探索全球前沿学术脉络

MegaFold: Efficient Training of Next-Generation 3D Attention Protein Models on Cross-Platform GPUs

arXiv:2506.20686v2 Announce Type: replace-cross Abstract: Recent advances in biomolecular modeling have been catalyzed by models such as AlphaFold3 (AF3), which introduce science-informed changes to the transformer architecture. Unlike transformers, a defining characteristic of AF3-style models is their 3D attention over 2D pairwise representations which produces tensors whose computation and memory costs scale cubically with sequence length. As a result, despite moderate parameter counts, AF3-style models are far more expensive to train than size-equivalent transformers, and are severely constrained by GPU memory capacity. Our characterization shows 3D attention fundamentally changes the training workload, causing massive 3D attention maps, complex inter-operator dependencies, kernel fragmentation, and heavy host-side data pipelines which differ substantially from LLM training, leading to poor utilization on modern GPU systems. Moreover, existing GPU optimizations do not adequately address these challenges due to complex cross-layer inter-operator dependencies introduced by 3D attention. Motivated by these challenges, we introduce MegaFold, a novel cross-platform system for efficient training of next-generation 3D-attention protein models. MegaFold combines a memory-efficient 3D-attention kernel, a communication-efficient sharding strategy for quadratic representations, fused operator implementations for critical execution paths, and a determinism-aware host-device pipeline that eliminates preprocessing stalls. Evaluation on both NVIDIA H200 and AMD MI250 GPUs shows that MegaFold enables training with up to 3.36$\times$ longer sequence lengths on 32 GPUs while reducing end-to-end execution time by up to 1.73$\times$ (NVIDIA) and 1.62$\times$ (AMD).

Quantization Inflates Reasoning: Token Inflation as a Hidden Cost of Low-Bit Reasoning Models

arXiv:2606.25519v1 Announce Type: new Abstract: 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.

VoltanaLLM: Energy-Efficient and SLO-Aware Disaggregated LLM Serving via Adaptive Frequency Control and State-Space Routing

arXiv:2509.04827v3 Announce Type: replace-cross Abstract: The energy cost of Large Language Model (LLM) inference is rapidly becoming a barrier to sustainable and scalable deployment. Although modern serving architectures expose distinct prefill and decode behaviors, existing systems fail to exploit these phase differences for energy-efficient serving under strict latency SLOs. This paper introduces VoltanaLLM, the first system that explicitly targets and reduces the energy bloat in modern prefill-decode (P/D) disaggregated LLM serving. Guided by a control-theory perspective, VoltanaLLM separates two levers: per-instance operating-point selection (GPU frequency per iteration) and system-level state-space routing of requests. We empirically observe that LLM inference exhibits a U-shaped energy-frequency curve creating "sweet spots" that depend on phase behavior and load. VoltanaLLM exploits this by combining phase-specific, iteration-level frequency selection driven by a lightweight, online-adaptive latency predictor, with a decode state-space guided router that avoids architectural granularity-induced inefficiencies, all while meeting desired SLOs. We implement VoltanaLLM using SGLang and evaluate it across multiple models and real-world workloads. Our results show VoltanaLLM reduces end-to-end energy by up to 36.3% versus a static max-frequency baseline while maintaining high SLO attainment, and generalizes to newer GPUs. These results point to sustainable LLM serving via phase-aware, iteration-level frequency selection coupled with architecture-aware routing. Source code is available in https://github.com/Supercomputing-System-AI-Lab/VoltanaLLM.