Design and Architecture

This document is intended for developers who want to understand the architecture of Apache TVM and/or actively develop on the project. This page is organized as follows:

• The Overall Flow gives an overview of the steps that TVM takes to turn a high level description of a model into a deployable module. To get started, please read this section first.

• Brief introduction to the key components of the TVM stack. Feel free to also check out the TensorIR Deep Dive and Relax Deep Dive for more details about the two major components in the TVM stack.

This guide provides a few complementary views of the architecture. First, we review a single end-to-end compilation flow and discuss the key data structures and the transformations. This runtime-based view focuses on the interactions of each components when running the compiler. Then we will review the logical modules of the codebase and their relationship. This part provides a static overarching view of the design.

Overall Flow

In this guide, we will study an example compilation flow in the compiler. The figure below shows the flow. At a high-level, it contains several steps:

• Model Creation: Create the IRModule to be optimized and compiled, which contains a collection of functions that internally represent the model. Users can manually construct IRModule via NNModule, TVMScript, or import a pre-trained model from Relax frontend.

• Transformation: The compiler transforms an IRModule to another functionally equivalent or approximately equivalent(e.g. in the case of quantization) IRModule. Many of the transformations are target (backend) independent. We also allow target to affect the configuration of the transformation pipeline.

• Target Translation: The compiler translates(codegen) the IRModule to an executable format specified by the target. The target translation result is encapsulated as a runtime.Module that can be exported, loaded, and executed on the target runtime environment.

• Runtime Execution: the user loads back a runtime.Module and runs the compiled functions in the supported runtime environment.

Key data structures

One of the best ways to design and understand a complex system is to identify the key data structures and APIs that manipulate (transform) these data structures. Once we identified the key data structures, we can then breakdown a system into logical components that either define a collection of key data structures or transformations among the data structures.

IRModule is the primary data structure used across the entire stack. An IRModule (intermediate representation module) contains a collection of functions. Currently, we support two primary variants of functions.

• relax::Function is a high-level functional program representation. A relax.Function represents high-level graph structure, usually corresponds to an end-to-end model or a sub-graph of the overall model. You can view a relax.Function as a computational graph with additional support for control-flow, and complex data structures.

• tirx::PrimFunc is a low-level program representation that contains elements including loop-nest choices, multi-dimensional load/store, threading, and vector/tensor instructions. It is usually used to represent an operator program that executes a (possibly-fused) layer in a model.

During the compilation and transformation, all relax operators are lowered to tirx::PrimFunc or TVM PackedFunc, which can be executed directly on the target device, while the calls to relax operators are lowered to calls to low-level functions (e.g. R.call_tir or R.call_dps).

Transformations

Now that we have covered the key data structures, let us talk about the transformations. Each transformation could serve one of the following purposes:

• optimization: transform a program to an equivalent, possibly more optimized version.

• lowering: transform a program to a lower-level representation that is closer to the target.

relax transformations

relax transformations contain a collection of passes that apply to relax functions. The optimizations include common graph-level optimizations such as constant folding and dead-code elimination for operators, and backend-specific optimizations such as library dispatch.

TensorIR transformations

• TensorIR schedule: TensorIR schedules are designed to optimize the TensorIR functions for a specific target, with user-guided instructions and control how the target code is generated. For CPU targets, a TensorIR PrimFunc can generate valid code and execute on the target device without schedule but with very-low performance. However, for GPU targets, the schedule is essential for generating valid code with thread bindings. For more details, please refer to the TensorIR Transformation section. Additionally, we provides MetaSchedule to automate the search of TensorIR schedule.

• Lowering Passes: These passes usually perform after the schedule is applied, transforming a TensorIR PrimFunc into another functionally equivalent PrimFunc, but closer to the target-specific representation. For example, there are passes to flatten multi-dimensional access to one-dimensional pointer access, to expand the intrinsics into target-specific ones, and to decorate the function entry to meet the runtime calling convention.

Many low-level optimizations can be handled in the target phase by the LLVM, CUDA C, and other target compilers. As a result, we leave low-level optimizations such as register allocation to the downstream compilers and only focus on optimizations that are not covered by them.

cross-level transformations

Apache TVM enables cross-level optimization of end-to-end models. As the IRModule includes both Relax and TensorIR functions, the cross-level transformations are designed to mutate the IRModule by applying different transformations to these two types of functions.

