arXiv subjects

Junyu Yin

Publications and source records attributed to Junyu Yin.

7 recordsLinked to original sources

arXiv2026

Warpline: Binary-Level Tracing and Unified Evidence for Agentic GPU Kernel Optimization

LLM agents are becoming GPU kernel engineers, but they still tune kernels without seeing where the time goes. Hardware profilers report how busy each unit was, not where each warp's time went; intra-kernel tracers can show it, but only with hand-placed probes and a recompilation that alters the very kernel being measured. We present Warpline, which gives agents this view automatically while leaving the kernel essentially unchanged. Warpline analyzes the binary to decide which events to record and inserts probes into the binary itself without recompilation or any manual step. It then aligns the trace with hardware profiling data on one unified timeline. Because it needs only the binary, it can regenerate this evidence for every new version of the kernel at whatever level the agent edits it (source, IR, or PTX), closing the optimization loop without human intervention. On NVIDIA B200, across 35 kernels in Triton, CuTe DSL, and CUDA C++, Warpline adds a median overhead of 1.2% and leaves a median 99.9% of executed instructions unchanged, whereas existing tracers have median overheads of 1.6% to 13.2% and leave only 4.6% to 32.7% unchanged. Warpline also traces closed-source kernels, which these tracers cannot. On 10 SOL-ExecBench problems, agents using Warpline spend less time and fewer model requests than a bare coding agent and NVIDIA's Kernel Design Agents (KDA), yet lead both after every round in geometric mean and reach a 7.69$\times$ geometric-mean speedup, against 6.94$\times$ and 6.13$\times$. Two case studies further demonstrate that Warpline reveals hidden bottlenecks and provides actionable optimization insights.

cs.PF↗

arXiv2026

HPC-Bench: A Comprehensive Benchmark for High Performance Computing Codes

Large Language Models (LLMs) have shown strong capabilities in code generation and reasoning, motivating extensive benchmarks for evaluating functional correctness. However, performance optimization in High-Performance Computing (HPC) poses fundamentally different challenges that are not captured by existing evaluations: HPC performance depends on low-level code, intricate memory access patterns, and computational kernels embedded within sophisticated execution contexts. We introduce HPC-Bench, a performance-grounded benchmark for evaluating LLMs as optimizers of computational kernels in realistic HPC workloads, including 114 workloads across 16 HPC computational motifs. HPC-Bench provides a fully automated, end-to-end evaluation pipeline that combines program-level correctness verification with two correctness-gated metrics, Fast@k and Speedup@k, to characterize reliability and speedup under a fixed sampling budget. Our evaluation of three representative LLMs across serial CPU, OpenMP, and CUDA settings shows that stricter speedup requirements substantially reduce the probability of finding a correct and fast optimization. HPC-Bench is available at https://github.com/Deep-Learning-Profiling-Tools/HPC-Bench.

cs.PF↗

arXiv2026

Elastic Spectral State Space Models for Train-Once Budgeted Inference

Modern sequence models are typically trained at a fixed computational capacity, while real-world applications require deployment across platforms with different resource constraints. Existing approaches either train or distill separate compact models for selected budgets, or use elastic architectures that shrink dimensions such as width, depth, Feed-Forward Network (FFN) size, or SSM state dimension. In this paper, we propose Elastic Spectral State Space Models (ES-SSM), a train-once, export-many sequence modeling framework that gains elasticity through spectral approximation of the SSM sequence operator. ES-SSM builds on Hankel spectral filtering for state space models, where long-range token mixing is represented through fixed Hankel spectral channels that define an operator-level approximation resolution. To make this spectral elasticity reliably deployable, ES-SSM combines input-adaptive channel-wise gates with budget dropout, training the same spectral prefixes that are used by compact models at inference time. This encourages low-index channels to become predictive on their own, while higher-index channels provide refinements for larger budgets. We evaluate ES-SSM across byte-level language modeling, Long Range Arena, Speech Commands V2, and offline reinforcement learning benchmarks. Across these settings, a single trained ES-SSM can be truncated to competitive compact models compared with modern Transformer and SSM baselines at similar parameter scales. Furthermore, by testing under various runtime budgets, we observe smooth and stable quality-cost curves over a wide range of truncation levels. Code is available at https://github.com/GMU-MICO/ES-SSM.

cs.LG↗

arXiv2026

TileBench: A Controlled Benchmark for Performance Evaluation and Bottleneck Diagnosis of Tile-Based Programming Models

