From 863897f92a1a759d77a3807a8cc601c19192c0d3 Mon Sep 17 00:00:00 2001
From: Frank <frank@dankware>
Date: Tue, 25 Aug 2026 19:07:46 -0700
Subject: [PATCH] Clarify llama-bench harness scope

---
 docs/.obsidian/app.json            |   2 +-
 docs/LLM/index.md                  |   2 +-
 docs/LLM/ubatch-benchmark-large.md |  47 +++---
 docs/index.md                      |   2 +-
 scripts/bench-fast-greedy.sh       |  77 ---------
 scripts/bench-kairic-ab.sh         | 153 ------------------
 scripts/bench-spec-stack.sh        | 191 ----------------------
 scripts/bench-ubatch-rocmi4.sh     | 245 -----------------------------
 scripts/bench-ubatch.sh            |  45 ++++++
 tests/frank_large_config.bats      |  32 ++--
 10 files changed, 78 insertions(+), 718 deletions(-)
 delete mode 100755 scripts/bench-fast-greedy.sh
 delete mode 100755 scripts/bench-kairic-ab.sh
 delete mode 100755 scripts/bench-spec-stack.sh
 delete mode 100755 scripts/bench-ubatch-rocmi4.sh
 create mode 100755 scripts/bench-ubatch.sh

diff --git a/docs/.obsidian/app.json b/docs/.obsidian/app.json
index c9ee2b7..baeee86 100644
--- a/docs/.obsidian/app.json
+++ b/docs/.obsidian/app.json
@@ -5,4 +5,4 @@
   "attachmentFolderPath": "_attachments",
   "useMarkdownLinks": false,
   "showUnsupportedFiles": true
-}
+}
\ No newline at end of file
diff --git a/docs/LLM/index.md b/docs/LLM/index.md
index b4ebf7d..836e177 100644
--- a/docs/LLM/index.md
+++ b/docs/LLM/index.md
@@ -50,7 +50,7 @@ MTP draft + Turbo4 V cache, 262K ctx / 1 slot]
 - [[frank-redo-2026-08-17]] — the 2026-08-17 topology change
 - [kingjones Qwen3.8 Uncensored ROCmFP4 STRIX MTP](https://huggingface.co/kingjones777/Qwen3.8-27B-Uncensored-ROCmFP4-STRIX-MTP-GGUF) — active frank-default-large model on Dankclaw; Q6 AGENT main model with standalone MTP
 - [[kingjones-strix-mtp]] — the active Q6 AGENT runbook: serving profile, standalone MTP, reasoning + preserve_thinking, 114 GiB TTM/GTT
-- [[ubatch-benchmark-large]] — active-model ubatch sweep plan and results
+- [[ubatch-benchmark-large]] — active-model llama-bench ubatch sweep plan and results
 - [[kingjones-q8agent-rocmfpx]] — the retired Q8_0-AGENT investigation and `set_rows` failure
 - [[eemin-rocmi4-mtp]] — the former caf-eemin Qwen3.8-27B embedded-MTP model
   on Dankclaw (retired)
diff --git a/docs/LLM/ubatch-benchmark-large.md b/docs/LLM/ubatch-benchmark-large.md
index 6fac208..68ed540 100644
--- a/docs/LLM/ubatch-benchmark-large.md
+++ b/docs/LLM/ubatch-benchmark-large.md
@@ -12,13 +12,13 @@ timestamp: 2026-09-01
 
 # frank-default-large ubatch sweep
 
-This benchmark sweeps the active `frank-default-large` model on Dankclaw using
-scratch ports and a 32K-context workload. The production service is stopped
-before the sweep and restarted by the benchmark's exit handler.
+This benchmark uses the active `frank-default-large` model on Dankclaw with
+`llama-bench`. It does not use `llama-server`, HTTP requests, or server-log
+scraping. Stop the production service before running it and start the service
+again afterward.
 
-The 2026-09-01 run used `ctx=32768` and `batch=32768`. It completed through
-`ubatch=3072`; `ubatch=4096` failed during initialization with a ROCm
-allocation error, so 4096, 5120, and 6144 were not run.
+The old custom harnesses are retained in `.AGENTTRASH/bench-scripts-20260901/`
+for auditability only. Their results are invalid and must not be used.
 
 ## Sweep
 
@@ -30,32 +30,25 @@ requests by default.
 Run on Dankclaw from the repository checkout:
 
 ```bash
-STOPSVC=1 ./scripts/bench-ubatch-rocmi4.sh
+sudo systemctl stop frank-default-large
+./scripts/bench-ubatch.sh
+sudo systemctl start frank-default-large
 ```
 
-The script uses the active Q6 AGENT model, standalone MTP draft, a fixed
-32K batch, and the tracked Sharp
-template, and `--reasoning-format deepseek`. It reports server timing lines and
-per-run generation throughput. Run only when the production service can be
-interrupted; the GPU is shared with live workloads.
+The script uses the active Q6 AGENT model, a fixed 16K prompt workload, and
+`llama-bench` JSONL output. The initial ubatch-only sweep is:
+```text
+256, 512, 768, 1024, 1280, 1536, 1792, 2048, 3072, 4096, 5120, 6144
+```
+
+The GPU is shared with live workloads; do not run with the production service
+up.
 
 ## Results
 
-Record results below after the sweep, including failures or OOM points.
-
-| ubatch | cold prefill | generation runs 1/2/3 | notes |
-| ---: | ---: | ---: | --- |
-| 256 | 214.73 tok/s | 17.68 / 17.15 / 17.12 tok/s | completed |
-| 512 | 213.47 tok/s | 17.91 / 17.29 / 17.72 tok/s | completed |
-| 768 | 215.60 tok/s | 17.96 / 18.17 / 17.98 tok/s | completed |
-| 1024 | 222.20 tok/s | 18.07 / 18.10 / 18.12 tok/s | completed |
-| 1280 | 219.13 tok/s | 17.47 / 17.69 / 17.49 tok/s | completed |
-| 1536 | 220.52 tok/s | 16.85 / 16.86 / 16.86 tok/s | completed |
-| 2048 | 60.97 tok/s | 6.24 / 4.94 / 11.19 tok/s | severe slowdown |
-| 3072 | 28.50 tok/s | 1.45 / 5.03 / 4.44 tok/s | severe slowdown |
-| 4096 | — | — | initialization OOM: rs-cache allocation |
-| 5120 | — | — | not reached |
-| 6144 | — | — | not reached |
+No valid results are recorded here. All results from the retired custom
+harnesses are explicitly discarded. Record the raw `llama-bench` JSONL output
+from the new harness here after a clean run.
 
