docs: frank-secondary at ubatch 2048 — Measured section, rollback ladder, fleet docs

23,120 used / 1,005 free at boot, 763 settled after a vision probe
(encoder buffers resident). Rung cost 746 MiB, not the transferable
616 prediction. Prefill baseline 1,209 t/s @ 3,760 recorded; decode
unchanged as the law says. 2048 = ceiling for this shape.

AGENTS.md

@@ -27,7 +27,7 @@ merge absorbed the former `vllmScripts` repo, history preserved).
 |--------|---------|---------|-------|------|
 | Dankserver | `frank-primary` | **radiance** (StillDeadcode v1.0.7 line; image `stilldeadcode/radiance@sha256:4c3f484564…` digest-pinned INDEX digest, 287 MiB, `libr4d` gfx1201 kernels, plain `podman run`, root systemd unit) | `Qwen/Qwen3.8-Flash-Next-FP8` official (172.8 GB, 131 shards, sha256-verified) converted **no-recipe** to `qwen4exp-fp8-native.rad` (177,109 MiB measured = 173.0 GiB) — **native modelopt block-FP8, no requantisation** — on **BOTH R9700s TP=2** (`--tp-wire wht6`), experts in a **40 GiB/rank pinned host pool** + VRAM cache (`--placement expert_tiered`, `--expert-vs-cache-ratio 0.82` = the 99 %-hit point), **PLE 47.7 GiB pinned to host RAM** (`--ngram-placement ram`; root unit + `LimitMEMLOCK=infinity`), **fp8 KV 7.32 KiB/token**, **MTP spec decode depth 3**, *…
 | Dankclaw | `frank-large` | **halogen-flash-server 0.16.2** (`ghcr.io/peonist-ai/halogen-flash-server@sha256:0c61bf84…`, digest-pinned, plain `podman run` via `scripts/serve-frank-large.sh`, root systemd unit) | Quat3rnion/halogen-swift1.5-qwen3.8-flash-next-v2-abliterated **`qwen38-flash-next-v2-swift15-abliterated.hgn`** (66,687,678,464 B = compacted r2 of the card's 89,775,868,032 B original, ~62 GiB resident; the v2 `.hgn` format loads in NO other engine), **+ REQUIRED `qwen38-flash-next-ngram.hgn`** (51,200,246,144 B, peonist @5cc17cea — read beside the checkpoint) **+ vision tower** (897,916,416 B, @e053f488 — **vision ON**, `image_url` parts only, ≥256 px probes), **MTP spec decode ALWAYS-ON (`HALOGEN_SPEC_ADAPT=0` — the adaptive default disables the head per request and HALVES decode)**, MTP head INSIDE the checkpoint (`HALOGEN_MTP_HEAD` unset; no sidecar fetched), **2 slots × 200,192 ctx** (KV pool 400,384 positions ≈ 10.7 GiB at ~28 KiB/position; the 27B slots>1-disables-drafting law does NOT carry over — batching and speculation coexist, the card's quickstart ships 2 with the head), **server-side sampler 1.0/0.95/20** (card recipe; models.yml extraBody matches), coupled cap/timeout 65536/7200, prompt cache ON by default (warm 100k turn ≈ 2 s vs ≈ 88 s cold) | 7770 |
-| Dankclaw | `frank-secondary` | llama.cpp master (0.5.0-dev, commit `5fc4f3c`), CUDA sm_86 build (root systemd unit, relinked) | ajgazin/Swift-1.5-Qwen3.8-27B-Uncensored-Dynamic-MTP-GGUF **RCO-IQ3_XXS -mtp** (27B dense qwen3_5 hybrid, 10,442,827,392 B; a per-tensor allocation mix of stock types; the MTP head in the file **drafted** — **spec ON** (draft-mtp depth 3) since 2026-10-06 evening, ngram skipped on the recurrent-rollback law) + the publisher's own **BF16 mmproj** for **vision**, on the RTX 3090, **2 × 114,688** (Frank's 112k cap each since the 2026-10-06 late reshape, cut from 131,072 to fund **ubatch 1024** — **1,751 MiB measured free**) with **q8_0 K + q8_0 V (FP8 class, 34 KiB/token)** under flash attention, batch 4096 / ubatch 1024; **58.2 t/s recipe decode / 66.9 (temp 0)** (MTP acceptance 42.8–54.4 %). Renamed from `frank-small` 2026-10-06 (87721ab). Supersedes the same-morning IQ3_S-mtp payloads: the two-slot no-spec shape and the one-slot MTP shape (both 2026-10-06; payload on disk).|7771|
+| Dankclaw | `frank-secondary` | llama.cpp master (0.5.0-dev, commit `5fc4f3c`), CUDA sm_86 build (root systemd unit, relinked) | ajgazin/Swift-1.5-Qwen3.8-27B-Uncensored-Dynamic-MTP-GGUF **RCO-IQ3_XXS -mtp** (27B dense qwen3_5 hybrid, 10,442,827,392 B; a per-tensor allocation mix of stock types; the MTP head in the file **drafted** — **spec ON** (draft-mtp depth 3) since 2026-10-06 evening, ngram skipped on the recurrent-rollback law) + the publisher's own **BF16 mmproj** for **vision**, on the RTX 3090, **2 × 114,688** (Frank's 112k cap each since the 2026-10-06 late reshape, cut from 131,072 to fund the **ubatch ladder 512 → 1024 → 2048** — **1,005 MiB free at boot, 763 settled after a vision probe** (encoder buffers resident)) with **q8_0 K + q8_0 V (FP8 class, 34 KiB/token)** under flash attention, batch 4096 / **ubatch 2048** (the ceiling for this shape — 4096 would not fit); **prefill 1,209 t/s @ 3,760 tokens; ~52–58 t/s recipe decode** (MTP acceptance 34–54 %, temp-1.0 noise; decode is ubatch-invariant). Renamed from `frank-small` 2026-10-06 (87721ab). Supersedes the same-morning IQ3_S-mtp payloads: the two-slot no-spec shape and the one-slot MTP shape (both 2026-10-06; payload on disk).|7771|
 
