Switch frank-default-small to Ornith Q6_K

docs/LLM/Dankserver System Specs.md

@@ -60,12 +60,12 @@ Dankserver is an NVIDIA GPU server for vLLM inference. The host runs Ubuntu 26.0
 
 ### Active Services
 
-The active service on Dankserver is `frank-default-small.service`: the llama.cpp Ornith-1.5-35B-A3B CRACK (Q8_0) on port 7770, with MoE expert weights held on CPU via `--override-tensor`. The earlier PP worker (`frank-pp-worker.service`) was retired with the 2026-08-17 redo. See [[deployment]].
+The active service on Dankserver is `frank-default-small.service`: the llama.cpp Ornith-1.5-35B-A3B CRACK (Q6_K) on port 7770, with expert tensors from blocks 0 through 36 held on CPU via `--override-tensor`. The earlier PP worker (`frank-pp-worker.service`) was retired with the 2026-08-17 redo. See [[deployment]].
 
 ## Hardware Relevance
 
-- 24 GB VRAM is close to saturated with this config. At idle ~22.1 GB used, leaving ~2.0 GiB headroom at `--n-cpu-moe 36` (4 layers' experts on GPU) with `ubatch 3072` and a dual-f16 KV pool of 13680 MiB at 700416 cells (2026-08-14 refit; the same pool at 800768 cells OOM'd the load warmup). A llama-bench sweep ([[frank-default-small-benchmark]]) guided this state: prefill is the priority, so `ubatch 3072` stays (top prefill tier), and `ncmoe 36` was raised from 35 to fix a cuBLAS OOM-class crash during big prefills by freeing ~0.75 GiB peak VRAM at only -2% prefill. If it still crashes under load, add `GGML_CUDA_FORCE_MMQ=1`. If decode speed matters more, use `ncmoe 32 - ubatch 1024` (gen ~48.7 t/s). The tuning guidance applied to the retired `frank-default-small` service; the PP worker's allocations are fixed at load.
-- 62 GB system RAM holds the ~29 GB of MoE expert weights with room to spare (~57 GB available after load).
+- 24 GB VRAM hosts the Q6_K non-expert tensors and expert tensors from blocks 37 through 40. The new context pool contains 589824 f16 cells across three 196608-token slots. The configuration targets about 1.5 to 2 GiB of free VRAM for checkpoint overhead. Verify the actual headroom after each model restart with `nvidia-smi` and the service journal.
+- 62 GB system RAM holds the ~22.76 GiB of CPU expert tensors with room to spare.
 - 127 GB swap is available for CPU KV cache paging if GPU KV offload is disabled.
 - Root disk has approximately 140 GB free. Store the GGUF models under `/home/dank/models` (about 32 GB total).
 

docs/LLM/dealign-ornith-crack.md

@@ -1,7 +1,7 @@
 ---
 type: reference
 title: dealign Ornith-1.5-35B-A3B CRACK on Dankserver
-description: dealignai/Ornith-1.5-35B-A3B-UNCENSORED (CRACK Q8_0) served by upstream llama.cpp on the RTX 3090. MoE model with the MTP head bundled; expert weights offloaded to CPU with an explicit --override-tensor. frank-default-small.
+description: dealignai/Ornith-1.5-35B-A3B-UNCENSORED (CRACK Q6_K) served by upstream llama.cpp on the RTX 3090. MoE model with the MTP head bundled; selected expert weights offloaded to CPU with an explicit --override-tensor. frank-default-small.
 tags:
   - llamacpp
   - cuda
@@ -14,12 +14,13 @@ timestamp: 2026-08-24
 
 # dealign Ornith-1.5-35B-A3B CRACK on Dankserver
 
-> **Active 2026-08-24.** `frank-default-small` serves this model. It
-> replaced the HauhauCS Qwen3.6-35B-A3B Q6_K_P service on the same upstream
-> llama.cpp binary (b10121).
+> **Active 2026-08-25.** `frank-default-small` serves this model. It
+> replaced the CRACK Q8_0 file on the same upstream llama.cpp binary (b10121).
+> The Q6_K file was downloaded from the model repository and validated with
+> `llama-gguf` before installation.
 