 ## See also
 
diff --git a/docs/index.md b/docs/index.md
index 504cad6..0537829 100644
--- a/docs/index.md
+++ b/docs/index.md
@@ -119,7 +119,7 @@ The LLM inference fleet on the two hosts. Full index:
 ### Benchmarks and Sizing
 
 - [[frank-default-large-benchmark]] — backend/spec benchmark behind the old head
-- [[ubatch-benchmark-large]] — active-model ubatch sweep plan and results
+- [[ubatch-benchmark-large]] — active-model llama-bench ubatch sweep plan and results
 - [[frank-default-small-benchmark]] — the small-model benchmark suite
 - [[Running Qwen3.6 35B A3B FP8 on Dankserver]] — Qwen3.6-35B-A3B hardware notes
 - [[Running Qwen3.6 27B FP8 on Dankclaw]] — Qwen3.6-27B sizing on the iGPU
diff --git a/scripts/bench-fast-greedy.sh b/scripts/bench-fast-greedy.sh
deleted file mode 100755
index 19f4d5a..0000000
--- a/scripts/bench-fast-greedy.sh
+++ /dev/null
@@ -1,77 +0,0 @@
-#!/usr/bin/env bash
-# Bench the Kairic Edge "fast greedy (default)" profile.
-# Usage: ./bench-fast-greedy.sh
-#
-# Config: reasoning off, argmax fastpath=1, deterministic sampler.
-# Sends a coding prompt, captures full stream, computes decode tokens and time.
-set -euo pipefail
-
-HOST="127.0.0.1"
-PORT=7770
-ALIAS="frank-default-large"
-
-# A coding prompt representative of the card's 164-task suite.
-PROMPT='### Instruction
-Write a Python class that implements a priority queue using a binary heap. Include methods for enqueue, dequeue, peek, and is_empty. Add type hints, docstrings, and a usage example. The heap should support arbitrary comparable elements.
-### Response'
-
-DECODE_TOKENS=0
-START_SEC=0
-END_SEC=0
-
-echo "=== Fast-greedy benchmark (FASTGREEDY=1) ==="
-echo "Model: $ALIAS on $HOST:$PORT"
-echo "Prompt tokens: $(echo "$PROMPT" | wc -c) chars"
-echo ""
-
-# Use /v1/completions with streaming, capture tokens+time from usage.
-RESPONSE=$(curl -sf http://$HOST:$PORT/v1/completions \
-  -H 'Content-Type: application/json' \
-  -d "{
-    \"model\": \"$ALIAS\",
-    \"prompt\": $(python3 -c "import json,sys; print(json.dumps(sys.stdin.read()))" <<< "$PROMPT"),
-    \"max_tokens\": 512,
-    \"temperature\": 0,
-    \"top_p\": 1.0,
-    \"top_k\": 0,
-    \"min_p\": 0.0,
-    \"n\": 1,
-    \"stream\": true
-  }" 2>/dev/null)
-
-# Extract final usage from the last JSON line (finish_reason=stop or length).
-# Also count output text tokens.
-LAST_LINE=$(echo "$RESPONSE" | grep 'finish_reason' | tail -1)
-
-# Parse timing from the last chunk (has full usage)
-if echo "$LAST_LINE" | grep -q 'timings'; then
-    COMP_TOKENS=$(echo "$LAST_LINE" | python3 -c "
-import sys, json
-data = json.loads(sys.stdin.read())
-print(data.get('usage', {}).get('completion_tokens', 0))
-print(data.get('timings', {}).get('predicted_per_second', 0))
-print(data.get('timings', {}).get('prompt_per_second', 0))
-" 2>/dev/null)
-    echo "$COMP_TOKENS" | {
-        read CT
-        read TPS
-        read PPS
-        echo "Completion tokens: $CT"
-        echo "Predicted (decode) tokens/sec: $TPS"
-        echo "Prompt tokens/sec: $PPS"
-    }
-else
-    echo "WARNING: No timing data found in response"
-    echo "$LAST_LINE" | head -5
-fi
-
-# Count the actual text output
-OUTPUT_TEXT=$(echo "$RESPONSE" | grep '"text"' | head -1 | sed 's/.*"text":"//;s/","index.*//')
-OUTPUT_TOKENS=$(echo -n "$OUTPUT_TEXT" | wc -w)
-echo "Output text (rough token count): ~$OUTPUT_TOKENS words"
-echo ""
-
-# Check journal for draft acceptance and final timing
-echo "=== Draft acceptance & final timing ==="
-sleep 2
-ssh -p 2222 dank@dankclaw "journalctl -u frank-default-large --since '1 minute ago' --no-pager 2>/dev/null | grep -E 'draft acceptance|print_timing.*n_decoded|statistics.*draft-mtp' | tail -4" 2>/dev/null || echo "(could not read journal)"
diff --git a/scripts/bench-kairic-ab.sh b/scripts/bench-kairic-ab.sh
deleted file mode 100755
index fe1f4e0..0000000
--- a/scripts/bench-kairic-ab.sh
+++ /dev/null
@@ -1,153 +0,0 @@
-#!/usr/bin/env bash
-# bench-kairic-ab.sh — A/B perf isolation for the Kairic Edge deployment on
-# Dankclaw. Run this ON DANKCLAW (paths are absolute target paths).
-#
-# Configs:
-#   runner        pinned runner profile (b2048/ub512, reasoning off, target
-#                 greedy argmax fast path ON)
-#   nommap        runner + --no-mmap
-#   nofast        runner + fast path OFF
-#   prod          production profile (b4096/ub4096, --no-mmap, reasoning on/deepseek,
-#                 fast path OFF, preserve_thinking)
-#   prod2048      prod with the pinned runner's batch sizes (b2048/ub512)
-#   prodnommap    (obsolete: identical to prod since ubatch 4096 + --no-mmap
-#                 moved into the production profile)
-#
-# Usage: ./scripts/bench-kairic-ab.sh <cfg> [ctx] [prompt_chars]
-# Example: ./scripts/bench-kairic-ab.sh prod 16384 53000
-#
-# The server runs on scratch port 7893. STOP frank-default-large before
