Fix [[nodiscard]] build errors and BUCK deps across comms, gloo, caffe2 by gyllstromk · Pull Request #494 · pytorch/gloo

gyllstromk · 2026-03-06T01:15:56Z

Summary:
ROCm 7.0+ HIP headers annotate API functions (hipStreamDestroy,
hipMemcpyAsync, hipStreamSynchronize, hipSetDevice, hipGetDevice, hipFree,
hipHostUnregister, hipDeviceEnablePeerAccess, cuGetErrorString) with
[[nodiscard]]. Combined with -Werror, this causes build failures wherever
return values are discarded.

Originally discovered building with ROCm 7.2 headers, but confirmed to
also affect ROCm 7.0 builds (reported independently by yvliu and hqguo).
The [[nodiscard]] attribute is present in both ROCm 7.0 and 7.2 HIP
headers — the fix is the same for both versions.

Changes:

Add (void) casts to suppress [[nodiscard]] warnings across comms/
(tcp_devmem, ctran, rcclx), gloo/, and caffe2/ (nativert) — 12 C++ files
Fix BUCK dependency issues in comms/tcp_devmem/nccl (replace devmgr-client
with common:common) and comms/tcp_devmem/unpack (explicit glog dep path)
that surface when building these targets under ROCm constraints

The (void) casts are no-ops on CUDA and older ROCm — safe to land
regardless of ROCm version.

Reviewed By: bbeckca

Differential Revision: D93759269

meta-codesync · 2026-03-06T23:59:08Z

@gyllstromk has exported this pull request. If you are a Meta employee, you can view the originating Diff in D93759269.

…omms, gloo, caffe2 (#176671) Summary: X-link: pytorch/gloo#494 X-link: meta-pytorch/torchcomms#960 ROCm 7.0+ HIP headers annotate API functions (hipStreamDestroy, hipMemcpyAsync, hipStreamSynchronize, hipSetDevice, hipGetDevice, hipFree, hipHostUnregister, hipDeviceEnablePeerAccess, cuGetErrorString) with [[nodiscard]]. Combined with -Werror, this causes build failures wherever return values are discarded. Originally discovered building with ROCm 7.2 headers, but confirmed to also affect ROCm 7.0 builds (reported independently by yvliu and hqguo). The [[nodiscard]] attribute is present in both ROCm 7.0 and 7.2 HIP headers — the fix is the same for both versions. Changes: - Add (void) casts to suppress [[nodiscard]] warnings across comms/ (tcp_devmem, ctran, rcclx), gloo/, and caffe2/ (nativert) — 12 C++ files - Fix BUCK dependency issues in comms/tcp_devmem/nccl (replace devmgr-client with common:common) and comms/tcp_devmem/unpack (explicit glog dep path) that surface when building these targets under ROCm constraints The (void) casts are no-ops on CUDA and older ROCm — safe to land regardless of ROCm version. Test Plan: ## Reproducing the [[nodiscard]] build errors (ROCm 7.0+) ROCm 7.0 HIP headers annotate CUDA-mapped API functions with `[[nodiscard]]` (e.g. `hipStreamDestroy`, `hipSetDevice`, `hipFree`). With `-Werror` enabled, any call site that discards the return value fails to compile: ``` error: ignoring return value of function declared with 'nodiscard' attribute [-Werror,-Wunused-result] cudaSetDevice(device); ^~~~~~~~~~~~~~~~~~~~~~ ``` To reproduce, build any affected target with ROCm >= 7.0. Example using the gloo HIP collectives (which calls `cudaSetDevice` and `cudaDeviceEnablePeerAccess` without checking the return value): ```bash hipcc -std=c++17 -Werror \ -I<pytorch_root> -I<rocm_7.0_path>/include \ -c gloo/cuda_collectives_native.h # → error: ignoring return value ... [-Werror,-Wunused-result] ``` The fix adds `(void)` casts to explicitly discard the return value, which is the standard C++ pattern for suppressing `[[nodiscard]]` warnings. The casts are no-ops on CUDA and older ROCm versions. ## Verification (fbcode, mode/amd-gpu with ROCm 7.0 headers) ```bash # ctran (ibutils.cc, LogInit.cc) buck2 build mode/amd-gpu fbcode//comms/ctran/backends/ib:ib buck2 build mode/amd-gpu fbcode//comms/ctran/utils:utils # rcclx (cudawrap.cc, register.cc, RcclxScubaLogger.h) buck2 build mode/amd-gpu fbcode//comms/rcclx:rcclx-dev # tcp_devmem (batch_unpack_producer.cc, shared_region.cc) buck2 build mode/amd-gpu fbcode//comms/tcp_devmem/unpack:batch_unpack_producer buck2 build mode/amd-gpu fbcode//comms/tcp_devmem/common:common # gloo (cuda_collectives_native.h) buck2 build mode/amd-gpu fbcode//gloo:gloo_gpu_hip ``` All targets build successfully. The BUCK dep fixes for `comms/tcp_devmem/nccl` (nccl-gen, nccl-sim) and `comms/tcp_devmem/unpack` resolve link-time errors that surface when building under ROCm constraints. Reviewed By: bbeckca Differential Revision: D93759269

