1  #!/bin/bash
 2  set -euo pipefail
 3  this=$(realpath "$0"); readonly this
 4  cd "$(dirname "$this")"
 5  shellcheck "$this"
 6  
 7  if (( $# != 1 && $# != 2  )); then
 8      cat >&2 <<'EOF'
 9  usage:
10      ci-run.sh <tmp_dir> [<cache_dir>]
11  
12  This script wraps ci/run.sh:
13  * If <tmp_dir> is a ramdisk, you can reduce writes to your SSD. If <tmp_dir> is not a ramdisk, keep in mind that total writes will increase by the size of <cache_dir>.
14      (openllama_3b_v2: quantized models are about 30GB)
15  * Persistent model and data files are synced to and from <cache_dir>,
16      excluding generated .gguf files.
17      (openllama_3b_v2: persistent files are about 6.6GB)
18  * <cache_dir> defaults to  ~/.cache/llama.cpp
19  EOF
20      exit 1
21  fi
22  
23  cd .. # => llama.cpp repo root
24  
25  tmp="$1"
26  mkdir -p "$tmp"
27  tmp=$(realpath "$tmp")
28  echo >&2 "Using tmp=$tmp"
29  
30  cache="${2-$HOME/.cache/llama.cpp}"
31  mkdir -p "$cache"
32  cache=$(realpath "$cache")
33  echo >&2 "Using cache=$cache"
34  
35  _sync() {
36      local from="$1"; shift
37      local to="$1"; shift
38  
39      echo >&2 "Syncing from $from to $to"
40      mkdir -p "$from" "$to"
41      rsync -a "$from" "$to" --delete-during "$@"
42  }
43  
44  _sync "$(realpath .)/" "$tmp/llama.cpp"
45  _sync "$cache/ci-mnt/models/" "$tmp/llama.cpp/ci-mnt/models/"
46  
47  cd "$tmp/llama.cpp"
48  bash ci/run.sh ci-out ci-mnt
49  
50  _sync 'ci-mnt/models/' "$cache/ci-mnt/models/" --exclude='*.gguf' -P