Gengyu Rao, Ghadeer Ahmed H Alabandi, Su Ann Chong, Thomas Huber, Kailash Gogineni, Subhajit Dutta Chowdhury, Saleel Kudchadker, Suchita Pati 等 31 位作者
Large language models (LLM) have been the key focus of recent machine learning breakthroughs. As the model size grows exponentially, the computation needed for training is astronomical. A GPU can do a huge amount of arithmetic. However, the data its compute units need is not always available when they are ready. This data may still be in slow memory or on another GPU. While the chip waits, expensive compute sits idle, and at scale even a small drop in utilization translates into substantial cost. In this paper, we treat this waiting time as the primary bottleneck. We focus on hardware features designed for explicit data movement. These features load data from off-chip memory onto the chip, rearrange it into the layout the computation requires as it is loaded, and transfer it between GPUs. By orchestrating them, we keep a steady pipeline of data flowing to the compute units so that they remain busy. We present kernel and framework implementations that address this bottleneck using three key data-movement instructions and hardware features. The Tensor Data Mover (TDM) moves tensor data from high-bandwidth memory (HBM) into on-chip memory. Transpose-load instructions transpose tensor data into the layout that LLM computations require as it is loaded on-chip. Finally, System DMA (SDMA) and Remote DMA (RDMA) move data between GPUs over the scale-up and scale-out networks, respectively. Our results show that TDM and transpose-load provide on average 1.88x and 1.19x speedup at the kernel level, respectively. The SDMA and RDMA lead to on average 1.16x and 1.33x speedup for computation overlapped with communication. This is the first AMD LLM training system to utilize all the GPU data pipelining features. Our implementation has been integrated into open source framework and library for AMD GPUs. https://github.com/AMD-AGI/Primus-Turbo https://github.com/AMD-AGI/Primus
Anshu Raina, Yao Fu, Zhang Jiejing, Peyman Razaghi, Shiksha Patel, Ravi Gupta, Deval Shah, Adeem Jassani 等 14 位作者
Serving an LLM efficiently depends on the engine, how the model is split across GPUs, precision, scheduler settings, speculative decoding, KV-cache management, and prefill-decode disaggregation. The best configuration depends on the model, hardware, workload, and objective. Operators usually find it by trial and error. Exhaustively measuring the candidates in our study would take about 25,000 GPU-hours. ROCm InferaSim predicts serving performance from a few short measurements. Choices such as batch size, prompt length, and GPU count change the amount of work per GPU without changing the kernels. ROCm InferaSim benchmarks each distinct set of kernels once, on at most one node. We call this measurement an anchor. Analytical laws carry that anchor to other deployments that run the same kernels. For multi-node deployments, ROCm InferaSim reduces the configuration to a single-node equivalent, measures it, and scales the result analytically. A simulator replays multi-turn traffic, and a tuning agent searches configurations using the predictions. On AMD Instinct MI355X GPUs with vLLM, SGLang, and ATOM, 11 anchors took 9.1 GPU-hours, against 204 GPU-hours to measure the 226 fixed-length configurations they cover. On 172 configurations, time per output token has 8.1% median error. On a multi-turn coding workload, throughput is within 8%. For 159 disaggregated deployments, none used as an anchor, throughput is within 5.5%. A five-candidate short list contains a deployment within 12% of the best feasible pairing of model and latency target.
Vansh Bhatia, Mehdi Rezagholizadeh, Parsa Ashrafi Fashi, Igor Gilitschenski, Vikram Appia, Emad Barsoum
Final-answer accuracy reports whether a vision-language model was right, but not whether its answer was grounded in the image, whether its stated reasoning reflects that evidence, or what the answer cost to produce. Those are the properties that decide whether a system can be deployed. We present PRISM-BENCH, a benchmark of 1,000 expert-curated questions spanning five under-represented reasoning categories, in which every item carries an idealized reasoning trace written by one human annotator and validated by another. The traces support two diagnostics that accuracy cannot provide. A perception probe, in which the model answers from a description it wrote itself, attributes roughly 85% of frontier-model failures to misperception and places the bottleneck at grounding. Trace alignment predicts answer correctness for all 25 systems tested (macro AUROC 0.66) and survives length control, so it measures reasoning content and not verbosity. We further introduce CORE, an efficiency-aware metric that reorders 24.2% of model pairs relative to accuracy while remaining uncorrelated with response length. Across 25 open and closed systems, including GPT-5 and Claude Opus 4.6, the best accuracy is 66.0%. On the four categories with a human baseline the best system trails humans by 18.4 points, and recent frontier gains fall on recognition and not on spatial grounding.
