#!/usr/bin/env bash
# Complete the NATIVE-config RULER curve for the 128k models: T=131072 YaRN x32
# ORIG=4096 (factor = 131072/4096, exactly Qwen's target/original pattern), held
# FIXED across all lengths -- which is what RULER does via vLLM static YaRN.
# Missing cells: 8,192 and 16,360.
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:$PATH; export PY=/root/Adhi/BluTrain/envv/bin/python
L=eval_native_logs; mkdir -p "$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; }
: # ordering handled by eval_chain.sh
say "=== native x32 fill-in: c300 / d360 at 8,192 and 16,360 ==="
for PL in 8192 16360; do
  O="en_${PL}"
  A1B="$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" CKPTS="c300 d360" \
  TASKS="niah_single" N=100 T=131072 YARN=32 YARN_ORIG=4096 PROMPT_LEN="$PL" \
  GPUS="0,1,2,3" NP=1 JOBS_PER_GPU=2 OUT="$O" ./run_longeval.sh > "$L/len_${PL}.log" 2>&1
  say "  $PL: $(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 "EVAL_NATIVE_DONE"