replicate-javascript VS pytorch-lightning

Replicate (YC W20) | San Francisco, CA + Remote | https://replicate.com/

Replicate makes it easy to run AI in the cloud. You can run a big library of open source models with a few lines of code, or deploy your own models at scale.

We're an experienced team from Spotify, Docker, GitHub, Heroku, Apple, and various other places. We're backed by a16z, Sequoia, Andrej Karpathy, Dylan Field, Guillermo Rauch.

We're hiring:

- An infrastructure engineer

- An expert at deploying and optimizing language models

- An engineer who is good at humans to look after our customers

... and more: https://replicate.com/about#join-us

Email us: [email protected]

import { experimental_buildLlama2Prompt } from 'ai/prompts'; // Now use Replicate LLAMA 70B streaming to perform the autocompletion with context const response = await replicate.predictions.create({ // You must enable streaming. stream: true, // The model must support streaming. See https://replicate.com/docs/streaming model: 'meta/llama-2-70b-chat', // Format the message list into the format expected by Llama 2 // @see https://github.com/vercel/ai/blob/99cf16edf0a09405d15d3867f997c96a8da869c6/packages/core/prompts/huggingface.ts#L53C1-L78C2 input: { prompt: experimental_buildLlama2Prompt([ { // create a system content message to be added as // the llama2prompt generator will supply it as the context with the API role: 'system', content: systemContext }, { // create a system instruction // make sure to wrap code blocks with ``` {% endraw %} so that the svelte markdown picks it up correctly role: 'assistant', content: {% raw %}`When creating repsonses sure to wrap any code blocks that you output as code blocks and not text so that they can be rendered beautifully.`{% endraw %} }, // also, pass the whole conversation! ...messages ]) } }); {% raw %}

We used Replicate to run the models and the cost so far is 26 cents for 90 cards — which means it’s less than a third of a cent per card!

Basically https://replicate.com/

Because it happens when running your own models on localhost too. I have ollama and all the ones they support, but there are some on HuggingFace I run through llama.cpp inside apps where I won't have ollama installed, Replicate also has Stable Diffusion models, not just chat ones, and OpenAI which is its own thing. So it could potentially all be unified under a provider like that.

Haven't actually tried Replicate because I'm just running locally for free, but probably would try to find a single cloud provider for all deployments, like a Heroku of LLMs.

It's very easy to get started, right in your Terminal, no fees! No credit card at all.

And there are cloud providers like https://replicate.com/ and https://lightning.ai/ that will let you use your LLM via an API key just like you did with OpenAI if you need that.

You don't need OpenAI - nobody does.

In this section, we’ll show you how to generate MDE depth map predictions with both DPT and Marigold. In both cases, you can optionally run the model locally with the respective Hugging Face library, or run remotely with Replicate.

Built with Supabase, Astro, Unreal Speech, Stable Diffusion, Replicate, Metropolitan Museum of Art

Now that we understand the process, let’s take a look at the actual code. The first step is simply importing our dependencies. Note that we will be using the replicate npm package, which you can install with npm i replicate.

It's very easy to get started, right in your Terminal, no fees! No credit card at all.

And there are cloud providers like https://replicate.com/ and https://lightning.ai/ that will let you use your LLM via an API key just like you did with OpenAI if you need that.

You don't need OpenAI - nobody does.

https://see.stanford.edu/Course/CS229 https://lightning.ai/ https://www.youtube.com/watch?v=00s9ireCnCw&t=57s https://towardsdatascience.com/

I've been wondering on what is the recommended method of saving logits/activations using PyTorch Lightning. I've looked at Callbacks, Loggers and ModelHooks but none of the use-cases seem to be for this kind of activity (even if I were to create my own custom variants of each utility). The ModelCheckpoint Callback in its utility makes me feel like custom Callbacks would be the way to go but I'm not quite sure. This closed GitHub issue does address my issue to some extent.

Our codebase for the diffusion models builds heavily on OpenAI's ADM codebase , lucidrains, Stable Diffusion, Lightning and Hugging Face. Thanks for open-sourcing!

For lightweight use, pytorch-lightning is too heavy, and its source code will be very difficult for beginners to read, at least for me.

For the A100 GPU, theoretical performance is the same for FP16/BF16 and both rely on the same number of bits, meaning memory should be the same. However since it's quite newly added to PyTorch, performance seems to still be dependent on underlying operators used (pytorch lightning debugging in progress here).