Alexander Korolev
|
3e63d19947
state doc: image build v2 (clone-at-pin) contract
|
1 месяц назад |
Alexander Korolev
|
9118a4648b
r27: WORKDIR in builder stages (stage split dropped inherited cwd)
|
1 месяц назад |
Alexander Korolev
|
97329c6b10
r27 build dir: clone-at-pin (no tarballs), gated vllm/b12x/lmcache stages, in-build deps fetch w/ pin gates
|
1 месяц назад |
Alexander Korolev
|
0a1a5dbcc1
typo
|
1 месяц назад |
Alexander Korolev
|
b564bd561d
ops layout: 2nodes-dac = production (operator agent), ds4 = dev lane
|
1 месяц назад |
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 месяц назад |
Alexander Korolev
|
fe78fec205
add nvfp4kv spark variants (2x + 4x 16G, -Spark checkpoint on r18 img)
|
1 месяц назад |
Alexander Korolev
|
c296f735a3
4x cells: KV pin 10 -> 16 GiB (NVFP4 4x has the headroom; 2x stays 10)
|
1 месяц назад |
Alexander Korolev
|
79c40c9482
add r18 2x nvfp4_ds_mla KV lane (parity with the 4x cell)
|
1 месяц назад |
Alexander Korolev
|
cc19b5f581
persist JIT caches per node (flashinfer/torchinductor/vllm mounts in all glm53 lanes)
|
1 месяц назад |
Alexander Korolev
|
14b02ac5f5
history/night-shifts/2026-09-05.md: full combined score+perf table of the r18 campaign
|
1 месяц назад |
Alexander Korolev
|
24cfdc3a76
state doc: add --perf comparison table (all 7 lanes, llama-benchy pp/tg/TTFT)
|
1 месяц назад |
Alexander Korolev
|
5d587215dd
separation complete: r18 dir carries source tarballs; state doc records r12/r18 split + r17 rebuild caveat
|
1 месяц назад |
Alexander Korolev
|
b3d61b7af0
separate r18 from r12: dedicated docker-images/glm53-jovian-r18 dir + glm53-jovian-r18 tag; r12 dir restored to r17 pins; recipes renamed to r18 lanes
|
1 месяц назад |
Alexander Korolev
|
6327352f09
untrack regenerated data/benchmarks.sqlite (gitignore)
|
1 месяц назад |
Alexander Korolev
|
fe070bebdd
night shift: r18 built + full lane matrix (DCP4 161/176 best; eugr/DFlash2/dsv4fv-eugr documented)
|
1 месяц назад |
Alexander Korolev
|
4e6b20c673
dflash2 lane: drafter kv_cache_dtype auto (fp8 KV invalid on SM121 outside B12X)
|
1 месяц назад |
Alexander Korolev
|
a8ed91b8b5
dflash2 lane: pin drafter attention_backend FLASH_ATTN (auto-select poisoned by global B12X)
|
1 месяц назад |
Alexander Korolev
|
3976ebe72f
dflash2 lane: drop drafter B12X (kv_cache_dtype constraint; main keeps B12X)
|
1 месяц назад |
Alexander Korolev
|
bddf8d552e
add DFlash2 k=7 drafter lane (r18 img, A/B vs MTP)
|
1 месяц назад |
Alexander Korolev
|
fb4692fdab
dcp4 probe: add --cp-kv-cache-interleave-size 4 (GLM C4 DCP requirement)
|
1 месяц назад |
Alexander Korolev
|
2191c7018b
fix 4x recipes: tensor_parallel_size 4 (world size must cover nnodes)
|
1 месяц назад |
Alexander Korolev
|
b2135bab66
add r18 4x probe variants: DCP4 (TP4+decode-CP4) and nvfp4_ds_mla KV
|
1 месяц назад |
Alexander Korolev
|
eb8646ed43
add r17abi-4x probe recipe (4-node r18 cells: DCP4/RoCE/#646/nvfp4 KV)
|
1 месяц назад |
Alexander Korolev
|
d6076f1d8e
deps re-stage note: MSA cutlass pin (eb61c911) staged too; AppleDouble junk purged
|
1 месяц назад |
Alexander Korolev
|
bd1dc24418
deps re-stage note: FA's own cutlass pin (nvidia/cutlass 62750a2b) now in deps-src.tar.gz; vllm staged tree 67e7e6e
|
1 месяц назад |
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 месяц назад |
Alexander Korolev
|
54dfffeb24
eugr GLM lane: document forum 381350#364 parity + DFlash2 alt spec (night queue 3b: k=7 incoai drafter A/B vs MTP k=5)
|
1 месяц назад |
Alexander Korolev
|
82a8368e5b
night queue: DCP4 probe (sparkring reference, r17abi-first, no build dep) + #646 geometry check; refresh r18 pins to 343661fe/1adb8a38
|
1 месяц назад |
Alexander Korolev
|
250194c1c0
r18: add PR #646 (fine-grained SWA prefix hits + decoupled GLM53 target/recurrent blocks, 5.4x KV capacity); advance bases to new tips (vllm a8c796f3 w/ #649, b12x b58f34ea w/ #315)
|
1 месяц назад |