dankclaw: Ling-3.0-flash Q6_K trial config (Vulkan+GTT+draft-mtp)

Trial replaces frank-default-large temporarily via frank-large-ling.service
(port 7770, same alias). Requires the aetherbird bailingmoe3-support llama.cpp
fork at /home/dank/llama.cpp-ling (stock llama.cpp cannot load bailingmoe3).
Memory: 8x16GB LPDDR5 = 128GB physical; Vulkan exposes ~111.2GiB heap (UMA
96GiB + GTT) vs ROCm's 96GiB. Q6_K = 98.26GiB; MLA/KDA KV is tiny so 256k
ctx x 2 slots fits with ~10GiB headroom. Verified: loads ~45s, concurrent
streams both complete (no DFlash-style kill), ~7.7 t/s gen.
Restore Qwen: systemctl stop frank-large-ling; systemctl start
frank-default-large.

scripts/frank-default-large.ling.env

@@ -0,0 +1,39 @@
+# Tunables for the Ling-3.0-flash TRAIL on dankclaw (frank-large-ling).
+# This REPLACES frank-default-large (Qwen3.6-27B) temporarily. To switch
+# back: systemctl stop frank-large-ling && systemctl start frank-default-large.
+# Runtime: llama.cpp FORK bailingmoe3-support; Vulkan + AMD GTT.
+#
+# Memory model (amdgpu Strix Halo gfx1151, 8x16GB LPDDR5 = 128 GiB physical):
+#   - firmware UMA carve-out gives the iGPU ~96 GiB VRAM (rocm-smi "VRAM total").
+#   - CPU system RAM is the remainder (~30 GiB; kernel MemTotal).
+#   - Vulkan (RADV) reports an ~111.2 GiB heap (113910 MiB) - larger than ROCm's
+#     96 GiB, because Vulkan can also reach system memory via GTT.
+#   - Q6_K is 98.26 GiB, leaving only ~13 GiB of heap for KV + graphs, so keep
+#     the context small (Ling uses MLA/KDA, so KV stays small). Tune CTX_SIZE.
+MODEL_DIR=/home/dank/models
+MODEL_FILE=Ling-3.0-flash-Q6_K.gguf
+ALIAS=frank-default-large   # keep the API alias so clients are unaffected
+HOST=0.0.0.0
+PORT=7770
+# Context: Q6_K uses ~98.7 GiB of the ~111.2 GiB Vulkan heap; Ling's MLA/KDA
+# KV is tiny (~2.5 GiB for 256k), so a large context fits. 262144 total = 2
+# slots x 131072, ~10 GiB heap headroom after load. Verified 2026-08-14.
+CTX_SIZE=262144
+PARALLEL=2
+# MTP is bundled in the Ling GGUF itself (block 42) - no separate draft file.
+SPEC_TYPE=draft-mtp
+SPEC_DRAFT_N_MAX=3
+# Vulkan exposes the larger heap (UMA + GTT) vs ROCm. Use Vulkan0.
+DEVICE=Vulkan0
+N_GPU_LAYERS=all
+FLASH_ATTN=auto
+# Batch kept small-leaning for the tight memory budget; tune with CTX_SIZE.
+BATCH_SIZE=512
+UBATCH_SIZE=512
+# Sampling defaults (model card: reasoning-capable; leave temp/top to default)
+TEMPERATURE=1.0
+TOP_K=20
+TOP_P=0.95
+MIN_P=0.0
+THREADS=12
+THREADS_HTTP=1

scripts/frank-large-ling.service

@@ -0,0 +1,17 @@
+[Unit]
+Description=llama.cpp Ling-3.0-flash Q6_K Vulkan+MTP (frank-large-ling TRIAL)
+After=network.target
+
+[Service]
+Type=simple
+User=dank
+Group=dank
+Environment="HOST=0.0.0.0"
+Environment="PORT=7770"
+Environment="FRANK_LING_ENV_FILE=/home/dank/vllmScripts/scripts/frank-default-large.ling.env"
+ExecStart=/home/dank/vllmScripts/scripts/serve-ling-flash.sh
+Restart=on-failure
+RestartSec=10
+
+[Install]
+WantedBy=multi-user.target