 Retired endpoints (`frank-npu`, and every prior runtime) are in
 the [runbook index](#runbook-index). The full swap history lives in
@@ -310,7 +310,7 @@ are the engine core; `--mm-max-patches` available),
 and `frank-secondary` through
 the publisher's own `mmproj-BF16.gguf` passed as `--mmproj` — the Swift
 GGUF itself carries no vision tensors. `frank-secondary`'s `contextWindow` is
-**114688** (cut from 131072 on 2026-10-06 late to fund ubatch 1024), the
+**114688** (cut from 131072 on 2026-10-06 late to fund the ubatch ladder to 2048), the
 per-slot window, not the model's native 262144: the harness must compact
 before llama.cpp answers a longer prompt with a 400. Note also
 that
@@ -426,7 +426,8 @@ context cap, KV at FP8, vision on, as many 128k contexts as fit**;
 reshaped three times on 2026-10-06 — morning to two 131,072 slots
 no-spec, evening to this UNCERSED RCO-IQ3_XXS payload with MTP back
 on, late the RENAME to **frank-secondary** (87721ab) at two 114,688
-slots with ubatch 1024. The
+slots with the ubatch ladder climbing 512 → 1024 → 2048 (f2575a1,
+d8da7a9). The
 same-day payloads it supersedes: the LIVE MiMo-V2.6 Q6_K shape and the
 Ornith-1.5-9B Q8_0 swap that was committed (77decb5..380aafa) but never
 deployed (blocked at the docs gate; the docs debt is what this section
@@ -440,36 +441,36 @@ at q8_0**. The DeltaNet state is a fixed per-slot buffer (~150 MiB/slot at
 48 linear layers × 48 value heads). Header facts were read from the Hub
 file by HTTP range + struct parse before anything was downloaded.
 