-# benchmarking (see AGENTS.md). The server is killed on exit.
-set -uo pipefail
-CFG="${1:?runner|nofast|prod}"
-CTX="${2:-32768}"
-PCHARS="${3:-11000}"
-PORT=7893
-BIN=/home/dank/ciru-qwen38/ROCmFPX-kairic/build-kairic/bin/llama-server
-MDIR=/home/dank/models/Kairic-Qwen3.8-27B-IU4-Edge
-SDK=/home/dank/ling3-ciru/Ling-3.0-Flash-CIRU-int4-Strix-native/.venv/lib/python3.12/site-packages
-LOG=/tmp/kairic-ab-$CFG.log
-: > "$LOG"
-
-export ROCM_PATH="$SDK/_rocm_sdk_devel"
-export LD_LIBRARY_PATH="$(dirname "$BIN"):$SDK/_rocm_sdk_libraries/lib:$SDK/_rocm_sdk_core/lib:$SDK/_rocm_sdk_devel/lib"
-export HIP_VISIBLE_DEVICES=0
-export HSA_OVERRIDE_GFX_VERSION=11.5.1
-export GGML_CUDA_GRAPH_OPT=0
-export LLAMA_MTP_CPU_ARGMAX_FASTPATH=1
-export PROMPTFORGE_TARGET_ONLY=0
-export PROMPTFORGE_MODE=iu4_ffn
-export PROMPTFORGE_IU4_SIDECAR="$MDIR/Qwen3.8-27B-Kairic-IU4-FFN.pfs"
-export PROMPTFORGE_SIDECAR="$MDIR/Qwen3.8-27B-Kairic-IU4-FFN.pfs"
-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="$MDIR/Qwen3.8-27B-Kairic-IU4-GDN.pfs"
-export PROMPTFORGE_GDN_SIDECAR="$MDIR/Qwen3.8-27B-Kairic-IU4-GDN.pfs"
-export PROMPTFORGE_GDN_IU4_HADAMARD=1
-export PROMPTFORGE_ENABLE_GDN_KEEPERS=0
-export PROMPTFORGE_ENABLE_GDN_OUTPUT=1
-export PROMPTFORGE_GDN_OUTPUT_SIDECAR="$MDIR/Qwen3.8-27B-Kairic-IU4-GDN-Output.pfs"
-export PROMPTFORGE_GDN_OUTPUT_KEEPERS=v3_lateq6
-export PROMPTFORGE_ENABLE_SMALLM_IU4=1
-export PROMPTFORGE_ENABLE_SMALLM_GDN_IU4=1
-export PROMPTFORGE_ENABLE_SMALLM_GDN_OUTPUT_IU4=0
-export PROMPTFORGE_ENABLE_IU4_DECODE_M2_M5=0
-export PROMPTFORGE_OUTPUT_K8_STRICT_GREEDY=0
-
-MMAP=""
-case "$CFG" in
-  runner)  export LLAMA_TARGET_GREEDY_ARGMAX_FASTPATH=1; B=2048; UB=512;  REA=off;  RFMT=none;     EXTRA="" ;;
-  nommap)  export LLAMA_TARGET_GREEDY_ARGMAX_FASTPATH=1; B=2048; UB=512;  REA=off;  RFMT=none;     EXTRA=""; MMAP="--no-mmap" ;;
-  nofast)  export LLAMA_TARGET_GREEDY_ARGMAX_FASTPATH=0; B=2048; UB=512;  REA=off;  RFMT=none;     EXTRA="" ;;
-  prod)    export LLAMA_TARGET_GREEDY_ARGMAX_FASTPATH=0; B=4096; UB=4096; REA=on;   RFMT=deepseek; EXTRA="--chat-template-kwargs {\"preserve_thinking\":\"true\"}"; MMAP="--no-mmap" ;;
-  prod2048)  export LLAMA_TARGET_GREEDY_ARGMAX_FASTPATH=0; B=2048; UB=512;  REA=on;   RFMT=deepseek; EXTRA="--chat-template-kwargs {\"preserve_thinking\":\"true\"}"; MMAP="--no-mmap" ;;
-  prodnommap) export LLAMA_TARGET_GREEDY_ARGMAX_FASTPATH=0; B=4096; UB=2048; REA=on; RFMT=deepseek; EXTRA="--chat-template-kwargs {\"preserve_thinking\":\"true\"}"; MMAP="--no-mmap" ;;
-esac
-
-"$BIN" -m "$MDIR/Qwen3.8-27B-IU4-Kairic-Edge.gguf" \
-  --alias bench --host 127.0.0.1 --port "$PORT" --jinja \
-  -dev ROCm0 -ngl 999 -c "$CTX" -b "$B" -ub "$UB" -fa on -ctk f16 -ctv f16 \
-  -t 16 -tb 32 -np 1 -ctxcp 32 --cache-ram 8192 --cache-prompt --cache-idle-slots $MMAP \
-  --metrics --kairic-edge \
-  --spec-type draft-mtp --spec-draft-device ROCm0 --spec-draft-ngl 999 \
-  --spec-draft-type-k f16 --spec-draft-type-v f16 \
-  --spec-draft-n-max 4 --spec-draft-n-min 0 --spec-draft-p-min 0.0 --spec-draft-p-split 0.10 \
-  --spec-draft-backend-sampling \
-  --temp 0 --top-p 1.0 --top-k 0 --min-p 0.0 \
-  --reasoning "$REA" --reasoning-format "$RFMT" --reasoning-budget -1 \
-  $EXTRA >> "$LOG" 2>&1 &
-SPID=$!
-trap 'kill $SPID 2>/dev/null; wait $SPID 2>/dev/null' EXIT
-
-for i in $(seq 1 300); do
-  kill -0 $SPID 2>/dev/null || { echo "SERVER DIED"; tail -20 "$LOG"; exit 2; }
-  curl -sf "http://127.0.0.1:$PORT/health" >/dev/null 2>&1 && break
-  sleep 2
-done
-echo "=== CFG=$CFG ready ==="
-
-python3 - "$PORT" "$CFG" "$PCHARS" <<'PY'
-import sys, json, time, urllib.request
-port, cfg, pchars = sys.argv[1], sys.argv[2], int(sys.argv[3])
-VAR = [
-    "narrative nonfiction blends observation with reflection, ",
-    "and a careful writer lets each clause earn its place ",
-    "in the sentence's rising line, so meaning accumulates slowly. ",
-    "The reader is not hurried; the reader is led. Loose and periodic, ",
-    "the style turns on concrete nouns and verbs that carry weight. ",
-    "A paragraph is not a list but a single turn of thought that ",
-    "opens, develops, and lands, and clarity is a courtesy owed ",
-    "to the willing reader who gives the writer their attention. ",
-]
-parts = []
-t = 0
-while len("".join(parts)) < pchars:
-    for k in range(3):
-        parts.append(VAR[(t + k) % len(VAR)])
-    parts.append(f" Note {t}. ")
-    t += 1
-longtext = "".join(parts)
-prompt = longtext + "\n\nNow write a long, detailed continuation of this essay (at least five paragraphs) about the philosophy of language.\n"
-body = {
-    "model": "bench",
-    "prompt": prompt,
-    "max_tokens": 512,
-    "temperature": 0,
-    "top_p": 1.0,
-    "top_k": 0,
-    "min_p": 0.0,
-    "stream": False,
-    "cache_prompt": True,
-}
-# ignore_eos is rejected by the exact-Q8 greedy argmax fast path (runner cfg).
-if cfg not in ("runner",):
