frank-default-small: swap to ornith-ai/Ornith-1.5-35B-A3B-FP8 + Qwen-Sharp chat template

- MODEL_DIR -> /home/dank/models/ornith-1.5-35b-a3b-fp8 (37 GB,
  compressed-tensors per-channel FP8, same qwen3_5_moe arch/mass as
  Huihui so the pool plan carries unchanged)
- CHAT_TEMPLATE_FILE -> scripts/chat-template.jinja (Qwen-Sharp, shared
  with frank-default-large); the checkpoint's stock template
  mis-serializes reasoning per its own model card
- env/serve/unit headers updated; loader-risk + BF16-does-not-fit note

scripts/frank-default-small.env

@@ -1,30 +1,42 @@
-# Tunables for frank-default-small (batsclamp/Huihui-Qwen3.6-35B-A3B-
+# Tunables for frank-default-small (ornith-ai/Ornith-1.5-35B-A3B-FP8,
-# abliterated-FP8, qwen3_5_moe hybrid MoE, Dankserver, RTX 3090).
+# qwen3_5_moe hybrid MoE, Dankserver, RTX 3090).
 #
 # Runtime: FreeToken v0.1.2 (edge-native MoE serving engine).
 # Installed via uv into /home/dank/freetoken/.venv (CPython 3.12).
 # FreeToken keeps expert weights in host RAM with an LRU GPU cache;
-# KV is paged with radix prefix sharing.  Only the 10 full-attention
+# KV is paged with radix prefix sharing.  Only the full-attention
-# layers page KV (~20 KiB/token); the 30 GDN layers use constant-size
+# layers page KV (~20 KiB/token); the GDN layers use constant-size
 # per-sequence state (mamba slot pool).
 #
-# Model: batsclamp/Huihui-Qwen3.6-35B-A3B-abliterated-FP8 (HuggingFace
+# Model: ornith-ai/Ornith-1.5-35B-A3B-FP8 (HuggingFace safetensors,
-# safetensors).  Abliterated (refusal-removed) Huihui finetune of
+# compressed-tensors FP8, ~35 GB on disk).  Ornith-1.5 is a
-# Qwen3.6-35B-A3B, quantized to the SAME native format the previous
+# self-improvement RL-trained coding/agentic model built on Qwen3.5 +
-# official checkpoint used: modelopt block-fp8 e4m3, dynamic
+# Gemma4; ~35B total / ~3B active (A3B), same qwen3_5_moe arch and the
-# activations, 128x128 weight_block_size, bf16 weight_scale_inv;
+# same ~35 GB FP8 weight mass as the Huihui FP8 it replaces, so the
-# lm_head/embeds/norms/router gates/visual/MTP kept BF16.  Un-fused
+# FreeToken memory/pool plan carries unchanged.  Native 262144 context.
-# per-expert layout matches the official Qwen FP8, so the FreeToken
+# Reasoning model: each assistant turn opens with a <think> block,
-# loader path is identical; ~35 GB on disk.  Ships the fixed Qwen3.6
+# surfaced separately via the qwen3 reasoning parser (REASONING=on).
-# chat template (froggeric/Qwen-Fixed-Chat-Templates) — FreeToken
+# The RTX 3090 has no FP8 tensor cores, so FP8 runs via upcast (as
-# reads it from the checkpoint; no template override needed.
+# Huihui did).  Replaced batsclamp/Huihui-Qwen3.6-35B-A3B-abliterated-
-# Replaced Qwen/Qwen3.6-35B-A3B-FP8 on 2026-08-30; pool numbers
+# FP8 on 2026-09-01.
-# unchanged — same arch, same weight mass.
+#
+# RISK: this is compressed-tensors PER-CHANNEL FP8 (weights strategy
+# "channel"), the one FP8 layout the runbook flags as a FreeToken v0.1.2
+# loader bug.  Load is being verified live; 37 GB fits RAM so a failure is
+# a clean crash-loop, not a swap-thrash.  The BF16 sibling
+# (ornith-ai/Ornith-1.5-35B-A3B, ~70 GB) does NOT fit this 62 GB box —
+# never fall back to it (that is the 2026-09-01 Thomson outage mode).
+#
+# Chat template: the checkpoint's stock Qwen template is WRONG for this
+# model (its own model card says so and ships a modified
+# chat_template.jinja for the agentic benchmarks).  The fleet standard is
+# the Qwen-Sharp template (scripts/chat-template.jinja, repo-tracked and
+# shared with frank-default-large) — see CHAT_TEMPLATE_FILE below.
 #
 # Pool sizing: GPU budget (~17.6 GiB after weights + fixed caches) is
 # split MoE-first.  MOE_CACHE_SIZE pins resident expert slots; NUM_TOKENS
 # pins the KV pool.  At 2048 slots / 453005 tokens the split is ~20%
-# residency (2048/10240) and ~8.6 GiB KV.  Live resize via
+# residency and ~8.6 GiB KV.  Live resize via
 # `ft ctl cache --kv N --moe M` can grow KV by shrinking MoE without
 # restart (the reverse requires a restart).
 #
@@ -32,8 +44,14 @@
 # The ROCm HIP runtime reads those bare env vars (see the 2026-08-13 TEMP
 # incident).  This box is NVIDIA/CUDA, but the rule keeps the env portable.
 
