fpga-archive-reports-process · EAGLE4 FPGA archive

Source docs/handoff_20260924/00_START_HERE.md · source md5 e80b414ee9d113e53abf41b325a37a44 · archived 2026-11-18

Download PDFOpen PDFRaw markdown

HANDOFF 2026-09-24 — START HERE (orchestrator session 87e06065, 2026-09-23 ~09:00 → 2026-09-24 ~02:30 EDT)

Written by the orchestrator at the user's request ("record all you learned so far … so a fresh you with no knowledge on this subject can continue"). Read the pages in order. Everything here cites a file, commit or measurement; if a number has no source, treat it as a projection.

Page What it covers
00_START_HERE.md (this) how to resume, who the user is, standing rules, the first 10 minutes
01_PROJECT_AND_ARCHITECTURE.md what the drafter is, the split pair, engines vs phases, registers, measured time budgets
02_2BIT_ENGINE_RESEARCH.md the 2-bit GEMV array: width, adder, DSP packing, LUT-GEMM, systolic, locality, fast clock
03_ENGINE_SPLIT_DESIGN_D.md the next architecture: GEMV card + state card + link transport; what is proven
04_TAIL_TREE_WIDTH_BATCHING.md fast TAIL, runtime tree width 3, sibling batching W, state-card work items
05_MODELS_ACCEPT_AND_WIN_CONDITION.md depth×width model, one-pair win condition vs GPU, accept-gap RCA, GPU grid, overlap, scale-out
06_IMAGES_BUILDS_FLASH_TEST.md every image id, fork, pipeline, status; flash + cold-AC + post-flash test procedure
07_INFRA_MACHINES_MESH_SITES.md machines, access, licenses, the mesh dispatcher, monitoring + research sites
08_AGENTS_OPEN_ITEMS_NEXT.md agent roster (ids, deliverables), running jobs + where to harvest, decisions waiting on the user
09_LESSONS_AND_TRAPS.md everything that bit tonight, and the rules they produced

The 10-minute resume checklist

  1. CronList; if empty, re-arm the hourly heartbeat (CLAUDE.md: "FIRST ACTION OF EVERY RESUME"). Crons die with the session.
  2. Read CLAUDE.md (pinned: GPU golden command, the live tracklist, the compute-mesh table), then TRACKLIST_20260923.md (the live board, ~40 KB, one row per job/experiment), then USER_DIALOGUE_LOG.md tail (the user's own words tonight), then reports/results/SESSION_CONTEXT_BACKUP_20260923.md tail (timestamped facts with evidence).
  3. Check reality by ARTIFACT, not memory: ~/mesh/mesh_status.sh (all 3 mesh hosts), the VM build logs (page 06), ami_tool overview (card state), git log --oneline -15.
  4. Page 08 lists every agent that was alive and what it owned. Agents die with the session: re-launch the ones whose deliverable is not DONE, briefed from their report file + the tracklist row. Never redo finished work.
  5. If box1 was cold-AC'd: run the post-cold-AC sequence (page 06 §4) before anything else touches the cards.

Who the user is and how they work

Hard rules (the ones that were live tonight — full list in CLAUDE.md)

One-paragraph state of the world at handoff (2026-09-24 02:30 EDT)

Live silicon pair: HEAD 0xE4B711C9 (61:00.0) + TAIL 0xE4B711B0 (41:00.0) = 102.13 tok/s @ accept 2.22 on mt_bench/32-token/depth 2/width 8/UNPRIMED (humaneval regime is ~28–39 tok/s). The TAIL card is FLASHED with the new TAIL 0xE4B711CA (attention removed) but NOT yet cold-AC'd (user deciding). Builds in flight: HEAD 0xE4B711C8 (HEAD_PAR 8) routing on Oleg's VM; HEAD 0xE4B711CD (C8 + 2-bit array on a 133 MHz clock) packaged, moving to asgard; TAIL 0xE4B711CB (fast TREE) queued on the VM; TAIL 0xE4B711CF (runtime tree width, for width 3) gating on box1. The one-pair win-condition model says the pair reaches ~parity (0.95×) with GPU-only at bs=1 with every lever, not a win; the biggest pro-FPGA unknown is SGLang 0.5.16 on the pair; the structural win path is serving 2 requests (FPGA drafts B while GPU verifies A). The accept gap (2.22 vs GPU 2.78) is 92% the unprimed regime.

Skills (added 2026-09-24 ~02:45)