-    body["ignore_eos"] = True
-def run(tag):
-    req = urllib.request.Request(f"http://127.0.0.1:{port}/v1/completions",
-                                 data=json.dumps(body).encode(),
-                                 headers={"Content-Type": "application/json"})
-    t0 = time.time()
-    try:
-        with urllib.request.urlopen(req, timeout=3600) as resp:
-            out = json.load(resp)
-    except Exception as e:
-        body_txt = getattr(e, "read", lambda: b"")()
-        print(f"[{tag}] REQUEST FAILED: {e}")
-        print(f"[{tag}] body: {body_txt.decode('utf-8', 'ignore')[:500]}")
-        raise SystemExit(3)
-    wall = time.time() - t0
-    u = out.get("usage", {})
-    tm = out.get("timings", {})
-    pt = u.get("prompt_tokens", 0); ct = u.get("completion_tokens", 0)
-    pps = tm.get("prompt_per_second", 0); tps = tm.get("predicted_per_second", 0)
-    print(f"[{tag}] cfg={cfg} prompt={pt} completion={ct} prefill={pps:.1f} t/s decode={tps:.2f} t/s wall={wall:.1f}s")
-run("cold")
-run("warm")
-PY
-
-grep -E "print_timing|statistics" "$LOG" | tail -6
-exit 0
diff --git a/scripts/bench-spec-stack.sh b/scripts/bench-spec-stack.sh
deleted file mode 100755
index 857a003..0000000
--- a/scripts/bench-spec-stack.sh
+++ /dev/null
@@ -1,191 +0,0 @@
-#!/usr/bin/env bash
-# bench-spec-stack.sh — compare MTP-only vs MTP+ngram-mod spec on the v3
-# CIRU prompt-forge stack (Dankclaw, gfx1151), at 8k / 16k / 64k context.
-#
-# Reads authoritative prefill/generation tok/s from llama-server's own
-# `print_timing` lines (prompt eval time / eval time). Runs one server
-# instance per (spec, ctx) on a scratch port.
-#
-# Usage:
-#   ./scripts/bench-spec-stack.sh <draft-mtp|draft-mtp,ngram-mod> <ctx> [out_tokens]
-# Example:
-#   ./scripts/bench-spec-stack.sh draft-mtp,ngram-mod 16384 256
-set -euo pipefail
-
-SPEC="${1:?usage: bench-spec-stack.sh <spec> <ctx> [out_tokens]}"
-CTX="${2:?usage: bench-spec-stack.sh <spec> <ctx> [out_tokens]}"
-OUT_TOKENS="${3:-256}"
-
-ENV_FILE=/home/dank/serverAppConfigs/scripts/frank-default-large.env
-source "$ENV_FILE"
-
-# This harness benchmarks the retired v3 stack. Force the v3 values so the
-# harness works regardless of the live service env (now Kairic).
-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
-
-# Derive the PromptForge sidecar paths from the env base values.
-export PROMPTFORGE_SIDECAR="$MODEL_DIR/$FFN_PFS"
-export PROMPTFORGE_GDN_SIDECAR="$MODEL_DIR/$GDN_PFS"
-export PROMPTFORGE_MTP_OUTPUT_K8_PROXY="$MODEL_DIR/$OUTPUT_K8_PFS"
-
-MODEL="$MODEL_DIR/$MODEL_FILE"
-LLAMA_BIN=/home/dank/ciru-qwen38/ROCmFPX/build-promptforge/bin/llama-server
-PORT=7893
-PARALLEL=1
-BATCH=4096
-UBATCH=4096
-LOG="/tmp/specstack-$(echo "$SPEC" | tr ',' '-')-ctx$CTX.log"
-: > "$LOG"
-
-for f in "$MODEL" "$PROMPTFORGE_SIDECAR" "$PROMPTFORGE_GDN_SIDECAR" "$PROMPTFORGE_MTP_OUTPUT_K8_PROXY" "$MODEL_DIR/$MMPROJ_FILE"; do
-    [[ -f "$f" ]] || { echo "missing: $f" >&2; exit 1; }
-done
-[[ -x "$LLAMA_BIN" ]] || { echo "missing: $LLAMA_BIN" >&2; exit 1; }
-
-export PROMPTFORGE_MODE="m2048_fused_tail1476"
-export GGML_CUDA_GRAPH_OPT=0
-export PROMPTFORGE_ENABLE_SMALLM_W8=1
-export PROMPTFORGE_MTP_OUTPUT_K8=1
-export PROMPTFORGE_MTP_OUTPUT_K8_VALIDATE=0
-export PROMPTFORGE_MTP_OUTPUT_K8_PROXY
-unset LLAMA_MTP_CPU_ARGMAX_FASTPATH LLAMA_MTP_BACKEND_GREEDY || true
-unset PROMPTFORGE_IU4_SIDECAR PROMPTFORGE_GDN_IU4_SIDECAR || true
-unset PROMPTFORGE_ENABLE_NGRAM_MOD PROMPTFORGE_ENABLE_NGRAM_M65_IU4 || true
-
-export ROCM_PATH="$ROCM_DEVEL"
-export LD_LIBRARY_PATH="${ROCM_LIB}/lib:${ROCM_CORE}/lib:${ROCM_DEVEL}/lib${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}"
-
-echo "=== spec-stack bench: spec=$SPEC ctx=$CTX model=$MODEL_FILE ==="
-echo "=== env: OUTPUT_K8=$PROMPTFORGE_MTP_OUTPUT_K8 smallm=$PROMPTFORGE_ENABLE_SMALLM_W8 ==="
-
-# ngram-mod tuning matches the retired fallback line: n_min 48 / n_max 64 / n_match 24
-"$LLAMA_BIN" \
-    --model "$MODEL" \
-    --mmproj "$MODEL_DIR/$MMPROJ_FILE" \
-    --alias bench \
-    --host 127.0.0.1 \
-    --port "$PORT" \
-    -dev ROCm0 \
-    -ngl 999 \
-    -c "$CTX" \
-    -b "$BATCH" \
-    -ub "$UBATCH" \
-    -fa on \
-    -ctk "$CACHE_TYPE_K" \
-    -ctv "$CACHE_TYPE_V" \
-    -np "$PARALLEL" \
-    -ctxcp "$CTX_CHECKPOINTS" \
-    --cache-ram "$CACHE_RAM" \
-    --cache-prompt \
-    --no-mmap \
-    --spec-type "$SPEC" \
-    --spec-draft-device ROCm0 \
-    --spec-draft-ngl 999 \
-    --spec-draft-type-k f16 \
-    --spec-draft-type-v f16 \
-    --spec-draft-n-max 4 \
-    --spec-draft-n-min 0 \
-    --spec-draft-p-min 0.0 \
-    --spec-draft-p-split 0.10 \
-    --spec-draft-backend-sampling \
-    --spec-ngram-mod-n-min 48 \
-    --spec-ngram-mod-n-max 64 \
-    --spec-ngram-mod-n-match 24 \
-    --temp 0.7 \
-    --top-p 0.95 \
-    -lv 4 \
-    >> "$LOG" 2>&1 &
-SERVER_PID=$!
-trap 'kill $SERVER_PID 2>/dev/null; wait $SERVER_PID 2>/dev/null' EXIT
