feat(frank-dragoon): double the KV pool to 2 x ctx for slot parity
frank-dragoon's total in-flight context was HALF frank-zealot's: zealot runs one 401,408-position pool split into two 200,704 conversations, while dragoon ran ONE 200,704-position pool behind its two slots, so a single maximum-length request filled it and two long conversations had to stay under ~100k each. HALOGEN_KV_POOL_POSITIONS 200704 -> 401408 = slots x ctx, the top of the seam's legal band (ctx <= pool <= slots * ctx). The PER-REQUEST ceiling does not move: HALOGEN_CTX stays 200704, because the checkpoint's native window is 262,144 and the engine's own 400 text refuses to exceed it without HALOGEN_ROPE_YARN. What this buys is concurrency, not a longer conversation, which is why the harness contextWindow (155648 = ceiling - 45056) is untouched. Cost: ~10.8 GiB of KV where 200,704 cost 5.5 (262,144 cost 7.2; the 2026-10-06 shape's 524,288 measured 14.6), taken from the page cache the 51.2 GB n-gram table reads through — the warm prompt-cache ratio is the number to re-measure after the boot. Engine-side total goes 74.1 -> ~79.4 GiB of 121.2. Also carries scripts/download-frank-zealot-bonsai2-ptq1.sh: the PTQ1_0 candidate payload fetcher (5,946,648,928 B, LFS sha256-verified, mmproj hardlinked), with the publisher's own throughput table in its header — the packing is smaller but the card measures it SLOWER on Ampere, so the swap is decided on our card, not on the card's A100 row.
This commit is contained in:
Frank
@@ -1097,8 +1097,9 @@ deploy_anubis() {
|
||||
# conversion (2026-10-11) of the abliterated Qwen3.8-Flash-Next —
|
||||
# qwen38-flash-next-v2-abliterated.hgn (66,687,678,432 B, ~62 GiB resident)
|
||||
# + the REQUIRED n-gram lookup table + the vision tower — MTP spec decode
|
||||
# (HALOGEN_SPEC_ADAPT=0), 2 slots over ONE 200,704-position KV pool at the
|
||||
# matching ctx 200704 ceiling, image input — on the Strix Halo iGPU. It stays a
|
||||
# (HALOGEN_SPEC_ADAPT=0), 2 slots over ONE 401,408-position KV pool (2 × ctx,
|
||||
# raised 2026-10-18 so BOTH slots can hold a maximum-length request) at the
|
||||
# ctx 200704 ceiling, image input — on the Strix Halo iGPU. It stays a
|
||||
# ROOT unit because the serve script drives `podman run` directly (the frank-archon pattern):
|
||||
# relink_unit points /etc/systemd/system/frank-dragoon.service at the repo
|
||||
# file, boot persistence is that unit's [Install], and LimitMEMLOCK=infinity
|
||||
|
||||
+124
@@ -0,0 +1,124 @@
|
||||
#!/usr/bin/env bash
|
||||
# Candidate payload for frank-zealot: prism-ml/Ternary-Bonsai-2-27B-gguf at
|
||||
# **PTQ1_0** — the DENSE ternary packing (1.75 bits/weight, 5.95 GB) of the same
|
||||
# weights the live PQ2_0 endpoint serves (2.13 bits/weight, 7.21 GB). Public
|
||||
# repo: no auth header, and never a token on the command line (a Bearer token is
|
||||
# world-readable in /proc/<pid>/cmdline).
|
||||
#
|
||||
# WHY THIS ARM EXISTS. The whole argument is footprint: 5,946,648,928 B against
|
||||
# the live 7,206,168,928 B = **1,201 MiB freed on the card**. On this endpoint
|
||||
# KV bills 34,816 B/token (16 full-attention layers x 4 KV heads x 256 dim at
|
||||
# q8_0 K + q8_0 V), so that one gap is ~37,000 tokens of pool per 1,201 MiB —
|
||||
# context is exactly what the freed weight bytes buy.
|
||||
#
|
||||
# WHY IT MIGHT LOSE. The publisher's own throughput table (README, llama-bench,
|
||||
# batch 1, depth 0): PTQ1_0 wins decode only on Ada/L4 (RTX 4090 91.1 vs 81.2,
|
||||
# RTX 6000 Ada 90.4 vs 82.8, L4 32.1 vs 29.8) and LOSES on H100 (86.9 vs 113.9)
|
||||
# and A100 (54.7 vs 73.9); prompt processing "favors PQ2_0 everywhere" (A100
|
||||
# 706 vs 1328). The card text says explicitly that dense-trit unpacking costs
|
||||
# arithmetic and the pack is "slower on Ampere, Hopper, and Blackwell". There is
|
||||
# NO GA102/RTX 3090 row: the Ampere statement generalizes from an A100, which is
|
||||
# instruction-bound at batch 1 with 2 TB/s of HBM, while the 3090 is the
|
||||
# bandwidth-starved member of the family (936 GB/s) — the one regime where
|
||||
# moving 17 % less weight per step can pay. That is an argument, not a
|
||||
# measurement: this script exists so the decision is made on OUR card, against
|
||||
# the PQ2_0 record (76.6 t/s single-stream decode, 1,258 t/s prefill at 35k).
|
||||
#
|
||||
# THE RUNTIME LOCK CARRIES OVER. Both packings are custom GGML tensor types and
|
||||
# stock llama.cpp rejects them ("It rejects PQ2_0 and PTQ1_0 as unknown types"
|
||||
# — the card), so the PrismML fork build stays mandatory. Confirm the loadability
|
||||
# against the installed binary BEFORE pointing any unit at this file.
|
||||
#
|
||||
# The vision projector is the SAME file the PQ2_0 endpoint already uses (Hub LFS
|
||||
# oid 6807ede61d57… matches byte-for-byte), so it is HARDLINKED from the existing
|
||||
# payload dir instead of re-downloaded — a hardlink is a regular file, so the
|
||||
# serve gates' no-symlink rule stays satisfied.
|
||||
#
|
||||
# Usage:
|
||||
# scripts/download-frank-zealot-bonsai2-ptq1.sh [DEST_DIR] [--verify-only]
|
||||
set -euo pipefail
|
||||
|
||||
DEST=""
|
||||
VERIFY_ONLY=0
|
||||
for arg in "$@"; do
|
||||
case "$arg" in
|
||||
--verify-only) VERIFY_ONLY=1 ;;
|
||||
*) DEST="$arg" ;;
|
||||
esac
|
||||
done
|
||||
DEST="${DEST:-/home/dank/models/Bonsai2-PTQ1_0}"
|
||||
LINK_SRC="${LINK_SRC:-/home/dank/models/Bonsai2-PQ2_0}"
|
||||
|
||||
REPO="prism-ml/Ternary-Bonsai-2-27B-gguf"
|
||||
MMPROJ="Ternary-Bonsai-2-27B-mmproj-Q8_0.gguf"
|
||||
MMPROJ_SIZE=629246976
|
||||
MMPROJ_SHA="6807ede61d570bb86ba34b756a0fa109edc33668604de867c6ea6d8f1d631903"
|
||||
|
||||
# hub-name | dest-name | bytes | sha256 (Hub LFS oid)
|
||||
FILES=(
|
||||
"Ternary-Bonsai-2-27B-PTQ1_0.gguf|Ternary-Bonsai-2-27B-PTQ1_0.gguf|5946648928|53107f530aa52eb00912263ab1ee29bd199261c87cd7b4ad4ca1318c1fe33ee3"
|
||||
)
|
||||
|
||||
mkdir -p "$DEST"
|
||||
status=0
|
||||
|
||||
check() { # file size sha -> 0 when byte- and hash-exact
|
||||
[[ -f "$1" ]] && [[ "$(stat -c%s "$1")" == "$2" ]] \
|
||||
&& sha256sum -c --quiet <(printf '%s %s\n' "$3" "$1")
|
||||
}
|
||||
|
||||
for entry in "${FILES[@]}"; do
|
||||
IFS='|' read -r hubname destname size sha <<<"$entry"
|
||||
f="$DEST/$destname"
|
||||
if check "$f" "$size" "$sha"; then
|
||||
echo "OK $destname"
|
||||
continue
|
||||
fi
|
||||
if (( VERIFY_ONLY )); then
|
||||
echo "MISSING $destname"
|
||||
status=1
|
||||
continue
|
||||
fi
|
||||
echo "FETCHING $destname (from $REPO)"
|
||||
tmp="$f.part"
|
||||
# No resume: `curl -C -` once produced an oversized file on a 31 GiB payload.
|
||||
# -f/--fail-with-body so an HTTP 401/403 body is never written as success.
|
||||
curl -fsSL --fail-with-body --retry 3 --retry-delay 5 \
|
||||
-o "$tmp" "https://huggingface.co/${REPO}/resolve/main/${hubname}"
|
||||
mv "$tmp" "$f"
|
||||
if ! check "$f" "$size" "$sha"; then
|
||||
echo "FAIL $destname (size $(stat -c%s "$f" 2>/dev/null || echo none) != $size, or sha mismatch)"
|
||||
status=1
|
||||
continue
|
||||
fi
|
||||
echo "VERIFIED $destname"
