Status board (auto-updated) · 작업 현황 #40

        *    .    '     *    .
    .        \  .  |  '  /       *
      '     *  \  \ | /  /    .
                 .-^-^-.
                /  ) (  \       ___
               |  (   )  |=====<   >
               |   ) (   |      \_/
               |  .---.  |
              _|_/_____\_|_
             [_____________]
q75: 1/2 green, running

Rewritten on every train, batch and worker change by tools/board/board.py: state files, gh and git log only, no LLM. Rendered 2026-10-22 07:38 +0900.

Release

0.3.1: tagged 2026-10-12 19:12

next: not opened

cut window 2026-10-13 22:00–24:00 KST; plan updated 10-12 07:04; 1 landed, 3 in a train now, 3 pending, 0 deferred to the next release

Landed (1)

• v0.3.0 re-cut: host prompt cache restores a shared prefix (quant) — 8d93782, tarball/tap/tag/docker re-cut 2026-10-12

In a train now (3)

• Common prompt stream for Body35/Body38 (B-1b, B-2a, B-2b) (common) — train2
• GLM vision seat (G3a, G2, G3b) (vision) — G3a/G2 in train2; G3b next
• serve: 12 GB 카드에서 qwen35moe 기본 ctx가 슬롯당 2048이라 6K 토큰 prompt를 거부 #28 12 GB defaults verified (gate) (quant) — exp0212b; reporter asked to confirm

Pending (with ETA) (3)

• GLM IQ3_S experts on the card (0.3.0 off cell) (quant) — round A kernels running; B after the qw2 train
• NVMe tier: R4d-c records + refill (+7.07 % decode) (unitier) — refill landed; R4d-c batch running
• Workflow fixes: box reservation file, per-track narrow dirs, gate-batch --wait-hold (lead, toktape) — from the 10-12 workflow reports

Deferred to the next release (0)

None.

Trains

as of 07:38 (the newest train)

q75: batch running

updated 07:38; branch q75 on base -; from batch events, no train file

• batch real: running, 1/2 done (1 green, 0 red), predicted end 07:40

raonpf2: batch green

updated 06:54; branch raonpf2 on base -; from batch events, no train file

• batch b627c: ended, 1/1 done (1 green, 0 red)

rbugs: batch green

updated 05:50; branch rbugs on base -; from batch events, no train file

• batch rbugs-r2: ended, 40/40 done (17 green, 23 deferred, 0 red)

c612: landed

updated 04:51; branch c612 on base -; on origin/main at 1e8e923e; from batch events, no train file

• batch c612-real: ended, 2/2 done (2 green, 0 red)

c605: landed

updated 04:24; branch c605 on base -; on origin/main at ba2057ce; from batch events, no train file

• batch c605-tp: ended, 1/1 done (1 green, 0 red)

lift3: batch green

updated 03:11; branch lift3 on base -; from batch events, no train file

• batch lift-fix: ended, 4/4 done (4 green, 0 red)

v41enc: landed

updated 02:56; branch v41enc on base -; on origin/main at 1f3e69a7; from batch events, no train file

• batch 20261021-165112: ended, 24/24 done (13 green, 10 deferred, 1 skip, 0 red)

raonpf: batch green

updated 02:43; branch raonpf on base -; from batch events, no train file

• batch raonpf-rerun: ended, 2/2 done (2 green, 0 red)

dflash2: batch red

updated 10-21 23:15; branch dflash2 on base -; from batch events, no train file

• batch fix-batch: ended, 9/9 done (8 green, 1 red)

mac-m0b: batch green

updated 10-21 22:51; branch mac-m0b on base -; from batch events, no train file

• batch batch4: ended, 1/1 done (1 green, 0 red)

fix594: batch green

updated 10-21 18:16; branch fix594 on base -; from batch events, no train file

• batch 20261021-180551: ended, 4/4 done (4 green, 0 red)

mac-portmap: batch green

updated 10-21 17:31; branch mac-portmap on base -; from batch events, no train file

• batch batch: ended, 7/7 done (7 green, 0 red)

• … and 9 more

Batches

as of 07:38 (the newest event)