-# Model directory (safetensors checkpoint, ~35 GB, block-fp8).
+# Model directory (safetensors checkpoint, ~35 GB, compressed-tensors FP8).
-MODEL_DIR=/home/dank/models/huihui-qwen36-35b-a3b-abliterated-fp8
+MODEL_DIR=/home/dank/models/ornith-1.5-35b-a3b-fp8
+
+# Chat template: the fleet Qwen-Sharp template (repo-tracked), shared with
+# frank-default-large.  The checkpoint's stock template mis-serializes
+# reasoning for this model; the Sharp template renders the <think> trail and
+# the Qwen3.5-family tool-call shape correctly.
+CHAT_TEMPLATE_FILE=/home/dank/serverAppConfigs/scripts/chat-template.jinja
 
 # FreeToken venv (uv-managed CPython 3.12, ft CLI, flashinfer, torch+cu130).
 FT_VENV=/home/dank/freetoken/.venv

scripts/frank-default-small.service

@@ -1,5 +1,5 @@
 [Unit]
-Description=FreeToken Huihui-Qwen3.6-35B-A3B-abliterated-FP8 qwen3_5_moe hybrid MoE (dankserver, frank-default-small)
+Description=FreeToken Ornith-1.5-35B-A3B-FP8 qwen3_5_moe hybrid MoE (dankserver, frank-default-small)
 After=network-online.target
 Wants=network-online.target
 

scripts/serve-frank-small.sh

@@ -1,6 +1,6 @@
 #!/usr/bin/env bash
-# Serve batsclamp/Huihui-Qwen3.6-35B-A3B-abliterated-FP8 as
+# Serve ornith-ai/Ornith-1.5-35B-A3B-FP8 as frank-default-small via
-# frank-default-small via FreeToken on Dankserver (RTX 3090, 60 GB RAM).
+# FreeToken on Dankserver (RTX 3090, 62 GB RAM).
 #
 # Runtime: FreeToken v0.1.2 (edge-native MoE serving engine).
 #   - Experts live in host RAM; an LRU cache of expert slots is resident
@@ -10,14 +10,14 @@
 #     hold constant-size per-sequence state (mamba slots).
 #   - OpenAI-compatible API (/v1/chat/completions, /v1/models).
 #
-# Model: batsclamp/Huihui-Qwen3.6-35B-A3B-abliterated-FP8 — Huihui
+# Model: ornith-ai/Ornith-1.5-35B-A3B-FP8 — Ornith-1.5 (self-improvement
-# abliterated finetune of Qwen3.6-35B-A3B in the same native modelopt
+# RL-trained coding/agentic model over Qwen3.5 + Gemma4), 35B total / 3B
-# block-fp8 format (e4m3, 128x128 weight_block_size, un-fused per-expert
+# active, in compressed-tensors per-channel FP8 (float-quantized, dynamic
-# layout) as the official FP8 checkpoint this service ran before; the
+# token activations).  Same qwen3_5_moe arch and ~35 GB FP8 mass as the
-# FreeToken loader path is unchanged.  40 layers (10 attention + 30
+# Huihui FP8 it replaces: 40 layers (10 full-attention + 30 GDN), 256
-# GDN), 256 experts / 8 used, 2 KV heads × 256 dim.  ~35 GB on disk.
+# experts / 8 used, 2 KV heads × 128 dim, native 262144 context.  FreeToken
-# The checkpoint ships the fixed Qwen3.6 chat template; FreeToken reads
+# loads a Qwen-Sharp chat template over it (see CHAT_TEMPLATE_FILE in the
-# it from the model directory.
+# env file); the checkpoint's own stock template mis-serializes reasoning.
 #
 # Pool sizing: the GPU budget is split MoE-first between expert residency
 # and KV pages.  --moe-cache-size pins the number of resident expert
@@ -39,7 +39,7 @@ if [[ -f "$ENV_FILE" ]]; then
 fi
 
 # --- defaults (used when the env file does not set them) -----------------
-: "${MODEL_DIR:=/home/dank/models/huihui-qwen36-35b-a3b-abliterated-fp8}"
+: "${MODEL_DIR:=/home/dank/models/ornith-1.5-35b-a3b-fp8}"
 : "${ALIAS:=frank-default-small}"
 : "${BIND_ADDR:=0.0.0.0}"
 : "${LISTEN_PORT:=7770}"