Skip to content

Latest commit

 

History

History
86 lines (77 loc) · 15.2 KB

File metadata and controls

86 lines (77 loc) · 15.2 KB
title Documentation Index
slug readme
type index
status canonical
created 2026-04-07
updated 2026-06-05
owner Team 13
scope team-repo
canonical true

Documentation Index

Living, authored documentation for the SmartGridBench project. Everything in this directory is a doc that evolves with the project - domain background, setup guides, architecture notes, methodology. Historical planning records live in ../planning/archive/; current task truth lives in GitHub Issues / Projects and active docs. Frozen deliverables (shipped PDFs, slide decks) live in ../reports/. Historical supporting notes that are no longer live move into archive/. Lower-churn class / mentor / setup references now live under knowledge-base/reference/. Durable closeouts and extracted lessons live under retrospectives/ and knowledge-base/.

Document index

File Purpose Start here if…
project_synopsis.md Cold-start project overview with full domain background, problem statement, team roles, timeline, current status You're new to the project and want the complete picture in ~10 minutes
knowledge-base/reference/2026-04-05_project_reference.md Class requirements, grading rubric, mentor guidance, course context, report templates You need to know what HPML class/Dhaval expects as deliverables
execution_plan.md Task dependency map (Tier 1-5 critical path) + benchmarking operations (async batch workflow, 5-cell experimental grid, role clarifications) You want to know what blocks what, who owns what, and what running experiments actually looks like operationally
runbook.md Canonical end-to-end reproducibility runbook for the infra side — preconditions, first-time setup, submitting benchmark cells, profiling workflow, troubleshooting decision tree, pointers to detailed runbooks You need to stand up the serving / benchmark / profiling pipeline from scratch without verbal help
infra_profiling_serving_brief.md One-page paper-bound fact pack for #43: model IDs + version pins + Slurm run shape + profiling instrumentation + WandB linkage + canonical run IDs + GCP A100 path + known limitations. Verbatim-quotable bullets, no narrative You're filling §3 System Design or the §infra paragraphs of #39 / #40 and need concrete cite-by-job-id facts without rereading every detailed runbook
validation_log.md Canonical log of concrete serve / benchmark / profiling proofs, including run IDs, artifacts, what each run actually proved, and caveats You need the durable record of live validation runs rather than the how-to runbooks
compute_plan.md Phase-by-phase GPU allocation across Insomnia (H100, A6000) and $500/person GCP budget, hardware strategy decisions You need to spin up an environment or pick a GPU for a workload
gcp_fallback.md Emergency GCP A100 spin-up runbook — when to use it, instance selection, env setup, artifact persistence, shutdown, spot preemption handling, budget tracking, known GPU differences from Insomnia Insomnia is down / queue-saturated and you need GPUs now, or you're considering a one-off A100 run
insomnia_runbook.md Verified Insomnia setup notes - Slurm account/partition/QoS, scratch storage, CUDA/cuDNN workarounds, vLLM Python-version gotcha, login-node etiquette, queue tips, foreground-debug recipe You're hitting weird Slurm/CUDA/vLLM behavior on Insomnia and need verified working settings
governance/model_registry.yaml Canonical registry for the current local-vLLM and WatsonX model contracts, including served model names, repo-facing model IDs, runtime pins, and the standardized MODEL_REVISION for the local Llama mirror You need the quick source of truth for which model names/IDs/runtime pins the repo is actually supposed to use
slurm_cheatsheet.md Command-first Slurm reference for submit, watch, estimate start, inspect failures, and historical timing on Insomnia You need the exact command for sbatch, srun, squeue, sacct, scontrol, or scancel without rereading the longer runbook
orchestration_wiring.md Current repo-side orchestration status for Plan-Execute, Agent-as-Tool, and Hybrid, including what is genuinely wired now versus only adapter-ready You need to know what #22 / #62 cover in this repo and what is still upstream or mentor-gated
experiment_matrix.md Sharp statement of the core experiment grid, trial policy, Self-Ask tracking, and which optional follow-on cells are worth adding later You want one defensible answer to "what exactly are we running?" before the matrix sprawls
experiment1_capture_plan.md Capture plan for Experiment 1 (#25): Cell A/B/C config layout, fairness contract, runner requirements, team dependencies, and proposed run sequence through Apr 22 You need to know what blocks the Direct / MCP-baseline / MCP-optimized captures and how the artifacts feed Alex's Notebook 02
lane2_int8_kv_status.md Lane 2 status for #29 (INT8) and #30 (KV-cache) — vLLM 0.19.0 quantization landscape, INT8 deferral rationale + smoke evidence (8979660), chosen KV-cache knob (prefix caching only — fp8 KV tested in smoke 8979532 and deferred due to a vLLM 0.19.0 FA3 kernel constraint under FP16 weights), and recorded Insomnia smoke results You need to know what optimizations Cell C bundles, why INT8 is deferred, or how to read the Lane 2 smoke evidence
failure_analysis_scaffold.md Before/after metric pack for #36: outcome / failure-shape / latency / profiling field list, comparison ledger, export contract, comparison-ready status labels You need the canonical contract for what the #36 rerun lane has to produce so notebook and paper exports stay joinable
failure_taxonomy_evidence.md Failure taxonomy classification for #35: Berkeley categories, decision ladder, evidence schema, populated Apr 22 evidence pass, paper-safe wording guide, classification workflow You need to label observed failures with concrete artifact backing instead of vibes, and you want one evidence row per (run_name, scenario_id, trial_index)
failure_visuals_mitigation.md Visuals scaffold + figure-ready aggregation contract + mitigation ranking rubric + Apr 22 mitigation ranking + promotion gate into #65 / #66 for #64 You need to decide which mitigation goes first into implementation/rerun and what figure each visual should render from
mitigation_recovery_adjudication.md Implementation-ready spec for the retry/replan recovery and explicit fault/risk adjudication mitigation rungs You need to implement or review the follow-on mitigation ladder without creating a full cell-by-mitigation grid
scenario_realism_validation.md Mentor-facing realism-validation pack for the current Smart Grid scenario families, including representative scenarios, known realism gaps, and the concrete questions to send Dhaval You need to sanity-check whether the current Smart Grid scenarios read like believable transformer maintenance work
ps_b_evaluation_methodology.md Validation rubric for comparing generated PS B scenarios against the hand-crafted Smart Grid set, including duplication checks, acceptance thresholds, and explicit circularity handling You need the concrete standard Akshat should apply when validating generated scenarios
neurips_abstract_outline.md Working title list, abstract skeleton, and evidence map for the NeurIPS paper lane You need a prepared outline for the final abstract rather than drafting from scratch under deadline pressure
neurips_draft.md Live NeurIPS paper-writing scaffold for #5 / #39: title, one-paragraph claim, draft abstract, working contribution list, claim ledger, section scaffold, draft prose ready to lift into Overleaf You want the paper effort to move beyond outline-only planning into a real draft surface with reusable section text
neurips_submission_packet.md Deadline-facing NeurIPS control packet for #5 / #39 / #47 / #48: submission surface, exact abstract candidate, claim tiers, current result snapshot, figure queue, and blockers You need the shortest source of truth for what can be submitted now and what still blocks the full paper
neurips_overleaf_transfer_plan.md Repo-to-Overleaf sync plan for the NeurIPS 2026 project: what was copied, what remains caveated, and which metric sources back paper tables and figures You need to keep Overleaf aligned with repo evidence without re-deriving the copy order
final_report_backport_scaffold.md Class final report back-port scaffold for #40: IEEE section map, drift-control rules, figure/table requirements, and conversion checklist from the NeurIPS draft You need to turn the NeurIPS draft into the class final report without inventing a second source of truth
final_presentation_deck.md Slide-by-slide final presentation draft for #44: deck spine, claims, proof objects, result tables, and backup Q&A prompts You need to convert the current paper/report story into the class presentation deck
auto_scenario_generation_runbook.md Runbook for scripts/generate_scenarios.py (#2 prototype): how the generator consumes the support data, output layout under data/scenarios/generated/<batch_id>/, dry-run vs live invocation, promotion path from generated to canonical, scope deferred to #68 scale-up You want to produce a candidate Smart Grid scenario batch or read what a generated scenario file actually contains

