[in-person] Abu Dhabi Machine Learning Meetup Season 6 Episode 2

🚀 Abu Dhabi Machine Learning Majlis

🚨 Emirates ID required for entry
(or passport, if not resident)

Logistics:
1. Fill the form: https://forms.gle/cd7Cxc3E3o2teTkR8
2. Show up on time at ADGM Al Maqam tower - 20th floor @ 5pm.

Looking forward to see you at the next ADML Abu Dhabi Machine Learning Majlis.

Large-Scale, High-Throughput LLM Inference: The Art & Science of Building AI Systems at Scale

This meetup will bring together engineers, researchers, founders, and practitioners to discuss:

• High-throughput LLM inference architectures
• Batch processing and large-scale data enrichment
• vLLM, SGLang, TensorRT-LLM, Ray, and modern inference stacks
• GPU utilization, scheduling, batching, and cost optimization
• Structured extraction and derived data systems
• Production lessons learned from deploying AI at scale
• Open-source and frontier model ecosystems
• Building AI-native companies in the UAE and beyond

Whether you’re working on AI products, research, data platforms, inference infrastructure, or simply interested in learning from others in the ecosystem, we’d love to have you join us.

📍 Abu Dhabi
🗓 June 24, 2026
🌇 Evening event: 5 - 8 PM

Talk 1: SlimQwen: Exploring the Pruning and Distillation in Large MoE Model Pre-training

Talk 2: Fast Inference with Cerebras

Talk 3: Inference Arithmetic for GPU throughput optimization: From Dense Models to Mixture-of-Experts

Talk 4: AI Inference at the Edge: Running Small Models at Scale

More details below...

Talk 1: SlimQwen: Exploring the Pruning and Distillation in Large MoE Model Pre-training

Abstract:
Structured pruning and knowledge distillation (KD) are typical techniques for compressing large language models, but it remains unclear how they should be applied at pretraining scale, especially to recent mixture-of-experts (MoE) models. In this work, we systematically study MoE compression in large-scale pretraining, focusing on three key questions: whether pruning provides a better initialization than training from scratch, how expert compression choices affect the final model after continued training, and which training strategy is most effective. We have the following findings: First, across depth, width, and expert compression, pruning a pretrained MoE consistently outperforms training the target architecture from scratch under the same training budget. Second, different one-shot expert compression methods converge to similar final performance after large-scale continual pretraining. Motivated by this, we introduce a simple partial-preservation expert merging strategy that improves downstream performance across most benchmarks. Third, combining KD with the language modeling loss outperforms KD alone, particularly on knowledge-intensive tasks. We further propose multi-token prediction (MTP) distillation, which yields consistent gains. Finally, given the same training tokens, progressive pruning schedules outperform one-shot compression, suggesting that gradual architecture transitions lead to better optimization trajectories. Putting it all together, we compress Qwen3-Next-80A3B to a 23A2B model that retains competitive performance. These results offer practical guidance for efficient MoE compression at scale.

Bio:
Shengkun Tang is a PhD student of Machine Learning in MBZUAI, under the supervision of Prof. Zhiqiang Shen. He was a research intern in Alibaba Qwen Pretraining Team, exploring the structured pruning and knowledge distillation to obtain powerful small LLMs. His research focuses on building efficient machine learning systems. He is interested in improving the full pipeline of modern foundation models, including inference, training, data efficiency, and novel efficiency architectures.

Talk 2: Fast Inference with Cerebras
Fares Hassanieh (Cerebras)
Ahmed Frikha (Cerebras)

Abstract
This talk explains how Cerebras delivers industry-leading inference speed by rethinking the hardware through its Wafer Scale Engine.
LLM inference is fundamentally a memory-bound problem: Generation speed is gated by memory bandwidth, not raw FLOPs. Conventional GPUs separate a large pool of compute from off-chip HBM; the Cerebras Wafer-Scale Engine (WSE-3) instead meshes 900,000 processing elements with 44 GB of on-chip SRAM, delivering roughly 3,100× the memory bandwidth of a leading GPU.
We show how this architecture maps inference onto the wafer, placing model layers into a low-latency on-chip pipeline while keeping weights and KV cache local to compute. The session will close with current ML research directions and publications in inference optimization—including speculative decoding, quantization, and pruning—that further improve speed and reduce memory footprint, unlocking new classes of real-time AI applications and markets.