For example, relax.LegalizeOps pass mutates the IRModule by lowering relax operators, adding corresponding TensorIR PrimFunc into the IRModule, and replacing the relax operators with calls to the lowered TensorIR PrimFunc. Another example is operator fusion pipeline in relax (including relax.FuseOps and relax.FuseTIR), which fuses multiple consecutive tensor operations into one. Different from the previous implementations, relax fusion pipeline analyzes the pattern of TensorIR functions and detects the best fusion rules automatically rather than human-defined operator fusion patterns.

Target Translation

The target translation phase transforms an IRModule to the corresponding target executable format. For backends such as x86 and ARM, we use the LLVM IRBuilder to build in-memory LLVM IR. We can also generate source-level languages such as CUDA C and OpenCL. Finally, we support direct translations of a Relax function (sub-graph) to specific targets via external code generators. It is important that the final code generation phase is as lightweight as possible. Vast majority of transformations and lowering should be performed before the target translation phase.

We also provide a Target structure to specify the compilation target. The transformations before the target translation phase can also be affected by the target — for example, a target’s vector length would change the vectorization behavior.

Runtime Execution

The main goal of TVM’s runtime is to provide a minimal API for loading and executing the compiled artifact in a language of their choice, including Python, C++, Rust, Go, Java, and JavaScript. The code snippet below shows such an example in Python:

import tvm
# Example runtime execution program in python, with type annotated
mod: tvm.runtime.Module = tvm.runtime.load_module("compiled_artifact.so")
arr: tvm.runtime.Tensor = tvm.runtime.tensor([1, 2, 3], device=tvm.cuda(0))
fun: tvm.runtime.PackedFunc = mod["addone"]
fun(arr)
print(arr.numpy())

tvm.runtime.Module encapsulates the result of compilation. A runtime.Module contains a GetFunction method to obtain PackedFuncs by name.

tvm.runtime.PackedFunc is a type-erased function interface for both the generated functions. A runtime.PackedFunc can take arguments and return values with the following types: POD types(int, float), string, runtime.PackedFunc, runtime.Module, runtime.Tensor, and other sub-classes of runtime.Object.

tvm.runtime.Module and tvm.runtime.PackedFunc are powerful mechanisms to modularize the runtime. For example, to get the above addone function on CUDA, we can use LLVM to generate the host-side code to compute the launching parameters(e.g. size of the thread groups) and then call into another PackedFunc from a CUDAModule that is backed by the CUDA driver API. The same mechanism can be used for OpenCL kernels.

The above example only deals with a simple addone function. The code snippet below gives an example of an end-to-end model execution using the same interface:

import tvm
# Example runtime execution program in python, with types annotated
factory: tvm.runtime.Module = tvm.runtime.load_module("resnet18.so")
# Create a stateful graph execution module for resnet18 on cuda(0)
gmod: tvm.runtime.Module = factory["resnet18"](tvm.cuda(0))
data: tvm.runtime.Tensor = get_input_data()
# set input
gmod["set_input"](0, data)
# execute the model
gmod["run"]()
# get the output
result = gmod["get_output"](0).numpy()

The main take away is that runtime.Module and runtime.PackedFunc are sufficient to encapsulate both operator level programs (such as addone), as well as the end-to-end models.

Summary and Discussions

In summary, the key data structures in the compilation flows are:

• IRModule: contains relax.Function and tirx.PrimFunc

• runtime.Module: contains runtime.PackedFunc

Most parts of the compilation are transformations among the key data structures.

• relax/transform and tirx/transform are deterministic rule-based transformations

• meta-schedule contains the search-based transformations

Finally, the compilation flow example is only a typical use-case of the TVM stack. We expose these key data structures and transformations to python and C++ APIs. As a result, you can use TVM just like the way you use numpy, except that the data structure of interest changes from the numpy.ndarray to tvm.IRModule. Here are some example use-cases:

• Directly construct IRModule using the python API.

• Compose a custom set of transformations(e.g. customize quantization).

• Manipulate the IR directly using TVM’s python API.

tvm/support

The support module contains the most common utilities for the infrastructure, such as generic arena allocator, socket, and logging.

tvm/runtime

The runtime serves as the foundation of the TVM stack. It provides the mechanism to load and execute compiled artifacts. The runtime defines a stable standard set of C APIs to interface with frontend languages such as Python and Rust.

runtime::Object is one of the primary data structures in TVM runtime besides the ffi::Function. It is a reference-counted base class with a type index to support runtime type checking and downcasting. The object system allows the developer to introduce new data structures to the runtime, such as Array, Map, and new IR data structures.

