Pruna AI

Every single day, we hear from ML Engineers comparing Pruna AI to torch.compile and TensorRT. So, we decided to help them save time. We tried to be as neutral as possible. Nils Fleischmann decided to take it from scratch, putting himself in the situation of an engineer starting a new project, without prior configuration. And since we’re obsessed with image generation optimization, we ran this using Flux Dev from Black Forest Labs on NVIDIA hardware. 🔥 Pruna delivers SOTA 2.69× speedup (more than TensorRT and torch.compile)… ⏳ Pruna provides easy setup in <5min and instant compression <1min (significantly easier that TensorRT). … When it comes to tech debt and productivity, there’s no match! 🔗 You can check the details of the study here: https://lnkd.in/gePrYj8C

💜 𝗣𝗿𝘂𝗻𝗮 𝗶𝘀 𝗚𝗼𝗶𝗻𝗴 𝗢𝗽𝗲𝗻 𝗦𝗼𝘂𝗿𝗰𝗲! 💜 Our vision has always been to make AI 𝗮𝗰𝗰𝗲𝘀𝘀𝗶𝗯𝗹𝗲 𝗮𝗻𝗱 𝘀𝘂𝘀𝘁𝗮𝗶𝗻𝗮𝗯𝗹𝗲 Today, we’re taking a big step toward that vision. 🚀 𝗢𝗻 𝗠𝗮𝗿𝗰𝗵 𝟮𝟬𝘁𝗵, 𝗣𝗿𝘂𝗻𝗮 𝗴𝗼𝗲𝘀 𝗢𝗽𝗲𝗻 𝗦𝗼𝘂𝗿𝗰𝗲! To celebrate, we’re hosting 𝗟𝗮𝘂𝗻𝗰𝗵 𝗘𝘃𝗲𝗻𝘁𝘀 in 𝗠𝘂𝗻𝗶𝗰𝗵 & 𝗣𝗮𝗿𝗶𝘀! 🎉 🔹 Connect with ML engineers & AI researchers 🔹 See 𝗣𝗿𝘂𝗻𝗮 𝗢𝗽𝗲𝗻 𝗦𝗼𝘂𝗿𝗰𝗲 in action with live demos 🔹 Join discussions on AI efficiency & optimization 🎟️ 𝗦𝗮𝘃𝗲 𝘆𝗼𝘂𝗿 𝘀𝗽𝗼𝘁 𝗻𝗼𝘄: https://www.pruna.ai/party Let’s build the future of AI—𝗳𝗮𝘀𝘁𝗲𝗿, 𝘀𝗺𝗮𝗿𝘁𝗲𝗿, 𝗮𝗻𝗱 𝗼𝗽𝗲𝗻. See you there! ✨ #MachineLearning #OpenSource #AI #PrunaAI

A warm welcome to Nils Fleischmann, our new 𝗠𝗮𝗰𝗵𝗶𝗻𝗲 𝗟𝗲𝗮𝗿𝗻𝗶𝗻𝗴 𝗥𝗲𝘀𝗲𝗮𝗿𝗰𝗵 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿 at Pruna AI! 🎉 Nils has a background in math and recently completed his Master’s in Mathematics and Data Science, where he worked on applying generative models to aid particle physics research. When he isn’t coding, you can find him on the tennis court practicing his serve or hitting the pavement to train for his first marathon. Rumor has it he can’t wait to show off his racket skills in our legendary Pruna paddle matches once summer rolls around. Welcome aboard, Nils! 🤗

🚀 𝟭𝟬,𝟬𝟬𝟬 𝗠𝗼𝗱𝗲𝗹𝘀 𝗦𝗺𝗮𝘀𝗵𝗲𝗱 𝗼𝗻 Hugging Face 🤗 (+2,500 in just the last two months 📈 all of you are keeping us busy!) To celebrate, we gave our HF space a little makeover… and snuck in an Easter Egg 🥚 👀 Think you can find it? Drop a comment if you do!

ML Engineers: When should optimization happen in an ML pipeline? Are you team “𝗮𝘁 𝘁𝗵𝗲 𝗲𝗻𝗱 𝗼𝗳 𝗱𝗲𝘃𝗲𝗹𝗼𝗽𝗺𝗲𝗻𝘁”… optimizing only after the model has been trained and validated but still in dev. Or… Team “𝗮𝘁 𝘁𝗵𝗲 𝗯𝗲𝗴𝗶𝗻𝗻𝗶𝗻𝗴 𝗼𝗳 𝗽𝗿𝗼𝗱𝘂𝗰𝘁𝗶𝗼𝗻”… because “real-world” efficiency gains depend on production context? Drop by and share your take!

