6 papers
Architecting for Hardware Diversity: Meta's End-to-End Co-Design for High-Performance AI Training on AMD GPUs
Haoyu Zhang, Arshad Rahman, Adnan Aziz +55
Meta deploys a heterogeneous fleet comprising NVIDIA GPUs, AMD GPUs, and proprietary MTIA accelerators to reduce costs and mitigate vendor lock-in. However, the extensive history o…
ARC-KV: Amortizing Anchor Search for Reconstruction-Based KV Cache Compaction
Zheyu Shen, Guanhua Wang, Dezhan Tu +5
Long-context large language model inference is bottlenecked by KV caches that grow linearly with sequence length. This burden is especially severe for long, reusable context prefix…
ROCS: Request-Oriented Compute Sharing for Efficient Large-Scale Recommendation
Yuxin Chen, Liang Luo, Buyun Zhang +44
Modern recommendation models gain prediction quality by scaling feature-interaction and sequence modules, but production cost constraints cap how far systems can scale. In this wor…
LoKA: Low-precision Kernel Applications for Recommendation Models At Scale
Liang Luo, Yinbin Ma, Quanyu Zhu +21
Recent GPU generations deliver significantly higher FLOPs using lower-precision arithmetic, such as FP8. While successfully applied to large language models (LLMs), its adoption in…
Scaling Multi-Node Mixture-of-Experts Inference Using Expert Activation Patterns
Abhimanyu Bambhaniya, Geonhwa Jeong, Jason Park +6
Most recent state-of-the-art (SOTA) large language models (LLMs) use Mixture-of-Experts (MoE) architectures to scale model capacity without proportional per-token compute, enabling…
Training LLMs with Fault Tolerant HSDP on 100,000 GPUs
Omkar Salpekar, Rohan Varma, Kenny Yu +20
Large-scale training systems typically use synchronous training, requiring all GPUs to be healthy simultaneously. In our experience training on O(100K) GPUs, synchronous training r…