My attempt to run Gemma AI model locally

Recently, Google/Deep Mind announced Gemma:

A family of lightweight, state-of-the art open models built from the same research and technology used to create the Gemini models

I was curious if I could ever run these models on a local server. Gemma features both a 7B and a 2B token model, so I thought I'd have a shot on one of my cheap 4-CPU Intel servers that has 16GB of RAM.

The 2B parameter model, at 2 Bytes / token, should be around 4GB.

2 x 1024 x 1024 x 1024 tokens x 2 Bytes / token = 4,294,967,296 Bytes

Downloading the model

I'm following the instructions on the Kaggle page.

I also had to download my Kaggle API token into ~/.kaggle/kaggle.json. This was required in order to run the Keras option below.


The first option is to use the Python keras package, which I've used in the past.

The total space taken by the model is 4.7G.

Python setup:

pyenv install 3.10.10
pyenv virtualenv 3.10.10 gemma-keras
pyenv activate gemma-keras
pip install --upgrade keras-nlp-nightly
# Alternatively, from
# pip install --upgrade keras-nlp
# pip install --upgrade keras>=3

Sample script: I needed to create this script inside the downloaded model directory

import keras_nlp

gemma_lm = keras_nlp.models.GemmaCausalLM.from_preset("gemma_2b_en")
print(gemma_lm.generate("Keras is a", max_length=30))

I needed to create and run this script inside the downloaded model directory, or I would see errors when importing the keras_nlp library in python.

When this script was run the first time, it re-downloaded the 4.7G model into a new directory--super annoying.


Invoking the script:

$ time python
Keras is a popular deep learning library for Python. It is easy to use and has a wide range of features. One of the features of

real	1m38.011s
user	1m52.609s
sys	0m17.086s

So it took 98 seconds to produce 30 words, including the model load time. Ha! I was surprised that it finally worked. Not all the other methods worked out.


I also tried the Flax option, because this page said that it could be run with either a CPU, GPU or TPU.

I downloaded the 2b model, which takes 3.7GB:

$ du -sh
3.7G	.

Python setup:

pyenv install 3.10.10
pyenv virtualenv 3.10.10 gemma
pyenv activate gemma
pip install git+

Invoking the sample script

git clone
cd gemma
python examples/ \
    --path_checkpoint=../model/2b/ \
    --path_tokenizer=../model/tokenizer.model \
    --string_to_sample="Where is Salt Lake City?"

which looked good at first:

Loading the parameters from /home/schaubj/work/jschaub30/gemma/model/2b/
I0223 14:37:24.434225 135111827680128] Restoring item from /home/schaubj/work/jschaub30/gemma/model/2b.
I0223 14:37:34.196148 135111827680128] Unable to initialize backend 'cuda':
I0223 14:37:34.196315 135111827680128] Unable to initialize backend 'rocm': module 'jaxlib.xla_extension' has no attribute 'GpuAllocatorConfig'
I0223 14:37:34.198884 135111827680128] Unable to initialize backend 'tpu': INTERNAL: Failed to open cannot open shared object file: No such file or directory
I0223 14:37:34.299071 135111827680128] Finished restoring checkpoint from /home/schaubj/work/jschaub30/gemma/model/2b.
Parameters loaded.

...but crashed after about 30sec with this helpful error message:


I decided to run my viz_workload profiler to try and uncover any problems.

The results are here.

The memory profile immediately stood out: Gemma profile.

The used memory (in red) peaked at 15.1 right when the workload crashed.

I then checked the dmesg log and saw many memory related errors:

[ 330.616577] Out of memory: Killed process 2646 ( total-vm:24639840kB, anon-rss:15396924kB, file-rss:256kB, shmem-rss:0kB, UID:1000 pgtables:35264kB oom_score_adj:0


The transformers description looks promising:

A transformers implementation of Gemma-2b-instruct. It is a good choice for Python developers that want a quick and easy way to prompt a model, or even begin developing LLM applications. It is a 2B parameter, instruction-tuned LLM.

The downloaded model size was

$ du -sh
15G  .

This is too large--maybe a mistake?

The setup is easy enough:

pyenv virtualenv 3.10.10 gemma-transform
pyenv activate gemma-transform
pip install --upgrade pip
pip install transformers
pip install torch  # not documented!  Took forever

Sample script--I modified it to add the access API token.

from transformers import AutoTokenizer, AutoModelForCausalLM

with open(".token", "r") as fid:
    access_token =

# tokenizer = AutoTokenizer.from_pretrained("google/gemma-2b-it", token=access_token)
tokenizer = AutoTokenizer.from_pretrained("google/", token=access_token)
model = AutoModelForCausalLM.from_pretrained("google/", token=access_token)

input_text = "Write me a poem about Machine Learning."
input_ids = tokenizer(input_text, return_tensors="pt")

outputs = model.generate(**input_ids, max_new_tokens=30)

Invoking the script:

$ time python
Loading checkpoint shards: 100%|█████████████████████████████████████████████████| 2/2 [00:02<00:00,  1.47s/it]
<bos>Write me a poem about Machine Learning.

Machines, they weave and they learn,
From the data, they discern.
Algorithms, a symphony,
Unleash the power of

real	0m28.090s
user	1m12.143s
sys	0m22.906s

So this is much quicker than keras (28 seconds for 18 words).