…e2 (pytorch#494) Summary: X-link: meta-pytorch/torchcomms#960 X-link: pytorch/pytorch#176671 ROCm 7.0+ HIP headers annotate API functions (hipStreamDestroy, hipMemcpyAsync, hipStreamSynchronize, hipSetDevice, hipGetDevice, hipFree, hipHostUnregister, hipDeviceEnablePeerAccess, cuGetErrorString) with [[nodiscard]]. Combined with -Werror, this causes build failures wherever return values are discarded. Originally discovered building with ROCm 7.2 headers, but confirmed to also affect ROCm 7.0 builds (reported independently by yvliu and hqguo). The [[nodiscard]] attribute is present in both ROCm 7.0 and 7.2 HIP headers — the fix is the same for both versions. Changes: - Add (void) casts to suppress [[nodiscard]] warnings across comms/ (tcp_devmem, ctran, rcclx), gloo/, and caffe2/ (nativert) — 12 C++ files - Fix BUCK dependency issues in comms/tcp_devmem/nccl (replace devmgr-client with common:common) and comms/tcp_devmem/unpack (explicit glog dep path) that surface when building these targets under ROCm constraints The (void) casts are no-ops on CUDA and older ROCm — safe to land regardless of ROCm version. Reviewed By: bbeckca Differential Revision: D93759269

…e2 (meta-pytorch#960) Summary: X-link: pytorch/gloo#494 X-link: pytorch/pytorch#176671 ROCm 7.0+ HIP headers annotate API functions (hipStreamDestroy, hipMemcpyAsync, hipStreamSynchronize, hipSetDevice, hipGetDevice, hipFree, hipHostUnregister, hipDeviceEnablePeerAccess, cuGetErrorString) with [[nodiscard]]. Combined with -Werror, this causes build failures wherever return values are discarded. Originally discovered building with ROCm 7.2 headers, but confirmed to also affect ROCm 7.0 builds (reported independently by yvliu and hqguo). The [[nodiscard]] attribute is present in both ROCm 7.0 and 7.2 HIP headers — the fix is the same for both versions. Changes: - Add (void) casts to suppress [[nodiscard]] warnings across comms/ (tcp_devmem, ctran, rcclx), gloo/, and caffe2/ (nativert) — 12 C++ files - Fix BUCK dependency issues in comms/tcp_devmem/nccl (replace devmgr-client with common:common) and comms/tcp_devmem/unpack (explicit glog dep path) that surface when building these targets under ROCm constraints The (void) casts are no-ops on CUDA and older ROCm — safe to land regardless of ROCm version. Reviewed By: bbeckca Differential Revision: D93759269

