From c60156f3c6213df5a0c3a50f1fde781e4fdad5dd Mon Sep 17 00:00:00 2001
From: Frank <frank@dankware>
Date: Sat, 22 Aug 2026 14:20:11 -0700
Subject: [PATCH] rollback frank-default-large to v3 CIRU ActiveFPX PromptForge
 (operator decision 2026-08-22)

Restores the script pair verbatim from 132e147. v3 settings: PromptForge
m2048_fused_tail1476 full-W8 views, Output-K8 draft proxy, BF16 mmproj
vision, TEMPERATURE 0.7 / TOP_P 0.95, batch 4096/4096, ctxcp 16, reasoning
on/deepseek, no-mmap. Runtime: build-promptforge llama-server + v3 model
files in /home/dank/models/CIRU-Qwen3.8-27B-ActiveFPX-v3. Kairic Edge
files remain on the machine for rollback (docs/kairic-edge.md).
---
 scripts/frank-default-large.env | 121 +++++++++-----------
 scripts/serve-frank-large.sh    | 189 ++++++++++----------------------
 2 files changed, 110 insertions(+), 200 deletions(-)

diff --git a/scripts/frank-default-large.env b/scripts/frank-default-large.env
index 4686d10..ded4d07 100644
--- a/scripts/frank-default-large.env
+++ b/scripts/frank-default-large.env
@@ -1,39 +1,40 @@
-# Tunables for frank-default-large (Kairic Edge llama.cpp, dankclaw,
-# Strix Halo gfx1151).
+# Tunables for frank-default-large (CIRU llama.cpp, dankclaw, Strix Halo
+# gfx1151).
 #
-# Runtime: the llama-server built from the pinned ciru-ai/ROCmFPX tag
-# kairic-edge-qwen38-27b-v1 (Kairic Edge IU4 lane: PromptForge + Dual
-# View + native RDNA 3.5 IU4 matrix compute), built against the private
-# TheRock ROCm 7.15 SDK kept in the retired ling3-ciru venv.
-# Model: jcbtc/Qwen3.8-27B-IU4-Kairic-Edge, four-artifact payload
-# (GGUF + FFN/GDN/GDN-Output .pfs sidecars, 26.05 GiB total). Alias
-# frank-default-large, port 52625. See docs/kairic-edge.md.
+# Runtime: the PromptForge llama-server built from ciru-ai/ROCmFPX tag
+# qwen3.8-activefpx-promptforge-v2.3 plus the v3 output-K8 runtime patch
+# serving jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge (v3). v3 tensor
+# allocation (FFN-equalized, late-Q6 writers, full-Q8 output) with the
+# full-W8 PromptForge companions, the Output-K8 draft proxy, and the BF16
+# vision projector. Alias frank-default-large, port 52625.
 #
-# This runtime replaced the retired v3 CIRU ActiveFPX PromptForge
-# service. The v3 model files remain in
-# /home/dank/models/CIRU-Qwen3.8-27B-ActiveFPX-v3/ and the v3 binary in
-# /home/dank/ciru-qwen38/ROCmFPX/build-promptforge/ for rollback.
+# This runtime replaced the retired vLLM service that previously occupied
+# this slot. See docs/ciru-activefpx-promptforge.md.
 #
-# WARNING: do not name any variable TEMP, HOST, or GPU_MAX_HEAP_SIZE. The
-# ROCm HIP runtime reads those bare env vars (2026-08-13 TEMP incident).
+# 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).
 # llama.cpp reads --host/--port/--temp as CLI flags, never as env vars.
 
