TensorRT-LLM Coding Guidelines

This document outlines the coding standards and guidelines for contributing to TensorRT-LLM. Following these guidelines ensures consistency, maintainability, and quality across the codebase.

C++ Coding Guidelines

The TensorRT-LLM C++ Coding Guidelines are mainly derived from the Google C++ Style Guide.

Namespaces

Closing braces of namespaces should have a comment saying the namespace it closes:

namespace foo
{
...
} // namespace foo

Constants

Prefer const or constexpr variables over #defines whenever possible, as the latter are not visible to the compiler.
A variable that is not modified after its initialization should be declared as const.
For naming of constants, see the Naming section.

Literals

Except 0 (only used in comparison for checking signness/existence/emptiness) and nullptr, true, false, all other literals should only be used for variable initialization. Example:

// Bad
if (nbInputs == 2U){/*...*/}

// Good
constexpr size_t kNbInputsWBias = 2U;
if (nbInputs == kNbInputsWBias) {/*...*/}

Brace Notation

Use the Allman indentation style.
Put the semicolon for an empty for or while loop in a new line.
The statement forming the body of a switch, while, do .. while or for statement shall be a compound statement (use brace-delimited statements).
if and else should always be followed by brace-delimited statements, even if empty or a single statement.

Naming

1. Filenames

Camel case with first letter lowercase: thisIsASubDir and thisIsAFilename.cpp
All files involved in the compilation of a compilation target (.exe/.so) must have filenames that are case-insensitive unique.

2. Types

All types (including class names) are camel case with uppercase first letter.
Example: FooBarClass

3. Local Variables, Methods and Namespaces

Camel case with first letter lowercase.
Example: localFooBar

4. Non-magic-number Global Variables (non-static, not in anonymous namespace)

Camel case prefixed by a lower case ‘g’.
Example: gDontUseGlobalFoos

5. Non-magic-number Global Variables (static or in anonymous namespace)

Camel case prefixed by a lower case ‘s’.
Example: sMutableStaticGlobal

6. Locally Visible Static Variable

Camel case with lowercase prefix ‘s’ as the first letter of the name.
Example: static std::once_flag sFlag;

7. Class Member Variables

Camelcase prefixed with an ‘m’: mNbFooValues.
Public member variables do not require the ‘m’ prefix but it is highly encouraged to use the prefix when needed to improve code clarity.

8. Constants

Enumerations, global constants, static constants at class-scope and function-scope magic-number/literal constants are uppercase snakecase with prefix ‘k’:

int const kDIGIT_NUM = 10;

Function-scope constants that are not magic numbers or literals are named like non-constant variables:

bool const pass = a && b;

9. Macros

See Constants section (preferred over #define).
If you must use macros, follow uppercase snakecase: FOO_VERSION

Naming Notes

In general don’t use hungarian notation, except for ‘apps hungarian’ in some cases such as ‘nb’ in a variable name to indicate count: mNbLayers
If a constructor’s parameter name foo conflicts with a public member name foo, add a trailing underscore to the parameter name: foo_.
Literal suffixes should be upper case. For example, use 1234L instead of 1234l.

Tabs vs Spaces

Use only spaces. Do not use tabs.
Indent 4 spaces at a time. This is enforced automatically if you format your code using our clang-format config.

Formatting

Use the LLVM clang-format tool for formatting your changes prior to submitting the PR.
Use a maximum of 120 characters per line. The auto formatting tool will wrap longer lines.
Exceptions to formatting violations must be justified on a per-case basis. Bypassing the formatting rules is discouraged, but can be achieved for exceptions as follows:

// clang-format off
// .. Unformatted code ..
// clang-format on

Pointers and Memory Allocation

Use smart pointers for allocating objects on the heap.
When picking a smart pointer, prefer unique_ptr for single resource ownership and shared_ptr for shared resource ownership. Use weak_ptr only in exceptional cases.
Do not use smart pointers that have been deprecated in C++11.

Comments

C++ comments are required. C comments are not allowed except for special cases (inline).
C++ style for single-line comments: // This is a single line comment
In function calls where parameters are not obvious from inspection, it can be helpful to use an inline C comment to document the parameter for readers:

doSomeOperation(/* checkForErrors = */ false);

If the comment is a full sentence, it should be capitalized and punctuated properly.
Follow Doxygen rules for documenting new class interfaces and function prototypes:
- For C++-style single-line comments use //!.
- For class members, use //!<.

