 Alexander Korolev
|
9cf76da709
mxfp4 probe: KV pin 4->8 GiB - engine needs 6.58 GiB/rank for one 131072-token request (51.6 KB/token/rank measured)
|
1 месяц назад |
|
Alexander Korolev
|
57e6336047
mxfp4 recipes: --linear-backend b12x (native SM120 GEMM) - the flashinfer mxfp4 linear kernel wrongly claims sm_12x support and crashes in cute-dsl mm_fp4; --moe-backend is inert for the compressed-tensors MoE scheme (it picks MarlinExperts itself)
|
1 месяц назад |
|
Alexander Korolev
|
17672b1322
mxfp4 recipes: --moe-backend flashinfer_cutlass (the Sept-tree CLI family; _afp8 was not a valid choice on r35 - first boot argparse)
|
1 месяц назад |
|
Alexander Korolev
|
87e866512c
night shift: P0 probe recipe (no DCP, 131K, 4 GiB KV) - step 1 of the full-model boot ladder
|
1 месяц назад |
|
Alexander Korolev
|
501e627b2b
night plan: R37 image built+distributed (label 9350429b) and the MXFP4 checkpoint complete on the NAS - both lanes ready, switch needs the go-ahead
|
1 месяц назад |
|
Alexander Korolev
|
beda9d96ae
web: persist window layout (geometry/minimized/closed/maximized + open optional windows) in localStorage across reloads; recipes panel drops its own title and live chip when hosted (body.embed) and reopens from the menu when minimized; closed named windows keep a dock chip
|
1 месяц назад |
|
Alexander Korolev
|
b5e08df460
glm53 full MXFP4: explicit --moe-backend flashinfer_cutlass_afp8 + document the MXFP4 ingestion path
|
1 месяц назад |
|
Alexander Korolev
|
e3e9cbd3b1
web header: guard #maxinfo[hidden] (display:flex defeated the hidden attribute, so the maximize chrome sat in the toolbar permanently); assert header visibility, not the .hidden property, in playwright
|
1 месяц назад |
|
Alexander Korolev
|
8433106ff4
night-shift 2026-09-13 plan + R37 (jovian) build profile
|
1 месяц назад |
|
Alexander Korolev
|
5c54bf0af5
web windows: maximized window drops its titlebar too — header carries the title plus restore/minimize/close for that window
|
1 месяц назад |
|
Alexander Korolev
|
305f9066a6
web windows: maximized window is frameless and flush to the layer (title moves into the header with a restore button), retire the dead dock reserve, clamp px geometry to the layer when the dock appears or the page resizes
|
1 месяц назад |
|
Alexander Korolev
|
aa49a8b61e
web windows: dock lists every window (recipes/raw/logs no longer minimize into nowhere), add a maximize/restore button to every titlebar, size the Topology window to the map aspect and reflow the default layout (topo+jobs left, state full height right) so nothing is dead space
|
1 месяц назад |
|
Alexander Korolev
|
07055b199f
web: recipes panel window — the report --serve page mounted at /recipes (one implementation, prefix-rewritten paths/SSE/viewer) + iframe theme inheritance; fix stale raw-window responses overwriting a newer selection; docs + playwright checks
|
1 месяц назад |
|
Alexander Korolev
|
7535a7d2c4
glm53 full-model lane: retarget to RedHatAI/GLM-5.3-MXFP4 on the r35 jovian stack; delete the OOM-proven NVFP4-Spark variants
|
1 месяц назад |
|
Alexander Korolev
|
4ca149b977
web: replay cached snapshot on SSE connect (live indicator no longer waits a poll tick) + jobs window groups running containers one row per launcher job with its node list
|
1 месяц назад |
|
Alexander Korolev
|
f9dd6c7f18
web topo: legible devices in both themes — theme-safe outlines (fg-mixed, not --border), soft drop shadow via --shade, stronger vents/plate, unpowered = hollow + dashed + grayscale at .7 instead of .32 opacity; contrast probe + skill note
|
1 месяц назад |
|
Alexander Korolev
|
8c1ca0d747
r36 recipe header: final A/B table (512K run) + candidate status
|
1 месяц назад |
|
Alexander Korolev
|
2bde637398
r36 @512K + Engram NVMe: 154/176, 3,774 pp / 822 ms TTFT (2.3x prod prefill, decode -3%) — candidate lane; + forum 381350#521 b12x loader finding (verified: 320B/18B from the checkpoint config)
|
1 месяц назад |
|
Alexander Korolev
|
6798a62b8a
r36 recipe: 512K context (our floor; their 131K profile is not usable) + note the b12x vs instanttensor load-format finding (forum 381350 #521: -35% prefill on the b12x loader at TP4, and it cannot load a DFlash2 drafter)
|
1 месяц назад |
|
Alexander Korolev
|
f042556790
r36 dev65 + local Engram: prefill 3,654 pp / 826 ms TTFT (9.1x the NFS run, 2.2x prod) — storage suspect confirmed; 149/176 = variance + K7 stack gap; lane is now a candidate
|
1 месяц назад |
|
Alexander Korolev
|
eb63d88c6b
add composed_model.local_files: hardlink node-local files into composed dir
|
1 месяц назад |
|
Alexander Korolev
|
a02c574b51
dsv41 lab r36 A/B: builds + works on GB10 (DSpark K7, Engram disk/io_uring, 3.38M KV) but loses — 158/176 vs 167 prod, prefill 401 vs 1,642 pp (suspect: io_uring over NFS); eugr stays prod
|
1 месяц назад |
|
Alexander Korolev
|
b9d4309084
r36 build: add liburing-dev (b12x PLE loader compiles io_uring support at wheel-build time via pkg-config; without it ENGRAM disk mode hard-errors at boot)
|
1 месяц назад |
|
Alexander Korolev
|
09a88f2ced
r36 build: update b12x gate marker (B12X_PCIE_ONESHOT_PDL removed in b12x 323107ff; use _PREPARED_ONESHOT_LAUNCHERS — same file, r36 oneshot launcher tables)
|
1 месяц назад |
|
Alexander Korolev
|
1d6625344d
dsv41 lab r36: build profile (jovian-r36-ds41: vLLM 202a11a98eb + b12x 323107ff + liburing 2.5, rest = r35 pins) + 4x DCP1 recipe (DSpark K7 greedy, Engram disk, their qualified profile, GB10 deviations documented)
|
1 месяц назад |
|
Alexander Korolev
|
743b5a7706
dflash2 night A/B: 4x 165/176 (+2, prefill +27%, TTFT -31%) vs MTP3; 2x Spark blocked (C4 page-stride + interleave%4 gates); history + spec
|
1 месяц назад |
|
Alexander Korolev
|
a2d87c7423
dflash2 4x A/B report: 165/176 (94) vs MTP3 163 (93), prefill +27% / TTFT -31%, decode -4.6%; fix utc_now_iso (secs→days in civil_date)
|
1 месяц назад |
|
Alexander Korolev
|
3e409b5836
2x Spark DFlash2 K7 recipe (r35 img, NVFP4-Spark model, r31-2x shape + speculative swap) — night A/B part 2
|
1 месяц назад |
|
Alexander Korolev
|
23c8fb202f
web topo: cluster chips render as the topmost SVG layer with opaque fill; down state mutes saturation instead of alpha (no frame border shows through badges)
|
1 месяц назад |
|
Alexander Korolev
|
e32d2f054f
skill: web-topology — working manual for tuning the dashboard topology map (layout anchors, frame/hue algorithms, probe ladder, gotchas)
|
1 месяц назад |