-# Kairic Edge llama-server binary (pinned tag build; see
-# docs/kairic-edge.md for the build runbook and binary identity).
-LLAMA_BIN=/home/dank/ciru-qwen38/ROCmFPX-kairic/build-kairic/bin/llama-server
+# CIRU llama-server binary (ROCmFPX v2.3 + v3 output-K8 patch build, see
+# docs/ciru runtime page). Built against the private TheRock ROCm 7.15 SDK
+# kept in the retired ling3-ciru venv. ROCm 7.15 is required for the
+# advertised Strix Halo performance; the system ROCm 7.2.4 selects
+# slower/fallback kernels.
+LLAMA_BIN=/home/dank/ciru-qwen38/ROCmFPX/build-promptforge/bin/llama-server
 
-# Model files (Kairic Edge four-artifact payload). The sidecars are
-# accelerator-specific prepacked copies of selected weights; they are
-# required, not optional.
-MODEL_DIR=/home/dank/models/Kairic-Qwen3.8-27B-IU4-Edge
-MODEL_FILE=Qwen3.8-27B-IU4-Kairic-Edge.gguf
-FFN_PFS=Qwen3.8-27B-Kairic-IU4-FFN.pfs
-GDN_PFS=Qwen3.8-27B-Kairic-IU4-GDN.pfs
-GDN_OUTPUT_PFS=Qwen3.8-27B-Kairic-IU4-GDN-Output.pfs
+# Model files (v3). The five-file payload (GGUF + FFN/GDN .pfs + Output-K8
+# proxy + BF16 vision projector) is the full v3 configuration. The v3 files
+# live in a separate directory so the v2 artifacts stay for rollback.
+MODEL_DIR=/home/dank/models/CIRU-Qwen3.8-27B-ActiveFPX-v3
+MODEL_FILE=Qwen3.8-27B-CIRU-ActiveFPX-PromptForge-v3-Q8-FFNEQ-LateQ6.gguf
+FFN_PFS=Qwen3.8-27B-CIRU-ActiveFPX-PromptForge-FFN.pfs
+GDN_PFS=Qwen3.8-27B-CIRU-ActiveFPX-PromptForge-GDN.pfs
+OUTPUT_K8_PFS=Qwen3.8-27B-CIRU-ActiveFPX-PromptForge-Output-K8.pfs
+MMPROJ_FILE=Qwen3.8-27B-mmproj-BF16.mmproj
 
-# ROCm 7.15 SDK paths (inside the retired ling3-ciru venv). The serve
-# script exports ROCM_PATH and LD_LIBRARY_PATH from these so the 7.15
-# libraries precede any system ROCm 7.2.4.
+# ROCm 7.15 SDK paths (inside the retired ling3-ciru venv). The serve script
+# exports ROCM_PATH and LD_LIBRARY_PATH from these so the 7.15 libraries
+# precede any system ROCm 7.2.4.
 ROCM_DEVEL=/home/dank/ling3-ciru/Ling-3.0-Flash-CIRU-int4-Strix-native/.venv/lib/python3.12/site-packages/_rocm_sdk_devel
 ROCM_LIB=/home/dank/ling3-ciru/Ling-3.0-Flash-CIRU-int4-Strix-native/.venv/lib/python3.12/site-packages/_rocm_sdk_libraries
 ROCM_CORE=/home/dank/ling3-ciru/Ling-3.0-Flash-CIRU-int4-Strix-native/.venv/lib/python3.12/site-packages/_rocm_sdk_core
@@ -42,54 +43,38 @@ ALIAS=frank-default-large
 BIND_ADDR=0.0.0.0
 LISTEN_PORT=52625
 
-# Serving: the qualified Kairic profile (pinned runner defaults). Full
-# native 262144 context, 1 slot, batch 2048 / ubatch 512, 16 target /
-# 32 batch threads, f16 target and draft KV, 32 context checkpoints,
-# 8192 MiB prompt cache with idle-slot persistence. Context checkpoints
-# are required for this hybrid recurrent model; do not remove CTX_CHECKPOINTS
-# while the prompt cache is enabled.
+# Serving: full native 262144 context (n_ctx_train), 1 slot (user request
+# 2026-08-19: parallel slots degraded performance). -c is the TOTAL
+# context; with PARALLEL=1 it equals the per-slot context. Concurrent
+# requests queue on the single slot; the server never preempts a running
+# generation. f16 KV, context checkpoints, prompt cache, native MTP
+# depth-4. The Strix Halo is unified memory; rocm-smi reports the 64 GiB
+# budget. KV at 262144 x 64 KiB/cell is ~16 GiB.
 CTX_SIZE=262144
 PARALLEL=1
-# Batch profile: the pinned runner's qualified sizes (2048/512). Benchmarks
-# on 2026-08-22: ubatch 512 gives the fastest prefill at every depth
-# (524/485/323 t/s at 8k/16k/64k vs 266/258/203 for ubatch 4096) with no
-# decode difference. Do not raise ubatch without re-benching prefill.
-BATCH_SIZE=2048
-UBATCH_SIZE=512
-TARGET_THREADS=16
-BATCH_THREADS=32
+BATCH_SIZE=4096
+UBATCH_SIZE=4096
 CACHE_TYPE_K=f16
 CACHE_TYPE_V=f16
-CTX_CHECKPOINTS=32
+CTX_CHECKPOINTS=16
 CACHE_RAM=8192
 
