#!/usr/bin/env bash
# The config the 98 came from: T=262,144 YaRN x64 ORIG=4096 -> virtual 1,023 @65,512.
# Every t128 eval so far used T=131,072 x32 -> virtual 2,047, where even t64 only
# scores 63. Same checkpoint, 35 points apart. This is the apples-to-apples test
# against t64=98 / old-t128=54 / base680=75.
set -uo pipefail; cd "$(dirname "$0")"
export LD_LIBRARY_PATH="$(cd ../.. && pwd)/Tensor-Implementations/lib:$(cd ../.. && pwd)/Profiler/lib:${LD_LIBRARY_PATH:-}"
export PATH=/usr/local/cuda/bin:$PATH; export PY=/root/Adhi/BluTrain/envv/bin/python
L=eval_x64_logs; mkdir -p "$L"
A680="CP_N_EMBD=1536 CP_N_LAYER=24 CP_N_HEAD=12 CP_N_KVHEAD=4 CP_FFN=4096 CP_WEIGHT_TYING=1"
say(){ echo "[$(date -u +%H:%M:%S)] $*" | tee -a "$L/master.log" >&2; }
# wait for the x32 evals to release the GPUs
: # ordering handled by eval_chain.sh
say "=== x64 / T=262144: c300 d360 t64 at 65,512 then 131,000 ==="
for PL in 65512 131000; do
  O="ex64_${PL}"
  A1B="$A680" HAY="data/mix128_b6_60/val_shard_00000.bin" GEN_DATA_ROOT="data/mix128_b6_60" \
  CKPT_EXTRA="c300=t128c_s300.ckpt d360=t128d_s360.ckpt t64=t64.ckpt" CKPTS="c300 d360 t64" \
  TASKS="niah_single" N=100 T=262144 YARN=64 YARN_ORIG=4096 PROMPT_LEN="$PL" \
  GPUS="0,1,2,3" NP=1 JOBS_PER_GPU=1 OUT="$O" ./run_longeval.sh > "$L/len_${PL}.log" 2>&1
  say "  $PL (virtual $((PL/64))): $(head -1 $O/summary.csv 2>/dev/null|tr -d '\r') -> $(grep -h '^niah_single,' $O/summary.csv 2>/dev/null|tr -d '\r')"
done
say "ref @x64: t64=98  base680=75  old-t128=54  (@65,512, virtual 1,023)"
say "EVAL_X64_DONE"