pMB activation capacity is read as total memory, hiding HBM, workspace, buffers or per-rank imbalance",
+ "Removing the optimizer barrier drops global norm, clipping or NaN decisions, rollback and re-update, or the correctness of work that depends on a tentative update",
+ "Optimizer post-validation is conflated with accepting stale-gradient delayed updates"
+ ],
+ "refusal_point": "Withholds an enablement recommendation under frequent clipping/NaNs, small available memory, strongly unequal passes, unsupported split kernels, hidden collectives or dynamic shapes, and retains the synchronized fallback while optimizer reversibility and rollback coverage are unverified"
+ },
"recommendations": [
- "Read the pinned SKILL.md and its prerequisites before choosing this procedure.",
- "Read the packaged references and owner-local cases alongside the procedure."
+ "Compare ordinary 1F1B with a memory-capped split schedule first; add V-shaped two-chunk placement only when memory permits, and at matched global batch also try smaller microbatches with more microbatches.",
+ "Decide optimizer post-validation separately from the schedule change.",
+ "In a later authorized comparison, hold model, global batch and precision fixed and compare real with calculated idle intervals."
],
"limitations": [
"This generated abstract points to the complete packaged skill; external evidence and upstream repositories remain source links.",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-fixed-workload-claim.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-fixed-workload-claim.json
index fac9a5b67..204c4778a 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-fixed-workload-claim.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-fixed-workload-claim.json
@@ -20,8 +20,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/methods/audit-fixed-workload-claim/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/methods/audit-fixed-workload-claim/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/methods/audit-fixed-workload-claim/SKILL.md",
"src/inferrsi/data/skill_library/methods/audit-fixed-workload-claim/metadata.json",
@@ -37,7 +37,7 @@
"checkout_ref": "checkout:audit-fixed-workload-claim",
"source_files": {
"src/inferrsi/data/skill_library/methods/audit-fixed-workload-claim/SKILL.md": "2be89b1340a3deb978c8e1792d18aeecbe8f5de81d4daf542389f53826b8ebee",
- "src/inferrsi/data/skill_library/methods/audit-fixed-workload-claim/metadata.json": "81edd75fa16939807548d0944e9203cfd7c0cb17e63463872d96ca2c1273d1d8",
+ "src/inferrsi/data/skill_library/methods/audit-fixed-workload-claim/metadata.json": "d36e8a6148b7ab830f1ddf6e573a865965502709a835ce460c37e3d4c386de95",
"src/inferrsi/data/skill_library/skill-index.json": "b5a9f05bd8179644f3fea79d7eb04fe4a8baad9e8d68aa8d0f1db213cf633317"
},
"skill_metadata": {
@@ -53,14 +53,47 @@
"measurement"
],
"empirical_validation": "not_run",
+ "applicability": {
+ "scope": "Auditing a performance claim against supplied fixed-workload reports and an explicit measurement contract: aggregation, retained observations, comparability and quality evidence",
+ "input_contract": "Baseline and candidate fixed-workload reports with request observations grouped by trial and workload; the measurement contract stating metric, unit, timing boundary, improvement direction, aggregation rule and order, fixed exclusions, threshold and comparison operator; the declared quality and completion evidence; the requested output schema",
+ "produced_artifact": "An audit conclusion showing baseline and candidate aggregates, the calculation, the threshold comparison and the evidence or missing information that controls the verdict; it interprets supplied evidence and does not itself establish a measured improvement",
+ "failure_modes": [
+ "A pooled request median silently replaces a median of per-trial means, or the reverse, reweighting trials whose request counts differ",
+ "Late requests, failures or outliers are dropped, or requests move between trials, outside the exclusions the contract fixes",
+ "Values are rounded before the threshold comparison",
+ "A missing rule that permits different verdicts is resolved toward the favorable one",
+ "A speed improvement is read as task success, or a quality pass as equivalent measured work",
+ "Different workloads or incomplete observations are repaired by a favorable aggregate"
+ ],
+ "refusal_point": "Does not replace missing evidence with a pass, does not support a percentage claim on an undefined denominator, and does not launch a new benchmark or alter the contract to resolve an unfavorable result; a failed threshold is reported apart from an unverified comparison or missing quality evidence"
+ },
+ "recommendations": [
+ "Recover the contract's metric, direction, aggregation order and comparison operator before reading the report's own verdict.",
+ "Treat a named minimum bound as >= and a maximum bound as <= unless the contract requires a strict comparison.",
+ "Keep separate regimes and metrics separate unless the contract defines their aggregation."
+ ],
"limitations": [
"Guidance for supplied reports; no local efficacy validation or measured performance gain is established."
]
},
- "applicability": {},
+ "applicability": {
+ "scope": "Auditing a performance claim against supplied fixed-workload reports and an explicit measurement contract: aggregation, retained observations, comparability and quality evidence",
+ "input_contract": "Baseline and candidate fixed-workload reports with request observations grouped by trial and workload; the measurement contract stating metric, unit, timing boundary, improvement direction, aggregation rule and order, fixed exclusions, threshold and comparison operator; the declared quality and completion evidence; the requested output schema",
+ "produced_artifact": "An audit conclusion showing baseline and candidate aggregates, the calculation, the threshold comparison and the evidence or missing information that controls the verdict; it interprets supplied evidence and does not itself establish a measured improvement",
+ "failure_modes": [
+ "A pooled request median silently replaces a median of per-trial means, or the reverse, reweighting trials whose request counts differ",
+ "Late requests, failures or outliers are dropped, or requests move between trials, outside the exclusions the contract fixes",
+ "Values are rounded before the threshold comparison",
+ "A missing rule that permits different verdicts is resolved toward the favorable one",
+ "A speed improvement is read as task success, or a quality pass as equivalent measured work",
+ "Different workloads or incomplete observations are repaired by a favorable aggregate"
+ ],
+ "refusal_point": "Does not replace missing evidence with a pass, does not support a percentage claim on an undefined denominator, and does not launch a new benchmark or alter the contract to resolve an unfavorable result; a failed threshold is reported apart from an unverified comparison or missing quality evidence"
+ },
"recommendations": [
- "Read the pinned SKILL.md and its prerequisites before choosing this procedure.",
- "Read the packaged references and owner-local cases alongside the procedure."
+ "Recover the contract's metric, direction, aggregation order and comparison operator before reading the report's own verdict.",
+ "Treat a named minimum bound as >= and a maximum bound as <= unless the contract requires a strict comparison.",
+ "Keep separate regimes and metrics separate unless the contract defines their aggregation."
],
"limitations": [
"This generated abstract points to the complete packaged skill; external evidence and upstream repositories remain source links.",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-h3-metal-fasth3-semantics.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-h3-metal-fasth3-semantics.json
index f0f81d491..54b00d96c 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-h3-metal-fasth3-semantics.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-h3-metal-fasth3-semantics.json
@@ -29,8 +29,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/repositories/h3-metal/audit-h3-metal-fasth3-semantics/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/repositories/h3-metal/audit-h3-metal-fasth3-semantics/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/repositories/h3-metal/audit-h3-metal-fasth3-semantics/LICENSE",
"src/inferrsi/data/skill_library/repositories/h3-metal/audit-h3-metal-fasth3-semantics/SKILL.md",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-h3-runtime-residency.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-h3-runtime-residency.json
index 9def23f7b..d2bd86525 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-h3-runtime-residency.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-h3-runtime-residency.json
@@ -30,8 +30,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/repositories/h3-c/audit-h3-runtime-residency/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/repositories/h3-c/audit-h3-runtime-residency/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/repositories/h3-c/audit-h3-runtime-residency/LICENSE",
"src/inferrsi/data/skill_library/repositories/h3-c/audit-h3-runtime-residency/SKILL.md",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-llama-runtime-state.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-llama-runtime-state.json
index 5a176dc23..911b8e55f 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-llama-runtime-state.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-llama-runtime-state.json
@@ -23,8 +23,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/repositories/llama-cpp/audit-llama-runtime-state/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/repositories/llama-cpp/audit-llama-runtime-state/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/repositories/llama-cpp/audit-llama-runtime-state/SKILL.md",
"src/inferrsi/data/skill_library/repositories/llama-cpp/audit-llama-runtime-state/metadata.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-recurrent-kv-finalization.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-recurrent-kv-finalization.json
index 9f292aefb..a87627aa7 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-recurrent-kv-finalization.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-recurrent-kv-finalization.json
@@ -26,8 +26,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/repositories/vllm-rlt/audit-recurrent-kv-finalization/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/repositories/vllm-rlt/audit-recurrent-kv-finalization/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/repositories/vllm-rlt/audit-recurrent-kv-finalization/LICENSE",
"src/inferrsi/data/skill_library/repositories/vllm-rlt/audit-recurrent-kv-finalization/SKILL.md",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-sd-cpp-residency.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-sd-cpp-residency.json
index 89f11e181..b4e3f7cf1 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-sd-cpp-residency.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-sd-cpp-residency.json
@@ -25,8 +25,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/repositories/stable-diffusion-cpp/audit-sd-cpp-residency/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/repositories/stable-diffusion-cpp/audit-sd-cpp-residency/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/repositories/stable-diffusion-cpp/audit-sd-cpp-residency/LICENSE",
"src/inferrsi/data/skill_library/repositories/stable-diffusion-cpp/audit-sd-cpp-residency/SKILL.md",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-sglang-diffusion-lifetimes.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-sglang-diffusion-lifetimes.json
index d88bd22f4..1e97e502c 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-sglang-diffusion-lifetimes.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-sglang-diffusion-lifetimes.json
@@ -23,8 +23,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/repositories/sglang/audit-sglang-diffusion-lifetimes/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/repositories/sglang/audit-sglang-diffusion-lifetimes/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/repositories/sglang/audit-sglang-diffusion-lifetimes/LICENSE",
"src/inferrsi/data/skill_library/repositories/sglang/audit-sglang-diffusion-lifetimes/SKILL.md",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-vllm-kv-lifetimes.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-vllm-kv-lifetimes.json
index 672929502..d13ea4060 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-vllm-kv-lifetimes.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/audit-vllm-kv-lifetimes.json
@@ -21,8 +21,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/repositories/vllm/audit-vllm-kv-lifetimes/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/repositories/vllm/audit-vllm-kv-lifetimes/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/repositories/vllm/audit-vllm-kv-lifetimes/SKILL.md",
"src/inferrsi/data/skill_library/repositories/vllm/audit-vllm-kv-lifetimes/metadata.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/bound-windowed-attention-replay.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/bound-windowed-attention-replay.json
index 05840745c..d4ebb1f0a 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/bound-windowed-attention-replay.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/bound-windowed-attention-replay.json
@@ -23,8 +23,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/repositories/deepinfer/bound-windowed-attention-replay/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/repositories/deepinfer/bound-windowed-attention-replay/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/repositories/deepinfer/bound-windowed-attention-replay/SKILL.md",
"src/inferrsi/data/skill_library/repositories/deepinfer/bound-windowed-attention-replay/metadata.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/budget-speculative-verification.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/budget-speculative-verification.json
index ceed92610..afa4bc20c 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/budget-speculative-verification.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/budget-speculative-verification.json
@@ -23,8 +23,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/repositories/deepinfer/budget-speculative-verification/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/repositories/deepinfer/budget-speculative-verification/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/repositories/deepinfer/budget-speculative-verification/SKILL.md",
"src/inferrsi/data/skill_library/repositories/deepinfer/budget-speculative-verification/metadata.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/calibrate-serving-simulation.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/calibrate-serving-simulation.json
index d7040f1c6..dfd66207b 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/calibrate-serving-simulation.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/calibrate-serving-simulation.json
@@ -16,8 +16,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/methods/calibrate-serving-simulation/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/methods/calibrate-serving-simulation/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/methods/calibrate-serving-simulation/SKILL.md",
"src/inferrsi/data/skill_library/methods/calibrate-serving-simulation/metadata.json",
@@ -34,7 +34,7 @@
"checkout_ref": "checkout:calibrate-serving-simulation",
"source_files": {
"src/inferrsi/data/skill_library/methods/calibrate-serving-simulation/SKILL.md": "ac773c6f1c110a664dfca28a3ae9129b10c07614e60ca3640dea776220e7b352",
- "src/inferrsi/data/skill_library/methods/calibrate-serving-simulation/metadata.json": "261ce9f644ea6eebb90c830f41fb8b320d2b401456e5bd8d494095a44b8620d0",
+ "src/inferrsi/data/skill_library/methods/calibrate-serving-simulation/metadata.json": "96771fb6fecf5fc7773bb7cf84038d8eaf436add1a0362dc26110c82037f8320",
"src/inferrsi/data/skill_library/methods/calibrate-serving-simulation/references/source-notes.md": "9696d3c0ef96b6611787446bd6f96d4d6255ca8d107105de8c1d9af83edb9fa5",
"src/inferrsi/data/skill_library/skill-index.json": "b5a9f05bd8179644f3fea79d7eb04fe4a8baad9e8d68aa8d0f1db213cf633317"
},
@@ -63,6 +63,26 @@
"readset": "references/source-notes.md"
}
],
+ "applicability": {
+ "scope": "Building an evidence-gated serving-configuration shortlist when a profiled simulator or cost model is proposed to reduce deployment experiments; extends performance-ceiling-analysis, optimization-search-policy and inference-benchmark-design without replacing their ceilings or measurement contract",
+ "prerequisites": "The packaged Vidur evidence and evidence-type notes are read; reading the skill does not authorize execution, and no validation check runs under it without separate execution authorization",
+ "input_contract": "Model, shape and dtype; actual kernel/runtime revision; graph mode; GPU, host and topology; request length and arrival distribution; total resources and cost basis; memory limits; latency definitions and percentile SLOs",
+ "produced_artifact": "A bounded decision memo with input identities, model assumptions, evidence type, exact metrics and denominators, unsupported dimensions, nondominated candidates with predicted ranges, joint constraint checks, the next discriminating trial and rollback/finalization requirements, keeping prediction and measurement columns apart; a shortlist for future measurement, not a measured result",
+ "failure_modes": [
+ "Learned runtime output is labeled a physical bound",
+ "A missing identity is filled with a paper default, or topology is substituted, instead of being marked unknown",
+ "A low-load check stands in for an SLO claim at the intended load regime and capacity boundary",
+ "A predictor that relies on summed KV volume hides batches with equal volume but unequal context distributions",
+ "Capacity binary search is applied without verifying its monotonic/stationary premise or treating overloaded samples explicitly",
+ "One predicted winner is reported when uncertainty exceeds rank separation"
+ ],
+ "refusal_point": "Stops predictive certification when profiles are missing, topology is substituted, error grows near capacity or host overhead is omitted; labels an ordering unresolved when uncertainty exceeds rank separation; asserts neither universal error below 9% nor current vLLM, asynchronous PP or NPU support from the historical simulator paper"
+ },
+ "recommendations": [
+ "Classify every estimate as a physical bound, measured envelope, Amdahl-style conditional impact or empirical prediction before comparing configurations.",
+ "Keep prefill and decode separate in the operator signatures, and record how sharding and memory planning enter the model.",
+ "Recalibrate after workload drift, kernel/graph/runtime changes, new hardware or topology, or mismatched residuals."
+ ],
"limitations": [
"Source-derived candidate; no local efficacy validation."
],
@@ -71,10 +91,25 @@
],
"consumption_mode": "packaged"
},
- "applicability": {},
+ "applicability": {
+ "scope": "Building an evidence-gated serving-configuration shortlist when a profiled simulator or cost model is proposed to reduce deployment experiments; extends performance-ceiling-analysis, optimization-search-policy and inference-benchmark-design without replacing their ceilings or measurement contract",
+ "prerequisites": "The packaged Vidur evidence and evidence-type notes are read; reading the skill does not authorize execution, and no validation check runs under it without separate execution authorization",
+ "input_contract": "Model, shape and dtype; actual kernel/runtime revision; graph mode; GPU, host and topology; request length and arrival distribution; total resources and cost basis; memory limits; latency definitions and percentile SLOs",
+ "produced_artifact": "A bounded decision memo with input identities, model assumptions, evidence type, exact metrics and denominators, unsupported dimensions, nondominated candidates with predicted ranges, joint constraint checks, the next discriminating trial and rollback/finalization requirements, keeping prediction and measurement columns apart; a shortlist for future measurement, not a measured result",
+ "failure_modes": [
+ "Learned runtime output is labeled a physical bound",
+ "A missing identity is filled with a paper default, or topology is substituted, instead of being marked unknown",
+ "A low-load check stands in for an SLO claim at the intended load regime and capacity boundary",
+ "A predictor that relies on summed KV volume hides batches with equal volume but unequal context distributions",
+ "Capacity binary search is applied without verifying its monotonic/stationary premise or treating overloaded samples explicitly",
+ "One predicted winner is reported when uncertainty exceeds rank separation"
+ ],
+ "refusal_point": "Stops predictive certification when profiles are missing, topology is substituted, error grows near capacity or host overhead is omitted; labels an ordering unresolved when uncertainty exceeds rank separation; asserts neither universal error below 9% nor current vLLM, asynchronous PP or NPU support from the historical simulator paper"
+ },
"recommendations": [
- "Read the pinned SKILL.md and its prerequisites before choosing this procedure.",
- "Read the packaged references and owner-local cases alongside the procedure."
+ "Classify every estimate as a physical bound, measured envelope, Amdahl-style conditional impact or empirical prediction before comparing configurations.",
+ "Keep prefill and decode separate in the operator signatures, and record how sharding and memory planning enter the model.",
+ "Recalibrate after workload drift, kernel/graph/runtime changes, new hardware or topology, or mismatched residuals."
],
"limitations": [
"This generated abstract points to the complete packaged skill; external evidence and upstream repositories remain source links.",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/cgroup-resource-controls.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/cgroup-resource-controls.json
index 973710200..eedc29fad 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/cgroup-resource-controls.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/cgroup-resource-controls.json
@@ -25,8 +25,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/hardware/host/cgroup-resource-controls/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/hardware/host/cgroup-resource-controls/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/hardware/host/cgroup-resource-controls/SKILL.md",
"src/inferrsi/data/skill_library/hardware/host/cgroup-resource-controls/metadata.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/checkpoint-recovery-consistency.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/checkpoint-recovery-consistency.json
index b8cb4e5e9..c238e965a 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/checkpoint-recovery-consistency.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/checkpoint-recovery-consistency.json
@@ -22,8 +22,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/methods/checkpoint-recovery-consistency/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/methods/checkpoint-recovery-consistency/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/methods/checkpoint-recovery-consistency/SKILL.md",
"src/inferrsi/data/skill_library/methods/checkpoint-recovery-consistency/metadata.json",
@@ -41,7 +41,7 @@
"checkout_ref": "checkout:checkpoint-recovery-consistency",
"source_files": {
"src/inferrsi/data/skill_library/methods/checkpoint-recovery-consistency/SKILL.md": "6b7db1eb2780dec27858cb7aa0a3c093c5b5dfa4c30261aebfeaa255213ff688",
- "src/inferrsi/data/skill_library/methods/checkpoint-recovery-consistency/metadata.json": "4dc8659806539b4784ea9348a01d4194ac70d0bb297a916f90ff6419efa5c328",
+ "src/inferrsi/data/skill_library/methods/checkpoint-recovery-consistency/metadata.json": "4185ff9f8b4c06d2a984753127780f29ceebf50d87b225a7697bfecccc2a4a60",
"src/inferrsi/data/skill_library/methods/checkpoint-recovery-consistency/references/dlrover-persistence-replicas.md": "055f4357f79026fcbcc766bbe77db610376ec353623f8898f5d14a30ee69e7bc",
"src/inferrsi/data/skill_library/methods/checkpoint-recovery-consistency/references/sources.md": "9425b1aab963a2c8d35f3db897bd15ee69699bf5b0d0c8425c3c8bba3e9a0233",
"src/inferrsi/data/skill_library/skill-index.json": "b5a9f05bd8179644f3fea79d7eb04fe4a8baad9e8d68aa8d0f1db213cf633317"
@@ -64,6 +64,28 @@
"external"
],
"empirical_validation": "not_run",
+ "applicability": {
+ "scope": "Assessing distributed checkpoint recovery through DLRover memory agreement, persistence completion and replica transport, and NVIDIA Resiliency Extension complete-iteration selection; distinguishes worker, host and replica loss and separates queued saves, publication, application-state validity and durable recovery. Save performance belongs to dcp-async-save-performance",
+ "prerequisites": "The pinned sources and limits are read; reading authorizes no launchers, failure injection or checkpoint loading, and restore evidence needs separate authorization",
+ "input_contract": "The declared failure domain; actual memory, storage and replica placement and helper lifetime; step or iteration identity, the expected shard/rank set and finalization result; the application's model, optimizer and caller-required RNG/scheduler/data-position state contract; the target environment's format and world-size/replication assumptions",
+ "produced_artifact": "An answer to each of the four acceptance questions (snapshot survival, cross-participant completeness, application-state completeness, restorability) at its own evidence level, a proposed per-generation recovery evidence record including errors and skipped saves, and a proposed verification split by worker, host, replica and persistent-storage loss; no recovery success, durability, NPU support or measured gain",
+ "failure_modes": [
+ "Equal steps or complete rank IDs are taken to verify tensor contents or whole application semantics",
+ "A skipped save, where not every participant acquired shared-memory access, is reported as a new recovery point",
+ "A returned step paired with an empty state dictionary is read as a successful restore",
+ "A local maximum filename or one clean shard substitutes for global coverage",
+ "The .dirty-then-rename publication is read as power-loss durability",
+ "An asynchronous save request is treated as a completed checkpoint, or queue concurrency is assumed unlimited",
+ "One failure domain is tested and recovery from all of them is claimed"
+ ],
+ "refusal_point": "Claims no recovery success, durability, NPU support or measured gain, does not upgrade allocation, metadata or a completed collective into proven recovery, and neither injects failures nor authorizes them"
+ },
+ "recommendations": [
+ "Answer the four acceptance questions separately; record each at its own evidence level.",
+ "Count replication as added availability only when copies occupy independent failure domains, and budget clean, dirty and replicated shards on each host.",
+ "Keep the one-outstanding asynchronous save restriction until the target queue version is checked.",
+ "Send save-throughput questions to dcp-async-save-performance; a faster return or higher write throughput does not establish recoverability."
+ ],
"limitations": [
"Source-derived candidate; no local efficacy validation.",
"Original typed source inventory is not normalized; retain source identities and limits in the procedure/readsets."
@@ -73,10 +95,27 @@
],
"consumption_mode": "packaged"
},
- "applicability": {},
+ "applicability": {
+ "scope": "Assessing distributed checkpoint recovery through DLRover memory agreement, persistence completion and replica transport, and NVIDIA Resiliency Extension complete-iteration selection; distinguishes worker, host and replica loss and separates queued saves, publication, application-state validity and durable recovery. Save performance belongs to dcp-async-save-performance",
+ "prerequisites": "The pinned sources and limits are read; reading authorizes no launchers, failure injection or checkpoint loading, and restore evidence needs separate authorization",
+ "input_contract": "The declared failure domain; actual memory, storage and replica placement and helper lifetime; step or iteration identity, the expected shard/rank set and finalization result; the application's model, optimizer and caller-required RNG/scheduler/data-position state contract; the target environment's format and world-size/replication assumptions",
+ "produced_artifact": "An answer to each of the four acceptance questions (snapshot survival, cross-participant completeness, application-state completeness, restorability) at its own evidence level, a proposed per-generation recovery evidence record including errors and skipped saves, and a proposed verification split by worker, host, replica and persistent-storage loss; no recovery success, durability, NPU support or measured gain",
+ "failure_modes": [
+ "Equal steps or complete rank IDs are taken to verify tensor contents or whole application semantics",
+ "A skipped save, where not every participant acquired shared-memory access, is reported as a new recovery point",
+ "A returned step paired with an empty state dictionary is read as a successful restore",
+ "A local maximum filename or one clean shard substitutes for global coverage",
+ "The .dirty-then-rename publication is read as power-loss durability",
+ "An asynchronous save request is treated as a completed checkpoint, or queue concurrency is assumed unlimited",
+ "One failure domain is tested and recovery from all of them is claimed"
+ ],
+ "refusal_point": "Claims no recovery success, durability, NPU support or measured gain, does not upgrade allocation, metadata or a completed collective into proven recovery, and neither injects failures nor authorizes them"
+ },
"recommendations": [
- "Read the pinned SKILL.md and its prerequisites before choosing this procedure.",
- "Read the packaged references and owner-local cases alongside the procedure."
