Skip to content

Latest commit

 

History

History
263 lines (227 loc) · 9.49 KB

export-guide.md

File metadata and controls

263 lines (227 loc) · 9.49 KB
Raw

Automation and Reproducibility

Overview

This guide hopefully makes it easier to reproduce and export 3d visualizations and animations on your own hardware.

It also assumes you have already set these up:

Datasets

How do I Convert an Existing AI Model File to a HuggingFace SafeTensors Format?

For now the only supported model file is a HuggingFace safetensors file (a GPTQ model file). Below are one-time options to download example model.safetensors file(s). HuggingFace can also convert model bin files to safetensors with these supported options:

GPT2 model.safetensors

export MODEL=./gpt2.model.safetensors
wget https://huggingface.co/gpt2/resolve/main/model.safetensors -O "${MODEL}"

Llama 2 7B Chat GPTQ model.safetensors

export MODEL=./llama-2-7b-chat.model.safetensors
wget https://huggingface.co/TheBloke/Llama-2-7B-Chat-GPTQ/resolve/main/model.safetensors -O "${MODEL}"

Mistral 7B OpenOrca GPTQ model.safetensors

export MODEL=./mistral-7b-openorca.model.safetensors
wget https://huggingface.co/TheBloke/Mistral-7B-OpenOrca-GPTQ/resolve/main/model.safetensors -O "${MODEL}"

Dolphin 2.5 Mixtral - 8x7B GPTQ model.safetensors

export MODEL=./dolphin-2.5-mixtral-8x7b.model.safetensors
wget https://huggingface.co/TheBloke/dolphin-2.5-mixtral-8x7b-GPTQ/resolve/main/model.safetensors?download=true -O "${MODEL}"

Phind 34B v2 GPTQ model.safetensors

export MODEL=./phind-34b-v2.model.safetensors
wget https://huggingface.co/TheBloke/Phind-CodeLlama-34B-v2-GPTQ/resolve/main/model.safetensors?download=true -O "${MODEL}"

DeepSeek Coder 33B Instruct GPTQ model.safetensors

export MODEL=./deepseek-coder-33B-instruct.model.safetensors
wget https://huggingface.co/TheBloke/deepseek-coder-33B-instruct-GPTQ/resolve/main/model.safetensors?download=true -O "${MODEL}"

Visualizations, Animations and Exporting to STL and glTF Binary (glb) Files

Run Blender Automation and Export to an Animated GIF

Note: gif creation is more resource intensive than just exporting to STL and glTF. Please be patient.

# pick a model safetensors file
# name="gpt2"
# name="llama-2-7b-chat"
name="mistral-7b-openorca"
echo "start: ${name}"
export ROWS=256
export COLS=256
export FACES=10000
# how many layers do you want to extract
export MAX_LAYERS=50
# automatic shutdown once done
export SHUTDOWN_ENABLED=1
export FRAMES=50
mkdir -p "/tmp/blender/${name}"
export MODEL="./${name}.model.safetensors"
export GIF="/tmp/blender/${name}/${name}-${ROWS}x${COLS}"
unset GLTF STL
rm -rf .tmp/*; time blender -p 400 400 800 800 --python ./examples/show-model.py

Run Blender Automation and Export to STL and glTF Binary

Run this following code to generate your own artifacts from a list of models using bash into the output_dir directory.

ls -lrth blender/*
blender/gpt2:
total 473M
-rw-r--r-- 1 matlok matlok 210M Dec 15 08:26 gpt2-256x256.stl
-rw-r--r-- 1 matlok matlok 264M Dec 15 08:26 gpt2-256x256-gltf.glb

blender/llama-2-7b-chat:
total 210M
-rw-r--r-- 1 matlok matlok  96M Dec 15 08:27 llama-2-7b-chat-256x256.stl
-rw-r--r-- 1 matlok matlok 115M Dec 15 08:27 llama-2-7b-chat-256x256-gltf.glb

blender/mistral-7b-openorca:
total 219M
-rw-r--r-- 1 matlok matlok 101M Dec 15 08:27 mistral-7b-openorca-256x256.stl
-rw-r--r-- 1 matlok matlok 118M Dec 15 08:27 mistral-7b-openorca-256x256-gltf.glb

Once you generate the artifacts, you can view each visualization with the Viewing AI Models with a Blender Container Image.

Analyze and Export 128x128 Five Layer Mesh Demos with Configurable Extracted Shapes

