Ionut-Vlad Modoranu

If you’re at the conference and want to geek out over LLM optimizers, send me a DM. I'll be around and happy to meet. I’m also nearing the end of my PhD and actively exploring Research Scientist opportunities. If you're building a great team, let's chat! [3/3]

Training LLMs just got faster I'm at @icmlconf in Seoul this week presenting DASH - our latest improvement to the well-known Shampoo optimizer. By using DASH, we're achieving 5x lower running time and higher accuracy than the standard Distributed Shampoo implementation! [1/3]

Poster details: 🗓️ Thursday, July 9th ⏱️ 2:30PM - 4:15PM 📍 Hall A, Session 7, Poster 405 [2/3]

We're releasing DASH (Distributed Accelerated Shampoo), an improved implementation of the Shampoo optimizer that achieves up to 4.83× faster optimizer steps, while matching or improving final model quality. [1/6]

@OncelTuzel Hi, @OncelTuzel! Thank you for posting this announcement! I have just applied to this job as it fits my PhD expertise.

Come work with us! The Machine Learning Research (MLR) team at Apple is seeking a passionate AI researcher to work on Efficient ML algorithms, including models optimized for fast inference and efficient training methods. Apply here: jobs.apple.com/en-us/details/…

@calcsam BOOK

icymi we wrote a new agents book: patterns for building ai agents it has everything you need to take your agents from prototype to production, like agent design patterns, the basics of security, etc reply to this tweet with BOOK and we'll dm you so you can get a copy

@asmah2107 Hi! I'm currently a PhD student working in optimizers for deep learning. I'm focused on decreasing the memory usage and speeding up the runtime.

Twitter is cool. But it’s 10x better when you connect with people who like building and scaling GenAI systems. If you’re into LLMs, GenAI, Distributed Systems or backend. say hi.

@iamgrigorev This is a great tip for productivity! Please check out our GridSearcher project, it is completely written in Python and allows you to run jobs in parallel or sequentially using dictionaries for hyper-parameters by employing a basic scheduling: github.com/IST-DASLab/Gri…

one interesting trick that i constantly use when running experiments: 1. I create bash script with a command to launch an experiment 2. while it's running, I do some data work, implement another feature in training framework etc 3. I then add new line to the existing bash script that is currently running with new experiment launch 4. when current experiment finishes the new one launches automatically 5. you could continue on and on and play catch-up with your runs. You try to setup a new experiment before the current one finishes. (wonder if there's a better way - like an ongoing queue) P.S. There's an alternative -- sometimes I open a new pane in tmux and write sleep X && bash Y but you need to calculate time

Our QuEST paper was selected for Oral Presentation at ICLR @sparseLLMs workshop! QuEST is the first algorithm with Pareto-optimal LLM training for 4bit weights/activations, and can even train accurate 1-bit LLMs. Paper: arxiv.org/abs/2502.05003 Code: github.com/IST-DASLab/QuE…

We're releasing the DASLab GGUF Quantization Toolkit! 🚀 First open-source toolkit bringing GPTQ + EvoPress to @ggerganov's GGUF format, enabling heterogeneous quantization based on importance. Result: Better models at the same file size. [1/5]

🚀 We are releasing state-of-the-art post-training quantization (PTQ) algorithms for Microscaling FP4, together with kernels: - First study focused on MXFP4/NVFP4 PTQ for LLMs - New Micro-Rotated (MR) format and GPTQ algorithm - QuTLASS GPU kernels with up to 3.6x speedups.

@KwangjunA This seems to be similar to our recently shared work, where we use Discrete Cosine Transform (DCT) to perform a cheap low-rank projection of the momentum buffer, followed by Newton-Schulz orthogonalization. Check out our paper here: arxiv.org/abs/2505.17967

New improvement in Dion leads to a speedup that makes orthonormal updates (eg. Muon) more scalable for larger matrices. The trick: carefully using Newton-Schulz (on smaller matrices) as Dion's backend. Updates to our microsoft/dion codebase are coming soon---stay tuned!

Paper: arxiv.org/abs/2509.23202 We release vLLM & HuggingFace integrations. Code: github.com/IST-DASLab/FP-… Kernels: github.com/IST-DASLab/qut… Credit goes to the team: @RobertoL_Castro, Vage, Denis, @black_samorez and @AshkboosSaleh, with support from @RedHat_AI and @thoefler

@OptionsBuffett Trade

I’m officially restarting the $100 to $10k Account Challenge for 2025 Tomorrow 🚨 This challenge is going to retire dozens and will change lives 💰 Like & comment “Trade” to join! ❤️ (Notifications must be on 🔔)

@deedydas DeepSeek gets it right:

Tried this 5 times on ChatGPT 4.5 and it gets it wrong every time.

Introducing Panza, a personalized LLM email assistant, running entirely on-device! [1/6] * Panza adapts LLaMA-3-8B to match your unique writing style; * Can be fine-tuned and executed on a single GPU (free Colab version available!). Give it a try: github.com/IST-DASLab/Pan…