Skip to content
Draft
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 6 additions & 0 deletions docs/reference-matrix.md
Original file line number Diff line number Diff line change
@@ -1,5 +1,11 @@
# InferRSI Reference Matrix

## EPD cold-serving allocation case

| Source | Inspected scope | Extraction and qualification boundary |
| --- | --- | --- |
| Kelvin-Al/epd-provisioning cold comparison and integer allocator | Twelve published files cover the allocator, original comparison and position-balanced follow-up. The [typed source record](../src/inferrsi/knowledge/data/skill_library/casebooks/inference/inference-optimization-casebook/cases/epd-cold-provisioning.json) owns both immutable source revisions and the rehashed file identities. | [EPD cold-provisioning case](../src/inferrsi/knowledge/data/skill_library/casebooks/inference/inference-optimization-casebook/cases/epd-cold-provisioning.md). External author-reported four-H800, preprocessed MMMU observation. Formal acceptance INCONCLUSIVE. Answer quality and sustained service remain unqualified. |

**Branch:** `opus/meta-skill-bootstrap` | **Integration date:** 2026-10-16 | **Comparison anchor:** `d96ff8d3`

Current skill links resolve to the complete categorized
Expand Down
6 changes: 6 additions & 0 deletions docs/searchspace-matrix.md
Original file line number Diff line number Diff line change
@@ -1,5 +1,11 @@
# InferRSI SearchSpace Matrix

## EPD cold-serving configuration precedent

| Resource layer | Lifecycle use | Owner | Qualification |
| --- | --- | --- | --- |
| Accelerator / Host | Diagnose / Configure / Verify stage instance counts using independent capacity estimates and matched serving windows | [EPD cold-provisioning case](../src/inferrsi/knowledge/data/skill_library/casebooks/inference/inference-optimization-casebook/cases/epd-cold-provisioning.md) | External author-reported four-H800 cold MMMU observation. Formal acceptance INCONCLUSIVE. Transfer and queue behavior remain deployment-specific. |

**Branch:** `opus/meta-skill-bootstrap` | **Integration date:** 2026-10-16 | **Comparison anchor:** `d96ff8d3`

Current skill links resolve to the complete categorized
Expand Down
2 changes: 1 addition & 1 deletion docs/skill-coverage.md
Original file line number Diff line number Diff line change
Expand Up @@ -4,7 +4,7 @@ Generated from the registry, SKILL.md frontmatter and the two maintained
matrices. **Do not edit by hand**; regenerate after any registry, skill,
matrix or frozen-plan change.

- Input digest: `sha256:93ad79f2b36af340d3c1c819dcb2f9b651d0f6078913aa726bd9ee1f01e9df54`
- Input digest: `sha256:1b437e45037eced08be2653c0577e1d756c1bf12ac137d074da9a4bc3c005fd1`
- Regenerate: `python tools/build_skill_coverage.py`
- Verify without writing: `python tools/build_skill_coverage.py --check`

Expand Down
Original file line number Diff line number Diff line change
@@ -1,7 +1,7 @@
{
"schema_version": 1,
"repository": "https://github.com/InferMatrixLab/InferRSI",
"source_revision": "217d0b386d9c8d53bc06fc0d485f2e99722d5847",
"source_revision": "55cb1c5a945bae7e9dbd211b9f6662af143472dd",
"registry": {
"path": "src/inferrsi/knowledge/data/skill_library/skill-index.json",
"sha256": "c6c235483034dfce3c0a4ec3bf7d68046ea2cf8c933f069833ecaef65cdf2d70"
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -594,11 +594,13 @@
"description": "How to structure and use a casebook of past optimization decisions as a decision-support tool. Extracts the casebook pattern from TensorRT-LLM's perf-optimization-casebook skill. Use when building a record of optimization experiments or when deciding which optimization to try next based on prior art.\n"
},
"files": {
"SKILL.md": "cec75f3a721dd13e6913b597d5373833a14c39d9048951a30c4864dd2c49aac8",
"SKILL.md": "2f6061305991384727d6d0d4519ad9b230e19c49bb784e0bdc2ae5306d66e27d",
"cases/admission-drain-occupancy.json": "4f8544292b2cc3391f4d4d75994cb81fdd295eff99e55b7e60374c6c122de09e",
"cases/admission-drain-occupancy.md": "e88bbac5d253a8e7e3a3d5e21047f23f2970c11ac23887d4303ab94fca52dd27",
"cases/drama3-rejected-directions.json": "a2dc5db8aba62081c5743b12f854e74ae500c53bd57e2486260d0b5ae478638a",
"cases/drama3-rejected-directions.md": "6f143bc363a3290fc008c08e57cb075f1776d1c8cca8a6a59297d5b1fbbf61bf",
"cases/epd-cold-provisioning.json": "68e471335193e2027779e1377ee22e3384f6c3c66034f0a7b7631e1b67d2e245",
"cases/epd-cold-provisioning.md": "9d1931e74585a57d824f9c48003aa1ad155d9d7c7d0d605f17c7d807e81233c4",
"cases/qwen-ascend-rejected-directions.json": "b2c1808919ae525efb82148a24b465459799422fefc6d420891d7dbeb36cd31a",
"cases/qwen-ascend-rejected-directions.md": "be396983812262515899a3e2ba4f01b23ff92e1b836881882c7580372b8ae552",
"cases/qwen-b300-rejected-directions.json": "90e3cdca085e2d122e648aae527bd6a61de6d4fddb66c393ca25ac4541cd9122",
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -45,6 +45,7 @@ metadata:
"Source-derived candidate; no local efficacy validation.",
"Original typed source inventory is not normalized; retain source identities and limits in the procedure/readsets."
],
"references": ["cases/epd-cold-provisioning.md"],
"provenance_readsets": [],
"consumption_mode": "packaged"
}
Expand Down Expand Up @@ -123,6 +124,11 @@ Source: perf-optimization-casebook @ f33e0e5f, §how-to-use.