• real: running, 1/2 done (1 green, 0 red), predicted end 07:40 — tree bloomery-q75, branch q75, real tier, lead ledger
• lib: ended, 1/1 done (1 green, 0 red) — tree bloomery-q75, branch q75, fixture tier, lead ledger
  • DONE total=1 red=0 tier=fixture skipped=0 wall=309s laneA=307s laneB=0s deferred=0
• b629c: ended, 9/9 done (7 green, 2 deferred, 0 red) — tree bloomery-q75, branch q75, fixture tier, lead ledger
  • DONE total=9 red=0 tier=fixture skipped=0 wall=444s laneA=443s laneB=428s deferred=12 real_only=1 real_file=1
• 20261022-065726: stopped (signal) at 06:57, 0/1 done (0 green, 0 red), predicted end 06:58 — tree bloomery-mac-m21b1, branch mac-m21b1, real tier
• 20261022-065613: ended, 1/1 done (0 green, 1 red) — tree bloomery-mac-base, branch HEAD, real tier
  • red: gate-gpu-qwen35moe-e2e rc 1, lane B, 70 s
  • DONE total=1 red=1 tier=real wall=71s laneA=0s laneB=71s
• 20261022-065459: ended, 1/1 done (1 green, 0 red) — tree bloomery-mac-m21b1, branch mac-m21b1, real tier
  • DONE total=1 red=0 tier=real wall=72s laneA=0s laneB=71s
• 20261022-065344: ended, 1/1 done (1 green, 0 red) — tree bloomery-mac-base, branch HEAD, real tier
  • DONE total=1 red=0 tier=real wall=73s laneA=0s laneB=72s
• b627c: ended, 1/1 done (1 green, 0 red) — tree bloomery-raonpf2, branch raonpf2, fixture tier, lead ledger
  • DONE total=1 red=0 tier=fixture skipped=0 wall=336s laneA=335s laneB=0s deferred=0
• 20261022-065228: ended, 1/1 done (1 green, 0 red) — tree bloomery-mac-m21b1, branch mac-m21b1, real tier
  • DONE total=1 red=0 tier=real wall=74s laneA=0s laneB=74s
• 20261022-065105: ended, 1/1 done (0 green, 1 red) — tree bloomery-mac-base, branch HEAD, real tier
  • red: gate-gpu-qwen35moe-e2e rc 1, lane B, 80 s
  • DONE total=1 red=1 tier=real wall=81s laneA=0s laneB=80s
• … and 75 more

Workers

from the Claude session registry; a note file is the session's own words

vision: no session

• (note from 10-13 13:40, stale) vision — handed over

unitier: no session

• (note from 10-13 13:37, stale) unitier — handed over, see specs/unitier/HANDOFF.md

common: no session

• (note from 10-13 13:39, stale) # common

quant: no session

• (note from 10-13 13:41, stale) quant — handed over, see specs/quant/HANDOFF.md

toktape: no session

• (note from 10-13 13:37, stale) # toktape

b2: no session

• (note from 10-20 03:00, stale) # b2 (Sonnet 5.5) — 2026-10-20 ~05:05

bmat: no session

• (note from 10-12 11:25, stale) round bmat (unitier): done, report delivered

cevid: no session

• (note from 10-17 06:44, stale) round: cevid — card evidence for 16 contributor needs-gpu PRs

engine: no session

• (note from 10-17 14:41, stale) # engine (GLM-5.3, Bolt PR gpu: Bolt prompt path — the nvfp4p GEMM, the barrierless Mamba-2 scan, the relu_sqr fold, the common router #396)

g5: no session

• (note from 10-17 14:44, stale) # g5 — lever4 (IQ2_XXS plane)

h1: no session

• (note from 10-19 22:55, stale) round: none. model: the host union's ungated relu² form, for Bolt's routed experts #355 closed as superseded by lead (r5 8924d06 is in main); nothing open for h1

lead: no session

• (note from 10-21 09:19, stale) lead — 2026-10-21 09:30

lintchk2: no session

• (note from 10-20 03:35, stale) lintchk2: done, report sent to lead

mocc: no session

models: no session

• (note from 10-21 20:29, stale) # models (MiniMax-M3, Solar-Open2)

offload: no session

• (note from 10-17 15:24, stale) # offload (gpu: the host tier nine-slot pieces, handoff_9 and the card sum Nine entry (Bolt offload) #391 merged; round closed)

perf: no session

• (note from 10-17 15:26, stale) # perf

promo: no session

• (note from 10-21 13:31, stale) promo — round done

sysone: no session

• (note from 10-20 18:48, stale) sysone: /v1/systemone on generative seats (0.3.6) — DONE, report sent

ubimpl: no session

• (note from 10-17 10:06, stale) round: ubimpl-cont (Bolt load-time pass size, finish the proof; Sonnet 5.5)

vs6stage: no session

• (note from 10-21 08:43, stale) # vs6stage — the class stage in affected (2026-10-21)

