Link torch_cpu_library, not torch, so the CUDA wheel imports with a CPU-only torch

.ci/scripts/wheel/test_cuda_linux.py

@@ -143,6 +143,27 @@ def test_cuda_libraries_resolve_relatively() -> None:
         print(f"✓ {name} resolves the CUDA runtime relatively ({relative[0]})")
 
 
+def test_no_library_needs_cuda_torch() -> None:
+    """No shipped library may record a dependency on PyTorch's own CUDA libraries.
+
+    The CUDA delegate needs only the CUDA runtime. A library that names libtorch_cuda or
+    libc10_cuda cannot load next to a CPU-only torch, so the whole wheel then fails to import
+    there even though nothing in it calls into those libraries.
+    """
+    package_dir = _package_dir()
+    cuda_torch = {"libtorch_cuda.so", "libc10_cuda.so"}
+    offenders = {}
+    for library in test_shared_libraries._shipped_shared_objects(package_dir):
+        needed = test_shared_libraries._recorded_dependencies(library) & cuda_torch
+        if needed:
+            offenders[str(library.relative_to(package_dir))] = sorted(needed)
+    assert not offenders, (
+        f"these libraries depend on PyTorch's CUDA libraries, so the wheel cannot be imported "
+        f"with a CPU-only torch: {offenders}"
+    )
+    print("✓ no shipped library depends on PyTorch's CUDA libraries")
+
+
 def _row_architectures() -> list[str]:
     """The architectures this row claims, from the same script the build uses.
 
@@ -586,6 +607,7 @@ def test_a_model_runs_through_the_delegate() -> None:
     test_cuda_libraries_are_shipped()
     test_cuda_runtime_is_declared()
     test_cuda_libraries_resolve_relatively()
+    test_no_library_needs_cuda_torch()
     test_device_code_covers_the_row()
     test_portable_device_code_is_present()
     test_the_delegate_registers()

CMakeLists.txt

@@ -1450,11 +1450,16 @@ if(EXECUTORCH_BUILD_PYBIND)
       aten_bridge PUBLIC C10_USING_CUSTOM_GENERATED_MACROS
     )
     target_compile_options(aten_bridge PUBLIC ${_pybind_compile_options})
+    # torch_cpu_library is the CPU half of torch. Linking torch would also force
+    # libtorch_cuda into DT_NEEDED in a CUDA build, and a CPU-only torch could
+    # then not load this library.
     if(EXECUTORCH_BUILD_SHARED)
-      target_link_libraries(aten_bridge PRIVATE torch c10 executorch_shared)
+      target_link_libraries(
+        aten_bridge PRIVATE torch_cpu_library c10 executorch_shared
+      )
     else()
       target_link_libraries(
-        aten_bridge PRIVATE torch c10 executorch extension_tensor
+        aten_bridge PRIVATE torch_cpu_library c10 executorch extension_tensor
       )
     endif()
   endif()

extension/llm/custom_ops/CMakeLists.txt

@@ -230,8 +230,11 @@ if(EXECUTORCH_BUILD_KERNELS_LLM_AOT)
   endif()
   executorch_target_link_shared_runtime(custom_ops_aot_lib)
 
+  # torch_cpu_library is the CPU half of torch. Linking torch would also force
+  # libtorch_cuda into DT_NEEDED in a CUDA build, and a CPU-only torch could
+  # then not load this library.
   target_link_libraries(
-    custom_ops_aot_lib PUBLIC cpublas torch extension_tensor
+    custom_ops_aot_lib PUBLIC cpublas torch_cpu_library extension_tensor
                               extension_threadpool
   )
   if(EXECUTORCH_BUILD_KERNELS_LLM_QUANTIZED_MOE_OPTIMIZED)

extension/llm/runner/CMakeLists.txt

@@ -120,10 +120,14 @@ if(EXECUTORCH_BUILD_PYBIND)
   find_library(
     TORCH_PYTHON_LIBRARY torch_python PATHS "${TORCH_INSTALL_PREFIX}/lib"
   )
-  # Link with the extension_llm_runner library and its dependencies
+  # Link with the extension_llm_runner library and its dependencies.
+  #
+  # torch_cpu_library is the CPU half of torch. Linking torch would also force
+  # libtorch_cuda into DT_NEEDED in a CUDA build, and a CPU-only torch could
+  # then not load this library.
   target_link_libraries(
     _llm_runner PRIVATE extension_llm_runner tokenizers::tokenizers
-                        portable_lib ${TORCH_PYTHON_LIBRARY} ${TORCH_LIBRARIES}
+                        portable_lib ${TORCH_PYTHON_LIBRARY} torch_cpu_library
   )
   executorch_target_link_shared_runtime(_llm_runner)
 

extension/training/CMakeLists.txt

@@ -52,13 +52,18 @@ if(EXECUTORCH_BUILD_PYBIND)
   # Pybind library. When the consolidated shared runtime is built, the runtime
   # is resolved from it rather than from the whole-archive-forcing static
   # `executorch` target, so this module shares the one backend registry.
+  #
+  # torch_cpu_library is the CPU half of torch. Linking torch would also force
+  # libtorch_cuda into DT_NEEDED in a CUDA build, and a CPU-only torch could
+  # then not load this library.
   if(EXECUTORCH_BUILD_SHARED)
     set(_pybind_training_dep_libs ${TORCH_PYTHON_LIBRARY} etdump aten_bridge
-                                  torch extension_training
+                                  torch_cpu_library extension_training
     )
   else()
-    set(_pybind_training_dep_libs ${TORCH_PYTHON_LIBRARY} etdump executorch
-                                  aten_bridge torch extension_training
+    set(_pybind_training_dep_libs
+        ${TORCH_PYTHON_LIBRARY} etdump executorch aten_bridge torch_cpu_library
+        extension_training
     )
   endif()
 

tools/cmake/Codegen.cmake

@@ -334,7 +334,10 @@ function(gen_custom_ops_aot_lib)
   set_target_properties(${GEN_LIB_NAME} PROPERTIES CXX_STANDARD 20)
   target_compile_definitions(${GEN_LIB_NAME} PRIVATE USE_ATEN_LIB=1)
   include_directories(${TORCH_INCLUDE_DIRS})
-  target_link_libraries(${GEN_LIB_NAME} PRIVATE torch)
+  # torch_cpu_library is the CPU half of torch. Linking torch would also force
+  # libtorch_cuda into DT_NEEDED in a CUDA build, and a CPU-only torch could
+  # then not load this library.
+  target_link_libraries(${GEN_LIB_NAME} PRIVATE torch_cpu_library)
 
   executorch_target_link_options_shared_lib(${GEN_LIB_NAME})
   if(TARGET portable_lib)