arXiv subjects

Soojung Ryu

Publications and source records attributed to Soojung Ryu.

4 recordsLinked to original sources

arXiv2026

Phase-Decoupled, Model-Calibrated Power Control for Disaggregated LLM Serving

Datacenter GPU power is the binding constraint on LLM serving capacity, and production serving has shifted to prefill/decode (PD) disaggregation. Deploying NVIDIA's Max-Q inference profile on a disaggregated B200 system, we found its realized gain modest (+8.6% tokens/J), model-dependent, and carrying a mean end-to-end latency cost (+5.2%) that throughput-only evaluation does not surface; the profile also applies one setting to prefill and decode GPUs that operate in opposite hardware regimes. We hypothesize that the optimal power setting is a property of the deployed (model, quantization, engine, hardware) combination rather than of the GPU class, that each lane warrants its own profile, and that converting SLO headroom into energy safely requires latency-gated calibration under a runtime SLO guard rather than a fixed recipe. We present a phase-decoupled, model-calibrated controller: the prefill lane runs under an SM-clock window whose floor is a latency guarantee by construction, and the decode lane under a power cap placed by automatic calibration just above a measured throughput/latency cliff. Because a disaggregated decode lane draws flat, memory-bound power, the cap binds continuously, the reactive-overshoot weakness that led POLCA to reject capping is absent, and the GPU's own power manager retains throughput under the cap. On an 8x B200 node serving Qwen3-Coder-480B (FP8) under agentic load, our balanced mode delivers +20.4% tokens/J at +3.5% mean e2e versus +8.6% at +5.2% for Max-Q, a Pareto improvement on both axes. On Qwen3-235B-A22B (NVFP4) every operating mode meets the ITL-p99 SLO in every repetition; both vendor profiles miss it. A decode-actuator A/B shows the calibrated cap beats static clock locks, and a three-day sustained run saves 32.3% of a lane pair's electricity. Both models are MoE; a dense model recovers roughly 5x less, so we scope our claims to MoE serving.

cs.LG↗

arXiv2026

ReSET: Accurate Latency-Critical NVFP4 Reasoning via Step-Aware Temperature Scaling

Large reasoning models (LRMs) improve complex problem-solving by generating long intermediate reasoning traces, but this substantially increases inference costs. NVFP4 inference offers a promising approach to reduce both computational and memory costs through hardware-supported low-precision execution. However, directly applying NVFP4 to LRMs introduces two practical limitations: reasoning accuracy degrades under quantization, and existing NVFP4 kernels do not fully realize latency benefits in small-batch autoregressive decoding. In this work, we analyze the effect of NVFP4 quantization on token-level uncertainty during reasoning. We show that quantization increases incorrect sampling at low-entropy symbolic tokens, while causing over-concentration on a small set of tokens in high-uncertainty reasoning steps. Based on this observation, we propose \textbf{ReSET}, a reasoning-step entropy-based temperature-scaling method that estimates step-level uncertainty online and adapts the decoding temperature using both token-level and step-level entropy signals. To address the latency gap, we further design a CUDA-core small-$M$ NVFP4 kernel for latency-critical autoregressive decoding. Across reasoning benchmarks and model scales, ReSET improves NVFP4 reasoning accuracy by up to $\sim\!$2 points over the NVFP4 baseline. Our CUDA-core small-$M$ kernel further improves latency-critical decoding, delivering up to $2.5\!\times$ kernel-level speedup over NVFP4 vLLM and approximately $2\!\times$ end-to-end decoding speedup over BF16. Code is available at https://github.com/aiha-lab/ReSET.

cs.LG↗

arXiv2022

ShortcutFusion: From Tensorflow to FPGA-based accelerator with reuse-aware memory allocation for shortcut data

Residual block is a very common component in recent state-of-the art CNNs such as EfficientNet or EfficientDet. Shortcut data accounts for nearly 40% of feature-maps access in ResNet152 [8]. Most of the previous DNN compilers, accelerators ignore the shortcut data optimization. This paper presents ShortcutFusion, an optimization tool for FPGA-based accelerator with a reuse-aware static memory allocation for shortcut data, to maximize on-chip data reuse given resource constraints. From TensorFlow DNN models, the proposed design generates instruction sets for a group of nodes which uses an optimized data reuse for each residual block. The accelerator design implemented on the Xilinx KCU1500 FPGA card 2.8x faster and 9.9x more power efficient than NVIDIA RTX 2080 Ti for 256x256 input size. . Compared to the result from baseline, in which the weights, inputs, and outputs are accessed from the off-chip memory exactly once per each layer, ShortcutFusion reduces the DRAM access by 47.8-84.8% for RetinaNet, Yolov3, ResNet152, and EfficientNet. Given a similar buffer size to ShortcutMining [8], which also mine the shortcut data in hardware, the proposed work reduces off-chip access for feature-maps 5.27x while accessing weight from off-chip memory exactly once.

cs.DC↗

arXiv2021

GradPIM: A Practical Processing-in-DRAM Architecture for Gradient Descent

In this paper, we present GradPIM, a processing-in-memory architecture which accelerates parameter updates of deep neural networks training. As one of processing-in-memory techniques that could be realized in the near future, we propose an incremental, simple architectural design that does not invade the existing memory protocol. Extending DDR4 SDRAM to utilize bank-group parallelism makes our operation designs in processing-in-memory (PIM) module efficient in terms of hardware cost and performance. Our experimental results show that the proposed architecture can improve the performance of DNN training and greatly reduce memory bandwidth requirement while posing only a minimal amount of overhead to the protocol and DRAM area.

cs.LG↗