//! This is a Doxygen comment
//! in C++ style

struct Foo
{
    int x; //!< This is a Doxygen comment for members
}

Preprocessor Directives

#define and #undef of macros should be done only at global namespace.
Avoid the use of #ifdef and #ifndef directives (except in the case of header include guards). Prefer to use #if defined(...) or #if !defined(...) instead:

#if defined(FOO) || defined(BAR)
void foo();
#endif // defined(FOO) || defined(BAR)

When nesting preprocessor directives, use indentation after the hash mark (#):

#if defined(FOO)
# if FOO == 0
#  define BAR 0
# elif FOO == 1
#  define BAR 5
# else
#  error "invalid FOO value"
# endif
#endif

Use a preprocessor guard:
- The guard name must have prefix TRTLLM_ followed by the filename, all in caps.
- For a header file named FooBarHello.h, name the symbol as TRTLLM_FOO_BAR_HELLO_H.
- Only use the file name to create the symbol.
- Do not use prefix with underscore or trailing underscore.

#ifndef TRTLLM_FOO_BAR_HELLO_H
#define TRTLLM_FOO_BAR_HELLO_H
// ...
#endif // TRTLLM_FOO_BAR_HELLO_H

Additional C++ Guidelines

Exceptions: Must not be thrown across library boundaries.
Casts: Use the least forceful cast necessary. Prefer static_cast over reinterpret_cast. Avoid dynamic_cast.
Expressions: Do not use assignment operator in subexpressions.
Switch Statements: Should be well-structured with proper breaks or throws.
Functions: Avoid large inline functions. Use anonymous namespaces instead of static for internal linkage.
Signed vs Unsigned: Use signed integers by default, except for bitmaps or when interfacing with libraries that expect unsigned.

Common Pitfalls

C headers: Use C++ headers instead (e.g., <cstdint> instead of <stdint.h>).
C library functions: Avoid when possible. Use brace initialization or std::fill_n() instead of memset().

Python Coding Guidelines

Python Standard

The code developed for TensorRT-LLM should conform to Python 3.8+.

Indentation

Indent code with 4 spaces. Do not use tabs.

Imports

Always maintain the namespace when importing, even if only one class or function from a module is used. Example:

# Bad
from package.subpackage.foo import SomeClass
SomeClass()

# Bad
import package
package.subpackage.foo.SomeClass()

# Good
from package.subpackage import foo
foo.SomeClass()

Naming

Identifier Format

Files: snake_case - some_file.py
Classes: PascalCase - class SomeClass
Functions and Methods: snake_case - def my_awesome_function():
Local Variables: snake_case - my_variable = ...
- Prefix k for variable names that start with a number: k_99th_percentile = ...
Global Variables: upper snake_case and prefix G - G_MY_GLOBAL = ...
Constants: upper snake_case - MY_CONSTANT = ...

Identifier Guidelines

Avoid shadowing variables declared in an outer scope.
Initialize all externally visible members of a class in the constructor.

Comments and Docstrings

For interfaces that may be used outside a file, prefer docstrings over comments.
Comments should be reserved for code within a function, or interfaces that are local to a file.

Docstring Syntax

Use the Google style, which can be parsed by Sphinx. Example:

class MyClass:
    """
    Description of the class.
    """
    def __init__(self):
        self.x = 5
        """<type>: Description of 'x'"""

y = 2
"""<type>: Description of 'y'"""

Pydantic Guidelines

When defining any user-facing configuration classes (particularly LlmArgs or any class used in its fields), always use Pydantic classes rather than dataclasses or vanilla classes.

Model Structure

Inherit from StrictBaseModel (which sets extra="forbid") to fail fast when users specify invalid field names
Use discriminated unions when a field needs to accept one of several possible config classes
Do not define __init__ methods - this bypasses Pydantic’s validation and type coercion. Instead:
- For validation logic, use @field_validator or @model_validator
- For post-validation initialization, use model_post_init()
- For custom construction patterns, use classmethods (e.g. from_yaml())

Field Definitions

Add descriptions to all user-facing fields via Field(description="..."). Avoid using comments for descriptions.
Avoid dict, object, Any as field types - use properly typed alternatives
Avoid defining mutable defaults directly; use default_factory instead:
- Good: Field(default_factory=list), Field(default_factory=dict)
- Bad: Field(default=[]), Field(default={})
Use Literal["value1", "value2"] instead of str when a field should only accept certain values
Prefer PositiveInt, NonNegativeInt, NonNegativeFloat, PositiveFloat, Field(gt=0), Field(ge=0) for numeric constraints
Use Field(min_length=1) to enforce minimum length of a list

Validation

Use @field_validator and @model_validator instead of manual validate() or is_valid() methods
Raise ValueError instead of using assertions
Co-locate validation logic within the class itself rather than in a parent class

Serialization

Avoid defining to_dict() methods - prefer Pydantic’s built-in model_dump():
- Good: MyModel.model_dump()
- Bad: MyModel.to_dict()
Avoid defining from_dict() / from_kwargs() methods - prefer constructing the class directly:
- Good: MyModel(**kwargs), MyModel(**my_dict)
- Bad: MyModel.from_dict(kwargs)

