Upload scripts/ctl_106b.sh with huggingface_hub

scripts/ctl_106b.sh

@@ -0,0 +1,72 @@
+#!/usr/bin/env bash
+# =============================================================================
+# THE CONTROL: is the ceiling a property of the CONFIG or of the 680M?
+#
+# mix64k_step1200 (1.06B, best of the 3-stage ladder) was TRAINED at
+# ORIG=32768 SCALE=2, which puts a 16,360 prompt at virtual 8,180 -- twice
+# beyond the 4,096 range it was pretrained on. Re-parameterising to
+# ORIG=4096 SCALE=16 (same T=65,536) puts it at virtual 1,022.
+#
+# mixcont_g1400_cliff is the COLLAPSED checkpoint. If the collapse is partly a
+# config artefact it should recover too; if it is real damage it should not.
+#
+# PRE-REGISTERED PREDICTIONS (mix64k), written before the run:
+#   16,360 @ x16 -> 85-92   (known 70 at its trained config)
+#   32,744 @ x32 -> 70-80   (known 50)
+#   65,512 @ x32 -> 40-50   (known 31)
+#   16,360 @ x2/ORIG=32768 -> 66-74  <- REPRODUCTION CHECK of the prior 70.
+#      If this does not reproduce, the harness differs from the prior session
+#      and NONE of the cross-session comparisons in this analysis are valid.
+# =============================================================================
+set -uo pipefail
+cd "$(dirname "$0")"
+export PATH=/root/Adhi/BluTrain/envv/bin:$PATH
+source /root/Adhi/BluTrain/envv/bin/activate 2>/dev/null
+L=ctl_logs; mkdir -p "$L"; M="$L/master.log"
+say(){ echo "[$(date -u +%H:%M:%S)] $*" | tee -a "$M" >&2; }
+CK=checkpoints_bluscriptcp/rank_0
+HAY=Data_Loader/Data/b2_8192_16384/val_shard_00000.bin
+[ "$(stat -c%s "$HAY")" = 199754694 ] || { say "FATAL haystack differs"; exit 1; }
+
+for f in mix64k_step_1200 mixcont_g1400_cliff; do
+  [ -f "$CK/$f.ckpt" ] && continue
+  say "downloading $f.ckpt"
+  hf download unparallelled/BluScriptCP --quiet --include "checkpoints/$f.ckpt" --local-dir dl_c || exit 1
+  mv -f "dl_c/checkpoints/$f.ckpt" "$CK/" || exit 1
+done
+rm -rf dl_c
+for f in mix64k_step_1200 mixcont_g1400_cliff; do
+  [ "$(stat -c%s "$CK/$f.ckpt")" = 3750897182 ] || { say "FATAL $f wrong size"; exit 1; }
+done
+say "both 1.06B checkpoints verified at 3,750,897,182 B"
+
+# 1.06B arch is run_longeval.sh's DEFAULT A1B (36 layers, weight_tying=0) -- do not override
+run(){ # T SCALE ORIG PL tag
+  local T=$1 S=$2 OR=$3 PL=$4 TAG=$5
+  local O="ctl_${TAG}_len${PL}"
+  [ $((S*OR)) -eq $T ] || { say "YaRN invariant $S*$OR != $T"; return 1; }
+  [ -s "$O/summary.csv" ] && [ "$(wc -l < "$O/summary.csv")" -gt 1 ] && {
+    say "  $TAG len $PL cached: $(grep '^niah_single,' $O/summary.csv)"; return 0; }
+  say "  $TAG len $PL  (T=$T scale=$S orig=$OR -> virtual $((PL/S)))"
+  HAY="$HAY" GEN_DATA_ROOT="Data_Loader/Data/evalsrc" \
+  CKPT_EXTRA="mix1200=mix64k_step_1200.ckpt cliff1400=mixcont_g1400_cliff.ckpt" \
+  CKPTS="mix1200 cliff1400" TASKS=niah_single N=100 \
+  T=$T YARN=$S YARN_ORIG=$OR PROMPT_LEN=$PL GPUS=0,1 NP=1 OUT="$O" \
+    ./run_longeval.sh >> "$L/${TAG}_${PL}.log" 2>&1
+  if [ -s "$O/summary.csv" ] && [ "$(wc -l < "$O/summary.csv")" -gt 1 ]; then
+    say "    -> $(head -1 $O/summary.csv) | $(grep '^niah_single,' $O/summary.csv)"
+  else say "    FAILED"; grep -ahA2 'what()' "$O"/gen/*.log 2>/dev/null|head -3|tee -a "$M" >&2; fi
+}
+say "=== 1.06B CONTROL: config vs model ==="
+say "REPRODUCTION CHECK first -- if 16,360 @ x2/orig32768 is not ~70, stop trusting everything"
+run 65536  2 32768 16360 astrained
+say "now the corrected parameterisation (ORIG held at the 4096 pretrain length)"
+run 65536 16  4096 16360 x16
+run 131072 32 4096 32744 x32
+run 131072 32 4096 65512 x32
+say "=== CONTROL RESULTS (cols: mix1200, cliff1400) ==="
+for t in astrained_16360 x16_16360 x32_32744 x32_65512; do
+  f="ctl_${t%_*}_len${t##*_}/summary.csv"
+  [ -f "$f" ] && say "  $t  $(head -1 $f) -> $(grep '^niah_single,' $f)"
+done
+say "DONE"