Contributor PRs and issues

as of 07:38 (queried now)

Open PRs

• app, gpu, gpu-gates: a window no verify takes leaves no stale draft state #631 app, gpu, gpu-gates: a window no verify takes leaves no stale draft state by juliankang4: waiting (changes requested), updated 06:19
• quant, card leg: Qwen3.5/3.6's placed card leg runs the common leg's Q8_0, Q6_K, IQ3_XXS and IQ4_XS experts (#75) #629 quant, card leg: Qwen3.5/3.6's placed card leg runs the common leg's Q8_0, Q6_K, IQ3_XXS and IQ4_XS… by midagedev: waiting (changes requested), updated 05:51
• AGENTS.md, deny.toml: the Mac's Metal gate, the device, metal and record crates, one copy of each objc2 crate #628 AGENTS.md, deny.toml: the Mac's Metal gate, the device, metal and record crates, one copy of each o… by midagedev: not yet picked, updated 02:36
• raon1: the prompt in wide units on the common GEMM, KDA and route owners (#616) #627 raon1: the prompt in wide units on the common GEMM, KDA and route owners (#616) by midagedev: not yet picked, updated 07:02
• recipes: key reached inputs and persist structural reach #587 recipes: key reached inputs and persist structural reach by Burntgogi: not yet picked, updated 10-21 21:30
• gpu, gpu-minimaxm3: MiniMax-M3's block-sparse attention on a common KeyWalk in flash_gqa #566 gpu, gpu-minimaxm3: MiniMax-M3's block-sparse attention on a common KeyWalk in flash_gqa by midagedev: waiting (changes requested), updated 10-21 16:02

Issues with a contributor comment in the last 7 days

• Compose a model's body from common blocks read off its file; refuse a missing block by name #609 Compose a model's body from common blocks read off its file; refuse a missing block by name: last comment by Burntgogi 00:25
• gpu/rope_neox: safe head-256 launcher permits duplicate cache writers #580 gpu/rope_neox: safe head-256 launcher permits duplicate cache writers: last comment by Burntgogi 10-21 23:58
• tools, ci, hf: problems found on the side of the DGX Spark work, by severity #524 tools, ci, hf: problems found on the side of the DGX Spark work, by severity: last comment by Burntgogi 10-21 21:59
• gate-refset: the hf-qwen4exp-nvfp4 sets are not on the maintainer's box #596 gate-refset: the hf-qwen4exp-nvfp4 sets are not on the maintainer's box: last comment by juliankang4 10-21 16:45
• tools: gate keys per recipe from what its bins reach, not whole crates #529 tools: gate keys per recipe from what its bins reach, not whole crates: last comment by Burntgogi 10-21 14:22
• NVIDIA FP4-Local checkpoints on DGX Spark (TP=1, TP=2): format, model and speed plan in rounds #565 NVIDIA FP4-Local checkpoints on DGX Spark (TP=1, TP=2): format, model and speed plan in rounds: last comment by juliankang4 10-21 09:32
• measure: GLM-5.3-Flash with the MTP draft on and off (same binary) #76 measure: GLM-5.3-Flash with the MTP draft on and off (same binary): last comment by juliankang4 10-18 03:12
• qwen35/27B native generation: existing Body35, frontend dispatch and mixed-IQ admission gaps #88 qwen35/27B native generation: existing Body35, frontend dispatch and mixed-IQ admission gaps: last comment by Burntgogi 10-15 16:12

Recently landed

as of 07:38; origin/main as this checkout has it, newest commit 07:24

• 687b0f9 docs: Sage-A1.5 Q4_K_M runs on the qwen3 seat
• b63aac1 nextn: target rows whose end wraps are refused by name; Held::holds and shown checked against overf…
• c615afd router: a non-finite router logit raises the Router fault through one score owner (router_score), i…
• 1e8e923 Merge #612 (Burntgogi): gpu: make the partial head-256 cache append unsafe
• 76387c8 docs: census facts and glue for layer composition
• ba2057c Merge #605 (Burntgogi): hf: resolve compressed-tensors targets by checkpoint module class
• 374273d model: one ATTENTION table with per-head norms in stems.rs, which MiniMax-M3 and qwen3moe share; qw…
• 47064b2 model: noaux's block, GLM's shared expert and qwen3moe's layer names from the common stem tables; t…
• 8afe453 docs/contrib/new-model.md: the common walkers and tables a new model reuses
• d23118e gate-paths: rows for the raon1, minimaxm3 and solar2 readers, so a change to them narrows

1 lines withheld