…omms, gloo, caffe2 (pytorch#176671) Summary: X-link: pytorch/gloo#494 X-link: meta-pytorch/torchcomms#960 ROCm 7.0+ HIP headers annotate API functions (hipStreamDestroy, hipMemcpyAsync, hipStreamSynchronize, hipSetDevice, hipGetDevice, hipFree, hipHostUnregister, hipDeviceEnablePeerAccess, cuGetErrorString) with [[nodiscard]]. Combined with -Werror, this causes build failures wherever return values are discarded. Originally discovered building with ROCm 7.2 headers, but confirmed to also affect ROCm 7.0 builds (reported independently by yvliu and hqguo). The [[nodiscard]] attribute is present in both ROCm 7.0 and 7.2 HIP headers — the fix is the same for both versions. Changes: - Add (void) casts to suppress [[nodiscard]] warnings across comms/ (tcp_devmem, ctran, rcclx), gloo/, and caffe2/ (nativert) — 12 C++ files - Fix BUCK dependency issues in comms/tcp_devmem/nccl (replace devmgr-client with common:common) and comms/tcp_devmem/unpack (explicit glog dep path) that surface when building these targets under ROCm constraints The (void) casts are no-ops on CUDA and older ROCm — safe to land regardless of ROCm version. Test Plan: ## Reproducing the [[nodiscard]] build errors (ROCm 7.0+) ROCm 7.0 HIP headers annotate CUDA-mapped API functions with `[[nodiscard]]` (e.g. `hipStreamDestroy`, `hipSetDevice`, `hipFree`). With `-Werror` enabled, any call site that discards the return value fails to compile: ``` error: ignoring return value of function declared with 'nodiscard' attribute [-Werror,-Wunused-result] cudaSetDevice(device); ^~~~~~~~~~~~~~~~~~~~~~ ``` To reproduce, build any affected target with ROCm >= 7.0. Example using the gloo HIP collectives (which calls `cudaSetDevice` and `cudaDeviceEnablePeerAccess` without checking the return value): ```bash hipcc -std=c++17 -Werror \ -I<pytorch_root> -I<rocm_7.0_path>/include \ -c gloo/cuda_collectives_native.h # → error: ignoring return value ... [-Werror,-Wunused-result] ``` The fix adds `(void)` casts to explicitly discard the return value, which is the standard C++ pattern for suppressing `[[nodiscard]]` warnings. The casts are no-ops on CUDA and older ROCm versions. ## Verification (fbcode, mode/amd-gpu with ROCm 7.0 headers) ```bash # ctran (ibutils.cc, LogInit.cc) buck2 build mode/amd-gpu fbcode//comms/ctran/backends/ib:ib buck2 build mode/amd-gpu fbcode//comms/ctran/utils:utils # rcclx (cudawrap.cc, register.cc, RcclxScubaLogger.h) buck2 build mode/amd-gpu fbcode//comms/rcclx:rcclx-dev # tcp_devmem (batch_unpack_producer.cc, shared_region.cc) buck2 build mode/amd-gpu fbcode//comms/tcp_devmem/unpack:batch_unpack_producer buck2 build mode/amd-gpu fbcode//comms/tcp_devmem/common:common # gloo (cuda_collectives_native.h) buck2 build mode/amd-gpu fbcode//gloo:gloo_gpu_hip ``` All targets build successfully. The BUCK dep fixes for `comms/tcp_devmem/nccl` (nccl-gen, nccl-sim) and `comms/tcp_devmem/unpack` resolve link-time errors that surface when building under ROCm constraints. Reviewed By: bbeckca Differential Revision: D93759269

…e2 (meta-pytorch#960) Summary: X-link: pytorch/gloo#494 Pull Request resolved: meta-pytorch#960 X-link: pytorch/pytorch#176671 ROCm 7.0+ HIP headers annotate API functions (hipStreamDestroy, hipMemcpyAsync, hipStreamSynchronize, hipSetDevice, hipGetDevice, hipFree, hipHostUnregister, hipDeviceEnablePeerAccess, cuGetErrorString) with [[nodiscard]]. Combined with -Werror, this causes build failures wherever return values are discarded. Originally discovered building with ROCm 7.2 headers, but confirmed to also affect ROCm 7.0 builds (reported independently by yvliu and hqguo). The [[nodiscard]] attribute is present in both ROCm 7.0 and 7.2 HIP headers — the fix is the same for both versions. Changes: - Add (void) casts to suppress [[nodiscard]] warnings across comms/ (tcp_devmem, ctran, rcclx), gloo/, and caffe2/ (nativert) — 12 C++ files - Fix BUCK dependency issues in comms/tcp_devmem/nccl (replace devmgr-client with common:common) and comms/tcp_devmem/unpack (explicit glog dep path) that surface when building these targets under ROCm constraints The (void) casts are no-ops on CUDA and older ROCm — safe to land regardless of ROCm version. Reviewed By: bbeckca Differential Revision: D93759269

