arXiv subjects

Yingyi Huang

Publications and source records attributed to Yingyi Huang.

17 recordsLinked to original sources

arXiv2026

CAKE: Compiler-Agent Co-Design for Frontier Kernel Evolution

GPU kernel agents and GPU programming languages have advanced separately, leaving expert kernels difficult to reproduce. Agents usually treat the compiler as a fixed black box and receive only errors, correctness outcomes, and timing, while existing DSLs either hide critical scheduling decisions or expose them through difficult layout abstractions. We present CAKE, a compiler-agent co-design in which agents author CAKE IR, a typed, hardware-explicit schedule representation. CAKE exposes warp roles, memory movement, synchronization, and pipelines while supporting verification, cost modeling, and localized diagnostics. The harness itself evolves: recurring failures become verifier rules, IR primitives, model calibrations, and reusable optimization tactics. In matched implementation-hidden Flash-KMeans clean starts on B200, the best CAKE IR candidate at an 80-million-token budget runs at 1.144x the tuned FlashML baseline, compared with 0.928x for direct CUDA/PTX. Beyond this benchmark, agent-generated Kimi Delta Attention achieves a 2.05x geometric-mean speedup over official FlashKDA and passes end-to-end serving validation. Dispatcher-backed KNN and KMeans improve performance by 1.42x to 2.12x across more than 400 shapes, and four kernel changes are available as upstream PRs. CAKE targets NVIDIA GPUs from Ampere through Blackwell and separates single-shape evolution from library generalization and dispatch.

cs.LG↗

arXiv2026

Event Tensor: A Unified Abstraction for Compiling Dynamic Megakernel

Modern GPU workloads, especially large language model (LLM) inference, suffer from kernel launch overheads and coarse synchronization that limit inter-kernel parallelism. Recent megakernel techniques fuse multiple operators into a single persistent kernel to eliminate launch gaps and expose inter-kernel parallelism, but struggle to handle dynamic shapes and data-dependent computation in real workloads. We present Event Tensor, a unified compiler abstraction for dynamic megakernels. Event Tensor encodes dependencies between tiled tasks, and enables first-class support for both shape and data-dependent dynamism. Built atop this abstraction, our Event Tensor Compiler (ETC) applies static and dynamic scheduling transformations to generate high-performance persistent kernels. Evaluations show that ETC achieves state-of-the-art LLM serving latency while significantly reducing system warmup overhead.

cs.DC↗

arXiv2026

FlashInfer-Bench: Building the Virtuous Cycle for AI-driven LLM Systems

Recent advances show that large language models (LLMs) can act as autonomous agents capable of generating GPU kernels, but integrating these AI-generated kernels into real-world inference systems remains challenging. FlashInfer-Bench addresses this gap by establishing a standardized, closed-loop framework that connects kernel generation, benchmarking, and deployment. At its core, FlashInfer Trace provides a unified schema describing kernel definitions, workloads, implementations, and evaluations, enabling consistent communication between agents and systems. Built on real serving traces, FlashInfer-Bench includes a curated dataset, a robust correctness- and performance-aware benchmarking framework, a public leaderboard to track LLM agents' GPU programming capabilities, and a dynamic substitution mechanism (apply()) that seamlessly injects the best-performing kernels into production LLM engines such as SGLang and vLLM. Using FlashInfer-Bench, we further evaluate the performance and limitations of LLM agents, compare the trade-offs among different GPU programming languages, and provide insights for future agent design. FlashInfer-Bench thus establishes a practical, reproducible pathway for continuously improving AI-generated kernels and deploying them into large-scale LLM inference.

