 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 månad sedan |
|
Alexander Korolev
|
82a8368e5b
night queue: DCP4 probe (sparkring reference, r17abi-first, no build dep) + #646 geometry check; refresh r18 pins to 343661fe/1adb8a38
|
1 månad sedan |
|
Alexander Korolev
|
f8ab660d6c
new lane dsv4-flash-vision-eugr (eugr b12x nightly + Vision-Exp, DSpark-3 adaptive, plain-B12X attn); night-shift queue: r18 build, redhat MTP, eugr GLM, dsv4-vision eugr
|
1 månad sedan |
|
Alexander Korolev
|
f1d9e635f0
new lane glm53-flash-nvfp4-eugr: eugr B12X full-cluster dual-Spark variant (nightly-20260904 image, RoCE one-shot, MTP k=5 humming) — A/B cell vs r18; dry-run verified
|
1 månad sedan |
|
Alexander Korolev
|
02b94db96f
redhat lane: add MTP k=5 — model_mtp.safetensors (7.6 GB) now in the main NVFP4 repo; card 'Enable Speculative Decoding' section (2026-09-05)
|
1 månad sedan |
|
Alexander Korolev
|
509654fbed
ledger: glm53-flash-aiden-4x row (89/156-176, 37.6 tg c1 flat to d8192, prefill 2,185 = his ref) + TC-43 note
|
1 månad sedan |
|
Alexander Korolev
|
6922cb8584
ledger: glm53-flash-aiden-2x row (aiden production-1.1 + incoai DFlash2 k=7, 89/157-176)
|
1 månad sedan |
|
Alexander Korolev
|
fd4c74dcc1
checkpoint A/B on ds4: NVFP4 lab ckpt = same 88 quality on r17 stack but +4-7 tok/s; lab day-0 stack retains 93; r17abi recipe added
|
1 månad sedan |
|
Alexander Korolev
|
04984a169d
ledger: one-backend bisect solved (linear b12x = the toxin); ship config locked
|
1 månad sedan |
|
Alexander Korolev
|
582343c9c6
r17 composition ship anchor: backend-flags are the quality toxin (bisect 46/47/74 vs 87); knobs env clean; graphs+MTP fixed on public r17 refs; ledger
|
1 månad sedan |
 kilork
|
08a930e477
r12 image: adopt eugr lab build patches (12x patch_vllm_*.py) + pre-staged build sub-deps (SRC_DIR escapes) for flaky node github; ledger: lil render reference + perf push state
|
1 månad sedan |
|
kilork
|
5283d11ce1
clean r12 image shipped: 90/100 c1 (was 88 day-0-base), entrypoint+MAX_JOBS boot fixes, ledger
|
1 månad sedan |
|
kilork
|
2d17cd0356
r12 clean build: SHA-pinned PR heads (550 closed), deepgemm pre-clone, lil launch-policy diff recorded
|
1 månad sedan |
|
kilork
|
8de6340763
clean r12 rebuild: raw-cuda base per solar pattern (r12-specific refs, torch 2.13, flashinfer 0.6.17, b12x PR-head composition); docker_image_dir fixes; bench 88/100 (69-suite)
|
1 månad sedan |
|
kilork
|
ac5c6c9108
NVFP4-Spark serving on our r12 fork build: spec-config attention_backend=B12X fixes MTP draft selection; intel recipe retargeted; ledger
|
1 månad sedan |
|
kilork
|
344479943a
recipes: intel glm53 W4A16-AutoRound (tjstewart image, gptq_marlin, entrypoint) + lab NVFP4-Spark (mtp-5 embedded; blocked on whole-network W4A4 loader); ledger: rtx6kpro source contract for own fork build
|
1 månad sedan |
|
kilork
|
de62327a2b
ledger: dsv4vision tool-eval-bench c1 (81/100, tg 34.1/31.0/29.1, deployability 74, 3 safety-critical fails)
|
1 månad sedan |
|
kilork
|
8d49ca7159
dsv4vision: env-gated openai mm branch (VLLM_DSV4_MM), worker shape-follow patches, local-NVMe weights (dgx3+dgx4), cache-clear runbook; text lane verified exact on new boots
|
1 månad sedan |
|
kilork
|
d34e7758bc
dsv4-vision lane: text serving verified on eugr fork (34.9 tok/s, accept 0.714); OpenAI-mm processor gap documented
|
1 månad sedan |
|
kilork
|
1617621e8c
dsv4-flash-vision lane (sfxnz day-0 plugin port) + MNBT1024 assertion result
|
1 månad sedan |
|
kilork
|
65916245e2
lab-quant variant of 0rand winning config (forum 381350 #119): GMU 0.89, MNBT 2048, KV auto-size
|
1 månad sedan |
|
kilork
|
5e3c35617b
EXL3 lane benched: 88/100 quality, 23.1/24.1 t/s flat, deployability 77 (best of four glm53 lanes)
|
1 månad sedan |
|
kilork
|
e81ae9b4ec
EXL3 recipe: verified serving via instanttensor loader overlay
|
1 månad sedan |
|
kilork
|
25f2710be9
EXL3 recipe boot diagnosis + sparkrun bench numbers in spec ledger
|
1 månad sedan |
|
kilork
|
caae1b0ead
4p67: eager vs CUDA-graphs A/B measured (tool-eval-bench, seed 42)
|
1 månad sedan |
|
Alexander Korolev
|
96c6bfe777
feat(recipes): lab defaults to --enforce-eager — MTP+CUDA-graphs quality bug confirmed by A/B bench
|
1 månad sedan |
|
Alexander Korolev
|
28e08afa73
feat(recipes): redhat @ 524288 ctx — pinned 10 GiB KV, autotune off, 4096 chunks, lab image for CC-12.0 topk guards
|
1 månad sedan |
|
Alexander Korolev
|
3d30d85655
docs(specs): lab /174 explained — TC-53 excluded on 120s infra timeout (score unchanged at 89 either way)
|
1 månad sedan |
|
Alexander Korolev
|
a0383302e3
docs(specs): redhat MTP reality check — RedHatAI ships mtp as separate -mtp-FP8 variant repos (GLM-5.2 pattern); none for 5.3-Flash yet
|
1 månad sedan |
|
Alexander Korolev
|
5730abbfee
docs(specs): GLM-5.3-Flash three-way quant benchmark on ds4 (measured)
|
1 månad sedan |