…e2 (pytorch#494) Summary: Pull Request resolved: pytorch#494 X-link: meta-pytorch/torchcomms#960 X-link: pytorch/pytorch#176671 ROCm 7.0+ HIP headers annotate API functions (hipStreamDestroy, hipMemcpyAsync, hipStreamSynchronize, hipSetDevice, hipGetDevice, hipFree, hipHostUnregister, hipDeviceEnablePeerAccess, cuGetErrorString) with [[nodiscard]]. Combined with -Werror, this causes build failures wherever return values are discarded. Originally discovered building with ROCm 7.2 headers, but confirmed to also affect ROCm 7.0 builds (reported independently by yvliu and hqguo). The [[nodiscard]] attribute is present in both ROCm 7.0 and 7.2 HIP headers — the fix is the same for both versions. Changes: - Add (void) casts to suppress [[nodiscard]] warnings across comms/ (tcp_devmem, ctran, rcclx), gloo/, and caffe2/ (nativert) — 12 C++ files - Fix BUCK dependency issues in comms/tcp_devmem/nccl (replace devmgr-client with common:common) and comms/tcp_devmem/unpack (explicit glog dep path) that surface when building these targets under ROCm constraints The (void) casts are no-ops on CUDA and older ROCm — safe to land regardless of ROCm version. Reviewed By: bbeckca Differential Revision: D93759269

…omms, gloo, caffe2 (pytorch#176671) Summary: X-link: pytorch/gloo#494 X-link: meta-pytorch/torchcomms#960 Pull Request resolved: pytorch#176671 ROCm 7.0+ HIP headers annotate API functions (hipStreamDestroy, hipMemcpyAsync, hipStreamSynchronize, hipSetDevice, hipGetDevice, hipFree, hipHostUnregister, hipDeviceEnablePeerAccess, cuGetErrorString) with [[nodiscard]]. Combined with -Werror, this causes build failures wherever return values are discarded. Originally discovered building with ROCm 7.2 headers, but confirmed to also affect ROCm 7.0 builds (reported independently by yvliu and hqguo). The [[nodiscard]] attribute is present in both ROCm 7.0 and 7.2 HIP headers — the fix is the same for both versions. Changes: - Add (void) casts to suppress [[nodiscard]] warnings across comms/ (tcp_devmem, ctran, rcclx), gloo/, and caffe2/ (nativert) — 12 C++ files - Fix BUCK dependency issues in comms/tcp_devmem/nccl (replace devmgr-client with common:common) and comms/tcp_devmem/unpack (explicit glog dep path) that surface when building these targets under ROCm constraints The (void) casts are no-ops on CUDA and older ROCm — safe to land regardless of ROCm version. Test Plan: ## Reproducing the [[nodiscard]] build errors (ROCm 7.0+) ROCm 7.0 HIP headers annotate CUDA-mapped API functions with `[[nodiscard]]` (e.g. `hipStreamDestroy`, `hipSetDevice`, `hipFree`). With `-Werror` enabled, any call site that discards the return value fails to compile: ``` error: ignoring return value of function declared with 'nodiscard' attribute [-Werror,-Wunused-result] cudaSetDevice(device); ^~~~~~~~~~~~~~~~~~~~~~ ``` To reproduce, build any affected target with ROCm >= 7.0. Example using the gloo HIP collectives (which calls `cudaSetDevice` and `cudaDeviceEnablePeerAccess` without checking the return value): ```bash hipcc -std=c++17 -Werror \ -I<pytorch_root> -I<rocm_7.0_path>/include \ -c gloo/cuda_collectives_native.h # → error: ignoring return value ... [-Werror,-Wunused-result] ``` The fix adds `(void)` casts to explicitly discard the return value, which is the standard C++ pattern for suppressing `[[nodiscard]]` warnings. The casts are no-ops on CUDA and older ROCm versions. ## Verification (fbcode, mode/amd-gpu with ROCm 7.0 headers) ```bash # ctran (ibutils.cc, LogInit.cc) buck2 build mode/amd-gpu fbcode//comms/ctran/backends/ib:ib buck2 build mode/amd-gpu fbcode//comms/ctran/utils:utils # rcclx (cudawrap.cc, register.cc, RcclxScubaLogger.h) buck2 build mode/amd-gpu fbcode//comms/rcclx:rcclx-dev # tcp_devmem (batch_unpack_producer.cc, shared_region.cc) buck2 build mode/amd-gpu fbcode//comms/tcp_devmem/unpack:batch_unpack_producer buck2 build mode/amd-gpu fbcode//comms/tcp_devmem/common:common # gloo (cuda_collectives_native.h) buck2 build mode/amd-gpu fbcode//gloo:gloo_gpu_hip ``` All targets build successfully. The BUCK dep fixes for `comms/tcp_devmem/nccl` (nccl-gen, nccl-sim) and `comms/tcp_devmem/unpack` resolve link-time errors that surface when building under ROCm constraints. Reviewed By: bbeckca Differential Revision: D93759269