cs.AI↗

arXiv2025

MPK: A Compiler and Runtime for Mega-Kernelizing Tensor Programs

We introduce Mirage Persistent Kernel (MPK), the first compiler and runtime system that automatically transforms multi-GPU model inference into a single high-performance mega-kernel. MPK introduces an SM-level graph representation that captures data dependencies at the granularity of individual streaming multiprocessors (SMs), enabling cross-operator software pipelining, \rev{fine-grained overlap of computation and communication, and other optimizations that are infeasible under the conventional kernel-per-operator execution model}. The MPK compiler lowers tensor programs into optimized SM-level task graphs and generates fast CUDA implementations for each task, while the MPK in-kernel parallel runtime executes these tasks within a single persistent mega-kernel using decentralized scheduling across SMs. Together, these components provide end-to-end kernel fusion with minimal developer effort, while preserving the flexibility of existing programming models. Our evaluation shows that MPK significantly outperforms existing kernel-per-operator LLM serving systems, achieving up to 1.7$\times$ lower end-to-end inference latency and pushing LLM inference performance close to the limits of the underlying hardware. MPK is publicly available at https://github.com/mirage-project/mirage.

cs.DC↗

arXiv2024

Exceptional topology in non-Hermitian twisted bilayer graphene

Twisted bilayer graphene (TBG) has extraordinary electronic properties at the magic angle along with an isolated flat band at the magic angle. However, the non-Hermitian phenomena in twisted bilayer graphene remain unexplored. In this work, we study a non-Hermitian TBG formed by one-layer graphene twisted relative to another layer with gain and loss. Using a non-Hermitian generalization of the Bistritzer-MacDonald model, we find Dirac cones centered at only the $K_M$ ($K'_M$) corner of the moir\'e Brillouin zone at the $K'$ ($K$) valley deform into rings of exceptional points in the presence of non-Hermiticity, which is different from single-layer graphene with gain and loss, where exceptional rings appear in both $K$ and $K'$ corners of the Brillouin zone. We show that the exceptional rings are protected by non-Hermitian chiral symmetry. More interestingly, at an ``exceptional magic angle" larger than the Hermitian magic angle, the exceptional rings coincide and form non-Hermitian flat bands with zero energy and a finite lifetime. These non-Hermitian flat bands in the moir\'e system, which are isolated from dispersive bands, are distinguished from those in non-Hermitian frustrated lattices. In addition, we find that the non-Hermitian flat band has topological charge conserved in the moir\'e Brillouin zone, which is allowed for analogs of non-Hermitian fractional quantum Hall states.

cond-mat.mes-hall↗

arXiv2024

Topological superconductivity in superconducting chiral topological semimetals with parallel spin-momentum locking

In contrast to conventional Weyl semimetals in achiral crystals, chiral topological semimetals in chiral crystals exhibit Weyl nodes at time-reversal-invariant momenta. A Fermi surface spin texture with parallel spin-momentum locking in these material has been observed by a recent experiment [Nat. Comm. 15,3720(2024)]. We find that the Weyl nodes location and the Fermi surface spin texture lead to gapped zero-momenta intranode superconductivity (SC), which is absent in achiral Weyl semimetals. Through self-consistent mean-field calculations, we find that a cubic lattice system in general favors a mixture of spin-singlet $s_\pm$ and $d+id$-wave pairings. In the presence of only the $s_\pm$-wave pairing, we identify a first-order time-reversal invariant topological SC phase. Notably, an SC phase with two Majorana cones for opened Fermi surfaces is energetically favorable. In addition, a second-order topological superconductor with chiral Majorana states can be realized in the presence of a mixture of $s\pm$- and $d+id$-wave pairing. We show that chiral topological semimetals in cubic lattice are fascinating platforms for exploring intrinsic unconventional superconductivity and topological superconductivity.

cond-mat.supr-con↗

arXiv2024

FlexLLM: Token-Level Co-Serving of LLM Inference and Finetuning with SLO Guarantees