+ "Answer the four acceptance questions separately; record each at its own evidence level.",
+ "Count replication as added availability only when copies occupy independent failure domains, and budget clean, dirty and replicated shards on each host.",
+ "Keep the one-outstanding asynchronous save restriction until the target queue version is checked.",
+ "Send save-throughput questions to dcp-async-save-performance; a faster return or higher write throughput does not establish recoverability."
],
"limitations": [
"This generated abstract points to the complete packaged skill; external evidence and upstream repositories remain source links.",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/choose-parallel-training-plan.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/choose-parallel-training-plan.json
index c77e42442..6ef486f00 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/choose-parallel-training-plan.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/choose-parallel-training-plan.json
@@ -16,8 +16,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/methods/choose-parallel-training-plan/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/methods/choose-parallel-training-plan/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/methods/choose-parallel-training-plan/SKILL.md",
"src/inferrsi/data/skill_library/methods/choose-parallel-training-plan/metadata.json",
@@ -34,7 +34,7 @@
"checkout_ref": "checkout:choose-parallel-training-plan",
"source_files": {
"src/inferrsi/data/skill_library/methods/choose-parallel-training-plan/SKILL.md": "25d04aed9782b2a8ff1186828becdc7fce93c4418039ceaef9b8dcc5ec9bcd24",
- "src/inferrsi/data/skill_library/methods/choose-parallel-training-plan/metadata.json": "c6acae57d70190cea0e18c71c79127962181e58169e011df75d8635b3db4daa7",
+ "src/inferrsi/data/skill_library/methods/choose-parallel-training-plan/metadata.json": "4990d0a2f1a269111a680f45a5d91221af37e3f8b3ce514fea84e97878e767cd",
"src/inferrsi/data/skill_library/methods/choose-parallel-training-plan/references/source-notes.md": "9f9184d46b19312138a6e805e439e43dc0594a0cc6b46675255a582dc946ca5b",
"src/inferrsi/data/skill_library/skill-index.json": "b5a9f05bd8179644f3fea79d7eb04fe4a8baad9e8d68aa8d0f1db213cf633317"
},
@@ -49,6 +49,26 @@
"origin": [
"external"
],
+ "applicability": {
+ "scope": "Comparing stage, device-mesh and operator-sharding candidates for static distributed training when heterogeneous stage memory or work, or communication cost, makes a uniform data/tensor/pipeline plan suspect; derived from historical Alpa, not a current compiler configuration guide",
+ "prerequisites": "The packaged source notes are read; the skill authorizes no compiler or solver install or run, parallelism reconfiguration, repository edit, hardware task or benchmark",
+ "input_contract": "Model operator graph and shapes; global batch, microbatch and accumulation; precision and optimizer; physical links and contention; candidate shard layouts; stage latency, workspace and in-flight activation memory; the objective and expected run length. A device count without topology is insufficient",
+ "produced_artifact": "The best supported candidate and a simpler fallback with a memory and critical-path ledger, planning cost and amortization, semantic invariants, counter-signals, missing evidence and a proposed matched comparison; a plan, not a reconfiguration",
+ "failure_modes": [
+ "Equal arithmetic work is taken to mean equal kernel time, carrying over Alpa's zero compute-cost simplification",
+ "Mean memory replaces the largest per-rank peak with multiple in-flight activation sets",
+ "Inter-stage transfer and microbatch count stay outside the objective, as in historical Alpa, without the gap being reported",
+ "Repeated slow-link sends to replicated destination tiles are not compared with one cross-mesh transfer plus local allgather",
+ "Inter-node Network traffic is merged with intra-node accelerator links",
+ "Static cost estimates are applied to dynamic shapes, asymmetric devices or changing communication contention"
+ ],
+ "refusal_point": "Rejects infeasible stage/mesh pairs, claims no global optimality outside the declared search space, and stops at the plan in paper-only mode"
+ },
+ "recommendations": [
+ "Keep a simple reference plan with the same training semantics and global batch as the comparison point.",
+ "Separate fixed-model scaling from model-size weak scaling before quoting a speedup; Alpa's 9.7x MoE ratio belongs to a specific weak-scaling workload, baseline and 25-Gbps cluster.",
+ "Weigh planning cost against the expected run length; a shorter job may not amortize it."
+ ],
"limitations": [
"Source-derived candidate; no local efficacy validation.",
"Original typed source inventory is not normalized; retain source identities and limits in the procedure/readsets."
@@ -58,10 +78,25 @@
],
"consumption_mode": "packaged"
},
- "applicability": {},
+ "applicability": {
+ "scope": "Comparing stage, device-mesh and operator-sharding candidates for static distributed training when heterogeneous stage memory or work, or communication cost, makes a uniform data/tensor/pipeline plan suspect; derived from historical Alpa, not a current compiler configuration guide",
+ "prerequisites": "The packaged source notes are read; the skill authorizes no compiler or solver install or run, parallelism reconfiguration, repository edit, hardware task or benchmark",
+ "input_contract": "Model operator graph and shapes; global batch, microbatch and accumulation; precision and optimizer; physical links and contention; candidate shard layouts; stage latency, workspace and in-flight activation memory; the objective and expected run length. A device count without topology is insufficient",
+ "produced_artifact": "The best supported candidate and a simpler fallback with a memory and critical-path ledger, planning cost and amortization, semantic invariants, counter-signals, missing evidence and a proposed matched comparison; a plan, not a reconfiguration",
+ "failure_modes": [
+ "Equal arithmetic work is taken to mean equal kernel time, carrying over Alpa's zero compute-cost simplification",
+ "Mean memory replaces the largest per-rank peak with multiple in-flight activation sets",
+ "Inter-stage transfer and microbatch count stay outside the objective, as in historical Alpa, without the gap being reported",
+ "Repeated slow-link sends to replicated destination tiles are not compared with one cross-mesh transfer plus local allgather",
+ "Inter-node Network traffic is merged with intra-node accelerator links",
+ "Static cost estimates are applied to dynamic shapes, asymmetric devices or changing communication contention"
+ ],
+ "refusal_point": "Rejects infeasible stage/mesh pairs, claims no global optimality outside the declared search space, and stops at the plan in paper-only mode"
+ },
"recommendations": [
- "Read the pinned SKILL.md and its prerequisites before choosing this procedure.",
- "Read the packaged references and owner-local cases alongside the procedure."
+ "Keep a simple reference plan with the same training semantics and global batch as the comparison point.",
+ "Separate fixed-model scaling from model-size weak scaling before quoting a speedup; Alpa's 9.7x MoE ratio belongs to a specific weak-scaling workload, baseline and 25-Gbps cluster.",
+ "Weigh planning cost against the expected run length; a shorter job may not amortize it."
],
"limitations": [
"This generated abstract points to the complete packaged skill; external evidence and upstream repositories remain source links.",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/cite-check-downstream-pr.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/cite-check-downstream-pr.json
index 234c28387..9f61b1d91 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/cite-check-downstream-pr.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/cite-check-downstream-pr.json
@@ -22,8 +22,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/workflows/cite-check-downstream-pr/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/workflows/cite-check-downstream-pr/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/skill-index.json",
"src/inferrsi/data/skill_library/workflows/cite-check-downstream-pr/SKILL.md",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/claim-contract.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/claim-contract.json
index 99507873a..8a53c56bd 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/claim-contract.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/claim-contract.json
@@ -22,8 +22,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/workflows/claim-contract/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/workflows/claim-contract/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/skill-index.json",
"src/inferrsi/data/skill_library/workflows/claim-contract/SKILL.md",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/comfy-ui.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/comfy-ui.json
index e8be332c3..dc6b366a6 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/comfy-ui.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/comfy-ui.json
@@ -22,8 +22,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/repositories/comfy-ui/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/repositories/comfy-ui/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/repositories/comfy-ui/LICENSE",
"src/inferrsi/data/skill_library/repositories/comfy-ui/SKILL.md",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/competitor-scan.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/competitor-scan.json
index a6123d5db..78717465d 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/competitor-scan.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/competitor-scan.json
@@ -21,8 +21,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/workflows/competitor-scan/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/workflows/competitor-scan/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/skill-index.json",
"src/inferrsi/data/skill_library/workflows/competitor-scan/SKILL.md",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/dcp-async-save-performance.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/dcp-async-save-performance.json
index 268a99d95..abaca753f 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/dcp-async-save-performance.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/dcp-async-save-performance.json
@@ -23,8 +23,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/frameworks/pytorch/dcp-async-save-performance/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/frameworks/pytorch/dcp-async-save-performance/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/frameworks/pytorch/dcp-async-save-performance/SKILL.md",
"src/inferrsi/data/skill_library/frameworks/pytorch/dcp-async-save-performance/metadata.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/dense-performance-evidence.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/dense-performance-evidence.json
index 163a91576..01e03d395 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/dense-performance-evidence.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/dense-performance-evidence.json
@@ -21,8 +21,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/methods/dense-performance-evidence/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/methods/dense-performance-evidence/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/methods/dense-performance-evidence/SKILL.md",
"src/inferrsi/data/skill_library/methods/dense-performance-evidence/metadata.json",
@@ -38,7 +38,7 @@
"checkout_ref": "checkout:dense-performance-evidence",
"source_files": {
"src/inferrsi/data/skill_library/methods/dense-performance-evidence/SKILL.md": "53cc1867a13f54d1006d50517dea8848869ca50df3f0cd610b2424cc8d18c933",
- "src/inferrsi/data/skill_library/methods/dense-performance-evidence/metadata.json": "a0d1e3b17140295510a764045d1b222e19be3a3406e6d1e7dddaad49ff474490",
+ "src/inferrsi/data/skill_library/methods/dense-performance-evidence/metadata.json": "041780985dcaf76809d29dc0d4e242359b17b6bcfaadab2d05674fa4699a5427",
"src/inferrsi/data/skill_library/skill-index.json": "b5a9f05bd8179644f3fea79d7eb04fe4a8baad9e8d68aa8d0f1db213cf633317"
},
"skill_metadata": {
@@ -58,6 +58,26 @@
"external"
],
"empirical_validation": "not_run",
+ "applicability": {
+ "scope": "Deciding which performance evidence to collect next and which claim each evidence level supports, from correctness through microbenchmark, component profile, isolated scenario benchmark and multi-regime serving benchmark to a production A/B experiment; also structuring a measurement campaign",
+ "prerequisites": "Level 0 correctness evidence is collected before any performance measurement; a Level 5 experiment needs production infrastructure, traffic splitting and statistical power",
+ "input_contract": "The claim to be made or the question to answer, and the evidence already collected with its level, workload, hardware/software context and controls",
+ "produced_artifact": "The minimum evidence level each claim requires, the next evidence to collect under the escalation rule, and negative-result records with hypothesis, evidence level, the metric that did not improve, context and the reason if known; a classification of evidence, not a measurement",
+ "failure_modes": [
+ "An md5 match or a test pass is read as semantically correct output or as support for a performance claim",
+ "A microbenchmark improvement is reported as an inference throughput gain",
+ "Absolute kernel durations or single collective latencies are compared across devices or sessions",
+ "A win on one fixed workload or concurrency is generalized to production traffic",
+ "Worse metrics, per-task regressions or negative results are discarded, or a shared task budget is taken as equal total cost"
+ ],
+ "refusal_point": "Withholds any claim above the level its evidence reaches, leaves an unresolved source protocol or denominator mismatch unresolved, and lets neither upstream scores nor local packaging checks promote an InferRSI candidate"
+ },
+ "recommendations": [
+ "Collect Level 0 first and profile at Level 2 before implementing; escalate to Level 3 only after a promising Level 2 result, and to Level 4 only after a Level 3 win worth validating across regimes.",
+ "Collect a Level 1 microbenchmark only for an isolated-kernel question after Level 2 has confirmed that kernel is on the critical path.",
+ "For a Level 1 claim, pin input shape and dtype, exclude warmup, synchronize before timing, run at least 10 repetitions and report median and IQR; for a Level 3 claim, match hardware, software stack, model revision and inputs, run at least 3 repetitions and alternate A/B/A order.",
+ "Read the admission/drain occupancy case when interpreting finite service cohorts; phase-selected occupancy is not end-to-end gain."
+ ],
"limitations": [
"Source-derived candidate; no local efficacy validation.",
"Original typed source inventory is not normalized; retain source identities and limits in the procedure/readsets."
@@ -69,10 +89,25 @@
],
"consumption_mode": "packaged"
},
- "applicability": {},
+ "applicability": {
+ "scope": "Deciding which performance evidence to collect next and which claim each evidence level supports, from correctness through microbenchmark, component profile, isolated scenario benchmark and multi-regime serving benchmark to a production A/B experiment; also structuring a measurement campaign",
+ "prerequisites": "Level 0 correctness evidence is collected before any performance measurement; a Level 5 experiment needs production infrastructure, traffic splitting and statistical power",
+ "input_contract": "The claim to be made or the question to answer, and the evidence already collected with its level, workload, hardware/software context and controls",
+ "produced_artifact": "The minimum evidence level each claim requires, the next evidence to collect under the escalation rule, and negative-result records with hypothesis, evidence level, the metric that did not improve, context and the reason if known; a classification of evidence, not a measurement",
+ "failure_modes": [
+ "An md5 match or a test pass is read as semantically correct output or as support for a performance claim",
+ "A microbenchmark improvement is reported as an inference throughput gain",
+ "Absolute kernel durations or single collective latencies are compared across devices or sessions",
+ "A win on one fixed workload or concurrency is generalized to production traffic",
+ "Worse metrics, per-task regressions or negative results are discarded, or a shared task budget is taken as equal total cost"
+ ],
+ "refusal_point": "Withholds any claim above the level its evidence reaches, leaves an unresolved source protocol or denominator mismatch unresolved, and lets neither upstream scores nor local packaging checks promote an InferRSI candidate"
+ },
"recommendations": [
- "Read the pinned SKILL.md and its prerequisites before choosing this procedure.",
- "Read the packaged references and owner-local cases alongside the procedure."
+ "Collect Level 0 first and profile at Level 2 before implementing; escalate to Level 3 only after a promising Level 2 result, and to Level 4 only after a Level 3 win worth validating across regimes.",
+ "Collect a Level 1 microbenchmark only for an isolated-kernel question after Level 2 has confirmed that kernel is on the critical path.",
+ "For a Level 1 claim, pin input shape and dtype, exclude warmup, synchronize before timing, run at least 10 repetitions and report median and IQR; for a Level 3 claim, match hardware, software stack, model revision and inputs, run at least 3 repetitions and alternate A/B/A order.",
+ "Read the admission/drain occupancy case when interpreting finite service cohorts; phase-selected occupancy is not end-to-end gain."
],
"limitations": [
"This generated abstract points to the complete packaged skill; external evidence and upstream repositories remain source links.",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/diagnose-vllm-admission-pressure.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/diagnose-vllm-admission-pressure.json
index 1b051cb95..16490d466 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/diagnose-vllm-admission-pressure.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/diagnose-vllm-admission-pressure.json
@@ -21,8 +21,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/repositories/vllm/diagnose-vllm-admission-pressure/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/repositories/vllm/diagnose-vllm-admission-pressure/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/repositories/vllm/diagnose-vllm-admission-pressure/SKILL.md",
"src/inferrsi/data/skill_library/repositories/vllm/diagnose-vllm-admission-pressure/metadata.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/diffusers-invocation-state.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/diffusers-invocation-state.json
index 345989799..3eb243d6f 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/diffusers-invocation-state.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/diffusers-invocation-state.json
@@ -22,8 +22,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/frameworks/diffusers/diffusers-invocation-state/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/frameworks/diffusers/diffusers-invocation-state/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/frameworks/diffusers/diffusers-invocation-state/SKILL.md",
"src/inferrsi/data/skill_library/frameworks/diffusers/diffusers-invocation-state/metadata.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/draft-acceptance-issue.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/draft-acceptance-issue.json
index c61f32fe3..204dd8a30 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/draft-acceptance-issue.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/draft-acceptance-issue.json
@@ -22,8 +22,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/workflows/draft-acceptance-issue/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/workflows/draft-acceptance-issue/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/skill-index.json",
"src/inferrsi/data/skill_library/workflows/draft-acceptance-issue/SKILL.md",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/dwarfstar.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/dwarfstar.json
index 5b0ff9d30..d1348c401 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/dwarfstar.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/dwarfstar.json
@@ -25,8 +25,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/repositories/dwarfstar/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/repositories/dwarfstar/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/repositories/dwarfstar/LICENSE",
"src/inferrsi/data/skill_library/repositories/dwarfstar/SKILL.md",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/evaluate-elastic-serving-placement.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/evaluate-elastic-serving-placement.json
index d462bacea..697011f94 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/evaluate-elastic-serving-placement.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/evaluate-elastic-serving-placement.json
@@ -16,8 +16,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/methods/evaluate-elastic-serving-placement/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/methods/evaluate-elastic-serving-placement/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/methods/evaluate-elastic-serving-placement/SKILL.md",
"src/inferrsi/data/skill_library/methods/evaluate-elastic-serving-placement/metadata.json",
@@ -34,7 +34,7 @@
"checkout_ref": "checkout:evaluate-elastic-serving-placement",
"source_files": {
"src/inferrsi/data/skill_library/methods/evaluate-elastic-serving-placement/SKILL.md": "82d0099d3c33706f46c657d4fa7f76d3d486e019b36760d1de001921328a60cc",
- "src/inferrsi/data/skill_library/methods/evaluate-elastic-serving-placement/metadata.json": "37379a4d4a058ca46da2b20a920dc01be9ec11e557c32b2332702ec76027f4b5",
+ "src/inferrsi/data/skill_library/methods/evaluate-elastic-serving-placement/metadata.json": "091a3b95d1a8996b5ad917e7948efb4ed598e31c7163709377f8e9199cebfe26",
"src/inferrsi/data/skill_library/methods/evaluate-elastic-serving-placement/references/source-notes.md": "ea416831fd7d0874e06d0a76fb0c7196df53072b717673a0cdf635db6bfd917b",
"src/inferrsi/data/skill_library/skill-index.json": "b5a9f05bd8179644f3fea79d7eb04fe4a8baad9e8d68aa8d0f1db213cf633317"
},
@@ -50,6 +50,26 @@
"origin": [
"external"
],
+ "applicability": {
+ "scope": "Comparing migration of a running LLM request between compatible model replicas with changing its sequence-parallel execution group, including state-copy, reservation, recomputation, interruption and priority costs, for cross-instance memory imbalance or highly variable long-context serving; not connector wiring or live scaling",
+ "prerequisites": "The relevant Llumnix or LoongServe branch and its readset in the packaged source notes are read; moving state, altering admission or priority, scaling replicas, changing parallelism, running experiments or editing frameworks needs separate authorization",
+ "input_contract": "Exact model and cache identity and format; weight residency; request phase, context and generated length; source and destination KV bytes; physical free/reserved memory and virtual accounting; batch state; current parallel group; queue and service times; actual topology and transport for each path; the target SLO statistic and affected priority or tenant cohorts",
+ "produced_artifact": "The smallest justified branch with missing capability or evidence, expected saved work or wait, added state, traffic and replica cost, affected SLO and cohorts, counter-signals, the proposed next gate and a future verification plan; an evidence-based recommendation, not a validated deployment rule",
+ "failure_modes": [
+ "Model-weight repartitioning, pipeline configuration, instance migration and sequence sharding are treated as interchangeable uses of parallelism",
+ "Aggregate free bytes are taken to mean a single blocked request fits, conflating internal KV-block slack with cross-instance fragmentation",
+ "Llumnix's 20-30 ms measured downtime is substituted for total transfer time",
+ "Virtual headroom is counted as extra memory or as a tenant-fairness guarantee",
+ "A zero-overhead transition label is read as every group change being free, or a token sum hides length-sensitive attention work",
+ "Ongoing distributed-attention communication goes uncharged because existing KV does not migrate"
+ ],
+ "refusal_point": "Withholds the candidate when every destination is full, pre-copy catch-up is unproven, model or cache compatibility is unknown, or the measured work saved does not exceed exposed costs; current-code review and target-system tests remain a later authorization gate"
+ },
+ "recommendations": [
+ "Decide first whether the question is replica relocation, a sequence-parallel group change, or prefix reuse and phase pools, which evaluate-kv-cache-transfer-gates and choose-prefill-decode-schedule cover when available.",
+ "Name each transfer path separately; only inter-node transport is Network.",
+ "Keep a feasible static, no-migration baseline in any later comparison."
+ ],
"limitations": [
"Source-derived candidate; no local efficacy validation.",
"Original typed source inventory is not normalized; retain source identities and limits in the procedure/readsets."
@@ -59,10 +79,25 @@
],
"consumption_mode": "packaged"
},
- "applicability": {},
+ "applicability": {
+ "scope": "Comparing migration of a running LLM request between compatible model replicas with changing its sequence-parallel execution group, including state-copy, reservation, recomputation, interruption and priority costs, for cross-instance memory imbalance or highly variable long-context serving; not connector wiring or live scaling",
+ "prerequisites": "The relevant Llumnix or LoongServe branch and its readset in the packaged source notes are read; moving state, altering admission or priority, scaling replicas, changing parallelism, running experiments or editing frameworks needs separate authorization",
+ "input_contract": "Exact model and cache identity and format; weight residency; request phase, context and generated length; source and destination KV bytes; physical free/reserved memory and virtual accounting; batch state; current parallel group; queue and service times; actual topology and transport for each path; the target SLO statistic and affected priority or tenant cohorts",
+ "produced_artifact": "The smallest justified branch with missing capability or evidence, expected saved work or wait, added state, traffic and replica cost, affected SLO and cohorts, counter-signals, the proposed next gate and a future verification plan; an evidence-based recommendation, not a validated deployment rule",
+ "failure_modes": [
+ "Model-weight repartitioning, pipeline configuration, instance migration and sequence sharding are treated as interchangeable uses of parallelism",
+ "Aggregate free bytes are taken to mean a single blocked request fits, conflating internal KV-block slack with cross-instance fragmentation",
+ "Llumnix's 20-30 ms measured downtime is substituted for total transfer time",
+ "Virtual headroom is counted as extra memory or as a tenant-fairness guarantee",
+ "A zero-overhead transition label is read as every group change being free, or a token sum hides length-sensitive attention work",
+ "Ongoing distributed-attention communication goes uncharged because existing KV does not migrate"
+ ],
+ "refusal_point": "Withholds the candidate when every destination is full, pre-copy catch-up is unproven, model or cache compatibility is unknown, or the measured work saved does not exceed exposed costs; current-code review and target-system tests remain a later authorization gate"
+ },
"recommendations": [
- "Read the pinned SKILL.md and its prerequisites before choosing this procedure.",
- "Read the packaged references and owner-local cases alongside the procedure."
+ "Decide first whether the question is replica relocation, a sequence-parallel group change, or prefix reuse and phase pools, which evaluate-kv-cache-transfer-gates and choose-prefill-decode-schedule cover when available.",
+ "Name each transfer path separately; only inter-node transport is Network.",
+ "Keep a feasible static, no-migration baseline in any later comparison."
