 Alexander Korolev
|
8f78bcdc45
serve --cold: recipe-declared cold_args overlay for first-launch safety
|
пре 6 дана |
|
Alexander Korolev
|
62bf7be418
power off --node: do NOT turn off cluster-level extras (fabric switch/fan)
|
пре 6 дана |
|
Alexander Korolev
|
0e826d4b35
bench: add --no-think and --timeout passthrough to tool-eval-bench — unblocks valid quality benches for thinking-by-default models (Ling, GLM-4.5-style) whose default reasoning blows the 120s/turn timeout and excludes most scenarios (the 10-16 fp8 Ling run scored 67/100 at 3.1% completion for exactly this reason)
|
пре 6 дана |
|
Alexander Korolev
|
8e3aff4749
upstream-watch 10-17: eugr/spark-vllm-b12x nightly (vLLM 0.1.dev21830, 10-15, 391 archs) checked — NO MuseSpark, NO Step5 (only MuseGlimmer + Step3 family). Confirms SGLang #44530 series is earliest path for Muse-816B. Bonus: b12x has DSpark speculators (Glm53DSpark, Qwen3OmniDSpark, K3DSpark, Gemma4DSpark)
|
пре 6 дана |
|
Alexander Korolev
|
183e6cfea5
upstream-watch 10-17: SGLang check — v0.5.21 has no MuseSpark/Step5 either; but MuseSpark landing via PR #44530 5-layer series (#44664 Docker + #44665 kernels merged, model layer in-flight) + merged cookbook; SGLang likely runs Muse-816B before vLLM. SGLang has bailing_moe_v3 (Ling) as an alt engine option
|
пре 6 дана |
|
Alexander Korolev
|
52fa899b5b
upstream-watch 10-17: Muse Spark Open 816B NVFP4 research — 497GB omni (text+img+video+audio), 1M ctx, NVFP4 g16; BLOCKED: MuseSpark arch not in vLLM v0.31.0/main (only MuseGlimmer 30B), PR #62064 is sm80+INT4 not NVFP4/SM121; ≥700GB VRAM; runnable Muse alt = Muse-Glimmer-30B (on NAS, vLLM-supported)
|
пре 6 дана |
|
Alexander Korolev
|
28ed284c75
Night-plan Task1: Ling fp4 38/54 done; NAS cold-read collapsed to 1.3MB/s (external, not our cluster) = blocker; converter fixes (copy_file_range NFS hang, 1-worker, NAS is the only bottleneck); resume command
|
пре 6 дана |
|
Alexander Korolev
|
802de90878
Night-shift agenda 2026-10-17: Task1 Ling fp4 finalize+serve+bench, Task2 GLM-5.3-Flash-FP4-Local 2x (A/B vs kindling 95/100), Task3 Qwen3.8-Flash-Next 1x, Task4 Nemotron; guardrails incl GB10 stuck-clock + GPU-mem-not-cgroup-contained
|
пре 6 дана |
|
Alexander Korolev
|
74ece11aca
upstream-watch 2026-10-17: v0.32.0 still rc (base stays v0.31.0); NVIDIA FP4-Local family (GLM-5.3/Qwen3.8-Flash-Next/Nemotron) all need b12x container (new lossless NVFP4-scale fmt); GB10 stuck-clock bug (386060) + GPU-mem-not-cgroup-contained (386057) — both bit us; Ling fp4 converter in progress
|
пре 6 дана |
|
Alexander Korolev
|
f4b4f62683
upstream-watch 10-16: Ling fp4 research — ling3 parser, fp8-as-source (Ling is mixed bf16+fp8 ckpt), fp4 format spec (mxfp4 experts+fp8 dense), single-Spark serving via cookbook, fp8->mxfp4 converter path + reference files, Qwen4-27B warm-up plan
|
пре 1 недеља |
|
Alexander Korolev
|
909162aef1
upstream-watch + shift log 10-16: live model probes — Ling fp8 feasible on GB10, Step-5 SM121 DSA blocker; entrypoint=[vllm,serve] gap noted
|
пре 1 недеља |
|
Alexander Korolev
|
621d14da70
night shift 2026-10-16 log: DGPP provisioning e2e (4 bugs fixed), Ling/Step-5 feasibility + prep recipes, repo/forum mining, cluster powered off
