From 3140fba53dd8c0264d32dd17de2fb5b0a4882ca0 Mon Sep 17 00:00:00 2001 From: Dong1017 <55723162+Dong1017@users.noreply.github.com> Date: Sun, 25 Oct 2026 13:10:03 +0900 Subject: [PATCH] docs: connect TensorRT-LLM methodology and disaggregation guidance Expose the retained source scopes and shared measurement assumptions through existing navigation entrypoints. Signed-off-by: GUOGUO Agent-assisted-by: Codex / gpt-6.1-sol [effort=xhigh] (author) --- docs/reference-matrix.md | 10 +++++++++ docs/searchspace-matrix.md | 10 +++++++++ src/inferrsi/data/knowledge_packs/README.md | 6 +++++ src/inferrsi/data/skill_library/README.md | 25 +++++++++++++++++++++ 4 files changed, 51 insertions(+) diff --git a/docs/reference-matrix.md b/docs/reference-matrix.md index 89ea0fee2..bff2d4d75 100644 --- a/docs/reference-matrix.md +++ b/docs/reference-matrix.md @@ -94,6 +94,16 @@ in (evidence/claim discipline, benchmark hygiene, RunSpec) — not an external mining source. The local sibling clone at `codex/skill-eval-bootstrap` (82218b4) is a **different fork lineage**; see §2. +### TensorRT-LLM retained source responsibilities + +The [source and shared-assumption map](../src/inferrsi/data/skill_library/README.md#tensorrt-llm-source-and-responsibility-routing) +connects the historical `f33e0e5fd` skill-document row to the separately pinned +`c9b18eefa731dc59b0091bdfff6d890aa9d1e98e` PD/KV-transfer peer card. +It distinguishes startup and request-time denominators, local H2D and network +transport, and the workload/layout/resource controls needed for a PD comparison. +Original prefix precision, missing methodology inspection dates, mixed-source +attribution, owner readsets and candidate/imported limits remain unchanged. + ### Reviewed systems-paper mechanisms `mined` below means pack-mediated extraction into the linked candidate artifacts. diff --git a/docs/searchspace-matrix.md b/docs/searchspace-matrix.md index 3f1428942..12cc5c7c7 100644 --- a/docs/searchspace-matrix.md +++ b/docs/searchspace-matrix.md @@ -313,6 +313,16 @@ English procedures. They do not qualify every backend listed by their upstream projects. Main-only contribution, dependency and peer material remains at its original owners rather than duplicated under local skills. +## TensorRT-LLM shared assumptions + +Use the [methodology and disaggregation map](../src/inferrsi/data/skill_library/README.md#tensorrt-llm-source-and-responsibility-routing) +for the TensorRT-LLM Accelerator, RAM/Host, SSD/Storage, Communication/Network +and OS/System entries. Its evidence row fixes the comparison contract; the five +resource rows assess host exposure, startup/request units, transport/layout and +queue costs, kernel correctness and placement controls against the existing +peer card. The retained methodology and PD pins are separate source scopes; +this adds no measured/green cell, new runtime mechanism or compatible-stack claim. + ## Issue #27: Drama3 NPU readiness (guidance only, no new coverage) The [Drama3 NPU readiness supplement](drama3-npu-readiness.md) is a scoped task diff --git a/src/inferrsi/data/knowledge_packs/README.md b/src/inferrsi/data/knowledge_packs/README.md index 12415f8e5..52ca1114f 100644 --- a/src/inferrsi/data/knowledge_packs/README.md +++ b/src/inferrsi/data/knowledge_packs/README.md @@ -15,6 +15,12 @@ Cache-DiT, for example, can also be a direct dependency of vLLM-Omni. Its entry is classified as a peer technique for this initial extraction; reclassify an entry when the learning use case changes instead of duplicating its claims. +For TensorRT-LLM, the [methodology and disaggregation map](../skill_library/README.md#tensorrt-llm-source-and-responsibility-routing) +routes the existing mixed-source performance methods alongside the +[PD/KV-transfer card](peer/tensorrt_llm/cuda/llm/prefill-decode-disaggregation.json). +It assesses shared measurement controls and separate startup, host and network +costs while retaining the card's source revision and transfer-hypothesis limits. + ## Storage and compatibility ```text diff --git a/src/inferrsi/data/skill_library/README.md b/src/inferrsi/data/skill_library/README.md index 004cac075..4d12d5ceb 100644 --- a/src/inferrsi/data/skill_library/README.md +++ b/src/inferrsi/data/skill_library/README.md @@ -65,6 +65,31 @@ and server task/slot cancellation. Its pinned source contracts are untested nati existing owners. The [scope and source map](repositories/llama-cpp/audit-llama-runtime-state/SKILL.md#choose-the-scope) links all three responsibilities and their separate source snapshots. +## TensorRT-LLM source and responsibility routing + +TensorRT-LLM's retained methodology row covers **skill documents**, not an +engine-runtime audit. It records prefix `f33e0e5fd`; procedures keep +`f33e0e5f`, while the [ceiling sidecar](methods/performance-ceiling-analysis/metadata.json) +retains the longer **prefix** `f33e0e5fd55b`. An original inspection date for +that methodology extraction is not individually recorded. These are partial +source identities, not a verified common runtime version. + +The separate [PD/KV-transfer peer card](../knowledge_packs/peer/tensorrt_llm/cuda/llm/prefill-decode-disaggregation.json) +records `c9b18eefa731dc59b0091bdfff6d890aa9d1e98e`, accessed 2026-10-15. +Its upstream-reported technique remains imported/unverified. The methods below +remain candidate / empirical validation `not_run`, with their other sources +and source notices retained. This cross-reading assesses when to combine the +guidance; it establishes no cross-pin implementation compatibility. + +| Decision / resource | Existing entry | Shared prerequisite or accounting judgment | Keep separate | +| --- | --- | --- | --- | +| Evidence and measurement | [Evidence levels](methods/dense-performance-evidence/SKILL.md#claim-evidence-matching), [headroom inputs](methods/performance-ceiling-analysis/SKILL.md#evidence-to-inspect-first) and [case reuse](methods/inference-optimization-casebook/SKILL.md#adapt-do-not-copy) | Apply the peer card's fixed input/output-length and arrival distributions, same total device budget and SLO workload; report TTFT, TPOT, tail latency, goodput and total GPU/network resources | Analytical/component evidence and a precedent do not establish service gain; re-derive model/hardware-specific values | +| Accelerator | [Kernel verification](methods/kernel-writing-guide/SKILL.md#workflow-verify-before-benchmark) | A kernel change still needs reference parity over target shapes/dtypes and critical-path profiling | Kernel correctness does not establish reduced prefill/decode interference or end-to-end PD benefit | +| RAM / Host | [Serving-trace prerequisites](hardware/cuda/host-overhead-profiling/SKILL.md#prerequisites) and [confirmed host bottleneck](hardware/host/host-overhead-optimization/SKILL.md#confirming-the-bottleneck) | Establish exposed host work from serving traces before attributing a PD regression to CPU preparation | A line profiler locates CPU time; it does not establish that CPU is limiting or measure the KV interconnect | +| SSD / Storage | [Startup accounting](methods/weight-loading-optimization/SKILL.md#2-account-startup-fixed-overhead-separately) | Loading/warmup is once per process; request KV delivery needs its own per-request bytes and timing | Do not combine those denominators; disk-loading mechanics retain their separately attributed sources | +| Communication / Network | [Disaggregation peer card](../knowledge_packs/peer/tensorrt_llm/cuda/llm/prefill-decode-disaggregation.json) | Consider PD with measured prefill/decode interference, compatible KV layouts and qualified NVIDIA GPU transport; include transfer, layout-conversion and queue costs, and verify backpressure, cancellation and failed-worker recovery | Host/device loading overlap and local H2D do not establish this interconnect path; diffusion-stage splitting does not inherit the autoregressive PD cost model | +| OS / System | [Placement discipline](hardware/host/os-process-placement/SKILL.md#1-cpu-affinity-check-then-pin-then-keep-constant) | Record and control affinity, environment and device contention across matched pool comparisons | Placement methodology controls an experiment; it does not implement PD routing or OS-priority scheduling | + ## Engine knowledge Five repository entry skills, 15 focused sub-skills and one research router