unparallelled commited on
Commit
ba6a188
·
verified ·
1 Parent(s): f6b525a

Upload scripts/dl_vb.sh with huggingface_hub

Browse files
Files changed (1) hide show
  1. scripts/dl_vb.sh +30 -0
scripts/dl_vb.sh ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env bash
2
+ # Only what the val_backward diagnostic needs.
3
+ set -uo pipefail; cd "$(dirname "$0")"
4
+ export PATH=/root/Adhi/BluTrain/envv/bin:$PATH
5
+ say(){ echo "[$(date -u +%H:%M:%S)] $*"; }
6
+ mkdir -p checkpoints_680m/rank_0 checkpoints_bluscriptcp/rank_0 \
7
+ Data_Loader/Data/b2_8192_16384 Data_Loader/Data/evalsrc
8
+ R=unparallelled/BluScriptCP; RD=unparallelled/BluScriptCP-data
9
+ say "1/4 base680 (varybatch)"
10
+ hf download navingv/varybatch --quiet --include 'varybatch.ckpt' --local-dir dlv || exit 1
11
+ mv -f dlv/varybatch.ckpt checkpoints_680m/rank_0/; rm -rf dlv
12
+ ln -sf "$PWD/checkpoints_680m/rank_0/varybatch.ckpt" checkpoints_bluscriptcp/rank_0/varybatch.ckpt
13
+ say "2/4 p1final (best model: 95/90/82 at x64)"
14
+ hf download $R --quiet --include 'checkpoints/p1_680m_step1200_rank0.ckpt' --local-dir dlp || exit 1
15
+ mv -f dlp/checkpoints/p1_680m_step1200_rank0.ckpt checkpoints_bluscriptcp/rank_0/p1final.ckpt; rm -rf dlp
16
+ say "3/4 haystack -- byte-identical to every prior eval"
17
+ hf download $RD --repo-type dataset --quiet \
18
+ --include 'cp_data_by_length/b2_8192_16384/val_shard_00000.bin' --local-dir dlh || exit 1
19
+ mv -f dlh/cp_data_by_length/b2_8192_16384/val_shard_00000.bin Data_Loader/Data/b2_8192_16384/; rm -rf dlh
20
+ say "4/4 loader feed (trainer builds a loader even for pure generation)"
21
+ hf download $RD --repo-type dataset --quiet \
22
+ --include 'cp_data_by_length/b1_4096_8192/train_shard_00000.bin' --local-dir dlt || exit 1
23
+ mv -f dlt/cp_data_by_length/b1_4096_8192/train_shard_00000.bin Data_Loader/Data/evalsrc/; rm -rf dlt
24
+ cp -f Data_Loader/Data/b2_8192_16384/val_shard_00000.bin Data_Loader/Data/evalsrc/
25
+ say "=== sizes ==="
26
+ stat -c ' %s %n' checkpoints_680m/rank_0/varybatch.ckpt \
27
+ checkpoints_bluscriptcp/rank_0/p1final.ckpt \
28
+ Data_Loader/Data/b2_8192_16384/val_shard_00000.bin
29
+ say "expected: varybatch 1362674564 | p1final 5141244514 | haystack 199754694"
30
+ say "DL_DONE"