Modal is a cloud deployment system that’s entirely programmatic. No yaml:

import modal

stub = modal.stub(gpu="a100")
@stub.function()
def fit(x):
    import whatever
    whatever.thing()

So think run.house, but they have the infra.

fine-tuning with Modal

https://github.com/modal-labs/llama-recipes

You can store the serverless functions, and Modal can serve stored serverless functions. Modal have web hooks as well to do inference at a front end.

Modal can serve most the management as well.

pricing

13B: 500 tokens/s on 40GB AA10 (3.73 / hour) 70B: 300 tok /s 80 GB( 2* 5.59/hour)

modalization is the

Goal: we need to find a model that is “expressive enough”: we need to have enough parameters to help match the shape of the data we collect. to help match the shape of the data we collect.

constituents

requirements

additional information

selecting parameters

see model fitting

increasing expressiveness

mixure model

We could mix distributions into a . See Gaussian mixture model.

transforming distributions

Suppose you start with:

\begin{equation} Z \sim \mathcal{N}\qty(0,1) \end{equation}

we can sample \(k\) points \(k \sim Z\), and then transform them across a function \(x_{j}=f(k_{j})\). We now want to know the destruction of \(x_{j}\). Turns out, if \(f\) is invertible and differential, we have:

see Maximum Likelihood Parameter Learning