build: select C++20 when building against torch 2.14+

torch 2.14 headers (NGC pytorch 26.09) require C++20, and setup.py hard-codes
-std=c++17, so the extension does not compile against it. Pick the standard
from the installed torch major/minor (nightly and local versions such as
2.14.0a0+nv parse the same way). Builds with torch < 2.14 get exactly the
previous flags.

The existing sm_100 cubin also runs on B300 (sm_103), so no new
architecture targets are needed.

Motif-Tests: wheel builds with torch 2.14 / CUDA 13 (NGC 26.09); full tests/ suite passes on 1x B300 with this change
Motif-Upstream: n/a (this is the upstream)
Motif-Remove-When: n/a

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

CLAUDE.md

@@ -21,7 +21,7 @@ python setup.py build_ext --inplace       # build only
 
 After `pip install -e .`, all imports work directly (`import activation`, `from activation.grouped_poly_norm import ...`). No PYTHONPATH manipulation needed.
 
-NVCC flags: `-O3 --use_fast_math -std=c++17`, targets sm_80 (A100), sm_89 (L40/4090), sm_90 (H100), sm_100 (B200, CUDA 12.8+).
+NVCC flags: `-O3 --use_fast_math -std=c++17` (`-std=c++20` with PyTorch >= 2.14), targets sm_80 (A100), sm_89 (L40/4090), sm_90 (H100), sm_100 (B200, CUDA 12.8+).
 
 ### CI / HuggingFace distribution build
 

setup.py

@@ -38,11 +38,16 @@ INCLUDE_DIRS = [
     str(ROOT / "torch-ext"),
 ]
 
+# PyTorch 2.14 headers require C++20; preserve C++17 for older releases.
+# Major/minor parsing also accepts nightly/local versions (e.g. 2.14.0a0+nv).
+torch_version = tuple(int(x) for x in torch.__version__.split(".")[:2])
+CXX_STD = "-std=c++20" if torch_version >= (2, 14) else "-std=c++17"
+
 # CUDA flags matching the existing kernel style
 NVCC_FLAGS = [
     "-O3",
     "--use_fast_math",
-    "-std=c++17",
+    CXX_STD,
     # Generate code for common architectures
     "-gencode=arch=compute_80,code=sm_80",  # A100
     "-gencode=arch=compute_89,code=sm_89",  # L40/4090
@@ -54,7 +59,7 @@ cuda_version = tuple(int(x) for x in torch.version.cuda.split(".")[:2])
 if cuda_version >= (12, 8):
     NVCC_FLAGS.append("-gencode=arch=compute_100,code=sm_100")
 
-CXX_FLAGS = ["-O3", "-std=c++17"]
+CXX_FLAGS = ["-O3", CXX_STD]
 
 ext_modules = [
     CUDAExtension(