[Bug] Native Jetson Orin (sm_87) build: explicit -DCMAKE_CUDA_ARCHITECTURES ignored + CuTe DSL --wrap/cudart shim not propagated to static-lib consumers

[suharvest]

Describe the bug

On a native Jetson Orin (sm_87, JetPack 6 / CUDA 12.6) build — compiling on the device itself with -DEMBEDDED_TARGET=jetson-orin, not the aarch64 cross toolchain — two build-system issues prevent the project from building/linking out of the box. Impact: blocker for native Orin builds.

1. The top-level default CUDA architecture list clobbers an explicit -DCMAKE_CUDA_ARCHITECTURES.
CMakeLists.txt sets CMAKE_CUDA_ARCHITECTURES = 80;86;89[;100a;120] whenever AARCH64_BUILD is undefined. A native Orin build is not a cross build, so AARCH64_BUILD is unset and the datacenter arch list is forced even when the user configures with -DCMAKE_CUDA_ARCHITECTURES=87. The build then targets the wrong SM.

2. The CUDA<12.8 _cudaLaunchKernelEx wrap and the cutedsl cudart shim are not propagated to the final executable when CuTe DSL is linked through a STATIC_LIBRARY.
cmake/CuteDsl.cmake (cute_dsl_setup()) attaches the cutedsl cudart shim and -Wl,--wrap=_cudaLaunchKernelEx with PRIVATE. A static archive performs no link step, so for STATIC_LIBRARY link targets neither reaches the consuming executable, and the CuTe DSL GEMM path fails to link/resolve on JetPack 6 / CUDA 12.6 (wrapped _cudaLaunchKernelEx / cudaLibrary_t).

Steps/Code to reproduce bug

Build configuration:

# On a Jetson Orin (sm_87, JetPack 6 / CUDA 12.6), native (on-device) build:
cmake -B build -DCMAKE_BUILD_TYPE=Release \
  -DEMBEDDED_TARGET=jetson-orin \
  -DCMAKE_CUDA_ARCHITECTURES=87 \
  -DTRT_PACKAGE_DIR=/path/to/TensorRT
cmake --build build -j
# (1) Configure still resolves CMAKE_CUDA_ARCHITECTURES to 80;86;89 (+100a;120), not 87.
# (2) Linking the CuTe DSL static-library consumer fails on the wrapped _cudaLaunchKernelEx.

Expected behavior

• An explicit -DCMAKE_CUDA_ARCHITECTURES (e.g. 87) is respected; the project's default arch set applies only when the user does not provide one. Regular x86 builds are unchanged.
• The cutedsl cudart shim and the --wrap linker option reach the final executable when CuTe DSL is consumed via a static library, so JetPack 6 / CUDA 12.6 builds link.

System information (Edge Device)

• Platform: NVIDIA Jetson Orin
• Software release: JetPack 6 (CUDA 12.6)
• CPU architecture: aarch64
• GPU compute capability: SM87
• Build type: Release
• Library versions:
  • TensorRT Edge-LLM version or commit hash: v0.8.0 (f9cc746)
  • CUDA: 12.6
• CMake options used:
  • EMBEDDED_TARGET: jetson-orin
  • CMAKE_CUDA_ARCHITECTURES: 87

A fix is ready and will be submitted as a PR referencing this issue.

[nvluxiaoz]

For native JP 6.x, please use this as outlined here: https://nvidia.github.io/TensorRT-Edge-LLM/user_guide/getting_started/installation.html

cmake .. \
    -DCMAKE_BUILD_TYPE=Release \
    -DTRT_PACKAGE_DIR=/usr \
    -DCMAKE_TOOLCHAIN_FILE=cmake/aarch64_linux_toolchain.cmake \
    -DEMBEDDED_TARGET=jetson-orin \
    -DCUDA_CTK_VERSION=12.6 \
    -DENABLE_CUTE_DSL=ALL

[rickwalder]

@nvluxiaoz That build command fails for me building both 0.8.0 and 0.9.0 on Jetson AGX Orin running JP6.2.1 -- see documented_build_command_output.txt (for 0.9.0)

The linked fix in #118 builds for me on 0.8.0 (suharvest's forked version) with the command:

mkdir -p build && cd build && \
    cmake .. \
      -DCMAKE_BUILD_TYPE=Release \
      -DTRT_PACKAGE_DIR=/usr \
      -DEMBEDDED_TARGET=jetson-orin \
      -DCUDA_CTK_VERSION=12.6 \
      -DCMAKE_CUDA_ARCHITECTURES=87 \
      -DCUTE_DSL_ARTIFACT_TAG=sm_87 \
      -DENABLE_CUTE_DSL=ALL \
      -DBUILD_PYTHON_BINDINGS=ON \
      -DCMAKE_C_COMPILER=/usr/bin/gcc \
      -DCMAKE_CXX_COMPILER=/usr/bin/g++ \
      -DCMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc \
      -DCUDA_DRIVER_LIB=/usr/local/cuda/targets/aarch64-linux/lib/stubs/libcuda.so && \
    cmake --build . -j"$(nproc)"

The above command ALSO fails while building vanilla 0.9.0 with what appears to be the same failure -- see modified_build_command_output.txt

So native JP6.2 support doesn't seem to work 100%, at least in our cases

Additionally, I've found i need to patch engine.py due to a signature mismatch -- see patch_engine.py

Here's the full dockerfile that works to build @suharvest's patched branch (the one in #118) for JP6.2.1 -- dockerfile.txt. A simple docker build . will build this without issue on an AGX running 6.2.1

Please let me know if I should create issues for any or all of this. Thank you!