ncompass 0.0.14

An API built to enable one-line-of-code access to accelerated open-source and custom AI models.

ncompass

A One-Line-Of-Code (OLOC) API that makes it easy for anyone to access low-latency open-source or custom AI models. This repo has some examples of the various ways in which you can use our API.

Installing

You can install our package via pip with the following command:

pip install ncompass

Getting your API Key

We're working on automating the process for you to select a model of your choice and getting an API key, but for now please email either [e-mail] or [e-mail] to be onboarded and provided with your API key. Thanks!

One-Line-Of-Code (OLOC)

We strive to make sure that all our capabilities can be exposed to you with as close to one line of code as possible. Below is an example of how you would do that with us for streaming prompt complettion. The following lines of code and other examples like it can be found in the examples/hello_world.py file.

from ncompass.client import nCompassOLOC

prompt = 'Give me 5 tools I can use to accelerate inference of my ML model?'
api_key = '<api_key>'
nCompassOLOC().start_session(api_key)
for i in range(3):
    nCompassOLOC().complete_prompt(api_key, prompt, max_tokens = 300, temperature = 0.5
                                   , top_p = 0.9, stream = True, pprint = True)
nCompassOLOC().stop_session(api_key)

The section below unwraps the code a bit more. It is still simple, but not quite OLOC.

More fine grained control

The following code (also found in examples/fine_grained_control.py) is an example of using the nCompass client with more fine grained control. For instance:

• you could replace client.wait_until_model_running() with the non-blocking version client.model_is_running() which just returns a boolean.
• use a custom way to iterate through the response by replacing client.print_prompt. The examples/hello_world.py file has an example (use_streaming_iterator) of extracting the iterator from our OLOC version as well

from ncompass.client import nCompass

# The client will automatically read the API key from the environment variable passed here.
# Alternatively, you can directly pass the api_key to the client (see README.md)
client = nCompass(custom_env_var = 'NCOMPASS_API_KEY')
client.start_session()
client.wait_until_model_running()
params = {'max_tokens':    300 # max output tokens requested
          , 'temperature': 0.5
          , 'top_p':       0.9
          , 'stream':      True}
prompt = 'Give me 5 tools I can use to accelerate inference of my ML model?'
response_iterator = client.complete_prompt(prompt, **params)
client.print_prompt(response_iterator)
client.stop_session()

Note that here we read the API key from a custom environment variable NCOMPASS_API_KEY which you can set with the token we provide you. However, the API also works if you directly provide the API key to the client directly as below.

client = nCompass(api_key = '<key we provide you>')

IMPORTANT: What are sessions?

Sessions are how nCompass keeps track of usage statistics. nCompass billing is not per_token but rather time_based. This means, your costs do not scale linearly with your inputs. We simply charge you for the time between calls to start and stop session.

There are two ways to start and stop sessions:

• Using the command line: When you install the library via pip, you will have 3 command line calls available to you.

  • nccli-help : prints a help msg on how to use the cli commands
  • nccli-start-session <api_key> : starts a session for that api_key
  • nccli-stop-session <api_key> : stops a session for that api_key

• Programatically: In all the examples above, the start_session and stop_session calls perform the tasks of starting and stopping sessions.

Please do not forget to stop sessions you have started as billing occurs between the start and stop of a session.

IMPORTANT: Cold-starts

Our models will automatically shut down after ~2hours of inactivity. The current cold start time after model shutdown is a few minutes, so please raise this as a concern when we onboard you if you think your workload is so sparse that queries are spaced out by more than 2 hours. There are a few mitigation strategies to deal with preventing these cold starts listed below.

CLI commands for starting and stopping models

If your model does get shut down due to lack of activity, use the following commands to launch your model up via our cli tools. These commands enable checking model status and starting and stopping models (not sessions). Starting and stopping models are slow commands and should not be used during runtime.

• nccli-start-model <api_key> : starts a model for that api_key
• nccli-stop-model <api_key> : stops a model for that api_key
• nccli-model-status <api_key> : gets model status for the api_key As they are slow running commands, they are by default not blocking. Which means after running nccli-start-model for instance, you have to keep querying the status using nccli-model-status in order to ensure that your model was started before running normal execution.

You don't really have to worry about nccli-stop-model as we will auto shutdown models after 2 hours of inactivity, but if you know you're not going to be using your model for a while, running nccli-stop-model out of good faith would be much appreciated :)

Booking timeslots so we can ensure you model is up and running

We are working on programatically exposing this, but for now if you know that you need a warm model for a certain time period during the day and definitely don't need traffic at other hours, just simply let us know when we onboard you / email us and we will ensure for those hours you don't have any cold starts.

Note on thread safety

nCompassOLOC client is not thread safe, but the nCompass client is thread safe. So if you would like to use the client inside threads that you are spawning (for eg. when using a wsgi web frameworks like flask), then please only use the nCompass client.

Chat Completions

The above examples focused on simple prompt completions. If you would like to use chat based completions with a custom template, please checkout the following example.

client = nCompass(api_key = <api_key>)
client.start_session()
client.wait_until_model_running()
params = {'max_tokens':    300 # max output tokens requested
          , 'temperature': 0.5
          , 'top_p':       0.9
          , 'stream':      True}
messages = [
    {
        "role": "assistant",
        "content": "Hello, how can I speed up your ML inference today?",
    },
    {
        "role": "user",
        "content": "What tweaks can I make to my model to make it faster?",
    },
    {
        "role": "assistant",
        "content": "Ah, the age old question.",
    },
]
response_iterator = client.complete_chat(messages=messages
                                         , add_generation_prompt=True
                                         , **params)
client.print_prompt(response_iterator)
client.stop_session()

The main changes from prompt completion is that we now use the complete_chat API instead of the complete_prompt API. For convenience, there is the equivalent nCompassOLOC().complete_chat(...) API as well. Also the argument name is no longer prompt, but rather messages. add_generation_prompt is also a boolean that can be passed through.

The chat_template used is the chat_template specified in the model's tokenizer (if HF tokenizer is used). If you have a custom template you would like to pass through, there is support for that that is in testing, so please contact either [e-mail] or [e-mail] for information.

Batch Requests

For both completions and chat, we support batch requests which is enabled by setting the stream argument in params to False. For prompt, the input can either be a single string or a list of strings. For messages, the input can either be a single list or a list of lists.