 Alexander Korolev
|
2aa9023c2f
bench: MiMo-V2.6-Pro-RL 8x first run — 165/176 (94/100), perf pp ~1.7K t/s
|
пре 4 недеља |
|
Alexander Korolev
|
759910ef68
history: 09-25 shift — MiMo-V2.6-Pro-RL 8x bring-up (build, sync, serve from NAS, smoke pass)
|
пре 4 недеља |
|
Alexander Korolev
|
d5663ec4d9
feat: MiMo-V2.6-Pro-RL 8x Spark recipe + build tree (SGLang a14, DFlash, RoCEnante); dry-run path-ref placeholder
|
пре 4 недеља |
|
Alexander Korolev
|
51b7404aa8
feat: serve --hf-cache <path> overrides the host-side HF cache dir per serve (config hf_cache_dir stays the global knob; mutual exclusion with --mirror)
|
пре 4 недеља |
|
Alexander Korolev
|
3f52b0072a
config: merge [mikrotik] from cluster_sources (with the HA endpoint it merges from) — a RouterOS-named extra_entity 404s as raw HA when the section is dropped
|
пре 1 месец |
|
Alexander Korolev
|
36872f238c
web: recipes iframe loads /recipes with the token (the page document itself is token-gated)
|
пре 1 месец |
|
Alexander Korolev
|
d1e9c05c35
web: dashboard token in the browser (?token= seed -> localStorage, auth() on all data URLs, 401 prompt; recipes iframe shares it)
|
пре 1 месец |
|
Alexander Korolev
|
c2e6c33d79
4x r39 perf: warm re-baseline via perf-only reruns (cold d0 c1 1772 -> 2018-2103), spec numbers corrected
|
пре 1 месец |
|
Alexander Korolev
|
314812116c
feat: sm121 bench --perf-only (skip scenarios, re-baseline the perf table after a cold first row)
|
пре 1 месец |
|
Alexander Korolev
|
5dc8f91220
4x r39 verified: boot clean (comm.roce 81/81), bench 164/176 v2.7.0, perf +53% prefill / 2x long-decode vs 2x
|
пре 1 месец |
|
Alexander Korolev
|
bb1b4c83fa
recipe: glm53-flash-nvfp4-lab-4x-r39 (r39 tip on the 4x 512K config)
|
пре 1 месец |
|
Alexander Korolev
|
f1e87ec828
history: night 09-21/22 shift log (r39 verified, r38 4x repro, control baseline)
|
пре 1 месец |
|
Alexander Korolev
|
ccaa32b296
night 09-21/22: r39 2x verified (deadlock fixed upstream, bench 164/176 v2.7.0), r38 4x reproduces the pool deadlock (0/1822), r35 rebuild + control baseline 165/174, libdw-dev apt pin for r39 DeepGEMM
|
пре 1 месец |
|
Alexander Korolev
|
9431a65802
feat: r39 image lineage (lab tip 2026-09-21: new external-dep contract, DeepSelect, cutlass v4.7.1); fix deps.lock header continuation line not being a comment
|
пре 1 месец |
|
Alexander Korolev
|
ecf4102e7e
docs: rewrite README to match current implementation (sm121 binary, all commands, serve flags, recipe fields, mikrotik/registry config, build flow)
|
пре 1 месец |
|
Alexander Korolev
|
c11e975b4e
refactor: rename base image family to jovian-judgement (dir + images + profiles); tag = pin cut + model, :local retired
|
пре 1 месец |
|
Alexander Korolev
|
8e20fd1dad
feat: versioned registry tags (build checksum) + latest; publish --push; duplicate guard via publication ledger with --force
|
пре 1 месец |
|
Alexander Korolev
|
9c39603968
feat: [registry] — push built images to a registry (ghcr.io) after build; docker push via tmux on the node, localized (6 locales), --no-push opt-out
|
пре 1 месец |
|
Alexander Korolev
|
69ba71ad59
feat: health probes [mikrotik] devices via RouterOS API login (ping = on, dead/rejected = off, blocking); power-on routes through on_entity relay
|
пре 1 месец |
|
Alexander Korolev
|
3717c3a4de
feat: deps.lock — fetch-deps.sh verifies staged deps against a committed fingerprint lock (regen via --regen-lock / REGEN_LOCK=1)
|
пре 1 месец |
|
Alexander Korolev
|
4e25ee4b61
fix: publish --gist — silence gh api child output (was leaking 540 KB JSON to stdout)
|
пре 1 месец |
|
Alexander Korolev
|
40f9e96bef
fix: publish — merge build profile pins, flatten all COPY-referenced dirs, gist via gh api
|
пре 1 месец |
|
Alexander Korolev
|
e3d6c0401b
feat: graceful MikroTik/RouterOS power-off via ROS API ([mikrotik] config)
|
пре 1 месец |
|
Alexander Korolev
|
b16bf62dee
bench: spark 2x @512K + 10 GiB KV (157/176, A/B vs re-cut 512K)
|
пре 1 месец |
|
Alexander Korolev
|
ed4bcc4369
2x re-cut: 512K context via max-num-seqs 4 + 10 GiB KV pin (bench 160/176)
|
пре 1 месец |
|
Alexander Korolev
|
ed49b2d94a
docker: image sync falls back 200G→10G with a warning (fabric ssh fix hint)
|
пре 1 месец |
|
Alexander Korolev
|
1b605f259b
docker: image sync inner hop over the fabric (ib_ip) when available
|
пре 1 месец |
|
Alexander Korolev
|
20a3c768c9
agents: re-cut measured results + node-to-node transfer / hf-cache / pkill gotchas
|
пре 1 месец |
|
Alexander Korolev
|
e9c009a7c3
bench: NVFP4-lab re-cut — 2x 164/176 (93), 4x 159/176 (90)
|
пре 1 месец |
|
Alexander Korolev
|
144077f8a4
recipes: nvfp4-lab 2x — 256K context (2.4 GiB KV budget at GMU 0.88)
|
пре 1 месец |