Finetuning large language models (LLMs) is essential for task adaptation, yet today's serving stacks isolate inference and finetuning on separate GPU clusters -- wasting resources and under-utilizing hardware. We introduce FlexLLM, the first system to co-serve LLM inference and PEFT-based finetuning on shared GPUs by fusing computation at the token level. FlexLLM's static compilation optimizations -- dependent parallelization and graph pruning significantly shrink activation memory, leading to end-to-end GPU memory savings by up to 80%. At runtime, a novel token-level finetuning mechanism paired with a hybrid token scheduler dynamically interleaves inference and training tokens within each co-serving iteration, meeting strict latency SLOs while maximizing utilization. In end-to-end benchmarks on LLaMA-3.1-8B, Qwen-2.5-14B, and Qwen-2.5-32B, FlexLLM maintains inference SLO compliance at up to 20 req/s, and improves finetuning throughput by $1.9-4.8\times$ under heavy inference workloads and $2.5-6.8\times$ under light loads, preserving over 76% of peak finetuning progress even at peak demand. FlexLLM is publicly available at https://flexllm.github.io.

cs.DC↗

arXiv2023

Topological Floquet Flat Bands in Irradiated Alternating Twist Multilayer Graphene

We study the appearance of topological Floquet flat bands in alternating-twist multilayer graphene, which has alternating relative twist angle $\pmθ$ near the first magic angle. While the system hosts both flat bands and a steep Dirac cone in the static case, the circularly polarized laser beam can open a gap at the Moiré $K$ point and create Floquet flat bands carrying nonzero Chern numbers. Considering recent lattice-relaxation results, we find that the topological flat band is well-isolated for the effective interlayer tunneling in $n=3, 4, 5$ layers. Such dynamically produced topological flat bands are potentially observed in the experiment and thus provide a feasible way to realize the fractional Chern insulator.

cond-mat.mes-hall↗

arXiv2021

Charge-$4e$ superconductivity from nematic superconductors in 2D and 3D

Charge-$4e$ superconductivity as a novel phase of matter remains elusive so far. Here we show that charge-$4e$ phase can arise as a vestigial order above the nematic superconducting transition temperature in time-reversal-invariant nematic superconductors. On the one hand, the nontrivial topological defect -- nematic vortex -- is energetically favored over the superconducting phase vortex when the nematic stiffness is less than the superfluid stiffness; consequently the charge-$4e$ phase emerges by proliferation of nematic vortices upon increasing temperatures. On the other hand, the Ginzburg-Landau theory of the nematic superconductors has two distinct decoupling channels to either charge-$4e$ orders or nematic orders; by analyzing the competition between the effective mass of the charge-$4e$ order and the cubic potential of the nematic order, we find a sizable regime where the charge-$4e$ order is favored. These two analysis consistently show that nematic superconductors can provide a promising route to realize charge-$4e$ phases, which may apply to candidate nematic superconductors such as PbTaSe$_2$ and twisted bilayer graphene.

cond-mat.str-el↗

arXiv2021

Three-dimensional time reversal invariant topological superconductivity in doped chiral topological semimetals

Chiral topological semimetals host multifold degenerate band crossing points under the protection of crystalline symmetries. In this paper, we suggest that the recently discovered chiral topological semimetals in space group 198, parts of which are superconducting upon doping, can be new candidates of time reversal invariant topological superconductors. By investigating the Fermi surfaces around the band crossing points that carry nonzero Chern numbers, we clarify how the nontrivial topology of chiral topological semimetals affects their superconducting state and show the existence of topological superconductivity in $s_\pm$-wave pairing with surface Majorana fermions. We further demonstrate that the topological superconductivity is favored by the inter-unit-cell phonon-mediated electron-electron interaction.