],
"limitations": [
"This generated abstract points to the complete packaged skill; external evidence and upstream repositories remain source links.",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/generation-cache-lifecycle.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/generation-cache-lifecycle.json
index 49d6f4f08..d0374a3e5 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/generation-cache-lifecycle.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/generation-cache-lifecycle.json
@@ -22,8 +22,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/repositories/mlx-lm/generation-cache-lifecycle/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/repositories/mlx-lm/generation-cache-lifecycle/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/repositories/mlx-lm/generation-cache-lifecycle/SKILL.md",
"src/inferrsi/data/skill_library/repositories/mlx-lm/generation-cache-lifecycle/metadata.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/host-memory-management.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/host-memory-management.json
index 2104c11e8..0bc3e959f 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/host-memory-management.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/host-memory-management.json
@@ -30,8 +30,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/hardware/host/host-memory-management/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/hardware/host/host-memory-management/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/hardware/host/host-memory-management/SKILL.md",
"src/inferrsi/data/skill_library/hardware/host/host-memory-management/metadata.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/host-overhead-optimization.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/host-overhead-optimization.json
index a9a84fed4..50c7c46e8 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/host-overhead-optimization.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/host-overhead-optimization.json
@@ -22,8 +22,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/hardware/host/host-overhead-optimization/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/hardware/host/host-overhead-optimization/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/hardware/host/host-overhead-optimization/SKILL.md",
"src/inferrsi/data/skill_library/hardware/host/host-overhead-optimization/metadata.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/husky.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/husky.json
index 8a38a23e4..29ea89b99 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/husky.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/husky.json
@@ -21,8 +21,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/repositories/husky/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/repositories/husky/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/repositories/husky/SKILL.md",
"src/inferrsi/data/skill_library/repositories/husky/metadata.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/inference-benchmark-design.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/inference-benchmark-design.json
index 128279b6b..074f45352 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/inference-benchmark-design.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/inference-benchmark-design.json
@@ -30,8 +30,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/methods/inference-benchmark-design/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/methods/inference-benchmark-design/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/methods/inference-benchmark-design/SKILL.md",
"src/inferrsi/data/skill_library/methods/inference-benchmark-design/metadata.json",
@@ -49,7 +49,7 @@
"checkout_ref": "checkout:inference-benchmark-design",
"source_files": {
"src/inferrsi/data/skill_library/methods/inference-benchmark-design/SKILL.md": "59b5656e7d989b3cee8d6f95cd43be03053b1fb11c3a82bb49ab3d402f7aacee",
- "src/inferrsi/data/skill_library/methods/inference-benchmark-design/metadata.json": "18310402a8ade03fd48ac687f79ba9d447871e90337de84c0e1bc15c0b827306",
+ "src/inferrsi/data/skill_library/methods/inference-benchmark-design/metadata.json": "d51b5b3600d13b38c2adbf6c6e149c7375c9f1db12bc078679cab67a32a3e9a7",
"src/inferrsi/data/skill_library/methods/inference-benchmark-design/references/benchmark-commands.md": "0898424e0bff3a6246e1eeeda6abd6c5bd974c047fe32f3bbd5ab9d31ce17cd3",
"src/inferrsi/data/skill_library/methods/inference-benchmark-design/references/task-benchmark-selection.md": "0d4bb080a754e26846e81680023db0dff5e6a4c96be4327c411c3b72ac4ef8e2",
"src/inferrsi/data/skill_library/skill-index.json": "b5a9f05bd8179644f3fea79d7eb04fe4a8baad9e8d68aa8d0f1db213cf633317"
@@ -200,6 +200,28 @@
}
],
"empirical_validation": "not_run",
+ "applicability": {
+ "scope": "Selecting task quality gates, reference parity checks and performance workloads for text, agentic coding, work agents, TTS, full-duplex audio/video, image and video inference, and designing goodput-based multi-regime comparisons that fit InferRSI's evidence trail",
+ "prerequisites": "The primary objective, timing boundary, quality gates, sample/failure accounting, required regimes, repetitions, seeds and decision rule are frozen before optimization; baseline and candidate share hardware, software stack, input distribution, benchmark configuration, warmup and measurement protocol, and on a shared machine also one time window and dedicated devices",
+ "input_contract": "The task family and deployment use case; candidate quality evaluators and reference parity; the declared SLO and the caller's policy on which regimes must improve or must not regress; for an audit, a supplied fixed-workload report with request-to-trial grouping, exclusion and failure counts and the declared aggregation rule",
+ "produced_artifact": "A frozen benchmark design with evaluators, reference parity, primary objective, timing boundary, quality gates, required regimes, repetition and seed rules and an explicit decision config, plus the evidence-tree layout a PASS claim must produce; for a supplied report, the computed threshold result and missing evidence. Candidate knowledge, not qualification of any model or backend",
+ "failure_modes": [
+ "Raw throughput is optimized in place of goodput, so requests served outside the SLO count toward a win",
+ "One concurrency level or one undisclosed aggregate score stands in for all required regimes",
+ "Recorded agent task scores or replay completion are read as current task success, or a VBench prompt workload as computed VBench quality",
+ "An upstream better/worse/inconclusive label is used as InferRSI's PASS/FAIL/INCONCLUSIVE verdict",
+ "Runs on the same hardware class, in different time windows or on shared devices are compared as if pinned",
+ "A quality pass is taken to certify that the timed work actually occurred",
+ "Best-seen results are reported as delivered without restarting and validating the final code/config artifact"
+ ],
+ "refusal_point": "Records missing coverage and high-variance results as INCONCLUSIVE rather than a win; a run without a frozen RunSpec, or whose validator changed after the plan was frozen, is INCONCLUSIVE regardless of metric values; sets no universal numeric gate or noise threshold"
+ },
+ "recommendations": [
+ "Freeze one primary objective, timing boundary, quality gates and sample/failure accounting before optimization.",
+ "Choose the required regimes from the deployment use case, report each separately, and record which must improve or must not regress in a versioned decision config.",
+ "Use the replay trace that matches the actual deployment pattern.",
+ "State the concurrency, trace, model, SKU and microbenchmark scope boundaries in any report that cites benchmark results."
+ ],
"limitations": [
"Source-derived candidate; no local efficacy validation.",
"Original typed source inventory is not normalized; retain source identities and limits in the procedure/readsets.",
@@ -216,10 +238,27 @@
],
"consumption_mode": "packaged"
},
- "applicability": {},
+ "applicability": {
+ "scope": "Selecting task quality gates, reference parity checks and performance workloads for text, agentic coding, work agents, TTS, full-duplex audio/video, image and video inference, and designing goodput-based multi-regime comparisons that fit InferRSI's evidence trail",
+ "prerequisites": "The primary objective, timing boundary, quality gates, sample/failure accounting, required regimes, repetitions, seeds and decision rule are frozen before optimization; baseline and candidate share hardware, software stack, input distribution, benchmark configuration, warmup and measurement protocol, and on a shared machine also one time window and dedicated devices",
+ "input_contract": "The task family and deployment use case; candidate quality evaluators and reference parity; the declared SLO and the caller's policy on which regimes must improve or must not regress; for an audit, a supplied fixed-workload report with request-to-trial grouping, exclusion and failure counts and the declared aggregation rule",
+ "produced_artifact": "A frozen benchmark design with evaluators, reference parity, primary objective, timing boundary, quality gates, required regimes, repetition and seed rules and an explicit decision config, plus the evidence-tree layout a PASS claim must produce; for a supplied report, the computed threshold result and missing evidence. Candidate knowledge, not qualification of any model or backend",
+ "failure_modes": [
+ "Raw throughput is optimized in place of goodput, so requests served outside the SLO count toward a win",
+ "One concurrency level or one undisclosed aggregate score stands in for all required regimes",
+ "Recorded agent task scores or replay completion are read as current task success, or a VBench prompt workload as computed VBench quality",
+ "An upstream better/worse/inconclusive label is used as InferRSI's PASS/FAIL/INCONCLUSIVE verdict",
+ "Runs on the same hardware class, in different time windows or on shared devices are compared as if pinned",
+ "A quality pass is taken to certify that the timed work actually occurred",
+ "Best-seen results are reported as delivered without restarting and validating the final code/config artifact"
+ ],
+ "refusal_point": "Records missing coverage and high-variance results as INCONCLUSIVE rather than a win; a run without a frozen RunSpec, or whose validator changed after the plan was frozen, is INCONCLUSIVE regardless of metric values; sets no universal numeric gate or noise threshold"
+ },
"recommendations": [
- "Read the pinned SKILL.md and its prerequisites before choosing this procedure.",
- "Read the packaged references and owner-local cases alongside the procedure."
+ "Freeze one primary objective, timing boundary, quality gates and sample/failure accounting before optimization.",
+ "Choose the required regimes from the deployment use case, report each separately, and record which must improve or must not regress in a versioned decision config.",
+ "Use the replay trace that matches the actual deployment pattern.",
+ "State the concurrency, trace, model, SKU and microbenchmark scope boundaries in any report that cites benchmark results."
],
"limitations": [
"This generated abstract points to the complete packaged skill; external evidence and upstream repositories remain source links.",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/inference-optimization-casebook.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/inference-optimization-casebook.json
index 78099b182..64ebace8c 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/inference-optimization-casebook.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/inference-optimization-casebook.json
@@ -21,8 +21,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/methods/inference-optimization-casebook/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/methods/inference-optimization-casebook/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/methods/inference-optimization-casebook/SKILL.md",
"src/inferrsi/data/skill_library/methods/inference-optimization-casebook/cases/admission-drain-occupancy.json",
@@ -54,7 +54,7 @@
"src/inferrsi/data/skill_library/methods/inference-optimization-casebook/cases/qwen-ascend-rejected-directions.md": "be396983812262515899a3e2ba4f01b23ff92e1b836881882c7580372b8ae552",
"src/inferrsi/data/skill_library/methods/inference-optimization-casebook/cases/qwen-b300-rejected-directions.json": "146ea1564a46ed03092c7dc5796b318fc05b04b1821105ee00eb7fa7fa1333e2",
"src/inferrsi/data/skill_library/methods/inference-optimization-casebook/cases/qwen-b300-rejected-directions.md": "cce2930f544a0f78f21797fba4d8ded298e4a81f76c34d7977b0e31d361bc738",
- "src/inferrsi/data/skill_library/methods/inference-optimization-casebook/metadata.json": "fb55067fb45970319787068a3206110641cceaf0ccbabd706be6d0b3d2afc2ed",
+ "src/inferrsi/data/skill_library/methods/inference-optimization-casebook/metadata.json": "963488658f327d2379183d70357eabf832d70a6a70db2aa67f548493f4088d97",
"src/inferrsi/data/skill_library/skill-index.json": "b5a9f05bd8179644f3fea79d7eb04fe4a8baad9e8d68aa8d0f1db213cf633317"
},
"skill_metadata": {
@@ -74,6 +74,27 @@
"external"
],
"empirical_validation": "not_run",
+ "applicability": {
+ "scope": "Structuring and querying a casebook of past optimization decisions as decision support, when recording optimization experiments or choosing the next optimization from prior art; the casebook is reference material and neither runs profiling nor makes decisions",
+ "prerequisites": "Whole-body text access to the case files for free-text search; applying a case first requires reading the nearest checked-in config and the model's deployment guide",
+ "input_contract": "For a query: natural search terms such as model name, knob, mechanism word or deployment feature, plus the current model and hardware. For a new case: mechanism, applies-when and does-not-apply conditions, procedure, expected effect, accuracy risk, verification, rollback and a real measurement trail with run ID, verdict, exact hardware SKU, model revision, workload and before/after result",
+ "produced_artifact": "Ranked matching cases read as precedents and diagnostic options, or a new or deprecated case record in the template; a precedent's procedure and expected direction of effect, not a guaranteed configuration or a measured result",
+ "failure_modes": [
+ "The query is pre-translated into canonical vocabulary or bounded by structured fields, so synonyms in the body text are missed",
+ "A case's specific values, such as batch size, tensor-parallel degree or kernel parameters, are copied instead of re-derived for the current model and hardware",
+ "The precedent's expected direction of effect stands in for measurement after the change is applied",
+ "A rejected direction is read as a permanent blacklist, or one discarded implementation as ruling out its mechanism",
+ "A confounded run, a missing experiment and a failed gate are recorded as the same evidence state",
+ "Cross-system cases are bootstrapped without re-verifying that the mechanism holds on the target hardware"
+ ],
+ "refusal_point": "Adds no case for an INCONCLUSIVE optimization-effect claim, a single-configuration result without plausible generalization, or a result without a real measurement trail, except a scoped diagnostic observation that keeps its verdict and configuration boundary; manufactures no PASS, and reading the records does not reopen a paused campaign"
+ },
+ "recommendations": [
+ "Search the full case bodies with free text first, then rank the hits by bottleneck class and signals.",
+ "Before retrying a rejected direction, state which condition changed or which uncertainty a bounded check resolves, run an unchanged control and choose the least expensive discriminating check.",
+ "Mark bootstrapped cases source_grounded or author_reported until a local measurement confirms or refutes them.",
+ "When a case's conditions no longer hold, mark it deprecated with the reason instead of deleting it."
+ ],
"limitations": [
"Source-derived candidate; no local efficacy validation.",
"Original typed source inventory is not normalized; retain source identities and limits in the procedure/readsets."
@@ -81,10 +102,26 @@
"provenance_readsets": [],
"consumption_mode": "packaged"
},
- "applicability": {},
+ "applicability": {
+ "scope": "Structuring and querying a casebook of past optimization decisions as decision support, when recording optimization experiments or choosing the next optimization from prior art; the casebook is reference material and neither runs profiling nor makes decisions",
+ "prerequisites": "Whole-body text access to the case files for free-text search; applying a case first requires reading the nearest checked-in config and the model's deployment guide",
+ "input_contract": "For a query: natural search terms such as model name, knob, mechanism word or deployment feature, plus the current model and hardware. For a new case: mechanism, applies-when and does-not-apply conditions, procedure, expected effect, accuracy risk, verification, rollback and a real measurement trail with run ID, verdict, exact hardware SKU, model revision, workload and before/after result",
+ "produced_artifact": "Ranked matching cases read as precedents and diagnostic options, or a new or deprecated case record in the template; a precedent's procedure and expected direction of effect, not a guaranteed configuration or a measured result",
+ "failure_modes": [
+ "The query is pre-translated into canonical vocabulary or bounded by structured fields, so synonyms in the body text are missed",
+ "A case's specific values, such as batch size, tensor-parallel degree or kernel parameters, are copied instead of re-derived for the current model and hardware",
+ "The precedent's expected direction of effect stands in for measurement after the change is applied",
+ "A rejected direction is read as a permanent blacklist, or one discarded implementation as ruling out its mechanism",
+ "A confounded run, a missing experiment and a failed gate are recorded as the same evidence state",
+ "Cross-system cases are bootstrapped without re-verifying that the mechanism holds on the target hardware"
+ ],
+ "refusal_point": "Adds no case for an INCONCLUSIVE optimization-effect claim, a single-configuration result without plausible generalization, or a result without a real measurement trail, except a scoped diagnostic observation that keeps its verdict and configuration boundary; manufactures no PASS, and reading the records does not reopen a paused campaign"
+ },
"recommendations": [
- "Read the pinned SKILL.md and its prerequisites before choosing this procedure.",
- "Read the packaged references and owner-local cases alongside the procedure."
+ "Search the full case bodies with free text first, then rank the hits by bottleneck class and signals.",
+ "Before retrying a rejected direction, state which condition changed or which uncertainty a bounded check resolves, run an unchanged control and choose the least expensive discriminating check.",
+ "Mark bootstrapped cases source_grounded or author_reported until a local measurement confirms or refutes them.",
+ "When a case's conditions no longer hold, mark it deprecated with the reason instead of deleting it."
],
"limitations": [
"This generated abstract points to the complete packaged skill; external evidence and upstream repositories remain source links.",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/kernel-writing-guide.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/kernel-writing-guide.json
index 6b42d688c..372fd2351 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/kernel-writing-guide.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/kernel-writing-guide.json
@@ -39,8 +39,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/methods/kernel-writing-guide/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/methods/kernel-writing-guide/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/methods/kernel-writing-guide/SKILL.md",
"src/inferrsi/data/skill_library/methods/kernel-writing-guide/cases/causal-conv1d-dispatch-state.md",
@@ -79,7 +79,7 @@
"src/inferrsi/data/skill_library/methods/kernel-writing-guide/cases/vllm-gdn-causal-conv-contract.md": "2f3dcee66e2dfe7a98bee21a36b892538998de5de3caf1bc63257137f9984236",
"src/inferrsi/data/skill_library/methods/kernel-writing-guide/cases/vllm-gdn-source-contract.json": "73f1a2d8fa89ce05aead2091a05a285b13cbdc314d05452c35c6a2ad15ffeff7",
"src/inferrsi/data/skill_library/methods/kernel-writing-guide/cases/vllm-gdn-source-contract.md": "7f9ebe8de7e55d3db2a119b6a56e15c37448fd7d4e022a47122a143206d31687",
- "src/inferrsi/data/skill_library/methods/kernel-writing-guide/metadata.json": "c86f93f4eacde789e5438327d85b1de0bfe623fb87acad49306e47b7e8ef1261",
+ "src/inferrsi/data/skill_library/methods/kernel-writing-guide/metadata.json": "766aa688d573c790789e2836f952315e91c4e06806fca3d04523a8a748bf9c32",
"src/inferrsi/data/skill_library/methods/kernel-writing-guide/references/akg-affine-tail-guards.md": "113deb2db10b171db3425bb3f1792bee763706b794c8cc2e222b4cc0d0114118",
"src/inferrsi/data/skill_library/methods/kernel-writing-guide/references/causal-conv1d-contract.md": "b3a6245cf5e9deeba0dffd4a15ddcf97da6a4315f4d3131921779438832663ad",
"src/inferrsi/data/skill_library/methods/kernel-writing-guide/references/hf-qwen35-gdn-binding.md": "848d3c99b47ef711e8820bbc4f6bcbd1db7211197811c9d855201025448d806b",
@@ -251,6 +251,37 @@
}
],
"empirical_validation": "not_run",
+ "applicability": {
+ "scope": "Deciding whether to write a custom inference kernel, routing a kernel pattern to Triton, CuTe DSL, CUDA C++, AscendC or TileIR, applying framework-independent correctness rules, and verifying the executing kernel against a reference before benchmarking; covers fused elementwise chains, reductions, tiled GEMM, attention variants, custom collectives, causal convolution and stateful serving-operator qualification",
+ "prerequisites": "The applicable operator, caller/state and layout contracts are read before substitution, including the causal-conv1d contract and HF Qwen3.5 GDN binding; the CuTe DSL environment uses Linux x86_64, Python 3.10–3.13 and Ampere+ (SM80+) with nvidia-cutlass-dsl installed and no NPU target; the CANN programming guide for the specific SoC is consulted for AscendC",
+ "input_contract": "The candidate operation or operation chain, its shapes, dtypes, quantization and layout requirements, and the target execution phase; the target accelerator and framework constraints; for serving operators, the concrete caller, effective phase and backend, state-row mapping, cache or buffer lifetime and in-place-write behavior; a reference implementation with the same signature, a callable kernel wrapper and fresh CUDA or NPU test tensors; verification shapes and dtypes plus profiler output for the target metric",
+ "produced_artifact": "A write-or-decline decision for the candidate operation with the profiled bottleneck, fusable pair or required dtype, shape or quantization variant unsupported by the standard implementation that justifies it, a framework choice routed by pattern and accelerator, a correctness checklist covering fp32 transcendental round-trips, sync-free wrappers, C division semantics, pre-allocated outputs, fp32 matmul accumulators and masked boundary loads and stores, a starting block or tile size with its autotune and warmup cost noted, a kernel shipping kernel_fn, reference_fn and get_inputs, a reference-parity result across the verification shapes and dtypes, and a post-verification profiler reading on the target metric; for an existing causal-convolution or stateful serving operator, a qualification record naming the caller, phase, backend, state rows and in-place writes, with source selection kept apart from executed dispatch; a routing and correctness decision, not a complete implementation or a measured speedup",
+ "failure_modes": [
+ "A kernel is written for a standalone elementwise op, a standalone matmul or standard attention where the vendor library already wins",
+ "Host-side scheduling cost is treated as a kernel problem instead of being routed to host-overhead-profiling",
+ "Fusion is justified by removed Python call overhead rather than by the DRAM traffic the shared data no longer makes",
+ "A transcendental runs on fp16/bf16 inputs without the fp32 cast round-trip, so the error stays silent through benchmarking",
+ "tl.dot accumulates in the input dtype because Triton is assumed to promote fp16/bf16 to fp32 automatically",
+ "Negative offsets are computed assuming Python floor division while Triton or AscendC rounds toward zero",
+ "An unmasked load or store is accepted because the benchmark shapes happen to divide the block size",
+ "A kernel is benchmarked before it matches its reference, or a shape-boundary, accumulator-dtype or precision failure is read as a tuning problem",
+ "A verification pass is reported as a win although the profiler metric did not move, so the kernel was never on the critical path",
+ "Autotune compile cost is measured as steady-state serving latency instead of being pre-warmed or cached",
+ "A CuTe DSL pattern is routed to an NPU target, or a CUDA tail-guard, attention primitive or SM-matched example is assumed to carry to Ascend",
+ "A causal-conv1d or GDN operator is substituted on build success, import success or matching names without resolving dtype, strides, packed boundaries, state rows and cursors",
+ "An in-tree vLLM wrapper, a GDN prefill flag or a cache-free classification adapter is read as establishing the executed convolution path or generation-state support"
+ ],
+ "refusal_point": "Declines to start a kernel until a profiler puts the target on the critical path near its hardware ceiling, a concrete fusable pair is identified, or the required dtype, shape or quantization variant is unsupported by the standard implementation, and sends host-bound cases elsewhere; refuses to benchmark an unverified kernel and withholds a gain whose profiler metric did not move; keeps operator source selection separate from executed dispatch, numerical parity and hardware qualification, so build or import success qualifies nothing; treats AscendC and TileIR content as source-inferred pending CANN and SoC verification with no local NPU run, treats the AKG Affine FileCheck expectations as IR structure rather than executed CUDA correctness, and reads author-reported overlap and FP8 cases as counterexamples to unconditional tuning rules; claims no Ascend support from NVIDIA primitives",
+ "hardware": "CUDA Ampere+ and Ascend NPU have separate framework routes for each kernel pattern; CuTe DSL requires Linux x86_64, Python 3.10-3.13 and SM80+ with nvidia-cutlass-dsl and is unavailable on NPU; profiling uses ncu on CUDA and msprof on Ascend; the AscendC and TileIR guidance is source-inferred with no local NPU verification and requires validation against the specific CANN version and SoC"
+ },
+ "recommendations": [
+ "Confirm with nsys/ncu or msprof that the target kernel sits on the critical path near its hardware ceiling, name the adjacent operations whose shared data would otherwise round-trip through DRAM, or identify a required dtype, shape or quantization variant unsupported by the standard implementation, before writing any kernel.",
+ "Ship every kernel with kernel_fn, reference_fn and get_inputs, pass reference parity across the verification shapes and dtypes first, and profile afterwards to confirm the target metric actually moved.",
+ "Cast fp16/bf16 to fp32 and back for any math beyond +, -, * and /, declare tl.dot accumulators as tl.float32, and mask every load and store whose shape does not divide the block size.",
+ "Start Triton at BLOCK_SIZE 1024 on Hopper and 512 on Ampere using powers of two, take GEMM tiles from the source skeleton configs (128/256/64 for Hopper, 64/128/32 for smaller cards) and autotune from there.",
+ "On Ascend, reach for TileIR before raw AscendC, size tiles to fit the AI Core Unified Buffer and check cube-unit alignment against the CANN guide for the specific SoC.",
+ "Match a CuTe DSL example to the GPU architecture (SM80 Ampere, SM90 Hopper, SM100 Blackwell) and study its structure instead of copying it, and pre-warm or cache autotuned kernels at startup so first-run compile cost stays out of serving latency."
