diff --git a/scripts/bench-batch-ubatch.sh b/scripts/bench-batch-ubatch.sh
new file mode 100755
index 0000000..47129da
--- /dev/null
+++ b/scripts/bench-batch-ubatch.sh
@@ -0,0 +1,170 @@
+#!/usr/bin/env bash
+#
+# bench-batch-ubatch.sh — how large may the logical batch and the micro-batch be?
+#
+# Sweeps llama.cpp's --batch-size against --ubatch-size with llama-bench and records, per
+# cell, the prefill rate, the decode rate and the PEAK VRAM of each card. The VRAM half is
+# not decoration: batch/ubatch is the knob that sizes the activation buffers, so a cell
+# that looks fast because it silently fell back to a smaller allocation would otherwise be
+# indistinguishable from a cell that is genuinely fast.
+#
+# ONE llama-bench INVOCATION PER CELL, on purpose. llama-bench calls exit(1) the moment a
+# test cannot run, so a single infeasible ubatch would take every later cell with it —
+# and "infeasible" is one of the things this sweep is looking for.
+#
+# RUN IT FROM A TRANSIENT UNIT, NOT FROM A SHELL. The deployment's unit sets
+# LimitMEMLOCK=infinity because the ROCm backend pins staging buffers; an SSH shell on
+# these hosts gets 8 MiB, and a load that fails on that proves nothing about the model:
+#
+#   sudo systemd-run --unit=fm-bench --collect --no-block --property=Type=oneshot \
+#        --property=LimitMEMLOCK=infinity --property=User=dank \
+#        /home/dank/serverAppConfigs/scripts/bench-batch-ubatch.sh
+#
+# It stops $SERVICE for the duration and starts it again on exit (trap), because a bench
+# beside a live service starves the same GPU. Results land in $OUT: one directory per cell
+# (bench.json, bench.log, vram.tsv) plus cells.tsv, appended as the sweep proceeds, so
+# progress can be read without waiting for the end. A re-run skips every cell that already
+# has a bench.json, so an interrupted sweep resumes.
+#
+set -uo pipefail
+
+BIN=${BIN:-/home/dank/GitPackages/llama.cpp-dflash2/build-rdna4-rocm714/bin}
+MODEL=${MODEL:-/home/dank/models/Swift-Qwen3.8-27B-Uncensored-Dynamic-MTP-GGUF/Swift-Qwen3.8-27B-Uncensored-Dynamic-MTP-UD-Q8_K_XL.gguf}
+OUT=${OUT:-/home/dank/models/logs/bench-bub-$(date +%Y%m%d-%H%M%S)}
+
+# what a cell measures: a prefill and a decode, both with a 64k KV already resident, so the
+# numbers describe the agent-loop shape (prompt landing in a long conversation) rather than
+# a cold start
+N_PROMPT=${N_PROMPT:-65536}
+N_GEN=${N_GEN:-128}
+N_DEPTH=${N_DEPTH:-65536}
+REPS=${REPS:-3}
+BATCHES=${BATCHES:-"2048 4096 6144 8192"}
+UBATCHES=${UBATCHES:-"512 768 1024 1280 1536 1792 2048"}
+SEED_CELLS=${SEED_CELLS:-"6144/1536"}   # the deployed cell first: it is the reference
+
+# how the model is loaded — must match the deployment, or the numbers describe a service
+# that does not exist
+NGL=${NGL:-999}
+DEVICES=${DEVICES:-ROCm0/ROCm1}        # llama-bench separates with '/'; ROCm2 is the iGPU
+SPLIT_MODE=${SPLIT_MODE:-layer}
+TENSOR_SPLIT=${TENSOR_SPLIT:-36/29}    # slash here — the llama-server unit takes commas
+FLASH_ATTN=${FLASH_ATTN:-on}
+KV_TYPE=${KV_TYPE:-f16}
+THREADS=${THREADS:-8}
+
+SERVICE=${SERVICE:-frank-medium}
+BOOTED=
+
+service_stop() {
+    if systemctl is-active --quiet "$SERVICE"; then
+        echo "stopping $SERVICE for the sweep" >&2
+        sudo systemctl stop "$SERVICE"
+        BOOTED=1
+    fi
+}
+
+service_start() {
+    if [ -n "$BOOTED" ]; then
+        echo "starting $SERVICE again" >&2
+        sudo systemctl start "$SERVICE"
+    fi
+}
+
+trap service_start EXIT
+
+run_cell() {
+    local b=$1 ub=$2 dir="$OUT/b${b}_ub${ub}" sampler start rc wall
+    mkdir -p "$dir"
+    if [ -s "$dir/bench.json" ]; then
+        echo "b=$b ub=$ub: bench.json present, skipping" >&2
+        return 0
+    fi
+    echo "b=$b ub=$ub: running" >&2
+
+    # 2 s sampling for the cell's whole life, load included
+    (
+        while :; do
+            rocm-smi --showmeminfo vram --csv 2>/dev/null |
+                awk -F, -v t="$(date +%s)" '/^card[01],/ { print t, $1, $2, $3 }'
+            sleep 2
+        done
+    ) > "$dir/vram.tsv" 2>/dev/null &
+    sampler=$!
