Add cuda-minver option to the general section

build2cmake/src/config/v2.rs

@@ -9,6 +9,8 @@ use eyre::{bail, Result};
 use itertools::Itertools;
 use serde::{Deserialize, Serialize};
 
+use crate::version::Version;
+
 use super::v1::{self, Language};
 
 #[derive(Debug, Deserialize, Serialize)]
@@ -44,6 +46,8 @@ pub struct General {
     pub name: String,
     #[serde(default)]
     pub universal: bool,
+
+    pub cuda_minver: Option<Version>,
 }
 
 #[derive(Debug, Deserialize, Clone, Serialize)]
@@ -207,6 +211,7 @@ impl General {
         Self {
             name: general.name,
             universal,
+            cuda_minver: None,
         }
     }
 }

build2cmake/src/main.rs

@@ -17,6 +17,8 @@ use config::{Backend, Build, BuildCompat};
 mod fileset;
 use fileset::FileSet;
 
+mod version;
+
 #[derive(Parser, Debug)]
 #[command(version, about, long_about = None)]
 struct Cli {

build2cmake/src/templates/cuda/preamble.cmake

@@ -36,6 +36,13 @@ endif()
 
 if (NOT HIP_FOUND AND CUDA_FOUND)
   set(GPU_LANG "CUDA")
+
+  {% if cuda_minver %}
+    if (CUDA_VERSION VERSION_LESS {{ cuda_minver }})
+      message(FATAL_ERROR "CUDA version ${CUDA_VERSION} is too old. "
+        "Minimum required version is {{ cuda_minver }}.")
+    endif()
+  {% endif %}
 elseif(HIP_FOUND)
   set(GPU_LANG "HIP")
 

build2cmake/src/torch/cuda.rs

@@ -8,6 +8,7 @@ use minijinja::{context, Environment};
 
 use super::kernel_ops_identifier;
 use crate::config::{Backend, Build, Dependencies, Kernel, Torch};
+use crate::version::Version;
 use crate::FileSet;
 
 static CMAKE_UTILS: &str = include_str!("../templates/utils.cmake");
@@ -163,7 +164,7 @@ fn write_cmake(
 
     let cmake_writer = file_set.entry("CMakeLists.txt");
 
-    render_preamble(env, name, cmake_writer)?;
+    render_preamble(env, name, build.general.cuda_minver.as_ref(), cmake_writer)?;
 
     render_deps(env, build, cmake_writer)?;
 
@@ -338,12 +339,18 @@ pub fn render_extension(env: &Environment, ops_name: &str, write: &mut impl Writ
     Ok(())
 }
 
-pub fn render_preamble(env: &Environment, name: &str, write: &mut impl Write) -> Result<()> {
+pub fn render_preamble(
+    env: &Environment,
+    name: &str,
+    cuda_minver: Option<&Version>,
+    write: &mut impl Write,
+) -> Result<()> {
     env.get_template("cuda/preamble.cmake")
         .wrap_err("Cannot get CMake prelude template")?
         .render_to_write(
             context! {
                 name => name,
+                cuda_minver => cuda_minver.map(|v| v.to_string()),
                 cuda_supported_archs => cuda_supported_archs(),
 
             },

build2cmake/src/version.rs

@@ -0,0 +1,68 @@
+use std::{fmt::Display, str::FromStr};
+
+use eyre::{ensure, Context};
+use itertools::Itertools;
+use serde::{de, Deserialize, Deserializer, Serialize, Serializer};
+
+#[derive(Clone, Debug, Eq, Hash, Ord, PartialEq, PartialOrd)]
+pub struct Version(Vec<usize>);
+
+impl<'de> Deserialize<'de> for Version {
+    fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
+    where
+        D: Deserializer<'de>,
+    {
+        let s = String::deserialize(deserializer)?;
+        FromStr::from_str(&s).map_err(de::Error::custom)
+    }
+}
+
+impl Serialize for Version {
+    fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
+    where
+        S: Serializer,
+    {
+        serializer.serialize_str(&self.0.iter().map(|v| v.to_string()).join("."))
+    }
+}
+
+impl Display for Version {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(
+            f,
+            "{}",
+            itertools::join(self.0.iter().map(|v| v.to_string()), ".")
+        )
+    }
+}
+
+impl From<Vec<usize>> for Version {
+    fn from(value: Vec<usize>) -> Self {
+        // Remove trailing zeros for normalization.
+        let mut normalized = value
+            .into_iter()
+            .rev()
+            .skip_while(|&x| x == 0)
+            .collect::<Vec<_>>();
+        normalized.reverse();
+        Version(normalized)
+    }
+}
+
+impl FromStr for Version {
+    type Err = eyre::Report;
+
+    fn from_str(version: &str) -> Result<Self, Self::Err> {
+        let version = version.trim().to_owned();
+        ensure!(!version.is_empty(), "Empty version string");
+        let mut version_parts = Vec::new();
+        for part in version.split('.') {
+            let version_part: usize = part
+                .parse()
+                .context(format!("Version must consist of numbers: {}", version))?;
+            version_parts.push(version_part);
+        }
+
+        Ok(Version::from(version_parts))
+    }
+}

docs/writing-kernels.md

@@ -96,6 +96,11 @@ depends = [ "torch" ]
   Universal kernels do not use the other sections described below.
   A good example of a universal kernel is a Triton kernel.
   Default: `false`
+- `cuda-minver`: the minimum required CUDA toolkit version. This option
+  _must not_ be set under normal circumstances, since it can exclude Torch
+  build variants that are [required for compliant kernels](https://github.com/huggingface/kernels/blob/main/docs/kernel-requirements.md).
+  This option is provided for kernels that require functionality only
+  provided by newer CUDA toolkits.
 
 ### `torch`
 

lib/build.nix

@@ -62,12 +62,20 @@ rec {
     buildConfig: buildSets:
     let
       backends' = backends buildConfig;
+      requiredCuda = buildConfig.general.cuda-minver or "11.8";
       supportedBuildSet =
         buildSet:
-        (buildSet.gpu == "cuda" && backends'.cuda)
-        || (buildSet.gpu == "rocm" && backends'.rocm)
-        || (buildSet.gpu == "metal" && backends'.metal)
-        || (buildConfig.general.universal or false);
+        let
+          backendSupported =
+            (buildSet.gpu == "cuda" && backends'.cuda)
+            || (buildSet.gpu == "rocm" && backends'.rocm)
+            || (buildSet.gpu == "metal" && backends'.metal)
+            || (buildConfig.general.universal or false);
+          cudaVersionSupported =
+            buildSet.gpu != "cuda"
+            || (lib.strings.versionAtLeast buildSet.pkgs.cudaPackages.cudaMajorMinorVersion requiredCuda);
+        in
+        backendSupported && cudaVersionSupported;
     in
     builtins.filter supportedBuildSet buildSets;