+ ],
"limitations": [
"Source-derived candidate; no local efficacy validation.",
"Original typed source inventory is not normalized; retain source identities and limits in the procedure/readsets.",
@@ -270,10 +301,36 @@
],
"consumption_mode": "packaged"
},
- "applicability": {},
+ "applicability": {
+ "scope": "Deciding whether to write a custom inference kernel, routing a kernel pattern to Triton, CuTe DSL, CUDA C++, AscendC or TileIR, applying framework-independent correctness rules, and verifying the executing kernel against a reference before benchmarking; covers fused elementwise chains, reductions, tiled GEMM, attention variants, custom collectives, causal convolution and stateful serving-operator qualification",
+ "prerequisites": "The applicable operator, caller/state and layout contracts are read before substitution, including the causal-conv1d contract and HF Qwen3.5 GDN binding; the CuTe DSL environment uses Linux x86_64, Python 3.10–3.13 and Ampere+ (SM80+) with nvidia-cutlass-dsl installed and no NPU target; the CANN programming guide for the specific SoC is consulted for AscendC",
+ "input_contract": "The candidate operation or operation chain, its shapes, dtypes, quantization and layout requirements, and the target execution phase; the target accelerator and framework constraints; for serving operators, the concrete caller, effective phase and backend, state-row mapping, cache or buffer lifetime and in-place-write behavior; a reference implementation with the same signature, a callable kernel wrapper and fresh CUDA or NPU test tensors; verification shapes and dtypes plus profiler output for the target metric",
+ "produced_artifact": "A write-or-decline decision for the candidate operation with the profiled bottleneck, fusable pair or required dtype, shape or quantization variant unsupported by the standard implementation that justifies it, a framework choice routed by pattern and accelerator, a correctness checklist covering fp32 transcendental round-trips, sync-free wrappers, C division semantics, pre-allocated outputs, fp32 matmul accumulators and masked boundary loads and stores, a starting block or tile size with its autotune and warmup cost noted, a kernel shipping kernel_fn, reference_fn and get_inputs, a reference-parity result across the verification shapes and dtypes, and a post-verification profiler reading on the target metric; for an existing causal-convolution or stateful serving operator, a qualification record naming the caller, phase, backend, state rows and in-place writes, with source selection kept apart from executed dispatch; a routing and correctness decision, not a complete implementation or a measured speedup",
+ "failure_modes": [
+ "A kernel is written for a standalone elementwise op, a standalone matmul or standard attention where the vendor library already wins",
+ "Host-side scheduling cost is treated as a kernel problem instead of being routed to host-overhead-profiling",
+ "Fusion is justified by removed Python call overhead rather than by the DRAM traffic the shared data no longer makes",
+ "A transcendental runs on fp16/bf16 inputs without the fp32 cast round-trip, so the error stays silent through benchmarking",
+ "tl.dot accumulates in the input dtype because Triton is assumed to promote fp16/bf16 to fp32 automatically",
+ "Negative offsets are computed assuming Python floor division while Triton or AscendC rounds toward zero",
+ "An unmasked load or store is accepted because the benchmark shapes happen to divide the block size",
+ "A kernel is benchmarked before it matches its reference, or a shape-boundary, accumulator-dtype or precision failure is read as a tuning problem",
+ "A verification pass is reported as a win although the profiler metric did not move, so the kernel was never on the critical path",
+ "Autotune compile cost is measured as steady-state serving latency instead of being pre-warmed or cached",
+ "A CuTe DSL pattern is routed to an NPU target, or a CUDA tail-guard, attention primitive or SM-matched example is assumed to carry to Ascend",
+ "A causal-conv1d or GDN operator is substituted on build success, import success or matching names without resolving dtype, strides, packed boundaries, state rows and cursors",
+ "An in-tree vLLM wrapper, a GDN prefill flag or a cache-free classification adapter is read as establishing the executed convolution path or generation-state support"
+ ],
+ "refusal_point": "Declines to start a kernel until a profiler puts the target on the critical path near its hardware ceiling, a concrete fusable pair is identified, or the required dtype, shape or quantization variant is unsupported by the standard implementation, and sends host-bound cases elsewhere; refuses to benchmark an unverified kernel and withholds a gain whose profiler metric did not move; keeps operator source selection separate from executed dispatch, numerical parity and hardware qualification, so build or import success qualifies nothing; treats AscendC and TileIR content as source-inferred pending CANN and SoC verification with no local NPU run, treats the AKG Affine FileCheck expectations as IR structure rather than executed CUDA correctness, and reads author-reported overlap and FP8 cases as counterexamples to unconditional tuning rules; claims no Ascend support from NVIDIA primitives",
+ "hardware": "CUDA Ampere+ and Ascend NPU have separate framework routes for each kernel pattern; CuTe DSL requires Linux x86_64, Python 3.10-3.13 and SM80+ with nvidia-cutlass-dsl and is unavailable on NPU; profiling uses ncu on CUDA and msprof on Ascend; the AscendC and TileIR guidance is source-inferred with no local NPU verification and requires validation against the specific CANN version and SoC"
+ },
"recommendations": [
- "Read the pinned SKILL.md and its prerequisites before choosing this procedure.",
- "Read the packaged references and owner-local cases alongside the procedure."
+ "Confirm with nsys/ncu or msprof that the target kernel sits on the critical path near its hardware ceiling, name the adjacent operations whose shared data would otherwise round-trip through DRAM, or identify a required dtype, shape or quantization variant unsupported by the standard implementation, before writing any kernel.",
+ "Ship every kernel with kernel_fn, reference_fn and get_inputs, pass reference parity across the verification shapes and dtypes first, and profile afterwards to confirm the target metric actually moved.",
+ "Cast fp16/bf16 to fp32 and back for any math beyond +, -, * and /, declare tl.dot accumulators as tl.float32, and mask every load and store whose shape does not divide the block size.",
+ "Start Triton at BLOCK_SIZE 1024 on Hopper and 512 on Ampere using powers of two, take GEMM tiles from the source skeleton configs (128/256/64 for Hopper, 64/128/32 for smaller cards) and autotune from there.",
+ "On Ascend, reach for TileIR before raw AscendC, size tiles to fit the AI Core Unified Buffer and check cube-unit alignment against the CANN guide for the specific SoC.",
+ "Match a CuTe DSL example to the GPU architecture (SM80 Ampere, SM90 Hopper, SM100 Blackwell) and study its structure instead of copying it, and pre-warm or cache autotuned kernels at startup so first-run compile cost stays out of serving latency."
],
"limitations": [
"This generated abstract points to the complete packaged skill; external evidence and upstream repositories remain source links.",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/kv-tier-management.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/kv-tier-management.json
index a212c83c7..b0ccffcc9 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/kv-tier-management.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/kv-tier-management.json
@@ -24,8 +24,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/repositories/lmcache/kv-tier-management/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/repositories/lmcache/kv-tier-management/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/repositories/lmcache/kv-tier-management/SKILL.md",
"src/inferrsi/data/skill_library/repositories/lmcache/kv-tier-management/metadata.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/lazy-evaluation-stream-lifetimes.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/lazy-evaluation-stream-lifetimes.json
index 9e7dd43e3..1be4faa41 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/lazy-evaluation-stream-lifetimes.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/lazy-evaluation-stream-lifetimes.json
@@ -22,8 +22,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/frameworks/mlx/lazy-evaluation-stream-lifetimes/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/frameworks/mlx/lazy-evaluation-stream-lifetimes/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/frameworks/mlx/lazy-evaluation-stream-lifetimes/SKILL.md",
"src/inferrsi/data/skill_library/frameworks/mlx/lazy-evaluation-stream-lifetimes/metadata.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/lily.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/lily.json
index 2e7ef494b..d78d76fd7 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/lily.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/lily.json
@@ -25,8 +25,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/repositories/lily/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/repositories/lily/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/repositories/lily/LICENSE",
"src/inferrsi/data/skill_library/repositories/lily/NOTICE.md",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/lithos-metal.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/lithos-metal.json
index 3c48b0523..de4b4d96b 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/lithos-metal.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/lithos-metal.json
@@ -21,8 +21,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/repositories/lithos-metal/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/repositories/lithos-metal/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/repositories/lithos-metal/LICENSE",
"src/inferrsi/data/skill_library/repositories/lithos-metal/SKILL.md",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/local-dream.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/local-dream.json
index 52ab293b9..ea0873dab 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/local-dream.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/local-dream.json
@@ -29,8 +29,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/repositories/local-dream/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/repositories/local-dream/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/repositories/local-dream/SKILL.md",
"src/inferrsi/data/skill_library/repositories/local-dream/metadata.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/mistral-rs.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/mistral-rs.json
index 3db94e91f..1e265df15 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/mistral-rs.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/mistral-rs.json
@@ -26,8 +26,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/repositories/mistral-rs/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/repositories/mistral-rs/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/repositories/mistral-rs/LICENSE",
"src/inferrsi/data/skill_library/repositories/mistral-rs/SKILL.md",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/mlc-llm.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/mlc-llm.json
index 74fcb5bf7..2d666edfd 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/mlc-llm.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/mlc-llm.json
@@ -27,8 +27,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/repositories/mlc-llm/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/repositories/mlc-llm/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/repositories/mlc-llm/LICENSE",
"src/inferrsi/data/skill_library/repositories/mlc-llm/SKILL.md",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/mlx-swift-lm.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/mlx-swift-lm.json
index ac17206b0..7d0c483e1 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/mlx-swift-lm.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/mlx-swift-lm.json
@@ -27,8 +27,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/repositories/mlx-swift-lm/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/repositories/mlx-swift-lm/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/repositories/mlx-swift-lm/LICENSE",
"src/inferrsi/data/skill_library/repositories/mlx-swift-lm/SKILL.md",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/npu-skill-factory.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/npu-skill-factory.json
index 6d2ee6691..c9bfe7cc2 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/npu-skill-factory.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/npu-skill-factory.json
@@ -22,8 +22,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/workflows/npu-skill-factory/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/workflows/npu-skill-factory/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/skill-index.json",
"src/inferrsi/data/skill_library/workflows/npu-skill-factory/SKILL.md",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/optimization-search-policy.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/optimization-search-policy.json
index 23a054eb2..507ecadaf 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/optimization-search-policy.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/optimization-search-policy.json
@@ -21,8 +21,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/methods/optimization-search-policy/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/methods/optimization-search-policy/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/methods/optimization-search-policy/SKILL.md",
"src/inferrsi/data/skill_library/methods/optimization-search-policy/cases/sarathi-chunk-budget.json",
@@ -50,7 +50,7 @@
"src/inferrsi/data/skill_library/methods/optimization-search-policy/SKILL.md": "0ba4ade1648764a7422f564abe2d4fc9f9a25b5943a6c70c81ce17b8e7fb48d9",
"src/inferrsi/data/skill_library/methods/optimization-search-policy/cases/sarathi-chunk-budget.json": "f8448973cf1c9e506179c02c5e55bd804ea039b7276cf08071952f6c0be656ad",
"src/inferrsi/data/skill_library/methods/optimization-search-policy/cases/sarathi-chunk-budget.md": "2c2a8147001d24d1abc8672f15cf0ca759ae461b734413c8dcb7497634bb6821",
- "src/inferrsi/data/skill_library/methods/optimization-search-policy/metadata.json": "d5aaaa4511c9287e1f3d1f07a25c36320a68cb35a5066778cbd3c2bbb5203c6a",
+ "src/inferrsi/data/skill_library/methods/optimization-search-policy/metadata.json": "6dc0c3b8ce13cff285ad99da4441f2fb6a5931247e7634e657275dd915854e02",
"src/inferrsi/data/skill_library/methods/optimization-search-policy/references/distserve-phase-placement/README.md": "56e16efce447dfd4990a024ff1d6f7cb428f8e3e1f7b32a7e4d05f57139b836a",
"src/inferrsi/data/skill_library/methods/optimization-search-policy/references/distserve-phase-placement/source-record.json": "21cc9562f1c19d6b57957c01b53e0adf8700d05ffaaa7764afe07c4da1284de1",
"src/inferrsi/data/skill_library/methods/optimization-search-policy/references/papers-20261015/INTEGRATION_NOTES.md": "157e74ba180426b279b548a729769f152bf18f9a78ffca4ebfcaf26b7d0c5292",
@@ -78,25 +78,27 @@
"external"
],
"empirical_validation": "not_run",
- "sources": [
- {
- "type": "paper",
- "id": "usenix:osdi24:zhong-yinmin",
- "version": "OSDI 24 proceedings, July 2024, pages 193-210",
- "url": "https://www.usenix.org/conference/osdi24/presentation/zhong-yinmin",
- "sections": [
- "3.3",
- "4.1-4.2",
- "6.1-6.5",
- "7",
- "Appendices A-C"
- ],
- "readset": "references/distserve-phase-placement/source-record.json"
- },
- {
- "type": "unknown",
- "scope": "Pre-existing sources remain in their original procedure/readsets; this amendment does not normalize that inventory"
- }
+ "applicability": {
+ "scope": "Selecting, ordering and stopping optimization experiments: candidate generation, exploration versus exploitation, mechanism diversity, negative experiment memory, cost-aware escalation, budget allocation and stopping criteria, drawn from vllm-evolve, the TRT-LLM casebook and SkillOpt",
+ "prerequisites": "The useful work, quality contract, allowed approximations, timing boundary and resource budget are frozen before search, and a complete restartable incumbent is preserved; reading the procedures authorizes no execution, framework change, new resource spending or publication",
+ "input_contract": "Profiling data, existing config, prior experiment and negative records and the casebook for the current bottleneck class, model and hardware; each target stage's fraction of E2E time; the target, locally measured uncertainty and the authorized budget",
+ "produced_artifact": "An ordered experiment plan: one candidate change per round with an explicit hypothesis and its motivating evidence, a bounded shortlist of distinct mechanisms with assumptions, uncertainty, expected impact and evaluation cost, negative records turned into search constraints, and a stop-or-switch decision; when authorized rounds run, a recorded verdict per round with best-seen and delivered outcomes kept apart",
+ "failure_modes": [
+ "The evaluation protocol, measurement setup or baseline changes in the same round as the candidate",
+ "Benchmark budget is spent on a candidate that fails correctness",
+ "A matching keyword in a rejection record is used as a causal diagnosis or automatic veto, or one failed implementation excludes a whole solution family",
+ "Expensive work starts on a stage without computing its Amdahl contribution, so a large local speedup yields little E2E gain",
+ "One mechanism family keeps being sampled only because budget remains, without re-profiling after the limiting stage moves",
+ "Roofline proximity or a utilization threshold is used as a universal stopping rule",
+ "Screening is allowed to skip the frozen acceptance regimes or required service-level evidence"
+ ],
+ "refusal_point": "Discards a correctness-breaking, worse or inconclusive change instead of integrating it, preserves no-change when candidates cannot be separated, does not count a shortcut outside the authorized semantics as an optimization, and does not let a successful probe qualify a broken final state"
+ },
+ "recommendations": [
+ "Before generating a candidate, consult the casebook, the negative record and the target stage's Amdahl contribution.",
+ "Record every worse or inconclusive result with hypothesis, metric, context, evidence level and failure type, and derive a search constraint from it.",
+ "When about three experiments on one hypothesis return inconclusive from measurement variance, fix the protocol or move to a different hypothesis.",
+ "Before reporting, restore and verify the best fully valid incumbent through its canonical launcher."
],
"limitations": [
"Source-derived candidate; no local efficacy validation.",
@@ -129,12 +131,49 @@
"../performance-ceiling-analysis/SKILL.md",
"../../repositories/vllm/audit-vllm-kv-lifetimes/SKILL.md"
],
- "consumption_mode": "packaged"
+ "consumption_mode": "packaged",
+ "sources": [
+ {
+ "type": "paper",
+ "id": "usenix:osdi24:zhong-yinmin",
+ "version": "OSDI 24 proceedings, July 2024, pages 193-210",
+ "url": "https://www.usenix.org/conference/osdi24/presentation/zhong-yinmin",
+ "sections": [
+ "3.3",
+ "4.1-4.2",
+ "6.1-6.5",
+ "7",
+ "Appendices A-C"
+ ],
+ "readset": "references/distserve-phase-placement/source-record.json"
+ },
+ {
+ "type": "unknown",
+ "scope": "Pre-existing sources remain in their original procedure/readsets; this amendment does not normalize that inventory"
+ }
+ ]
+ },
+ "applicability": {
+ "scope": "Selecting, ordering and stopping optimization experiments: candidate generation, exploration versus exploitation, mechanism diversity, negative experiment memory, cost-aware escalation, budget allocation and stopping criteria, drawn from vllm-evolve, the TRT-LLM casebook and SkillOpt",
+ "prerequisites": "The useful work, quality contract, allowed approximations, timing boundary and resource budget are frozen before search, and a complete restartable incumbent is preserved; reading the procedures authorizes no execution, framework change, new resource spending or publication",
+ "input_contract": "Profiling data, existing config, prior experiment and negative records and the casebook for the current bottleneck class, model and hardware; each target stage's fraction of E2E time; the target, locally measured uncertainty and the authorized budget",
+ "produced_artifact": "An ordered experiment plan: one candidate change per round with an explicit hypothesis and its motivating evidence, a bounded shortlist of distinct mechanisms with assumptions, uncertainty, expected impact and evaluation cost, negative records turned into search constraints, and a stop-or-switch decision; when authorized rounds run, a recorded verdict per round with best-seen and delivered outcomes kept apart",
+ "failure_modes": [
+ "The evaluation protocol, measurement setup or baseline changes in the same round as the candidate",
+ "Benchmark budget is spent on a candidate that fails correctness",
+ "A matching keyword in a rejection record is used as a causal diagnosis or automatic veto, or one failed implementation excludes a whole solution family",
+ "Expensive work starts on a stage without computing its Amdahl contribution, so a large local speedup yields little E2E gain",
+ "One mechanism family keeps being sampled only because budget remains, without re-profiling after the limiting stage moves",
+ "Roofline proximity or a utilization threshold is used as a universal stopping rule",
+ "Screening is allowed to skip the frozen acceptance regimes or required service-level evidence"
+ ],
+ "refusal_point": "Discards a correctness-breaking, worse or inconclusive change instead of integrating it, preserves no-change when candidates cannot be separated, does not count a shortcut outside the authorized semantics as an optimization, and does not let a successful probe qualify a broken final state"
},
- "applicability": {},
"recommendations": [
- "Read the pinned SKILL.md and its prerequisites before choosing this procedure.",
- "Read the packaged references and owner-local cases alongside the procedure."
+ "Before generating a candidate, consult the casebook, the negative record and the target stage's Amdahl contribution.",
+ "Record every worse or inconclusive result with hypothesis, metric, context, evidence level and failure type, and derive a search constraint from it.",
+ "When about three experiments on one hypothesis return inconclusive from measurement variance, fix the protocol or move to a different hypothesis.",
+ "Before reporting, restore and verify the best fully valid incumbent through its canonical launcher."
],
"limitations": [
"This generated abstract points to the complete packaged skill; external evidence and upstream repositories remain source links.",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/os-process-placement.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/os-process-placement.json
index f446824d3..e491a895a 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/os-process-placement.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/os-process-placement.json
@@ -24,8 +24,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/hardware/host/os-process-placement/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/hardware/host/os-process-placement/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/hardware/host/os-process-placement/SKILL.md",
"src/inferrsi/data/skill_library/hardware/host/os-process-placement/metadata.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/performance-ceiling-analysis.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/performance-ceiling-analysis.json
index a349b7279..f45160e4c 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/performance-ceiling-analysis.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/performance-ceiling-analysis.json
@@ -22,8 +22,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/methods/performance-ceiling-analysis/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/methods/performance-ceiling-analysis/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/methods/performance-ceiling-analysis/SKILL.md",
"src/inferrsi/data/skill_library/methods/performance-ceiling-analysis/metadata.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/plan-activation-residency.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/plan-activation-residency.json
index 86f1cb7ea..1c000bc57 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/plan-activation-residency.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/plan-activation-residency.json
@@ -16,8 +16,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/methods/plan-activation-residency/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/methods/plan-activation-residency/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/methods/plan-activation-residency/SKILL.md",
"src/inferrsi/data/skill_library/methods/plan-activation-residency/metadata.json",
@@ -34,7 +34,7 @@
"checkout_ref": "checkout:plan-activation-residency",
"source_files": {
"src/inferrsi/data/skill_library/methods/plan-activation-residency/SKILL.md": "542d8a1e6963f5019986e1fbd79d3a247a045a51487c09313026087ac457c15e",
- "src/inferrsi/data/skill_library/methods/plan-activation-residency/metadata.json": "c6acae57d70190cea0e18c71c79127962181e58169e011df75d8635b3db4daa7",
+ "src/inferrsi/data/skill_library/methods/plan-activation-residency/metadata.json": "0289c62f93d087cd6a4ba4140487925c4b3258a9db1f7004ff1dfb22fd6d4582",
"src/inferrsi/data/skill_library/methods/plan-activation-residency/references/source-notes.md": "5ff52655cfd06fcf970ed2a7d4d4fe1c31115b102af2d41497de0c74848ed2b3",
"src/inferrsi/data/skill_library/skill-index.json": "b5a9f05bd8179644f3fea79d7eb04fe4a8baad9e8d68aa8d0f1db213cf633317"
},
@@ -49,6 +49,27 @@
"origin": [
"external"
],
+ "applicability": {
+ "scope": "Proposing training activation and gradient memory reductions when pressure appears to come from live tensors or allocator-reserved memory, by distinguishing lifetime reordering, address packing and selective rematerialization; covers activation checkpointing for recomputation, not persistent recovery checkpoints",
+ "prerequisites": "The packaged evidence and limits are read before citing paper results; installing or running a solver or framework, modifying allocation or checkpoint policies, reproducing workloads or provisioning hardware needs separate authorization",
+ "input_contract": "The actual forward/backward/update DAG, shapes and precision, tensor sizes and consumer lifetimes; resident versus reserved peak bytes; operator, workspace and temporary-output peaks; batch and accumulation policy; optimizer and collective barriers; representative compute and transfer costs; whether the goal is fit, step time or time to a fixed training objective",
+ "produced_artifact": "A preferred candidate, simpler fallback, resource ledger, critical-path prediction, assumptions, counter-signals and unresolved-evidence list, with source claims kept apart from proposed decisions, plus a proposed matched comparison; a reviewable proposal, not an implementation",
+ "failure_modes": [
+ "Fragmentation and live-byte excess are not distinguished at the same phase",
+ "OLLA address packing is read as tensor placement across CPU/SSD tiers",
+ "Reordering moves a release or update past a dependency such as shared weights, accumulation, gradient clipping, optimizer state or collectives",
+ "A static packing plan ignores alignment, aliasing, workspace, asynchronous overlap or shape variability",
+ "Added recomputation consumes the compute window that previously hid transfers",
+ "Checkmate's rounding is taken to guarantee the memory cap, or its 10% allowance is used as a portable safety margin",
+ "A solver timeout is read as proof of infeasibility"
+ ],
+ "refusal_point": "Returns the missing classification instead of a preferred checkpoint interval when only total OOM memory is known, and escalates when a legal dependency graph, memory margin or correctness model is unavailable"
+ },
+ "recommendations": [
+ "Where ordering is legal, try earlier last-use and gradient release before adding computation or transfer.",
+ "Check dependency validity and memory feasibility separately for an approximate schedule.",
+ "Account for planning cost over the expected number of repeated steps."
+ ],
"limitations": [
"Source-derived candidate; no local efficacy validation.",
"Original typed source inventory is not normalized; retain source identities and limits in the procedure/readsets."
@@ -58,10 +79,26 @@
],
"consumption_mode": "packaged"
},
- "applicability": {},
+ "applicability": {
+ "scope": "Proposing training activation and gradient memory reductions when pressure appears to come from live tensors or allocator-reserved memory, by distinguishing lifetime reordering, address packing and selective rematerialization; covers activation checkpointing for recomputation, not persistent recovery checkpoints",
+ "prerequisites": "The packaged evidence and limits are read before citing paper results; installing or running a solver or framework, modifying allocation or checkpoint policies, reproducing workloads or provisioning hardware needs separate authorization",
+ "input_contract": "The actual forward/backward/update DAG, shapes and precision, tensor sizes and consumer lifetimes; resident versus reserved peak bytes; operator, workspace and temporary-output peaks; batch and accumulation policy; optimizer and collective barriers; representative compute and transfer costs; whether the goal is fit, step time or time to a fixed training objective",
+ "produced_artifact": "A preferred candidate, simpler fallback, resource ledger, critical-path prediction, assumptions, counter-signals and unresolved-evidence list, with source claims kept apart from proposed decisions, plus a proposed matched comparison; a reviewable proposal, not an implementation",
+ "failure_modes": [
+ "Fragmentation and live-byte excess are not distinguished at the same phase",
+ "OLLA address packing is read as tensor placement across CPU/SSD tiers",
+ "Reordering moves a release or update past a dependency such as shared weights, accumulation, gradient clipping, optimizer state or collectives",
+ "A static packing plan ignores alignment, aliasing, workspace, asynchronous overlap or shape variability",
+ "Added recomputation consumes the compute window that previously hid transfers",
+ "Checkmate's rounding is taken to guarantee the memory cap, or its 10% allowance is used as a portable safety margin",
+ "A solver timeout is read as proof of infeasibility"
+ ],
+ "refusal_point": "Returns the missing classification instead of a preferred checkpoint interval when only total OOM memory is known, and escalates when a legal dependency graph, memory margin or correctness model is unavailable"
+ },
"recommendations": [
- "Read the pinned SKILL.md and its prerequisites before choosing this procedure.",
- "Read the packaged references and owner-local cases alongside the procedure."
+ "Where ordering is legal, try earlier last-use and gradient release before adding computation or transfer.",
+ "Check dependency validity and memory feasibility separately for an approximate schedule.",
+ "Account for planning cost over the expected number of repeated steps."
