Parallelism introduces collective communication that is both expensive and represents a phase when hardware resources are underuti-lized.

Test-time scaling (TTS) has recently emerged as a promising direction to exploit the hidden reasoning capabilities of pre-trained large language models (LLMs). However, existing scaling methods narrowly focus on the compute-optimal Pareto-frontier, ignoring the simple fact that compute-optimal is not always system-optimal. In this work, we propose a system-driven perspective on TTS, analyzing how reasoning models scale against practical metrics, such as latency and cost-per-token. By evaluating the impact of popular optimizations such as tensor parallelism and speculative decoding, our preliminary analysis reveals the limitations of current methods and calls for a paradigm shift toward holistic, system-aware evaluations that capture the true essence of scaling laws at inference time.

Multi-Head Latent Attention (MLA), introduced in DeepSeek-V2, compresses key-value states into a low-rank latent vector, caching only this vector to reduce memory. In tensor parallelism (TP), however, attention heads are computed across multiple devices, and each device must load the full cache, eroding the advantage of MLA over Grouped Query Attention (GQA). We propose Tensor-Parallel Latent Attention (TPLA): a scheme that partitions both the latent representation and each head's input dimension across devices, performs attention independently per shard, and then combines results with an all-reduce. TPLA preserves the benefits of a compressed KV cache while unlocking TP efficiency. Unlike Grouped Latent Attention (GLA), every head in TPLA still leverages the full latent representation, maintaining stronger representational capacity. TPLA is drop-in compatible with models pre-trained using MLA: it supports MLA-style prefilling and enables efficient tensor-parallel decoding without retraining. Applying simple orthogonal transforms -- e.g., the Hadamard transform or PCA -- before TP slicing further mitigates cross-shard interference, yielding minimal accuracy degradation. By reducing the per-device KV cache for DeepSeek-V3 and Kimi-K2, we achieve 1.79x and 1.93x speedups, respectively, at a 32K-token context length while maintaining performance on commonsense and LongBench benchmarks. TPLA can be implemented with FlashAttention-3, enabling practical end-to-end acceleration.

In the "modern era", such model-parallel training techniques trace their roots back to AlexNet [

Energy efficiency of training and inferencing with large neural network models is a critical challenge facing the future of sustainable large-scale machine learning workloads. This paper introduces an alternative strategy, called phantom parallelism, to minimize the net energy consumption of traditional tensor (model) parallelism, the most energy-inefficient component of large neural network training. The approach is presented in the context of feed-forward network architectures as a preliminary, but comprehensive, proof-of-principle study of the proposed methodology. We derive new forward and backward propagation operators for phantom parallelism, implement them as custom autograd operations within an end-to-end phantom parallel training pipeline and compare its parallel performance and energy-efficiency against those of conventional tensor parallel training pipelines. Formal analyses that predict lower bandwidth and FLOP counts are presented with supporting empirical results on up to 256 GPUs that corroborate these gains. Experiments are shown to deliver ~50% reduction in the energy consumed to train FFNs using the proposed phantom parallel approach when compared with conventional tensor parallel methods. Additionally, the proposed approach is shown to train smaller phantom models to the same model loss on smaller GPU counts as larger tensor parallel models on larger GPU counts offering the possibility for even greater energy savings.

This paper presents the Photonic FabricTM and the Photonic Fabric ApplianceTM (PFA), a photonic-enabled switch and memory subsystem that delivers low latency, high bandwidth, and low per-bit energy. By integrating high-bandwidth HBM3E memory, an on-module photonic switch, and external DDR5 in a 2.5D electro-optical system-in-package, the PFA offers up to 32 TB of shared memory alongside 115 Tbps of all-to-all digital switching. The Photonic FabricTM enables distributed AI training and inference to execute parallelism strategies more efficiently. The Photonic Fabric removes the silicon beachfront constraint that limits the fixed memory-to-compute ratio observed in virtually all current XPU accelerator designs. Replacing a local HBM stack on an XPU with a chiplet that connects to the Photonic Fabric increases its memory capacity and correspondingly its memory bandwidth by offering a flexible path to scaling well beyond the limitations of on-package HBM alone. We introduce CelestiSim, a lightweight analytical simulator validated on NVIDIA H100 and H200 systems. It is used to evaluate the performance of LLM reference and energy savings on PFA, without any significant change to the GPU core design. With the PFA, the simulation results show that up to 3.66x throughput and 1.40x latency improvements in LLM inference at 405B parameters, up to 7.04x throughput and 1.41x latency improvements at 1T parameters, and 60-90% energy savings in data movement for heavy collective operations in all LLM training scenarios. While these results are shown for NVIDIA GPUs, they can be applied similarly to other AI accelerator designs (XPUs) that share the same fundamental limitation of fixed memory to compute.

Reasoning models represented by the Deepseek-R1-Distill series have been widely adopted by the open-source community due to their strong performance in mathematics, science, programming, and other domains. However, our study reveals that their benchmark evaluation results are subject to significant fluctuations caused by various factors. Subtle differences in evaluation conditions can lead to substantial variations in results. Similar phenomena are observed in other open-source inference models fine-tuned based on the Deepseek-R1-Distill series, as well as in the QwQ-32B model, making their claimed performance improvements difficult to reproduce reliably. Therefore, we advocate for the establishment of a more rigorous paradigm for model performance evaluation and present our empirical assessments of the Deepseek-R1-Distill series models.

Low-latency decoding for large language models (LLMs) is crucial for applications like chatbots and code assistants, yet generating long outputs remains slow in single-query settings. Prior work on speculative decoding (which combines a small draft model with a larger target model) and tensor parallelism has each accelerated decoding. However, conventional approaches fail to apply both simultaneously due to imbalanced compute requirements (between draft and target models), KV-cache inconsistencies, and communication overheads under small-batch tensor-parallelism. This paper introduces SwiftSpec, a system that targets ultra-low latency for LLM decoding. SwiftSpec redesigns the speculative decoding pipeline in an asynchronous and disaggregated manner, so that each component can be scaled flexibly and remove draft overhead from the critical path. To realize this design, SwiftSpec proposes parallel tree generation, tree-aware KV cache management, and fused, latency-optimized kernels to overcome the challenges listed above. Across 5 model families and 6 datasets, SwiftSpec achieves an average of 1.75x speedup over state-of-the-art speculative decoding systems and, as a highlight, serves Llama3-70B at 348 tokens/s on 8 Nvidia Hopper GPUs, making it the fastest known system for low-latency LLM serving at this scale.