-
-echo "server pid $SERVER_PID on 127.0.0.1:$PORT"
-echo -n "waiting for server"
-ready=0
-crashed=0
-for i in $(seq 1 900); do
-    if ! kill -0 "$SERVER_PID" 2>/dev/null; then
-        crashed=1
-        break
-    fi
-    if curl -sf "http://127.0.0.1:$PORT/health" >/dev/null 2>&1; then
-        echo " [ready]"
-        ready=1
-        break
-    fi
-    [ $((i % 10)) = 0 ] && echo -n .
-    sleep 2
-done
-if [ "$crashed" = "1" ]; then
-    echo " BACKEND_CRASHED (server died before ready)"
-    tail -30 "$LOG"
-    exit 2
-fi
-[ "$ready" = "1" ] || { echo " FAIL (timeout)"; tail -25 "$LOG"; exit 1; }
-
-# Print the implementations that initialized (feasibility evidence)
-grep -E "adding speculative implementation|ngram_mod|draft-mtp" "$LOG" | head -6
-
-python3 - "$SPEC" "$CTX" "$PORT" "$OUT_TOKENS" "$LOG" <<'PY'
-import sys, time, json, urllib.request
-spec, ctx, port, out_tokens, log = sys.argv[1], int(sys.argv[2]), sys.argv[3], int(sys.argv[4]), sys.argv[5]
-
-# Flowing prose (varied) at ~75% of ctx so MTP acceptance is realistic.
-VAR = (
-    "narrative nonfiction blends observation with reflection, ",
-    "and a careful writer lets each clause earn its place ",
-    "in the sentence's rising line, so meaning accumulates slowly. ",
-    "The reader is not hurried; the reader is led. Loose and periodic, ",
-    "the style turns on concrete nouns and verbs that carry weight, ",
-    "and it resists the abstract noun that hides a vague claim. ",
-    "A paragraph is not a list. It is a single turn of thought that ",
-    "opens, develops, and lands. This is the craft of prose. Choose, ",
-    "arrange, and above all omit. Omission is the discipline that ",
-    "keeps the whole clear, and clarity is the courtesy we owe the ",
-    "willing reader."
-)
-n = len(VAR)
-parts = []
-t = 0
-while t * 76 < (ctx * 3) // 4:
-    for k in range(5):
-        parts.append(VAR[(t + k) % n])
-    parts.append(f" This is cued by note {t}. ")
-    t += 1
-prompt = "".join(parts)
-prompt += ("\n\nNow, in a detailed and flowing paragraph, continue this essay about "
-           "the philosophy of language and its influence on modern computation. ")
-
-body = json.dumps({
-    "prompt": prompt,
-    "max_tokens": out_tokens,
-    "temperature": 0.0,
-    "stream": False,
-}).encode()
-req = urllib.request.Request(f"http://127.0.0.1:{port}/v1/completions",
-                             data=body, headers={"Content-Type": "application/json"})
-t0 = time.time()
-with urllib.request.urlopen(req, timeout=7200) as resp:
-    out = json.load(resp)
-dt = time.time() - t0
-pt = out.get("usage", {}).get("prompt_tokens", 0)
-ct = out.get("usage", {}).get("completion_tokens", 0)
-print(f"request done: prompt_tokens={pt} completion_tokens={ct} wall={dt:.1f}s", flush=True)
-
-lines = open(log, encoding="utf-8", errors="ignore").read().splitlines()
-prompt_evals = [l for l in lines if "prompt eval time" in l]
-evals       = [l for l in lines if "eval time" in l and "prompt eval" not in l]
-
-def grab(lines, label):
-    if lines:
-        print(label + ": " + lines[-1].strip())
-
-print(f"== RESULT spec={spec} ctx={ctx} ==")
-grab(prompt_evals, "PREFILL")
-grab(evals, "GENERATION")
-print(f"prompt_tokens={pt} completion_tokens={ct} wall={dt:.1f}s")
-PY
\ No newline at end of file
diff --git a/scripts/bench-ubatch-rocmi4.sh b/scripts/bench-ubatch-rocmi4.sh
deleted file mode 100755
index 997efcd..0000000
--- a/scripts/bench-ubatch-rocmi4.sh
+++ /dev/null
@@ -1,245 +0,0 @@
-#!/usr/bin/env bash
-# bench-ubatch-rocmi4.sh — sweep batch size with a fixed ubatch at 16k context
-# on the frank-default-large ROCmFPX standalone-MTP stack (dankclaw, gfx1151).
-#
-# Boots one llama-server instance per batch on a scratch port with a fixed
-# ubatch and -c 16384, runs a standardized ~75%-of-ctx prose workload, and reads
-# authoritative prefill/generation tok/s from the server's own print_timing
-# lines. Steps from UBATCH_MIN to UBATCH_MAX inclusive by UBATCH_STEP.
-#
-# Usage:
-#   ./scripts/bench-ubatch-rocmi4.sh
-#   UBATCH_MIN=512 UBATCH_MAX=4096 UBATCH_STEP=512 CTX=8192 RUNS=3 \
-#     ./scripts/bench-ubatch-rocmi4.sh
-#
-# Benchmarking policy: stop the production service first so the GPU is not
-# starved. With STOPSVC=1 the script restarts frank-default-large on exit.
-set -euo pipefail
-
-SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-ENV_FILE="${FRANK_LARGE_ENV_FILE:-$SCRIPT_DIR/frank-default-large.env}"
-if [[ -f "$ENV_FILE" ]]; then
-    set -a
-    # shellcheck disable=SC1090
-    source "$ENV_FILE"
-    set +a
-fi
-
-# --- sweep parameters -------------------------------------------------------
-: "${UBATCH_FIXED:=512}"
-: "${BATCH_VALUES:=512 1024 1536 2048}"
-: "${CTX:=16384}"                # 16k context
-: "${OUT_TOKENS:=256}"
-: "${RUNS:=3}"                   # decode runs per ubatch