-# Native MTP draft (depth 4, p_min 0, backend sampling) — the pinned
-# runner's explicit MTP4.
+# Native MTP draft (depth-4, p_min 0, backend sampling). The qualified v3
+# profile drives the draft via the Output-K8 proxy; the CPU argmax fast
+# path and IU4/ngram lanes are disabled in the serve script.
 SPEC_DRAFT_N_MAX=4
 SPEC_DRAFT_N_MIN=0
 SPEC_DRAFT_P_MIN=0.0
 SPEC_DRAFT_P_SPLIT=0.10
 
-# Sampling: the pinned runner keeps the deterministic evaluation
-# profile (temp 0 / top-p 1 / top-k 0 / min-p 0) as the server default.
-# Clients override per request; the model card recommends Qwen live
-# values for ordinary chat (temp 0.7, top-p 0.8, top-k 20, min-p 0,
-# presence_penalty 1.5).
-#
-# Reasoning: the qualified profile runs reasoning off, but production
-# chat runs --reasoning on --reasoning-format deepseek (thinking goes to
-# message.reasoning_content, matching the retired v3 service) with
-# --reasoning-budget -1 (unrestricted) and chat_template_kwargs
-# preserve_thinking=true (preserve thinking tokens across turns).
-#
-# FASTGREEDY mode: set to 1 for fast-greedy benchmark (card's default profile)
-# FASTGREEDY=0: production chat config (reasoning on, argmax off)
-# FASTGREEDY=1: reasoning off, deterministic sampler, argmax fastpath off
-#   (compatible mode). Mirrors the card's 'compatible mode' profile — the
-#   same as production minus reasoning. Non-greedy requests work fine.
-# 2026-08-22: flipped to 1 for real (operator decision).
-FASTGREEDY=1
+# Sampling (model card defaults).
+TEMPERATURE=0.7
+TOP_P=0.95
 