],
"limitations": [
"This generated abstract points to the complete packaged skill; external evidence and upstream repositories remain source links.",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/project-maya.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/project-maya.json
index 5e09a5383..5345db2b5 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/project-maya.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/project-maya.json
@@ -25,8 +25,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/repositories/project-maya/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/repositories/project-maya/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/repositories/project-maya/LICENSE",
"src/inferrsi/data/skill_library/repositories/project-maya/SKILL.md",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/repository-evolution-analysis.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/repository-evolution-analysis.json
index f083d22ba..4ddf37b27 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/repository-evolution-analysis.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/repository-evolution-analysis.json
@@ -21,8 +21,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/methods/repository-evolution-analysis/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/methods/repository-evolution-analysis/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/methods/repository-evolution-analysis/SKILL.md",
"src/inferrsi/data/skill_library/methods/repository-evolution-analysis/metadata.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/rsi-evidence.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/rsi-evidence.json
index 420673822..b1c230d1f 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/rsi-evidence.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/rsi-evidence.json
@@ -22,8 +22,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/workflows/rsi-evidence/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/workflows/rsi-evidence/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/skill-index.json",
"src/inferrsi/data/skill_library/workflows/rsi-evidence/SKILL.md",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/skill-evolution-playbook.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/skill-evolution-playbook.json
index a4ce44c6e..fcc3f3c47 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/skill-evolution-playbook.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/skill-evolution-playbook.json
@@ -21,8 +21,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/methods/skill-evolution-playbook/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/methods/skill-evolution-playbook/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/methods/skill-evolution-playbook/SKILL.md",
"src/inferrsi/data/skill_library/methods/skill-evolution-playbook/metadata.json",
@@ -38,7 +38,7 @@
"checkout_ref": "checkout:skill-evolution-playbook",
"source_files": {
"src/inferrsi/data/skill_library/methods/skill-evolution-playbook/SKILL.md": "9cd37f21fe0d862069e4798edca8a04d59f8b840e55d521901594a63dddd9ba3",
- "src/inferrsi/data/skill_library/methods/skill-evolution-playbook/metadata.json": "92b36338b9257e2e41b95e4370694d7de52ed84bc7971056b3bfced07b0993d8",
+ "src/inferrsi/data/skill_library/methods/skill-evolution-playbook/metadata.json": "d3e790614e09d2d57acb24bfe8d2c2a524cc36bd667bd3867c557554bcbb605e",
"src/inferrsi/data/skill_library/skill-index.json": "b5a9f05bd8179644f3fea79d7eb04fe4a8baad9e8d68aa8d0f1db213cf633317"
},
"skill_metadata": {
@@ -58,6 +58,28 @@
"external"
],
"empirical_validation": "not_run",
+ "applicability": {
+ "scope": "Design choices for evolving agent skills from execution traces, drawn from Trace2Skill and SkillOpt: when a trajectory should update a skill, how conflicting lessons are handled, how negative evidence is preserved, how large an edit should be and when skills split or merge; for designing or operating a skill evolution pipeline",
+ "prerequisites": "A final-test partition kept outside edit selection and the incumbent skill retained for rollback; changing a skill's scope first needs explicit review of its identity and consumers",
+ "input_contract": "Execution trajectories or trace IDs with the diagnosed causal failure and the successful behavior an amendment must preserve; the incumbent skill and its revision lineage; held-out evaluation results for candidate edits",
+ "produced_artifact": "A scoped case or amendment proposal, or a consolidated candidate edit, with rejection records and the versioned revision lineage; a candidate for independently evaluated acceptance, not a replacement decision or an efficacy claim",
+ "failure_modes": [
+ "An anecdote, or agreement among analysts, is taken as efficacy",
+ "Consolidation output is treated as a replacement decision instead of a candidate awaiting independent acceptance",
+ "A tie or uncertain result is reported as improvement",
+ "A conflict is resolved by voting for whichever outcome appeared more often in the pool",
+ "A rejection is applied to every future use of an idea instead of the comparison it constrains",
+ "Paper-specific section counts or length budgets are used as universal defaults, or a skill is split on length alone",
+ "A stronger task solver is assumed to be a stronger improver, or replay rankings and private inner scores are read as outer generalization"
+ ],
+ "refusal_point": "Records a conflict that conditioning cannot resolve, where the same context produces different outcomes across runs, as INCONCLUSIVE and flags it for human review instead of choosing a lesson; claims no universal minimum trace count, paper-to-code equivalence or local efficacy"
+ },
+ "recommendations": [
+ "Condition a conflicting lesson on its context (when X, do A; when Y, avoid A) before choosing one side.",
+ "Store each rejection record with its evaluation evidence, outside the deployed task procedure, and give it a retirement condition.",
+ "Bound the target passage and intended behavior before editing, inspect the actual diff and preserve rollback state.",
+ "Prefer an amendment when the existing skill owns the workflow; create a new candidate only for a genuinely distinct capability."
+ ],
"limitations": [
"Source-derived candidate; no local efficacy validation.",
"Original typed source inventory is not normalized; retain source identities and limits in the procedure/readsets."
@@ -65,10 +87,27 @@
"provenance_readsets": [],
"consumption_mode": "packaged"
},
- "applicability": {},
+ "applicability": {
+ "scope": "Design choices for evolving agent skills from execution traces, drawn from Trace2Skill and SkillOpt: when a trajectory should update a skill, how conflicting lessons are handled, how negative evidence is preserved, how large an edit should be and when skills split or merge; for designing or operating a skill evolution pipeline",
+ "prerequisites": "A final-test partition kept outside edit selection and the incumbent skill retained for rollback; changing a skill's scope first needs explicit review of its identity and consumers",
+ "input_contract": "Execution trajectories or trace IDs with the diagnosed causal failure and the successful behavior an amendment must preserve; the incumbent skill and its revision lineage; held-out evaluation results for candidate edits",
+ "produced_artifact": "A scoped case or amendment proposal, or a consolidated candidate edit, with rejection records and the versioned revision lineage; a candidate for independently evaluated acceptance, not a replacement decision or an efficacy claim",
+ "failure_modes": [
+ "An anecdote, or agreement among analysts, is taken as efficacy",
+ "Consolidation output is treated as a replacement decision instead of a candidate awaiting independent acceptance",
+ "A tie or uncertain result is reported as improvement",
+ "A conflict is resolved by voting for whichever outcome appeared more often in the pool",
+ "A rejection is applied to every future use of an idea instead of the comparison it constrains",
+ "Paper-specific section counts or length budgets are used as universal defaults, or a skill is split on length alone",
+ "A stronger task solver is assumed to be a stronger improver, or replay rankings and private inner scores are read as outer generalization"
+ ],
+ "refusal_point": "Records a conflict that conditioning cannot resolve, where the same context produces different outcomes across runs, as INCONCLUSIVE and flags it for human review instead of choosing a lesson; claims no universal minimum trace count, paper-to-code equivalence or local efficacy"
+ },
"recommendations": [
- "Read the pinned SKILL.md and its prerequisites before choosing this procedure.",
- "Read the packaged references and owner-local cases alongside the procedure."
+ "Condition a conflicting lesson on its context (when X, do A; when Y, avoid A) before choosing one side.",
+ "Store each rejection record with its evaluation evidence, outside the deployed task procedure, and give it a retirement condition.",
+ "Bound the target passage and intended behavior before editing, inspect the actual diff and preserve rollback state.",
+ "Prefer an amendment when the existing skill owns the workflow; create a new candidate only for a genuinely distinct capability."
],
"limitations": [
"This generated abstract points to the complete packaged skill; external evidence and upstream repositories remain source links.",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/splash.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/splash.json
index 58cdafbfc..f151ead72 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/splash.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/splash.json
@@ -25,8 +25,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/repositories/splash/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/repositories/splash/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/repositories/splash/LICENSE",
"src/inferrsi/data/skill_library/repositories/splash/SKILL.md",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/storage-io-measurement.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/storage-io-measurement.json
index 2e8ddd500..bcefba3e6 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/storage-io-measurement.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/storage-io-measurement.json
@@ -24,8 +24,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/hardware/host/storage-io-measurement/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/hardware/host/storage-io-measurement/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/hardware/host/storage-io-measurement/SKILL.md",
"src/inferrsi/data/skill_library/hardware/host/storage-io-measurement/metadata.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/training-input-pipeline.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/training-input-pipeline.json
index 7f72fa1f8..0227a13cf 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/training-input-pipeline.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/training-input-pipeline.json
@@ -23,8 +23,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/methods/training-input-pipeline/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/methods/training-input-pipeline/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/methods/training-input-pipeline/SKILL.md",
"src/inferrsi/data/skill_library/methods/training-input-pipeline/metadata.json",
@@ -41,7 +41,7 @@
"checkout_ref": "checkout:training-input-pipeline",
"source_files": {
"src/inferrsi/data/skill_library/methods/training-input-pipeline/SKILL.md": "64bd2c6b795d966045d1d3cca55976d3d49b08b06df7c7e06a70c2824cf55b7e",
- "src/inferrsi/data/skill_library/methods/training-input-pipeline/metadata.json": "80659099fce6a6ec05e9084b75452c877a5812e30f60a78f31ef8c7efcdc1bac",
+ "src/inferrsi/data/skill_library/methods/training-input-pipeline/metadata.json": "debad47896f600c7a1e19e477e34026c975dfde320a786dca1ddd02030a5feb1",
"src/inferrsi/data/skill_library/methods/training-input-pipeline/references/sources.md": "60e0f3e88655d8e1d5eb4797f39ca4370deda01614f265b8d2ea4559d8ac020f",
"src/inferrsi/data/skill_library/skill-index.json": "b5a9f05bd8179644f3fea79d7eb04fe4a8baad9e8d68aa8d0f1db213cf633317"
},
@@ -64,6 +64,28 @@
"external"
],
"empirical_validation": "not_run",
+ "applicability": {
+ "scope": "Diagnosing training input stalls and choosing bounded queue, shard-cache and cache-ownership changes from DALI, Mosaic Streaming and FFCV, while preserving execution-mode gates, sample ordering and per-host memory budgets and separating remote downloads, local storage reads and host/device transfers",
+ "prerequisites": "The pinned readset and its license constraints are read; reading authorizes no commands, cache manipulation or experiment",
+ "input_contract": "Dataset format, sample-size distribution, workers and ranks per host, ordering/shuffle, local-cache state and executor mode; a per-host memory budget that includes checkpoint staging and runtime offload; timings that separate remote fetch, decompression/decoding, transformation, batch readiness and device transfer",
+ "produced_artifact": "One proposed queue, shard-cache or cache-ownership change with its source-backed rule and boundary, the evidence proposed to confirm it, and a verification contract for a separately authorized comparison; no input-throughput, model-quality or hardware result",
+ "failure_modes": [
+ "Low accelerator utilization is taken to identify which producer stage is late",
+ "Missing legacy DALI statistics are read as zero use, or the execution mode is switched silently to obtain them, which changes the comparison",
+ "DALI run() returning is read as output-ready timing instead of the outputs() synchronization boundary",
+ "Mosaic predownload is treated as a byte cap or request count, or per-worker sample targets are multiplied into an exact network-byte prediction",
+ "Look-ahead is increased when the local cache cannot retain the working set, so evicted shards are downloaded again",
+ "cache_limit is read as a bound on total process RAM, or the constructor minimum as an optimum or a guarantee against thrashing",
+ "An FFCV memory map is taken as proof that its pages are resident"
+ ],
+ "refusal_point": "Rejects changes that violate sample coverage or the stated reproducibility contract even if batches arrive faster, does not report queue depth alone as a performance result, and draws no NPU acceleration claim from the CUDA-oriented DALI source"
+ },
+ "recommendations": [
+ "Record dataset format, sample sizes, workers per host, ordering, cache state and executor mode before proposing one change.",
+ "Increase Mosaic look-ahead only when readiness starvation is observed and cache headroom exists; otherwise address shard and working-set residency first.",
+ "Check the exact DALI configuration surface before use, because the constructor and deserialization paths use different queue keys.",
+ "In a later comparison, separate warm-cache throughput from cold retrieval and initialization, and report step time with input wait, host/HBM peaks and actual I/O."
+ ],
"limitations": [
"Source-derived candidate; no local efficacy validation.",
"Original typed source inventory is not normalized; retain source identities and limits in the procedure/readsets."
@@ -73,10 +95,27 @@
],
"consumption_mode": "packaged"
},
- "applicability": {},
+ "applicability": {
+ "scope": "Diagnosing training input stalls and choosing bounded queue, shard-cache and cache-ownership changes from DALI, Mosaic Streaming and FFCV, while preserving execution-mode gates, sample ordering and per-host memory budgets and separating remote downloads, local storage reads and host/device transfers",
+ "prerequisites": "The pinned readset and its license constraints are read; reading authorizes no commands, cache manipulation or experiment",
+ "input_contract": "Dataset format, sample-size distribution, workers and ranks per host, ordering/shuffle, local-cache state and executor mode; a per-host memory budget that includes checkpoint staging and runtime offload; timings that separate remote fetch, decompression/decoding, transformation, batch readiness and device transfer",
+ "produced_artifact": "One proposed queue, shard-cache or cache-ownership change with its source-backed rule and boundary, the evidence proposed to confirm it, and a verification contract for a separately authorized comparison; no input-throughput, model-quality or hardware result",
+ "failure_modes": [
+ "Low accelerator utilization is taken to identify which producer stage is late",
+ "Missing legacy DALI statistics are read as zero use, or the execution mode is switched silently to obtain them, which changes the comparison",
+ "DALI run() returning is read as output-ready timing instead of the outputs() synchronization boundary",
+ "Mosaic predownload is treated as a byte cap or request count, or per-worker sample targets are multiplied into an exact network-byte prediction",
+ "Look-ahead is increased when the local cache cannot retain the working set, so evicted shards are downloaded again",
+ "cache_limit is read as a bound on total process RAM, or the constructor minimum as an optimum or a guarantee against thrashing",
+ "An FFCV memory map is taken as proof that its pages are resident"
+ ],
+ "refusal_point": "Rejects changes that violate sample coverage or the stated reproducibility contract even if batches arrive faster, does not report queue depth alone as a performance result, and draws no NPU acceleration claim from the CUDA-oriented DALI source"
+ },
"recommendations": [
- "Read the pinned SKILL.md and its prerequisites before choosing this procedure.",
- "Read the packaged references and owner-local cases alongside the procedure."
+ "Record dataset format, sample sizes, workers per host, ordering, cache state and executor mode before proposing one change.",
+ "Increase Mosaic look-ahead only when readiness starvation is observed and cache headroom exists; otherwise address shard and working-set residency first.",
+ "Check the exact DALI configuration surface before use, because the constructor and deserialization paths use different queue keys.",
+ "In a later comparison, separate warm-cache throughput from cold retrieval and initialization, and report step time with input wait, host/HBM peaks and actual I/O."
],
"limitations": [
"This generated abstract points to the complete packaged skill; external evidence and upstream repositories remain source links.",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/training-resource-budget.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/training-resource-budget.json
index 63dab161e..664e488d1 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/training-resource-budget.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/training-resource-budget.json
@@ -23,8 +23,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/methods/training-resource-budget/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/methods/training-resource-budget/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/methods/training-resource-budget/SKILL.md",
"src/inferrsi/data/skill_library/methods/training-resource-budget/metadata.json",
@@ -49,7 +49,7 @@
"checkout_ref": "checkout:training-resource-budget",
"source_files": {
"src/inferrsi/data/skill_library/methods/training-resource-budget/SKILL.md": "d3805c5f605faae663ad5a674a56c1e1e5fdd211cd58833c008e557cfd7e4c37",
- "src/inferrsi/data/skill_library/methods/training-resource-budget/metadata.json": "adb0473945920794e07f62e19230683e6f8b2a0f07f91d41a451374f919d343b",
+ "src/inferrsi/data/skill_library/methods/training-resource-budget/metadata.json": "9ae014b0ca5ca5d32e31e05b52e7436ad1c197fdce8d731fd4d9c79965fd4f2d",
"src/inferrsi/data/skill_library/methods/training-resource-budget/references/mindspeed-historical.md": "ad65c27f278c3d0ca3599a7630c02d3e529477a05b71b7703bd2d75f0d8fdd2c",
"src/inferrsi/data/skill_library/methods/training-resource-budget/references/mindspeed-source-readset.md": "4462d69ed6a941a03639c756968be3ce54f947e51f6ca618f4fb1cd43b643b35",
"src/inferrsi/data/skill_library/methods/training-resource-budget/references/papers-20261015/INTEGRATION_NOTES.md": "c55c6ad2131ef788866411dcc7056f221420ebff30c95c6bc919b09a4818cb73",
@@ -80,6 +80,28 @@
"external"
],
"empirical_validation": "not_run",
+ "applicability": {
+ "scope": "Choosing training offload and resource-budget changes from pinned DeepSpeed configuration, attributed expert diagnosis and a historical MindSpeed case, balancing accelerator capacity, CPU work, pinned RAM and local storage while keeping host/device transfer exposure distinct from collective network bottlenecks",
+ "prerequisites": "The source boundaries are read; garbage-collection or affinity changes, privileged memlock changes and delayed parameter updates each need separate approval, and the skill authorizes no newest-wheel installation, scheduler-priority change, benchmark or failure injection; historical MindSpeed reuse requires compatible allocator symbols, tracked swappable storage, sufficient host capacity and verified device/NUMA/CPU mapping, not merely the presence of npu-smi and lspci",
+ "input_contract": "State attributed to parameters, optimizer state, gradients or activations with its current memory tier, lifetime and next consumer; whether the failure is allocation capacity or throughput; a per-host budget for input buffers, checkpoint snapshots, offloaded state, OS cache and CPU workers with the actual ranks sharing the host; the host/device link recorded apart from collective traffic; the exact source version and enabled execution mode",
+ "produced_artifact": "A candidate decision record per proposal: the observed limiting interval and owning state or resource, source version and execution mode, capacity released and resources consumed with units and uncertainty, overlap readiness and mutation boundaries with expected contention, and the fixed-workload step and correctness evidence required for later validation",
+ "failure_modes": [
+ "A change that lets the model fit is presented as a speedup, although it can slow each step",
+ "A tensor copied to CPU is taken as an InfiniBand/RoCE bottleneck, or CPU/NVMe offload is counted as network evidence",
+ "Parameter max_in_cpu is read as a byte or total-host-memory cap instead of an element count",
+ "NVMe optimizer placement is assumed to remove CPU optimizer demand",
+ "The parameter buffer-pool count is transplanted to the optimizer, or all optimizers are assumed to hold identical states",
+ "Configuration presence or schema documentation is taken as evidence that a knob or overlap actually ran",
+ "Ordinary overlap is taken as permission for delayed, stale parameter updates"
+ ],
+ "refusal_point": "Claims no hardware validation, measured gain or current NPU compatibility, keeps source support and extracted coverage separate from measured improvement, and does not reuse the MindSpeed headline gain; historical source observations do not establish current-stack compatibility, whole-model correctness or measured benefit"
+ },
+ "recommendations": [
+ "Attribute the state and separate a capacity failure from a throughput bottleneck before choosing an offload surface.",
+ "Verify the consumer of buffer_size before calculating bytes or tuning it, and measure the executing path before claiming a knob ran.",
+ "Compare the whole step's critical path, not host capacity alone.",
+ "Use plan-activation-residency, choose-parallel-training-plan and assess-split-backward-schedule for live-range, parallel-plan and backward-ordering questions."
+ ],
"limitations": [
"Source-derived candidate; no local efficacy validation.",
"Original typed source inventory is not normalized; retain source identities and limits in the procedure/readsets."
@@ -97,10 +119,27 @@
],
"consumption_mode": "packaged"
},
- "applicability": {},
+ "applicability": {
+ "scope": "Choosing training offload and resource-budget changes from pinned DeepSpeed configuration, attributed expert diagnosis and a historical MindSpeed case, balancing accelerator capacity, CPU work, pinned RAM and local storage while keeping host/device transfer exposure distinct from collective network bottlenecks",
+ "prerequisites": "The source boundaries are read; garbage-collection or affinity changes, privileged memlock changes and delayed parameter updates each need separate approval, and the skill authorizes no newest-wheel installation, scheduler-priority change, benchmark or failure injection; historical MindSpeed reuse requires compatible allocator symbols, tracked swappable storage, sufficient host capacity and verified device/NUMA/CPU mapping, not merely the presence of npu-smi and lspci",
+ "input_contract": "State attributed to parameters, optimizer state, gradients or activations with its current memory tier, lifetime and next consumer; whether the failure is allocation capacity or throughput; a per-host budget for input buffers, checkpoint snapshots, offloaded state, OS cache and CPU workers with the actual ranks sharing the host; the host/device link recorded apart from collective traffic; the exact source version and enabled execution mode",
+ "produced_artifact": "A candidate decision record per proposal: the observed limiting interval and owning state or resource, source version and execution mode, capacity released and resources consumed with units and uncertainty, overlap readiness and mutation boundaries with expected contention, and the fixed-workload step and correctness evidence required for later validation",
+ "failure_modes": [
+ "A change that lets the model fit is presented as a speedup, although it can slow each step",
+ "A tensor copied to CPU is taken as an InfiniBand/RoCE bottleneck, or CPU/NVMe offload is counted as network evidence",
+ "Parameter max_in_cpu is read as a byte or total-host-memory cap instead of an element count",
+ "NVMe optimizer placement is assumed to remove CPU optimizer demand",
+ "The parameter buffer-pool count is transplanted to the optimizer, or all optimizers are assumed to hold identical states",
+ "Configuration presence or schema documentation is taken as evidence that a knob or overlap actually ran",
+ "Ordinary overlap is taken as permission for delayed, stale parameter updates"
+ ],
+ "refusal_point": "Claims no hardware validation, measured gain or current NPU compatibility, keeps source support and extracted coverage separate from measured improvement, and does not reuse the MindSpeed headline gain; historical source observations do not establish current-stack compatibility, whole-model correctness or measured benefit"
+ },
"recommendations": [
- "Read the pinned SKILL.md and its prerequisites before choosing this procedure.",
- "Read the packaged references and owner-local cases alongside the procedure."
+ "Attribute the state and separate a capacity failure from a throughput bottleneck before choosing an offload surface.",
+ "Verify the consumer of buffer_size before calculating bytes or tuning it, and measure the executing path before claiming a knob ran.",
+ "Compare the whole step's critical path, not host capacity alone.",
+ "Use plan-activation-residency, choose-parallel-training-plan and assess-split-backward-schedule for live-range, parallel-plan and backward-ordering questions."
