Alexander Korolev
|
64fb5a2457
docs(flare-storage): ranks have no NAS dependency (proved with the NAS unmounted); package_cache is only a CLI fallback; fleet drift recorded
|
1 hete |
Alexander Korolev
|
40a98e5400
docs(progress): R6 round - bind design proven with a serving TP1 rank on one node; negative control replaces a traffic counter; option D off the table; fstab bind survives reboot; two of my own earlier conclusions corrected (store-root scaffolding, 2564 MiB)
|
1 hete |
Alexander Korolev
|
1cf8f3c54b
docs(flare-storage): R6b closed without a window on a 25 GB TP1 core; option D not needed; fstab bind survives reboot; R5's store-root claim corrected
|
1 hete |
Alexander Korolev
|
b4a80f084d
docs(flare-storage): store recovered - re-linked from local blobs for 34 MB of network; version id is mutable, deployment id is not
|
1 hete |
Alexander Korolev
|
84c8d632ac
docs(progress): R5 done - pre-cutover runtime table, the dpkg-does-not-own-the-agent-copy mechanism, flare-swap.sh tested on dgx4+dgx8, and the 470 G weights/ trap for the window
|
1 hete |
Alexander Korolev
|
94ca6b831a
docs(flare-storage): pre-window disk math - new core weights/ is 470 G (not the old 260 GiB), dgx1 root is the only tight node (760 G free only after its 477 G store is released); measure the per-rank footprint on dgx4 first
|
1 hete |
Alexander Korolev
|
a5caf79247
feat(host-scripts): flare-swap.sh — the cutover binary swap as one reviewed, idempotent, rank-safe command
|
1 hete |
Alexander Korolev
|
f392e4764c
docs(flare-storage): R5 pre-flight done — swap rehearsed on idle dgx4; the agent copy is what runs and dpkg does not own it
|
1 hete |
Alexander Korolev
|
3a9d1219b8
docs(flare-storage): Phase 3 bind gate half-passes; a standalone model-dir copy is invalid in v0.1.0; rm -rf deletes through live binds
|
1 hete |
Alexander Korolev
|
a46f69a3bf
docs(flare-storage): measured NAS read of a real rank slice - 1.3 GiB direct in 1.29 s at 1.1 GB/s (nconnect=8, 4 MiB reads), so a NAS-backed deployments/ costs ~1.3 s per rank at load
|
1 hete |
Alexander Korolev
|
11d1a02211
docs(flare-storage): gb10 core acquired and verified on the NAS (R3 gate passed)
|
1 hete |
Alexander Korolev
|
5a1374505a
docs(flare-storage): v0.1.0 renames package->core; CLI store now on the NAS (R4)
|
1 hete |
Alexander Korolev
|
e5d03f2f07
docs(flare-storage): correct acquisition rate/ETA from the log timer (55→90 MiB/s, ~1.5 h) and note node clocks are UTC
|
1 hete |
Alexander Korolev
|
3a5ecd1750
docs(flare-storage): the Nov-5 runtime cannot acquire the current package; v0.1.0 staged fleet-wide
|
1 hete |
Alexander Korolev
|
6a341ad022
docs(flare-storage): run the latest release build; take launch params from the package
|
1 hete |
Alexander Korolev
|
cc2e7d5dbf
docs(flare-storage): purge stale data (783 GiB freed); replace, don't migrate
|
1 hete |
Alexander Korolev
|
4ace161940
ops(nas): phase 1 applied on dgx1-4 + dgx8, with a per-node protocol script
|
1 hete |
Alexander Korolev
|
ae6ea2f26e
docs(spark): Ripple-65B NVFP4 running on one GB10, measured
|
1 hete |
Alexander Korolev
|
6fa147f042
docs(flare): plan for weights-local / everything-else-on-NAS storage
|
1 hete |
Alexander Korolev
|
9c20c4f9ff
docs(nvme): swap procedure - the 2TB replacement is 48 GB smaller, so no raw clone
|
2 hete |
Alexander Korolev
|
06d4fad12b
docs(nvme): record the ordered MP600 Mini 2TB and its acceptance test
|
2 hete |
Alexander Korolev
|
8f891c6358
docs: dgx4 NVMe investigation - DRAM-less FTL thrash, not a fault
|
2 hete |
Alexander Korolev
|
2977785ed3
docs: current FLARE lane state, uniform-binary rule, refreshed upstream pin
|
2 hete |
Alexander Korolev
|
06b53e0c78
flare: the dashboard/probe controller pin also resolves outside the lane cluster
|
2 hete |
Alexander Korolev
|
7271c95ae2
release: sm121 0.2.7 — FLARE runtime override, controller outside the launch cluster
|
2 hete |
Alexander Korolev
|
0ebe2ce33e
flare: prod-no4 is production - dgx4 fails sustained Engram duty even with the reader patch
|
2 hete |
Alexander Korolev
|
7d0dc801cd
flare: resolve the pinned controller across clusters; prod-no4 lane config
|
2 hete |
Alexander Korolev
|
cc0c08393f
flare: first-class --executable override + Engram-stall post-mortem
|
2 hete |
Alexander Korolev
|
a1d58a7374
docs(AGENTS): FLARE engine deploy, orbit/controller mapping and the two-store disk budget
|
2 hete |
Alexander Korolev
|
f3548d85e8
flare: honour the pinned orbit controller and match orbit node names by name (0.2.6)
|
2 hete |