weight compression script

compress_weights.bash

@@ -0,0 +1,38 @@
+#!/usr/bin/env bash
+
+# this script generates compressed blobs for a model into huggingface's cache, to be transferred elsewhere and decompressed
+# the size is reduced by about 22%
+
+LEVEL=-4
+#LEVEL=-9
+#LEVEL=--adapt=min=1,max=22
+
+model_id="$1"
+revision="$2"
+repo_url=https://huggingface.co/"$model_id"
+clonepath="$(basename "$repo_url")"
+cachepath="${HOME}/.cache/huggingface/hub/models--${model_id//\//--}/blobs"
+HF_TOKEN="$(<${HOME}/.cache/huggingface/token)"
+GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/"$1" "$clonepath"
+mkdir -p "$cachepath"
+cd "$clonepath"
+git checkout "$revision" || exit -1
+filecount=$(ls *.safetensors | wc -l)
+for file in *.safetensors
+do
+	file_url="$repo_url"/resolve/"$revision"/"$file"
+	export $(cat "$file" | tr \  =) # sets oid and size
+	blobname=${oid##*:}
+	outfile="$cachepath"/"$blobname".zst
+	if [ -e "$outfile" ]; then
+		zstd -l "$outfile" && { echo "$file" 1>&3; continue; }
+		echo
+		echo "# Recompressing $file #"
+		echo
+	fi
+	curl -s -L -H "Authorization: Bearer $HF_TOKEN" "$file_url" | pv -cpteabm $((60*60*24)) -s "$size" -N "$file" | zstd -q --no-progress --force --size-hint="$size" --stream-size="$size" --ultra --long=31 --memory=2048MB --threads=1024 "$LEVEL" | pv -cpteabm $((60*60*24)) -s "$((size * 100/129))" -N "$blobname".zst > "$outfile"
+	zstd -l "$outfile"
+	echo
+	echo
+	echo "$file" 1>&3
+done 3>&1 1>&2 | pv -lcpteabm $((60*60*24)) -s "$filecount" -N "$clonepath"