…e2 (pytorch#494) Summary: X-link: meta-pytorch/torchcomms#960 X-link: pytorch/pytorch#176671 ROCm 7.0+ HIP headers annotate API functions (hipStreamDestroy, hipMemcpyAsync, hipStreamSynchronize, hipSetDevice, hipGetDevice, hipFree, hipHostUnregister, hipDeviceEnablePeerAccess, cuGetErrorString) with [[nodiscard]]. Combined with -Werror, this causes build failures wherever return values are discarded. Originally discovered building with ROCm 7.2 headers, but confirmed to also affect ROCm 7.0 builds (reported independently by yvliu and hqguo). The [[nodiscard]] attribute is present in both ROCm 7.0 and 7.2 HIP headers — the fix is the same for both versions. Changes: - Add (void) casts to suppress [[nodiscard]] warnings across comms/ (tcp_devmem, ctran, rcclx), gloo/, and caffe2/ (nativert) — 12 C++ files - Fix BUCK dependency issues in comms/tcp_devmem/nccl (replace devmgr-client with common:common) and comms/tcp_devmem/unpack (explicit glog dep path) that surface when building these targets under ROCm constraints The (void) casts are no-ops on CUDA and older ROCm — safe to land regardless of ROCm version. Reviewed By: bbeckca Differential Revision: D93759269

…e2 (meta-pytorch#960) Summary: X-link: pytorch/gloo#494 X-link: pytorch/pytorch#176671 ROCm 7.0+ HIP headers annotate API functions (hipStreamDestroy, hipMemcpyAsync, hipStreamSynchronize, hipSetDevice, hipGetDevice, hipFree, hipHostUnregister, hipDeviceEnablePeerAccess, cuGetErrorString) with [[nodiscard]]. Combined with -Werror, this causes build failures wherever return values are discarded. Originally discovered building with ROCm 7.2 headers, but confirmed to also affect ROCm 7.0 builds (reported independently by yvliu and hqguo). The [[nodiscard]] attribute is present in both ROCm 7.0 and 7.2 HIP headers — the fix is the same for both versions. Changes: - Add (void) casts to suppress [[nodiscard]] warnings across comms/ (tcp_devmem, ctran, rcclx), gloo/, and caffe2/ (nativert) — 12 C++ files - Fix BUCK dependency issues in comms/tcp_devmem/nccl (replace devmgr-client with common:common) and comms/tcp_devmem/unpack (explicit glog dep path) that surface when building these targets under ROCm constraints The (void) casts are no-ops on CUDA and older ROCm — safe to land regardless of ROCm version. Reviewed By: bbeckca Differential Revision: D93759269

