Ollama Model Export Script

ollama-export.sh

#!/bin/bash
# Ollama Model Export Script
# Usage: bash ollama-export.sh vicuna:7b
# SPDX-License-Identifier: MIT (https://ncurl.xyz/s/o_o6DVqIR)
# https://gist.github.com/supersonictw/f6cf5e599377132fe5e180b3d495c553

# Interrupt if any error occurred
set -e

# Declare
echo "Ollama Model Export Script"
echo "License: MIT (https://ncurl.xyz/s/o_o6DVqIR)"
echo ""

# OS-specific
case "$OSTYPE" in
linux*)
    HOME="$(echo ~ollama)"
    ;;
esac

# Define variables
OLLAMA_HOME="${OLLAMA_HOME:="$HOME/.ollama"}"
BLOBS_FILE_BASE_PATH="$OLLAMA_HOME/models/blobs"
MANIFESTS_FILE_BASE_PATH="$OLLAMA_HOME/models/manifests"

# Define constants
SUCCESS_PREFIX="\033[1;32mSuccess:\033[0m"
FAILED_PREFIX="\033[0;31mFailed:\033[0m"

# Read arguments
IFS='/' read -ra NAME_ARGS <<< "${1/://}"
case "${#NAME_ARGS[@]}" in
4)
    MANIFESTS_REGISTRY_NAME="${NAME_ARGS[0]}"
    MANIFESTS_LIBRARY_NAME="${NAME_ARGS[1]}"
    MANIFESTS_MODEL_NAME="${NAME_ARGS[2]}"
    MANIFESTS_PARAMS_NAME="${NAME_ARGS[3]}"
    ;;
3)
    MANIFESTS_LIBRARY_NAME="${NAME_ARGS[0]}"
    MANIFESTS_MODEL_NAME="${NAME_ARGS[1]}"
    MANIFESTS_PARAMS_NAME="${NAME_ARGS[2]}"
    ;;
2)
    MANIFESTS_MODEL_NAME="${NAME_ARGS[0]}"
    MANIFESTS_PARAMS_NAME="${NAME_ARGS[1]}"
    ;;
1)
    MANIFESTS_MODEL_NAME="${NAME_ARGS[0]}"
    ;;
esac

# Define variables
MANIFESTS_REGISTRY_NAME="${MANIFESTS_REGISTRY_NAME:="registry.ollama.ai"}"
MANIFESTS_LIBRARY_NAME="${MANIFESTS_LIBRARY_NAME:="library"}"
MANIFESTS_MODEL_NAME="${MANIFESTS_MODEL_NAME:="vicuna"}"
MANIFESTS_PARAMS_NAME="${MANIFESTS_PARAMS_NAME:="latest"}"

# Echo the model full name
MODEL_FULL_NAME="$MANIFESTS_MODEL_NAME:$MANIFESTS_PARAMS_NAME"
echo "Exporting model \"$MODEL_FULL_NAME\"..."
echo ""

# Make sure manifests exist
MANIFESTS_FILE_PATH="$MANIFESTS_FILE_BASE_PATH/$MANIFESTS_REGISTRY_NAME/$MANIFESTS_LIBRARY_NAME/$MANIFESTS_MODEL_NAME/$MANIFESTS_PARAMS_NAME"
if [ ! -f "$MANIFESTS_FILE_PATH" ]; then
    echo -e "$FAILED_PREFIX \"$MANIFESTS_FILE_PATH\" not exists, the model \"$MODEL_FULL_NAME\" you requested is not found."
    exit 1
fi

# Make sure dist not exist
EXPORT_DST_BASE_PATH="${EXPORT_DST_BASE_PATH:="$PWD/${MODEL_FULL_NAME/:/-}"}"
if [ -d "$EXPORT_DST_BASE_PATH" ]; then
    echo -e "$FAILED_PREFIX \"$EXPORT_DST_BASE_PATH\" already exists, exits for preventing from unexpected operations."
    exit 1
fi

# Create dist directory
mkdir -p "$EXPORT_DST_BASE_PATH"
printf "%s" "$MANIFESTS_REGISTRY_NAME/$MANIFESTS_LIBRARY_NAME/$MANIFESTS_MODEL_NAME:$MANIFESTS_PARAMS_NAME" >"$EXPORT_DST_BASE_PATH/source.txt"

