diff --git a/build_tools/build_backend.py b/build_tools/build_backend.py
new file mode 100644
index 0000000000..f340b47ba9
--- /dev/null
+++ b/build_tools/build_backend.py
@@ -0,0 +1,261 @@
+# Copyright (c) 2022-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+#
+# See LICENSE for license information.
+
+"""PEP 517 backend with CUDA-version-specific build requirements."""
+
+import os
+import re
+import shutil
+import subprocess
+from pathlib import Path
+from typing import List, Mapping, Optional, Union
+
+from setuptools import build_meta
+
+
+_SETUPTOOLS_BACKEND = build_meta.__legacy__
+_CUDA_BUILD_PACKAGES = {
+    "12": [
+        # Also contains nvvm and crt
+        "nvidia-cuda-nvcc-cu12",
+        "nvidia-cuda-runtime-cu12",
+        "nvidia-cuda-cccl-cu12",
+        "nvidia-cuda-profiler-api-cu12",
+        "nvidia-nvml-dev-cu12",
+    ],
+    "13": [
+        "nvidia-cuda-nvcc",
+        "nvidia-cuda-runtime",
+        "nvidia-cuda-crt",
+        "nvidia-cuda-cccl",
+        "nvidia-cuda-profiler-api",
+        "nvidia-nvml-dev",
+        "nvidia-nvvm",
+    ],
+}
+_DEFAULT_CUDA_VERSION = "13.3"
+
+ConfigValue = Union[str, List[str]]
+ConfigSettings = Optional[Mapping[str, ConfigValue]]
+
+
+def _setuptools_config_settings(config_settings: ConfigSettings) -> ConfigSettings:
+    """Remove settings that are private to this backend."""
+    if config_settings is None:
+        return None
+
+    settings = dict(config_settings)
+    settings.pop("cuda-version", None)
+    return settings or None
+
+
+def _normalize_cuda_version(value: str) -> str:
+    """Validate and normalize a CUDA major/minor version."""
+    value = value.lower().removeprefix("cu")
+    if re.fullmatch(r"\d+\.\d+", value) is None:
+        raise ValueError(
+            f"Invalid CUDA version {value!r}; expected <major>.<minor>, for example 12.8"
+        )
+
+    major = value.split(".", maxsplit=1)[0]
+    if major not in _CUDA_BUILD_PACKAGES:
+        supported = ", ".join(sorted(_CUDA_BUILD_PACKAGES))
+        raise ValueError(f"Unsupported CUDA major version {major!r}; expected one of: {supported}")
+
+    return value
+
+
+def _config_cuda_version(config_settings: ConfigSettings) -> Optional[str]:
+    """Get the CUDA version from PEP 517 config settings."""
+    settings: Mapping[str, ConfigValue] = config_settings or {}
+    value = settings.get("cuda-version")
+    if value is None:
+        return None
+
+    if isinstance(value, list):
+        if not value:
+            raise ValueError("CUDA package version cannot be empty")
+        value = value[-1]
+
+    return _normalize_cuda_version(str(value))
+
+
+def _environment_cuda_version() -> Optional[str]:
+    """Get the CUDA version from the environment."""
+    value = os.getenv("NVTE_CUDA_VERSION")
+    return _normalize_cuda_version(value) if value is not None else None
+
+
+def _system_cuda_version() -> Optional[str]:
+    """Get the version reported by a system-installed NVCC."""
+    candidates = []
+    if cuda_compiler := os.getenv("CUDACXX"):
+        candidates.append(Path(cuda_compiler))
+    if cuda_home := os.getenv("CUDA_HOME"):
+        candidates.append(Path(cuda_home) / "bin" / "nvcc")
+    if nvcc := shutil.which("nvcc"):
+        candidates.append(Path(nvcc))
+    candidates.append(Path("/usr/local/cuda/bin/nvcc"))
+
+    checked = set()
+    for candidate in candidates:
+        if candidate in checked or not candidate.is_file():
+            continue
+
+        checked.add(candidate)
+        try:
+            result = subprocess.run(
+                [candidate, "--version"],
+                capture_output=True,
+                check=True,
+                text=True,
+            )
+        except (OSError, subprocess.CalledProcessError):
+            continue
+
+        match = re.search(r"release\s+(\d+\.\d+)", result.stdout)
+        if match is not None:
+            return _normalize_cuda_version(match.group(1))
+
+    return None
+
+
+def _torch_cuda_version() -> Optional[str]:
+    """Get the CUDA version required by PyTorch, if available."""
+    try:
+        import torch
+    except (ImportError, OSError):
+        return None
+    value = getattr(getattr(torch, "version", None), "cuda", None)
+    return _normalize_cuda_version(value) if value else None
+
+
+def _framework_cuda_version() -> Optional[str]:
+    """Get a CUDA version required by the selected framework."""
+    frameworks = {
+        framework.strip().lower()
+        for framework in os.getenv("NVTE_FRAMEWORK", "").split(",")
+        if framework.strip()
+    }
+
+    if not frameworks or frameworks.intersection({"all", "pytorch"}):
+        if cuda_version := _torch_cuda_version():
+            return cuda_version
+
+    # JAX CUDA plugin names identify the CUDA major version, but not the
+    # major/minor toolkit release needed to select component packages.
+    return None
+
+
+def _cuda_version(config_settings: ConfigSettings) -> str:
+    """Resolve the CUDA version in descending order of precedence."""
+    return (
+        _config_cuda_version(config_settings)
+        or _environment_cuda_version()
+        or _system_cuda_version()