## Recorded diagnostic options

The [EPD cold-provisioning case](cases/epd-cold-provisioning.md) records
integer E:PD GPU allocation, applicability and matched cold MMMU comparisons
on four H800 GPUs, including a position-balanced follow-up. Its source record retains external author-reported scope,
published artifact hashes and INCONCLUSIVE formal acceptance.

Current owner-local records retain rejected directions alongside their limits:

- [Qwen B300](cases/qwen-b300-rejected-directions.md): full-image rejection,
Expand Down
Original file line number Diff line number Diff line change
@@ -0,0 +1,116 @@
{
"id": "case.epd-cold-provisioning",
"kind": "case",
"origin": [
"external"
],
"state": "author_reported",
"content": "epd-cold-provisioning.md",
"required_resources": [
"epd-cold-provisioning.json"
],
"sources": [
{
"type": "git",
"url": "https://github.com/Kelvin-Al/epd-provisioning/blob/3ed184e3a83229806a9fa1adb2bc75f1744c6518/planner.py",
"commit": "3ed184e3a83229806a9fa1adb2bc75f1744c6518",
"paths": [
"planner.py",
"experiments/ec-cache/cold_sweep.py",
"experiments/ec-cache/audit.py",
"experiments/ec-cache/results/cold/plan.json",
"experiments/ec-cache/results/cold/results.json",
"experiments/ec-cache/results/failed-cold-attempt/status.json",
"experiments/four-gpu/results/mmmu/plan.json"
],
"sha256": {
"planner.py": "6bf4ea101fc12da1ca9166121a09f362b22feff3bf45ccf94b037dec1d93ce68",
"experiments/ec-cache/cold_sweep.py": "fcffb83d0351bf36387a971622edd92667838112fc737f68f1c6acff46082853",
"experiments/ec-cache/audit.py": "56be9d553ba04ac1a279035cea95eabe4892b5225eaabb9d604b20d057350816",
"experiments/ec-cache/results/cold/plan.json": "88941e0f5969ce25f94774a0a4c95c6204b9bc0d48669c4d18547017104c2542",
"experiments/ec-cache/results/cold/results.json": "dc5c1ecb7d3f92708a3ea7078d65ab575854a350235d5d14d7257bda956f4f2b",
"experiments/ec-cache/results/failed-cold-attempt/status.json": "33f44ead48d151e3b2b4527e12c2f89085651ed332ddf5ffd8aac7223c538806",
"experiments/four-gpu/results/mmmu/plan.json": "93a5b729357464f9ae2d0b62c120e5e5f22ff2e82db92a3a6afa10eae773c7d5"
},
"integrity_status": "published_bytes_retrieved_and_rehashed",
"issue": "https://github.com/Kelvin-Al/epd-provisioning/issues/1",
"evidence_url": "https://github.com/Kelvin-Al/epd-provisioning/blob/3ed184e3a83229806a9fa1adb2bc75f1744c6518/experiments/ec-cache/results/cold/results.json"
},
{
"type": "git",
"url": "https://github.com/Kelvin-Al/epd-provisioning/blob/a0037f7ddc4b1af5f59fbef0a86b2a5627b69feb/experiments/ec-cache/results/cold-order-followup/plan.json",
"commit": "a0037f7ddc4b1af5f59fbef0a86b2a5627b69feb",
"paths": [
"experiments/ec-cache/order_followup/cold_sweep.py",
"experiments/ec-cache/order_followup/audit.sbatch",
"experiments/ec-cache/results/cold-order-followup/plan.json",
"experiments/ec-cache/results/cold-order-followup/results.json",
"experiments/ec-cache/results/cold-order-followup/interrupted-results.json"
],
"sha256": {
"experiments/ec-cache/order_followup/cold_sweep.py": "f031d7c6f354461310eecdf8a74798c31615c4f1aad806187671ac3ba147e32f",
"experiments/ec-cache/order_followup/audit.sbatch": "40c3064acd59f1255c62ed53effb4a7be834d1dbc32546eff59387133288d128",
"experiments/ec-cache/results/cold-order-followup/plan.json": "c7fe3d8cbd2cf4cabccfe38bf9a7428ab980af591ed76e423128a2031b42fb95",