cond-mat.mes-hall↗

arXiv2019

Scalable Majorana vortex modes in iron-based superconductors

A vortex in an s-wave superconductor with a surface Dirac cone can trap a Majorana bound state with zero energy leading to a zero-bias peak (ZBP) of tunneling conductance. The iron-based superconductor FeTe$_x$Se$_{1-x}$ is one of the material candidates hosting these Majorana vortex modes. It has been observed by recent scanning tunneling spectroscopy measurement that the fraction of vortex cores possessing ZBPs decreases with increasing magnetic field on the surface of this iron-based superconductor. We construct a three-dimensional tight-binding model simulating the physics of over a hundred Majorana vortex modes in FeTe$_x$Se$_{1-x}$ with realistic physical parameters. Our simulation shows that the Majorana hybridization and disordered vortex distribution can explain the decreasing fraction of the ZBPs observed in the experiment. Furthermore, we find the statistics of the energy peaks off zero energy in our simulation with the Majorana physics in agreement with the analyzed peak statistics in the vortex cores from the experiment. This agreement and the explanation of the decreasing ZBP fraction lead to an important indication of scalable Majorana vortex modes in the iron-based superconductor. Thus, FeTe$_x$Se$_{1-x}$ can be one promising platform possessing scalable Majorana qubits for quantum computing. In addition, we further show the interplay of the ZBP presence and the vortex locations qualitatively agrees with our additional experimental observation and predict the universal spin signature of the hybridized multiple Majorana vortex modes.

cond-mat.supr-con↗

arXiv2018

Quasiparticle gaps in multiprobe Majorana nanowires

We theoretically study a spin-orbit-coupled nanowire proximitized by a superconductor in the presence of an externally applied Zeeman field ("Majorana nanowire") with zero-energy Majorana bound states localized at the two ends of the wire when the Zeeman spin splitting is large enough for the system to enter the topological phase. The specific physics of interest in the current work is the effect of having several tunnel probes attached to the wire along its length. Such tunnel probes should allow, as a matter of principle, one to observe both the predicted bulk superconducting gap closing and opening associated with the topological quantum phase transition as well as the Majorana bound states at the wire ends showing up as zero-bias conductance peaks, depending on which probes are used for the tunneling spectroscopy measurement. Because of the possible invasive nature of the tunnel probes, producing local potential fluctuations in the nanowire, we find the physical situation to be quite complex. In particular, depending on the details of the tunnel barrier operational at the probes, the Majorana nanowire could manifest additional low-energy Andreev bound states which will manifest their own almost-zero-bias peaks, complicating the interpretation of the tunneling data in multiprobe Majorana nanowires. We use two complementary microscopic models to simulate the probes, finding that the tunneling conductance spectrum depends rather sensitively on the details of the tunnel barriers at the probes. We apply our general analysis to simulate a recent multiprobe nanowire experiment commenting on the nature of the quasiparticle gaps likely controlling the experimental observations.

cond-mat.mes-hall↗

arXiv2018

Metamorphosis of Andreev bound states into Majorana bound states in pristine nanowires

We show theoretically that in the generic finite chemical potential situation, the clean superconducting spin-orbit-coupled nanowire has two distinct nontopological regimes as a function of Zeeman splitting (below the topological quantum phase transition): one is characterized by finite-energy in-gap Andreev bound states, while the other has only extended bulk states. The Andreev bound state regime is characterized by strong features in the tunneling spectra creating a "gap closure" signature, but no "gap reopening" signature should be apparent above the topological quantum phase transition, in agreement with most recent experimental observations. The gap closure feature is actually the coming together of the Andreev bound states at high chemical potential rather than a simple trivial gap of extended bulk states closing at the transition. Our theoretical finding establishes the generic intrinsic Andreev bound states on the trivial side of the topological quantum phase transition as the main contributors to the tunneling conductance spectra, providing a generic interpretation of existing experiments in clean Majorana nanowires. Our work also explains why experimental tunnel conductance spectra generically have gap closing features below the topological quantum phase transition, but no gap opening features above it.

cond-mat.mes-hall↗

arXiv2018

Helical Majorana edge mode in a superconducting antiferromagnetic quantum spin Hall insulator

A two-dimensional time-reversal symmetric topological superconductor is a fully gapped system possessing a helical Majorana mode on the edges. This helical Majorana edge mode (HMEM), which is a Kramer's pair of two chiral Majorana edge modes in the opposite propagating directions, is robust under time-reversal symmetry protection. We propose a feasible setup and accessible measurement to provide the preliminary step of the HMEM realization by studying superconducting antiferromagnetic quantum spin Hall insulators. Since this antiferromagnetic topological insulator hosts a helical electron edge mode and preserves effective time-reversal symmetry, which is the combination of time-reversal symmetry and crystalline symmetry, the proximity effect of the conventional s-wave superconducting pairing can directly induce a single HMEM. We further show the HMEM leads to the observation of an $e^2/h$ conductance, and this quantized conductance survives even in the presence of small symmetry-breaking disorders.