-: "${PORT_BASE:=7894}"           # scratch ports 7894..
-: "${STOPSVC:=1}"                # stop/restart the live unit for a clean GPU
-
-# --- resolved paths (fall back to serve-script defaults) --------------------
-: "${LLAMA_BIN:=/home/dank/ciru-qwen38/ROCmFPX/build-promptforge/bin/llama-server}"
-: "${MODEL_DIR:=/home/dank/models/kingjones-Qwen3.8-27B-STRIX-MTP}"
-: "${MODEL_FILE:=Qwen3.8-27B-Uncensored-Q6_0_ROCMFPX_AGENT.gguf}"
-: "${MTP_DRAFT_FILE:=$MODEL_DIR/mtp-Qwen3.8-27B-Q4_0.gguf}"
-: "${MMPROJ_FILE:=mmproj-Qwen3.8-27B-Q8_0.gguf}"
-: "${CHAT_TEMPLATE_FILE:=$SCRIPT_DIR/chat-template.jinja}"
-: "${ALIAS:=frank-default-large}"
-: "${HSA_OVERRIDE_GFX_VERSION:=11.5.1}"
-: "${GGML_HIP_ENABLE_UNIFIED_MEMORY:=1}"
-: "${MTP_N_MAX:=16}"
-: "${MTP_P_MIN:=0.60}"
-
-MODEL="$MODEL_DIR/$MODEL_FILE"
-MMPROJ="$MODEL_DIR/$MMPROJ_FILE"
-
-for f in "$MODEL" "$CHAT_TEMPLATE_FILE"; do
-    if [[ ! -f "$f" ]]; then
-        echo "missing required file: $f" >&2
-        exit 1
-    fi
-done
-if [[ ! -x "$LLAMA_BIN" ]]; then
-    echo "missing llama-server binary: $LLAMA_BIN" >&2
-    exit 1
-fi
-
-MMPROJ_ARGS=()
-if [[ -f "$MMPROJ" ]]; then
-    MMPROJ_ARGS=(--mmproj "$MMPROJ")
-fi
-
-STopped=0
-SCRATCH_PID=""
-cleanup() {
-    if [[ -n "$SCRATCH_PID" ]] && kill -0 "$SCRATCH_PID" 2>/dev/null; then
-        echo "=== stopping scratch llama-server (pid $SCRATCH_PID) ==="
-        kill "$SCRATCH_PID" 2>/dev/null || true
-        wait "$SCRATCH_PID" 2>/dev/null || true
-        SCRATCH_PID=""
-    fi
-    if [ "$STopped" = 1 ]; then
-        echo "=== restarting frank-default-large ==="
-        sudo systemctl restart frank-default-large || true
-        STopped=0
-    fi
-}
-trap cleanup EXIT
-
-stop_live_service() {
-    if [ "$STOPSVC" = 1 ]; then
-        echo "=== stopping frank-default-large for a clean GPU ==="
-        sudo systemctl stop frank-default-large || true
-        STopped=1
-    fi
-}
-
-# wait_health <pid> <port> — poll /health until ready or the server dies.
-wait_health() {
-    local pid="$1" port="$2" i ready=0 crashed=0
-    for i in $(seq 1 900); do
-        if ! kill -0 "$pid" 2>/dev/null; then crashed=1; break; fi
-        if curl -sf "http://127.0.0.1:${port}/health" >/dev/null 2>&1; then
-            ready=1; break
-        fi
-        [ $((i % 10)) = 0 ] && echo -n .
-        sleep 2
-    done
-    [ "$crashed" = 1 ] && { echo " BACKEND_CRASHED (before ready)"; return 2; }
-    [ "$ready" = 1 ] || { echo " FAIL (timeout)"; return 1; }
-    echo " [ready]"
-    return 0
-}
-
-# run_one <batch> <port> <out_tokens> <runs> — boot one server instance, run
-# the workload, parse print_timing, print a RESULT line.
-run_one() {
-    local batch="$1" port="$2" out_tokens="$3" runs="$4"
-    local ubatch="$UBATCH_FIXED"
-    local log="/tmp/ubatch-${ubatch}-batch-${batch}-p${port}.log"
-    : > "$log"
-
-    echo "=== ubatch=${ubatch} ctx=${CTX} batch=${batch} port=${port} runs=${runs} ==="
-    "$LLAMA_BIN" \
-        --model "$MODEL" \
-        "${MMPROJ_ARGS[@]}" \
-        --chat-template-file "$CHAT_TEMPLATE_FILE" \
-        --chat-template-kwargs '{"enable_thinking":true,"preserve_thinking":true,"preserve_reasoning":true}' \
-        --alias bench-ubatch \
-        --host 127.0.0.1 --port "$port" \
-        -dev ROCm0 -ngl 999 -np 1 \
-        -t 16 -tb 32 \
-        -c "$CTX" -b "$batch" -ub "$ubatch" \
-        -fa on -ctk f16 -ctv f16 -ctxcp 16 \
-        --cache-ram 8192 --cache-prompt --metrics --no-mmap \
-        --spec-type draft-mtp --model-draft "$MTP_DRAFT_FILE" \
-        --spec-draft-device ROCm0 --spec-draft-ngl 999 \
-        --spec-draft-type-k f16 --spec-draft-type-v f16 \
-        --spec-draft-n-max "$MTP_N_MAX" --spec-draft-n-min 0 \
-        --spec-draft-p-min "$MTP_P_MIN" \
-        --spec-draft-backend-sampling \
-        --temp 0.7 --top-p 0.95 \
-        --reasoning "$REASONING" --reasoning-format "$REASONING_FORMAT" --reasoning-budget -1 \
-        -lv 4 >>"$log" 2>&1 &
-    local pid=$!
-    SCRATCH_PID="$pid"
-    if ! wait_health "$pid" "$port"; then
-        tail -30 "$log"
-        kill "$pid" 2>/dev/null || true
-        wait "$pid" 2>/dev/null || true
-        SCRATCH_PID=""
-        return 2
-    fi
-
-    # one python pass per run; reuse the same boot across runs
-    local r
-    for r in $(seq 1 "$runs"); do
-        echo "--- run ${r}/${runs} ---"
-        if ! python3 - "$ubatch" "$port" "$out_tokens" "$CTX" "$log" "$r" <<'PY'
-import sys, time, json, urllib.request
-ubatch, port, out_tokens, ctx, log, run_number = sys.argv[1], sys.argv[2], int(sys.argv[3]), int(sys.argv[4]), sys.argv[5], sys.argv[6]
-
-# Flowing prose at ~75% of ctx so MTP acceptance is realistic.
-VAR = (
-    "narrative nonfiction blends observation with reflection, ",
-    "and a careful writer lets each clause earn its place ",
-    "in the sentence's rising line, so meaning accumulates slowly. ",
