--- a/python/xgrammar/kernels/apply_token_bitmask_inplace_cuda.py
+++ b/python/xgrammar/kernels/apply_token_bitmask_inplace_cuda.py
@@ -50,6 +50,17 @@ def _check_cuda_toolchain() -> None:
         raise ImportError(f"nvcc compiler not found at {nvcc_path}")
 
 
+def _select_cuda_host_compiler() -> None:
+    """Select the compiler declared by the CUDA-enabled Gentoo package."""
+    import os
+
+    # torch.utils.cpp_extension uses CXX for its C++ shim and CC to select
+    # nvcc's host compiler.  Translate the standard CUDA overrides and
+    # otherwise use the GCC slot pulled in by USE=cuda.
+    os.environ["CC"] = os.environ.get("CUDAHOSTCC", "/usr/bin/gcc-15")
+    os.environ["CXX"] = os.environ.get("CUDAHOSTCXX", "/usr/bin/g++-15")
+
+
 def _remove_torch_nvcc_flags() -> None:
     REMOVE_NVCC_FLAGS = [
         "-D__CUDA_NO_HALF_OPERATORS__",
@@ -91,6 +102,7 @@ def _load_torch_ops() -> None:
 
 
 _check_cuda_toolchain()
+_select_cuda_host_compiler()
 _remove_torch_nvcc_flags()
 _load_torch_ops()