-# For fast-greedy bench: set FASTGREEDY=1 above. This flips to: reasoning off,
-# argmax fastpath=1, sampling defaults. Deploy with
-# FASTGREEDY=1 ./deploy.sh dankclaw to test the fast-greedy profile.
+# Thinking enabled (user request, 2026-08-18). deepseek format moves the
+# <think> reasoning into message.reasoning_content (OpenAI-compatible);
+# none would leave the thought tags unparsed in content.
+REASONING=on
+REASONING_FORMAT=deepseek
+
+VERBOSITY=4
diff --git a/scripts/serve-frank-large.sh b/scripts/serve-frank-large.sh
index 576fad7..ba09b77 100755
--- a/scripts/serve-frank-large.sh
+++ b/scripts/serve-frank-large.sh
@@ -1,25 +1,19 @@
 #!/usr/bin/env bash
-# Serve jcbtc/Qwen3.8-27B-IU4-Kairic-Edge as frank-default-large via the
-# Kairic Edge CIRU ROCmFPX llama.cpp runtime on Dankclaw (Strix Halo
-# gfx1151).
+# Serve jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge (v3) as
+# frank-default-large via the CIRU ROCmFPX llama.cpp runtime on Dankclaw
+# (Strix Halo gfx1151).
 #
-# Runtime: the llama-server built from the pinned ciru-ai/ROCmFPX tag
-# kairic-edge-qwen38-27b-v1 (commit e97b324) against the private TheRock
-# ROCm 7.15 SDK kept in the retired ling3-ciru venv. The model is the
-# Kairic Edge GGUF plus the three native IU4 sidecars (FFN, GDN, GDN
-# Output); the runtime routes qualified prompt and multi-token
-# verification shapes through AMD RDNA 3.5 native IU4 matrix compute.
-# See docs/kairic-edge.md for the build runbook and binary identity.
+# Runtime: the PromptForge llama-server built from ciru-ai/ROCmFPX tag
+# qwen3.8-activefpx-promptforge-v2.3 (commit 103af8b41) plus the v3 output-K8
+# runtime patch (runtime/qwen38-v3-output-k8-runtime.patch, commit af0eac6 on
+# Dankclaw) against the private TheRock ROCm 7.15 SDK (kept in the retired
+# ling3-ciru venv). The model is the v3 GGUF with the full-W8 PromptForge
+# companion views (FFN/GDN .pfs), the Output-K8 draft proxy, and the BF16
+# vision projector. Native depth-4 MTP speculative decoding.
 #
-# This script replicates the release pinned runner
-# (ROCmFPX-kairic/scripts/run-kairic-edge-gfx1151.sh). The deliberate
-# differences: --host/--port come from BIND_ADDR/LISTEN_PORT (safe names;
-# the ROCm HIP runtime reads a bare HOST env var, see the 2026-08-13 TEMP
-# incident), LD_LIBRARY_PATH uses this machine's split 7.15 SDK layout,
-# the profile runs --no-mmap, reasoning on with deepseek format and
-# preserve_thinking, and the target greedy argmax fast path is disabled
-# for non-greedy chat sampling. Batch 2048 / ubatch 512 are the runner's
-# qualified sizes.
+# The model card requires ROCm 7.15 (not the system ROCm 7.2.4) for the
+# advertised Strix Halo performance. LD_LIBRARY_PATH therefore points at the
+# 7.15 SDK libraries first. An older system ROCm must not precede them.
 #
 # Tunables come from frank-default-large.env in this directory unless
 # FRANK_LARGE_ENV_FILE overrides that path.
@@ -35,40 +29,55 @@ if [[ -f "$ENV_FILE" ]]; then
 fi
 