"experiments/ec-cache/results/cold-order-followup/results.json": "dc4f3d4f492f60171a5124e0d97a1a52c8dd6134072251666f55b7ff157a9c24",
"experiments/ec-cache/results/cold-order-followup/interrupted-results.json": "834d616d8b1dbd04bb08a28a8e4b19a9128692b342b8714062d26a9c8a248456"
},
"integrity_status": "published_bytes_retrieved_and_rehashed",
"issue": "https://github.com/Kelvin-Al/epd-provisioning/issues/1",
"evidence_url": "https://github.com/Kelvin-Al/epd-provisioning/blob/a0037f7ddc4b1af5f59fbef0a86b2a5627b69feb/experiments/ec-cache/results/cold-order-followup/results.json"
}
],
"applicability": [
"Cold E+PD serving with one GPU per effective stage instance",
"Four H800, Qwen2.5-VL-3B-Instruct, prepared vLLM 0.33.0 overlay, filesystem EC connector",
"720 single-image preprocessed MMMU validation requests, fixed 64 output tokens, zero embedding reuse"
],
"tags": [
"EPD",
"GPU-allocation",
"capacity-planning",
"cold-serving",
"MMMU",
"H800",
"prefill-decode",
"queueing"
],
"limitations": [
"External author report. InferRSI campaign acceptance INCONCLUSIVE.",
"Published JSON and code hashes checked. Raw worker traces, image bytes, model weights and container were not independently replayed or verified.",
"Model weight revision absent. Overlay identity is partly represented by recorded script digests.",
"Original reverse-order rounds fixed 2E:2PD in the middle. The nine-wave follow-up balances all positions with one observation per split and position, with limited predecessor and variability coverage.",
"One offered rate in each single-node comparison, with a central arrival window of about 12.5 seconds. The interrupted six-wave follow-up is retained separately from the complete rerun.",
"700 of 720 images resized into calibration range. Fixed-length completion does not evaluate MMMU answer accuracy.",
"First cold attempt failed on inconsistent media keys. Complete rerun uses corrected keys and explicitly forwards seed 9197."
],
"attribution": "Kelvin-Al/epd-provisioning, Apache-2.0 source and published cold-serving observations.",
"findings": [
{
"direction": "Choose integer E:PD counts using independently estimated stage capacity",
"outcome": "inconclusive",
"reason": "The minimum of E-pool and PD-pool capacity selected 2E:2PD. It ranked first in the two original cold rounds and all three position-balanced follow-up rounds, with 26.04, 26.44 and 26.67 completions/s at the first, second and third positions.",
"revisit_when": "Re-estimate stage rates and test alternatives when model, workload, hardware, transfer path or worker configuration changes.",
"keywords": [
"EPD",
"integer allocation",
"stage capacity",
"2E:2PD"
],
"scope": "regime_bound"
},
{
"direction": "Use actual visual work and paired queue observations to diagnose the limiting stage",
"outcome": "caveat",
"reason": "Each wave encoded 720 images with zero EC and resident hits. One E completed about 23 requests/s with queueing, while two E workers kept up with arrivals and two PD workers completed about 26 requests/s.",
"revisit_when": "Align encoder, decoder and client counters to the same arrival interval before inferring bottlenecks.",
"keywords": [
"visual forward",
"EC",
"cold cache",
"encoder queue",
"PD queue"
],
"scope": "regime_bound"
}
]
}
Loading
Loading