Error Handling

When using try-except blocks, limit the except to the smallest set of errors possible:

# Bad
try:
    open(path, "r").read()
except:
    print("Failed to open file")

# Good
try:
    open(path, "r").read()
except FileNotFoundError:
    print("Failed to open file")

When using try-except blocks to handle multiple possible variable types, keep the body of the try as small as possible:

# Bad
try:
    f.seek(0)
    f.read()
except AttributeError:
    ... # Not a file-like object

# Good
try:
    f.seek  # Do not call to minimize chance of unrelated failure
except AttributeError:
    ... # Not a file-like object
else:
    f.seek(0)
    f.read()

Avoid Reflection

Avoid using reflection when functionality can be easily achieved without reflection.

# Bad
def make_complex(*args):
    x, y = args
    return dict(**locals())

# Good
def make_complex(x, y):
    return {'x': x, 'y': y}

Pre-commit Hooks

We use pre-commit for automatic code formatting and validation. Install and set up pre-commit hooks:

pip install pre-commit
pre-commit install

Pre-commit will automatically run the following tools on every commit:

isort: Import sorting
yapf: Python code formatting
clang-format: C++ code formatting
cmake-format: CMake file formatting
autoflake: Remove unused imports
ruff: Python linting
codespell: Spell checking
mdformat: Markdown formatting

If any files are modified by pre-commit hooks, you will need to stage and commit them again.

Documentation Guidelines

CLI Options in Documentation

When documenting CLI commands for trtllm-serve, trtllm-bench, trtllm-eval, or similar tools, prefer using --config over --extra_llm_api_options for specifying configuration files. Example:

# Preferred
trtllm-serve --model <model_path> --config config.yaml

# Avoid
trtllm-serve --model <model_path> --extra_llm_api_options config.yaml

NVIDIA Copyright

All TensorRT-LLM Open Source Software code should contain an NVIDIA copyright header that includes the year of its latest meaningful modification. The following block of text should be prepended to the top of all files (.cpp, .h, .cu, .py, and other source files):

/*
 * Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 *     http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */

Update the copyright year to the current year when modifying existing files.

Summary

Following these coding guidelines ensures:

Consistency: Code is uniform across the entire codebase
Maintainability: Code is easier to read, understand, and modify
Quality: Automated tools catch common errors and enforce best practices
Collaboration: Team members can work together more effectively

For more information, refer to:

Contributing

Extending

Advanced

​TensorRT-LLM Coding Guidelines

​C++ Coding Guidelines

​Namespaces

​Constants

​Literals

​Brace Notation

​Naming

​1. Filenames

​2. Types

​3. Local Variables, Methods and Namespaces

​4. Non-magic-number Global Variables (non-static, not in anonymous namespace)

​5. Non-magic-number Global Variables (static or in anonymous namespace)

​6. Locally Visible Static Variable

​7. Class Member Variables

​8. Constants

​9. Macros

​Naming Notes

​Tabs vs Spaces

​Formatting

​Pointers and Memory Allocation

​Comments

​Preprocessor Directives

​Additional C++ Guidelines

​Common Pitfalls

​Python Coding Guidelines

​Python Standard

​Indentation

​Imports

​Naming

​Identifier Format

​Identifier Guidelines

​Comments and Docstrings

​Docstring Syntax

​Pydantic Guidelines

​Model Structure

​Field Definitions

​Validation

​Serialization

​Error Handling

​Avoid Reflection

​Pre-commit Hooks

​Documentation Guidelines

​CLI Options in Documentation

​NVIDIA Copyright

​Summary

Build docs developers (and LLMs) love

TensorRT-LLM Coding Guidelines

C++ Coding Guidelines

Namespaces

Constants

Literals

Brace Notation

Naming

1. Filenames

2. Types

3. Local Variables, Methods and Namespaces

4. Non-magic-number Global Variables (non-static, not in anonymous namespace)

5. Non-magic-number Global Variables (static or in anonymous namespace)

6. Locally Visible Static Variable

7. Class Member Variables

8. Constants

9. Macros

Naming Notes

Tabs vs Spaces

Formatting

Pointers and Memory Allocation

Comments

Preprocessor Directives

Additional C++ Guidelines

Common Pitfalls

Python Coding Guidelines

Python Standard

Indentation

Imports

Naming

Identifier Format

Identifier Guidelines

Comments and Docstrings

Docstring Syntax

Pydantic Guidelines

Model Structure

Field Definitions

Validation

Serialization

Error Handling

Avoid Reflection

Pre-commit Hooks

Documentation Guidelines

CLI Options in Documentation

NVIDIA Copyright

Summary