Thanh Vu

🚀 Multi-GPU CUDA Programming in Algorithmic Trading In modern markets, speed is strategy, milliseconds can decide profits. That’s why traders and researchers are adopting multi-GPU CUDA computing to supercharge their models. ⚡ With multiple GPUs working in parallel, we can: 🧠 Train deep learning & reinforcement models 10× faster 📈 Backtest hundreds of strategies simultaneously 💹 Run real-time predictions across multiple instruments 🔄 Use GPU-to-GPU communication (NCCL) for low-latency data flow The result? Faster experimentation, smarter strategies, and near-instant decisions. GPU computing isn’t just for AI anymore — it’s the new backbone of quantitative finance. #CUDA #AlgoTrading #AI #Fintech #DeepLearning #NVIDIA #QuantFinance #GPUComputing #MachineLearning https://lnkd.in/d7w6ADMG

2

DeepSeek just released DeepSeek-OCR – an open-source model that investigates vision encoders from an LLM-centric perspective, featuring dynamic resolution support and ultra-efficient visual-text compression (as low as 64 tokens for 512×512 images). From document-to-markdown conversion to figure parsing and grounding-based OCR, this model pushes the boundaries of how efficiently we can process visual information. For enterprises drowning in PDFs and scanned documents, this means faster processing, lower API costs, and seamless conversion to structured markdown at ~2500 tokens/second. Check it out: https://lnkd.in/gXRdPEpa

6

Incredible milestone for our EMOTIV research team! Our paper has been accepted at NeurIPS 2025, and it represents something far bigger than an academic achievement. This research brings us closer to a future where brain–computer interfaces can truly understand us. For decades, P300 spellers have held the promise of giving a voice to those who can’t speak, but they’ve also faced real barriers: time-consuming calibration, noisy signals, and limited generalization. With this work, our team proved that foundation model approaches can help BCIs learn universally across users and tasks. Much like how language models transformed the way we communicate with machines. 94% accuracy. 60% less calibration. A leap toward effortless, human-centered neurotechnology. I'm so proud of Soheila Ghane, PhD, Jiazhen Hong, Geoff Mackellar, and the entire Emotiv research team for pushing the boundaries of what brain–computer interfaces can achieve. Not just in the lab, but in the real world. This is how we move closer to a future where neurotechnology is not only powerful, but practical, inclusive, and life-changing. #NeurIPS2025 #AIforGood #Neurotech #BrainComputerInterface #Emotiv #Innovation #HumanPotential

37

2 Comments

We’re excited to welcome Mooncake to the PyTorch Ecosystem! Mooncake is designed to solve the “memory wall” in LLM serving. By integrating Mooncake’s high performance KVCache transfer and storage capabilities with PyTorch native inference engines like SGLang, vLLM, and TensorRT-LLM, it unlocks new levels of throughput and scalability for large language model deployments. Mooncake enables prefill decode disaggregation, global KVCache reuse, elastic expert parallelism, and serves as a fault-tolerant PyTorch distributed backend. 🔗 https://lnkd.in/g3Dm4mN6 #PyTorch #OpenSourceAI #LLM #AIInfrastructure

101

Hey folks — I’m really happy to finally share something we’ve been working on for a while. We just released a new preprint introducing Contraction Dynamics Model (CDM) — a framework for bringing formal nonlinear stability guarantees into model-based reinforcement learning. MBRL is powerful and sample-efficient, but anyone who has worked with it knows the core pain point: small model errors → compounding uncertainty → unstable rollouts → brittle policies. CDM tackles this at the root by learning the geometry of stability itself. Instead of treating stability as a constraint or a post-training check, we jointly learn: • system dynamics • control policies • and state-dependent Riemannian contraction metrics This allows the learner to actively prefer regions of the state space where trajectories are provably contracting. ⸻ What’s inside • Differentiable positive-definite metric learning via softplus-Cholesky parameterization • Stability optimization using virtual displacement dynamics • An adaptive contraction-based regularizer embedded directly into the policy objective • Formal guarantees: exponential trajectory convergence (in expectation), robustness bounds to model error, and sample complexity characterization ⸻ Results On continuous control benchmarks (Pendulum, CartPole, HalfCheetah): • 38.9% fewer environment steps to reach comparable performance • 78% performance retention under 10% model noise (vs 52% for strong MBRL baselines like PETS / MBPO) • 10–40% gains from learned metrics in ablations, with ~20% compute overhead ⸻ Why this matters This shows that learning contraction metrics is a practical and scalable way to embed nonlinear stability directly into deep RL systems — not just in theory, but in real training pipelines. If you’re working on: model-based RL, robotics, control theory, safe learning, or reliable AI systems, I’d love to hear your thoughts. Preprint link below 👇 https://lnkd.in/gZ2qUEEF

3

Today’s guide by Reza Bahmanzadeh, AI Engineer on our Professional Services Team, teaches the essentials of serving large models with vLLM: https://lnkd.in/eHK9nwZm It builds from foundational neural network concepts, like transformers and attention, to introduce practical inference workflows, explore useful vLLM features and offer hands-on, operational guidance for efficient deployments. #LLMs #Serving #Inference #TechGuides

36

2 Comments

In 2025, PyTorch Foundation became an umbrella foundation, bringing PyTorch and related frameworks together across training, inference, and distributed compute. As Chair of the PyTorch Technical Advisory Council, Lightning AI CTO Luca Antiga helped push the hardest work forward: shared CI at massive scale, accelerator support beyond CUDA, ecosystem curation, and security governance. This is how open, production-grade AI actually scales. Read the full PyTorch Foundation year in review → https://lnkd.in/ehNWW93P

27