-**The shape (2026-10-06 late; Frank's tight-fit order, MTP back on in
-the evening, per-slot cut + ubatch raised late):** `--ctx-size 229376
---parallel 2` = **TWO 114,688-token conversations** (Frank's 112k cap
-each; `n_ctx_slot = CTX_SIZE / PARALLEL`; cut from 131,072 to fund
-ubatch 1024), `--flash-attn on` pinned, **q8_0 K + q8_0 V** (llama.cpp's
-FP8 class — no true fp8 KV dtype exists; 8.5 effective bits; supersedes
-the same-day f16/f16 rule written for the Ornith 9B fit), all layers on
-the GPU, `--mmproj` for vision, batch 4096 / ubatch 1024. **Speculative
-decoding is ON** (draft-mtp, depth 3): the -mtp file's head is drafted
-against the main model's own weights; ngram-mod skipped on purpose — on
-a qwen35 hybrid ngram disables recurrent-state rollback, so a rejected
-draft corrupts the DeltaNet state.
+**The shape (2026-10-06 midnight; Frank's tight-fit order, MTP back on
+in the evening, per-slot cut + ubatch climbed to 2048):** `--ctx-size
+229376 --parallel 2` = **TWO 114,688-token conversations** (Frank's
+112k cap each; `n_ctx_slot = CTX_SIZE / PARALLEL`; cut from 131,072 to
+fund the ubatch rungs), `--flash-attn on` pinned, **q8_0 K + q8_0 V**
+(llama.cpp's FP8 class — no true fp8 KV dtype exists; 8.5 effective
+bits; supersedes the same-day f16/f16 rule written for the Ornith 9B
+fit), all layers on the GPU, `--mmproj` for vision, batch 4096 /
+ubatch 2048. **Speculative decoding is ON** (draft-mtp, depth 3): the
+-mtp file's head is drafted against the main model's own weights;
+ngram-mod skipped on purpose — on a qwen35 hybrid ngram disables
+recurrent-state rollback, so a rejected draft corrupts the DeltaNet
+state.
 
-**Measured at the late boot (2026-10-06, 2 × 114,688 / ubatch 1024):
-22,374 used / 1,751 MiB free** (`nvidia-smi` free field) — the per-slot
-cut freed ~1.05 GiB of KV, which paid for the ubatch-1024 buffers and
-netted ~1.25 GiB back over the evening's tightest-ever fit (23,636 used
-/ 489 MiB free at 2 × 131,072 / ubatch 512, under the old 2-3 GiB floor
-BY ORDER; a 512×512 solid-red image still answered `Red` there, so the
-fit carries the vision transients). The MTP draft ctx costs **~2.4 GiB
+**Measured at the midnight boot (2026-10-06, 2 × 114,688 / ubatch
+2048): 23,120 used / 1,005 MiB free, settling 763 MiB after a vision
+probe** (`nvidia-smi` free; the encoder buffers stay resident on this
+boot) — the per-slot cut freed ~1.05 GiB of KV and both ubatch rungs
+spent it back (~0.3 + ~0.75 GiB measured; the 1024 rung had landed
+22,374/1,751, the 2048 rung cost 746 MiB, ~130 over the 616 the rung
+measured on the MiMo shape). The MTP draft ctx costs **~2.4 GiB
 measured** (23,636 − 22,728 + 1,561 at the 262,144 draft window) — ~6×
 the ~0.4 GiB the one-slot era assumed, because the drafter's KV is
 built against the target's FULL context, not the 3-token proposal
 depth; the lighter RCO-IQ3_XXS quant (−1.56 GiB vs IQ3_S-mtp) funds it.
-**Decode at the late boot: 58.2 tok/s (recipe temp 1.0) / 66.9 (temp
-0)**, acceptance 42.8–54.4 % (the 2 × 131,072 boot read 63.3 / 72.9 at
-49.8–62.4 % — acceptance is text-dependent and the prompts differ: the
-comparison is confounded, not a shape verdict). `reasoning_effort`
-probe unchanged: `xhigh` 200, `high` → 500 Jinja raise (the models.yml
-fold stays mandatory).
+**Prefill 1,209 t/s @ 3,760 tokens; decode 52.0–58.2 tok/s (recipe
+temp 1.0) / 66.9 (temp 0)**, acceptance 34.2–54.4 % — decode is
+ubatch-invariant, the spread is temp-1.0 acceptance noise.
+`reasoning_effort` probe unchanged: `xhigh` 200, `high` → 500 Jinja
+raise (the models.yml fold stays mandatory).
 
 **Vision with the publisher's own projector**: ajgazin's repo SHIPS
 `mmproj-BF16.gguf` in the same repo (the IQ3_S era's unsloth mmproj was