Oscar Rovira

Exploring.
-
Prev. co-founder @ Mystic AI, YCW21

[Damien C. Tanner @dctanner]

We're brining a slice of @swyx's AI Engineer World's Fair to London!

Evening of 12 September is the first AI Engineer London Meetup.

Hear from 4 amazing speakers: @maximelabonne, @roviosc, @BruverisMartins and Chris Bull.

See you there! Registration link below 👇

[Oscar Rovira @roviosc]

@mysticdotai @bfl_ml

[Oscar Rovira @roviosc]

You can now access the Flux models commercially on @mysticdotai
Run here: https://t.co/BNAYhwnTw4

Excited to announce our new official partnership with @bfl_ml, creators of Flux; The new state-of-the-art image generation model, also used in Grok to generate images.

[Oscar Rovira @roviosc]

We’re one of only three official API providers! 🎉 And the only provider that charges for inference time, so better pricing for varying quality/speed.

The models included are:

- Flux Pro (Highest quality but slower)
- Flux Dev (Best quality/performance)
- Flux Schnell (Good quality & fast, specially with low steps)

[Oscar Rovira @roviosc]

Tutorial here: https://t.co/QvzsvP6NKA

[Oscar Rovira @roviosc]

You can deploy your own ComfyUI workflows as an API on @mysticdotai

This is a generated video of my dream home, made with ComfyUI and Mystic using Flux, image upsampler and the video generated with SDV. All open-weights!

[Oscar Rovira @roviosc]

✅ We optimised this at Mystic AI with,

Running your workflow on spot GPUs (Up to 10x cheaper than on-demand GPUs)
Our Rust-based Turbo Registry, lowering cold-start by 14x, from minutes to seconds!

[Oscar Rovira @roviosc]

The problem comes at scale. Say you’ve build an awesome workflow and you want 100s of users to access it too, now you have to figure out how to scale this ComfyUI workflow.

[Mystic @mysticdotai]

🔥The Mystic Turbo Registry, is now live on ProductHunt: https://t.co/ZMaGNV7gJ6

[Oscar Rovira @roviosc]

@mysticdotai This means that if you deploy your ML model with Mystic, going from 0 GPU to 1 would take a few seconds instead of close to 1 minute. That is 1 minute less your customers have to wait until they get a response back from your AI product.

[Oscar Rovira @roviosc]

Every time I look at this video I’m amazed at the work @mysticdotai team keeps doing to push the boundaries of ML infrastructure. I'm talking about our new Turbo Registry - Our optimized Rust-based Docker registry.

[Oscar Rovira @roviosc]

@mysticdotai This is a 13GB Docker Image (i.e., SDXL and all libraries and packages required to run) loaded into a new GPU in 12s! 💥 (Instead of the standard 54s 💩). The best news is that it scales nicely, with improvements of up to 15x for larger Docker images.

[Oscar Rovira @roviosc]

@mysticdotai

[Oscar Rovira @roviosc]

As you know I'm always on the lookout for the best library to run open-source LLMs, well let me tell you my new not-secret, exllamav2. Honestly, 10/10.

Got Mixtral-8x7B up and running in a single A100 GPU with 40GB while achieving an avg. latency of 45 tok/s, (single prompt, no batch and without any sort of speculative sampling) and by using Mystic AI BYOC (more on this soon!) I can instantly deploy it using spot instances while running it on my own secure cloud. Will be writing a tutorial on this shortly.

If you wanna give the model a go on our serverless offering or are interested in instantly deploying this on your own cloud, I’ll add links in comments. Follow for more!

[Oscar Rovira @roviosc]

Exllamav2: https://t.co/PcFWOQJqpr
Serverless API: https://t.co/avmCPbrdew
Deploy on your own cloud: https://t.co/pHbLsJtNJy

[Mystic @mysticdotai]

With its unique sparse mixture-of-experts network, it offers unparalleled efficiency and multilingual capabilities.

Read the blog here: https://t.co/VdaI9wmToN

[Oscar Rovira @roviosc]

Google's new music model -> MusicLM: Generating music from text.

Excited to see restaurants adding a custom sound experience for a particular dish. Or background music for MMORPG generated on the fly.

https://t.co/JBTSkxX43C

[Mystic @mysticdotai]

#StableDiffusion2 API is now available at https://t.co/vZTmew5YLU!
1️⃣ Create an account
2️⃣ Start using Stable Diffusion 2 in production

👀 Got your own model? PipelineAI takes care of your production infrastructure with our serverless GPU inference system.