Frameworks for writing, compiling, and optimizing deep learning (DL) models have recently enabled progress in areas like computer vision and natural language processing. Extending these frameworks to accommodate the rapidly diversifying landscape of DL models and hardware platforms presents challenging tradeoffs between expressivity, composability, and portability. We present Relay, a new compiler framework for DL. Relay's functional, statically typed intermediate representation (IR) unifies and generalizes existing DL IRs to express state-of-the-art models. The introduction of Relay's expressive IR requires careful design of domain-specific optimizations, addressed via Relay's extension mechanisms. Using these extension mechanisms, Relay supports a unified compiler that can target a variety of hardware platforms. Our evaluation demonstrates Relay's competitive performance for a broad class of models and devices (CPUs, GPUs, and emerging accelerators). Relay's design demonstrates how a unified IR can provide expressivity, composability, and portability without compromising performance.

Programmers can use gradual types to migrate programs to have more precise type annotations and thereby improve their readability, efficiency, and safety. Such migration requires an exploration of the migration space and can benefit from tool support, as shown in previous work. Our goal is to provide a foundation for better tool support by settling decidability questions about migration with gradual types. We present three algorithms and a hardness result for deciding key properties and we explain how they can be useful during an exploration. In particular, we show how to decide whether the migration space is finite, whether it has a top element, and whether it is a singleton. We also show that deciding whether it has a maximal element is NP-hard. Our implementation of our algorithms worked as expected on a suite of microbenchmarks.

Gradual typing systems ensure type soundness by transforming static type annotations into run-time checks. These checks provide semantic guarantees, but may come at a large cost in performance. In particular, recent work by Takikawa et al. suggests that enforcing a conventional form of type soundness may slow a program by two orders of magnitude.