…omms, gloo, caffe2 (pytorch#176671) Summary: X-link: pytorch/gloo#494 X-link: meta-pytorch/torchcomms#960 ROCm 7.0+ HIP headers annotate API functions (hipStreamDestroy, hipMemcpyAsync, hipStreamSynchronize, hipSetDevice, hipGetDevice, hipFree, hipHostUnregister, hipDeviceEnablePeerAccess, cuGetErrorString) with [[nodiscard]]. Combined with -Werror, this causes build failures wherever return values are discarded. Originally discovered building with ROCm 7.2 headers, but confirmed to also affect ROCm 7.0 builds (reported independently by yvliu and hqguo). The [[nodiscard]] attribute is present in both ROCm 7.0 and 7.2 HIP headers — the fix is the same for both versions. Changes: - Add (void) casts to suppress [[nodiscard]] warnings across comms/ (tcp_devmem, ctran, rcclx), gloo/, and caffe2/ (nativert) — 12 C++ files - Fix BUCK dependency issues in comms/tcp_devmem/nccl (replace devmgr-client with common:common) and comms/tcp_devmem/unpack (explicit glog dep path) that surface when building these targets under ROCm constraints The (void) casts are no-ops on CUDA and older ROCm — safe to land regardless of ROCm version. Test Plan: ## Reproducing the [[nodiscard]] build errors (ROCm 7.0+) ROCm 7.0 HIP headers annotate CUDA-mapped API functions with `[[nodiscard]]` (e.g. `hipStreamDestroy`, `hipSetDevice`, `hipFree`). With `-Werror` enabled, any call site that discards the return value fails to compile: ``` error: ignoring return value of function declared with 'nodiscard' attribute [-Werror,-Wunused-result] cudaSetDevice(device); ^~~~~~~~~~~~~~~~~~~~~~ ``` To reproduce, build any affected target with ROCm >= 7.0. Example using the gloo HIP collectives (which calls `cudaSetDevice` and `cudaDeviceEnablePeerAccess` without checking the return value): ```bash hipcc -std=c++17 -Werror \ -I<pytorch_root> -I<rocm_7.0_path>/include \ -c gloo/cuda_collectives_native.h # → error: ignoring return value ... [-Werror,-Wunused-result] ``` The fix adds `(void)` casts to explicitly discard the return value, which is the standard C++ pattern for suppressing `[[nodiscard]]` warnings. The casts are no-ops on CUDA and older ROCm versions. ## Verification (fbcode, mode/amd-gpu with ROCm 7.0 headers) ```bash # ctran (ibutils.cc, LogInit.cc) buck2 build mode/amd-gpu fbcode//comms/ctran/backends/ib:ib buck2 build mode/amd-gpu fbcode//comms/ctran/utils:utils # rcclx (cudawrap.cc, register.cc, RcclxScubaLogger.h) buck2 build mode/amd-gpu fbcode//comms/rcclx:rcclx-dev # tcp_devmem (batch_unpack_producer.cc, shared_region.cc) buck2 build mode/amd-gpu fbcode//comms/tcp_devmem/unpack:batch_unpack_producer buck2 build mode/amd-gpu fbcode//comms/tcp_devmem/common:common # gloo (cuda_collectives_native.h) buck2 build mode/amd-gpu fbcode//gloo:gloo_gpu_hip ``` All targets build successfully. The BUCK dep fixes for `comms/tcp_devmem/nccl` (nccl-gen, nccl-sim) and `comms/tcp_devmem/unpack` resolve link-time errors that surface when building under ROCm constraints. Reviewed By: bbeckca Differential Revision: D93759269

…e2 (pytorch#494) Summary: Pull Request resolved: pytorch#494 X-link: meta-pytorch/torchcomms#960 X-link: pytorch/pytorch#176671 ROCm 7.0+ HIP headers annotate API functions (hipStreamDestroy, hipMemcpyAsync, hipStreamSynchronize, hipSetDevice, hipGetDevice, hipFree, hipHostUnregister, hipDeviceEnablePeerAccess, cuGetErrorString) with [[nodiscard]]. Combined with -Werror, this causes build failures wherever return values are discarded. Originally discovered building with ROCm 7.2 headers, but confirmed to also affect ROCm 7.0 builds (reported independently by yvliu and hqguo). The [[nodiscard]] attribute is present in both ROCm 7.0 and 7.2 HIP headers — the fix is the same for both versions. Changes: - Add (void) casts to suppress [[nodiscard]] warnings across comms/ (tcp_devmem, ctran, rcclx), gloo/, and caffe2/ (nativert) — 12 C++ files - Fix BUCK dependency issues in comms/tcp_devmem/nccl (replace devmgr-client with common:common) and comms/tcp_devmem/unpack (explicit glog dep path) that surface when building these targets under ROCm constraints The (void) casts are no-ops on CUDA and older ROCm — safe to land regardless of ROCm version. Reviewed By: bbeckca Differential Revision: D93759269