-    "The reader is not hurried; the reader is led. Loose and periodic, ",
-    "the style turns on concrete nouns and verbs that carry weight, ",
-    "and it resists the abstract noun that hides a vague claim. ",
-    "A paragraph is not a list. It is a single turn of thought that ",
-    "opens, develops, and lands. This is the craft of prose. Choose, ",
-    "arrange, and above all omit. Omission is the discipline that ",
-    "keeps the whole clear, and clarity is the courtesy we owe the ",
-    "willing reader."
-)
-n = len(VAR)
-parts = []
-t = 0
-while t * 76 < (ctx * 3) // 4:
-    for k in range(5):
-        parts.append(VAR[(t + k) % n])
-    parts.append(f" This is cued by note {t}. ")
-    t += 1
-prompt = "".join(parts)
-prompt += ("\n\nNow, in a detailed and flowing paragraph, continue this essay about "
-           "the philosophy of language and its influence on modern computation. ")
-
-body = json.dumps({
-    "prompt": prompt, "max_tokens": out_tokens,
-    "temperature": 0.0, "stream": False,
-}).encode()
-req = urllib.request.Request(f"http://127.0.0.1:{port}/v1/completions",
-                             data=body, headers={"Content-Type": "application/json"})
-t0 = time.time()
-try:
-    with urllib.request.urlopen(req, timeout=7200) as resp:
-        out = json.load(resp)
-except Exception:
-    print("request failed; preserving benchmark cleanup", file=sys.stderr)
-    raise
-dt = time.time() - t0
-pt = out.get("usage", {}).get("prompt_tokens", 0)
-ct = out.get("usage", {}).get("completion_tokens", 0)
-
-lines = open(log, encoding="utf-8", errors="ignore").read().splitlines()
-prefill = [l for l in lines if "prompt eval time" in l]
-evals = [l for l in lines if "eval time" in l and "prompt eval" not in l]
-
-def grab(arr):
-    return arr[-1].strip() if arr else "n/a"
-
-g_tps = ""
-if evals and ct:
-    # eval timing line: "eval time = .. ms / N tokens (... tokens per second)"
-    import re
-    match = re.search(r"eval time =\s*([0-9.]+) ms", evals[-1])
-    if match:
-        ms = float(match.group(1))
-        g_tps = f"{ct / (ms / 1000):.2f}"
-print(f"== RESULT ubatch={ubatch} run={run_number} "
-      f"PREFILL={grab(prefill)} GEN_TPS={g_tps} "
-      f"pt={pt} ct={ct} wall={dt:.1f}s")
-PY
-        then
-            kill "$pid" 2>/dev/null || true
-            wait "$pid" 2>/dev/null || true
-            SCRATCH_PID=""
-            return 2
-        fi
-    done
-    kill "$pid" 2>/dev/null || true
-    wait "$pid" 2>/dev/null || true
-    SCRATCH_PID=""
-}
-
-main() {
-    stop_live_service
-    local idx=0 ubatch port
-    printf "%-10s %-14s %-10s\n" "UBATCH" "PREFILL" "GEN_TPS"
-    for batch in $BATCH_VALUES; do
-        port=$((PORT_BASE + idx))
-        run_one "$batch" "$port" "$OUT_TOKENS" "$RUNS"
-        idx=$((idx + 1))
-    done
-    echo "=== sweep complete ==="
-}
-
-main "$@"
diff --git a/scripts/bench-ubatch.sh b/scripts/bench-ubatch.sh
new file mode 100755
index 0000000..991a02c
--- /dev/null
+++ b/scripts/bench-ubatch.sh
@@ -0,0 +1,45 @@
+#!/usr/bin/env bash
+# Benchmark ubatch with llama-bench for frank-default-large.
+#
+# This is intentionally a ubatch-only benchmark. It uses one llama-bench
+# process, a fixed 16K prompt workload, and a fixed batch size large enough to
+# hold that workload. No server process, HTTP requests, or log scraping are used.
+#
+# Run on Dankclaw after stopping frank-default-large:
+#   sudo systemctl stop frank-default-large
+#   ./scripts/bench-ubatch.sh
+#   sudo systemctl start frank-default-large
+set -euo pipefail
+
+: "${LLAMA_BENCH:=/home/dank/GitPackages/ROCmFPX/build-strix-rocmfp4/bin/llama-bench}"
+: "${MODEL:=/home/dank/models/kingjones-Qwen3.8-27B-STRIX-MTP/Qwen3.8-27B-Uncensored-Q6_0_ROCMFPX_AGENT.gguf}"
+: "${UBATCH_VALUES:=256,512,768,1024,1280,1536,1792,2048,3072,4096,5120,6144}"
+: "${CONTEXT:=16384}"
+: "${GEN_TOKENS:=128}"
+: "${REPETITIONS:=3}"
+: "${DEVICE:=ROCm0}"
+: "${OUTPUT:=jsonl}"
+
+if [[ ! -x "$LLAMA_BENCH" ]]; then
+    printf 'missing executable llama-bench: %s\n' "$LLAMA_BENCH" >&2
+    exit 1
+fi
+if [[ ! -r "$MODEL" ]]; then
+    printf 'missing or unreadable model: %s\n' "$MODEL" >&2
+    exit 1
+fi
+
+exec "$LLAMA_BENCH" \
+    --model "$MODEL" \
+    --n-prompt "$CONTEXT" \
+    --n-gen "$GEN_TOKENS" \
+    --batch-size "$CONTEXT" \
+    --ubatch-size "$UBATCH_VALUES" \
+    --device "$DEVICE" \
+    --n-gpu-layers 999 \
+    --cache-type-k f16 \
+    --cache-type-v f16 \
+    --flash-attn on \
+    --mmap 0 \
+    --repetitions "$REPETITIONS" \
+    --output "$OUTPUT"
diff --git a/tests/frank_large_config.bats b/tests/frank_large_config.bats
index 5154010..69fc2a0 100644
--- a/tests/frank_large_config.bats
+++ b/tests/frank_large_config.bats
@@ -63,28 +63,16 @@ TEMPLATE_FILE="$REPO_ROOT/scripts/chat-template.jinja"
     grep -q -- '-ub "$UBATCH_SIZE"' "$SERVE_FILE"
 }
 
