[leconcio]

Why

PyTorch 2.14 is now an official ROCm Docker release for ROCm 10.1 and above, and on it apex ops fail to build, whether they are JIT-built on first use or prebuilt with APEX_BUILD_*=1.

torch.utils.cpp_extension compiles extensions with -std=c++20 since PyTorch 2.12, but CUDAOpBuilder appends -std=c++17 to the host C++ flags (cxx_args()) and to the ROCm hipcc flags (nvcc_args()). The later flag wins, so every op is compiled as C++17. That still worked on 2.12 and 2.13, whose headers accept C++17, but the PyTorch 2.14 headers require C++20:

ATen/ATen.h: error: #error C++20 or later compatible compiler is required to use ATen.
c10/util/intrusive_ptr.h:775:27: error: 'strong_ordering' in namespace 'std' does not name a type
ATen/core/TensorBase.h:1031:5: error: 'requires' does not name a type

How

Remove every hard-coded -std flag from CUDAOpBuilder, and let torch.utils.cpp_extension choose the standard that matches the installed PyTorch headers:

• cxx_args(): drop -std=c++17.
• ROCm branch of nvcc_args(): drop -std=c++17.
• CUDA branch of nvcc_args(): drop the CUDA-version-based selection (c++20 on CUDA 12.5+, c++17 on CUDA 11 to 12.4, c++14 on older).

torch.utils.cpp_extension adds its own -std whenever none is passed, on both the JIT path and the setup.py path. This is the approach upstream apex took in NVIDIA#2006, and apex no longer needs updating each time PyTorch raises its required standard.

After this change apex sets no -std anywhere. Builders that override cxx_args() or nvcc_args() without calling super() (bnp, and the hipcc flags of amp_C, fused_adam, syncbn, ...) already passed none.

Testing

All runs on MI300X (gfx942) with ROCm images, comparing master (2dfd342) with this branch.

Setup	master	This PR
PyTorch 2.14, AOT: APEX_BUILD_CPP_OPS=1 APEX_BUILD_CUDA_OPS=1 python setup.py build_ext	Fails with the errors above, 0 of 28 ops built	28 of 28 ops built, 0 errors; all 77 compile commands use PyTorch's -std=c++20
PyTorch 2.14, JIT: pip install ., then jit_load() on every builder	27 of 28 fail with the errors above (only bnp, which passes no -std, builds)	28 of 28 ops built; the generated build.ninja files contain only -std=c++20
amp_C.multi_tensor_l2norm on a 1024-element tensor of ones, PyTorch 2.14	Doesn't build	Returns 32.0 (expected)

L0 tests on PyTorch 2.14 with this branch (JIT install, APEX_TEST_WITH_ROCM=1 APEX_SKIP_FLAKY_TEST=1):

Suite	Result
run_amp	70 ran, OK (9 skipped)
run_fp16util	2 ran, OK
run_optimizers	39 ran, OK (9 skipped)
run_fused_layer_norm	84 of 84 non-ONNX tests passed
run_mlp	6 ran, OK (4 skipped)
run_fused_dense	2 ran, OK
run_transformer	102 ran, OK (44 skipped), including the GPT, BERT and dynamic batch size minimal tests on 8 GPUs

Image: rocm/frameworks-internal:rocm10.2.0a20261008_ubuntu24.04_py3.12_pytorch_2.14.0_5d7f2f0 (torch 2.14.0+rocm10.2.0a20261008, HIP 7.17).

Not tested: CUDA builds. On CUDA 12.0 to 12.4 this changes the standard from c++17 to whatever the installed PyTorch defaults to.

🤖 Generated with Claude Code

Cherry-picked to release/1.14.0 branch via #342

Cherry-picked to release/1.15.0 branch via #343

[jithunnair-amd]

! cherry-pick --onto release/1.14.0,release/1.15.0

[rocm-repo-management-api-6]

Target branch release/1.14.0,release/1.15.0 doesn't exist

Comment processed by Build

[jithunnair-amd]

! cherry-pick --onto release/1.14.0 release/1.15.0

[rocm-repo-management-api-6]

Created branch autogenerated/release/1.14.0_cherry-pick_pr-341 and #342

Created branch autogenerated/release/1.15.0_cherry-pick_pr-341 and #343

Comment processed by Build