- AWQ from https://github.com/mit-han-lab/llm-awq (installed under
/opt/awq
) - AWQ's CUDA kernels require a GPU with
sm_75
or newer (so for Jetson, Orin only)
Follow the instructions from https://github.com/mit-han-lab/llm-awq#usage to quantize your model of choice. Or use awq/quantize.py
./run.sh --env HUGGINGFACE_TOKEN=<YOUR-ACCESS-TOKEN> $(./autotag awq) /bin/bash -c \
'/opt/awq/quantize.py --model=$(huggingface-downloader meta-llama/Llama-2-7b-hf) \
--output=/data/models/awq/Llama-2-7b'
If you downloaded a model from the AWQ Model Zoo that already has the AWQ search results applied, you can load that with --load_awq
and skip the search step (which can take a while and use lots of memory)
./run.sh --env HUGGINGFACE_TOKEN=<YOUR-ACCESS-TOKEN> $(./autotag awq) /bin/bash -c \
'/opt/awq/quantize.py --model=$(huggingface-downloader meta-llama/Llama-2-7b-hf) \
--output=/data/models/awq/Llama-2-7b \
--load_awq=/data/models/awq/Llama-2-7b/llama-2-7b-w4-g128.pt'
This process will save the model with the real quantized weights (to a file like $OUTPUT/w4-g128-awq.pt
)
You can use the awq/benchmark.py
tool to gather performance and memory measurements:
./run.sh --env HUGGINGFACE_TOKEN=<YOUR-ACCESS-TOKEN> $(./autotag awq) /bin/bash -c \
'/opt/awq/benchmark.py --model=$(huggingface-downloader meta-llama/Llama-2-7b-hf) \
--quant=/data/models/awq/Llama-2-7b/w4-g128-awq.pt'
Make sure that you load the output from the quantization steps above with --quant
(use the model that ends with -awq.pt
)
CONTAINERS
awq:0.1.0 |
|
---|---|
Aliases | awq |
Requires | L4T ['>=34.1.0'] |
Dependencies | build-essential cuda cudnn python numpy cmake onnx pytorch:2.2 torchvision huggingface_hub rust transformers |
Dockerfile | Dockerfile |
CONTAINER IMAGES
Repository/Tag | Date | Arch | Size |
---|---|---|---|
dustynv/awq:r35.2.1 |
2023-12-14 |
arm64 |
6.1GB |
dustynv/awq:r35.3.1 |
2023-12-15 |
arm64 |
6.1GB |
dustynv/awq:r35.4.1 |
2023-12-12 |
arm64 |
6.1GB |
dustynv/awq:r36.2.0 |
2023-12-15 |
arm64 |
7.8GB |
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 awq)
# or explicitly specify one of the container images above
jetson-containers run dustynv/awq:r35.3.1
# or if using 'docker run' (specify image and mounts/ect)
sudo docker run --runtime nvidia -it --rm --network=host dustynv/awq:r35.3.1
jetson-containers run
forwards arguments todocker 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 awq)
To launch the container running a command, as opposed to an interactive shell:
jetson-containers run $(autotag awq) 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 awq
The dependencies from above will be built into the container, and it'll be tested during. Run it with --help
for build options.