File size: 7,888 Bytes
510314e | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 | #!/bin/bash
# Pilota GPU della correzione 2 (card t_caa8225d, BRIEF-PILOT-C2-GPU): UN blocco GPU su PC4070TI <= MAXSEC (4 h).
# Adattato da agentlongbench/scripts/gpu_block_alb_seq.sh a run_alb_v2.py (+ meter.py interno) e proxy 0.3.5.
# 1) controlli: GPU, RAM disponibile >= 20 GB, disco libero su /mnt/mneme-nvme;
# 2) ctl-k15.sh stop (Strata a 15 esperti, prova non di produzione) PRIMA del lock; poi flock del lock GPU;
# 3) per ogni finestra (nell'ordine della lista) un motore Strata 0.1.40 di produzione (strata-prod.json, 10 esperti,
# STRATA_EXPERTS_USED tolto dall'ambiente) su :8297 con --max-context = finestra, KV int8,
# --kv-resident = min(finestra, 32768);
# 4) lavori "<policy> <window> <file> <qid>" in ordine; result.json presente = saltato (ripresa);
# un lavoro non parte se il tempo residuo < START_MARGIN;
# 5) dopo ogni lavoro check_c2.py (regole d'arresto del brief) e /health del motore; alla prima: STOP;
# 6) trap EXIT: chiude banco/proxy/motore del blocco, rilascia il lock, ctl-prod.sh start (PRODUZIONE a 10 esperti),
# verifica :8095/health, :8096/health e una domanda di prova (19x21) attraverso :8096 (proxy 0.3.4 installato,
# NON modificato).
# Uso: gpu_block_c2.sh LISTA OUTDIR (DRY=1: a secco, non tocca k15, produzione, lock vero né GPU)
set -u
export ALB=/mnt/mneme-nvme/proxy-ideas/alb-bench-v2/alb-clar
B=/mnt/mneme-nvme/proxy-ideas/alb-bench-v2
S=/mnt/mneme-nvme/strata-0140
K15=/mnt/mneme-nvme/strata-k15-run/ctl-k15.sh
LOCK=/mnt/mneme-nvme/.flash-next-gpu.lock
LIST=$1; OUT=$2
NAME=$(basename $LIST .txt)
MAXSEC=${MAXSEC:-14400}
START_MARGIN=${START_MARGIN:-900}
DRY=${DRY:-0}
EPORT=8297; PPORT=8496
P=$B/venv/bin/python
export PROXY_SRC=$B/alb/src/mnemonic-proxy-035
L=$B/logs; mkdir -p $L $OUT
LOG=$L/gpu-$NAME.log; [ "$DRY" = 1 ] && LOG=$L/gpu-$NAME-dry.log
exec >> $LOG 2>&1
echo "=== $(date -Is) blocco $NAME (DRY=$DRY, MAXSEC=$MAXSEC) -> $OUT"
echo "proxy: $(git -C $PROXY_SRC log --oneline -1)"
STATE=$OUT/.check-state.json; rm -f $STATE
cleanup() {
rc=$?
echo "=== $(date -Is) cleanup rc=$rc"
pkill -f "alb-bench-v2/scripts/run_alb_v2.py" 2>/dev/null
pkill -f "alb-bench-v2/venv/bin/mnemonic-proxy.*--port $PPORT" 2>/dev/null
if [ -f $L/eng-$NAME.pid ]; then
E=$(cat $L/eng-$NAME.pid); kill -- -$E 2>/dev/null
for i in $(seq 30); do kill -0 $E 2>/dev/null || break; sleep 1; done
kill -9 -- -$E 2>/dev/null; rm -f $L/eng-$NAME.pid
fi
sleep 3
exec 9>&-
if [ "$DRY" = 0 ]; then
nvidia-smi --query-gpu=memory.used --format=csv,noheader
echo "=== riavvio PRODUZIONE (10 esperti)"
env -u STRATA_EXPERTS_USED $S/ctl-prod.sh start
ok=0; for i in $(seq 180); do curl -s -m 2 127.0.0.1:8095/health | grep -q ok && { ok=1; break; }; sleep 5; done
$S/ctl-prod.sh status
echo "8095 ok=$ok"
tr '\0' ' ' < /proc/$(pgrep -f "strata-0140/src/build-cuda/strata" | head -1)/cmdline 2>/dev/null | head -c 600; echo
grep -c STRATA_EXPERTS_USED /proc/$(pgrep -f "strata-0140/src/build-cuda/strata" | head -1)/environ 2>/dev/null
echo "8096 health: $(curl -s -m 5 127.0.0.1:8096/health)"
curl -s -m 300 127.0.0.1:8096/v1/chat/completions -H 'Content-Type: application/json' \
-d '{"messages":[{"role":"user","content":"Quanto fa 19x21? Rispondi solo con il numero."}],"max_tokens":512}' \
> $L/prova-8096-$NAME.json
echo "prova 8096: $(python3 -c "import json;d=json.load(open('$L/prova-8096-$NAME.json'));print(repr(d['choices'][0]['message'].get('content')))" 2>&1)"
fi
echo "=== $(date -Is) fine blocco $NAME"
}
trap cleanup EXIT
trap 'exit 4' INT TERM
nvidia-smi --query-gpu=memory.used,utilization.gpu,power.limit,driver_version --format=csv
MEMA=$(awk '/MemAvailable/ {print int($2/1048576)}' /proc/meminfo)
DISK=$(df -B1G --output=avail /mnt/mneme-nvme | tail -1 | tr -d ' ')
echo "RAM disponibile ${MEMA} GB, disco libero ${DISK} GB"
