571 citations · 1.7k across the 33 of their papers we have counts for
11 papers · 1 filter
Helix Parallelism: Rethinking Sharding Strategies for Interactive Multi-Million-Token LLM Decoding
Nidhi Bhatia, Ankit More, Ritika Borkar +7
As LLMs scale to multi-million-token KV histories, real-time autoregressive decoding under tight Token-to-Token Latency (TTL) constraints faces growing pressure. Two core bottlenec…
Beyond the Buzz: A Pragmatic Take on Inference Disaggregation
Tiyasa Mitra, Ritika Borkar, Nidhi Bhatia +10
As inference scales to multi-node deployments, disaggregation - splitting inference into distinct phases - offers a promising path to improving the throughput-interactivity Pareto…
Nonuniform-Tensor-Parallelism: Mitigating GPU failure impact for Scaled-up LLM Training
Daiyaan Arfeen, Dheevatsa Mudigere, Ankit More +3
LLM training is scaled up to 10Ks of GPUs by a mix of data-(DP) and model-parallel (MP) execution. Critical to achieving efficiency is tensor-parallel (TP; a form of MP) execution…
Software-Hardware Co-design for Fast and Scalable Training of Deep Learning Recommendation Models
Dheevatsa Mudigere, Yuchen Hao, Jianyu Huang +50
Deep learning recommendation models (DLRMs) are used across many business-critical services at Facebook and are the single largest AI application in terms of infrastructure demand…
Deep Learning Training in Facebook Data Centers: Design of Scale-up and Scale-out Systems
Maxim Naumov, John Kim, Dheevatsa Mudigere +12
Large-scale training is important to ensure high performance and accuracy of machine-learning models. At Facebook we use many different models, including computer vision, video and…
RecNMP: Accelerating Personalized Recommendation with Near-Memory Processing
Liu Ke, Udit Gupta, Carole-Jean Wu +18
Personalized recommendation systems leverage deep learning models and account for the majority of data center AI cycles. Their performance is dominated by memory-bound sparse embed…