Upload scripts/dl_64k.sh with huggingface_hub
Browse files- scripts/dl_64k.sh +46 -0
scripts/dl_64k.sh
ADDED
|
@@ -0,0 +1,46 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env bash
|
| 2 |
+
# Everything the 64k extension needs. Explicit --include on every call.
|
| 3 |
+
set -uo pipefail; cd "$(dirname "$0")"
|
| 4 |
+
export PATH=/root/Adhi/BluTrain/envv/bin:$PATH
|
| 5 |
+
say(){ echo "[$(date -u +%H:%M:%S)] $*"; }
|
| 6 |
+
mkdir -p checkpoints_680m/rank_0 checkpoints_bluscriptcp/rank_0 \
|
| 7 |
+
Data_Loader/Data/b2_8192_16384 Data_Loader/Data/evalsrc
|
| 8 |
+
R=unparallelled/BluScriptCP; RD=unparallelled/BluScriptCP-data
|
| 9 |
+
|
| 10 |
+
say "1/5 p1final (the parent -- best model we have, 95/90/82 at x64)"
|
| 11 |
+
hf download $R --quiet --include 'checkpoints/p1_680m_step1200_rank0.ckpt' --local-dir dlp || exit 1
|
| 12 |
+
mv -f dlp/checkpoints/p1_680m_step1200_rank0.ckpt checkpoints_bluscriptcp/rank_0/p1final.ckpt; rm -rf dlp
|
| 13 |
+
|
| 14 |
+
say "2/5 base680 (the control -- without it we cannot read the result)"
|
| 15 |
+
hf download navingv/varybatch --quiet --include 'varybatch.ckpt' --local-dir dlv || exit 1
|
| 16 |
+
mv -f dlv/varybatch.ckpt checkpoints_680m/rank_0/; rm -rf dlv
|
| 17 |
+
ln -sf "$PWD/checkpoints_680m/rank_0/varybatch.ckpt" checkpoints_bluscriptcp/rank_0/varybatch.ckpt
|
| 18 |
+
|
| 19 |
+
say "3/5 haystack -- MUST be byte-identical to every prior eval"
|
| 20 |
+
hf download $RD --repo-type dataset --quiet \
|
| 21 |
+
--include 'cp_data_by_length/b2_8192_16384/val_shard_00000.bin' --local-dir dlh || exit 1
|
| 22 |
+
mv -f dlh/cp_data_by_length/b2_8192_16384/val_shard_00000.bin Data_Loader/Data/b2_8192_16384/; rm -rf dlh
|
| 23 |
+
|
| 24 |
+
say "4/5 long buckets for a 64k window"
|
| 25 |
+
get(){ for i in $(seq -f '%05g' "$2" "$3"); do
|
| 26 |
+
hf download $RD --repo-type dataset --quiet \
|
| 27 |
+
--include "cp_data_by_length/$1/train_shard_$i.bin" --local-dir dld || return 1; done; }
|
| 28 |
+
get b5_65536_131072 0 8 || exit 1 # docs >= the 64k window: contiguous
|
| 29 |
+
get b4_32768_65536 0 5 || exit 1
|
| 30 |
+
get b3_16384_32768 0 3 || exit 1
|
| 31 |
+
get b1_4096_8192 0 2 || exit 1
|
| 32 |
+
for d in dld/cp_data_by_length/*; do n=$(basename "$d"); mkdir -p "src64/$n"
|
| 33 |
+
for f in "$d"/*.bin; do [ -e "$f" ] && mv -n "$f" "src64/$n/"; done; done
|
| 34 |
+
rm -rf dld
|
| 35 |
+
|
| 36 |
+
say "5/5 loader feed (trainer builds a data loader even for pure generation)"
|
| 37 |
+
cp -f src64/b1_4096_8192/train_shard_00000.bin Data_Loader/Data/evalsrc/ 2>/dev/null
|
| 38 |
+
cp -f Data_Loader/Data/b2_8192_16384/val_shard_00000.bin Data_Loader/Data/evalsrc/
|
| 39 |
+
|
| 40 |
+
say "=== sizes ==="
|
| 41 |
+
stat -c ' %s %n' checkpoints_bluscriptcp/rank_0/p1final.ckpt \
|
| 42 |
+
checkpoints_680m/rank_0/varybatch.ckpt \
|
| 43 |
+
Data_Loader/Data/b2_8192_16384/val_shard_00000.bin 2>/dev/null
|
| 44 |
+
say "expected: p1final 5141244514 | varybatch 1362674564 | haystack 199754694"
|
| 45 |
+
du -sh src64/* ; df -h . | tail -1
|
| 46 |
+
say "DL_DONE"
|