# Read manifests and handle them
while IFS= read -r layer; do
    BLOB_FILE_NAME="${layer/:/-}"
    BLOB_FILE_PATH="$BLOBS_FILE_BASE_PATH/$BLOB_FILE_NAME"
    BLOB_TYPE_NAME=$(jq -r --arg layer "$layer" '.layers[] | select(.digest == $layer) | .mediaType' "$MANIFESTS_FILE_PATH" | sed 's|.*\.ollama\.image\.\(.*\)|\1|')

    EXPORT_MODEL_FILE_PATH="$EXPORT_DST_BASE_PATH/Modelfile"
    EXPORT_MODEL_BIN_PATH="$EXPORT_DST_BASE_PATH/model.bin"

    case "$BLOB_TYPE_NAME" in
    model)
        cp "$BLOB_FILE_PATH" "$EXPORT_MODEL_BIN_PATH"
        printf "%s\n" "FROM ./model.bin" >>"$EXPORT_MODEL_FILE_PATH"
        ;;
    params)
        PARAMS_JSON="$(cat "$BLOB_FILE_PATH")"
        printf "%s" "$(jq -r 'keys[] as $key | .[$key][] | "PARAMETER \($key) \"\(.)\"" ' <<<"$PARAMS_JSON")" >>"$EXPORT_MODEL_FILE_PATH"
        ;;
    *)
        TYPE_NAME="$(echo "$BLOB_TYPE_NAME" | tr '[:lower:]' '[:upper:]')"
        FILE_CONTENT="$(cat "$BLOB_FILE_PATH")"
        printf "%s\n" "$TYPE_NAME \"\"\"$FILE_CONTENT\"\"\"" >>"$EXPORT_MODEL_FILE_PATH"
        ;;
    esac
done < <(jq -r '.layers[].digest' "${MANIFESTS_FILE_PATH}")

# Echo success message
echo -e "$SUCCESS_PREFIX Model \"$MODEL_FULL_NAME\" has been exported to \"$EXPORT_DST_BASE_PATH\"!"

@sokovnich
The patch already applied!
Linux can recognize their OLLAMA_HOME automatically. 🐧

@supersonictw thank you

i got an warning while exporting llava:34b model:
/ollama-export.sh: line 103: warning: command substitution: ignored null byte in input

i inspected the llava model manifest file, and the warning is maybe caused by the mediaType: projector:

{
      "mediaType": "application/vnd.ollama.image.projector",
      "digest": "sha256:83720bd8438ccdc910deba5efbdc3340820b29258d94a7a60d1addc9a1b5f095",
      "size": 699956416
    }

there is no blob file corresponding to the sha256 digest under ollama models/blob folder

Hi, I can use this script to export the model that I have already downloaded using for example "ollama pull" right?
This exported model I can then use to fine tune using LoRA?
I understood that I cannot fine tune using ollama and hence have to export the model, fine tune outside, then import it back to Ollama to use it with ollama inference?

(I am totally new to AI models)

Hello! @vasanthnagkv

This exported model I can then use to fine tune using LoRA?

Yes if the model is llama model based. It's related to ggml-org/llama.cpp#6680

import it back to Ollama to use it with ollama inference

The script can be worked with ollama create for importing, it's compatible.
https://github.com/ollama/ollama/blob/main/docs/import.md

Hi!
Thanks for your work!

I ve got this weird trace just before success message : jq: error (at <stdin>:1): Cannot iterate over number (8)

It happens with few models like codellama:34b and nomic-embed-text

I still manage to import these exported models into my offline ollama instance

If you have any idea ?

Am i allowed to add this to my package manager, scriptgrab?

Am i allowed to add this to my package manager, scriptgrab?

Welcome! 🚀

According to SPDX-License-Identifier: MIT (https://ncurl.xyz/s/o_o6DVqIR), it's allowed.

Please leave this line in the description/source code of the script. @Rays-Robotics

@xuegl @wingertjp

Sorry for replying late.
I'll check these problems as soon as possible. 🫡

@xuegl @wingertjp

Sorry for replying late. I'll check these problems as soon as possible. 🫡

Great 👍