 `frank-default-small` (port 7770) serves
-`Ornith-1.5-35B-A3B-CRACK-Q8_0.gguf` from `dealignai` on Hugging Face. It is
+`Ornith-1.5-35B-A3B-CRACK-Q6_K.gguf` from `dealignai` on Hugging Face. It is
 a **MoE** model (256 experts, 8 active) with a hybrid GatedDeltaNet (SSM) +
 attention architecture, abliterated (uncensored). The MTP speculative head
 is **bundled in the GGUF**, so `draft-mtp` auto-detects it from the target
@@ -29,26 +30,36 @@ model.
 
 | File | Size |
 |------|------|
-| `Ornith-1.5-35B-A3B-CRACK-Q8_0.gguf` | 37.2 GB |
-| `mmproj-Ornith-1.5-35B-A3B-f16.gguf` (vision) | 0.90 GB |
+| `Ornith-1.5-35B-A3B-CRACK-Q6_K.gguf` | 27.2 GiB |
+| `mmproj-Ornith-1.5-35B-A3B-f16.gguf` (vision) | 0.84 GiB |
 
 Store the files under `/home/dank/models/Ornith-1.5-35B-A3B-CRACK/`.
 
-## Expert Offload — `--override-tensor`, not `--n-cpu-moe`
+## Layer Audit and Expert Offload
 
-The service pins the MoE expert weights to CPU with an **explicit tensor
-override** rather than the layer-counter `--n-cpu-moe`. The override uses
-the same internal regex that `--cpu-moe` / `--n-cpu-moe` expand:
+The Q6_K file contains 41 transformer blocks and 753 tensors. Each block
+contains 630 MiB of expert tensors. Blocks 3, 7, 11, 15, 19, 23, 27, 31,
+35, and 39 contain full attention. Block 40 contains the bundled MTP head.
+
+The non-expert tensors total 2.02 GiB. The layer totals are 660.96 MiB for
+ordinary SSM blocks, 655.81 MiB for full-attention blocks, and 662.40 MiB for
+block 40. The expert portion is 630 MiB in every block.
+
+The service keeps all non-expert tensors on CUDA. It also keeps expert tensors
+from blocks 37 through 40 on CUDA. Those blocks sit next to the output and
+include the bundled MTP head. Expert tensors from blocks 0 through 36 move to
+CPU through this explicit override:
 
 ```
-OVERRIDE_EXPERT=\.ffn_(up|down|gate|gate_up)_(ch|)exps
+OVERRIDE_EXPERT='blk\.(0|1|2|3|4|5|6|7|8|9|10|11|12|13|14|15|16|17|18|19|20|21|22|23|24|25|26|27|28|29|30|31|32|33|34|35|36)\.ffn_(up|down|gate|gate_up)_(ch|)exps'
 OVERRIDE_BUFT=CPU
 ```
 
-This is passed as `--override-tensor "$OVERRIDE_EXPERT=$OVERRIDE_BUFT"`.
-Only the FFN expert tensors move to CPU; every other tensor stays on the
-24 GB card up to `--gpu-layers`. To list the buffer type names on the host,
-run a bogus override and read the error:
+This places 3.08 GiB of expert weights on CUDA and 22.76 GiB on CPU. The
+selection leaves the output-adjacent blocks and MTP path on CUDA while the
+f16 context cache uses the remaining GPU budget.
+
+To list the buffer type names on the host, run a bogus override:
 
 ```
 llama-server --model <ANY> -c 64 -ot foo=BADTYPE
@@ -72,22 +83,27 @@ proven best. `--spec-draft-n-max 4` is the knee.
 
