bloomery — status board · 2026-10-15 13:40 KST

Generated by tools/board/now.py every 10 min from git, gate-batch events, locks and GitHub. Only Now is hand-written (lead). Workers: read this first; report via specs/<role>/NOW.

Now (lead, updated 26h33 ago) ⚠ stale

Goal

• 0.3.3. v0.3.2 shipped. Cutoff = when every must-have is done (scope: rig-log plan-triage 0793990).

Next

• gpu-gates, model: the Qwen3.6 fixture tier #128 rebase → force-with-lease to its branch → Shingler + CI → merge (AGENTS.md one-liner approved)
• q38pg (qwen38-serve paged reds, the 0.3.2 known issue) → PR
• gpu-vision, vision, serve: MiniCPM-V 4.7 images end to end, a tile capped at 70 tokens #147 MiniCPM-V images: MUST_FIX fold, then the board reds below
• HF dataset refresh (dgx-spark: what is in progress on the Spark side, as a checklist #53 promise, today)
• ctx-first placement: the user's slots and context come first; experts take what is left (ctx-first) #155: 3060 PC 12 GB rerun → default-change PR

Scope in

• Must-have: placement: the user's slots and context come first; experts take what is left (ctx-first) #155 ctx-first (release waits for it) · MTP: NextN layer가 있는 qwen35moe 파일에 draft가 없음 (Role::Unused) #30 qwen35moe MTP · gpu-vision, vision, serve: MiniCPM-V 4.7 images end to end, a tile capped at 70 tokens #147 MiniCPM-V images · MiniCPM draft serving (q36int) · gpu-gates, model: the Qwen3.6 fixture tier #128 · q38pg · LES C3 prefetch lever (default ON only on a measured win)
• Ride-along if green: T5 resident-first · mixed precision (opt-in; official if gates pass; default unchanged) · Nemotron 4B after MTP: NextN layer가 있는 qwen35moe 파일에 draft가 없음 (Role::Unused) #30 (MTP as draft) · q6down, ubgemm, tilekern, dspaths · docs: a placed qwen35moe prompt walks units of 512, not 8-row passes #154 · contributor PRs

Scope out

• → 0.3.4: Nemotron 4B NVFP4, query-tiled prefill flash, PF-6 (unless it lands with T5)
• PF-1 stopped (tile arm slower on A6000), PF-4 closed; PF-2/PF-5 wait for a Spark

Held for the user

• ImageNumber 500 vs 400 (vision) · tile cap (vision) · per-image replay signature (vision) — gate gpu-vision, vision, serve: MiniCPM-V 4.7 images end to end, a tile capped at 70 tokens #147
• ctx-first default change: OK after the 3060 PC rerun

Notes

• Red gates to clear: b-real gate-gpu-minicpmv-tower, gate-gpu-qwen3-serve · batch128 gate-gpu-ds41-nvtier, gate-gpu-xstream, gate-gpu-glm5next-nvtier
• Resolved: dsfullr-real1 red superseded by dsfullr-real2 green and gpu-deepseek41: V4.1 draft runs on the DSpark owner with held bytes from owner and layer buffers #126 merged; card_leg.rs gate-paths row closed
• Builds: Mac via tools/mac-remote.sh; Linux on the lead box under the build-slot flock; GPU via tools/box.sh.

Needs attention

• ⚠ batch mimoscr-batch4 (mimoscratch): red weekly-gpu-mimo2-e2e
• ⚠ batch mimoscr-batch3 (mimoscratch): red weekly-gpu-mimo2-serve
• ⚠ batch mimoscr-batch2 (mimoscratch): red gate-gpu-routed-legs, weekly-gpu-mimo2-e2e, gate-gpu-glm5next-serve, weekly-gpu-mimo2-serve
• ⚠ batch mimoscr-mut1 (HEAD): red weekly-gpu-mimo2-serve

Roles

Resources

• GPU host holds (local mirror): none

Gate batches (last 12 h)

