Olatunji Ruwase

Zhaoyuan Su, Olatunji Ruwase, Karthik Ganesan et al.

Production LLM workloads increasingly serve discriminative tasks, such as classification, recommendation, and verification, whose answers are read from the logits of a single prefill pass with no autoregressive decoding. Serving these prefill-only workloads on mixture-of-experts (MoE) models is bottlenecked not by compute but by the distributed execution required to fit the model: existing parallel strategies (tensor, expert, and pipeline parallelism) trade memory pressure for redundant computation, communication, and synchronization, severely degrading MoE prefill serving efficiency. We observe that these overheads stem from coupling expert placement with synchronous activation routing -- a design inherited from the decoding era. The long, compute-bound forward passes of large-batch prefill open a per-layer window wide enough to stream expert weights in the background, replacing per-layer activation AllToAll with asynchronous weight AllGather fully overlapped with computation. We propose ZeRO-Prefill, a prefill-only serving system whose backend, AsyncEP (Asynchronous Expert Parallelism), gathers experts by weight rather than routing them by activation, and whose frontend co-enforces a physically-derived saturation threshold through prefix-aware routing and true-FLOPs load tracking. On Qwen3-235B-A22B across four hardware/precision configurations, ZeRO-Prefill delivers 1.35-1.37x throughput over the strongest distributed baseline on real-world workloads and up to 1.59x on long-context synthetic workloads, sustaining 29.8-36.2% per-GPU model FLOPs utilization.

Mahmoud Ahmed, Sameh Abdulah, Olatunji Ruwase et al.

Multimodal deep learning models enable joint learning across heterogeneous data sources, including text, images, and video, but their rapid scaling introduces significant memory and communication bottlenecks. As model sizes and sequence lengths increase, training performance becomes increasingly impacted by data movement rather than computation. Frameworks such as DeepSpeed mitigate these challenges through CPU offloading, activation checkpointing, and communication optimizations. However, these techniques introduce additional system activity, which may affect energy efficiency. Meanwhile, tightly integrated heterogeneous architectures, such as the NVIDIA Grace Hopper (GH200) superchip, provide high-bandwidth CPU-GPU interconnects and unified memory, thereby reducing data transfer overhead. In this work, we present a cross-layer analysis of energy and performance trade-offs in multimodal training on GH200 systems, explicitly characterizing the interactions between application, runtime, and hardware layers. Leveraging high-bandwidth CPU-GPU interconnects, our results show that energy efficiency is primarily governed by data movement and overlap rather than raw compute utilization, and that configurations optimized for runtime are not necessarily optimal for energy. Based on these findings, we distill a set of actionable guidelines for practitioners that demonstrate how to balance offloading strategies, sequence parallelism, and hardware-aware scheduling to achieve energy-efficient training. Our results demonstrate that leveraging high-bandwidth CPU-GPU interconnects enables offloading strategies and sequence parallelism, achieving a strong balance among energy efficiency, runtime performance, and computational throughput, providing practical guidelines for efficient multimodal training on modern heterogeneous systems.

Xinyu Lian, Walid Krichene, Beichen Huang et al.

Quantization is widely used to reduce the inference cost of large language models, but its effect on reasoning models is not fully captured by final-answer accuracy or per-token latency. We show that low-bit post-training quantization can introduce a hidden test-time compute cost: quantized reasoning models often generate longer chains of thought even when they still answer correctly. Across mathematical reasoning, code generation, scientific question answering, and agentic tool-use benchmarks, we find that INT4/INT3 quantization can preserve accuracy but increase reasoning-token usage, offsetting the expected per-token speedup. To measure this effect, we introduce the CoT Token Inflation Ratio, which compares reasoning length between quantized and full-precision models averaged across all evaluation benchmarks. We further show that token inflation is accompanied by behavioral changes in the reasoning trace, including more intermediate steps and greater semantic repetition. These changes translate into measurable end-to-end real-world serving penalties. Finally, we evaluate mitigation strategies and find that prompting and decoding-time sampling offer inconsistent accuracy-length trade-offs, while quantization-aware training shows more promise in reducing both accuracy degradation and token inflation. Our results suggest that reasoning-token usage should be reported alongside accuracy when evaluating quantized reasoning models.