 ```
 llama-server \
-  -m Ornith-1.5-35B-A3B-CRACK-Q8_0.gguf \
+  -m Ornith-1.5-35B-A3B-CRACK-Q6_K.gguf \
   --mmproj mmproj-Ornith-1.5-35B-A3B-f16.gguf \
   --chat-template-file chat-template.jinja \
   --spec-type draft-mtp \
-  --override-tensor "\.ffn_(up|down|gate|gate_up)_(ch|)exps=CPU" \
-  -ngl all -c 688128 -np 3 \
+  --override-tensor "blk\\.(0|1|2|3|4|5|6|7|8|9|10|11|12|13|14|15|16|17|18|19|20|21|22|23|24|25|26|27|28|29|30|31|32|33|34|35|36)\\.ffn_(up|down|gate|gate_up)_(ch|)exps=CPU" \
+  -ngl all -c 589824 -np 3 \
   --cache-type-k f16 --cache-type-v f16 --no-mmap \
   --temp 1.0 --top-p 0.95 --top-k 20
 ```
 
 ## Known Issues
 
-1. The 37.2 GB Q8_0 file is large for the 62 GB RAM box; lower `PARALLEL`
-   or switch quant if VRAM or RAM is tight.
-2. `--n-cpu-moe` remains supported; this service deliberately prefers the
-   explicit tensor override so only the expert tensors reach CPU.
+1. The 27.2 GiB Q6_K file and 0.84 GiB vision projector fit the 62 GiB RAM
+   box with the selected expert offload.
+2. The context pool contains 589824 cells. Three slots receive 196608 cells
+   each, which is 64K below the model context per stream.
+3. `--n-cpu-moe` remains supported. This service prefers the explicit tensor
+   override because it preserves the final four expert blocks on CUDA.
+4. The service uses the Qwen-Sharp template from
+   `peculiar-ragdoll/Qwen-Sharp-Chat-Templates` through
+   `scripts/chat-template.jinja`.
 
 ## See Also
 

docs/LLM/deployment.md

@@ -17,7 +17,7 @@ repository to a running service.
 
 | Target | Host | Runtime | Service | Port |
 |--------|------|---------|---------|------|
-| Dankserver | dank@dankserver | upstream llama.cpp (Ornith-1.5-35B-A3B-CRACK Q8_0, native MTP draft, MoE experts offloaded via --override-tensor) | frank-default-small | 7770 |
+| Dankserver | dank@dankserver | upstream llama.cpp (Ornith-1.5-35B-A3B-CRACK Q6_K, native MTP draft, selected MoE experts offloaded via --override-tensor) | frank-default-small | 7770 |
 | Dankclaw | dank@dankclaw | CIRU ROCmFPX llama.cpp (Qwen3.8-27B-Uncensored-ROCmFP4 Q6 AGENT, standalone MTP Q4_0 head, 262K ctx / 1 slot) | frank-default-large | 7770 |
 
 ## Workflow
@@ -122,17 +122,17 @@ curl -sf http://127.0.0.1:7770/health
 ### Dankserver — frank-default-small
 
 The active service on Dankserver is upstream llama.cpp (b10121) serving
-`Ornith-1.5-35B-A3B-CRACK-Q8_0.gguf` under the alias `frank-default-small`
+`Ornith-1.5-35B-A3B-CRACK-Q6_K.gguf` under the alias `frank-default-small`
 on port 7770. It runs the stock `/home/dank/llama.cpp/build/bin/llama-server`
-binary with the MoE expert weights pinned to CPU by an explicit
-`--override-tensor` regex (the expert `\.ffn_(up|down|gate|gate_up)_(ch|)exps`
-pattern, buffer type `CPU`) and 7 CPU threads, native `draft-mtp`
-speculative decoding (bundled MoE MTP head, n_max 4), f16 KV, context
-688128 (3 slots of ~229K each, 32k headroom per stream), batch 4096 /
-ubatch 2048, `--no-mmap`. The
-unit is `scripts/frank-default-small.service`. It launches through
+binary with expert tensors from blocks 0 through 36 pinned to CPU by an
+explicit `--override-tensor` regex. Expert tensors from blocks 37 through 40
+stay on CUDA. The service uses 7 CPU threads, native `draft-mtp`
+speculative decoding, f16 KV, context 589824 (3 slots of 196608 each), batch
+4096, ubatch 2048, and `--no-mmap`. The unit is
+`scripts/frank-default-small.service`. It launches through
 `scripts/serve-frank-small.sh`, which reads tunables from
-`scripts/frank-default-small.env`. See [[dealign-ornith-crack]].
