diff --git a/Cargo.toml b/Cargo.toml
index 79216152ea..4c9e81909f 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -132,15 +132,15 @@ tracel-llvm = { version = "20.1.4-6", features = ["mlir-helpers"] }
 ### For development, if you uncomment, don't forget do to it in `crates/cubecl-cpu/Cargo.toml` [build-dependencies] too to avoid race conditions during build
 # tracel-llvm = { path = "../tracel-llvm/crates/tracel-llvm", features = ["mlir-helpers"] }
 
-cudarc = { version = "0.19.0", features = [
+cudarc = { version = "0.19.0", default-features = false, features = [
     "std",
     "driver",
+    "runtime",
     "nvrtc",
     "nccl",
-    "fallback-dynamic-loading",
-    "cuda-version-from-build-system",
-    "fallback-latest",
-], default-features = false } # CubeCL-CUDA
+    "dynamic-loading",
+    "cuda-12080",
+] } # CubeCL-CUDA
 
 # CubeCL-SPIR-V
 rspirv = { package = "tracel-rspirv", version = "0.12.1" }
diff --git a/crates/cubecl-core/src/frontend/polyfills.rs b/crates/cubecl-core/src/frontend/polyfills.rs
index e600fa5997..38cf58a2fa 100644
--- a/crates/cubecl-core/src/frontend/polyfills.rs
+++ b/crates/cubecl-core/src/frontend/polyfills.rs
@@ -29,7 +29,7 @@ pub mod set_polyfill {
 #[cube]
 pub fn erf<F: Float, N: Size>(x: Vector<F, N>) -> Vector<F, N> {
     let erf = erf_positive(x.abs());
-    select_many(x.less_than(Vector::new(F::new(0.0))), -erf, erf)
+    select_many(x.less_than(Vector::new(F::new(0.0_f32))), -erf, erf)
 }
 
 /// An approximation of the error function: <https://en.wikipedia.org/wiki/Error_function#Numerical_approximations>
@@ -38,13 +38,13 @@ pub fn erf<F: Float, N: Size>(x: Vector<F, N>) -> Vector<F, N> {
 /// > All of these approximations are valid for x ≥ 0. To use these approximations for negative x, use the fact that erf x is an odd function, so erf x = −erf(−x).
 #[cube]
 fn erf_positive<F: Float, N: Size>(x: Vector<F, N>) -> Vector<F, N> {
-    let p = Vector::new(F::new(0.3275911));
-    let a1 = Vector::new(F::new(0.2548296));
-    let a2 = Vector::new(F::new(-0.28449674));
-    let a3 = Vector::new(F::new(1.4214137));
-    let a4 = Vector::new(F::new(-1.453152));
-    let a5 = Vector::new(F::new(1.0614054));
-    let one = Vector::new(F::new(1.0));
+    let p = Vector::new(F::new(0.3275911_f32));
+    let a1 = Vector::new(F::new(0.2548296_f32));
+    let a2 = Vector::new(F::new(-0.28449674_f32));
+    let a3 = Vector::new(F::new(1.4214137_f32));
+    let a4 = Vector::new(F::new(-1.453152_f32));
+    let a5 = Vector::new(F::new(1.0614054_f32));
+    let one = Vector::new(F::new(1.0_f32));
 
     let t = one / (one + p * x);
     let tmp = ((((a5 * t + a4) * t) + a3) * t + a2) * t + a1;
diff --git a/crates/cubecl-cpu/src/compute/runner.rs b/crates/cubecl-cpu/src/compute/runner.rs
index 0c67f542d9..26c80c4021 100644
--- a/crates/cubecl-cpu/src/compute/runner.rs
+++ b/crates/cubecl-cpu/src/compute/runner.rs
@@ -62,7 +62,7 @@ impl Debug for CpuKernel {
 
 impl Debug for KernelRunner {
     fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        write!(f, "{:?}", &self.workers)
+        write!(f, "{:?}", self.workers)
     }
 }
 
diff --git a/crates/cubecl-cuda/Cargo.toml b/crates/cubecl-cuda/Cargo.toml
index 642363a38a..91098e057c 100644
--- a/crates/cubecl-cuda/Cargo.toml
+++ b/crates/cubecl-cuda/Cargo.toml
@@ -74,4 +74,4 @@ pretty_assertions = { workspace = true }
 test-log = { workspace = true, features = ["trace"] }
 
 [build-dependencies]
-cudarc = { workspace = true }
+cudarc = { version = "0.19.0", default-features = false, features = ["std", "driver", "dynamic-loading", "cuda-12080"] }
diff --git a/crates/cubecl-cuda/tests/cudarc_feature_contract.rs b/crates/cubecl-cuda/tests/cudarc_feature_contract.rs
new file mode 100644
index 0000000000..fcdd21e395
--- /dev/null
+++ b/crates/cubecl-cuda/tests/cudarc_feature_contract.rs
@@ -0,0 +1,22 @@
+#[test]
+fn cudarc_uses_the_tensor4all_cuda_floor_without_fallback_detection() {
+    let manifest = include_str!("../../../Cargo.toml");
+    let normalized = manifest.split_whitespace().collect::<Vec<_>>().join(" ");
+
+    assert!(normalized.contains(
+        r#"cudarc = { version = "0.19.0", default-features = false, features = [ "std", "driver", "runtime", "nvrtc", "nccl", "dynamic-loading", "cuda-12080", ] }"#
+    ));
+    assert!(!normalized.contains("fallback-dynamic-loading"));
+    assert!(!normalized.contains("cuda-version-from-build-system"));
+    assert!(!normalized.contains("fallback-latest"));
+}
+
+#[test]
+fn cudarc_build_dependency_only_enables_build_script_requirements() {
+    let manifest = include_str!("../Cargo.toml");
+    let normalized = manifest.split_whitespace().collect::<Vec<_>>().join(" ");
+
+    assert!(normalized.contains(
+        r#"[build-dependencies] cudarc = { version = "0.19.0", default-features = false, features = ["std", "driver", "dynamic-loading", "cuda-12080"] }"#
+    ));
+}
diff --git a/crates/cubecl-opt/src/analyses/mod.rs b/crates/cubecl-opt/src/analyses/mod.rs
index 90ee54a856..bf8845dc81 100644
--- a/crates/cubecl-opt/src/analyses/mod.rs
+++ b/crates/cubecl-opt/src/analyses/mod.rs
@@ -1,5 +1,8 @@
 mod base;
 pub mod dominance;
+// The range analysis is implemented for index-bound optimization but is not
+// wired into the active pass pipeline yet.
+#[allow(dead_code)]
 pub mod integer_range;
 pub mod liveness;
 pub mod post_order;
diff --git a/crates/cubecl-opt/src/lib.rs b/crates/cubecl-opt/src/lib.rs
index 5028ed7dfe..6b0e506d44 100644
--- a/crates/cubecl-opt/src/lib.rs
+++ b/crates/cubecl-opt/src/lib.rs
@@ -147,6 +147,7 @@ pub struct Optimizer {
     /// Root scope to allocate variables on
     pub root_scope: Scope,
     /// The `CubeDim` used for range analysis
+    #[allow(dead_code)]
     pub(crate) cube_dim: CubeDim,
     pub(crate) transformers: Vec<Rc<dyn IrTransformer>>,
     pub(crate) processors: Rc<Vec<Box<dyn Processor>>>,