Training large language models (LLMs) in sub-8-bit arithmetic promises substantial gains in memory bandwidth and compute throughput, yet FP4 formats introduce quantization error severe enough to destabilize optimization if left unaddressed. We present ALTO, an open-source training recipe for GPT-OSS-20B, a 20-billion-parameter mixture-of-experts (MoE) model, on the MLPerf Small MoE Training benchmark under the MXFP4 microscaling format. The recipe combines established techniques from Abecassis, Felix et al. [arXiv:2509.25149] (hybrid 1D/2D block quantization, randomized Hadamard transforms, and stochastic rounding) with a simplified weight de-oscillation scheme adapted from TetraJet-v2 [arXiv:2510.27527]. On C4 validation at global batch size 16, MXFP4 + RHT + SR + de-oscillation reaches a validation loss of 3.3350 at 16,128 steps, only 0.007 above the BF16 baseline (3.3283), closing roughly 50% of the gap left by MXFP4 without de-oscillation (3.3418). Measured by steps to reach the MLPerf quality target (validation loss 3.34), de-oscillation cuts the FP4 convergence overhead relative to BF16 from +20.0% to +5.0%. We also report negative end-to-end results from differential gradient estimation, outlier clipping, and macro-block scaling, despite each showing operator-level SNR gains. All experiments use simulated MXFP4 kernels on AMD MI300X hardware and wall-clock speed has not been optimized.
Embodied agents in open-ended environments such as Minecraft increasingly adopt planner–controller architectures, with large language models acting as high-level planners. While planning has advanced rapidly, control remains underexplored. Existing systems commonly rely on a monolithic policy to execute subgoals across varying contexts, forcing incompatible behaviors into a shared parameter space and causing interference that scaling only partially mitigates. To address this, we propose MoEC, a Memory-Routed Mixture-of-Experts Controller for Adaptive Minecraft Control. MoEC routes via a subgoal-indexed, non-parametric expert memory and regulates capacity through failure-triggered expert growth and redundancy-aware consolidation. This design enables continual adaptation without full retraining, while maintaining parameter efficiency and with bounded inference cost. We evaluate MoEC on diverse and compositional Minecraft tasks, demonstrating significant gains in adaptability, robustness, and execution consistency over strong baselines, yielding a scalable and efficient alternative for open-ended control.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Xiusheng Huang, Zhe Li AMD, Xuanwu Yin Advanced Micro Devices, Lu Wang China Three Gorges Corporation, Yequan Wang Peking University and Beijing Academy of Artificial Intelligence, Dong Li, Emad Barsoum AMD, Kang Liu Institute of automation, Chinese academy of science, Chinese Academy of Sciences
Post-training quantization has emerged as a widely adopted technique for compressing and accelerating the inference of Large Language Models (LLMs). The primary challenges in LLMs quantization stem from activation outliers, which significantly degrade model performance especially at lower bit precision. While recent approaches attempt to mitigate outliers through linear transformations across feature dimensions, our analysis reveals that the transformed weights and activations still exhibit persistent outlier patterns with concentrated magnitude distributions. In this paper, we first model the mathematical relationship between quantization error and outliers, and then introduce a new metric Flatness to quantify the distribution of outliers. Based on this, we derive the theoretical optimal solution with respect to Flatness. Building on these insights, we propose Bidirectional Diagonal Quantization (BDQ), a novel post-training quantization framework that effectively disperses outlier patterns through optimized matrix transformations. BDQ strategically distributes outlier magnitudes across matrix dimensions via learned diagonal operations. Extensive experiments demonstrate that BDQ establishes a new quantization benchmark. It achieves less than 1% accuracy drop in W4A4 quantization on the LLaMA-3-8B model. In the more challenging W2A4KV16 experiment, compared to state-of-the-art approaches, BDQ reduces the performance gap by 39.1% on the DeepSeek-R1-Distill-LLaMA-70B model.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表