|
пре 1 недеља |
|
Alexander Korolev
|
5178315eb0
upstream-watch 10-16: DGPP forum mining — HEAD release notes add dense Qwen3.8-27B + GLM-5.3 quant + DGPP_NO_SWAP (pin-bump candidate); community 2x RoCE single-lane GID3
|
пре 1 недеља |
|
Alexander Korolev
|
accf2728ae
upstream-watch 2026-10-16: Ling BailingMoeV3 supported on v0.31.0 (no nightly) but NVFP4 cut empty; Step-5 is step5_vl VL needing dedicated image; dgpp HEAD moved; v0.32.0 still rc; DGPP w2 deploy prereqs
|
пре 1 недеља |
|
Alexander Korolev
|
28f909758a
engine-drivers spec: correct DGPP provisioning model (rank0-only install, idempotent, space-nodes, script-file tee) + DGPP deploy prereqs from live w2 run
|
пре 1 недеља |
|
Alexander Korolev
|
89013fbb75
native/dgpp: fix provisioning from the first live DGPP w2 run
|
пре 1 недеља |
|
Alexander Korolev
|
7e0fefcbb2
night-shift plan 2026-10-16: DGPP w2 finish, Ling-3.1-flash + Step-5-Preview tests, forum/repo mining, end-of-shift 8x power-off
|
пре 1 недеља |
|
Alexander Korolev
|
1d6fbf9da2
native/dgpp: on-demand DGPP provisioning (clone-at-pin + release.sh + install + rev-marker)
|
пре 1 недеља |
|
Alexander Korolev
|
503fe88e78
web/report: recipe logs open a dashboard WINDOW, not an inline stream
|
пре 1 недеља |
|
Alexander Korolev
|
5194805e77
native/dgpp: emit paths.resident_cache in cluster JSON; exempt Native deploy from multi-node-tp1
|
пре 1 недеља |
|
Alexander Korolev
|
06a4103738
shift log: first-class gr8 + DGPP engine drivers (Phases 1-3)
|
пре 1 недеља |
|
Alexander Korolev
|
fe1d70ada3
engine: DGPP driver + native deploy (Phase 3) — head-coordinated, generated cluster JSON/.env
|
пре 1 недеља |
|
Alexander Korolev
|
f772f3d5e6
engine: gr8 driver (Phase 2) — --tp/--peers rendezvous, --model, fixed API port
|
пре 1 недеља |
|
Alexander Korolev
|
5be604ccfe
engine: EngineDriver trait — first-class engine abstraction (Phase 1)
|
пре 1 недеља |
|
Alexander Korolev
|
28af1b0d4d
upstream-watch 10-15: MiMo 4x rebase done (90/100, +19% decode)
|
пре 1 недеља |
|
Alexander Korolev
|
f6e334a07b
shift log 2026-10-15: MiMo 4x rebase to SparkRing 2026.10.4 (dgx5 disk recovery, digest re-attach sync, 90/100 bench, +19% decode)
|
пре 1 недеља |
|
Alexander Korolev
|
ef7298e1ee
upstream watch 10-15: v0.32.0rc2 imminent (flashinfer 0.7.1 + SM12x NVFP4 KV + DSpark scheduler fix), SparkRing 2026.10.4 validates the MiMo rebase image + adds TP2-DCP2, DGPP adds Qwen3.8-27B, Nemotron-3.5-Bolt 24GB single-Spark, Ling-3.1-flash 572B day-0 NVFP4, net-sircl DAC plugin, CRS812 fabric resolution
|
пре 1 недеља |
|
Alexander Korolev
|
fd3acf21a7
shift log: add-on 1.1.12 — row controls in the expanded section (serve/stop, live logs, bench job)
|
пре 1 недеља |
|
Alexander Korolev
|
606670fa28
recipes panel: move row controls into the expanded section (before the toml) — serve/stop, a live log tail (▤ logs, /api/logs attach + /events stream, node picker) for running recipes, and ⚡ bench (new POST /api/bench tracked job, standard seed-42 campaign, report lands in the recipe repo)
|
пре 1 недеља |
|
Alexander Korolev
|
8e849667ac
shift log: add-on 1.1.11 built + deployed to the Pi (V1 /addons update path, HAOS 18.3 layout, fleet_up verified live)
|
пре 1 недеља |