• seat-batch1 [nemo7seat, real] running — 1/22 green, 2 skip/deferred · 3m ago
• n7b-batch3 [nemo7b, real] green — 10/10 green · 2h21 ago
• n7b-batch2 [nemo7b, real] green — 1/1 green · 2h40 ago
• n7b-batch1 [nemo7b, real] red — 7/8 green, red: gate-gpu-nemotron-e2e · 2h41 ago
• mimoscr-batch4 [mimoscratch, real] red — 1/2 green, red: weekly-gpu-mimo2-e2e · 8h22 ago
• mimoscr-batch3 [mimoscratch, real] red — 1/2 green, red: weekly-gpu-mimo2-serve · 8h42 ago
• mimoscr-batch2 [mimoscratch, real] red — 2/6 green, red: gate-gpu-routed-legs, weekly-gpu-mimo2-e2e, gate-gpu-glm5next-serve, weekly-gpu-mimo2-serve · 9h03 ago
• mimoscr-mut1 [HEAD, real] red — 0/1 green, red: weekly-gpu-mimo2-serve · 9h28 ago
• ds41rung2-batch4 [ds41rung2, real] green — 1/1 green · 10h44 ago
• ctxdel-batch1 [ctxsearch-del, real] green — 1/1 green · 10h52 ago

Contributors (open PRs; non-maintainer issues active in 24 h)

• tokenizer: the gemma4 model and pre-tokenizer, BPE over metaspace text as the reference runs it #241 tokenizer: the gemma4 model and pre-tokenizer, BPE over metaspace text as the re — @juliankang4 · updated 73m ago
• model, models: gemma-embedding2's header reader, with an EncoderSpec on ModelSpec #238 model, models: gemma-embedding2's header reader, with an EncoderSpec on ModelSpe — @juliankang4 · updated 2h31 ago
• gpu: geglu_tanh, the GELU-gated unit over f32 rows #237 gpu: geglu_tanh, the GELU-gated unit over f32 rows — @juliankang4 · updated 2h33 ago
• refset: Nemotron 3 Nano last-position logits, read from one reference set #236 refset: Nemotron 3 Nano last-position logits, read from one reference set — @midagedev · updated 21m ago
• gpu, model, gpu-gates, tools: the common pieces a Nemotron-H card body attaches to #235 gpu, model, gpu-gates, tools: the common pieces a Nemotron-H card body attaches — @midagedev · updated 2h28 ago
• gpu: gqa_prefill_flash_256_p2_lower, a first key a row #234 gpu: gqa_prefill_flash_256_p2_lower, a first key a row — @juliankang4 · updated 3h09 ago
• gpu, gpu-gates: the head-512 rope append and prefill flash of EmbeddingGemma 2's global layers, with their gate #233 gpu, gpu-gates: the head-512 rope append and prefill flash of EmbeddingGemma 2's — @juliankang4 · updated 3h06 ago
• tools: two pre-review slots, a device-dir cache, and a claude that does not answer --version refused #229 tools: two pre-review slots, a device-dir cache, and a claude that does not answ — @midagedev · updated 55m ago
• gpu, model, gpu-gates: the Qwen3.5 in-file draft runs one step on the card (M2a-1) #226 gpu, model, gpu-gates: the Qwen3.5 in-file draft runs one step on the card (M2a- — @midagedev · updated 3h09 ago
• tools/boxq.py: an attach rides out a daemon blip before exit 84 #225 tools/boxq.py: an attach rides out a daemon blip before exit 84 — @midagedev · updated 58m ago
• app, gpu, gpu-deepseek41, gpu-gates, gpu-glm5next, levers, model, placement: node 1 serves its expert tier to node 0 over a control channel, and a lost node ends both sides #219 app, gpu, gpu-deepseek41, gpu-gates, gpu-glm5next, levers, model, placement: nod — @juliankang4 · updated 7h22 ago
• tools: an ended gate run takes its binary's timeout with it #201 tools: an ended gate run takes its binary's timeout with it — @juliankang4 · updated 6h40 ago
• verbs, gpu, gpu-gates, levers, tools, docs: the tier link's RDMA end: each page held by its own end, pump thread and queue pair, its posts sent as ordered RDMA writes; verbs.rs is its own crate #180 verbs, gpu, gpu-gates, levers, tools, docs: the tier link's RDMA end: each page — @juliankang4 · updated 2m ago
• issue hidden_ref: invalid numeric options silently select defaults #240 hidden_ref: invalid numeric options silently select defaults — @app/github-actions · updated 2h27 ago
• issue gpu-gates: chain FFN gate duplicates the shared rounding model #239 gpu-gates: chain FFN gate duplicates the shared rounding model — @app/github-actions · updated 2h33 ago
• issue boxq: signals during the final output flush return the job's success code #214 boxq: signals during the final output flush return the job's success code — @app/github-actions · updated 5h09 ago
• issue model: EmbeddingGemma and EmbeddingGemma 2 on a side seat #190 model: EmbeddingGemma and EmbeddingGemma 2 on a side seat — @juliankang4 · updated 2h03 ago
• issue gpu-gate self-test: the outer timeout leaves the inner gate alive #171 gpu-gate self-test: the outer timeout leaves the inner gate alive — @app/github-actions · updated 4h54 ago
• issue tools, gpu-gates, gpu, engram, gguf, qdot, refset: problems found on the side of the DGX Spark work, by severity #137 tools, gpu-gates, gpu, engram, gguf, qdot, refset: problems found on the side of — @juliankang4 · updated 21h53 ago
• issue 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 ad — @Burntgogi · updated 4h21 ago
• issue native AMD: Rust/HIP Q4_K and FFN feasibility on RX 7900 GRE (operators only) #87 native AMD: Rust/HIP Q4_K and FFN feasibility on RX 7900 GRE (operators only) — @Burntgogi · updated 4h21 ago
• issue dgx-spark: what is in progress on the Spark side, as a checklist #53 dgx-spark: what is in progress on the Spark side, as a checklist — @juliankang4 · updated 10h10 ago
• issue dgx-spark: P0 builds and runs on a DGX Spark, with readings for the port map #13 dgx-spark: P0 builds and runs on a DGX Spark, with readings for the port map — @juliankang4 · updated 16h29 ago

