Held in conjunction with PACT 2026
Monday, October 19, 2026 (Afternoon)
Illini Center, Chicago, Illinois, USA
Machine learning and AI are increasingly influencing how compilers, runtime systems, and computer architectures are designed, optimized, and used. Learned techniques can replace or complement hand-designed heuristics, guide complex optimization decisions, support adaptive execution, and help researchers and developers explore increasingly large design spaces.
At the same time, AI is changing the workflow through which compiler and systems technologies are developed. Learning-based tools can assist with performance analysis, optimization discovery, autotuning, debugging, and experimental exploration while remaining grounded in compiler analyses, system measurements, and correctness constraints.
The Learning-Augmented Compilers & Systems (LACS) workshop brings together researchers and practitioners working at the intersection of:
| Time | Speakers | Affiliation | Talk title and abstract |
|---|---|---|---|
| 1:30–1:35 PM | Opening remarks | ||
| 1:35–2:05 PM |
Nikhil Rout, Blaise Tine |
UCLA |
TinyVortex: A Learning-Guided Kernel Scheduling Engine for Configurable SIMT GPU Co-Design Space ExplorationAbstract: Tensor compilers typically optimize kernels for fixed hardware targets, overlooking opportunities for joint hardware–software design-space exploration. To this end, we present TinyVortex, an open-source fork of the tinygrad Tensor compiler targeting the configurable RISC-V Vortex GPGPU as the backend. TinyVortex enables end-to-end PyTorch-like Tensor program execution, Vortex C++ kernel generation, and performance evaluation across various GPU microarchitectural configurations across the SIMT hierarchy (threads/warps/issue-width/cores). Building on this infrastructure, we introduce a learning-guided autotuning framework that jointly explores kernel scheduling decisions and hardware parameters, which we currently prototype using cycle-level simulation. Preliminary evaluations demonstrate 1.13x and 2.97x geometric-mean speedups over default kernel rendering schedules for vector addition and SGEMM, respectively, highlighting the potential of configuration-aware compilation for GPU hardware–software co-design. |
| 2:05–2:35 PM |
Xingjian Ding, Tejas Ramesh, Jonathan Ramirez, Anthony Ramirez, Henry Valladares, Shuyin Jiao, Keren Zhou, Pengfei Su |
UC Merced George Mason University UC Merced UC Merced UC Merced NCSU George Mason University, OpenAI UC Merced |
ProfileWeaver: Cost-Aware Profiler AI Harness for Automated Performance Analysis and Optimization of HPC ApplicationsAbstract: We study optimizers for constrained layers that consume the gradient at the constrained output and update the underlying parameters directly. For a doubly stochastic routing block parameterized by Sinkhorn iterations, training under such an optimizer leaves the score-to-router backward graph with no registered consumer. The runtime cannot discover this, because which gradients each optimizer consumes is recorded nowhere in the graph. We propose gradient-consumer declarations and dead-Jacobian elimination, which let the runtime skip building backward graphs that nothing will read. On our routing workload, the declaration cuts retained autograd state from 82 tensors to one per router; under torch.compile, the differentiated router remains 2.2× slower than the primal path and still retains its backward state. With the optimizer held identical and parameter updates numerically unchanged, eliminating the dead graph makes a full training step 6.7% faster. The optimizer’s own solves are small and dispatch-bound; batching them across routers and compiling the batched program shifts the lowest-latency execution from CPU to GPU as the number of routers grows, and a two-parameter cost model fitted per execution selects the fastest one at every held-out batch size. |
| 3:00–3:30 PM | Coffee break | ||
| 3:30–4:00 PM |
Chengyue Wang, Jason Cong |
UCLA |
Compiling Layout Transformations for NPUs and FPGAsAbstract:
Efficient execution on NPUs and FPGAs requires reconciling producer and consumer access patterns while sustaining parallelism and avoiding memory-bank conflicts. This talk presents our work on compiling layout transformations, from bit-level stream permutations to closed-form mixed-radix swizzles, with applications to streaming FFTs and NPU data movement.
|
| 4:00–4:30 PM | William Moses | UIUC, Google Deepmind |
Practical Perspectives on AI-Driven Software EngineeringAbstract: Abstract forthcoming. |
| 4:30–5:00 PM | Mary Hall | University of Utah |
What are the Right Abstractions to Support AI Collaboration on Performance Optimization?Abstract: Abstract forthcoming. |
Venue: Illini Center (University of Illinois)
Date: Monday, October 19 (afternoon)
LACS strongly encourages in-person presentations to facilitate discussion, networking, and engagement among participants. However, we recognize that travel may not be possible for every presenter.
In exceptional circumstances, we may consider remote presentations via a video-conferencing platform (for example, Zoom or Microsoft Teams). Please contact the workshop organizers as early as possible to discuss this option.
PACT 2026 is pleased to offer competitive travel support for students and early-career researchers attending the conference and its affiliated workshops, including LACS.
Travel support is competitive and subject to available funding. We encourage undergraduate students, graduate students, and early-career researchers—defined as researchers within three years of receiving their Ph.D.—to apply.
Application deadline: September 30, 2026
Apply here:
PACT 2026 Travel Support
Attendees who require a visa to travel to the United States should visit the PACT 2026 website for information on requesting a visa-support or invitation letter. Please begin this process well in advance, as visa processing times can vary.
Additional formatting and presentation instructions will be provided through the submission site.
LACS continues the legacy of:
MLSH – International Workshop on Machine Learning for Software Hardware Co-Design