feat: replace llama.cpp Tiel-Coder with FreeToken v0.1.2 serving Qwen3.6-35B-A3B-FP8

Replace the llama.cpp Tiel-Coder-35B-A3B (UD-Q6_K_XL) deployment with
FreeToken v0.1.2 — an edge-native MoE serving engine that keeps expert
weights in host RAM with an LRU GPU cache and paged KV with radix prefix
sharing.

Model: Qwen/Qwen3.6-35B-A3B-FP8 (modelopt block-fp8, 128×128 weight_block_size).
Architecture: qwen3_5_moe — 40 layers (10 attention + 30 GDN), 256 experts/8 used.

Production config:
- 2048 MoE cache slots (20% of 10240 expert instances, ~6.0 GiB VRAM)
- 453005 KV tokens (~8.6 GiB, ~20 KiB/token for 10 attn layers)
- 4 max-running-requests (mamba slot pool: 24 slots)
- bs4 decode: ~87 tok/s aggregate, ~30 tok/s bs1
- Prefill: ~1,130 tok/s at 4k prompt

Quirk: FreeToken v0.1.2 supports modelopt block-fp8, NVFP4, and per-tensor
FP8, but NOT compressed-tensors per-channel FP8 (Huihui/Ornith both failed).
Live KV resize via ft ctl cache grows KV by shrinking MoE; reverse needs restart.

Old llama.cpp scripts archived to .AGENTTRASH/.
Full benchmark sweep (bs1/bs3/bs4/bs6, 8k–700k KV) documented in
docs/LLM/freetoken-benchmark-sweep.md.

scripts/frank-default-small.env

@@ -1,71 +1,51 @@
-# Tunables for frank-default-small (Tiel-Coder-35B-A3B UD-Q6_K_XL,
-# qwen35moe hybrid MoE, Dankserver, RTX 3090).
+# Tunables for frank-default-small (Qwen3.6-35B-A3B-FP8,
+# qwen3_5_moe hybrid MoE, Dankserver, RTX 3090).
 #
-# Runtime: upstream llama.cpp (latest master, CUDA build with
-# -DGGML_CUDA_FA_ALL_QUANTS=ON -DCMAKE_CUDA_ARCHITECTURES=86) on RTX 3090
-# 24 GB; Ryzen 7 7700X; 62 GB RAM.  Replaced the ik_llama.cpp Ornith
-# CRACK Q8_0 deployment on 2026-09-01.
+# 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
+# layers page KV (~20 KiB/token); the 30 GDN layers use constant-size
+# per-sequence state (mamba slot pool).
 #
-# Model: peculiar-ragdoll/Tiel-Coder-35B-A3B-GGUF, UD-Q6_K_XL quant
-# (30 GB).  Architecture: qwen35moe — 40 layers, full_attention_interval=4
-# (10 attention + 30 SSM/Mamba), 256 experts / 8 used, 2 KV heads × 256
-# dim, embedding 2048, rope dimension_count 64, rope freq_base 10 000 000.
-# Text-only (no mmproj).  Built-in chat template in the GGUF.
+# Model: Qwen/Qwen3.6-35B-A3B-FP8 (HuggingFace safetensors, modelopt
+# block-fp8 quant, 128x128 weight_block_size, 35 GB on disk).
+# Architecture: qwen3_5_moe — 40 layers (10 attention + 30 GDN),
+# 256 experts / 8 used, 2 KV heads × 256 dim, full_attention_interval=4.
 #
-# Expert offload: --cpu-moe keeps all 256 expert weights in CPU RAM.
-# Non-expert weights (~1.6 GB) fit entirely on the 24 GB card.
+# Replaced the llama.cpp Tiel-Coder-35B-A3B (UD-Q6_K_XL) deployment on
+# 2026-09-02.  The old serve script and env are in .AGENTTRASH/.
 #
-# KV cache: Q8_0 for both K and V.  With CTX_SIZE=200700 (66 900 per slot
-# × 3 parallel) and 10 attention layers, the KV cache is ~2 GB — the
-# requested 2 GB VRAM buffer.
+# 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
+# `ft ctl cache --kv N --moe M` can grow KV by shrinking MoE without
+# restart (the reverse requires a restart).
 #
-# WARNING: do not name any variable TEMP, HOST, or GPU_MAX_HEAP_SIZE. 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.
+# WARNING: do not name any variable TEMP, HOST, or GPU_MAX_HEAP_SIZE.
+# 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 file (UD-Q6_K_XL, 30 GB, text-only MoE hybrid).
-MODEL_DIR=/home/dank/models/tiel-coder-35b
-MODEL_FILE=Tiel-Coder-35B-A3B-UD-Q6_K_XL.gguf
+# Model directory (safetensors checkpoint, 35 GB).
+MODEL_DIR=/home/dank/models/qwen36-35b-a3b-fp8
 
-# Upstream llama.cpp llama-server (CUDA build, latest master).
-# Built with -DGGML_CUDA_FA_ALL_QUANTS=ON -DCMAKE_CUDA_ARCHITECTURES=86
-# using the staged CUDA 13.2 toolkit at /home/dank/cuda13-stage.
-LLAMA_SERVER=/home/dank/llama.cpp/build/bin/llama-server
+# FreeToken venv (uv-managed CPython 3.12, ft CLI, flashinfer, torch+cu130).
+FT_VENV=/home/dank/freetoken/.venv
+CUDA_HOME_DIR=/home/dank/freetoken/.venv/lib/python3.12/site-packages/nvidia/cu13
 
 ALIAS=frank-default-small
 BIND_ADDR=0.0.0.0
 LISTEN_PORT=7770
 
-# Total context budget across all parallel slots.  200700 / 3 = 66 900
-# per slot.  The model's context_length is 262 144, so 66 900 is well
-# within range.
-CTX_SIZE=200700
-PARALLEL=3
+# Expert residency: 2048 of 10240 total expert instances (20%).
+# Each slot ~3 MiB → ~6.0 GiB VRAM for the MoE cache.
+MOE_CACHE_SIZE=2048
 
-# Expert offload: all MoE expert weights pinned to CPU RAM.
-CPU_MOE=1
+# KV pool: 453005 tokens (~8.6 GiB at ~20 KiB/token).
+# Covers 4 concurrent sessions at ~113k unique context each
+# (more with radix prefix sharing for common prefixes).
+NUM_TOKENS=453005
 
-# All non-expert layers on GPU (auto-fit will determine the exact count;
-# with --cpu-moe the non-expert footprint is ~1.6 GB, leaving ample room).
-N_GPU_LAYERS=999
-
-# KV cache: Q8_0 for both K and V.  ~2 GB total for 200700 ctx × 3 slots
-# × 10 attention layers × (2 KV heads × 256 dim × 1 byte) × 2 (K+V).
-CACHE_TYPE_K=q8_0
-CACHE_TYPE_V=q8_0
-
-VERBOSITY=2
-
-# Prefill batch sizing.
-BATCH_SIZE=4096
-UBATCH_SIZE=1024
-
-# Sampling.
-TEMPERATURE=1.0
-TOP_K=20
-TOP_P=0.95
-MIN_P=0.0
-
-# Compute threads lowered so other processes keep a core.
-THREADS=7
-THREADS_HTTP=8
+# Max concurrent decoding requests (also caps mamba slot allocation).
+MAX_RUNNING_REQUESTS=4