Streaming Tensor Programs: A Streaming Abstraction for Dynamic Parallelism

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Sohn, Gina, Zhang, Genghan, Hossfeld, Konstantin, Kim, Jungwoo, Sobotka, Nathan, Zhang, Nathan, Hsu, Olivia, Olukotun, Kunle
Natura: Preprint
Pubblicazione: 2025
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866915758373601280
author Sohn, Gina
Zhang, Genghan
Hossfeld, Konstantin
Kim, Jungwoo
Sobotka, Nathan
Zhang, Nathan
Hsu, Olivia
Olukotun, Kunle
author_facet Sohn, Gina
Zhang, Genghan
Hossfeld, Konstantin
Kim, Jungwoo
Sobotka, Nathan
Zhang, Nathan
Hsu, Olivia
Olukotun, Kunle
contents Dynamic behaviors are becoming prevalent in tensor applications, like machine learning, where many widely used models contain data-dependent tensor shapes and control flow. However, the limited expressiveness of prior programming abstractions for spatial dataflow accelerators (SDAs) forces these dynamic behaviors to be implemented statically and/or unoptimized. To address these challenges, we present Streaming Tensor Programs (STeP), a streaming abstraction that enables dynamic tensor workloads to run efficiently on SDAs. STeP introduces flexible routing operators, an explicit memory hierarchy, and symbolic-shape semantics that expose dynamic data rates and tensor dimensions. These capabilities unlock new optimizations, like dynamic tiling, dynamic parallelization, and configuration time-multiplexing, that adapt SDA execution to dynamic behaviors while preserving dataflow efficiency. Using a cycle-approximate simulator on representative LLM layers and a full model with real-world traces, STeP enables: dynamic tiling that breaks the Pareto-optimal frontier from prior work, dynamic parallelization that improves latency by ~2.72x, and configuration time-multiplexing that increases compute utilization by ~2.64x over prior SDA abstractions and their implementations.
format Preprint
id arxiv_https___arxiv_org_abs_2511_07776
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Streaming Tensor Programs: A Streaming Abstraction for Dynamic Parallelism
Sohn, Gina
Zhang, Genghan
Hossfeld, Konstantin
Kim, Jungwoo
Sobotka, Nathan
Zhang, Nathan
Hsu, Olivia
Olukotun, Kunle
Programming Languages
Hardware Architecture
Machine Learning
Dynamic behaviors are becoming prevalent in tensor applications, like machine learning, where many widely used models contain data-dependent tensor shapes and control flow. However, the limited expressiveness of prior programming abstractions for spatial dataflow accelerators (SDAs) forces these dynamic behaviors to be implemented statically and/or unoptimized. To address these challenges, we present Streaming Tensor Programs (STeP), a streaming abstraction that enables dynamic tensor workloads to run efficiently on SDAs. STeP introduces flexible routing operators, an explicit memory hierarchy, and symbolic-shape semantics that expose dynamic data rates and tensor dimensions. These capabilities unlock new optimizations, like dynamic tiling, dynamic parallelization, and configuration time-multiplexing, that adapt SDA execution to dynamic behaviors while preserving dataflow efficiency. Using a cycle-approximate simulator on representative LLM layers and a full model with real-world traces, STeP enables: dynamic tiling that breaks the Pareto-optimal frontier from prior work, dynamic parallelization that improves latency by ~2.72x, and configuration time-multiplexing that increases compute utilization by ~2.64x over prior SDA abstractions and their implementations.
title Streaming Tensor Programs: A Streaming Abstraction for Dynamic Parallelism
topic Programming Languages
Hardware Architecture
Machine Learning
url https://arxiv.org/abs/2511.07776