diff --git a/setup.py b/setup.py
index 3dedb83e66..522a90fe30 100644
--- a/setup.py
+++ b/setup.py
@@ -289,7 +289,8 @@ def remove_rpath(path: Path) -> None:
 
 class CMakeExtension(Extension):
     def __init__(self, name: str, cmake_lists_dir: str = ".", **kwa) -> None:
-        super().__init__(name, sources=[], py_limited_api=not is_freethreaded(), **kwa)
+        kwa.setdefault("py_limited_api", not is_freethreaded())
+        super().__init__(name, sources=[], **kwa)
         self.cmake_lists_dir = os.path.abspath(cmake_lists_dir)
 
 
@@ -1247,11 +1248,18 @@ def _read_requirements(filename: str) -> list[str]:
 if _is_cuda():
     if _cuda_arch_contains(7, 0):
         ext_modules.append(CMakeExtension(name="vllm._sm70_sampler_C"))
-        ext_modules.append(CMakeExtension(name="vllm._sm70_exact_reduce_C"))
-        ext_modules.append(CMakeExtension(name="vllm._h3_w8a16_C"))
-        ext_modules.append(CMakeExtension(name="vllm._h3_flashinfer_C"))
-        ext_modules.append(CMakeExtension(name="vllm._h3_flashattn_C"))
-        ext_modules.append(CMakeExtension(name="vllm._sm70_sparse_attention_C"))
+        # These five are pybind11 modules, which cannot be built against the
+        # limited API, so CMake defines them WITH_SOABI but without USE_SABI
+        # and they carry the full SOABI suffix. Declared limited-API, an
+        # in-place (editable) build copies an .abi3.so that was never built.
+        for soabi_module in (
+            "vllm._sm70_exact_reduce_C",
+            "vllm._h3_w8a16_C",
+            "vllm._h3_flashinfer_C",
+            "vllm._h3_flashattn_C",
+            "vllm._sm70_sparse_attention_C",
+        ):
+            ext_modules.append(CMakeExtension(name=soabi_module, py_limited_api=False))
     build_sm70_fa2 = _cuda_arch_contains(7, 0) and not _cuda_arch_at_least(8, 0)
     if _cuda_arch_at_least(8, 0) or build_sm70_fa2:
         ext_modules.append(CMakeExtension(name="vllm.vllm_flash_attn._vllm_fa2_C"))