 Alexander Korolev
|
7b9d588ac8
dsv41: b12x A/B — 167/176 (95, Excellent), +9 points and +18-33% speed
|
1 月之前 |
|
Alexander Korolev
|
59c79ebe43
docs: DSV4.1 reasoning_effort contract — low/high/xhigh/max/none or int 1-100 (JSON number)
|
1 月之前 |
|
Alexander Korolev
|
6a46ad87da
dsv41: 512K+thinking bench 158/176 (90, Excellent), b12x lane ready
|
1 月之前 |
|
Alexander Korolev
|
37f24518b8
dsv41: b12x moe-backend image (thin wrapper) + recipe, built on dgx5
|
1 月之前 |
|
Alexander Korolev
|
0a39b6fdec
dsv41: ctx 1M -> 512K — decode +50% (71.5 t/s counting), Aiden base analyzed
|
1 月之前 |
|
Alexander Korolev
|
093a945f1f
dsv41: 1M ctx, gmu 0.85, thinking+max-effort default, served as deepseek-v4-flash
|
1 月之前 |
|
Alexander Korolev
|
94e8109c00
dsv41-flash: 4x live on cluster — 156/176 (89), local engram, eugr 20260910 pulled
|
1 月之前 |
|
Alexander Korolev
|
dd86b8a715
dsv41-flash build fixes: cutlass extraction, build tools, NVRTC, no-GPU gate
|
1 月之前 |
|
Alexander Korolev
|
4eb845dec7
dsv41-flash image tree: tonyd2wild overlay chain as clone-at-pin build
|
1 月之前 |
|
Alexander Korolev
|
37e4be4a9f
r31 campaign: 4x A/B win (159/176, +10% decode), 2x below baseline, eugr recipe ready
|
1 月之前 |
|
Alexander Korolev
|
df77baae75
r31 version profile + r31/eugr recipes + night shift 09-10 plan
|
1 月之前 |
|
Alexander Korolev
|
f4abb53cbf
recipe [env] refs: values can resolve from recipe fields ({ from = "port" })
|
1 月之前 |
|
Alexander Korolev
|
95dfdf3001
history: note the PORT injection revert
|
1 月之前 |
|
Alexander Korolev
|
bd82f6f9ae
revert generic PORT injection; PORT is an explicit recipe [env] again
|
1 月之前 |
|
Alexander Korolev
|
74ad0dbd4b
repo language: history + spec sections to English (per operator)
|
1 月之前 |
|
Alexander Korolev
|
1c53f3c737
sparkring 4x results + recipe cleanup per operator
|
1 月之前 |
|
Alexander Korolev
|
87340498fb
sparkring-shared image dir: thin FROM wrapper over the pinned digest
|
1 月之前 |
|
Alexander Korolev
|
2c8328d8a9
Add sparkring shared-image TP2 recipe (2x profile, digest-pinned)
|
1 月之前 |
|
Alexander Korolev
|
0b41469beb
report page: update bell — SSE diff of rows snapshots, badge count, dropdown panel with per-update links, click-through to the row; node probe + playwright click-through tests
|
1 月之前 |
|
Alexander Korolev
|
476b5a5cde
report serve: event-driven refresh — notify watcher (FSEvents/inotify) + 500 ms debounce replaces the 5 s poll; 30 s keepalive as safety net; events filtered to reports/ + recipe.toml
|
1 月之前 |
|
Alexander Korolev
|
c0409a5a0c
report --serve: live updates over SSE — 5 s refresher re-scans reports/, pushes rows JSON on change; page re-renders keeping view state
|
1 月之前 |
|
Alexander Korolev
|
709a8ea749
history: production switched to full r28 4x; sparkring recon (DFlash2 retired, native MTP is the path)
|
1 月之前 |
|
Alexander Korolev
|
0d7c9f334e
upstream issue descriptions: b12x eh_proj loader + DFlash2 dtype (retired in sparkring, native MTP is the path); 386B capacity note
|
1 月之前 |
|
Alexander Korolev
|
73ac661b42
bench report: 2x r28 on 2nodes-dac — 158/176, DAC == switch (±2%); history: DAC A/B, pty fix, 386B OOM, DFlash2 blocked
|
1 月之前 |
|
Alexander Korolev
|
87ad741ec8
bench: run tool-eval-bench under a pty (script -q) — rich Live deadlocked on plain stdout pipes after the perf phase (2x lane, 3/3 repro; 4x got lucky)
|
1 月之前 |
|
Alexander Korolev
|
9d9c038ede
history: 386B OOM-loop required HA power cycle (respawning workers); cluster recovered clean
|
1 月之前 |
|
Alexander Korolev
|
06bf0e84da
history: 386B Spark model — 4x OOM-proven (101.5 GiB/node weights + double load); needs 5-6+ nodes
|
1 月之前 |
|
Alexander Korolev
|
0238fe8169
recipes: dflash2-2x BLOCKED on r28 too — DFlash kernel BF16-KV-only vs fp8_ds_mla engine KV (fork-side)
|
1 月之前 |
|
Alexander Korolev
|
9dcb240b67
bench+history: split-2048/256 env vars — boot clean, 156/176 neutral (KV +2.6%); DFlash2 prereq
|
1 月之前 |
|
Alexander Korolev
|
7d2bc49082
recipes: test variants — split2048 (Discord env vars, booted, bench running), spark-nvfp4-4x (386B boot test), r28-dflash2-2x (re-test), r28-spark (2x lane)
|
1 月之前 |