],
"limitations": [
"This generated abstract points to the complete packaged skill; external evidence and upstream repositories remain source links.",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/video-diffusion-optimization.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/video-diffusion-optimization.json
index 35fae44a7..fe864acfe 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/video-diffusion-optimization.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/video-diffusion-optimization.json
@@ -26,8 +26,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/methods/video-diffusion-optimization/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/methods/video-diffusion-optimization/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/methods/video-diffusion-optimization/SKILL.md",
"src/inferrsi/data/skill_library/methods/video-diffusion-optimization/assets/campaign.template.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/vllm-metal.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/vllm-metal.json
index ffeb4ece0..97c7fcb3d 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/vllm-metal.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/vllm-metal.json
@@ -26,8 +26,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/repositories/vllm-metal/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/repositories/vllm-metal/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/repositories/vllm-metal/LICENSE",
"src/inferrsi/data/skill_library/repositories/vllm-metal/SKILL.md",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/weight-loading-optimization.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/weight-loading-optimization.json
index 8897bac41..f85a58e45 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/weight-loading-optimization.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/common/skills/weight-loading-optimization.json
@@ -29,8 +29,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/methods/weight-loading-optimization/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/methods/weight-loading-optimization/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/methods/weight-loading-optimization/SKILL.md",
"src/inferrsi/data/skill_library/methods/weight-loading-optimization/metadata.json",
@@ -49,7 +49,7 @@
"checkout_ref": "checkout:weight-loading-optimization",
"source_files": {
"src/inferrsi/data/skill_library/methods/weight-loading-optimization/SKILL.md": "a1bf0e2ec066d5f30225c6cc5fb603f830f48eb5e2aee3fc3d084ed9d8aebe2f",
- "src/inferrsi/data/skill_library/methods/weight-loading-optimization/metadata.json": "6df47ea943e42a11807c54869774957ac2383266120caad730f788207403581c",
+ "src/inferrsi/data/skill_library/methods/weight-loading-optimization/metadata.json": "ebcc5425d7b6409e5c509122e9dc76bed1caab09d73daf013d96086d7faaba2b",
"src/inferrsi/data/skill_library/methods/weight-loading-optimization/references/fastsafetensors-transfer-readset.md": "5a39f14b102b0cee304541e011aa31e44e6b13527d1bee9987c3f3b66c064d3d",
"src/inferrsi/data/skill_library/methods/weight-loading-optimization/references/flare-preparation-boundary.md": "1d7922785a196bdbafdc9791fe79428465a5acc0292ca8f99446f8c764b35454",
"src/inferrsi/data/skill_library/methods/weight-loading-optimization/references/llama-loading-readset.md": "6c640f28ee3f0253a17b81461c865cb9f947367002cb75bf72550c4572d02e97",
@@ -136,6 +136,32 @@
}
],
"empirical_validation": "not_run",
+ "applicability": {
+ "scope": "Diagnosing the serving-level weight load path from storage to device and choosing changes to it when startup or loading latency dominates: locating the cost by source, format, placement and duplication, accounting once-per-process loading and warmup apart from per-request cost, verifying sharded downloads, restricting device-side prefetch to profiled memory-bound hot spots and ordering CPU-tier offload; also covers source-scoped disk-loading modes, pinned staging, synchronous checkpoint save semantics, fastsafetensors broadcast/scatter/coalesced delivery and the separation of model preparation from serving. Request-time host overhead belongs to host-overhead-profiling, DCP asynchronous save performance to dcp-async-save-performance and distributed recovery validity to checkpoint-recovery-consistency",
+ "prerequisites": "Flag names and API entry paths are re-verified against the pinned source revisions, and torch_npu.npu_prefetch details against the target CANN version, before they are relied on; card state is checked before launch and process placement fixed before measuring, through ascend-device-health and os-process-placement; the Diffusers enable_model_cpu_offload path requires accelerate",
+ "input_contract": "The current load path's byte source, file format, host-first or direct device placement and duplicated copies across processes or instances; startup timings for loading and warmup with byte and access counts per segment, and the expected warmup hit criteria; the safetensors index weight_map and the downloaded shard set; a profile showing memory-bound hot spots and the adjacent compute or communication segments; free device memory and whether an OOM occurs during construction or inference; for cross-rank delivery, the file-to-loader-group-rank map, tensor names, shapes, dtypes and call order, the required replicated or sliced output, the model's expected shard layout and the collective backend",
+ "produced_artifact": "The located load-path cost on the four dimensions with its matching lever, a startup ledger that keeps fixed overhead apart from per-request cost, a structural shard-completeness result naming the missing shards, a prefetch proposal behind an enable_prefetch switch defaulting to False or a recorded decision not to prefetch, an offload candidate order, the recorded loading mode, a distributed delivery choice with its proposed verification, and a preparation/serving record of serialized and rebuilt state, live resource owners and ABI/prepared-contract versions; proposed changes and verification targets, not a measured startup or delivery gain",
+ "failure_modes": [
+ "Loading and warmup are amortized into per-request latency, or a host segment without timing points is treated as costing nothing",
+ "Warmup is accepted on total time while mismatched warmup shapes have warmed a different set of cache entries",
+ "In-flight .incomplete shards are re-downloaded as corrupt, or gap-free shard numbering is accepted without matching the index weight_map",
+ "Prefetch is forced without a profiled memory-bound hot spot or a non-memory-bound dependency window",
+ "Resolution or CFG tier-downs are applied to a construction-phase OOM, or offload is adopted when compute already hides the transfer",
+ "A free-memory probe runs on an mmap-backed load, or a run whose first mlock failure latched with only a warning is treated as fully resident",
+ "Disk counters that include cache and pipes are read as physical writes, or a run without a page-cache drop is reported as a cold start",
+ "The ceil-block scatter partition is used without checking it against the model's expected shard layout, or ranks run different orders or plans so their collectives mismatch",
+ "A faster H2D copy, local clone or rank-local filtered read is counted as a cross-rank communication gain"
+ ],
+ "refusal_point": "Withholds a load-path win until the verification checklist holds and keeps the distributed-delivery checks as verification targets rather than results; does not force prefetch without a safe dependency window and disables it when neither a smaller size nor a moved position removes a regression; records unsupported save behaviors as unknown; rejects mismatched execution ABI or prepared-contract versions instead of editing their tags; and implies no Ascend support from the llama.cpp or fastsafetensors mechanisms",
+ "hardware": "NPU for the device-side prefetch API (torch_npu.npu_prefetch, checked against the target CANN version) and the Diffusers offload playbook; the prefetch and fastsafetensors methodology transfers while the fastsafetensors and llama.cpp APIs do not, and cross-rank completion follows each collective backend's own wait contract"
+ },
+ "recommendations": [
+ "Report preparation/startup/readiness, warmups and measured work as separate numbers, with load latency split into init, header, io and get phases, the cold-cache protocol stated and the effective loading mode identical across compared runs.",
+ "Place each prefetch behind an enable_prefetch switch defaulting to False, start max_size at 50% of the theoretical size, verify a single operator first and step up to 0.7 and then 1.0 only after verified wins.",
+ "Try offload candidates in order of change surface: batch size, resolution, CFG/feature flags, CPU or layered offload, tensor-parallel degree, then activation checkpointing; re-estimate the un-overlapped transfer share after adopting each tier.",
+ "Record whole-set atomic publication, asynchronous checkpointing, multi-rank save coordination and crash durability as unknown for the fastsafetensors saver instead of assuming them.",
+ "Before adopting coalesced delivery, compare it disabled and enabled with identical checkpoint, rank mapping, backend, topology, dtype, cache state, chunk/queue budgets and consumer retention, then measure complete startup under the same accounting protocol."
+ ],
"limitations": [
"Source-derived candidate; no local efficacy validation.",
"Normalized provenance covers the existing llama.cpp loading slice at 3e390af18de1131dda27913c00511725b4274c8a and the fastsafetensors distributed-delivery slice at 219e2561a41ede2d5d8eba611c248bb0effbf580; other inherited claims remain unnormalized with original identities and limits retained.",
@@ -150,10 +176,31 @@
],
"consumption_mode": "packaged"
},
- "applicability": {},
+ "applicability": {
+ "scope": "Diagnosing the serving-level weight load path from storage to device and choosing changes to it when startup or loading latency dominates: locating the cost by source, format, placement and duplication, accounting once-per-process loading and warmup apart from per-request cost, verifying sharded downloads, restricting device-side prefetch to profiled memory-bound hot spots and ordering CPU-tier offload; also covers source-scoped disk-loading modes, pinned staging, synchronous checkpoint save semantics, fastsafetensors broadcast/scatter/coalesced delivery and the separation of model preparation from serving. Request-time host overhead belongs to host-overhead-profiling, DCP asynchronous save performance to dcp-async-save-performance and distributed recovery validity to checkpoint-recovery-consistency",
+ "prerequisites": "Flag names and API entry paths are re-verified against the pinned source revisions, and torch_npu.npu_prefetch details against the target CANN version, before they are relied on; card state is checked before launch and process placement fixed before measuring, through ascend-device-health and os-process-placement; the Diffusers enable_model_cpu_offload path requires accelerate",
+ "input_contract": "The current load path's byte source, file format, host-first or direct device placement and duplicated copies across processes or instances; startup timings for loading and warmup with byte and access counts per segment, and the expected warmup hit criteria; the safetensors index weight_map and the downloaded shard set; a profile showing memory-bound hot spots and the adjacent compute or communication segments; free device memory and whether an OOM occurs during construction or inference; for cross-rank delivery, the file-to-loader-group-rank map, tensor names, shapes, dtypes and call order, the required replicated or sliced output, the model's expected shard layout and the collective backend",
+ "produced_artifact": "The located load-path cost on the four dimensions with its matching lever, a startup ledger that keeps fixed overhead apart from per-request cost, a structural shard-completeness result naming the missing shards, a prefetch proposal behind an enable_prefetch switch defaulting to False or a recorded decision not to prefetch, an offload candidate order, the recorded loading mode, a distributed delivery choice with its proposed verification, and a preparation/serving record of serialized and rebuilt state, live resource owners and ABI/prepared-contract versions; proposed changes and verification targets, not a measured startup or delivery gain",
+ "failure_modes": [
+ "Loading and warmup are amortized into per-request latency, or a host segment without timing points is treated as costing nothing",
+ "Warmup is accepted on total time while mismatched warmup shapes have warmed a different set of cache entries",
+ "In-flight .incomplete shards are re-downloaded as corrupt, or gap-free shard numbering is accepted without matching the index weight_map",
+ "Prefetch is forced without a profiled memory-bound hot spot or a non-memory-bound dependency window",
+ "Resolution or CFG tier-downs are applied to a construction-phase OOM, or offload is adopted when compute already hides the transfer",
+ "A free-memory probe runs on an mmap-backed load, or a run whose first mlock failure latched with only a warning is treated as fully resident",
+ "Disk counters that include cache and pipes are read as physical writes, or a run without a page-cache drop is reported as a cold start",
+ "The ceil-block scatter partition is used without checking it against the model's expected shard layout, or ranks run different orders or plans so their collectives mismatch",
+ "A faster H2D copy, local clone or rank-local filtered read is counted as a cross-rank communication gain"
+ ],
+ "refusal_point": "Withholds a load-path win until the verification checklist holds and keeps the distributed-delivery checks as verification targets rather than results; does not force prefetch without a safe dependency window and disables it when neither a smaller size nor a moved position removes a regression; records unsupported save behaviors as unknown; rejects mismatched execution ABI or prepared-contract versions instead of editing their tags; and implies no Ascend support from the llama.cpp or fastsafetensors mechanisms",
+ "hardware": "NPU for the device-side prefetch API (torch_npu.npu_prefetch, checked against the target CANN version) and the Diffusers offload playbook; the prefetch and fastsafetensors methodology transfers while the fastsafetensors and llama.cpp APIs do not, and cross-rank completion follows each collective backend's own wait contract"
+ },
"recommendations": [
- "Read the pinned SKILL.md and its prerequisites before choosing this procedure.",
- "Read the packaged references and owner-local cases alongside the procedure."
+ "Report preparation/startup/readiness, warmups and measured work as separate numbers, with load latency split into init, header, io and get phases, the cold-cache protocol stated and the effective loading mode identical across compared runs.",
+ "Place each prefetch behind an enable_prefetch switch defaulting to False, start max_size at 50% of the theoretical size, verify a single operator first and step up to 0.7 and then 1.0 only after verified wins.",
+ "Try offload candidates in order of change surface: batch size, resolution, CFG/feature flags, CPU or layered offload, tensor-parallel degree, then activation checkpointing; re-estimate the un-overlapped transfer share after adopting each tier.",
+ "Record whole-set atomic publication, asynchronous checkpointing, multi-rank save coordination and crash durability as unknown for the fastsafetensors saver instead of assuming them.",
+ "Before adopting coalesced delivery, compare it disabled and enabled with identical checkpoint, rank mapping, backend, topology, dtype, cache state, chunk/queue budgets and consumer retention, then measure complete startup under the same accounting protocol."
],
"limitations": [
"This generated abstract points to the complete packaged skill; external evidence and upstream repositories remain source links.",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/cuda/skills/audit-afd-p2p-exchanges.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/cuda/skills/audit-afd-p2p-exchanges.json
index 6444dc582..0c65f69a3 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/cuda/skills/audit-afd-p2p-exchanges.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/cuda/skills/audit-afd-p2p-exchanges.json
@@ -23,8 +23,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/repositories/afd-plugin/audit-afd-p2p-exchanges/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/repositories/afd-plugin/audit-afd-p2p-exchanges/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/repositories/afd-plugin/audit-afd-p2p-exchanges/SKILL.md",
"src/inferrsi/data/skill_library/repositories/afd-plugin/audit-afd-p2p-exchanges/metadata.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/cuda/skills/audit-flare-session-lifecycle.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/cuda/skills/audit-flare-session-lifecycle.json
index 2fe517e74..ce9eacbc7 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/cuda/skills/audit-flare-session-lifecycle.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/cuda/skills/audit-flare-session-lifecycle.json
@@ -29,8 +29,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/repositories/flare-local-inference/audit-flare-session-lifecycle/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/repositories/flare-local-inference/audit-flare-session-lifecycle/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/repositories/flare-local-inference/audit-flare-session-lifecycle/LICENSE",
"src/inferrsi/data/skill_library/repositories/flare-local-inference/audit-flare-session-lifecycle/SKILL.md",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/cuda/skills/diagnose-attention-bottlenecks.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/cuda/skills/diagnose-attention-bottlenecks.json
index 23c63d5cb..b631cb28f 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/cuda/skills/diagnose-attention-bottlenecks.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/cuda/skills/diagnose-attention-bottlenecks.json
@@ -16,8 +16,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/hardware/cuda/diagnose-attention-bottlenecks/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/hardware/cuda/diagnose-attention-bottlenecks/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/hardware/cuda/diagnose-attention-bottlenecks/SKILL.md",
"src/inferrsi/data/skill_library/hardware/cuda/diagnose-attention-bottlenecks/metadata.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/cuda/skills/host-overhead-profiling.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/cuda/skills/host-overhead-profiling.json
index 0449865e4..b1d173100 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/cuda/skills/host-overhead-profiling.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/cuda/skills/host-overhead-profiling.json
@@ -22,8 +22,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/hardware/cuda/host-overhead-profiling/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/hardware/cuda/host-overhead-profiling/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/hardware/cuda/host-overhead-profiling/SKILL.md",
"src/inferrsi/data/skill_library/hardware/cuda/host-overhead-profiling/metadata.json",
diff --git a/src/inferrsi/data/repo_packs/inferrsi/knowledge/cuda/skills/reason-hopper-kernels.json b/src/inferrsi/data/repo_packs/inferrsi/knowledge/cuda/skills/reason-hopper-kernels.json
index 6e886dd8f..c8d16ca99 100644
--- a/src/inferrsi/data/repo_packs/inferrsi/knowledge/cuda/skills/reason-hopper-kernels.json
+++ b/src/inferrsi/data/repo_packs/inferrsi/knowledge/cuda/skills/reason-hopper-kernels.json
@@ -36,8 +36,8 @@
"sources": [
{
"type": "git",
- "url": "https://github.com/InferMatrixLab/InferRSI/blob/ad322aab86d9a148bc4b378a15b2762e41b86908/src/inferrsi/data/skill_library/hardware/cuda/reason-hopper-kernels/SKILL.md",
- "commit": "ad322aab86d9a148bc4b378a15b2762e41b86908",
+ "url": "https://github.com/InferMatrixLab/InferRSI/blob/54d4c89d4db82cbfc10d6096fe49bb077136e8b1/src/inferrsi/data/skill_library/hardware/cuda/reason-hopper-kernels/SKILL.md",
+ "commit": "54d4c89d4db82cbfc10d6096fe49bb077136e8b1",
"paths": [
"src/inferrsi/data/skill_library/hardware/cuda/reason-hopper-kernels/SKILL.md",
"src/inferrsi/data/skill_library/hardware/cuda/reason-hopper-kernels/metadata.json",
diff --git a/src/inferrsi/data/skill_library/methods/assess-serving-preemption/metadata.json b/src/inferrsi/data/skill_library/methods/assess-serving-preemption/metadata.json
index 168c782db..d03ef788f 100644
--- a/src/inferrsi/data/skill_library/methods/assess-serving-preemption/metadata.json
+++ b/src/inferrsi/data/skill_library/methods/assess-serving-preemption/metadata.json
@@ -8,6 +8,27 @@
"origin": [
"external"
],
+ "applicability": {
+ "scope": "Diagnosing whether LLM serving queue delay justifies token-boundary priority preemption, including suspended KV, host swap or recomputation and long-request service guarantees; admission-versus-service-order diagnosis, not current scheduler implementation or tenant-fairness certification",
+ "prerequisites": "The packaged source notes are read before quoting a metric or comparing a baseline; installing, benchmarking, changing scheduling or admission, migrating state, editing a repository or running a reproduction needs a separately authorized task",
+ "input_contract": "The late boundary (first token, token gap, completion or a stated cohort's service) with its exact percentile or attainment target and denominator; model and KV format, engine revision, safe interruption boundary, request batch limit, input/output/context distributions, arrival bursts, device live/reserved/allocated bytes, active and suspended requests, host budget, transfer path, recovery time and queue/phase/service timings; per-tenant or priority-class obligations kept apart from aggregate latency",
+ "produced_artifact": "A read-only decision naming the observed failure boundary, the smallest candidate that addresses it, unchanged assumptions, missing evidence, resource cost, affected cohorts, counter-signal and next authorized gate, plus a proposed verification plan; not a validated configuration",
+ "failure_modes": [
+ "An ordinary iteration boundary is taken to permit interrupting a prefill halfway through",
+ "Unknown output length is treated as known from the prompt",
+ "KV is counted for the running batch only, omitting started requests that are suspended",
+ "Reservations are removed because some reserved capacity is not live",
+ "Recompute-based eviction is proposed without a progress invariant, so two jobs can repeatedly displace each other",
+ "Host-device copies or same-host IPC are charged to Network instead of their local resource"
+ ],
+ "refusal_point": "Withholds a preemption recommendation when the apparent gain can be explained by different kernels, unmatched parallelism or an incomparable latency metric; unknown safe boundaries, unsupported state recovery, uncalibrated timing or unclear service obligations are blockers"
+ },
+ "recommendations": [
+ "Name the late boundary and its exact target and denominator before choosing among batching, admission, preemption and state recovery.",
+ "Send prefill interference to choose-prefill-decode-schedule and physical KV slack to diagnose-kv-block-efficiency when those candidates are available; otherwise report the diagnosed boundary and the missing evidence.",
+ "In a later authorized test, isolate service ordering first and state recovery second, and evaluate their interaction only after both are understood.",
+ "Weigh FastServe's same-engine FCFS comparison above its historical-vLLM headline, and do not read its P95 aggregate gain as a starvation or tenant-fairness guarantee."
+ ],
"limitations": [
"Source-derived candidate; no local efficacy validation.",
"Original typed source inventory is not normalized; retain source identities and limits in the procedure/readsets."
diff --git a/src/inferrsi/data/skill_library/methods/assess-split-backward-schedule/metadata.json b/src/inferrsi/data/skill_library/methods/assess-split-backward-schedule/metadata.json
index 7cf649339..18173e4cf 100644
--- a/src/inferrsi/data/skill_library/methods/assess-split-backward-schedule/metadata.json
+++ b/src/inferrsi/data/skill_library/methods/assess-split-backward-schedule/metadata.json
@@ -7,6 +7,26 @@
"origin": [
"external"
],
+ "applicability": {
+ "scope": "Assessing memory-capped split-backward and V-shaped pipeline schedule proposals when a pipeline trace suggests input-gradient propagation waits behind deferrable weight-gradient work or an optimizer barrier, while preserving synchronous update, global-batch and optimizer-validation semantics",
+ "prerequisites": "The packaged evidence notes are read before citing the paper; installing or running frameworks, benchmarking schedules, changing training semantics, configuring hardware or modifying repositories needs separate authorization",
+ "input_contract": "Stage and microbatch counts, microbatch and global batch, precision, optimizer and clipping/NaN behavior; F, input-gradient B, weight-gradient W and transfer timings; per-pass retained and released bytes; total device peak and margin; collective ordering; kernel-splitting support; calculated bubble rate kept apart from measured idle time and useful samples/s",
+ "produced_artifact": "A candidate schedule and fallback with per-pass order rationale, peak-memory and transfer costs, unchanged semantic requirements, counter-signals and a future verification proposal; a reviewable plan, not an enabled schedule",
+ "failure_modes": [
+ "Completion of B is equated with releasing all activations, so W inputs retained by deferral go uncounted",
+ "The equal-time zero-bubble construction replaces observed F/B/W and transfer times, or kernel-splitting and synchronization overhead is left out",
+ "The schedule is chosen by minimum modeled bubble rate instead of whole-step throughput under the memory and quality constraint",
+ "pMB activation capacity is read as total memory, hiding HBM, workspace, buffers or per-rank imbalance",
+ "Removing the optimizer barrier drops global norm, clipping or NaN decisions, rollback and re-update, or the correctness of work that depends on a tentative update",
+ "Optimizer post-validation is conflated with accepting stale-gradient delayed updates"
+ ],
+ "refusal_point": "Withholds an enablement recommendation under frequent clipping/NaNs, small available memory, strongly unequal passes, unsupported split kernels, hidden collectives or dynamic shapes, and retains the synchronized fallback while optimizer reversibility and rollback coverage are unverified"
+ },
+ "recommendations": [
+ "Compare ordinary 1F1B with a memory-capped split schedule first; add V-shaped two-chunk placement only when memory permits, and at matched global batch also try smaller microbatches with more microbatches.",
+ "Decide optimizer post-validation separately from the schedule change.",
+ "In a later authorized comparison, hold model, global batch and precision fixed and compare real with calculated idle intervals."
+ ],
"limitations": [
"Source-derived candidate; no local efficacy validation.",
"Original typed source inventory is not normalized; retain source identities and limits in the procedure/readsets."
diff --git a/src/inferrsi/data/skill_library/methods/audit-fixed-workload-claim/metadata.json b/src/inferrsi/data/skill_library/methods/audit-fixed-workload-claim/metadata.json
index 81b8fdcfc..1e6acf5cc 100644
--- a/src/inferrsi/data/skill_library/methods/audit-fixed-workload-claim/metadata.json
+++ b/src/inferrsi/data/skill_library/methods/audit-fixed-workload-claim/metadata.json
@@ -3,6 +3,25 @@
"origin": ["internal"],
"tags": ["performance", "audit", "measurement"],
"empirical_validation": "not_run",
+ "applicability": {
+ "scope": "Auditing a performance claim against supplied fixed-workload reports and an explicit measurement contract: aggregation, retained observations, comparability and quality evidence",
+ "input_contract": "Baseline and candidate fixed-workload reports with request observations grouped by trial and workload; the measurement contract stating metric, unit, timing boundary, improvement direction, aggregation rule and order, fixed exclusions, threshold and comparison operator; the declared quality and completion evidence; the requested output schema",
+ "produced_artifact": "An audit conclusion showing baseline and candidate aggregates, the calculation, the threshold comparison and the evidence or missing information that controls the verdict; it interprets supplied evidence and does not itself establish a measured improvement",
+ "failure_modes": [
+ "A pooled request median silently replaces a median of per-trial means, or the reverse, reweighting trials whose request counts differ",
+ "Late requests, failures or outliers are dropped, or requests move between trials, outside the exclusions the contract fixes",
+ "Values are rounded before the threshold comparison",
+ "A missing rule that permits different verdicts is resolved toward the favorable one",
+ "A speed improvement is read as task success, or a quality pass as equivalent measured work",
+ "Different workloads or incomplete observations are repaired by a favorable aggregate"
+ ],
+ "refusal_point": "Does not replace missing evidence with a pass, does not support a percentage claim on an undefined denominator, and does not launch a new benchmark or alter the contract to resolve an unfavorable result; a failed threshold is reported apart from an unverified comparison or missing quality evidence"
+ },
+ "recommendations": [
+ "Recover the contract's metric, direction, aggregation order and comparison operator before reading the report's own verdict.",
+ "Treat a named minimum bound as >= and a maximum bound as <= unless the contract requires a strict comparison.",
+ "Keep separate regimes and metrics separate unless the contract defines their aggregation."
+ ],
"limitations": [
"Guidance for supplied reports; no local efficacy validation or measured performance gain is established."