Tile-based programming models, such as Triton and cuTile, aim to simplify high-performance kernel development, but their practical performance, tuning behavior, and usability remain difficult to compare systematically. We present TileBench, a controlled benchmark for evaluating Triton and cuTile on NVIDIA B200 GPUs under matched operator semantics and comparable implementation structures. TileBench contains 45 operators covering diverse AI-kernel patterns and memory/computation behaviors. Each task provides a PyTorch reference, verified Triton and cuTile implementations, standardized data-types (dtype) and input-size sweeps, default and autotuned configurations, roofline-based metrics, and profiling-guided diagnosis. Our evaluation shows that performance gaps are workload-dependent: cuTile excels on a small cluster of Tensor-Core/TMA-friendly kernels, while Triton is stronger on many irregular, streaming, and bandwidth-bound operators. We further evaluate LLM-generated cuTile and Triton kernels and find that Triton is consistently more token-efficient than cuTile under the same iterative refinement protocol. TileBench is publicly available at https://github.com/Deep-Learning-Profiling-Tools/Tilebench.

cs.PF↗

arXiv2026

Mergeable Model-Side Aggregation States for Long-Context Language Models

A known limitation of long-context language models is their increasingly unreliable performance in non-additive, set-based aggregation as context length grows. Examples include cardinality estimation, set relationships, and grouped statistics, which widely exist in logs, program outputs, tables, and multi-turn conversations. To provide the aggregation state required by these tasks, we introduce a model-side aggregation interface that maintains compact Hash-based HyperLogLog (HLL) sketch states alongside a frozen language model. While the model processes the context, an extractor maps each relevant record to a canonical identity. The identity is then hashed and updates the HLL state. These states can be merged across context segments and/or read out directly for downstream reasoning, avoiding an additional generate-execute-return cycle. We validate the proposed approach by setting the HLL state size as 2 KiB (2,048 registers), which does not increase with context length or set cardinality. In a distinct-count experiment involving one million records, the mean relative error was 1.6%. In a separate merge test, states built from as many as 256 segments produced exactly the same readout as a single pass over the same stream. On 3,969 aggregate-then-reason tasks from 174 source windows, the fixed-budget interface reached 99.2% accuracy on Gemma 4 (31B, BF16), compared with 100.0% under exact aggregation; the paired gap was 0.8 percentage points (95% window-cluster CI: 0.5-1.3 points). On a matched set of 174 items, our method improved over direct full-context reasoning by 63.2 points on Qwen and 56.3 points on Gemma. The corresponding gains over chain-of-thought (CoT) reasoning were 60.9 and 63.2 points, respectively. On a fixed 1,200-task Oolong-Synth subset, our method reached 91.1% on Qwen and 99.3% on Gemma. Code is available at https://github.com/songdc98/sketchops.

cs.CL↗

arXiv2026

Exploring Cross-Scenario Generality of Agentic Memory Systems: Diagnostics and a Strong Baseline

LLM agents accumulate histories that outgrow their context windows, motivating a growing literature on memory systems. Yet most existing designs are tuned to a single scenario (multi-session chat or a single trajectory format), and there is little evidence that they generalize across the heterogeneous trajectories agents encounter in deployment. We revisit eight memory systems plus an agentic harness for search problems, on five scenarios: single-turn QA, multi-session chat, agentic-trajectory QA, memory stress tests, and long-horizon agentic tasks. The harness, which self-manages flat text-file storage via tool calls, achieves the best cross-task ranking, suggesting that memory performance hinges on giving the agent active control over storage and retrieval rather than on a passive store behind a fixed pipeline. We instantiate this insight in AutoMEM, an agentic memory harness with a self-managed tool interface that achieves the best cross-scenario generality among the systems we evaluate.

cs.AI↗

arXiv2026

OpenRFM: Dissecting Relational In-Context Learning

Relational Foundation Models (RFMs) promise a single pre-trained predictor that, given any relational database, returns predictions in one forward pass via relational in-context learning (ICL). Yet a substantial gap separates open RFMs from their commercial counterparts, and the origin of this gap has not been systematically understood. We dissect a representative framework, the Relational Transformer (RT), from two perspectives. Model side: we show that RT performs relation-level ICL, and a kernel regression view shows it fails when sparse label-cell coverage yields an underdetermined regression. Data side: we ablate RT's pre-training source and find that existing synthetic-only pre-training and in-distribution pre-training drive the same architecture into different regimes, lazy vs. feature-learning. Probing this gap reveals that the missing ingredient is a support-identifiable relational latent in the label-generation process. These two diagnoses translate into (1) a dual-stage ICL architecture that combines the relational backbone with a batch-level ICL layer lifted from a pre-trained tabular foundation model to overcome relation-level label scarcity, and (2) a homophily-aware synthetic plus continual real-data pre-training mixture, augmented with a prototype-based regularization. These choices define OpenRFM, a simple yet effective RFM that improves average task performance by approximately 30% over the RT backbone and surpasses the commercial model KumoRFMv1 on a large set of evaluation tasks.

cs.LG↗