[Bug]: C++17 build cannot compile torch 2.12 headers with nvcc, but C++20 breaks float2(fp8) functional casts

System Info

• CPU architecture: x86_64
• GPU: not required to reproduce (compile-only)
• TensorRT-LLM commit: 4476ee3eacc1639c28f7e3e845b67ec246f2e482 (main)
• CUDA: 13.4, nvcc V13.4.59
• Host compiler: gcc 13.3.0 (Ubuntu 24.04)
• Container: nvcr.io/nvidia/cuda-dl-base:26.08-cuda13.4-devel-ubuntu24.04
• torch: 2.12.0+cu130 (from --extra-index-url https://download.pytorch.org/whl/cu130, as pinned by requirements.txt)

Who can help?

Build / C++ runtime.

Information about the problem

Building outside the prebuilt NGC container hits a C++ standard conflict that
originates entirely inside this repository:

• requirements.txt:29 pins torch>=2.12.0a0,<=2.13.0a0.
• torch 2.12 ships share/cmake/Torch/TorchConfig.cmake:158 with CXX_STANDARD 20,
  and share/cmake/Caffe2/public/utils.cmake:340 sets the same property on its targets.
  torch 2.12 therefore requires C++20.
• cpp/CMakeLists.txt:343 sets CMAKE_CXX_STANDARD 17, and line 346 derives
  CMAKE_CUDA_STANDARD from it.

The two requirements cannot both be satisfied:

At C++17, nvcc cannot compile torch 2.12 headers. Any .cu that transitively
includes ATen/core/function_schema.h fails:

ATen/core/List_inl.h:202:49: error: need 'typename' before
'decltype(((c10::List<T>*)this)->c10::List<T>::impl_->list)::difference_type'
because '...' is a dependent scope

This reaches several targets through pgUtils.h and thUtils.h, for example
tensorrt_llm/kernels/trtllmGenKernels/blockScaleMoe/runner.cu.

At C++20, torch compiles, but 45 functional casts in
cpp/tensorrt_llm/kernels/decoderMaskedMultiheadAttentionUtils.h break, producing
over 150 diagnostics:

decoderMaskedMultiheadAttentionUtils.h(1210): error: no suitable conversion function
from "const __nv_fp8x2_e4m3" to "float" exists

The cause is parenthesized aggregate initialization, added in C++20. float2(x) is a
functional cast in C++17 and selects __nv_fp8x2_e4m3::operator float2(). In C++20 the
same expression is an aggregate initialization of float2, so the compiler tries to
initialize the first member float x from the fp8 value and fails.

Reproduction

Three lines, no GPU and no TensorRT-LLM checkout required:

#include <cuda_fp8.h>
__global__ void k(const __nv_fp8x2_e4m3 a, float2* o) { *o = float2(a); }

nvcc -std=c++17 -arch=sm_90 -c repro.cu   # succeeds
nvcc -std=c++20 -arch=sm_90 -c repro.cu   # fails with the error above

The full build was run as:

python scripts/build_wheel.py --cpp_only --use_ccache \
  -a "90-real;100-real;103-real" -j 48 --no-venv

Suggested fix

Replacing the functional cast with an explicit static_cast compiles under both
standards and keeps the same conversion operator, so the generated code is unchanged:

-    float2 fb0 = float2(fp8_2[0]);
+    float2 fb0 = static_cast<float2>(fp8_2[0]);

Verified:

expression	C++17	C++20
float2(a)	compiles	fails
static_cast<float2>(a)	compiles	compiles

45 lines in decoderMaskedMultiheadAttentionUtils.h use the affected pattern. The
change is mechanical and safe at C++17, so it does not require the project to move to
C++20; it only removes the blocker if and when the standard is raised.

I am happy to open a PR with this change if the approach looks right.

Start in cpp/tensorrt_llm/kernels/decoderMaskedMultiheadAttentionUtils.h and review the 45 functional casts identified in the issue. Verify the minimal repro with nvcc under C++17 and C++20, update the affected casts as described, then run the cpp-only build to confirm both the torch headers and CUDA kernels compile.