Skip to content

Tutorial - LLaVA

LLaVA is a popular multimodal vision/language model that you can run locally on Jetson to answer questions about image prompts and queries. Llava uses the CLIP vision encoder to transform images into the same embedding space as its LLM (which is the same as Llama architecture). Below we cover different methods to run Llava on Jetson, with increasingly optimized performance:

  1. Chat with Llava using text-generation-webui
  2. Run from the terminal with llava.serve.cli
  3. Quantized GGUF models with llama.cpp
  4. Optimized Multimodal Pipeline with NanoVLM
Llava-13B (Jetson AGX Orin) Quantization Tokens/sec Memory
text-generation-webui 4-bit (GPTQ) 2.3 9.7 GB
llava.serve.cli FP16 (None) 4.2 27.7 GB
llama.cpp 4-bit (Q4_K) 10.1 9.2 GB
NanoVLM 4-bit (MLC) 21.1 8.7 GB

In addition to Llava, the NanoVLM pipeline supports VILA and mini vision models that run on Orin Nano as well.

1. Chat with Llava using text-generation-webui

What you need

  1. One of the following Jetson devices:

    Jetson AGX Orin (64GB) Jetson AGX Orin (32GB) Jetson Orin NX (16GB)

  2. Running one of the following versions of JetPack :

    JetPack 5 (L4T r35.x) JetPack 6 (L4T r36.x)

  3. Sufficient storage space (preferably with NVMe SSD).

    • 6.2GB for text-generation-webui container image
    • Space for models
      • CLIP model : 1.7GB
      • Llava-v1.5-13B-GPTQ model : 7.25GB

  4. Clone and setup jetson-containers : 

    git clone https://github.com/dusty-nv/jetson-containers
bash jetson-containers/install.sh

Download Model 

jetson-containers run --workdir=/opt/text-generation-webui $(autotag text-generation-webui) \
  python3 download-model.py --output=/data/models/text-generation-webui \
    TheBloke/llava-v1.5-13B-GPTQ

Start Web UI with Multimodal Extension 

jetson-containers run --workdir=/opt/text-generation-webui $(autotag text-generation-webui) \
  python3 server.py --listen \
    --model-dir /data/models/text-generation-webui \
    --model TheBloke_llava-v1.5-13B-GPTQ \
    --multimodal-pipeline llava-v1.5-13b \
    --loader autogptq \
    --disable_exllama \
    --verbose

Go to Chat tab, drag and drop an image into the Drop Image Here area, and your question in the text area and hit Generate :

Result

2. Run from the terminal with llava.serve.cli

What you need

  1. One of the following Jetson:

    Jetson AGX Orin 64GB Jetson AGX Orin (32GB)

  2. Running one of the following versions of JetPack :

    JetPack 5 (L4T r35.x) JetPack 6 (L4T r36.x)

  3. Sufficient storage space (preferably with NVMe SSD).

    • 6.1GB for llava container
    • 14GB for Llava-7B (or 26GB for Llava-13B)

This example uses the upstream Llava repo to run the original, unquantized Llava models from the command-line. It uses more memory due to using FP16 precision, and is provided mostly as a reference for debugging. See the Llava container readme for more info.

llava-v1.5-7b 

jetson-containers run $(autotag llava) \
  python3 -m llava.serve.cli \
    --model-path liuhaotian/llava-v1.5-7b \
    --image-file /data/images/hoover.jpg

llava-v1.5-13b 

jetson-containers run $(autotag llava) \
  python3 -m llava.serve.cli \
    --model-path liuhaotian/llava-v1.5-13b \
    --image-file /data/images/hoover.jpg

Unquantized 13B may run only on Jetson AGX Orin 64GB due to memory requirements.

3. Quantized GGUF models with llama.cpp

What you need

  1. One of the following Jetson devices:

    Jetson AGX Orin (64GB) Jetson AGX Orin (32GB) Jetson Orin NX (16GB)

  2. Running one of the following versions of JetPack :

    JetPack 5 (L4T r35.x) JetPack 6 (L4T r36.x)

llama.cpp is one of the faster LLM API's, and can apply a variety of quantization methods to Llava to reduce its memory usage and runtime. Despite its name, it uses CUDA. There are pre-quantized versions of Llava-1.5 available in GGUF format for 4-bit and 5-bit: 

jetson-containers run --workdir=/opt/llama.cpp/bin $(autotag llama_cpp:gguf) \
  /bin/bash -c './llava-cli \
    --model $(huggingface-downloader mys/ggml_llava-v1.5-13b/ggml-model-q4_k.gguf) \
    --mmproj $(huggingface-downloader mys/ggml_llava-v1.5-13b/mmproj-model-f16.gguf) \
    --n-gpu-layers 999 \
    --image /data/images/hoover.jpg \
    --prompt "What does the sign say"'
Quantization Bits Response Tokens/sec Memory
Q4_K 4 The sign says "Hoover Dam, Exit 9." 10.17 9.2 GB
Q5_K 5 The sign says "Hoover Dam exit 9." 9.73 10.4 GB

A lower temperature like 0.1 is recommended for better quality ( --temp 0.1 ), and if you omit --prompt it will describe the image: 

jetson-containers run --workdir=/opt/llama.cpp/bin $(autotag llama_cpp:gguf) \
  /bin/bash -c './llava-cli \
    --model $(huggingface-downloader mys/ggml_llava-v1.5-13b/ggml-model-q4_k.gguf) \
    --mmproj $(huggingface-downloader mys/ggml_llava-v1.5-13b/mmproj-model-f16.gguf) \
    --n-gpu-layers 999 \
    --image /data/images/lake.jpg'

In this image, a small wooden pier extends out into a calm lake, surrounded by tall trees and mountains. The pier seems to be the only access point to the lake. The serene scene includes a few boats scattered across the water, with one near the pier and the others further away. The overall atmosphere suggests a peaceful and tranquil setting, perfect for relaxation and enjoying nature.

You can put your own images in the mounted jetson-containers/data directory. The C++ code for llava-cli can be found here . The llama-cpp-python bindings also support Llava , however they are slower from Python (potentially handling of the tokens)

4. Optimized Multimodal Pipeline with NanoVLM

What's Next

This section got too long and was moved to the NanoVLM page - check it out there for performance optimizations, mini VLMs, and live streaming!