Draft:Apache TVM

• Comment: Red text, sentences missing citations, promotional language, lacks context for a general audience. Not sure if it sufficiently demonstrates notability. Caleb Stanford (talk) 05:48, 6 June 2025 (UTC)

• Comment: The topic may be notable but currently reads like an advertisement.
  Can you please revise the article to provide context to a general audience? The article focuses too much on architecture and applications. Why is this topic notable according to independent sources? Who has talked about it and what have they said?
  Also, the article is missing a lead. Include the "Apache TVM" section as a lead instead of as the first section.
  Please address the above and resubmit. Thanks! Caleb Stanford (talk) 02:01, 5 June 2025 (UTC)

---

Apache TVM

Developer(s)	Apache Software Foundation
Initial release	2017
Written in	Python (programming language), C++
Type	Compiler
Website	tvm.apache.org

Apache TVM (Tensor Virtual Machine) is an open source machine learning optimizing compiler. It supports machine learning models written in frameworks such as TensorFlow, PyTorch, and ONNX and compiles to target platforms including central processing units, graphics processing units, and field-programmable gate arrays.

Originally developed at the University of Washington, TVM has been adopted by several companies including Amazon Web Services , AMD, Nvidia , and Qualcomm.

History

TVM was developed as a research project at the University of Washington in 2017 by Tianqi Chen and the SAMPL group at the Paul G. Allen School of Computer Science and Engineering. It was published at the 2018 Symposium on Operating Systems Design and Implementation. The project entered the Apache Incubator in March 2019 as part of the process for becoming an official Apache Software Foundation project.

Architecture

TVM is composed of several major components:

Relay IR: Relay IR is a high-level functional intermediate representation (IR) to represent neural networks prior to low-level optimization and code generation. Introduced as a successor to NNVM IR, Relay encodes computation graphs as abstract syntax trees and includes features such as first-class functions, recursion, and a dependent-like type system that supports shape and tensor types.

Relay supports model transformation through a functional programming paradigm. It uses a Static single-assignment form in which expressions are bound to named variables. Relay includes a Python interface for building and manipulating computation graphs.

import tvm
from tvm import relay

# Define a simple function using Relay
def simple_addition(x, y):
    return relay.add(x, y)

# Create Relay variables for the function
x = relay.var("x", relay.TensorType((3, 3), dtype="float32"))
y = relay.var("y", relay.TensorType((3, 3), dtype="float32"))

# Call the function
add_fn = simple_addition(x, y)

Relay supports reverse-mode automatic differentiation by transforming functions to compute both output values values and corresponding partial derivatives by employing functional techniques such as dual numbers and dynamic closures for backpropagation. This enables differentiation for higher-order functions and models with complex control flow.

TE (Tensor Expression) language: A domain-specific language to represent low-level tensor computations. It includes transformations such as loop reordering, memory layout modifications, and parallel execution for performance optimization.

AutoTVM and Ansor: AutoTVM and Ansor are automated tuning systems used to optimize performance of tensor computations. AutoTVM uses machine learning models and statistical techniques to identify scheduling parameters that improve runtime efficiency across different hardware targets. Ansor extends this approach with search-based methods to explore a broader configuration space.

BYOC (Bring Your Own Codegen): A plugin mechanism enabling integration of external code generation backends. It supports the use of hardware-specific instruction sets, libraries, and custom optimizations for specialized or proprietary hardware platforms.

Applications

Apache TVM has been applied in embedded systems, data center inference workloads, and edge computing devices. Cloud providers and hardware vendors including AWS, AMD, ARM, and Qualcomm have adopted TVM for compiling deep learning workloads to run on specific hardware platforms.

In research contexts, TVM has been used for topics such as automatic scheduling, hardware-aware neural architecture search, and integration with compiler infrastructures such as LLVM and MLIR.

See also

• Halide (programming language)
• MLIR
• ONNX