]
diff --git a/src/inferrsi/data/skill_library/methods/calibrate-serving-simulation/metadata.json b/src/inferrsi/data/skill_library/methods/calibrate-serving-simulation/metadata.json
index f684ccde7..b9c515288 100644
--- a/src/inferrsi/data/skill_library/methods/calibrate-serving-simulation/metadata.json
+++ b/src/inferrsi/data/skill_library/methods/calibrate-serving-simulation/metadata.json
@@ -21,6 +21,26 @@
"readset": "references/source-notes.md"
}
],
+ "applicability": {
+ "scope": "Building an evidence-gated serving-configuration shortlist when a profiled simulator or cost model is proposed to reduce deployment experiments; extends performance-ceiling-analysis, optimization-search-policy and inference-benchmark-design without replacing their ceilings or measurement contract",
+ "prerequisites": "The packaged Vidur evidence and evidence-type notes are read; reading the skill does not authorize execution, and no validation check runs under it without separate execution authorization",
+ "input_contract": "Model, shape and dtype; actual kernel/runtime revision; graph mode; GPU, host and topology; request length and arrival distribution; total resources and cost basis; memory limits; latency definitions and percentile SLOs",
+ "produced_artifact": "A bounded decision memo with input identities, model assumptions, evidence type, exact metrics and denominators, unsupported dimensions, nondominated candidates with predicted ranges, joint constraint checks, the next discriminating trial and rollback/finalization requirements, keeping prediction and measurement columns apart; a shortlist for future measurement, not a measured result",
+ "failure_modes": [
+ "Learned runtime output is labeled a physical bound",
+ "A missing identity is filled with a paper default, or topology is substituted, instead of being marked unknown",
+ "A low-load check stands in for an SLO claim at the intended load regime and capacity boundary",
+ "A predictor that relies on summed KV volume hides batches with equal volume but unequal context distributions",
+ "Capacity binary search is applied without verifying its monotonic/stationary premise or treating overloaded samples explicitly",
+ "One predicted winner is reported when uncertainty exceeds rank separation"
+ ],
+ "refusal_point": "Stops predictive certification when profiles are missing, topology is substituted, error grows near capacity or host overhead is omitted; labels an ordering unresolved when uncertainty exceeds rank separation; asserts neither universal error below 9% nor current vLLM, asynchronous PP or NPU support from the historical simulator paper"
+ },
+ "recommendations": [
+ "Classify every estimate as a physical bound, measured envelope, Amdahl-style conditional impact or empirical prediction before comparing configurations.",
+ "Keep prefill and decode separate in the operator signatures, and record how sharding and memory planning enter the model.",
+ "Recalibrate after workload drift, kernel/graph/runtime changes, new hardware or topology, or mismatched residuals."
+ ],
"limitations": [
"Source-derived candidate; no local efficacy validation."
],
diff --git a/src/inferrsi/data/skill_library/methods/checkpoint-recovery-consistency/metadata.json b/src/inferrsi/data/skill_library/methods/checkpoint-recovery-consistency/metadata.json
index 0e000aa09..8ca288e82 100644
--- a/src/inferrsi/data/skill_library/methods/checkpoint-recovery-consistency/metadata.json
+++ b/src/inferrsi/data/skill_library/methods/checkpoint-recovery-consistency/metadata.json
@@ -14,6 +14,28 @@
"external"
],
"empirical_validation": "not_run",
+ "applicability": {
+ "scope": "Assessing distributed checkpoint recovery through DLRover memory agreement, persistence completion and replica transport, and NVIDIA Resiliency Extension complete-iteration selection; distinguishes worker, host and replica loss and separates queued saves, publication, application-state validity and durable recovery. Save performance belongs to dcp-async-save-performance",
+ "prerequisites": "The pinned sources and limits are read; reading authorizes no launchers, failure injection or checkpoint loading, and restore evidence needs separate authorization",
+ "input_contract": "The declared failure domain; actual memory, storage and replica placement and helper lifetime; step or iteration identity, the expected shard/rank set and finalization result; the application's model, optimizer and caller-required RNG/scheduler/data-position state contract; the target environment's format and world-size/replication assumptions",
+ "produced_artifact": "An answer to each of the four acceptance questions (snapshot survival, cross-participant completeness, application-state completeness, restorability) at its own evidence level, a proposed per-generation recovery evidence record including errors and skipped saves, and a proposed verification split by worker, host, replica and persistent-storage loss; no recovery success, durability, NPU support or measured gain",
+ "failure_modes": [
+ "Equal steps or complete rank IDs are taken to verify tensor contents or whole application semantics",
+ "A skipped save, where not every participant acquired shared-memory access, is reported as a new recovery point",
+ "A returned step paired with an empty state dictionary is read as a successful restore",
+ "A local maximum filename or one clean shard substitutes for global coverage",
+ "The .dirty-then-rename publication is read as power-loss durability",
+ "An asynchronous save request is treated as a completed checkpoint, or queue concurrency is assumed unlimited",
+ "One failure domain is tested and recovery from all of them is claimed"
+ ],
+ "refusal_point": "Claims no recovery success, durability, NPU support or measured gain, does not upgrade allocation, metadata or a completed collective into proven recovery, and neither injects failures nor authorizes them"
+ },
+ "recommendations": [
+ "Answer the four acceptance questions separately; record each at its own evidence level.",
+ "Count replication as added availability only when copies occupy independent failure domains, and budget clean, dirty and replicated shards on each host.",
+ "Keep the one-outstanding asynchronous save restriction until the target queue version is checked.",
+ "Send save-throughput questions to dcp-async-save-performance; a faster return or higher write throughput does not establish recoverability."
+ ],
"limitations": [
"Source-derived candidate; no local efficacy validation.",
"Original typed source inventory is not normalized; retain source identities and limits in the procedure/readsets."
diff --git a/src/inferrsi/data/skill_library/methods/choose-parallel-training-plan/metadata.json b/src/inferrsi/data/skill_library/methods/choose-parallel-training-plan/metadata.json
index 7cf649339..afc792f30 100644
--- a/src/inferrsi/data/skill_library/methods/choose-parallel-training-plan/metadata.json
+++ b/src/inferrsi/data/skill_library/methods/choose-parallel-training-plan/metadata.json
@@ -7,6 +7,26 @@
"origin": [
"external"
],
+ "applicability": {
+ "scope": "Comparing stage, device-mesh and operator-sharding candidates for static distributed training when heterogeneous stage memory or work, or communication cost, makes a uniform data/tensor/pipeline plan suspect; derived from historical Alpa, not a current compiler configuration guide",
+ "prerequisites": "The packaged source notes are read; the skill authorizes no compiler or solver install or run, parallelism reconfiguration, repository edit, hardware task or benchmark",
+ "input_contract": "Model operator graph and shapes; global batch, microbatch and accumulation; precision and optimizer; physical links and contention; candidate shard layouts; stage latency, workspace and in-flight activation memory; the objective and expected run length. A device count without topology is insufficient",
+ "produced_artifact": "The best supported candidate and a simpler fallback with a memory and critical-path ledger, planning cost and amortization, semantic invariants, counter-signals, missing evidence and a proposed matched comparison; a plan, not a reconfiguration",
+ "failure_modes": [
+ "Equal arithmetic work is taken to mean equal kernel time, carrying over Alpa's zero compute-cost simplification",
+ "Mean memory replaces the largest per-rank peak with multiple in-flight activation sets",
+ "Inter-stage transfer and microbatch count stay outside the objective, as in historical Alpa, without the gap being reported",
+ "Repeated slow-link sends to replicated destination tiles are not compared with one cross-mesh transfer plus local allgather",
+ "Inter-node Network traffic is merged with intra-node accelerator links",
+ "Static cost estimates are applied to dynamic shapes, asymmetric devices or changing communication contention"
+ ],
+ "refusal_point": "Rejects infeasible stage/mesh pairs, claims no global optimality outside the declared search space, and stops at the plan in paper-only mode"
+ },
+ "recommendations": [
+ "Keep a simple reference plan with the same training semantics and global batch as the comparison point.",
+ "Separate fixed-model scaling from model-size weak scaling before quoting a speedup; Alpa's 9.7x MoE ratio belongs to a specific weak-scaling workload, baseline and 25-Gbps cluster.",
+ "Weigh planning cost against the expected run length; a shorter job may not amortize it."
+ ],
"limitations": [
"Source-derived candidate; no local efficacy validation.",
"Original typed source inventory is not normalized; retain source identities and limits in the procedure/readsets."
diff --git a/src/inferrsi/data/skill_library/methods/dense-performance-evidence/metadata.json b/src/inferrsi/data/skill_library/methods/dense-performance-evidence/metadata.json
index 7893a0b1f..14777a2c0 100644
--- a/src/inferrsi/data/skill_library/methods/dense-performance-evidence/metadata.json
+++ b/src/inferrsi/data/skill_library/methods/dense-performance-evidence/metadata.json
@@ -13,6 +13,26 @@
"external"
],
"empirical_validation": "not_run",
+ "applicability": {
+ "scope": "Deciding which performance evidence to collect next and which claim each evidence level supports, from correctness through microbenchmark, component profile, isolated scenario benchmark and multi-regime serving benchmark to a production A/B experiment; also structuring a measurement campaign",
+ "prerequisites": "Level 0 correctness evidence is collected before any performance measurement; a Level 5 experiment needs production infrastructure, traffic splitting and statistical power",
+ "input_contract": "The claim to be made or the question to answer, and the evidence already collected with its level, workload, hardware/software context and controls",
+ "produced_artifact": "The minimum evidence level each claim requires, the next evidence to collect under the escalation rule, and negative-result records with hypothesis, evidence level, the metric that did not improve, context and the reason if known; a classification of evidence, not a measurement",
+ "failure_modes": [
+ "An md5 match or a test pass is read as semantically correct output or as support for a performance claim",
+ "A microbenchmark improvement is reported as an inference throughput gain",
+ "Absolute kernel durations or single collective latencies are compared across devices or sessions",
+ "A win on one fixed workload or concurrency is generalized to production traffic",
+ "Worse metrics, per-task regressions or negative results are discarded, or a shared task budget is taken as equal total cost"
+ ],
+ "refusal_point": "Withholds any claim above the level its evidence reaches, leaves an unresolved source protocol or denominator mismatch unresolved, and lets neither upstream scores nor local packaging checks promote an InferRSI candidate"
+ },
+ "recommendations": [
+ "Collect Level 0 first and profile at Level 2 before implementing; escalate to Level 3 only after a promising Level 2 result, and to Level 4 only after a Level 3 win worth validating across regimes.",
+ "Collect a Level 1 microbenchmark only for an isolated-kernel question after Level 2 has confirmed that kernel is on the critical path.",
+ "For a Level 1 claim, pin input shape and dtype, exclude warmup, synchronize before timing, run at least 10 repetitions and report median and IQR; for a Level 3 claim, match hardware, software stack, model revision and inputs, run at least 3 repetitions and alternate A/B/A order.",
+ "Read the admission/drain occupancy case when interpreting finite service cohorts; phase-selected occupancy is not end-to-end gain."
+ ],
"limitations": [
"Source-derived candidate; no local efficacy validation.",
"Original typed source inventory is not normalized; retain source identities and limits in the procedure/readsets."
diff --git a/src/inferrsi/data/skill_library/methods/evaluate-elastic-serving-placement/metadata.json b/src/inferrsi/data/skill_library/methods/evaluate-elastic-serving-placement/metadata.json
index c87aaf462..312f66977 100644
--- a/src/inferrsi/data/skill_library/methods/evaluate-elastic-serving-placement/metadata.json
+++ b/src/inferrsi/data/skill_library/methods/evaluate-elastic-serving-placement/metadata.json
@@ -8,6 +8,26 @@
"origin": [
"external"
],
+ "applicability": {
+ "scope": "Comparing migration of a running LLM request between compatible model replicas with changing its sequence-parallel execution group, including state-copy, reservation, recomputation, interruption and priority costs, for cross-instance memory imbalance or highly variable long-context serving; not connector wiring or live scaling",
+ "prerequisites": "The relevant Llumnix or LoongServe branch and its readset in the packaged source notes are read; moving state, altering admission or priority, scaling replicas, changing parallelism, running experiments or editing frameworks needs separate authorization",
+ "input_contract": "Exact model and cache identity and format; weight residency; request phase, context and generated length; source and destination KV bytes; physical free/reserved memory and virtual accounting; batch state; current parallel group; queue and service times; actual topology and transport for each path; the target SLO statistic and affected priority or tenant cohorts",
+ "produced_artifact": "The smallest justified branch with missing capability or evidence, expected saved work or wait, added state, traffic and replica cost, affected SLO and cohorts, counter-signals, the proposed next gate and a future verification plan; an evidence-based recommendation, not a validated deployment rule",
+ "failure_modes": [
+ "Model-weight repartitioning, pipeline configuration, instance migration and sequence sharding are treated as interchangeable uses of parallelism",
+ "Aggregate free bytes are taken to mean a single blocked request fits, conflating internal KV-block slack with cross-instance fragmentation",
+ "Llumnix's 20-30 ms measured downtime is substituted for total transfer time",
+ "Virtual headroom is counted as extra memory or as a tenant-fairness guarantee",
+ "A zero-overhead transition label is read as every group change being free, or a token sum hides length-sensitive attention work",
+ "Ongoing distributed-attention communication goes uncharged because existing KV does not migrate"
+ ],
+ "refusal_point": "Withholds the candidate when every destination is full, pre-copy catch-up is unproven, model or cache compatibility is unknown, or the measured work saved does not exceed exposed costs; current-code review and target-system tests remain a later authorization gate"
+ },
+ "recommendations": [
+ "Decide first whether the question is replica relocation, a sequence-parallel group change, or prefix reuse and phase pools, which evaluate-kv-cache-transfer-gates and choose-prefill-decode-schedule cover when available.",
+ "Name each transfer path separately; only inter-node transport is Network.",
+ "Keep a feasible static, no-migration baseline in any later comparison."
+ ],
"limitations": [
"Source-derived candidate; no local efficacy validation.",
"Original typed source inventory is not normalized; retain source identities and limits in the procedure/readsets."
diff --git a/src/inferrsi/data/skill_library/methods/inference-benchmark-design/metadata.json b/src/inferrsi/data/skill_library/methods/inference-benchmark-design/metadata.json
index f9f841fd6..a8c90d7d4 100644
--- a/src/inferrsi/data/skill_library/methods/inference-benchmark-design/metadata.json
+++ b/src/inferrsi/data/skill_library/methods/inference-benchmark-design/metadata.json
@@ -142,6 +142,28 @@
}
],
"empirical_validation": "not_run",
+ "applicability": {
+ "scope": "Selecting task quality gates, reference parity checks and performance workloads for text, agentic coding, work agents, TTS, full-duplex audio/video, image and video inference, and designing goodput-based multi-regime comparisons that fit InferRSI's evidence trail",
+ "prerequisites": "The primary objective, timing boundary, quality gates, sample/failure accounting, required regimes, repetitions, seeds and decision rule are frozen before optimization; baseline and candidate share hardware, software stack, input distribution, benchmark configuration, warmup and measurement protocol, and on a shared machine also one time window and dedicated devices",
+ "input_contract": "The task family and deployment use case; candidate quality evaluators and reference parity; the declared SLO and the caller's policy on which regimes must improve or must not regress; for an audit, a supplied fixed-workload report with request-to-trial grouping, exclusion and failure counts and the declared aggregation rule",
+ "produced_artifact": "A frozen benchmark design with evaluators, reference parity, primary objective, timing boundary, quality gates, required regimes, repetition and seed rules and an explicit decision config, plus the evidence-tree layout a PASS claim must produce; for a supplied report, the computed threshold result and missing evidence. Candidate knowledge, not qualification of any model or backend",
+ "failure_modes": [
+ "Raw throughput is optimized in place of goodput, so requests served outside the SLO count toward a win",
+ "One concurrency level or one undisclosed aggregate score stands in for all required regimes",
+ "Recorded agent task scores or replay completion are read as current task success, or a VBench prompt workload as computed VBench quality",
+ "An upstream better/worse/inconclusive label is used as InferRSI's PASS/FAIL/INCONCLUSIVE verdict",
+ "Runs on the same hardware class, in different time windows or on shared devices are compared as if pinned",
+ "A quality pass is taken to certify that the timed work actually occurred",
+ "Best-seen results are reported as delivered without restarting and validating the final code/config artifact"
+ ],
+ "refusal_point": "Records missing coverage and high-variance results as INCONCLUSIVE rather than a win; a run without a frozen RunSpec, or whose validator changed after the plan was frozen, is INCONCLUSIVE regardless of metric values; sets no universal numeric gate or noise threshold"
+ },
+ "recommendations": [
+ "Freeze one primary objective, timing boundary, quality gates and sample/failure accounting before optimization.",
+ "Choose the required regimes from the deployment use case, report each separately, and record which must improve or must not regress in a versioned decision config.",
+ "Use the replay trace that matches the actual deployment pattern.",
+ "State the concurrency, trace, model, SKU and microbenchmark scope boundaries in any report that cites benchmark results."
+ ],
"limitations": [
"Source-derived candidate; no local efficacy validation.",
"Original typed source inventory is not normalized; retain source identities and limits in the procedure/readsets.",
diff --git a/src/inferrsi/data/skill_library/methods/inference-optimization-casebook/metadata.json b/src/inferrsi/data/skill_library/methods/inference-optimization-casebook/metadata.json
index b7db68e5a..0d321e5cb 100644
--- a/src/inferrsi/data/skill_library/methods/inference-optimization-casebook/metadata.json
+++ b/src/inferrsi/data/skill_library/methods/inference-optimization-casebook/metadata.json
@@ -13,6 +13,27 @@
"external"
],
"empirical_validation": "not_run",
+ "applicability": {
+ "scope": "Structuring and querying a casebook of past optimization decisions as decision support, when recording optimization experiments or choosing the next optimization from prior art; the casebook is reference material and neither runs profiling nor makes decisions",
+ "prerequisites": "Whole-body text access to the case files for free-text search; applying a case first requires reading the nearest checked-in config and the model's deployment guide",
+ "input_contract": "For a query: natural search terms such as model name, knob, mechanism word or deployment feature, plus the current model and hardware. For a new case: mechanism, applies-when and does-not-apply conditions, procedure, expected effect, accuracy risk, verification, rollback and a real measurement trail with run ID, verdict, exact hardware SKU, model revision, workload and before/after result",
+ "produced_artifact": "Ranked matching cases read as precedents and diagnostic options, or a new or deprecated case record in the template; a precedent's procedure and expected direction of effect, not a guaranteed configuration or a measured result",
+ "failure_modes": [
+ "The query is pre-translated into canonical vocabulary or bounded by structured fields, so synonyms in the body text are missed",
+ "A case's specific values, such as batch size, tensor-parallel degree or kernel parameters, are copied instead of re-derived for the current model and hardware",
+ "The precedent's expected direction of effect stands in for measurement after the change is applied",
+ "A rejected direction is read as a permanent blacklist, or one discarded implementation as ruling out its mechanism",
+ "A confounded run, a missing experiment and a failed gate are recorded as the same evidence state",
+ "Cross-system cases are bootstrapped without re-verifying that the mechanism holds on the target hardware"
+ ],
+ "refusal_point": "Adds no case for an INCONCLUSIVE optimization-effect claim, a single-configuration result without plausible generalization, or a result without a real measurement trail, except a scoped diagnostic observation that keeps its verdict and configuration boundary; manufactures no PASS, and reading the records does not reopen a paused campaign"
+ },
+ "recommendations": [
+ "Search the full case bodies with free text first, then rank the hits by bottleneck class and signals.",
+ "Before retrying a rejected direction, state which condition changed or which uncertainty a bounded check resolves, run an unchanged control and choose the least expensive discriminating check.",
+ "Mark bootstrapped cases source_grounded or author_reported until a local measurement confirms or refutes them.",
+ "When a case's conditions no longer hold, mark it deprecated with the reason instead of deleting it."
+ ],
"limitations": [
"Source-derived candidate; no local efficacy validation.",
"Original typed source inventory is not normalized; retain source identities and limits in the procedure/readsets."
diff --git a/src/inferrsi/data/skill_library/methods/kernel-writing-guide/metadata.json b/src/inferrsi/data/skill_library/methods/kernel-writing-guide/metadata.json
index 8d5b53808..cc50284f2 100644
--- a/src/inferrsi/data/skill_library/methods/kernel-writing-guide/metadata.json
+++ b/src/inferrsi/data/skill_library/methods/kernel-writing-guide/metadata.json
@@ -156,6 +156,37 @@
}
],
"empirical_validation": "not_run",
+ "applicability": {
+ "scope": "Deciding whether to write a custom inference kernel, routing a kernel pattern to Triton, CuTe DSL, CUDA C++, AscendC or TileIR, applying framework-independent correctness rules, and verifying the executing kernel against a reference before benchmarking; covers fused elementwise chains, reductions, tiled GEMM, attention variants, custom collectives, causal convolution and stateful serving-operator qualification",
+ "prerequisites": "The applicable operator, caller/state and layout contracts are read before substitution, including the causal-conv1d contract and HF Qwen3.5 GDN binding; the CuTe DSL environment uses Linux x86_64, Python 3.10–3.13 and Ampere+ (SM80+) with nvidia-cutlass-dsl installed and no NPU target; the CANN programming guide for the specific SoC is consulted for AscendC",
+ "input_contract": "The candidate operation or operation chain, its shapes, dtypes, quantization and layout requirements, and the target execution phase; the target accelerator and framework constraints; for serving operators, the concrete caller, effective phase and backend, state-row mapping, cache or buffer lifetime and in-place-write behavior; a reference implementation with the same signature, a callable kernel wrapper and fresh CUDA or NPU test tensors; verification shapes and dtypes plus profiler output for the target metric",
+ "produced_artifact": "A write-or-decline decision for the candidate operation with the profiled bottleneck, fusable pair or required dtype, shape or quantization variant unsupported by the standard implementation that justifies it, a framework choice routed by pattern and accelerator, a correctness checklist covering fp32 transcendental round-trips, sync-free wrappers, C division semantics, pre-allocated outputs, fp32 matmul accumulators and masked boundary loads and stores, a starting block or tile size with its autotune and warmup cost noted, a kernel shipping kernel_fn, reference_fn and get_inputs, a reference-parity result across the verification shapes and dtypes, and a post-verification profiler reading on the target metric; for an existing causal-convolution or stateful serving operator, a qualification record naming the caller, phase, backend, state rows and in-place writes, with source selection kept apart from executed dispatch; a routing and correctness decision, not a complete implementation or a measured speedup",
+ "failure_modes": [
+ "A kernel is written for a standalone elementwise op, a standalone matmul or standard attention where the vendor library already wins",
+ "Host-side scheduling cost is treated as a kernel problem instead of being routed to host-overhead-profiling",
+ "Fusion is justified by removed Python call overhead rather than by the DRAM traffic the shared data no longer makes",
+ "A transcendental runs on fp16/bf16 inputs without the fp32 cast round-trip, so the error stays silent through benchmarking",
+ "tl.dot accumulates in the input dtype because Triton is assumed to promote fp16/bf16 to fp32 automatically",
+ "Negative offsets are computed assuming Python floor division while Triton or AscendC rounds toward zero",
+ "An unmasked load or store is accepted because the benchmark shapes happen to divide the block size",
+ "A kernel is benchmarked before it matches its reference, or a shape-boundary, accumulator-dtype or precision failure is read as a tuning problem",
+ "A verification pass is reported as a win although the profiler metric did not move, so the kernel was never on the critical path",
+ "Autotune compile cost is measured as steady-state serving latency instead of being pre-warmed or cached",
+ "A CuTe DSL pattern is routed to an NPU target, or a CUDA tail-guard, attention primitive or SM-matched example is assumed to carry to Ascend",
+ "A causal-conv1d or GDN operator is substituted on build success, import success or matching names without resolving dtype, strides, packed boundaries, state rows and cursors",
+ "An in-tree vLLM wrapper, a GDN prefill flag or a cache-free classification adapter is read as establishing the executed convolution path or generation-state support"
+ ],
+ "refusal_point": "Declines to start a kernel until a profiler puts the target on the critical path near its hardware ceiling, a concrete fusable pair is identified, or the required dtype, shape or quantization variant is unsupported by the standard implementation, and sends host-bound cases elsewhere; refuses to benchmark an unverified kernel and withholds a gain whose profiler metric did not move; keeps operator source selection separate from executed dispatch, numerical parity and hardware qualification, so build or import success qualifies nothing; treats AscendC and TileIR content as source-inferred pending CANN and SoC verification with no local NPU run, treats the AKG Affine FileCheck expectations as IR structure rather than executed CUDA correctness, and reads author-reported overlap and FP8 cases as counterexamples to unconditional tuning rules; claims no Ascend support from NVIDIA primitives",
+ "hardware": "CUDA Ampere+ and Ascend NPU have separate framework routes for each kernel pattern; CuTe DSL requires Linux x86_64, Python 3.10-3.13 and SM80+ with nvidia-cutlass-dsl and is unavailable on NPU; profiling uses ncu on CUDA and msprof on Ascend; the AscendC and TileIR guidance is source-inferred with no local NPU verification and requires validation against the specific CANN version and SoC"
+ },
+ "recommendations": [
+ "Confirm with nsys/ncu or msprof that the target kernel sits on the critical path near its hardware ceiling, name the adjacent operations whose shared data would otherwise round-trip through DRAM, or identify a required dtype, shape or quantization variant unsupported by the standard implementation, before writing any kernel.",
+ "Ship every kernel with kernel_fn, reference_fn and get_inputs, pass reference parity across the verification shapes and dtypes first, and profile afterwards to confirm the target metric actually moved.",
+ "Cast fp16/bf16 to fp32 and back for any math beyond +, -, * and /, declare tl.dot accumulators as tl.float32, and mask every load and store whose shape does not divide the block size.",
+ "Start Triton at BLOCK_SIZE 1024 on Hopper and 512 on Ampere using powers of two, take GEMM tiles from the source skeleton configs (128/256/64 for Hopper, 64/128/32 for smaller cards) and autotune from there.",
+ "On Ascend, reach for TileIR before raw AscendC, size tiles to fit the AI Core Unified Buffer and check cube-unit alignment against the CANN guide for the specific SoC.",
+ "Match a CuTe DSL example to the GPU architecture (SM80 Ampere, SM90 Hopper, SM100 Blackwell) and study its structure instead of copying it, and pre-warm or cache autotuned kernels at startup so first-run compile cost stays out of serving latency."