…e2 (meta-pytorch#960) Summary: X-link: pytorch/gloo#494 Pull Request resolved: meta-pytorch#960 X-link: pytorch/pytorch#176671 ROCm 7.0+ HIP headers annotate API functions (hipStreamDestroy, hipMemcpyAsync, hipStreamSynchronize, hipSetDevice, hipGetDevice, hipFree, hipHostUnregister, hipDeviceEnablePeerAccess, cuGetErrorString) with [[nodiscard]]. Combined with -Werror, this causes build failures wherever return values are discarded. Originally discovered building with ROCm 7.2 headers, but confirmed to also affect ROCm 7.0 builds (reported independently by yvliu and hqguo). The [[nodiscard]] attribute is present in both ROCm 7.0 and 7.2 HIP headers — the fix is the same for both versions. Changes: - Add (void) casts to suppress [[nodiscard]] warnings across comms/ (tcp_devmem, ctran, rcclx), gloo/, and caffe2/ (nativert) — 12 C++ files - Fix BUCK dependency issues in comms/tcp_devmem/nccl (replace devmgr-client with common:common) and comms/tcp_devmem/unpack (explicit glog dep path) that surface when building these targets under ROCm constraints The (void) casts are no-ops on CUDA and older ROCm — safe to land regardless of ROCm version. Reviewed By: bbeckca Differential Revision: D93759269

…omms, gloo, caffe2 (pytorch#176671) Summary: X-link: pytorch/gloo#494 X-link: meta-pytorch/torchcomms#960 Pull Request resolved: pytorch#176671 ROCm 7.0+ HIP headers annotate API functions (hipStreamDestroy, hipMemcpyAsync, hipStreamSynchronize, hipSetDevice, hipGetDevice, hipFree, hipHostUnregister, hipDeviceEnablePeerAccess, cuGetErrorString) with [[nodiscard]]. Combined with -Werror, this causes build failures wherever return values are discarded. Originally discovered building with ROCm 7.2 headers, but confirmed to also affect ROCm 7.0 builds (reported independently by yvliu and hqguo). The [[nodiscard]] attribute is present in both ROCm 7.0 and 7.2 HIP headers — the fix is the same for both versions. Changes: - Add (void) casts to suppress [[nodiscard]] warnings across comms/ (tcp_devmem, ctran, rcclx), gloo/, and caffe2/ (nativert) — 12 C++ files - Fix BUCK dependency issues in comms/tcp_devmem/nccl (replace devmgr-client with common:common) and comms/tcp_devmem/unpack (explicit glog dep path) that surface when building these targets under ROCm constraints The (void) casts are no-ops on CUDA and older ROCm — safe to land regardless of ROCm version. Test Plan: ## Reproducing the [[nodiscard]] build errors (ROCm 7.0+) ROCm 7.0 HIP headers annotate CUDA-mapped API functions with `[[nodiscard]]` (e.g. `hipStreamDestroy`, `hipSetDevice`, `hipFree`). With `-Werror` enabled, any call site that discards the return value fails to compile: ``` error: ignoring return value of function declared with 'nodiscard' attribute [-Werror,-Wunused-result] cudaSetDevice(device); ^~~~~~~~~~~~~~~~~~~~~~ ``` To reproduce, build any affected target with ROCm >= 7.0. Example using the gloo HIP collectives (which calls `cudaSetDevice` and `cudaDeviceEnablePeerAccess` without checking the return value): ```bash hipcc -std=c++17 -Werror \ -I<pytorch_root> -I<rocm_7.0_path>/include \ -c gloo/cuda_collectives_native.h # → error: ignoring return value ... [-Werror,-Wunused-result] ``` The fix adds `(void)` casts to explicitly discard the return value, which is the standard C++ pattern for suppressing `[[nodiscard]]` warnings. The casts are no-ops on CUDA and older ROCm versions. ## Verification (fbcode, mode/amd-gpu with ROCm 7.0 headers) ```bash # ctran (ibutils.cc, LogInit.cc) buck2 build mode/amd-gpu fbcode//comms/ctran/backends/ib:ib buck2 build mode/amd-gpu fbcode//comms/ctran/utils:utils # rcclx (cudawrap.cc, register.cc, RcclxScubaLogger.h) buck2 build mode/amd-gpu fbcode//comms/rcclx:rcclx-dev # tcp_devmem (batch_unpack_producer.cc, shared_region.cc) buck2 build mode/amd-gpu fbcode//comms/tcp_devmem/unpack:batch_unpack_producer buck2 build mode/amd-gpu fbcode//comms/tcp_devmem/common:common # gloo (cuda_collectives_native.h) buck2 build mode/amd-gpu fbcode//gloo:gloo_gpu_hip ``` All targets build successfully. The BUCK dep fixes for `comms/tcp_devmem/nccl` (nccl-gen, nccl-sim) and `comms/tcp_devmem/unpack` resolve link-time errors that surface when building under ROCm constraints. Reviewed By: bbeckca Differential Revision: D93759269

