#!/usr/bin/env bash
# THE DECISIVE TEST: YaRN OFF (scale=1). virtual position == real position.
# Everything inside the 4,096 pretrain range, no compression, no interpolation.
#
# Every score we have is a function of VIRTUAL position and not token count:
#   v128 -> ~99,  v1023 -> ~92,  v2047 -> 45-63,  v4094 -> 41
# and t128c EQUALS t64 at v128 and v1023. So extension training added no new
# positional competence; only inference-time compression ever helped.
#
# If these models fail at REAL position 2,048-4,096 with scale=1, the ceiling is
# in the base model's pretraining (theta=500,000 at a 4,096 pretrain leaves 32 of
# 64 RoPE dims without a full rotation) and no extension recipe can fix it.
# If they score ~95 flat, positions 1-4096 are healthy and YaRN's ramp at
# scale>1 is what destroys v2047 -- a fixable extension problem.
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:$PATH; export PY=/root/Adhi/BluTrain/envv/bin/python
L=eval_scale1_logs; mkdir -p "$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 "=== scale=1 (YaRN OFF): p1_s200 / t64 / c300 at 1024, 2048, 4096 ==="
for PL in 1024 2048 4096; do
  O="es1_${PL}"
  A1B="$A680" HAY="data/mix128_b6_60/val_shard_00000.bin" GEN_DATA_ROOT="data/mix128_b6_60" \
  CKPT_EXTRA="p1=p1_s200.ckpt t64=t64.ckpt c300=t128c_s300.ckpt" CKPTS="p1 t64 c300" \
  TASKS="niah_single" N=100 T=8192 YARN=1 YARN_ORIG=4096 PROMPT_LEN="$PL" \
  GPUS="0,1,2,3" NP=1 JOBS_PER_GPU=3 OUT="$O" ./run_longeval.sh > "$L/len_${PL}.log" 2>&1
  say "  real pos $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 "compare: at virtual 1023 every model scores ~92; at v2047, 45-63"
say "EVAL_SCALE1_DONE"