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 месяц назад |
Alexander Korolev
|
9d9c038ede
history: 386B OOM-loop required HA power cycle (respawning workers); cluster recovered clean
|
1 месяц назад |
Alexander Korolev
|
06bf0e84da
history: 386B Spark model — 4x OOM-proven (101.5 GiB/node weights + double load); needs 5-6+ nodes
|
1 месяц назад |
Alexander Korolev
|
0238fe8169
recipes: dflash2-2x BLOCKED on r28 too — DFlash kernel BF16-KV-only vs fp8_ds_mla engine KV (fork-side)
|
1 месяц назад |
Alexander Korolev
|
9dcb240b67
bench+history: split-2048/256 env vars — boot clean, 156/176 neutral (KV +2.6%); DFlash2 prereq
|
1 месяц назад |
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 месяц назад |
Alexander Korolev
|
8bf398fb74
history: b12x loader test blocked on eh_proj routing bug (fork-side)
|
1 месяц назад |
Alexander Korolev
|
e6fc89ee5a
history: b12x loader test result — works but eh_proj routing bug blocks it (upstream)
|
1 месяц назад |
Alexander Korolev
|
d1f10dd1e7
bench: r28 full rerun 159/176 — 155 vs 160 was run-to-run variance (8 flips), not a regression
|
1 месяц назад |
Alexander Korolev
|
ff1f727ae3
recipes: GMU 0.88 in r27/r28 4x+2x (startup gate 107.0 vs 109.5 GiB; KV pinned, capacity unchanged)
|
1 месяц назад |
Alexander Korolev
|
3bb263d1af
history: r28 full-stack eval, build_profile feature + TOML trap, memory-gate flake, bench 155
|
1 месяц назад |
Alexander Korolev
|
01a63d9cc5
bench: full r28 4x DCP4 — 155/176 (vs r27 160, mixed 159); 8 hard-scenario flips under full-stack noise band
|
1 месяц назад |
Alexander Korolev
|
3669194b80
skills: document build_profile + TOML bare-key-after-[[table]] pitfall (r28 half-pin incident)
|
1 месяц назад |
Alexander Korolev
|
b5f066216d
recipes: fix build_profile misplaced under [[extra_mounts]] (top-level key); verify live merge
|
1 месяц назад |
Alexander Korolev
|
15e9b62ca7
build: named version profiles (build_profile) + full r28 pin set (b12x 7216f6b9, lmcache 934c287a)
|
1 месяц назад |
Alexander Korolev
|
52e1dc891d
bench: fix pump reading into empty buffer (EOF→EPIPE→silent exit 120)
|
1 месяц назад |
Alexander Korolev
|
95d4ce3dec
apt: optional snapshot pinning (APT_SOURCE=current default, ports.ubuntu.com for arm64; snapshot kept as opt-in)
|
1 месяц назад |
Alexander Korolev
|
3fa8ac759b
builds: persistent per-node buildkit cache (buildx --cache-from/to, mode=max) so version-independent apt/torch/CUDA stages are reused across recipes
|
1 месяц назад |
Alexander Korolev
|
448c3605f1
r18 16g 4x boot-verify passed; agents.md sm121-only rule; r28 4x dcp4 recipe (vllm pin override)
|
1 месяц назад |
Alexander Korolev
|
40d15606b1
report page: exact-TP filter select (options from data, deep-link tp=)
|
1 месяц назад |
Alexander Korolev
|
f121f5f873
report: decode speed columns (llama-benchy tg t/s, c1/c2 at d0) from md+log perf boxes
|
1 месяц назад |
Alexander Korolev
|
62e13c270e
report page: deep-link view state (filter, grouping, sort, columns, paging) via url query
|
1 месяц назад |
Alexander Korolev
|
d481969512
report page: sync js state with browser-restored form controls after reload
|
1 месяц назад |
Alexander Korolev
|
89894dff34
add sm121 report: recipe bench report table with json support
|
1 месяц назад |
Alexander Korolev
|
5cd023e547
roce A/B 2x: 162/176 vs 158 (quality +4, d0c1 decode +17%); rocenante default on for r27
|
1 месяц назад |
Alexander Korolev
|
1ef3eebed6
bench: add sm121 bench — tool-eval-bench with auto-landed reports; move stop handler to docker/stop
|
1 месяц назад |
Alexander Korolev
|
fb3fbf560e
r27 recipes: enable rocenante by default; consolidate -roce clone; 2x arm A report
|
1 месяц назад |
Alexander Korolev
|
a33cfa87a2
stop: wait until containers are actually stopped (60s verify); add --no-wait
|
1 месяц назад |
Alexander Korolev
|
7d850a136d
roce A/B: rocenante +31-90% decode on r27 4x dcp4 (160/176 vs 155), restore prod
|
1 месяц назад |
Alexander Korolev
|
98dd4e42b3
bench instructions: keep json report (--json-file) alongside the md
|
1 месяц назад |