Upload scripts/eval_long.sh with huggingface_hub

scripts/eval_long.sh

@@ -0,0 +1,31 @@
+#!/usr/bin/env bash
+# THE MISSING EVAL: 65,512 and 131,000 -- the lengths this rung actually targets.
+# Everything measured so far was 4,096 / 32,744, i.e. a quarter of the window.
+# T=131072 YaRN x32 ORIG=4096 -> virtual 2,047 @65,512 and 4,094 @131,000.
+# 131,000 sits at the very edge of the 4,096 pretrain range, so expect a drop
+# for every model; the comparison between models is what matters.
+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:$PATH; export PY=/root/Adhi/BluTrain/envv/bin/python
+L=eval_long_logs; mkdir -p "$L"; M="$L/master.log"
+A680="CP_N_EMBD=1536 CP_N_LAYER=24 CP_N_HEAD=12 CP_N_KVHEAD=4 CP_FFN=4096 CP_WEIGHT_TYING=1"
+EX="c300=t128c_s300.ckpt d300=t128d_s300.ckpt d360=t128d_s360.ckpt"
+NM="c300 d300 d360"
+say(){ echo "[$(date -u +%H:%M:%S)] $*" | tee -a "$M" >&2; }
+blk(){ local PL="$1"; local JPG="$2"; local O="el_${PL}"
+  [ -s "$O/summary.csv" ] && grep -q '^niah_single,' "$O/summary.csv" && { say "  $PL cached"; return; }
+  A1B="$A680" HAY="data/mix128_b6_60/val_shard_00000.bin" GEN_DATA_ROOT="data/mix128_b6_60" \
+  CKPT_EXTRA="$EX" CKPTS="$NM" TASKS="niah_single" N=100 \
+  T=131072 YARN=32 YARN_ORIG=4096 PROMPT_LEN="$PL" GPUS="0,1,2,3" NP=1 JOBS_PER_GPU="$JPG" OUT="$O" \
+    ./run_longeval.sh > "$L/len_${PL}.log" 2>&1
+  say "  len $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')"; }
+say "=== LONG eval: c300 / d300 / d360 at 65,512 then 131,000 ==="
+blk 65512 2                 # 3 jobs, 8 slots
+say "  65k done, starting 131k at JOBS_PER_GPU=1 (4x the activations)"
+blk 131000 1                # 3 jobs, 4 slots -- conservative on memory
+say "=== LONG RESULTS ==="
+for PL in 65512 131000; do
+  say "  $PL: $(head -1 el_${PL}/summary.csv 2>/dev/null|tr -d '\r') -> $(grep -h '^niah_single,' el_${PL}/summary.csv 2>/dev/null|tr -d '\r')"
+done
+say "ref @65,512 x32: t64=63  old-t128=31   (@131,000 never measured for anything)"
+say "EVAL_LONG_DONE"