[dmmdea]

Intent

blackwell-16 ships a 26B placement that fits a 16 GB card that is also the desktop's display card, decided by measurement on the tier's new reference box (docs/tiers/reports/blackwell-16-node-h.md), and the Blackwell build texts say what that measurement found instead of a finding from an earlier build.

Scope

• setup/templates/profiles.json: blackwell-16 goes from moe_26b: gpu to n_cpu_moe with N=14 and n_cpu_moe_max 14; its notes carry the measurements and the reference box; the n_cpu_moe_max doc line names the second tier that declares one. Because the RAM gate drops an n_cpu_moe 26B on a box under 28 GB, the agent seat binds by RAM band: config_seed.agent_model is an explicit blank (the lane follows the workhorse on a min box) and config_seed_ram_low_up names gemma-4-26b-agent on 28 GB and up.
• internal/tierseed/tierseed.go: Resolve no longer binds a vLLM seat's llama.cpp fallback when it names the 26B agent seat on a RAM band whose gate drops the 26B (ramDrops26B mirrors moePlacement / Resolve-ProfileParams); the seed layers decide the lane there. Only blackwell-16 declares a 26B fallback.
• setup/detect.ps1: the CUDA-13 branch no longer adds the "cuBLAS fallback, ~5.6x slower prefill" warning; the 12.8 and undetected texts match.
• setup/install.ps1: the build matrix comment and the Select-CudaBuild report lines carry the measured numbers.
• setup/SETUP-AGENT.md: the Blackwell note, same correction.
• docs/tiers/reports/blackwell-16-node-h.md: an "Operator measurements" section after the report proper (bench, the placement sweep, the live result).
• docs/tiers/README.md, docs/tiers/blackwell-16.md: regenerated (go generate .).
• Tests that pinned the old placement, each moved to the rule rather than deleted: TestNoOtherTierRendersTheSpillSeatOrAnNCPUMoESpill (a tier's own measured partial spill may carry --n-cpu-moe; the spill seat stays ampere-6's), TestNo26BAgentSeatOnRAMDroppable26B (evaluates the RAM bands the gate drops the 26B on, with the seed layers that apply there), TestInstallRenderRefusesNCPUMoEAboveTheTiersMeasuredSpill (clears the table's own fields per case), the two audit tests (the bare --cpu-moe control edit anchors on the new 26B line), setup/render.tests.ps1 (expects --n-cpu-moe 14 -ngl 999 on both 26B entries).

Left out on purpose: the 8 GB Blackwell tier's note still says it requires the 12.8 build (not measured here); no new seat for Qwen3.8-27B on Windows (measured as an option only); no template change (both 26B entries already take the placement token).

What changed

Three commits: f7589b0 (the placement, the texts, the report), c038049 (answers the first CI run: the band-aware agent seat, tierseed.Resolve, the tests above, the build-tag text the no-pins lint flagged) and 2da7bbb (the installer's config-seed self-test reads the agent seat by RAM band; ALL PASS, 760 checks, locally under Windows PowerShell 5.1). The second CI run's other red check was internal/mediaremote's TestAnUnknownNodeIsAContractRefusalListingTheRoster (a loopback test node did not answer its health GET on the runner); that package is untouched here and passed on the first run.

The real run and its output

The measurement (same box, llama.cpp win-cuda-13.4 prebuilt, desktop at 4.5 GB falling to 1.6 GB during the run):

placement	desktop before (MiB)	seat on the card (MiB)	free after load (MiB)	pp tok/s	tg tok/s
-ngl 99 (as rendered before)	4546	10997 + spill	760	351	17.0
--n-cpu-moe 4	1546	13065	1692	1428	121.2
--n-cpu-moe 8	1553	11563	3187	824	81.2
--n-cpu-moe 14	1556	9109	5638	535	62.5

llama-bench on the same build: E4B pp512 17694 tok/s / tg128 212, E2B 34454 / 345. No cuBLAS fallback.

The renderer, run on this branch against main as the control (go run . install render -profile blackwell-16 -os windows -ram-tier high ...):

branch: line 54  --n-cpu-moe 14 -ngl 999 --parallel 1 ${common}
        line 66  --n-cpu-moe 14 -ngl 999 --parallel 1 --ctx-size 32768 --flash-attn on
main:   0 occurrences of n-cpu-moe 14

The placement live on the box (both entries patched to the rendered form, llama-swap restarted while idle), one chat completion on gemma4-26b-a4b at the owner's desktop:

26B via llama-swap: wall=8.5 s (cold load included) pp=472.2 tok/s tg=58.2 tok/s prompt_n=796 gen=30
used with the 26B loaded: 10678 MiB, 16303 MiB

After the follow-up commit, locally: go test ./internal/tierseed/... ok; the root package's seed, tier, agent, closure, audit and no-pins tests ok (-run 'FleetCap|Seed|Tier|Agent|...', 17 s); setup/render.tests.ps1 under Windows PowerShell 5.1 with the branch renderer: ALL PASS (326 checks). Gates once: go build ./..., go vet, the profile parses, go generate . clean.

Blast radius

• Every blackwell-16 render: the 26B cascade and agent entries change their placement; about 5 GB of experts move to host RAM (the tier's RAM gate still drops the 26B below 28 GB of RAM, as for any n_cpu_moe tier, and such a box now runs the agent lane on the workhorse instead of a dropped seat). A headless 16 GB card loses the full-GPU speed it could reach on an empty card (noted in the profile; N=4 measured there).
• tierseed.Resolve: only a vLLM fallback naming a 26B seat on a band that drops the 26B changes, which is blackwell-16 on min RAM and nothing else in the table.
• detect.ps1 on a Blackwell box with CUDA 13.x: one warning fewer.
• Docs otherwise.

Risk

Low. jev review-needed = no_review (consequential 0.12, blast 1.75): standard tier, no review pass.

🤖 Generated with Claude Code