Summary: See CI signals on PR: pytorch#494 Runner: ubuntu-latest is a standard GitHub-hosted runner, which is a vanilla VM with no InfiniBand/RDMA hardware Build flags: -DUSE_IBVERBS=ON: compiles IBVERBS support and installs libibverbs-dev headers for compilation only `gloo_test` crashes with `terminate called recursively` (exit code 134) on CI runners that compile with IBVERBS/TLS support but lack the corresponding hardware at runtime (e.g., GitHub Actions ubuntu-latest with `-DUSE_IBVERBS=ON`). **Root cause:** `BaseTest::spawn()` calls `GTEST_SKIP()` from worker threads when `createDevice()` returns nullptr for an unavailable transport. GTest assertion/skip macros are not thread-safe — concurrent calls from multiple threads race on GTest's internal`TestPartResultReporterInterface`, corrupting state. This leads to an exception during stack unwinding, triggering recursive `std::terminate()`. The crash manifests at the first IBVERBS test case (`AllgatherRing/AllgatherTest.VarNumPointer/360`) because all prior transport tests (TCP, TCP_LAZY, TCP_TLS) succeed, and IBVERBS is the first transport where `createDevice()` returns nullptr on a machine without RDMA hardware. **Fix:** Probe transport availability from the main test thread before spawning workers. If the transport is unavailable, `GTEST_SKIP()` is called from the main thread (where it is safe) and the test returns early. Per-thread device creation is preserved for socket address isolation, with a silent early return as a defensive fallback. Differential Revision: D95934130

meta-cla bot added the CLA Signed label Mar 6, 2026

meta-codesync bot added fb-exported meta-exported labels Mar 6, 2026

weifengpy requested a review from d4l3k March 6, 2026 20:45

gyllstromk force-pushed the export-D93759269 branch 2 times, most recently from 63c8173 to 0a5ee91 Compare March 6, 2026 23:59

gyllstromk force-pushed the export-D93759269 branch from 0a5ee91 to 046964b Compare March 9, 2026 15:40

gyllstromk force-pushed the export-D93759269 branch from 046964b to d0f6d20 Compare March 9, 2026 16:28

gyllstromk force-pushed the export-D93759269 branch from d0f6d20 to 1704537 Compare March 9, 2026 16:31

gyllstromk force-pushed the export-D93759269 branch from 1704537 to 17f8ded Compare March 9, 2026 17:23

gyllstromk force-pushed the export-D93759269 branch from 17f8ded to 80cf7d5 Compare March 9, 2026 18:11

gyllstromk force-pushed the export-D93759269 branch from 80cf7d5 to eabb08c Compare March 9, 2026 18:11

gyllstromk force-pushed the export-D93759269 branch from eabb08c to eac7298 Compare March 9, 2026 18:15

kapilsh mentioned this pull request Mar 10, 2026

Probe device for skip on main thread #497

Open

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Fix [[nodiscard]] build errors and BUCK deps across comms, gloo, caffe2#494

Fix [[nodiscard]] build errors and BUCK deps across comms, gloo, caffe2#494
gyllstromk wants to merge 1 commit intopytorch:mainfrom
gyllstromk:export-D93759269

gyllstromk commented Mar 6, 2026

Uh oh!

meta-codesync bot commented Mar 6, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

gyllstromk commented Mar 6, 2026

Uh oh!

meta-codesync bot commented Mar 6, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant