live: heartbeat sync 2026-10-18T23:57:41-07:00 (Refs #11)

This commit is contained in:
Frank
2026-10-18 23:57:41 -07:00
parent 1e7e212e5c
commit 54f15e2a58
6 changed files with 26 additions and 58 deletions
@@ -1,4 +1,4 @@
{
"v": 1,
"lastConsolidatedAt": 1792389229679
"lastConsolidatedAt": 1792392829703
}
@@ -1,6 +1,6 @@
{
"v": 1,
"lastConsolidatedAt": 1792389351747,
"lastConsolidatedAt": 1792392891753,
"lastResult": {
"at": 1792379504653,
"sessions": 1,
@@ -1,4 +1,4 @@
{
"v": 1,
"lastConsolidatedAt": 1792389222321
"lastConsolidatedAt": 1792392882348
}
@@ -1,6 +1,6 @@
{
"v": 1,
"lastConsolidatedAt": 1792389181835,
"lastConsolidatedAt": 1792392961865,
"lastResult": {
"at": 1792381870670,
"sessions": 1,
+18 -50
View File
@@ -24,27 +24,13 @@ providers:
# compactionModel candidate keeps the wire model id valid there.
compactionModel: llamacpp (dankclaw-3090)/frank-zealot
name: "frank-archon"
# Dankserver: BOTH Radeon AI PRO R9700s (gfx1201, TP=2 over the
# 2026-09-20 board swap's root ports, wht6 wire). Engine:
# radiance (StillDeadcode, Apache-2.0) from the digest-pinned
# stilldeadcode/radiance image (index digest sha256:4c3f4845..., 287
# MiB, kernels prebuilt for gfx1201), plain `podman run` through
# scripts/serve-frank-archon.sh, root unit (User=dank, dank's
# rootless podman storage). Serving the OFFICIAL
# Qwen/Qwen3.8-Flash-Next-FP8 at NATIVE modelopt block-FP8
# precision — a rad-convert no-recipe pure copy into
# qwen4exp-fp8-native.rad (172.76 GiB measured by plan-only), built by
# scripts/build-frank-archon-rad.sh against a staging dir whose
# config.json drops quantization_config so the shipped qwen4exp
# plugin's (qwen4exp, "") selection key matches; per-weight
# encodings are untouched. Experts stream from a 40 GiB/rank
# pinned host pool; the 47.7 GiB PLE ngram table is pinned in RAM
# (Frank's call); KV fp8; MTP depth 3 (the checkpoint's own mtp.*
# head); prefix cache 4 GiB host + 64 GiB disk. Chat template:
# froggeric's fixed qwen3.8 v22.5
# (scripts/qwen38-froggeric-v22.5.jinja) via
# --override-chat-template. Rollback: git tag
# frank-archon-freetoken-pr587-20261005 (FreeToken, one card).
# Dankserver: BOTH Radeon AI PRO R9700s (gfx1201, TP=2, wht6 wire) on
# the digest-pinned radiance image, plain `podman run` under a root
# unit. Payload, shape, eras, laws, rollback:
# serverAppConfigs-docs/LLM/frank-archon.md — and
# scripts/frank-archon.env is the authority for every number there.
# contextWindow MUST quote the engine's --max-model-len: radiance
# refuses a longer prompt and never truncates.
contextWindow: 200000
# maxTokens must survive the harness compaction gap AND the chars/4
# token-estimate under-run on code-heavy prompts. The clamp is
@@ -105,40 +91,22 @@ providers:
# compaction model candidate so the wire model id stays valid
# there even from a frank-dragoon session.
compactionModel: llamacpp (dankclaw-3090)/frank-zealot
# halogen-flash-server 0.16.2 (ghcr.io/peonist-ai/
# halogen-flash-server@sha256:0c61bf84, digest-pinned) serving the
# Quat3rnion NATIVE v2 ABLITERATED conversion of orcarouter's pinned
# BF16 Qwen3.8-Flash-Next (qwen38-flash-next-v2-abliterated.hgn,
# ~62 GiB resident, + REQUIRED n-gram table + vision tower, both
# hardlinked from the swift15 dir) on the Strix Halo iGPU: TWO slots
# over ONE 401,408-position KV pool (= 2 x ctx, raised 2026-10-18) at
# HALOGEN_CTX=200704 (the ceiling lowered from the card's 262,144 on
# 2026-10-17; ~10.8 GiB reserved where 200,704 cost 5.5, 262,144 cost
# 7.2 and 524,288 measured 14.6), with the MTP head ALWAYS ON
# (HALOGEN_SPEC_ADAPT=0 — the default adaptive policy halves
# decode). Measured 2026-10-17: 50 t/s greedy decode, 1,604 t/s
# cold prefill at 59k, NIAH 18/18. ABLITERATED base (no refusals:
# 0/5 on the hostile census) and VISION ON (OpenAI image_url
# parts; >=256 px images or it answers "video"). Runbook:
# serverAppConfigs-docs/LLM/frank-dragoon-flash-server-v2-native.md.
# Window the harness may fill — deliberately BELOW the engine
# ceiling (HALOGEN_CTX=200704, lowered from 262,144 on 2026-10-17):
# the clamp below leaves the 45k gap that absorbs the estimate error.
# Both slots SHARE one 401,408-position KV pool — exactly two full
# windows — so TWO maximum-length conversations can be resident at the
# same time (at the old pool == ctx 200,704 only two of <=100k could).
# The per-request ceiling is unchanged, so this window is unchanged.
# halogen flash-server on the Strix Halo iGPU. Payload, shape, eras,
# laws, rollback: serverAppConfigs-docs/LLM/frank-dragoon.md — and
# scripts/frank-dragoon.env is the authority for every number there.
# contextWindow sits deliberately BELOW the engine ceiling
# (HALOGEN_CTX): the clamp below leaves the gap that absorbs the
# harness's chars/4 under-count, so a window equal to the ceiling
# would make every long request a guaranteed 400.
contextWindow: 155648
# The clamp is max_tokens = min(maxTokens, contextWindow − est −
# 4096); the chars/4 est UNDER-counts the rendered prompt by ~3 %
# on code/agent traffic (measured 2026-10-07: est 152,938 ->
# server 157,388). Worst-case request = window − 4096 + 3 % est
# error ≈ 156.2k, inside the 200,704 engine ceiling with ~44k
# margin (robust to a ~25 % under-count). Compaction fires at 85 %
# of the window = 132,300 est ≈ 136k measured — the 180k target of
# the 262,144 era is unreachable now (the window itself is 155,648)
# — and rides on frank-zealot: input ≈ threshold − keepRecent(20k)
# ≈ 112k, inside its 200,704 slot.
# error ≈ 156.2k, inside the engine ceiling with ~44k of margin
# (robust to a ~25 % under-count). Compaction fires at 85 % of the
# window and rides on frank-zealot: input ≈ threshold −
# keepRecent(20k), comfortably inside its per-slot window.
# HALOGEN_MAX_TOKENS_CAP in frank-dragoon.env must stay >= this.
maxTokens: 45056
reasoning: true