cond-mat.supr-con↗

arXiv2018

Disorder-induced half-integer quantized conductance plateau in quantum anomalous Hall insulator-superconductor structures

A weak superconducting proximity effect in the vicinity of the topological transition of a quantum anomalous Hall system has been proposed as a venue to realize a topological superconductor (TSC) with chiral Majorana edge modes (CMEMs). A recent experiment [Science 357, 294 (2017)] claimed to have observed such CMEMs in the form of a half-integer quantized conductance plateau in the two-terminal transport measurement of a quantum anomalous Hall-superconductor junction. Although the presence of a superconducting proximity effect generically splits the quantum Hall transition into two phase transitions with a gapped TSC in between, in this Rapid Communication we propose that a nearly flat conductance plateau, similar to that expected from CMEMs, can also arise from the percolation of quantum Hall edges well before the onset of the TSC or at temperatures much above the TSC gap. Our Rapid Communication, therefore, suggests that, in order to confirm the TSC, it is necessary to supplement the observation of the half-quantized conductance plateau with a hard superconducting gap (which is unlikely for a disordered system) from the conductance measurements or the heat transport measurement of the transport gap. Alternatively, the half-quantized thermal conductance would also serve as a smoking-gun signature of the TSC.

cond-mat.supr-con↗

arXiv2015

Kibble-Zurek mechanism beyond adiabaticity: Finite-time scaling with critical initial slip

The Kibble-Zurek mechanism demands an initial adiabatic stage before an impulse stage to have a frozen correlation length that generates topological defects in a cooling phase transition. Here we study such a driven critical dynamics but with an initial condition that is near the critical point and that is far away from equilibrium. In this case, there is no initial adiabatic stage at all and thus adiabaticity is broken. However, we show that there again exists a finite length scale arising from the driving that divides the evolution into three stages. A relaxation--finite-time scaling--adiabatic scenario is then proposed in place of the adiabatic--impulse--adiabatic scenario of the original Kibble-Zurek mechanism. A unified scaling theory, which combines finite-time scaling with critical initial slip, is developed to describe the universal behavior and is confirmed with numerical simulations of a two-dimensional classical Ising model.

cond-mat.stat-mech↗

arXiv2014

Kibble-Zurek Mechanism and Finite-Time Scaling

The Kibble-Zurek (KZ) mechanism has been applied to a variety of systems ranging from low temperature Bose-Einstein condensations to grand unification scales in particle physics and cosmology and from classical phase transitions to quantum phase transitions. Here we show that finite-time scaling (FTS) provides a detailed improved understanding of the mechanism. In particular, the finite time scale, which is introduced by the external driving (or quenching) and results in FTS, is the origin of the division of the adiabatic regimes from the impulse regime in the KZ mechanism. The origin of the KZ scaling for the defect density, generated during the driving through a critical point, is not that the correlation length ceases growing in the nonadiabatic impulse regime, but rather, is that it is taken over by the effective finite length scale corresponding to the finite time scale. We also show that FTS accounts well for and improves the scaling ansatz proposed recently by Liu, Polkovnikov, and Sandvik [Phys. Rev. B {\bf 89}, 054307 (2014)]. Further, we show that their universal power-law scaling form applies only to some observables in cooling but not to heating. Even in cooling, it is invalid either when an appropriate external field is present. However, this finite-time-finite-size scaling calls for caution in application of FTS. Detailed scaling behaviors of the FTS and finite-size scaling, along with their crossover, are explicitly demonstrated, with the dynamic critical exponent $z$ being estimated for two- and three-dimensional Ising models under the usual Metropolis dynamics. These values of $z$ are found to give rise to better data collapses than the extant values do in most cases but take on different values in heating and cooling in both two- and three-dimensional spaces.

cond-mat.stat-mech↗