 Alexander Korolev
|
72358e130c
spec: stock vLLM/SGLang vs ours for DeepSeek-V4.1-Flash-1017
|
9 цаг өмнө |
|
Alexander Korolev
|
12e79273ef
scripts: flare-upstream-watch — FLARE moves daily, check it before every experiment
|
9 цаг өмнө |
|
Alexander Korolev
|
f9606ab210
night report 2026-10-22/23: full GLM-5.3 at 90/100 with MTP, DCP impossible, prefill ceiling measured, FLARE brought up, add-on 1.3.4-1.3.9
|
9 цаг өмнө |
|
Alexander Korolev
|
60ffddaa38
build: bake SM121_VERSION through a generated source, not option_env alone
|
14 цаг өмнө |
|
Alexander Korolev
|
89553b26ac
build: rebuild when SM121_VERSION changes (option_env! alone does not invalidate cargo's cache)
|
14 цаг өмнө |
|
Alexander Korolev
|
a9cffba5fa
web: build-version badge under the dashboard title
|
15 цаг өмнө |
|
Alexander Korolev
|
b984f59269
night plan 2026-10-22/23: reproduction of the long-prefill host hang, the three envelopes to measure (KV-per-token, prefill scratch, spec depth), and the guardrails
|
22 цаг өмнө |
|
Alexander Korolev
|
4c4e608ec1
report/web: model-column copy + Hugging Face buttons, and a favicon
|
1 өдөр өмнө |
|
Alexander Korolev
|
73b27de372
spec: FLARE bring-up — DeepSeek-V4.1-Flash-1017 served by NVIDIA's flare on 4 Sparks
|
1 өдөр өмнө |
|
Alexander Korolev
|
87e885d967
night report 2026-10-21/22: dashboard fixes 1.3.0-1.3.3, recipe-relative hooks, TF 1.0.12 blocked (MissingEntry) with full diagnosis, GLM-5.3 full MXFP4xFP8 blocked at 4-node worker init, GLM-5.3-Flash 2x repro 95/100
|
1 өдөр өмнө |
|
Alexander Korolev
|
ff00fa4d1a
night report 2026-10-21/22 (in progress): dashboard fixes 1.3.0-1.3.3, recipe-relative hooks, TF 1.0.12 bring-up blocked with diagnosis, GLM-5.3 full MXFP4xFP8 lane staged
|
1 өдөр өмнө |
|
Alexander Korolev
|
cd1fc19c1e
specs: TensorFold 1.0.12 4-node bring-up — engine+kit+weights+rank-cache all verified, blocked on a loader MissingEntry (documented with the exact inputs and our deviations)
|
1 өдөр өмнө |
|
Alexander Korolev
|
73b8ee0800
night plan 2026-10-21: TF 1.0.12 TP4 (upstream moved from our pinned 0.6.0), RedHat GLM-5.3-MXFP4xFP8 (415GB/4 nodes), local-inference-lab GitHub, daily HF deltas
|
2 өдөр өмнө |
|
Alexander Korolev
|
f38b60111d
recipes panel: running state follows the run ledger (stale rows after stop)
|
2 өдөр өмнө |
|
Alexander Korolev
|
ebf461d422
recipe-relative hook scripts (host scripts belong to recipes), bench --name, 500-line splits
|
2 өдөр өмнө |
|
Alexander Korolev
|
a24a83696e
pre-serve hooks: serve-local NODE_RANK + TensorFold staging fixes; AGENTS.md monitoring rule
|
2 өдөр өмнө |
|
Alexander Korolev
|
882625898f
web: window drag/resize now survives an <iframe> (pointer capture)
|
2 өдөр өмнө |
|
Alexander Korolev
|
fd30bf160a
web/report: settings button — per-agent client config for a recipe's served model
|
2 өдөр өмнө |
|
Alexander Korolev
|
06722fefd9
night shift 2026-10-21: long-context results (131K/327K needle, both correct), depth sweep, 384K context
|
2 өдөр өмнө |
|
Alexander Korolev
|
a334763f95
night shift 2026-10-21: Beam DFlash2 speculation unlocked via FLASHINFER (+40% decode, turn 2.6s); real agentic task on Beam via pi (11/11 tests); DeepInfer assessed as SM100/offline-only
|
2 өдөр өмнө |
|
Alexander Korolev
|
5e0d7c2306
specs: DeepInfer assessment — offline-only + SM100-only, not runnable on GB10; the transferable ideas (SWA bounded replay, confidence-scheduled verification)
|
2 өдөр өмнө |
|
Alexander Korolev
|
7a8c147767
night shift 2026-10-20: FLARE 1017 tested and rejected (2x slower decode, crashes on agent prompts); Beam-1.0 NVFP4 lane added at 90/100 (fleet best); four-lane comparison recorded
|
2 өдөр өмнө |
|
Alexander Korolev
|
5a4eec608c
spec: 1017 TensorFold lane is now an sm121 recipe; record the TF+vLLM blocker
|
4 өдөр өмнө |
|
Alexander Korolev
|
198e6137dd
engine: tensorfold driver + tensorfold_1017_prepare built-in (NODE_RANK exported to hooks)
|
4 өдөр өмнө |
|
Alexander Korolev
|
8f734c7926
spec: TensorFold 2x 1017 lane measured — 89/100 with half the nodes, decode 50.4 vs 39-43, but turn latency 6.7s vs 1.4s
|
4 өдөр өмнө |
|
Alexander Korolev
|
b17dc421d2
spec: 1017 reasoning-effort A/B — 75 buys nothing over 100 on the new model (86 vs 87, same 1.4s turn); depth sweep closed
|
5 өдөр өмнө |
|
Alexander Korolev
|
0deb95f896
spec: 1017 depth sweep (16K-256K) on the EP2/1M config — single-stream and suffix decode hold, concurrent deep prefills starve decode
|
5 өдөр өмнө |
|
Alexander Korolev
|
000f022481
spec: dsv41-1017 tuning A/B (EP2 decode win, fp4-indexer negative, engram MAX_IDS vs chunked-prefill coupling, slow-NAS boot desync)
|
5 өдөр өмнө |
|
Alexander Korolev
|
cebff9abb2
style: cargo fmt on src/power (occupancy-guard helpers) — no behavior change
|
5 өдөр өмнө |
|
Alexander Korolev
|
d390fd2e87
audit: #15 resolved — the effort-mapping patch baked into the versioned dsv41-sglang-1017:v1 image (glm52 pattern), verified live in the running engine; TensorFold 1017 dual-Spark thread 386121 noted for the watch list
|
5 өдөр өмнө |