scripts/serve-ling-flash.sh

@@ -0,0 +1,80 @@
+#!/usr/bin/env bash
+# Serve inclusionAI Ling-3.0-flash (Q6_K) on dankclaw as a TRIAL, replacing
+# frank-default-large temporarily.
+#
+# Runtime: llama.cpp FORK (aetherbird bailingmoe3-support) - stock llama.cpp
+# cannot load the bailingmoe3 architecture. Backend: Vulkan (RADV) on the
+# Strix Halo APU; memory is the shared 128 GB pool (96 GiB UMA + GTT that
+# Vulkan exposes as an ~111 GiB heap). MTP drafter is bundled in the same
+# GGUF (block 42) - no separate --spec-draft-model needed.
+#
+# Tie-back: to restore the Qwen3.6-27B service, stop frank-large-ling and
+# start frank-default-large (its env/unit are committed and unchanged).
+#
+# Tunables come from frank-default-large.ling.env in this directory unless
+# FRANK_LING_ENV_FILE overrides that path.
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+ENV_FILE="${FRANK_LING_ENV_FILE:-$SCRIPT_DIR/frank-default-large.ling.env}"
+if [[ -f "$ENV_FILE" ]]; then
+    set -a
+    # shellcheck disable=SC1090
+    source "$ENV_FILE"
+    set +a
+fi
+
+# --- defaults -------------------------------------------------------------
+: "${MODEL_DIR:=/home/dank/models}"
+: "${MODEL_FILE:=Ling-3.0-flash-Q6_K.gguf}"
+: "${ALIAS:=frank-default-large}"   # keep the same API alias for clients
+: "${HOST:=0.0.0.0}"
+: "${PORT:=7770}"
+: "${CTX_SIZE:=32768}"
+: "${PARALLEL:=2}"
+: "${SPEC_TYPE:=draft-mtp}"         # MTP block is bundled in the model GGUF
+: "${SPEC_DRAFT_N_MAX:=3}"
+: "${DEVICE:=Vulkan0}"
+: "${N_GPU_LAYERS:=all}"
+: "${BATCH_SIZE:=512}"
+: "${UBATCH_SIZE:=512}"
+: "${TEMPERATURE:=1.0}"
+: "${TOP_K:=20}"
+: "${TOP_P:=0.95}"
+: "${MIN_P:=0.0}"
+: "${THREADS:=12}"
+: "${THREADS_HTTP:=1}"
+
+# Fork binary; stock llama.cpp cannot load bailingmoe3.
+LLAMA_SERVER="${LLAMA_SERVER:-/home/dank/llama.cpp-ling/build/bin/llama-server}"
+MODEL="$MODEL_DIR/$MODEL_FILE"
+
+if [[ ! -x "$LLAMA_SERVER" ]]; then
+    echo "missing fork binary: $LLAMA_SERVER (build required)" >&2
+    exit 1
+fi
+if [[ ! -f "$MODEL" ]]; then
+    echo "missing model: $MODEL" >&2
+    exit 1
+fi
+
+exec "$LLAMA_SERVER" \
+    --model "$MODEL" \
+    --spec-type "$SPEC_TYPE" \
+    --spec-draft-n-max "$SPEC_DRAFT_N_MAX" \
+    --device "$DEVICE" \
+    --gpu-layers "$N_GPU_LAYERS" \
+    --flash-attn auto \
+    --ctx-size "$CTX_SIZE" \
+    --parallel "$PARALLEL" \
+    --batch-size "$BATCH_SIZE" \
+    --ubatch-size "$UBATCH_SIZE" \
+    --temp "$TEMPERATURE" \
+    --top-k "$TOP_K" \
+    --top-p "$TOP_P" \
+    --min-p "$MIN_P" \
+    --threads "$THREADS" \
+    --threads-http "$THREADS_HTTP" \
+    --alias "$ALIAS" \
+    --host "$HOST" \
+    --port "$PORT"