 # --- defaults (used when the env file does not set them) -------------------
-# The defaults are the release pinned runner's qualified values.
-: "${LLAMA_BIN:=/home/dank/ciru-qwen38/ROCmFPX-kairic/build-kairic/bin/llama-server}"
-: "${MODEL_DIR:=/home/dank/models/Kairic-Qwen3.8-27B-IU4-Edge}"
-: "${MODEL_FILE:=Qwen3.8-27B-IU4-Kairic-Edge.gguf}"
-: "${FFN_PFS:=Qwen3.8-27B-Kairic-IU4-FFN.pfs}"
-: "${GDN_PFS:=Qwen3.8-27B-Kairic-IU4-GDN.pfs}"
-: "${GDN_OUTPUT_PFS:=Qwen3.8-27B-Kairic-IU4-GDN-Output.pfs}"
+: "${LLAMA_BIN:=/home/dank/ciru-qwen38/ROCmFPX/build-promptforge/bin/llama-server}"
+: "${MODEL_DIR:=/home/dank/models/CIRU-Qwen3.8-27B-ActiveFPX-v3}"
+: "${MODEL_FILE:=Qwen3.8-27B-CIRU-ActiveFPX-PromptForge-v3-Q8-FFNEQ-LateQ6.gguf}"
+: "${FFN_PFS:=Qwen3.8-27B-CIRU-ActiveFPX-PromptForge-FFN.pfs}"
+: "${GDN_PFS:=Qwen3.8-27B-CIRU-ActiveFPX-PromptForge-GDN.pfs}"
+: "${OUTPUT_K8_PFS:=Qwen3.8-27B-CIRU-ActiveFPX-PromptForge-Output-K8.pfs}"
+: "${MMPROJ_FILE:=Qwen3.8-27B-mmproj-BF16.mmproj}"
 : "${ROCM_DEVEL:=/home/dank/ling3-ciru/Ling-3.0-Flash-CIRU-int4-Strix-native/.venv/lib/python3.12/site-packages/_rocm_sdk_devel}"
 : "${ROCM_LIB:=/home/dank/ling3-ciru/Ling-3.0-Flash-CIRU-int4-Strix-native/.venv/lib/python3.12/site-packages/_rocm_sdk_libraries}"
 : "${ROCM_CORE:=/home/dank/ling3-ciru/Ling-3.0-Flash-CIRU-int4-Strix-native/.venv/lib/python3.12/site-packages/_rocm_sdk_core}"
 : "${ALIAS:=frank-default-large}"
 : "${BIND_ADDR:=0.0.0.0}"
 : "${LISTEN_PORT:=52625}"
-: "${CTX_SIZE:=262144}"
-: "${PARALLEL:=1}"
-: "${BATCH_SIZE:=2048}"
-: "${UBATCH_SIZE:=512}"
-: "${TARGET_THREADS:=16}"
-: "${BATCH_THREADS:=32}"
+: "${CTX_SIZE:=524288}"
+: "${PARALLEL:=2}"
+: "${BATCH_SIZE:=4096}"
+: "${UBATCH_SIZE:=4096}"
 : "${CACHE_TYPE_K:=f16}"
 : "${CACHE_TYPE_V:=f16}"
-: "${CTX_CHECKPOINTS:=32}"
+: "${CTX_CHECKPOINTS:=16}"
 : "${CACHE_RAM:=8192}"
 : "${SPEC_DRAFT_N_MAX:=4}"
 : "${SPEC_DRAFT_N_MIN:=0}"
 : "${SPEC_DRAFT_P_MIN:=0.0}"
 : "${SPEC_DRAFT_P_SPLIT:=0.10}"
+: "${TEMPERATURE:=0.7}"
+: "${TOP_P:=0.95}"
+: "${REASONING:=on}"
+: "${REASONING_FORMAT:=deepseek}"
+: "${VERBOSITY:=4}"
 
 MODEL="$MODEL_DIR/$MODEL_FILE"
-FFN="$MODEL_DIR/$FFN_PFS"
-GDN="$MODEL_DIR/$GDN_PFS"
-GDN_OUTPUT="$MODEL_DIR/$GDN_OUTPUT_PFS"
+export PROMPTFORGE_SIDECAR="$MODEL_DIR/$FFN_PFS"
+export PROMPTFORGE_GDN_SIDECAR="$MODEL_DIR/$GDN_PFS"
+# v3 qualified runtime: Output-K8 draft proxy + small-row W8 prompt route.
+export PROMPTFORGE_MODE="m2048_fused_tail1476"
+export GGML_CUDA_GRAPH_OPT=0
+export PROMPTFORGE_ENABLE_SMALLM_W8=1
+export PROMPTFORGE_MTP_OUTPUT_K8=1
+export PROMPTFORGE_MTP_OUTPUT_K8_VALIDATE=0
+export PROMPTFORGE_MTP_OUTPUT_K8_PROXY="$MODEL_DIR/$OUTPUT_K8_PFS"
+# The qualified v3 profile disables the CPU argmax draft fast path; the
+# Output-K8 proxy + backend sampling replace it. Also drop any IU4/ngram
+# companions from the non-IU4 development lane.
+unset LLAMA_MTP_CPU_ARGMAX_FASTPATH LLAMA_MTP_BACKEND_GREEDY || true
+unset PROMPTFORGE_IU4_SIDECAR PROMPTFORGE_GDN_IU4_SIDECAR || true
+unset PROMPTFORGE_ENABLE_NGRAM_MOD PROMPTFORGE_ENABLE_NGRAM_M65_IU4 || true
 
-for f in "$MODEL" "$FFN" "$GDN" "$GDN_OUTPUT"; do
+for f in "$MODEL" "$PROMPTFORGE_SIDECAR" "$PROMPTFORGE_GDN_SIDECAR" "$PROMPTFORGE_MTP_OUTPUT_K8_PROXY" "$MODEL_DIR/$MMPROJ_FILE"; do
     if [[ ! -f "$f" ]]; then
         echo "missing required file: $f" >&2
         exit 1
@@ -80,95 +89,15 @@ if [[ ! -x "$LLAMA_BIN" ]]; then
 fi
 