Besides deployment use-cases, the compiler itself also makes heavy use of TVM’s runtime mechanism. All of the IR data structures are subclasses of runtime::Object, as a result, they can be directly accessed and manipulated from the Python frontend. We use the PackedFunc mechanism to expose various APIs to the frontend.

Runtime support for different hardware backends are defined in subdirectories of runtime(e.g. runtime/opencl). These hardware-specific runtime modules define APIs for device memory allocation and device function serialization.

runtime/rpc implements an RPC support for PackedFunc. We can use the RPC mechanism to send a cross-compiled library to a remote device and benchmark the execution performance. The rpc infrastructure enables data collection from a wide range of hardware backends for learning-based optimizations.

• TVM Runtime System
• Runtime-Specific Information

• Introduction to Module Serialization
• Device/Target Interactions

Relax defines what to compute — it is a graph-level IR that describes the operators and dataflow of a model. The Relax Virtual Machine (VM) handles how to run it — it is the runtime component that executes the compiled result. During compilation, tvm.compile() invokes VMCodeGen to translate Relax functions into a compact bytecode representation. The resulting VMExecutable bundles the bytecode together with a constant pool and per-function metadata, and can be serialized to disk for deployment.

The VM uses a register-based interpreter with an intentionally minimal instruction set — only four opcodes: Call, Ret, Goto, and If. The VM itself performs no mathematical computation; it only orchestrates control flow (function calls, conditional branches, loops). The actual compute-intensive work — matrix multiplications, convolutions, and other operators — is carried out by TIR functions that have been compiled down to native GPU/CPU kernels, or by external libraries such as cuBLAS and cuDNN. The VM dispatches to them through the PackedFunc mechanism. Internally the VM recognizes three function kinds: PackedFunc for external C/C++ functions, VMFunc for bytecode-interpreted Relax functions, and VMTIRFunc for compiled TIR kernels.

On the Python side, users interact with the VM through relax.VirtualMachine(executable, device), which provides both a direct invocation interface and a stateful set-input / invoke / get-output interface suitable for RPC-based remote execution.

Disco is TVM’s distributed runtime for executing models across multiple devices. When a model is too large to fit on a single GPU, the relax.distributed module annotates how tensors should be partitioned and placed across a mesh of devices at compile time. Disco then takes over at runtime: it manages a group of workers, dispatches the compiled program to all of them simultaneously, and coordinates inter-device communication through collective operations such as allreduce, allgather, broadcast, and scatter.

The central abstraction is the Session, which owns the workers and exposes a SPMD-style programming interface. Every object that lives on workers is represented by a DRef — a distributed reference that maps to a concrete value on each worker. When the controller invokes a DPackedFunc through the session, all workers execute the same PackedFunc call synchronously, each operating on its own local shard. Compiled VM modules can be loaded into a session as DModule objects and called in the same fashion. The session also provides collective primitives backed by NCCL or RCCL, so that workers can exchange partial results without routing data through the controller.

Three session backends cover different deployment topologies. ThreadedSession spawns workers as threads within a single process — this is the most common choice for multi-GPU inference on a single machine. ProcessSession launches workers as separate OS processes connected by pipes, providing stronger isolation. SocketSession extends the model to multi-node clusters by connecting workers across machines via TCP sockets.

tvm/node

The node module adds additional features on top of the runtime::Object for IR data structures. The main features include reflection, serialization, structural equivalence, and hashing.

Thanks to the node module, we can directly access any field of the TVM’s IRNode by their name in Python.

x = tvm.tirx.Var("x", "int32")
y = tvm.tirx.Add(x, x)
# a and b are fields of a tirx.Add node
# we can directly use the field name to access the IR structures
assert y.a == x

We can also serialize arbitrary IR node into a JSON format, and load them back. The ability to save/store, and inspect an IR node provides a foundation for making the compiler more accessible.

tvm/ir

The tvm/ir folder contains the unified data structure and interfaces across all IR function variants. The components in tvm/ir are shared by tvm/relax and tvm/tirx, notable ones include

• IRModule

• Type

• PassContext and Pass

• Op

Different variants of functions(e.g. relax.Function and tirx.PrimFunc) can co-exist in an IRModule. While these variants may not have the same content representation, they use the same data structure to represent types. As a consequence, we use the same data structure to represent function (type) signatures of these variants. The unified type system allows one function variant to call another function once we clearly define the calling convention. This opens doors for future cross-function-variant optimizations.