profile_models="gpt2 llama-2-7b-chat mistral-7b-openorca dolphin-2.5-mixtral-8x7b phind-34b-v2 deepseek-coder-33B-instruct"
dims="128"
layers="5"
echo "profiling: ${profile_models} dimensions: ${dims} layers: ${layers}"
# automatic shutdown once done
export SHUTDOWN_ENABLED=1
# gif creation is more resource intense
unset GIF
label=""
# where to store the artifacts
output_dir="./blender"
if [[ ! -e "${output_dir}" ]]; then
    mkdir "${output_dir}"
fi
echo "processing models: ${profile_models}"
for use_layer in ${layers}; do
    export MAX_LAYERS="${use_layer}"
    for use_dim in ${dims}; do
        export ROWS="${use_dim}"
        export COLS="${use_dim}"
        if [[ "${use_dim}" == "128" ]]; then
            export FACES=200000
        elif [[ "${use_dim}" == "256" ]]; then
            export FACES=100000
        elif [[ "${use_dim}" == "512" ]]; then
            export FACES=20000
        elif [[ "${use_dim}" == "1024" ]]; then
            export FACES=10000
        fi
        label="dim_${use_dim}_shapes_${FACES}_layers_${MAX_LAYERS}"
        echo "processing ${use_dim} - ${label}"
        for name in ${profile_models}; do
            echo "starting: ${name} - ${label}"
            mkdir -p "${output_dir}/${name}"
            export MODEL="./${name}.model.safetensors"
            # GLTF supports colors
            export GLTF="${output_dir}/${name}/${name}-${label}-gltf"
            export GLTF="${output_dir}/demo-${name}-${label}-gltf"
            # STL does not support colors
            export STL="${output_dir}/${name}-${label}.stl"
            export STL="${output_dir}/demo-${name}-${label}.stl"
            # clear any previous temp images
            rm -rf .tmp/*; time blender -p 400 400 800 800 --python ./examples/show-model.py
            echo "done: ${name} - ${label}"
        done
    done
done

Analyze and Export Dense 512x512 Two Layer Mesh Demos (targeting 500k shapes extracted per model layer)

profile_models="gpt2 llama-2-7b-chat mistral-7b-openorca dolphin-2.5-mixtral-8x7b phind-34b-v2 deepseek-coder-33B-instruct"
layers="2"
dims="512"
echo "profiling: ${profile_models} dimensions: ${dims} layers: ${layers}"
# automatic shutdown once done
export SHUTDOWN_ENABLED=1
# gif creation is more resource intense
unset GIF
label=""
# where to store the artifacts
output_dir="./blender"
if [[ ! -e "${output_dir}" ]]; then
    mkdir "${output_dir}"
fi
echo "processing models: ${profile_models}"
for use_layer in ${layers}; do
    export MAX_LAYERS="${use_layer}"
    for use_dim in ${dims}; do
        export ROWS="${use_dim}"
        export COLS="${use_dim}"
        export FACES=500000
        label="dim_${use_dim}_shapes_${FACES}_layers_${MAX_LAYERS}"
        echo "processing ${use_dim} - ${label}"
        for name in ${profile_models}; do
            echo "starting: ${name} - ${label}"
            mkdir -p "${output_dir}/${name}"
            export MODEL="./${name}.model.safetensors"
            # GLTF supports colors
            export GLTF="${output_dir}/${name}/${name}-${label}-gltf"
            export GLTF="${output_dir}/demo-${name}-${label}-gltf"
            # STL does not support colors
            export STL="${output_dir}/${name}-${label}.stl"
            export STL="${output_dir}/demo-${name}-${label}.stl"
            # clear any previous temp images
            rm -rf .tmp/*; time blender -p 400 400 800 800 --python ./examples/show-model.py
            echo "done: ${name} - ${label}"
        done
    done
done
unset MAX_LAYERS
for use_dim in ${dims}; do
    export ROWS="${use_dim}"
    export COLS="${use_dim}"
    export FACES=200000
    label="dim_${use_dim}_shapes_${FACES}_layers_all"
    echo "processing ${use_dim} - ${label}"
    for name in ${profile_models}; do
        echo "starting: ${name} - ${label}"
        mkdir -p "${output_dir}/${name}"
        export MODEL="./${name}.model.safetensors"
        # GLTF supports colors
        export GLTF="${output_dir}/${name}/${name}-${label}-gltf"
        # STL does not support colors
        export STL="${output_dir}/${name}/${name}-${label}.stl"
        # clear any previous temp images
        rm -rf .tmp/*; time blender -p 400 400 800 800 --python ./examples/show-model.py
        echo "done: ${name} - ${label}"
    done
done

Analyze and Export 256x256 50 Layer Mesh Demos (targeting 10k shapes extracted per model layer)

profile_models="gpt2 llama-2-7b-chat mistral-7b-openorca"
echo "profiling: ${profile_models}"
export ROWS=256
export COLS=256
export FACES=10000
# how many layers do you want to extract
export MAX_LAYERS=50
# automatic shutdown once done
export SHUTDOWN_ENABLED=1
# gif creation is more resource intense
unset GIF
# where to store the artifacts
output_dir="./blender"
if [[ ! -e "${output_dir}" ]]; then
    mkdir "${output_dir}"
fi
echo "processing models: ${profile_models}"
for name in ${profile_models}; do
    echo "starting: ${name}"
    mkdir -p "${output_dir}/${name}"
    export MODEL="./${name}.model.safetensors"
    # GLTF supports colors
    export GLTF="${output_dir}/${name}/${name}-${ROWS}x${COLS}-gltf"
    # STL does not support colors
    export STL="${output_dir}/${name}/${name}-${ROWS}x${COLS}.stl"
    # clear any previous temp images
    rm -rf .tmp/*; time blender -p 400 400 800 800 --python ./examples/show-model.py
    echo "done: ${name}"
done