#!/usr/bin/env bash
# Measure, do not assume: is batch-1 decode latency-bound (concurrency helps) or
# compute-bound (it does not)? The script header claims ~1.8x at 2/GPU on an
# H100 for a 1.06B; this is a 680M on an H200 at a different length, so measure.
set -uo pipefail
cd "$(dirname "$0")"
export PATH=/root/Adhi/BluTrain/envv/bin:$PATH
source /root/Adhi/BluTrain/envv/bin/activate 2>/dev/null
A680="CP_N_EMBD=1536 CP_N_LAYER=24 CP_N_HEAD=12 CP_N_KVHEAD=4 CP_FFN=4096 CP_WEIGHT_TYING=1"
HAY=Data_Loader/Data/b2_8192_16384/val_shard_00000.bin
T=16384; S=8; PL=8192      # T must exceed PL+MAX_NEW; x8 -> virtual 1,024
TASKS4="niah_single niah_multikey niah_multivalue vt"
for J in 1 4; do
  rm -rf bench_j${J}; t0=$(date +%s)
  A1B="$A680" HAY="$HAY" GEN_DATA_ROOT="Data_Loader/Data/evalsrc" \
  CKPT_EXTRA="t64=t64.ckpt" CKPTS="t64" TASKS="$TASKS4" N=100 \
  T=$T YARN=$S YARN_ORIG=4096 PROMPT_LEN=$PL GPUS=0 NP=1 JOBS_PER_GPU=$J OUT="bench_j${J}" \
    ./run_longeval.sh > "bench_j${J}.log" 2>&1
  t1=$(date +%s)
  pk=$(grep -aoE 'mem: [0-9]+MB' bench_j${J}/gen/*.log 2>/dev/null | grep -oE '[0-9]+' | sort -n | tail -1)
  echo "  JOBS_PER_GPU=$J : $(( (t1-t0) )) s for 4 tasks, peak/job ${pk:-?} MB"
done