Short bio:
Fares Hassanieh is a Principal AI Architect at Cerebras Systems. He is a technology leader with more than seventeen years of experience spanning generative AI, AI inference and infrastructure, and cloud. Fares combines a unique blend of corporate, consulting, and entrepreneurial experience.
Fares partners with organizations across the region to design and deploy generative AI solutions at scale — from high-performance inference and model serving to the accelerated computing infrastructure that powers them. Over his career he has led multiple digital transformation projects and built partner ecosystems across the Middle East and Turkey.
Fares holds an MBA from Warwick Business School and a master's degree in Engineering from École Supérieure d'Ingénieurs de Beyrouth.

Ahmed Frikha is an Applied AI Scientist at Cerebras Systems. His current work there focuses on different stages of LLM post-training and synthetic data generation, and he has contributed to the development of the Jais 2 8B and 70B models. He previously held applied AI research roles at Huawei and Siemens in Germany, working on LLM privacy and finetuning, meta-learning, continual learning, domain adaptation, physics-informed neural networks, and applications across different data modalities. He holds a PhD in Computer Science from LMU University of Munich.

Talk 3: Inference Arithmetic for GPU throughput optimization: From Dense Models to Mixture-of-Experts

Abstract
Efficient LLM inference is often less about peak FLOPs and more about understanding the arithmetic of deployment: how weights, activations, KV cache, batching, and hardware bandwidth interact to determine real throughput and latency.
In this talk, we will build a practical mental model for inference optimization, starting with dense transformer models. We will examine the basic arithmetic behind prefill and decoding, why decoding is often memory-bandwidth bound, and how batch size changes the throughput–latency tradeoff. We will then extend this reasoning to Mixture-of-Experts models, where sparse activation reduces compute but introduces new constraints around expert placement, memory movement, and multi-device communication.
The goal is to give practitioners a simple but rigorous framework for diagnosing inference bottlenecks and making better deployment decisions beyond raw model size or theoretical FLOPs.

Short bio
Etienne Goffinet is an applied AI scientist currently working at Mistral AI, with a PhD in Computer Science and experience across LLM pretraining, model adaptation, time-series foundation models, and inference optimization. He previously worked as a Research Engineer and held AI research roles at Cerebras Systems and the Technology Innovation Institute. At Cerebras, he took part in the pretraining and post-training of elements of the Jais series of models; at TII, he worked on the Falcon 180B project. His work focuses on building and optimizing large-scale AI systems, from frontier model training to efficient inference in production.

Talk 4: AI Inference at the Edge: Running Small Models at Scale

Abstract
AI inference is moving closer to where data is created: on phones, laptops, embedded devices, robots, and AI PCs. This shift promises lower latency, lower cloud cost, better privacy, and offline reliability, but it also changes the engineering problem. At the edge, model performance is no longer defined only by benchmark scores or parameter count; it is shaped by memory bandwidth, thermal limits, battery constraints, KV-cache growth, quantization choices, runtime support, and the hardware backend available on each device.
This talk will unpack the practical systems behind edge inference for small language and multimodal models, from prefill and decode to low-bit quantization and hardware-aware deployment. The goal is to give ML engineers and researchers a clear framework for evaluating edge inference systems: what to measure, where bottlenecks appear, why the same model behaves differently across devices, and why “build once, deploy anywhere” remains one of the hardest unsolved problems in applied AI infrastructure.

Short bio:
Sami Moustachir is the co-founder of Xybrid, an on-device AI inference startup building a cross-platform SDK to run AI models across fragmented edge hardware. He is a full-stack data and machine learning engineer with 10 years of experience in distributed systems, data platforms, and applied AI. He has worked across French corporate data labs and fintech infrastructure in South East Asia, and now focuses on making local AI inference faster, portable, and easier to deploy.