 # ROCm 7.15 SDK libraries must precede any system ROCm in LD_LIBRARY_PATH.
-# The release runner assumes one ROCm tree; this machine keeps the 7.15
-# SDK in three split directories, so use the proven split layout plus the
-# server's own directory.
 export ROCM_PATH="$ROCM_DEVEL"
-export LD_LIBRARY_PATH="$(dirname "$LLAMA_BIN"):${ROCM_LIB}/lib:${ROCM_CORE}/lib:${ROCM_DEVEL}/lib${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}"
+export LD_LIBRARY_PATH="${ROCM_LIB}/lib:${ROCM_CORE}/lib:${ROCM_DEVEL}/lib${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}"
 
-# Pinned runner environment, verbatim (Kairic Edge IU4 lane, PromptForge,
-# Dual View sidecars).
-export HIP_VISIBLE_DEVICES=0
-export HSA_OVERRIDE_GFX_VERSION=11.5.1
-export GGML_CUDA_GRAPH_OPT=0
-# The pinned runner sets the target exact-Q8 greedy argmax fast path
-# (LLAMA_TARGET_GREEDY_ARGMAX_FASTPATH=1). That fast path fails closed on
-# any non-greedy request (temperature != 0, top_k > 1, top_p < 1, penalties,
-# grammar, logit bias, or reasoning budget). The model card's own chat
-# instructions override sampling per request with Qwen live values (temp
-# 0.7, top-p 0.8, top-k 20, min-p 0, presence_penalty 1.5), which the fast
-# path rejects. Production chat therefore forces it to 0. This does NOT
-# affect the Kairic Edge IU4 lane (prompt/verification route via
-# --kairic-edge) nor the MTP draft fast path below.
-# FASTGREEDY env var (0 = production chat, 1 = fast-greedy bench from
-# card's default profile). When FASTGREEDY=1: reasoning off, deterministic
-# sampler. The argmax fastpath is always off (compatible mode) so non-greedy
-# requests (temp != 0, penalties, reasoning budget) work correctly.
-# The card calls this "compatible mode" — it disables only the target greedy
-# argmax fast path and keeps everything else (reasoning, MTP, caching).
-FASTGREEDY="${FASTGREEDY:-0}"
-export LLAMA_TARGET_GREEDY_ARGMAX_FASTPATH=0
-export LLAMA_MTP_CPU_ARGMAX_FASTPATH=1
-export PROMPTFORGE_TARGET_ONLY=0
-export PROMPTFORGE_MODE=iu4_ffn
-export PROMPTFORGE_IU4_SIDECAR="$FFN"
-export PROMPTFORGE_SIDECAR="$FFN"
-export PROMPTFORGE_IU4_HADAMARD=1
-export PROMPTFORGE_IU4_SEGMENTED=0
-export PROMPTFORGE_ENABLE_FFN_KEEPERS=0
-export PROMPTFORGE_FFN_KEEPERS=late6
-export PROMPTFORGE_ENABLE_GDN=1
-export PROMPTFORGE_ENABLE_GDN_W8=0
-export PROMPTFORGE_GDN_SIDECAR_OVERRIDE="$GDN"
-export PROMPTFORGE_GDN_SIDECAR="$GDN"
-export PROMPTFORGE_GDN_IU4_HADAMARD=1
-export PROMPTFORGE_ENABLE_GDN_KEEPERS=0
-export PROMPTFORGE_ENABLE_GDN_OUTPUT=1
-export PROMPTFORGE_GDN_OUTPUT_SIDECAR="$GDN_OUTPUT"
-export PROMPTFORGE_GDN_OUTPUT_KEEPERS=v3_lateq6
-export PROMPTFORGE_ENABLE_SMALLM_IU4=1
-export PROMPTFORGE_ENABLE_SMALLM_GDN_IU4=1
-export PROMPTFORGE_ENABLE_SMALLM_GDN_OUTPUT_IU4=0
-export PROMPTFORGE_ENABLE_IU4_DECODE_M2_M5=0
-export PROMPTFORGE_OUTPUT_K8_STRICT_GREEDY=0
-
