Upload scripts/eval_collapse.sh with huggingface_hub

scripts/eval_collapse.sh

@@ -0,0 +1,40 @@
+#!/usr/bin/env bash
+# Did retrieval degrade across the loss decline? 4 ckpts x 2 lengths, 8 GPUs.
+# T=131072 YaRN x32 ORIG=4096, N=100, haystack 199,754,694 B (identical to all
+# prior evals). Three things that bit me: PY must be the venv (python3 has no
+# tiktoken), CKPT_EXTRA takes a BARE filename (the harness prepends rank_0/),
+# and `local A=$1 B="${A}"` expands A before local assigns it -> set -u abort.
+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:$PATH
+export PY=/root/Adhi/BluTrain/envv/bin/python
+L=eval_collapse_logs; mkdir -p "$L"; M="$L/master.log"
+HAY=data/mix128_b6_35_640M/val_shard_00000.bin
+A680="CP_N_EMBD=1536 CP_N_LAYER=24 CP_N_HEAD=12 CP_N_KVHEAD=4 CP_FFN=4096 CP_WEIGHT_TYING=1"
+say(){ echo "[$(date -u +%H:%M:%S)] $*" | tee -a "$M" >&2; }
+[ "$(stat -c%s "$HAY")" = 199754694 ] || { say "FATAL haystack differs"; exit 1; }
+$PY -c "import tiktoken,numpy" || { say "FATAL venv lacks tiktoken/numpy"; exit 1; }
+
+job(){
+  local S="$1"; local PL="$2"; local G="$3"; local O="ec_s${S}_${PL}"
+  [ -s "$O/summary.csv" ] && grep -q '^niah_single,' "$O/summary.csv" && { say "  s$S @$PL cached"; return; }
+  A1B="$A680" HAY="$HAY" GEN_DATA_ROOT="data/mix128_b6_35_640M" \
+  CKPT_EXTRA="s${S}=t128c_s${S}.ckpt" CKPTS="s${S}" TASKS="niah_single" N=100 \
+  T=131072 YARN=32 YARN_ORIG=4096 PROMPT_LEN="$PL" GPUS="$G" NP=1 OUT="$O" \
+    ./run_longeval.sh > "$L/s${S}_${PL}.log" 2>&1
+  say "  s$S @$PL -> $(grep -h '^niah_single,' "$O/summary.csv" 2>/dev/null | cut -d, -f2 | tr -d '\r')"
+}
+say "=== eval across the loss decline: 300/400/450/500 x {4096,32744} @x32 ==="
+job 300  4096 0 &  job 400  4096 1 &  job 450  4096 2 &  job 500  4096 3 &
+job 300 32744 4 &  job 400 32744 5 &  job 450 32744 6 &  job 500 32744 7 &
+wait
+say "=== RESULTS  niah_single, T=131072 x32 ORIG=4096, N=100 ==="
+printf "  %-6s %8s %8s\n" "step" "4096" "32744" | tee -a "$M"
+for s in 300 400 450 500; do
+  a=$(grep -h '^niah_single,' "ec_s${s}_4096/summary.csv"  2>/dev/null | cut -d, -f2 | tr -d '\r')
+  b=$(grep -h '^niah_single,' "ec_s${s}_32744/summary.csv" 2>/dev/null | cut -d, -f2 | tr -d '\r')
+  printf "  %-6s %8s %8s\n" "$s" "${a:-.}" "${b:-.}" | tee -a "$M"
+done
+say "ref: run2 s300 100/89  s500 95/77 | t64 100/93 | old-t128 98/69"
+say "EVAL DONE"