| knowledge-base/reference/2026-04-06_watsonx_access.md | WatsonX API setup walkthrough, available models, usage patterns, latency benchmark results (Maverick vs 70B) | You need to onboard your local .venv to hit the hosted Llama models | | eval_harness_readme.md | End-to-end Windows runbook for AssetOpsBench harness, WatsonX setup, Docker path, scenario-server grading flow, both CODS benchmark tracks (cods_track1, cods_track2), smoke script (../scripts/run_harness_smoke.cmd), and proof expectations for canonical runs | You need to quickly prove harness execution is working this week and run new scenario prompts | | data_pipeline.tex | Paper-ready LaTeX section describing dataset schemas, shared-key strategy, output formats, reproducibility | You're writing the paper/report and need the data-pipeline section | | dataset_visualization.png | Historical 6-panel sample visualization of the processed datasets (static smoke test only; ../notebooks/01_data_exploration.ipynb is the reproducible successor) | You want to compare the old static smoke-test image with the new notebook-backed exploration path | | hpml_datasets.pdf | Tanisha's reference writeup on the 5 Kaggle source datasets (formats, row counts, licensing) | You need background on where the data came from | | images/ | Inline figures referenced by the .md files | - | | archive/ | Historical docs that were once live but are now frozen reference artifacts | You need provenance, not the current operating picture |

Related directories

Conventions

  • One purpose per file - if a doc is doing two things, split it.
  • Date-stamped markdown updates - every doc should have a *Last updated: YYYY-MM-DD* line near the top. Stale dates are a smell.
  • Cross-reference other docs by relative path - e.g. compute plan, not absolute URLs. Keeps the repo portable.
  • Low-churn reference docs live under knowledge-base/reference/ - class requirements, mentor guidance, and setup references that change less often should live there rather than crowding the top-level index.
  • Paper-ready content (LaTeX) lives here - finished paper sections can be dropped into Overleaf as .tex files. Don't mix draft and final in the same file; use git history for versions.
  • No shipped deliverables here - if it's a frozen PDF/PPTX/Keynote export that was submitted or emailed, it belongs in ../reports/, not docs/.
  • No active planning artifacts here - historical planning records live in ../planning/archive/; agent-coordination surfaces are no longer part of this team repo's documentation tree.