+`scripts/frank-default-small.env`. The chat template comes from
+`peculiar-ragdoll/Qwen-Sharp-Chat-Templates`. See [[dealign-ornith-crack]].
 
 ### Dankclaw — frank-default-large
 

docs/LLM/index.md

@@ -19,8 +19,8 @@ graph LR
         A[RTX 3090 24GB + 62GB RAM]
         B[upstream llama.cpp b10121
 frank-default-small, port 7770]
-        C[Ornith-1.5-35B-A3B-CRACK Q8_0
-native MTP, MoE experts offloaded via override, 688128 ctx]
+        C[Ornith-1.5-35B-A3B-CRACK Q6_K
+native MTP, selected experts offloaded via override, 589824 ctx]
         A --> B --> C
     end
 

scripts/frank-default-small.env

@@ -1,34 +1,37 @@
-# Tunables for frank-default-small (Ornith-1.5-35B-A3B CRACK Q8_0, MTP MoE,
+# Tunables for frank-default-small (Ornith-1.5-35B-A3B CRACK Q6_K, MTP MoE,
 # Dankserver, RTX 3090).
 #
 # Runtime: upstream llama.cpp llama-server (b10121). RTX 3090 24 GB; Ryzen
 # 7 7700X; 62 GB RAM. MoE expert weights pinned to CPU with an explicit
 # --override-tensor regex (not --n-cpu-moe). Replaced the HauhauCS
-# Qwen3.6-35B-A3B model on 2026-08-24.
+# Qwen3.6-35B-A3B model on 2026-08-24. Switched from CRACK Q8_0 to Q6_K on
+# 2026-08-25.
 #
 # 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 files (Q8_0 CRACK GGUF + f16 mmproj vision encoder + MTP head bundled
+# Model files (Q6_K CRACK GGUF + f16 mmproj vision encoder + MTP head bundled
 # in the GGUF). Dealign2026-08-24.
 MODEL_DIR=/home/dank/models/Ornith-1.5-35B-A3B-CRACK
-MODEL_FILE=Ornith-1.5-35B-A3B-CRACK-Q8_0.gguf
+MODEL_FILE=Ornith-1.5-35B-A3B-CRACK-Q6_K.gguf
 MMPROJ_FILE=mmproj-Ornith-1.5-35B-A3B-f16.gguf
 # Upstream llama.cpp binary (b10121).
 LLAMA_SERVER=/home/dank/llama.cpp/build/bin/llama-server
 
-# Qwen-Sharp chat template (tracked in the repo, staged on the host).
+# Qwen-Sharp chat template from peculiar-ragdoll (tracked in the repo,
+# staged on the host).
+# Source: https://huggingface.co/peculiar-ragdoll/Qwen-Sharp-Chat-Templates
 CHAT_TEMPLATE_FILE=/home/dank/serverAppConfigs/scripts/chat-template.jinja
 
 ALIAS=frank-default-small
 BIND_ADDR=0.0.0.0
 LISTEN_PORT=7770
 
-# 3 parallel slots x 229376 ctx each (262144 lowered by 32k per stream for
-# cache headroom) = 688128 total context pool. The per-slot 229376 matches
-# the per-slot context advertised in models.json (229376).
-CTX_SIZE=688128
+# 3 parallel slots x 196608 ctx each (262144 lowered by 64k per stream for
+# cache and checkpoint headroom) = 589824 total context pool. The per-slot
+# 196608 matches the per-slot context advertised in models.json (196608).
+CTX_SIZE=589824
 PARALLEL=3
 
 # Spec: MoE model with a bundled MTP head -> native draft-mtp, n_max 4
@@ -38,11 +41,12 @@ SPEC_DRAFT_N_MAX=4
 SPEC_DRAFT_N_MIN=0
 SPEC_DRAFT_P_MIN=0.0
 
-# Expert offload via explicit tensor override (replaces --n-cpu-moe). Only
-# the MoE FFN exps tensors move to CPU; everything else up to --gpu-layers
-# stays on the 24 GB card. The regex is the same one the built-in --cpu-moe
-# expands. Buffer type "CPU" is the CPU buft name on this host.
-OVERRIDE_EXPERT='\.ffn_(up|down|gate|gate_up)_(ch|)exps'
+# Expert offload via explicit tensor override (replaces --n-cpu-moe). The Q6_K
+# layer audit found 630 MiB of expert tensors per block. Keep the four
+# output-adjacent blocks 37-40 on CUDA, including the bundled MTP block 40.
+# Move expert tensors from blocks 0-36 to CPU. All non-expert tensors stay on
+# the 24 GB card. Buffer type "CPU" is the CPU buft name on this host.
+OVERRIDE_EXPERT='blk\.(0|1|2|3|4|5|6|7|8|9|10|11|12|13|14|15|16|17|18|19|20|21|22|23|24|25|26|27|28|29|30|31|32|33|34|35|36)\.ffn_(up|down|gate|gate_up)_(ch|)exps'