if [ "$DRY" = 1 ]; then
LOCK=$L/dry.lock
else
echo "=== stop Strata k15 (prima del lock)"
$K15 stop; $K15 status
for i in $(seq 60); do pgrep -f "strata-k15/build-cuda/strata" > /dev/null || break; sleep 1; done
pgrep -f "strata-k15/build-cuda/strata" && { echo "k15 ANCORA VIVO"; exit 2; }
sleep 5
nvidia-smi --query-gpu=memory.used,utilization.gpu --format=csv,noheader
MEMA=$(awk '/MemAvailable/ {print int($2/1048576)}' /proc/meminfo)
echo "RAM disponibile dopo stop k15: ${MEMA} GB"
fi
[ "$MEMA" -ge 20 ] || { echo "RAM disponibile < 20 GB"; [ "$DRY" = 0 ] && exit 2; }
[ "$DISK" -ge 20 ] || { echo "disco libero < 20 GB"; exit 2; }
exec 9>$LOCK
flock -w 60 9 || { echo "LOCK OCCUPATO da $(fuser $LOCK 2>&1)"; exit 2; }
echo "lock preso"
if [ "$DRY" = 0 ]; then
pgrep -f "build-cuda/strata" && { echo "un motore Strata e' ancora vivo"; exit 2; }
nvidia-smi --query-gpu=memory.used --format=csv,noheader
fi
mkcfg() {
python3 - "$1" <<'EOF'
import json, sys
w = int(sys.argv[1])
d = json.load(open("/mnt/mneme-nvme/strata-0140/strata-prod.json"))
a = d["args"]
a[a.index("--max-context") + 1] = str(w)
a[a.index("--kv") + 1] = "int8"
a[a.index("--kv-resident") + 1] = str(min(w, 32768))
d.update(port=8297, log="/mnt/mneme-nvme/proxy-ideas/alb-bench-v2/logs/strata-engine-%d.log" % w)
json.dump(d, open("/mnt/mneme-nvme/proxy-ideas/alb-bench-v2/logs/strata-%d.json" % w, "w"), indent=1)
print("engine cfg", a)
EOF
}
wait_up() { for i in $(seq $(( $2 / 5 ))); do curl -s -m 2 $1/health | grep -q ok && return 0; sleep 5; done; echo "MOTORE NON PRONTO"; return 1; }
stop_engine() {
[ -f $L/eng-$NAME.pid ] || return 0
E=$(cat $L/eng-$NAME.pid); kill -- -$E 2>/dev/null
for i in $(seq 30); do kill -0 $E 2>/dev/null || break; sleep 1; done
kill -9 -- -$E 2>/dev/null; rm -f $L/eng-$NAME.pid; sleep 3
}
start_engine() {
local W=$1
mkcfg $W
if [ "$DRY" = 0 ]; then
cd $S/src
env -u STRATA_EXPERTS_USED setsid /mnt/mneme-nvme/strata-ab/Strata/.venv/bin/python serve/server.py --engine strata \
--config $L/strata-$W.json --host 127.0.0.1 --port $EPORT >> $L/strata-$NAME-$W.log 2>&1 9>&- &
echo $! > $L/eng-$NAME.pid
wait_up http://127.0.0.1:$EPORT 900 || return 1
curl -s 127.0.0.1:$EPORT/v1/status > $L/strata-status-$NAME-$W.json
nvidia-smi --query-gpu=memory.used --format=csv,noheader >> $L/vram-idle-$NAME-$W.txt
cd $B
fi
echo "=== $(date -Is) motore pronto W=$W"
}
work() {
CUR=
while read -r POL WW F Q <&3; do
[ -z "${POL:-}" ] && continue; case "$POL" in \#*) continue;; esac
R=$OUT/$POL/$(echo $F | cut -d/ -f1-2)/$Q/result.json
[ -f $R ] && { echo "salto $POL $F $Q"; continue; }
if [ $(( MAXSEC - SECONDS )) -lt $START_MARGIN ]; then
echo "=== $(date -Is) tempo residuo $(( MAXSEC - SECONDS ))s < START_MARGIN: blocco chiuso"; return 0
fi
if [ "$WW" != "$CUR" ]; then
[ "$DRY" = 0 ] && stop_engine
start_engine $WW || return 1
CUR=$WW
fi
echo "=== $(date -Is) $POL W=$WW $F $Q"
[ "$DRY" = 1 ] && { [ -f $B/alb-clar/data/benchmark/$F ] || { echo "FILE ASSENTE $F"; return 3; }; continue; }
nice -n 5 $P -W ignore $B/scripts/run_alb_v2.py --file $B/alb-clar/data/benchmark/$F --qid $Q --policy $POL \
--engine-url http://127.0.0.1:$EPORT --window $WW --out $OUT --proxy-port $PPORT --engine-port $EPORT \
9>&- 3<&- < /dev/null | tail -1 | tee -a $OUT/results.jsonl
[ -f $R ] || { echo "STOP: result.json assente ($POL $Q)"; return 5; }
python3 $B/scripts/check_c2.py $R $STATE 3<&- 9>&- || return 9
curl -s -m 5 127.0.0.1:$EPORT/health | grep -q ok || { echo "STOP: MOTORE GIU"; return 6; }
done 3< $LIST
[ "$DRY" = 0 ] && stop_engine
return 0
}
work &
WK=$!
( exec 9>&-; sleep $(( MAXSEC - SECONDS )); echo "TEMPO MASSIMO DEL BLOCCO"; kill -TERM $WK 2>/dev/null;
pkill -f "alb-bench-v2/scripts/run_alb_v2.py" ) &
T=$!
wait $WK; rc=$?
kill $T 2>/dev/null; pkill -P $T 2>/dev/null
echo "lavoro rc=$rc"
exit $rc
|