-@test "large ubatch bench defaults to the active model and requested sweep" {
-    local bench="$REPO_ROOT/scripts/bench-ubatch-rocmi4.sh"
-    grep -q ': "${UBATCH_FIXED:=512}"' "$bench"
-    grep -q 'BATCH_VALUES.*512 1024 1536 2048' "$bench"
-    grep -q ': "${STOPSVC:=1}"' "$bench"
-    grep -q ': "${CTX:=16384}"' "$bench"
-    grep -q 'MODEL_FILE' "$bench"
-    grep -q 'REASONING_FORMAT' "$bench"
-}
-
-@test "large ubatch bench cleans up scratch server on request failure" {
-    local bench="$REPO_ROOT/scripts/bench-ubatch-rocmi4.sh"
-    grep -q 'SCRATCH_PID=' "$bench"
-    grep -q 'kill.*SCRATCH_PID' "$bench"
-    grep -qF 'eval time =\s' "$bench"
-}
-
-@test "large ubatch bench supports fixed ubatch batch sweep" {
-    local bench="$REPO_ROOT/scripts/bench-ubatch-rocmi4.sh"
-    grep -q ': "${UBATCH_FIXED:=512}"' "$bench"
-    grep -q 'BATCH_VALUES' "$bench"
-    grep -q -- '-b "$batch" -ub "$ubatch"' "$bench"
+@test "llama-bench ubatch harness uses the active model" {
+    local bench="$REPO_ROOT/scripts/bench-ubatch.sh"
+    [ -x "$bench" ]
+    grep -q 'llama-bench' "$bench"
+    grep -q 'Qwen3\.8-27B-Uncensored-Q6_0_ROCMFPX_AGENT\.gguf' "$bench"
+    grep -q -- '--n-prompt "$CONTEXT"' "$bench"
+    grep -q -- '--ubatch-size "$UBATCH_VALUES"' "$bench"
+    grep -q -- '--batch-size "$CONTEXT"' "$bench"
+    grep -q -- '--mmap 0' "$bench"
+    ! grep -qE '(^|[[:space:]])llama-server|(^|[[:space:]])curl|print_timing' "$bench"
 }
 
 @test "frank-default-large uses the tracked Qwen-Sharp template on :7770" {