 OVERRIDE_BUFT=CPU
 N_GPU_LAYERS=all
 

scripts/frank-default-small.service

@@ -1,5 +1,5 @@
 [Unit]
-Description=llama.cpp Ornith-1.5-35B-A3B CRACK Q8_0 MTP + Sharp template (dankserver, frank-default-small)
+Description=llama.cpp Ornith-1.5-35B-A3B CRACK Q6_K MTP + Qwen-Sharp template (dankserver, frank-default-small)
 After=network-online.target
 Wants=network-online.target
 

scripts/serve-frank-small.sh

@@ -1,16 +1,14 @@
 #!/usr/bin/env bash
-# Serve dealignai/Ornith-1.5-35B-A3B-UNCENSORED (CRACK, Q8_0) as
+# Serve dealignai/Ornith-1.5-35B-A3B-UNCENSORED (CRACK, Q6_K) as
 # frank-default-small via upstream llama.cpp on Dankserver (RTX 3090, MoE
 # experts on CPU).
 #
 # Runtime: upstream llama.cpp llama-server (b10121). The model is a 35B-total /
 # 3B-active MoE-dense hybrid with the MTP speculative head bundled in the GGUF.
 #
-# Expert offload: instead of the layer-counter --n-cpu-moe, the expert (FFN
-# exps) tensors are pinned to CPU explicitly with --override-tensor. The
-# pattern is the internal MoE exps regex (the same one --n-cpu-moe expands),
-# so only the Mixture-of-Experts weight tensors move to CPU; every other
-# tensor stays on the GPU up to --gpu-layers.
+# Expert offload: the expert (FFN exps) tensors selected by the layer-bounded
+# --override-tensor pattern are pinned to CPU. The final output-adjacent
+# blocks stay on CUDA with all non-expert tensors, including the MTP head.
 #
 # Spec: MoE model with a bundled MTP head -> native draft-mtp (the head is
 # auto-detected from the target model; b10121 creates the MTP draft context
@@ -32,7 +30,7 @@ fi
 
 # --- defaults (used when the env file does not set them) -----------------
 : "${MODEL_DIR:=/home/dank/models}"
-: "${MODEL_FILE:=Ornith-1.5-35B-A3B-CRACK-Q8_0.gguf}"
+: "${MODEL_FILE:=Ornith-1.5-35B-A3B-CRACK-Q6_K.gguf}"
 : "${MMPROJ_FILE:=mmproj-Ornith-1.5-35B-A3B-f16.gguf}"
 : "${LLAMA_SERVER:=/home/dank/llama.cpp/build/bin/llama-server}"
 : "${CHAT_TEMPLATE_FILE:=$SCRIPT_DIR/chat-template.jinja}"

tests/frank_small_config.bats

@@ -0,0 +1,37 @@
+#!/usr/bin/env bats
+# Seam: frank-default-small model, context, tensor placement, and template.
+# This test covers only the Q6_K deployment contract.
+
+load helpers
+
+ENV_FILE="$REPO_ROOT/scripts/frank-default-small.env"
+SERVE_FILE="$REPO_ROOT/scripts/serve-frank-small.sh"
+UNIT_FILE="$REPO_ROOT/scripts/frank-default-small.service"
+TEMPLATE_FILE="$REPO_ROOT/scripts/chat-template.jinja"
+
+@test "frank-default-small selects the Ornith Q6_K model" {
+    grep -q '^MODEL_FILE=Ornith-1\.5-35B-A3B-CRACK-Q6_K\.gguf$' "$ENV_FILE"
+    grep -q 'Ornith-1\.5-35B-A3B-CRACK-Q6_K\.gguf' "$SERVE_FILE"
+}
+
+@test "frank-default-small uses 196608 context per slot across three slots" {
+    grep -q '^CTX_SIZE=589824$' "$ENV_FILE"
+    grep -q '196608 ctx each' "$ENV_FILE"
+    grep -q '^PARALLEL=3$' "$ENV_FILE"
+}
+
+@test "frank-default-small keeps final four expert blocks on CUDA" {
+    grep -q '^OVERRIDE_EXPERT=' "$ENV_FILE"
+    grep -Fq '35|36)\.ffn_' "$ENV_FILE"
+    grep -q '^OVERRIDE_BUFT=CPU$' "$ENV_FILE"
+    grep -q '^N_GPU_LAYERS=all$' "$ENV_FILE"
+}
+
+@test "frank-default-small uses the tracked Qwen-Sharp template" {
+    grep -q '^CHAT_TEMPLATE_FILE=' "$ENV_FILE"
+    grep -q 'peculiar-ragdoll/Qwen-Sharp-Chat-Templates' "$ENV_FILE"
+    grep -q 'chat-template\.jinja' "$SERVE_FILE"
+    [ -s "$TEMPLATE_FILE" ]
+    grep -q 'qwen3\.8-froggeric-v22\.4\.0' "$TEMPLATE_FILE"
+    grep -q 'Qwen-Sharp' "$UNIT_FILE"
+}