We also provide a unified PassContext for configuring the pass behavior, and common composite passes to execute a pass pipeline. The following code snippet gives an example of PassContext configuration.

# configure the behavior of the tirx.UnrollLoop pass
with tvm.transform.PassContext(config={"tirx.UnrollLoop": { "auto_max_step": 10 }}):
    # code affected by the pass context

Op is the common class to represent all system-defined primitive operator/intrinsics. Developers can register new Ops as well as their additional attributes(e.g. whether the Op is elementwise) to the system.

• Pass Infrastructure

tvm/target

The target module contains all the code generators that translate an IRModule to a target runtime.Module. It also provides a common Target class that describes the target.

Targets can be constructed from a registered tag, a configuration dictionary, or a tag with attribute overrides:

from tvm.target import Target

# From a registered tag
target = Target("nvidia/nvidia-a100")

# From a config dictionary
target = Target({"kind": "cuda", "arch": "sm_80"})

# From a tag with attribute overrides
target = Target({"tag": "nvidia/nvidia-a100", "l2_cache_size_bytes": 12345})

Use Target.list_kinds() to see all available target kinds, and target.attrs to inspect target attributes.

The compilation pipeline can be customized according to the target by querying the attribute information in the target and builtin information registered to each target id(cuda, opencl).

• Device/Target Interactions

tvm/relax

Relax is the high-level IR used to represent the computational graph of a model. Various optimizations are defined in relax.transform. Note that Relax usually works closely with the TensorIR IRModule, most of the transformations are applied on both Relax and TensorIR functions in the IRModule. Please refer to the Relax Deep Dive for more details.

tvm/tirx

tirx contains the core IR definitions and lowering infrastructure for TensorIR (split from the former tir module). tirx::PrimFunc represents low-level tensor functions that can be transformed by tirx passes.

The tirx module includes:

• IR data structures (PrimFunc, Buffer, SBlock, expressions, statements).

• Analysis passes in tirx/analysis.

• Transformation and lowering passes in tirx/transform.

tvm/s_tir

s_tir (Schedulable TIR, split from the former tir module) contains schedule primitives and auto-tuning tools that operate on tirx::PrimFunc:

• Schedule primitives to control code generation (tiling, vectorization, thread binding) in s_tir/schedule.

• Builtin tensor intrinsics in s_tir/tensor_intrin.

• MetaSchedule for automated performance tuning.

• DLight for pre-defined, high-performance schedules.

Please refer to the TensorIR Deep Dive for more details.

tvm/arith

This module is closely tied to TensorIR. One of the key problems in the low-level code generation is the analysis of the indices’ arithmetic properties — the positiveness, variable bound, and the integer set that describes the iterator space. arith module provides a collection of tools that do (primarily integer) analysis. A TensorIR pass can use these analyses to simplify and optimize the code.

tvm/te and tvm/topi

TE stands for Tensor Expression. TE is a domain-specific language (DSL) for describing tensor computations. Importantly, a tensor expression itself is not a self-contained function that can be stored into IRModule. We can use te.create_prim_func to convert a tensor expression to a tirx::PrimFunc and then integrate it into the IRModule.

While possible to construct operators directly via TensorIR or tensor expressions (TE) for each use case, it is tedious to do so. topi (Tensor operator inventory) provides a set of pre-defined operators defined by numpy and found in common deep learning workloads.

tvm/s_tir/meta_schedule

MetaSchedule is a system for automated search-based program optimization, and can be used to optimize TensorIR schedules. Note that MetaSchedule only works with static-shape workloads.

tvm/s_tir/dlight

DLight is a set of pre-defined, easy-to-use, and performant s_tir schedules. DLight aims:

• Fully support dynamic shape workloads.

• Light weight. DLight schedules provides tuning-free schedule with reasonable performance.

• Robust. DLight schedules are designed to be robust and general-purpose for a single rule. And if the rule is not applicable, DLight not raise any error and switch to the next rule automatically.