|
||||
done
|
||||
|
||||
# Vision projector: hardlink the identical file when a verified copy exists.
|
||||
mm="$DEST/$MMPROJ"
|
||||
if check "$mm" "$MMPROJ_SIZE" "$MMPROJ_SHA"; then
|
||||
echo "OK $MMPROJ"
|
||||
elif [[ -f "$LINK_SRC/$MMPROJ" ]] && check "$LINK_SRC/$MMPROJ" "$MMPROJ_SIZE" "$MMPROJ_SHA"; then
|
||||
ln "$LINK_SRC/$MMPROJ" "$mm" && echo "HARDLINKED $MMPROJ from $LINK_SRC"
|
||||
elif (( VERIFY_ONLY )); then
|
||||
echo "MISSING $MMPROJ"
|
||||
status=1
|
||||
else
|
||||
echo "FETCHING $MMPROJ (no verified source copy to link)"
|
||||
curl -fsSL --fail-with-body --retry 3 --retry-delay 5 \
|
||||
-o "$mm.part" "https://huggingface.co/${REPO}/resolve/main/${MMPROJ}"
|
||||
mv "$mm.part" "$mm"
|
||||
if ! check "$mm" "$MMPROJ_SIZE" "$MMPROJ_SHA"; then
|
||||
echo "FAIL $MMPROJ"
|
||||
status=1
|
||||
else
|
||||
echo "VERIFIED $MMPROJ"
|
||||
fi
|
||||
fi
|
||||
|
||||
if (( status == 0 )); then
|
||||
echo "DOWNLOAD_DONE $DEST"
|
||||
else
|
||||
echo "DOWNLOAD_FAILED $DEST" >&2
|
||||
fi
|
||||
exit "$status"
|
||||
+27
-18
@@ -20,9 +20,11 @@
|
||||
# thinking tokens 1,476 vs 784 mean — the finetune compressed reasoning, the
|
||||
# base thinks ~1.9x longer per answer.
|
||||
#
|
||||
# The A/B ran at the card's ctx/pool of 262,144. Afterwards Frank lowered both
|
||||
# to 200,704 (RAM + the harness's compaction budget); the rates above are the
|
||||
# ceiling-262,144 record and the production ledger is the 200,704 one.
|
||||
# The A/B ran at the card's ctx/pool of 262,144. Afterwards Frank lowered ctx to
|
||||
# 200,704 (RAM + the harness's compaction budget) and, on 2026-10-18, raised the
|
||||
# pool to 2 x ctx = 401,408 so BOTH slots can carry a full-length request at
|
||||
# once; the rates above are the ceiling-262,144 record and the production ledger
|
||||
# is the ctx-200,704 one.
|
||||
#
|
||||
# WHY THE RUNTIME MOVED WITH THE MODEL: the payload is .hgn — per its own
|
||||
# card it "will not load in transformers, vLLM, or llama.cpp" — and the v2
|
||||
@@ -57,15 +59,21 @@
|
||||
# est-error ≈ 156.2k <= 200,704 — ~9.5× the 3 % error of margin (robust to a
|
||||
# ~25 % under-count). Compaction rides on frank-zealot: input ≈ threshold −
|
||||
# keepRecent(20k) ≈ 112k, inside its 200,704 slot.
|
||||
# HALOGEN_KV_POOL_POSITIONS=200704 — ONE pool behind HALOGEN_KV_SLOTS=2
|
||||
# (ledger form: "2 slots over ONE <pool>-position KV pool, each request up to
|
||||
# HALOGEN_CTX"). With pool == ctx a single MAXIMUM-length request fills the
|
||||
# whole pool. ~28 KiB/position (FLAGS.md) ⇒ ~5.4 GiB, where 262144 cost 7.2
|
||||
# and the 2026-10-06 shape's 524,288 measured 14.6. The engine's own preflight
|
||||
# formula bills a little more than the reservation because it adds block
|
||||
# scratch and ~115 MiB/slot of O(1) state — quote the RESERVATION for the RAM
|
||||
# ledger. FLAGS.md law: "RAM the pool takes is RAM that table loses", so every
|
||||
# position cut is page cache for the 47.68 GiB n-gram table long prompts read.
|
||||
# HALOGEN_KV_POOL_POSITIONS=401408 = SLOTS x CTX (2 x 200,704), the top of the
|
||||
# legal band (the seam enforces ctx <= pool <= slots * ctx). Still ONE pool
|
||||
# behind HALOGEN_KV_SLOTS=2 (ledger form: "2 slots over ONE <pool>-position KV
|
||||
# pool, each request up to HALOGEN_CTX"), but sized so BOTH slots can hold a
|
||||
# MAXIMUM-length request at the same time — the parity with frank-zealot's
|
||||
# 401,408-position pool that Frank ordered on 2026-10-18. The PER-REQUEST
|
||||
# ceiling does NOT move with it: HALOGEN_CTX stays 200,704, because the
|
||||
# checkpoint's native window is 262,144 and the engine refuses to go past it
|
||||
# without HALOGEN_ROPE_YARN (its own 400 text says so). ~28 KiB/position
|
||||
# (FLAGS.md) ⇒ ~10.8 GiB, where pool == ctx (200,704) cost 5.5, 262,144 cost
|
||||
# 7.2 and the 2026-10-06 shape's 524,288 measured 14.6. The engine's preflight
|
||||
# formula bills a little more than the reservation (block scratch + ~115
|
||||
# MiB/slot of O(1) state) — quote the RESERVATION. FLAGS.md law: "RAM the pool
|
||||
# takes is RAM that table loses" — the 47.68 GiB n-gram table surrenders this
|
||||
# much page cache, so the warm prompt-cache ratio is what to re-measure.
|
||||
# HALOGEN_KV_SLOTS=2 restores the w4b-era shape on this box: the MTP
|
||||
# drafter is NOT slot-locked on this engine (batching and speculation
|
||||
# coexist — DFlash-era slots>1 law does NOT carry over; the card's own
|
||||
@@ -74,13 +82,14 @@
|
||||
# Memory ledger against the box's 121.2 GiB unified (the n-gram table is
|
||||
# mmap'd page cache beside the checkpoint, not a pinned charge — the boot
|
||||
# ledger agrees: "mapped 51.2 GB, NOT pinned (Pin::None)"):
|
||||
# weights ~62.9 GiB locked (0.8 of them the vision tower) + KV ~5.4 +
|
||||
# working ~5.7 (HALOGEN_MAX_TOK=8192) ≈ 74 GiB engine-side, with the
|
||||
# weights ~62.9 GiB locked (0.8 of them the vision tower) + KV ~10.8 +
|
||||
# working ~5.7 (HALOGEN_MAX_TOK=8192) ≈ 79.4 GiB engine-side, with the
|
||||
# prompt cache (~1.0 at this ctx; 1.3 was measured at 262144) booked
|
||||
# separately by the ledger; HALOGEN_HOST_RESERVE_GIB=20 keeps the host
|
||||
# honest; baseline co-tenancy (frank-zealot, ComfyUI, Caddy/observ)
|
||||
# ~10-15. ~30 GiB margin. The boot ledger is the referee — read it after
|
||||
# deploy and re-size if the engine prints a clamp.
|
||||
# honest. Against the 121.2 GiB box that leaves ~22 GiB over the reserve for
|
||||
# baseline co-tenancy (frank-zealot, ComfyUI, Caddy/observ ~10-15). The boot
|
||||
# ledger is the referee — read it after deploy and re-size if the engine
|
||||
# prints a clamp.
|
||||
# HALOGEN_SPEC_ADAPT=0 is MANDATORY (card): the default adaptive policy
|
||||
# turns the MTP head off for whole requests after low acceptance and
|
||||
# HALVES decode as a result. Frank accepted the head always-on trade
|
||||
@@ -174,7 +183,7 @@ FRANK_DRAGOON_TOKENIZER_FILES="tokenizer.json chat_template.jinja tokenizer_conf
|
||||
# model_ids for the :7770 backend MUST name this value.
|
||||
HALOGEN_MODEL_ID=frank-dragoon
|
||||
HALOGEN_CTX=200704
|
||||
HALOGEN_KV_POOL_POSITIONS=200704
|
||||
HALOGEN_KV_POOL_POSITIONS=401408
|
||||
HALOGEN_KV_SLOTS=2
|
||||
HALOGEN_HOST_RESERVE_GIB=20
|
||||
# Spec decode: MTP head is INSIDE the checkpoint (do NOT set HALOGEN_MTP_HEAD);
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
[Unit]
|
||||
Description=halogen-flash-server container (Quat3rnion NATIVE v2 abliterated Qwen3.8-Flash-Next, MTP spec decode, 2 slots over ONE 200704-position pool at ctx 200704, vision ON), root systemd unit (dankclaw, frank-dragoon, Strix Halo iGPU)
|
||||
Description=halogen-flash-server container (Quat3rnion NATIVE v2 abliterated Qwen3.8-Flash-Next, MTP spec decode, 2 slots over ONE 401408-position pool at ctx 200704, vision ON), root systemd unit (dankclaw, frank-dragoon, Strix Halo iGPU)
|
||||
After=network-online.target
|
||||
Wants=network-online.target
|
||||
|
||||
|
||||
Reference in New Issue
Block a user