Alexander Korolev
|
c2e6c33d79
4x r39 perf: warm re-baseline via perf-only reruns (cold d0 c1 1772 -> 2018-2103), spec numbers corrected
|
1 сар өмнө |
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
|
1 сар өмнө |
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
|
1 сар өмнө |
Alexander Korolev
|
c11e975b4e
refactor: rename base image family to jovian-judgement (dir + images + profiles); tag = pin cut + model, :local retired
|
1 сар өмнө |
Alexander Korolev
|
7535a7d2c4
glm53 full-model lane: retarget to RedHatAI/GLM-5.3-MXFP4 on the r35 jovian stack; delete the OOM-proven NVFP4-Spark variants
|
1 сар өмнө |
Alexander Korolev
|
743b5a7706
dflash2 night A/B: 4x 165/176 (+2, prefill +27%, TTFT -31%) vs MTP3; 2x Spark blocked (C4 page-stride + interleave%4 gates); history + spec
|
1 сар өмнө |
Alexander Korolev
|
d6bba73685
r35 promoted to GLM 4x production (operator decision); r31 stays as rollback lane
|
1 сар өмнө |
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 сар өмнө |
Alexander Korolev
|
37e4be4a9f
r31 campaign: 4x A/B win (159/176, +10% decode), 2x below baseline, eugr recipe ready
|
1 сар өмнө |
Alexander Korolev
|
74ad0dbd4b
repo language: history + spec sections to English (per operator)
|
1 сар өмнө |
Alexander Korolev
|
2c8328d8a9
Add sparkring shared-image TP2 recipe (2x profile, digest-pinned)
|
1 сар өмнө |
Alexander Korolev
|
448c3605f1
r18 16g 4x boot-verify passed; agents.md sm121-only rule; r28 4x dcp4 recipe (vllm pin override)
|
1 сар өмнө |
Alexander Korolev
|
5cd023e547
roce A/B 2x: 162/176 vs 158 (quality +4, d0c1 decode +17%); rocenante default on for r27
|
1 сар өмнө |
Alexander Korolev
|
7d850a136d
roce A/B: rocenante +31-90% decode on r27 4x dcp4 (160/176 vs 155), restore prod
|
1 сар өмнө |
Alexander Korolev
|
98dd4e42b3
bench instructions: keep json report (--json-file) alongside the md
|
1 сар өмнө |
Alexander Korolev
|
7b64ba7d1c
revert b12x bump: rocenante already in the r27 pin (env-toggle A/B only)
|
1 сар өмнө |
Alexander Korolev
|
8936fde0cd
queue rocenante research (b12x #295): zero-port, b12x pin bump + env toggle
|
1 сар өмнө |
Alexander Korolev
|
00099bcfff
rust frontend evaluated and deferred (600s registration timeout vs our boot times); fork stays selective
|
1 сар өмнө |
Alexander Korolev
|
c957beb710
eager A/B done (159 vs 158, rumor not confirmed); aiden 4x 163 = campaign record; night closed
|
1 сар өмнө |
Alexander Korolev
|
9c648a2ee4
state doc: night-3 close — final campaign table, DFlash2 blocker, queue
|
1 сар өмнө |
Alexander Korolev
|
9cb56de106
state doc: night 2 results (aiden 159 first canonical, r27 156), next queue
|
1 сар өмнө |
Alexander Korolev
|
3e63d19947
state doc: image build v2 (clone-at-pin) contract
|
1 сар өмнө |
Alexander Korolev
|
0a1a5dbcc1
typo
|
1 сар өмнө |
Alexander Korolev
|
b564bd561d
ops layout: 2nodes-dac = production (operator agent), ds4 = dev lane
|
1 сар өмнө |
Alexander Korolev
|
ca8bb6f448
campaign extended: 2x nvfp4 KV (154), 2x Spark (158, best 2x), 4x Spark blocked (mm_mxfp8 TP4); 4x 16G pins; JIT warm-start 9min
|
1 сар өмнө |
Alexander Korolev
|
cc19b5f581
persist JIT caches per node (flashinfer/torchinductor/vllm mounts in all glm53 lanes)
|
1 сар өмнө |
Alexander Korolev
|
24cfdc3a76
state doc: add --perf comparison table (all 7 lanes, llama-benchy pp/tg/TTFT)
|
1 сар өмнө |
Alexander Korolev
|
5d587215dd
separation complete: r18 dir carries source tarballs; state doc records r12/r18 split + r17 rebuild caveat
|
1 сар өмнө |
Alexander Korolev
|
fe070bebdd
night shift: r18 built + full lane matrix (DCP4 161/176 best; eugr/DFlash2/dsv4fv-eugr documented)
|
1 сар өмнө |
Alexander Korolev
|
ccd615fcad
fix hardmode for night shift + future: version-pin tool (2.6.1.dev45, --hardmode = 88 scenarios), re-baseline anchor lane before gating; reports commit to recipes/<lane>/reports/; new eugr-4x test recipe (2x stays production default until 4x dramatic)
|
1 сар өмнө |