Fix build failure with CUDA < 13.0: unused private fields in symmetric_tensor.h

[x41lakazam]

…cTensor

Building with CUDA 12.x and Clang (-Werror,-Wunused-private-field) fails with 5 errors in symmetric_tensor.h:

error: private field 'mcast_handle_' is not used [-Werror,-Wunused-private-field]
error: private field 'cu_dev_' is not used [-Werror,-Wunused-private-field]
error: private field 'mc_base_ptr_' is not used [-Werror,-Wunused-private-field]
error: private field 'exporter_rank_' is not used [-Werror,-Wunused-private-field]
error: private field 'peer_fd_' is not used [-Werror,-Wunused-private-field]

These five fields are only used inside #if (CUDA_VERSION >= 13000) blocks (in the destructor and setupMulticast), but they are declared unconditionally. When building with CUDA 12.x, the usage is compiled out but the declarations remain, triggering Clang's -Wunused-private-field.

[greptile-apps]

Greptile Summary

This PR fixes a build failure when compiling with CUDA < 13.0 and Clang's -Werror,-Wunused-private-field by wrapping the five private fields that are exclusively used inside #if (CUDA_VERSION >= 13000) blocks into matching preprocessor guards in the class declaration.

Key changes in csrc/multidevice/symmetric_tensor.h:

• mcast_handle_, cu_dev_, mc_base_ptr_, exporter_rank_, and peer_fd_ are now declared inside #if (CUDA_VERSION >= 13000), matching their actual usage in the destructor and setupMulticast() in the .cpp file.
• mc_ptr_ is correctly moved outside the guard — it is accessed unconditionally by multicastPtr(), which has no version check of its own (it relies on is_multicast_setup_ being false at runtime on CUDA < 13.0 to prevent returning the field).

Confidence Score: 5/5

• This PR is safe to merge — it is a minimal, correct fix for a Clang diagnostic with no behavioral changes.
• The change is purely a header-level preprocessor guard adjustment. The fields being guarded are only ever accessed inside #if (CUDA_VERSION >= 13000) blocks in the implementation, so guarding their declarations is correct. The one field that needs to remain unconditional (mc_ptr_) is correctly left outside the guard. There is no change to runtime behavior, memory layout concerns are limited to compile-time selection, and the fix exactly matches the stated problem.
• No files require special attention.

Important Files Changed

Filename	Overview
csrc/multidevice/symmetric_tensor.h	Wraps five private fields (mcast_handle_, cu_dev_, mc_base_ptr_, exporter_rank_, peer_fd_) inside #if (CUDA_VERSION >= 13000) guards while correctly keeping mc_ptr_ unconditional since it is returned by the non-guarded multicastPtr() accessor.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[SymmetricTensor fields] --> B{CUDA_VERSION >= 13000?}

    B -- Yes --> C[All fields available:\n- mcast_handle_\n- cu_dev_\n- mc_base_ptr_\n- exporter_rank_\n- peer_fd_\n- mc_ptr_]
    B -- No --> D[Only unconditional fields:\n- mc_ptr_\n- is_multicast_setup_]

    C --> E[setupMulticast runs fully]
    D --> F[setupMulticast throws NVF_ERROR\n'Multicast requires CUDA 13.0+']

    E --> G[multicastPtr returns mc_ptr_]
    F --> H[is_multicast_setup_ stays false\nmulticastPtr NVF_CHECK fails]

    G --> I[Destructor: cleans up\nmc_base_ptr_, mcast_handle_,\npeer_fd_ inside #if guard]
    H --> J[Destructor: skips multicast\ncleanup block entirely]

Loading

_{Last reviewed commit: d7ad13a}

[x41lakazam]

!test

[samnordmann]

!test