Alexander Korolev
|
1d73fd0945
serve/stop reliability + run ledger + archive curation
|
1 mēnesi atpakaļ |
Alexander Korolev
|
623a906d5f
recipes: set parent_model for all quant/fork repos (GLM 5.1/5.2/5.3/5.3-Flash, Qwen3.8 27B+Flash-Next, MiniMax-M3, MiMo-V2.5, Solar-Open2, DSV4-vision env-driven) — report medal/tree grouping now ranks quants inside the base model's category
|
1 mēnesi atpakaļ |
Alexander Korolev
|
662c7764a4
report viewer: copy button floats over the text field's top-right corner (position:fixed below the header) instead of the toolbar; half-transparent, opaque on hover
|
1 mēnesi atpakaļ |
Alexander Korolev
|
b34eb65752
report viewer: old pty-captured bench logs render like clean ones — collapse rich spinner redraw runs to the last frame per phase (wordish-token progress key, bar-tip glyphs vary), drop tty caret-notation (^D/^H) and backspace artifacts, dedup consecutive lines; copy button in the viewer header (top right, one click, clipboard API + fallback)
|
1 mēnesi atpakaļ |
Alexander Korolev
|
f5144f1f03
report: view state persists across bare reloads (localStorage layer under the URL params — server restart / new tab / binary update keep column visibility, medal filter, grouping; LS arrays normalized to strings on restore)
|
1 mēnesi atpakaļ |
Alexander Korolev
|
081a0e6a0c
bench run package: per-run recipe copy + provenance manifest (recipe/config sha, launcher git sha, tool version, per-node image digest+build label); report shows provenance chip + Provenance column with drift flag
|
1 mēnesi atpakaļ |
Alexander Korolev
|
e74a3ff5d1
report: 80-90 band recolored limegreen -> greenyellow (more yellow per spec)
|
1 mēnesi atpakaļ |
Alexander Korolev
|
597364832f
report: star grade semantics per spec — 10-color palette covers 0..100 a decade per color, always five stars in the decade's color, filled count = position within the decade (1 star = 2 pts, half = 1: 95 -> green 2.5 stars)
|
1 mēnesi atpakaļ |
Alexander Korolev
|
14f912bc94
report: star grade rescaled to 10 steps (1 star = 10 pts, half = 5) with a decade-per-star palette ramp red->firebrick->brown->sienna->chocolate->darkorange->gold->yellowgreen->limegreen->green
|
1 mēnesi atpakaļ |
Alexander Korolev
|
23241ae3dd
report: half-star fix (inline background shorthand reset background-clip:text -> square; use background-image longhand + dim CanvasText right half); uniform hint spacing (94/100, single-space perf separators)
|
1 mēnesi atpakaļ |
Alexander Korolev
|
48b7aa92f7
report: parent_model category field (medals/tree rank under base model); medals by parent x nodes (gold/silver/bronze, filterable); half-star grades (color bands 0-20/40/60/80, half-star precision, sort by absolute score, both-scores hint); report viewer page (rendered md + raw toggle, highlighted toml/json, recipe.toml section once per recipe, opens in new tab)
|
1 mēnesi atpakaļ |
Alexander Korolev
|
d6bba73685
r35 promoted to GLM 4x production (operator decision); r31 stays as rollback lane
|
1 mēnesi atpakaļ |
Alexander Korolev
|
e8ac5f1bd7
r35 A/B: 163/176 (dev65) vs r31 159 (dev46), perf flat, TC-33/80 fixed, TC-51 new; report + state
|
1 mēnesi atpakaļ |
Alexander Korolev
|
9bf6e76e6c
adopt r35 (latest lab rev): full 3-fork pin profile + r35 4x-dcp4 recipe; image built on dgx5 (71m, checksum 8bb87348), boot gates green
|
1 mēnesi atpakaļ |
Alexander Korolev
|
0d8166de95
port DCP2 onto our tree (22-file superset wrapper + a2a recipe); A/B: 162/176, 5.34M-token KV @512K, -5 vs b12x — the high-KV lane
|
1 mēnesi atpakaļ |
Alexander Korolev
|
2b94b23132
pin tool-eval-bench harness at 6be685f0e (dev65) + document cross-harness bias; extract Aiden DCP2 set (23 files, port assessment: 12 clean / 4 merge / 7 cherry-pick)
|
1 mēnesi atpakaļ |
Alexander Korolev
|
a902dd0124
night 09-12: eugr 166/176, sparkring 8b32ef2 reliability FAIL (again), Aiden 1.1 DCP2 A/B 158/176 (b12x lane wins); r31 back to prod, dgx5 off
|
1 mēnesi atpakaļ |
Alexander Korolev
|
13720ae96f
eugr 2x nightly-20260910 bench (166/176); sparkring re-run on 8b32ef2/4774a7c5 (+b12x LIBRARY_PATH fix); Aiden production-1.1 DCP2 recipe staged
|
1 mēnesi atpakaļ |
Alexander Korolev
|
e75f2ba7a2
fix publish install.sh template: GIST_ID/IMAGE swap, build-arg line continuations
|
1 mēnesi atpakaļ |
Alexander Korolev
|
d46cc512c7
report: strip ANSI from served logs; bench logs saved clean
|
1 mēnesi atpakaļ |
Alexander Korolev
|
7b9d588ac8
dsv41: b12x A/B — 167/176 (95, Excellent), +9 points and +18-33% speed
|
1 mēnesi atpakaļ |
Alexander Korolev
|
59c79ebe43
docs: DSV4.1 reasoning_effort contract — low/high/xhigh/max/none or int 1-100 (JSON number)
|
1 mēnesi atpakaļ |
Alexander Korolev
|
6a46ad87da
dsv41: 512K+thinking bench 158/176 (90, Excellent), b12x lane ready
|
1 mēnesi atpakaļ |
Alexander Korolev
|
37f24518b8
dsv41: b12x moe-backend image (thin wrapper) + recipe, built on dgx5
|
1 mēnesi atpakaļ |
Alexander Korolev
|
0a39b6fdec
dsv41: ctx 1M -> 512K — decode +50% (71.5 t/s counting), Aiden base analyzed
|
1 mēnesi atpakaļ |
Alexander Korolev
|
093a945f1f
dsv41: 1M ctx, gmu 0.85, thinking+max-effort default, served as deepseek-v4-flash
|
1 mēnesi atpakaļ |
Alexander Korolev
|
94e8109c00
dsv41-flash: 4x live on cluster — 156/176 (89), local engram, eugr 20260910 pulled
|
1 mēnesi atpakaļ |
Alexander Korolev
|
dd86b8a715
dsv41-flash build fixes: cutlass extraction, build tools, NVRTC, no-GPU gate
|
1 mēnesi atpakaļ |
Alexander Korolev
|
4eb845dec7
dsv41-flash image tree: tonyd2wild overlay chain as clone-at-pin build
|
1 mēnesi atpakaļ |
Alexander Korolev
|
37e4be4a9f
r31 campaign: 4x A/B win (159/176, +10% decode), 2x below baseline, eugr recipe ready
|
1 mēnesi atpakaļ |