Recently landed on main

• 9fc2cc0d 25m ago — model: format the inlined stage-row check
• 9576ba71 25m ago — model, gpu: the stage-row refusal stays in the crate, and each rope test builds its layout
• 57592af1 25m ago — model, gpu: re-pin the stage rope plans and share the table byte count
• f8636c09 25m ago — model, gpu: the Qwen plans count the stage rope table, and the placed loads hold it to the
• 13531b6c 2h54 ago — tools: stop pre-review when the ancestry walk fails
• 071acba2 2h54 ago — tools: refuse a pre-review base that is behind origin/main
• 805284f2 2h58 ago — gpu-gates: drop the qwen3 props byte-sum the serve API already pins
• fb42019a 2h58 ago — gpu-gates: the qwen3 seat reports its plan's placement in /props

Rules for workers

• Commit on your own branch only. No push, no land, no tag, no GitHub write — the lead does those after the user says go.
• Builds: Mac checks via tools/mac-remote.sh (one run per branch); Linux cargo/just on the lead box, one at a time, under flock /tmp/bloomery-boxbuild.lock; GPU only via tools/box.sh on the GPU host.
• Never remove another track's lock or hold. Release work (the open release's branch) goes first on every slot.
• Don't decide held user questions; write options and wait.
• Report: keep specs/<role>/NOW current (task:/next:/blocker:, 3 lines) whenever you start, finish or get stuck; evidence in specs/<role>/HANDOFF.md + evidence/.
• Never write hosts, LAN/tailnet addresses, home paths or tokens in anything that may become public.

1 line(s) withheld by the privacy filter.