Set ubatch sweep context and batch to 16K

docs/LLM/ubatch-benchmark-large.md

@@ -13,7 +13,7 @@ timestamp: 2026-09-01
 # frank-default-large ubatch sweep
 
 This benchmark sweeps the active `frank-default-large` model on Dankclaw using
-scratch ports and an 8K-context workload. The production service is stopped
+scratch ports and a 16K-context workload. The production service is stopped
 before the sweep and restarted by the benchmark's exit handler.
 
 ## Sweep
@@ -29,7 +29,8 @@ Run on Dankclaw from the repository checkout:
 STOPSVC=1 ./scripts/bench-ubatch-rocmi4.sh
 ```
 
-The script uses the active Q6 AGENT model, standalone MTP draft, tracked Sharp
+The script uses the active Q6 AGENT model, standalone MTP draft, a fixed
+16K 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.

scripts/bench-ubatch-rocmi4.sh

@@ -3,7 +3,7 @@
 # frank-default-large ROCmFPX standalone-MTP stack (dankclaw, gfx1151).
 #
 # Boots one llama-server instance per ubatch on a scratch port with a fixed
-# -b and -c 8192, runs a standardized ~75%-of-ctx prose workload, and reads
+# -b 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.
 #
@@ -30,8 +30,8 @@ fi
 : "${UBATCH_MAX:=6144}"
 : "${UBATCH_STEP:=256}"
 : "${UBATCH_VALUES:=256 512 768 1024 1280 1536 2048 3072 4096 5120 6144}"
-: "${CTX:=8192}"                 # 8k contexts
-: "${BENCH_BATCH:=$CTX}"         # fixed batch = ctx, so the full 8k prompt fits
+: "${CTX:=16384}"                # 16k contexts
+: "${BENCH_BATCH:=16384}"         # batch is fixed to the 16k context size
 : "${OUT_TOKENS:=256}"
 : "${RUNS:=3}"                   # decode runs per ubatch
 : "${PORT_BASE:=7894}"           # scratch ports 7894..
@@ -118,7 +118,7 @@ run_one() {
     local log="/tmp/ubatch-${ubatch}-p${port}.log"
     : > "$log"
 
-    echo "=== ubatch=${ubatch} ctx=${CTX} port=${port} runs=${runs} ==="
+    echo "=== ubatch=${ubatch} ctx=${CTX} batch=${BENCH_BATCH} port=${port} runs=${runs} ==="
     "$LLAMA_BIN" \
         --model "$MODEL" \
         "${MMPROJ_ARGS[@]}" \
@@ -142,13 +142,19 @@ run_one() {
         -lv 4 >>"$log" 2>&1 &
     local pid=$!
     SCRATCH_PID="$pid"
-    wait_health "$pid" "$port" || { tail -30 "$log"; return 2; }
+    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} ---"
-        python3 - "$ubatch" "$port" "$out_tokens" "$CTX" "$log" "$r" <<'PY'
+        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]
 
@@ -214,6 +220,12 @@ 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

tests/frank_large_config.bats

@@ -69,6 +69,8 @@ TEMPLATE_FILE="$REPO_ROOT/scripts/chat-template.jinja"
     grep -q ': "${UBATCH_MAX:=6144}"' "$bench"
     grep -q 'UBATCH_VALUES.*256 512 768 1024 1280 1536 2048 3072 4096 5120 6144' "$bench"
     grep -q ': "${STOPSVC:=1}"' "$bench"
+    grep -q ': "${CTX:=16384}"' "$bench"
+    grep -q ': "${BENCH_BATCH:=16384}"' "$bench"
     grep -q 'MODEL_FILE' "$bench"
     grep -q 'REASONING_FORMAT' "$bench"
 }