+ ],
"limitations": [
"Source-derived candidate; no local efficacy validation.",
"Original typed source inventory is not normalized; retain source identities and limits in the procedure/readsets.",
diff --git a/src/inferrsi/data/skill_library/methods/optimization-search-policy/metadata.json b/src/inferrsi/data/skill_library/methods/optimization-search-policy/metadata.json
index acdf3a6d4..e718b8d87 100644
--- a/src/inferrsi/data/skill_library/methods/optimization-search-policy/metadata.json
+++ b/src/inferrsi/data/skill_library/methods/optimization-search-policy/metadata.json
@@ -13,19 +13,27 @@
"external"
],
"empirical_validation": "not_run",
- "sources": [
- {
- "type": "paper",
- "id": "usenix:osdi24:zhong-yinmin",
- "version": "OSDI 24 proceedings, July 2024, pages 193-210",
- "url": "https://www.usenix.org/conference/osdi24/presentation/zhong-yinmin",
- "sections": ["3.3", "4.1-4.2", "6.1-6.5", "7", "Appendices A-C"],
- "readset": "references/distserve-phase-placement/source-record.json"
- },
- {
- "type": "unknown",
- "scope": "Pre-existing sources remain in their original procedure/readsets; this amendment does not normalize that inventory"
- }
+ "applicability": {
+ "scope": "Selecting, ordering and stopping optimization experiments: candidate generation, exploration versus exploitation, mechanism diversity, negative experiment memory, cost-aware escalation, budget allocation and stopping criteria, drawn from vllm-evolve, the TRT-LLM casebook and SkillOpt",
+ "prerequisites": "The useful work, quality contract, allowed approximations, timing boundary and resource budget are frozen before search, and a complete restartable incumbent is preserved; reading the procedures authorizes no execution, framework change, new resource spending or publication",
+ "input_contract": "Profiling data, existing config, prior experiment and negative records and the casebook for the current bottleneck class, model and hardware; each target stage's fraction of E2E time; the target, locally measured uncertainty and the authorized budget",
+ "produced_artifact": "An ordered experiment plan: one candidate change per round with an explicit hypothesis and its motivating evidence, a bounded shortlist of distinct mechanisms with assumptions, uncertainty, expected impact and evaluation cost, negative records turned into search constraints, and a stop-or-switch decision; when authorized rounds run, a recorded verdict per round with best-seen and delivered outcomes kept apart",
+ "failure_modes": [
+ "The evaluation protocol, measurement setup or baseline changes in the same round as the candidate",
+ "Benchmark budget is spent on a candidate that fails correctness",
+ "A matching keyword in a rejection record is used as a causal diagnosis or automatic veto, or one failed implementation excludes a whole solution family",
+ "Expensive work starts on a stage without computing its Amdahl contribution, so a large local speedup yields little E2E gain",
+ "One mechanism family keeps being sampled only because budget remains, without re-profiling after the limiting stage moves",
+ "Roofline proximity or a utilization threshold is used as a universal stopping rule",
+ "Screening is allowed to skip the frozen acceptance regimes or required service-level evidence"
+ ],
+ "refusal_point": "Discards a correctness-breaking, worse or inconclusive change instead of integrating it, preserves no-change when candidates cannot be separated, does not count a shortcut outside the authorized semantics as an optimization, and does not let a successful probe qualify a broken final state"
+ },
+ "recommendations": [
+ "Before generating a candidate, consult the casebook, the negative record and the target stage's Amdahl contribution.",
+ "Record every worse or inconclusive result with hypothesis, metric, context, evidence level and failure type, and derive a search constraint from it.",
+ "When about three experiments on one hypothesis return inconclusive from measurement variance, fix the protocol or move to a different hypothesis.",
+ "Before reporting, restore and verify the best fully valid incumbent through its canonical launcher."
],
"limitations": [
"Source-derived candidate; no local efficacy validation.",
@@ -58,5 +66,25 @@
"../performance-ceiling-analysis/SKILL.md",
"../../repositories/vllm/audit-vllm-kv-lifetimes/SKILL.md"
],
- "consumption_mode": "packaged"
+ "consumption_mode": "packaged",
+ "sources": [
+ {
+ "type": "paper",
+ "id": "usenix:osdi24:zhong-yinmin",
+ "version": "OSDI 24 proceedings, July 2024, pages 193-210",
+ "url": "https://www.usenix.org/conference/osdi24/presentation/zhong-yinmin",
+ "sections": [
+ "3.3",
+ "4.1-4.2",
+ "6.1-6.5",
+ "7",
+ "Appendices A-C"
+ ],
+ "readset": "references/distserve-phase-placement/source-record.json"
+ },
+ {
+ "type": "unknown",
+ "scope": "Pre-existing sources remain in their original procedure/readsets; this amendment does not normalize that inventory"
+ }
+ ]
}
diff --git a/src/inferrsi/data/skill_library/methods/plan-activation-residency/metadata.json b/src/inferrsi/data/skill_library/methods/plan-activation-residency/metadata.json
index 7cf649339..eafcb232c 100644
--- a/src/inferrsi/data/skill_library/methods/plan-activation-residency/metadata.json
+++ b/src/inferrsi/data/skill_library/methods/plan-activation-residency/metadata.json
@@ -7,6 +7,27 @@
"origin": [
"external"
],
+ "applicability": {
+ "scope": "Proposing training activation and gradient memory reductions when pressure appears to come from live tensors or allocator-reserved memory, by distinguishing lifetime reordering, address packing and selective rematerialization; covers activation checkpointing for recomputation, not persistent recovery checkpoints",
+ "prerequisites": "The packaged evidence and limits are read before citing paper results; installing or running a solver or framework, modifying allocation or checkpoint policies, reproducing workloads or provisioning hardware needs separate authorization",
+ "input_contract": "The actual forward/backward/update DAG, shapes and precision, tensor sizes and consumer lifetimes; resident versus reserved peak bytes; operator, workspace and temporary-output peaks; batch and accumulation policy; optimizer and collective barriers; representative compute and transfer costs; whether the goal is fit, step time or time to a fixed training objective",
+ "produced_artifact": "A preferred candidate, simpler fallback, resource ledger, critical-path prediction, assumptions, counter-signals and unresolved-evidence list, with source claims kept apart from proposed decisions, plus a proposed matched comparison; a reviewable proposal, not an implementation",
+ "failure_modes": [
+ "Fragmentation and live-byte excess are not distinguished at the same phase",
+ "OLLA address packing is read as tensor placement across CPU/SSD tiers",
+ "Reordering moves a release or update past a dependency such as shared weights, accumulation, gradient clipping, optimizer state or collectives",
+ "A static packing plan ignores alignment, aliasing, workspace, asynchronous overlap or shape variability",
+ "Added recomputation consumes the compute window that previously hid transfers",
+ "Checkmate's rounding is taken to guarantee the memory cap, or its 10% allowance is used as a portable safety margin",
+ "A solver timeout is read as proof of infeasibility"
+ ],
+ "refusal_point": "Returns the missing classification instead of a preferred checkpoint interval when only total OOM memory is known, and escalates when a legal dependency graph, memory margin or correctness model is unavailable"
+ },
+ "recommendations": [
+ "Where ordering is legal, try earlier last-use and gradient release before adding computation or transfer.",
+ "Check dependency validity and memory feasibility separately for an approximate schedule.",
+ "Account for planning cost over the expected number of repeated steps."
+ ],
"limitations": [
"Source-derived candidate; no local efficacy validation.",
"Original typed source inventory is not normalized; retain source identities and limits in the procedure/readsets."
diff --git a/src/inferrsi/data/skill_library/methods/skill-evolution-playbook/metadata.json b/src/inferrsi/data/skill_library/methods/skill-evolution-playbook/metadata.json
index b6b899a91..2363a96d2 100644
--- a/src/inferrsi/data/skill_library/methods/skill-evolution-playbook/metadata.json
+++ b/src/inferrsi/data/skill_library/methods/skill-evolution-playbook/metadata.json
@@ -13,6 +13,28 @@
"external"
],
"empirical_validation": "not_run",
+ "applicability": {
+ "scope": "Design choices for evolving agent skills from execution traces, drawn from Trace2Skill and SkillOpt: when a trajectory should update a skill, how conflicting lessons are handled, how negative evidence is preserved, how large an edit should be and when skills split or merge; for designing or operating a skill evolution pipeline",
+ "prerequisites": "A final-test partition kept outside edit selection and the incumbent skill retained for rollback; changing a skill's scope first needs explicit review of its identity and consumers",
+ "input_contract": "Execution trajectories or trace IDs with the diagnosed causal failure and the successful behavior an amendment must preserve; the incumbent skill and its revision lineage; held-out evaluation results for candidate edits",
+ "produced_artifact": "A scoped case or amendment proposal, or a consolidated candidate edit, with rejection records and the versioned revision lineage; a candidate for independently evaluated acceptance, not a replacement decision or an efficacy claim",
+ "failure_modes": [
+ "An anecdote, or agreement among analysts, is taken as efficacy",
+ "Consolidation output is treated as a replacement decision instead of a candidate awaiting independent acceptance",
+ "A tie or uncertain result is reported as improvement",
+ "A conflict is resolved by voting for whichever outcome appeared more often in the pool",
+ "A rejection is applied to every future use of an idea instead of the comparison it constrains",
+ "Paper-specific section counts or length budgets are used as universal defaults, or a skill is split on length alone",
+ "A stronger task solver is assumed to be a stronger improver, or replay rankings and private inner scores are read as outer generalization"
+ ],
+ "refusal_point": "Records a conflict that conditioning cannot resolve, where the same context produces different outcomes across runs, as INCONCLUSIVE and flags it for human review instead of choosing a lesson; claims no universal minimum trace count, paper-to-code equivalence or local efficacy"
+ },
+ "recommendations": [
+ "Condition a conflicting lesson on its context (when X, do A; when Y, avoid A) before choosing one side.",
+ "Store each rejection record with its evaluation evidence, outside the deployed task procedure, and give it a retirement condition.",
+ "Bound the target passage and intended behavior before editing, inspect the actual diff and preserve rollback state.",
+ "Prefer an amendment when the existing skill owns the workflow; create a new candidate only for a genuinely distinct capability."
+ ],
"limitations": [
"Source-derived candidate; no local efficacy validation.",
"Original typed source inventory is not normalized; retain source identities and limits in the procedure/readsets."
diff --git a/src/inferrsi/data/skill_library/methods/training-input-pipeline/metadata.json b/src/inferrsi/data/skill_library/methods/training-input-pipeline/metadata.json
index d5552e7bb..3a92cc24c 100644
--- a/src/inferrsi/data/skill_library/methods/training-input-pipeline/metadata.json
+++ b/src/inferrsi/data/skill_library/methods/training-input-pipeline/metadata.json
@@ -15,6 +15,28 @@
"external"
],
"empirical_validation": "not_run",
+ "applicability": {
+ "scope": "Diagnosing training input stalls and choosing bounded queue, shard-cache and cache-ownership changes from DALI, Mosaic Streaming and FFCV, while preserving execution-mode gates, sample ordering and per-host memory budgets and separating remote downloads, local storage reads and host/device transfers",
+ "prerequisites": "The pinned readset and its license constraints are read; reading authorizes no commands, cache manipulation or experiment",
+ "input_contract": "Dataset format, sample-size distribution, workers and ranks per host, ordering/shuffle, local-cache state and executor mode; a per-host memory budget that includes checkpoint staging and runtime offload; timings that separate remote fetch, decompression/decoding, transformation, batch readiness and device transfer",
+ "produced_artifact": "One proposed queue, shard-cache or cache-ownership change with its source-backed rule and boundary, the evidence proposed to confirm it, and a verification contract for a separately authorized comparison; no input-throughput, model-quality or hardware result",
+ "failure_modes": [
+ "Low accelerator utilization is taken to identify which producer stage is late",
+ "Missing legacy DALI statistics are read as zero use, or the execution mode is switched silently to obtain them, which changes the comparison",
+ "DALI run() returning is read as output-ready timing instead of the outputs() synchronization boundary",
+ "Mosaic predownload is treated as a byte cap or request count, or per-worker sample targets are multiplied into an exact network-byte prediction",
+ "Look-ahead is increased when the local cache cannot retain the working set, so evicted shards are downloaded again",
+ "cache_limit is read as a bound on total process RAM, or the constructor minimum as an optimum or a guarantee against thrashing",
+ "An FFCV memory map is taken as proof that its pages are resident"
+ ],
+ "refusal_point": "Rejects changes that violate sample coverage or the stated reproducibility contract even if batches arrive faster, does not report queue depth alone as a performance result, and draws no NPU acceleration claim from the CUDA-oriented DALI source"
+ },
+ "recommendations": [
+ "Record dataset format, sample sizes, workers per host, ordering, cache state and executor mode before proposing one change.",
+ "Increase Mosaic look-ahead only when readiness starvation is observed and cache headroom exists; otherwise address shard and working-set residency first.",
+ "Check the exact DALI configuration surface before use, because the constructor and deserialization paths use different queue keys.",
+ "In a later comparison, separate warm-cache throughput from cold retrieval and initialization, and report step time with input wait, host/HBM peaks and actual I/O."
+ ],
"limitations": [
"Source-derived candidate; no local efficacy validation.",
"Original typed source inventory is not normalized; retain source identities and limits in the procedure/readsets."
diff --git a/src/inferrsi/data/skill_library/methods/training-resource-budget/metadata.json b/src/inferrsi/data/skill_library/methods/training-resource-budget/metadata.json
index f1ac78396..089b91e4f 100644
--- a/src/inferrsi/data/skill_library/methods/training-resource-budget/metadata.json
+++ b/src/inferrsi/data/skill_library/methods/training-resource-budget/metadata.json
@@ -15,6 +15,28 @@
"external"
],
"empirical_validation": "not_run",
+ "applicability": {
+ "scope": "Choosing training offload and resource-budget changes from pinned DeepSpeed configuration, attributed expert diagnosis and a historical MindSpeed case, balancing accelerator capacity, CPU work, pinned RAM and local storage while keeping host/device transfer exposure distinct from collective network bottlenecks",
+ "prerequisites": "The source boundaries are read; garbage-collection or affinity changes, privileged memlock changes and delayed parameter updates each need separate approval, and the skill authorizes no newest-wheel installation, scheduler-priority change, benchmark or failure injection; historical MindSpeed reuse requires compatible allocator symbols, tracked swappable storage, sufficient host capacity and verified device/NUMA/CPU mapping, not merely the presence of npu-smi and lspci",
+ "input_contract": "State attributed to parameters, optimizer state, gradients or activations with its current memory tier, lifetime and next consumer; whether the failure is allocation capacity or throughput; a per-host budget for input buffers, checkpoint snapshots, offloaded state, OS cache and CPU workers with the actual ranks sharing the host; the host/device link recorded apart from collective traffic; the exact source version and enabled execution mode",
+ "produced_artifact": "A candidate decision record per proposal: the observed limiting interval and owning state or resource, source version and execution mode, capacity released and resources consumed with units and uncertainty, overlap readiness and mutation boundaries with expected contention, and the fixed-workload step and correctness evidence required for later validation",
+ "failure_modes": [
+ "A change that lets the model fit is presented as a speedup, although it can slow each step",
+ "A tensor copied to CPU is taken as an InfiniBand/RoCE bottleneck, or CPU/NVMe offload is counted as network evidence",
+ "Parameter max_in_cpu is read as a byte or total-host-memory cap instead of an element count",
+ "NVMe optimizer placement is assumed to remove CPU optimizer demand",
+ "The parameter buffer-pool count is transplanted to the optimizer, or all optimizers are assumed to hold identical states",
+ "Configuration presence or schema documentation is taken as evidence that a knob or overlap actually ran",
+ "Ordinary overlap is taken as permission for delayed, stale parameter updates"
+ ],
+ "refusal_point": "Claims no hardware validation, measured gain or current NPU compatibility, keeps source support and extracted coverage separate from measured improvement, and does not reuse the MindSpeed headline gain; historical source observations do not establish current-stack compatibility, whole-model correctness or measured benefit"
+ },
+ "recommendations": [
+ "Attribute the state and separate a capacity failure from a throughput bottleneck before choosing an offload surface.",
+ "Verify the consumer of buffer_size before calculating bytes or tuning it, and measure the executing path before claiming a knob ran.",
+ "Compare the whole step's critical path, not host capacity alone.",
+ "Use plan-activation-residency, choose-parallel-training-plan and assess-split-backward-schedule for live-range, parallel-plan and backward-ordering questions."
+ ],
"limitations": [
"Source-derived candidate; no local efficacy validation.",
"Original typed source inventory is not normalized; retain source identities and limits in the procedure/readsets."
diff --git a/src/inferrsi/data/skill_library/methods/weight-loading-optimization/metadata.json b/src/inferrsi/data/skill_library/methods/weight-loading-optimization/metadata.json
index 777bf518b..5efd71b48 100644
--- a/src/inferrsi/data/skill_library/methods/weight-loading-optimization/metadata.json
+++ b/src/inferrsi/data/skill_library/methods/weight-loading-optimization/metadata.json
@@ -77,6 +77,32 @@
}
],
"empirical_validation": "not_run",
+ "applicability": {
+ "scope": "Diagnosing the serving-level weight load path from storage to device and choosing changes to it when startup or loading latency dominates: locating the cost by source, format, placement and duplication, accounting once-per-process loading and warmup apart from per-request cost, verifying sharded downloads, restricting device-side prefetch to profiled memory-bound hot spots and ordering CPU-tier offload; also covers source-scoped disk-loading modes, pinned staging, synchronous checkpoint save semantics, fastsafetensors broadcast/scatter/coalesced delivery and the separation of model preparation from serving. Request-time host overhead belongs to host-overhead-profiling, DCP asynchronous save performance to dcp-async-save-performance and distributed recovery validity to checkpoint-recovery-consistency",
+ "prerequisites": "Flag names and API entry paths are re-verified against the pinned source revisions, and torch_npu.npu_prefetch details against the target CANN version, before they are relied on; card state is checked before launch and process placement fixed before measuring, through ascend-device-health and os-process-placement; the Diffusers enable_model_cpu_offload path requires accelerate",
+ "input_contract": "The current load path's byte source, file format, host-first or direct device placement and duplicated copies across processes or instances; startup timings for loading and warmup with byte and access counts per segment, and the expected warmup hit criteria; the safetensors index weight_map and the downloaded shard set; a profile showing memory-bound hot spots and the adjacent compute or communication segments; free device memory and whether an OOM occurs during construction or inference; for cross-rank delivery, the file-to-loader-group-rank map, tensor names, shapes, dtypes and call order, the required replicated or sliced output, the model's expected shard layout and the collective backend",
+ "produced_artifact": "The located load-path cost on the four dimensions with its matching lever, a startup ledger that keeps fixed overhead apart from per-request cost, a structural shard-completeness result naming the missing shards, a prefetch proposal behind an enable_prefetch switch defaulting to False or a recorded decision not to prefetch, an offload candidate order, the recorded loading mode, a distributed delivery choice with its proposed verification, and a preparation/serving record of serialized and rebuilt state, live resource owners and ABI/prepared-contract versions; proposed changes and verification targets, not a measured startup or delivery gain",
+ "failure_modes": [
+ "Loading and warmup are amortized into per-request latency, or a host segment without timing points is treated as costing nothing",
+ "Warmup is accepted on total time while mismatched warmup shapes have warmed a different set of cache entries",
+ "In-flight .incomplete shards are re-downloaded as corrupt, or gap-free shard numbering is accepted without matching the index weight_map",
+ "Prefetch is forced without a profiled memory-bound hot spot or a non-memory-bound dependency window",
+ "Resolution or CFG tier-downs are applied to a construction-phase OOM, or offload is adopted when compute already hides the transfer",
+ "A free-memory probe runs on an mmap-backed load, or a run whose first mlock failure latched with only a warning is treated as fully resident",
+ "Disk counters that include cache and pipes are read as physical writes, or a run without a page-cache drop is reported as a cold start",
+ "The ceil-block scatter partition is used without checking it against the model's expected shard layout, or ranks run different orders or plans so their collectives mismatch",
+ "A faster H2D copy, local clone or rank-local filtered read is counted as a cross-rank communication gain"
+ ],
+ "refusal_point": "Withholds a load-path win until the verification checklist holds and keeps the distributed-delivery checks as verification targets rather than results; does not force prefetch without a safe dependency window and disables it when neither a smaller size nor a moved position removes a regression; records unsupported save behaviors as unknown; rejects mismatched execution ABI or prepared-contract versions instead of editing their tags; and implies no Ascend support from the llama.cpp or fastsafetensors mechanisms",
+ "hardware": "NPU for the device-side prefetch API (torch_npu.npu_prefetch, checked against the target CANN version) and the Diffusers offload playbook; the prefetch and fastsafetensors methodology transfers while the fastsafetensors and llama.cpp APIs do not, and cross-rank completion follows each collective backend's own wait contract"
+ },
+ "recommendations": [
+ "Report preparation/startup/readiness, warmups and measured work as separate numbers, with load latency split into init, header, io and get phases, the cold-cache protocol stated and the effective loading mode identical across compared runs.",
+ "Place each prefetch behind an enable_prefetch switch defaulting to False, start max_size at 50% of the theoretical size, verify a single operator first and step up to 0.7 and then 1.0 only after verified wins.",
+ "Try offload candidates in order of change surface: batch size, resolution, CFG/feature flags, CPU or layered offload, tensor-parallel degree, then activation checkpointing; re-estimate the un-overlapped transfer share after adopting each tier.",
+ "Record whole-set atomic publication, asynchronous checkpointing, multi-rank save coordination and crash durability as unknown for the fastsafetensors saver instead of assuming them.",
+ "Before adopting coalesced delivery, compare it disabled and enabled with identical checkpoint, rank mapping, backend, topology, dtype, cache state, chunk/queue budgets and consumer retention, then measure complete startup under the same accounting protocol."
+ ],
"limitations": [
"Source-derived candidate; no local efficacy validation.",
"Normalized provenance covers the existing llama.cpp loading slice at 3e390af18de1131dda27913c00511725b4274c8a and the fastsafetensors distributed-delivery slice at 219e2561a41ede2d5d8eba611c248bb0effbf580; other inherited claims remain unnormalized with original identities and limits retained.",
diff --git a/tests/test_research.py b/tests/test_research.py
index f1dfa0951..ef927a995 100644
--- a/tests/test_research.py
+++ b/tests/test_research.py
@@ -29,7 +29,12 @@ def test_h3_retrieves_framework_and_task_knowledge_with_source_boundaries():
assert len(identities) == len(set(identities)) <= 12
assert "inferrsi.common.skills.lazy_evaluation_stream_lifetimes" in identities
assert set(result["profile"]["knowledge_ids"]) <= set(identities)
- assert any("diffusion" in item["record"]["scenario"] for item in items)
+ task_first = search_knowledge("diffusion", limit=1)[0]["id"]
+ assert any(
+ item["record"]["id"] == task_first
+ and {"group": "task", "query": "diffusion"} in item["matched_queries"]
+ for item in items
+ )
assert any(
match["group"] == "modality" and match["query"] == "video"
for item in items for match in item["matched_queries"]