-# Sampling & reasoning defaults — production (FASTGREEDY=0) vs fast-greedy
-# (FASTGREEDY=1). Both use argmax fastpath off (compatible mode); the only
-# difference is reasoning: off (FASTGREEDY=1) vs on/deepseek/budget-unlimited
-# (FASTGREEDY=0). Sampling defaults are identical (temp 0, top_p 1, top_k 0,
-# min_p 0) and clients override per request as the card recommends.
-if [ "$FASTGREEDY" = "1" ]; then
-    TEMP=0
-    TOP_P=1.0
-    TOP_K=0
-    MIN_P=0.0
-    REASONING_ARGS=""
-else
-    TEMP=0
-    TOP_P=1.0
-    TOP_K=0
-    MIN_P=0.0
-    REASONING_ARGS="--reasoning on --reasoning-format deepseek --reasoning-budget -1"
-fi
-
-# Pinned runner flags (qualified 262K one-slot profile, native MTP depth
-# 4, deterministic sampler; batch 2048 / ubatch 512 are the runner's
-# qualified sizes and give the fastest measured prefill). Local overrides
-# for production chat:
-#   - reasoning on + deepseek format (thinking goes to message.reasoning_content,
-#     matching the retired v3 service so the fleet's reasoning clients work)
-#   - --no-mmap (local choice; benchmarks show no perf difference versus
-#     mmap, kept per operator preference)
-#   - chat_template_kwargs preserve_thinking=true (preserve thinking tokens
-#     in the conversation history across turns)
-# Sampling is overridden per request by clients; the card recommends Qwen
-# live values (temp 0.7, top-p 0.8, top-k 20, min-p 0, presence_penalty 1.5).
 exec "$LLAMA_BIN" \
-    -m "$MODEL" \
+    --model "$MODEL" \
+    --mmproj "$MODEL_DIR/$MMPROJ_FILE" \
     --alias "$ALIAS" \
     --host "$BIND_ADDR" \
     --port "$LISTEN_PORT" \
-    --jinja \
     -dev ROCm0 \
     -ngl 999 \
     -c "$CTX_SIZE" \
@@ -177,29 +106,25 @@ exec "$LLAMA_BIN" \
     -fa on \
     -ctk "$CACHE_TYPE_K" \
     -ctv "$CACHE_TYPE_V" \
-    -t "$TARGET_THREADS" \
-    -tb "$BATCH_THREADS" \
     -np "$PARALLEL" \
     -ctxcp "$CTX_CHECKPOINTS" \
     --cache-ram "$CACHE_RAM" \
     --cache-prompt \
-    --cache-idle-slots \
-    --no-mmap \
     --metrics \
-    --kairic-edge \
+    --no-mmap \
     --spec-type draft-mtp \
     --spec-draft-device ROCm0 \
     --spec-draft-ngl 999 \
-    --spec-draft-type-k "$CACHE_TYPE_K" \
-    --spec-draft-type-v "$CACHE_TYPE_V" \
+    --spec-draft-type-k f16 \
+    --spec-draft-type-v f16 \
     --spec-draft-n-max "$SPEC_DRAFT_N_MAX" \
     --spec-draft-n-min "$SPEC_DRAFT_N_MIN" \
     --spec-draft-p-min "$SPEC_DRAFT_P_MIN" \
     --spec-draft-p-split "$SPEC_DRAFT_P_SPLIT" \
     --spec-draft-backend-sampling \
-    --temp "$TEMP" \
+    --temp "$TEMPERATURE" \
     --top-p "$TOP_P" \
-    --top-k "$TOP_K" \
-    --min-p "$MIN_P" \
-    $REASONING_ARGS \
-    --chat-template-kwargs '{"preserve_thinking":"true"}'
+    --reasoning "$REASONING" \
+    --reasoning-format "$REASONING_FORMAT" \
+    --reasoning-budget -1 \
+    -lv "$VERBOSITY"