exllama

exllama

CONTAINERS IMAGES RUN BUILD

This package provides containers for both ExLlama and ExLlamaV2:

• exllama container uses the https://github.com/jllllll/exllama fork of https://github.com/turboderp/exllama (installed under /opt/exllama)
• exllama:v2 container uses https://github.com/turboderp/exllamav2 (installed under /opt/exllamav2)

Both loaders are also supported in the oobabooga text-generation-webui container.

Inference Benchmark

Substitute the GPTQ model from HuggingFace Hub that you want to run (see exllama compatible models)

./run.sh --workdir=/opt/exllama $(./autotag exllama) /bin/bash -c \
  'python3 test_benchmark_inference.py --perf --validate -d $(huggingface-downloader TheBloke/Llama-2-7B-GPTQ)'

If the model repository is private or requires authentication, add --env HUGGINGFACE_TOKEN=<YOUR-ACCESS-TOKEN>

Memory Usage

Model	Memory (MB)
TheBloke/Llama-2-7B-GPTQ	5,200
TheBloke/Llama-2-13B-GPTQ	9,135
TheBloke/LLaMA-30b-GPTQ	20,206
TheBloke/Llama-2-70B-GPTQ	35,462

CONTAINERS

exllama:0.0.15
Aliases	exllama
Requires	L4T ['>=36']
Dependencies	build-essential cuda cudnn python numpy cmake onnx pytorch huggingface_hub
Dependants	text-generation-webui:1.7 text-generation-webui:6a7cd01 text-generation-webui:main
Dockerfile	Dockerfile

exllama:0.0.14
Aliases	exllama
Requires	L4T ['==35.*']
Dependencies	build-essential cuda cudnn python numpy cmake onnx pytorch huggingface_hub
Dockerfile	Dockerfile

CONTAINER IMAGES

Repository/Tag	Date	Arch	Size
dustynv/exllama:r35.2.1	2023-12-15	arm64	5.5GB
dustynv/exllama:r35.3.1	2023-12-11	arm64	5.5GB
dustynv/exllama:r35.4.1	2023-12-14	arm64	5.4GB
dustynv/exllama:v1-r36.2.0	2023-12-15	arm64	7.2GB
dustynv/exllama:v2-r35.2.1	2023-12-15	arm64	5.5GB
dustynv/exllama:v2-r35.3.1	2023-12-14	arm64	5.5GB
dustynv/exllama:v2-r35.4.1	2023-12-12	arm64	5.5GB
dustynv/exllama:v2-r36.2.0	2023-12-15	arm64	7.2GB

_{Container images are compatible with other minor versions of JetPack/L4T:
    • L4T R32.7 containers can run on other versions of L4T R32.7 (JetPack 4.6+)
    • L4T R35.x containers can run on other versions of L4T R35.x (JetPack 5.1+)}

RUN CONTAINER

To start the container, you can use jetson-containers run and autotag, or manually put together a docker run command:

# automatically pull or build a compatible container image
jetson-containers run $(autotag exllama)

# or explicitly specify one of the container images above
jetson-containers run dustynv/exllama:v1-r36.2.0

# or if using 'docker run' (specify image and mounts/ect)
sudo docker run --runtime nvidia -it --rm --network=host dustynv/exllama:v1-r36.2.0

^{jetson-containers run forwards arguments to docker run with some defaults added (like --runtime nvidia, mounts a /data cache, and detects devices)
autotag finds a container image that's compatible with your version of JetPack/L4T - either locally, pulled from a registry, or by building it.}

To mount your own directories into the container, use the -v or --volume flags:

jetson-containers run -v /path/on/host:/path/in/container $(autotag exllama)

To launch the container running a command, as opposed to an interactive shell:

jetson-containers run $(autotag exllama) my_app --abc xyz

You can pass any options to it that you would to docker run, and it'll print out the full command that it constructs before executing it.

BUILD CONTAINER

If you use autotag as shown above, it'll ask to build the container for you if needed. To manually build it, first do the system setup, then run:

jetson-containers build exllama

The dependencies from above will be built into the container, and it'll be tested during. Run it with --help for build options.