fjcloud / llamapod

Llamapod

Drop easily your LLM API on an OpenShift Cluster

Getting Started

oc new-project llm-api
oc new-app https://github.com/fjcloud/llamapod
oc expose svc/llamapod

Play with it

generate_news.sh

In this example you will run one time a prompt for getting a creative news title and again another less creative, more limited prompt for getting tags in json

./generate_news.sh

Expected output

{
  "tags": [
    "technology",
    "electricity",
    "toasting"
  ],
  "title": "Tesla Model 3: It's not just a car, but also the world's best electric toaster"
}

webserver.sh

In this example we will emulate a web server using llm, a good way for creating a honeypot

nc -l -p 3000 -k -e ./webserver.sh

Expected response from netcat :

curl localhost:3000/api/ping -v -H "Host: hello.llm"
{
    "message": "Pong!"
}
curl localhost:3000/api/bye -H "Host: hello.llm"
{
    "message": "Bye!"
}
curl -X POST localhost:3000/api/v1
{
    "message": "success"
}

Benchmark

oc exec deploy/llamapod -- /bin/sh ./TinyLlama-1.1B-Chat-v1.0.F16.llamafile -f getty.txt -t 0 -n 22

Example output on ARM Ampere Altra 16 cores :

...
llama_print_timings:        load time =      83.73 ms
llama_print_timings:      sample time =       1.06 ms /    22 runs   (    0.05 ms per token, 20793.95 tokens per second)
llama_print_timings: prompt eval time =    1063.97 ms /   217 tokens (    4.90 ms per token,   203.95 tokens per second)
llama_print_timings:        eval time =    1146.98 ms /    21 runs   (   54.62 ms per token,    18.31 tokens per second)
llama_print_timings:       total time =    2219.06 ms /   238 tokens

Uninstall

oc delete all -l app=llamapod

That's all for the moment!