We’re thrilled to welcome Gabriel Trégoat, our new 𝗦𝗼𝗳𝘁𝘄𝗮𝗿𝗲 𝗟𝗲𝗮𝗱, to the team! 🎉 Gabriel specializes in creating and deploying impactful AI assets. His journey began in research, linking brainwaves to music, following his double degree in business and engineering. At Shell Energy and Ekimetrics, he deployed thousands of models and created AI products that made a significant impact both internally and for his customers. As the former AI in Production Lead at Ekimetrics, he ensured state-of-the-art standards in every deployment. Outside of his work in ML, Gabriel is a passionate musician and a sports addict. We’re excited to see how his unique perspective and skills will shape the future of AI at Pruna AI! Welcome, Gabriel Trégoat! 🤗

Using NVIDIA TritonServer? We put together a quick guide on getting Pruna running with TritonServer in 5 steps: 1. Preparing the Environment 2. Build the Triton + Pruna Docker Image 3. Configure the Model for Triton 4. Run the Triton Server 5. Run the Client Script There’s also a Dockerfile that sets everything up: Triton, Pruna with full GPU support, and extra dependencies for Stable Diffusion. We’re live on Discord next Tuesday if you wanna chat about serving platforms!

𝘁𝗼𝗿𝗰𝗵.𝗰𝗼𝗺𝗽𝗶𝗹𝗲 is great, but sometimes the compilation just goes brrrrrrr... 🤯 Ever waited minutes for your model to compile, only to get barely any inference speed-up? Toongether by Kartoon asked us to take a look. And here is what we got in… 48 hours: • 𝟭𝟮× 𝗳𝗮𝘀𝘁𝗲𝗿 𝗰𝗼𝗺𝗽𝗶𝗹𝗮𝘁𝗶𝗼𝗻 • 𝟰𝟬% 𝗳𝗮𝘀𝘁𝗲𝗿 𝗶𝗻𝗳𝗲𝗿𝗲𝗻𝗰𝗲 • No more warm-up bottlenecks If you’re curious about what went down, check out the full breakdown here: https://lnkd.in/ez8RM9H2 And if you’re hitting similar performance walls, let’s chat. Happy to help you unlock the same kind of speed-up.

Pruna AI + Replicate = 🔥🔥🔥 Replicate users, we’re here to make your models 2-5x faster & cheaper! How? • If you’re all about the details, check out this step-by-step guide: https://lnkd.in/d-SP2fyF • If you just wanna see it in action, we made this demo-only public space for you to try it out: https://lnkd.in/dmEVCbiy Not in the mood to set things up yourself? Ping Rayan Nait Mazi and we can run some tests and send you back optimized 𝗰𝗼𝗴.𝘆𝗮𝗺𝗹 & 𝗽𝗿𝗲𝗱𝗶𝗰𝘁.𝗽𝘆 files, ready to plug into your Replicate account.

📄 𝗣𝗼𝘀𝗶𝘁𝗶𝗼𝗻 𝗣𝗮𝗽𝗲𝗿: 𝗞𝗲𝘆 𝗰𝗵𝗮𝗹𝗹𝗲𝗻𝗴𝗲𝘀 𝗳𝗼𝗿 𝘁𝗵𝗲 𝗲𝗻𝘃𝗶𝗿𝗼𝗻𝗺𝗲𝗻𝘁𝗮𝗹 𝗽𝗲𝗿𝗳𝗼𝗿𝗺𝗮𝗻𝗰𝗲 𝗼𝗳 𝗔𝗜 Joining a coalition for sustainable AI is great, taking action is even better! Over the past few weeks, we’ve actively participated in consultation sessions to highlight the key challenges that need to be tackled to improve AI’s environmental impact. Following our official membership in the 𝗖𝗼𝗮𝗹𝗶𝘁𝗶𝗼𝗻 𝗳𝗼𝗿 𝗦𝘂𝘀𝘁𝗮𝗶𝗻𝗮𝗯𝗹𝗲 AI at the AI Action Summit, we’re excited to share the position paper: a collective effort from 𝟳𝟬 𝗲𝘅𝗽𝗲𝗿𝘁𝘀 to establish strong 𝗴𝘂𝗶𝗱𝗲𝗹𝗶𝗻𝗲𝘀 𝗳𝗼𝗿 𝘀𝗵𝗮𝗽𝗶𝗻𝗴 𝗽𝘂𝗯𝗹𝗶𝗰 𝗽𝗼𝗹𝗶𝗰𝗶𝗲𝘀 𝗮𝗻𝗱 𝗶𝗻𝗱𝘂𝘀𝘁𝗿𝗶𝗮𝗹 𝘀𝘁𝗿𝗮𝘁𝗲𝗴𝗶𝗲𝘀. Kudos to Inria and the French Ministry of Ecological Transition (Gouvernement) for leading this initiative! 🔗 Link to the report: https://lnkd.in/emhsNbia Grégory Lebourg, Caroline Marcouyoux, Ludovic ARGA, Jean-Philippe Bourgoin, Franck Coisnon, Claire Dorville,PhD, Sabrina STANISLAS-BOUMIER… and many more!