+        or _framework_cuda_version()
+        or _DEFAULT_CUDA_VERSION
+    )
+
+
+def _cuda_build_requirements(config_settings: ConfigSettings) -> List[str]:
+    """Get build requirements for the requested CUDA package version."""
+    cuda_version = _cuda_version(config_settings)
+    cuda_major = cuda_version.split(".", maxsplit=1)[0]
+    component_constraint = f"=={cuda_version}.*"
+
+    packages = _CUDA_BUILD_PACKAGES[cuda_major]
+    requirements = [f"{package}{component_constraint}" for package in packages]
+    requirements.append(f"nvidia-nccl-cu{cuda_major}>=2")
+    return requirements
+
+
+###################################################################################################
+# PEP 517 and PEP 660 defined functions
+###################################################################################################
+
+
+# Defined by PEP 517
+def get_requires_for_build_wheel(config_settings: ConfigSettings = None) -> List[str]:
+    """Get CUDA requirements for building a wheel."""
+    return _cuda_build_requirements(config_settings)
+
+
+# Defined by PEP 517
+def get_requires_for_build_sdist(config_settings: ConfigSettings = None) -> List[str]:
+    """Get CUDA requirements needed while evaluating setup.py for an sdist."""
+    # CUDA build requirements are needed even for sdist because setup.py currently always
+    # depends on these packages. It could be refactored to only depend on them when building
+    # a wheel, but that would require a more invasive refactor.
+    return _cuda_build_requirements(config_settings)
+
+
+# Defined by PEP 517
+def build_wheel(
+    wheel_directory: str,
+    config_settings: ConfigSettings = None,
+    metadata_directory: Optional[str] = None,
+) -> str:
+    """Build a wheel with setuptools."""
+    return _SETUPTOOLS_BACKEND.build_wheel(
+        wheel_directory,
+        _setuptools_config_settings(config_settings),
+        metadata_directory,
+    )
+
+
+# Defined by PEP 517
+def prepare_metadata_for_build_wheel(
+    metadata_directory: str,
+    config_settings: ConfigSettings = None,
+) -> str:
+    """Prepare wheel metadata with setuptools."""
+    return _SETUPTOOLS_BACKEND.prepare_metadata_for_build_wheel(
+        metadata_directory,
+        _setuptools_config_settings(config_settings),
+    )
+
+
+# Defined by PEP 517
+def build_sdist(
+    sdist_directory: str,
+    config_settings: ConfigSettings = None,
+) -> str:
+    """Build an sdist with setuptools."""
+    return _SETUPTOOLS_BACKEND.build_sdist(
+        sdist_directory,
+        _setuptools_config_settings(config_settings),
+    )
+
+
+if hasattr(_SETUPTOOLS_BACKEND, "build_editable"):
+
+    # Defined by PEP 660
+    def get_requires_for_build_editable(config_settings: ConfigSettings = None) -> List[str]:
+        """Get CUDA requirements for building an editable wheel."""
+        return _cuda_build_requirements(config_settings)
+
+    # Defined by PEP 660
+    def build_editable(
+        wheel_directory: str,
+        config_settings: ConfigSettings = None,
+        metadata_directory: Optional[str] = None,
+    ) -> str:
+        """Build an editable wheel with setuptools."""
+        return _SETUPTOOLS_BACKEND.build_editable(
+            wheel_directory,
+            _setuptools_config_settings(config_settings),
+            metadata_directory,
+        )
+
+    # Defined by PEP 660
+    def prepare_metadata_for_build_editable(
+        metadata_directory: str,
+        config_settings: ConfigSettings = None,
+    ) -> str:
+        """Prepare editable-wheel metadata with setuptools."""
+        return _SETUPTOOLS_BACKEND.prepare_metadata_for_build_editable(
+            metadata_directory,
+            _setuptools_config_settings(config_settings),
+        )
diff --git a/docs/envvars.rst b/docs/envvars.rst
index b3765a06bd..014720206b 100644
--- a/docs/envvars.rst
+++ b/docs/envvars.rst
@@ -78,6 +78,12 @@ Build Configuration
    :Default: ``10``
    :Description: Number of Philox4x32 rounds used by stochastic rounding kernels. Must be a positive integer.
 
+.. envvar:: NVTE_CUDA_VERSION
+
+   :Type: ``str``
+   :Default: Auto-detected based on system CUDA version, falling back to ``"13.3"``
+   :Description: CUDA version to build for. Valid values: ``"12.*"``, ``"13.*"``. If not set, automatically determined based on the installed CUDA Toolkit version. Not intended for external use.
+
 Optional Dependencies
 ^^^^^^^^^^^^^^^^^^^^^
 
diff --git a/pyproject.toml b/pyproject.toml
index 2c9f224c14..4f7278a7fc 100755
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -3,7 +3,8 @@
 # See LICENSE for license information.
 
 [build-system]
-requires = ["setuptools>=61.0", "cmake>=3.21", "wheel", "pybind11[global]", "ninja", "pip", "torch>=2.1", "jax>=0.5.0", "flax>=0.7.1", "nvidia-cudnn-frontend>=1.25.0"]
+# CUDA build-time requirements are defined in `build_tools/build_backend.py`.
 
-# Use legacy backend to import local packages in setup.py
-build-backend = "setuptools.build_meta:__legacy__"
+# Select build dependencies with --config-setting cuda-version=<version>.
+build-backend = "build_tools.build_backend"
+backend-path = ["."]