bench: the preflight observes the load mode it asked for, not just prints it (engine #1668)

tools/bench-preflight.sh

@@ -167,7 +167,36 @@ case "$lm" in
     *) fatal "PREFLIGHT_LM='$lm' is not a llama-bench load mode (none|mmap|mlock|mmap+mlock|dio); the mode is part of the measurement, so it is not guessed (engine #1668)" ;;
 esac
 GGML_HRX_LOG_DISPATCH=1 "$bench" -m "$model" -p "$tokens" -n 0 -ngl 99 --device HRX0 -r 3 -lm "$lm" \
-    >/dev/null 2>"$log"
+    >/dev/null 2>"$log" &
+bench_pid=$!
+# The mode is REQUESTED above and OBSERVED here (engine #1668): llama.cpp puts a model it mmaps in `/proc/<pid>/maps` and a
+# model it only reads is not there, so the run can be checked against the flag instead of trusted to honour it. A bench that
+# ignores `-lm` yields a number the header mislabels, which is the same defect as an unstated mode -- one variable measured,
+# two reported.
+mapped=no; alive=0
+while :; do
+    grep -qF -- "$(basename "$model")" "/proc/$bench_pid/maps" 2>/dev/null && { mapped=yes; break; }
+    kill -0 "$bench_pid" 2>/dev/null || break
+    alive=$((alive + 1))
+    sleep 0.1
+done
+bench_rc=0
+wait "$bench_pid" || bench_rc=$?
+[ "$bench_rc" -eq 0 ] || fatal "the bench exited $bench_rc; see $log"
+# Only the modes whose mapping is unambiguous are asserted: `mmap`/`mmap+mlock` map, `none`/`dio` do not, and `mlock` alone is
+# not asserted because whether it maps depends on the loader -- a gate that guessed would fail a correct run. An OBSERVED
+# mapping is trusted on its own; an ABSENT one is only trusted when the process was seen alive after the sample, since a
+# bench that maps and exits between polls would otherwise be reported as a mismatch.
+if [ "$mapped" = no ] && [ "$alive" -eq 0 ]; then
+    printf '  WARN the bench finished before the load mode could be observed; -lm %s not checked\n' "$lm"
+else
+    case "$lm:$mapped" in
+        mmap:no|mmap+mlock:no)
+            fatal "the load mode is -lm $lm but the model is not in /proc/$bench_pid/maps; the run is not the measurement the header claims (engine #1668)" ;;
+        none:yes|dio:yes)
+            fatal "the load mode is -lm $lm but the model IS in /proc/$bench_pid/maps; the run is not the measurement the header claims (engine #1668)" ;;
+    esac
+fi
 
 lines=$(grep -c 'ggml_hrx dispatch:' "$log" || true)
 # A silent bench is the failure this flag exists to catch: an empty tally looks like "nothing declined"

tools/tests/test-bench-preflight.sh

@@ -167,5 +167,55 @@ else
     printf 'FAIL %-38s header does not name the load mode\n%s\n' "tally header states -lm" "$(cat "$tmp/lastout")"; fail=$((fail+1))
 fi
 
+# --dispatch must also OBSERVE the load mode, not only request it (engine #1668): llama.cpp puts a model it mmaps in
+# `/proc/<pid>/maps` and a model it only reads is not there. The bench below really maps or really reads the model and then
+# sleeps, so the observation has something to see; FAKE_BENCH_LM lets a case make the bench ignore `-lm`, which is the defect
+# the check exists for (a number the header mislabels).
+head -c 65536 /dev/urandom > "$tmp/models/big.gguf"      # non-empty: mmap of a zero-length file fails
+cat > "$tmp/mapbench" <<'PY'
+#!/usr/bin/env python3
+import mmap, os, sys, time
+a = sys.argv[1:]
+lm = a[a.index('-lm') + 1] if '-lm' in a else 'mmap'
+model = a[a.index('-m') + 1]
+mode = os.environ.get('FAKE_BENCH_LM') or lm          # a case sets this to make the bench disagree with -lm
+handle = keep = None
+if mode in ('mmap', 'mmap+mlock'):
+    handle = open(model, 'rb')
+    keep = mmap.mmap(handle.fileno(), 0, access=mmap.ACCESS_READ)
+else:
+    open(model, 'rb').read(4096)
+sys.stderr.write('ggml_hrx dispatch: common.mul_mat.q4_k_q8_1_x4_prefill\n')
+sys.stderr.write('| qwen3-4b | 2.30 GiB | 4B | HRX | 99 | 1 | pp512 | 3181.97 t/s |\n')
+time.sleep(0.7)
+assert keep is None or len(keep) == 65536          # the mapping must outlive the sleep
+PY
+chmod +x "$tmp/mapbench"
+
+lm_case() {  # lm_case <want: ok|no> <name> <PREFLIGHT_LM> [FAKE_BENCH_LM]
+    local want=$1 name=$2 lm=$3 fake_lm=${4:-} root out rc got
+    root=$(fake 3600 0 60)
+    out=$(PREFLIGHT_LM=$lm FAKE_BENCH_LM=$fake_lm PREFLIGHT_MODEL=$tmp/models/big.gguf \
+          patched "$root" --dispatch "$tmp/mapbench" 512 2>&1); rc=$?
+    got=ok; [ $rc -ne 0 ] && got=no
+    if [ "$got" = "$want" ]; then
+        printf 'ok   %-38s rc=%d\n' "$name" "$rc"; pass=$((pass+1))
+    else
+        printf 'FAIL %-38s rc=%d want=%s\n%s\n' "$name" "$rc" "$want" "$out"; fail=$((fail+1))
+    fi
+    printf '%s' "$out" > "$tmp/lastlm"
+}
+
+lm_case ok "observed mmap agrees with -lm mmap"        mmap
+lm_case ok "observed read agrees with -lm none"        none
+lm_case no "-lm none that actually mmapped is refused"  none mmap
+lm_case no "-lm mmap that only read is refused"         mmap none
+# The refusal has to say what it saw, or the reader cannot tell a mapping mismatch from a bench that never started.
+if grep -q 'is not the measurement the header claims' "$tmp/lastlm"; then
+    printf 'ok   %-38s names the mismatch\n' "the refusal names its reason"; pass=$((pass+1))
+else
+    printf 'FAIL %-38s missing reason\n%s\n' "the refusal names its reason" "$(cat "$tmp/lastlm")"; fail=$((fail+1))
+fi
+
 printf '\nbench-preflight test: %d passed, %d failed\n' "$pass" "$fail"
 [ "$fail" -eq 0 ]