+
+    start=$(date +%s)
+    "$BIN/llama-bench" -m "$MODEL" \
+        -p "$N_PROMPT" -n "$N_GEN" -d "$N_DEPTH" \
+        -b "$b" -ub "$ub" \
+        -ngl "$NGL" -dev "$DEVICES" -sm "$SPLIT_MODE" -ts "$TENSOR_SPLIT" \
+        -fa "$FLASH_ATTN" -ctk "$KV_TYPE" -ctv "$KV_TYPE" -t "$THREADS" \
+        -r "$REPS" --no-warmup --progress -o json \
+        > "$dir/bench.json" 2> "$dir/bench.log"
+    rc=$?
+
+    kill "$sampler" 2>/dev/null
+    wait "$sampler" 2>/dev/null
+    wall=$(( $(date +%s) - start ))
+
+    if [ "$rc" != 0 ]; then
+        grep -iE "error|failed|abort" "$dir/bench.log" | tail -3 >> "$OUT/failures.txt"
+        echo "b=$b ub=$ub: exit $rc" >&2
+    fi
+
+    python3 - "$dir/bench.json" "$dir/vram.tsv" "$b" "$ub" "$rc" "$wall" <<'PY'
+import json, sys
+
+json_path, vram_path, b, ub, rc, wall = sys.argv[1:7]
+
+def rates(tests, key):
+    hit = [t for t in tests if t.get(key)]
+    if not hit:
+        return "-\t-"
+    return "%.2f\t%.2f" % (hit[0]["avg_ts"], hit[0]["stddev_ts"])
+
+try:
+    tests = json.load(open(json_path))
+except Exception:
+    tests = []
+
+peak = {}
+for line in open(vram_path):
+    parts = line.split()
+    if len(parts) == 4:
+        _, card, total, used = parts
+        if card not in peak or int(used) > peak[card][1]:
+            peak[card] = (int(total), int(used))
+
+def free_gib(card):
+    if card not in peak:
+        return "-"
+    total, used = peak[card]
+    return "%.2f" % ((total - used) / 2**30)
+
+row = [b, ub, rc, wall, rates(tests, "n_prompt"), rates(tests, "n_gen"),
+       free_gib("card0"), free_gib("card1"), len(tests)]
+print("\t".join(str(f) for f in row))
+PY
+}
+
+mkdir -p "$OUT"
+echo "output directory: $OUT" >&2
+if [ ! -s "$OUT/cells.tsv" ]; then
+    printf 'batch\tubatch\trc\twall_s\tpp_tok_s\tpp_sd\ttg_tok_s\ttg_sd\tgpu0_free_GiB\tgpu1_free_GiB\ttests\n' \
+        > "$OUT/cells.tsv"
+fi
+
+service_stop
+
+for cell in $SEED_CELLS; do
+    run_cell "${cell%%/*}" "${cell##*/}" >> "$OUT/cells.tsv"
+done
+
+for b in $BATCHES; do
+    for ub in $UBATCHES; do
+        run_cell "$b" "$ub" >> "$OUT/cells.tsv"
+    done
+done
+
+echo "sweep complete: $OUT/cells.tsv" >&2
+cat "$OUT/cells.tsv" >&2