[chore] Allow configuring linking of NVRTC wrapper

[AlessioNetti]

This PR introduces two changes related to linking of the NVRTC wrapper:

1. Bugfix: PR 4898 introduced a dependency on the NVPTX compiler library within the NVRTC wrapper, but no linking options were added to the respective CMakeLists.txt files. This PR adds the appropriate linking flags.
2. Change: The NVRTC wrapper is linked against the static NVRTC libraries. These libraries are usually not available in general-purpose images (e.g., the PyTorch or TensorRT NGC images) but need to be installed explicitly via the CUDA toolkit. The dynamic NVRTC libraries, on the other hand, are usually provided out-of-the-box. As such, this PR introduces a --nvrtc_dynamic_linking flag for the build_wheel.py script, which (optionally) allows using the dynamic NVRTC libraries as opposed to the static ones.

[Copilot]

Pull Request Overview

This PR addresses linking issues with the NVRTC wrapper by (1) adding the missing NVPTX dependency to the NVRTC wrapper target and (2) switching from static to dynamic linking for NVRTC libraries to better align with available toolkit distributions.

• Added NVPTX_LIB to the NVRTC wrapper's target_link_libraries
• Updated global CMake variables to use dynamic NVRTC libraries and introduced NVPTX_LIB for proper linkage

Reviewed Changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
cpp/tensorrt_llm/kernels/decoderMaskedMultiheadAttention/decoderXQAImplJIT/nvrtcWrapper/CMakeLists.txt	Added NVPTX_LIB in the target link libraries to fix missing dependency linking
cpp/CMakeLists.txt	Changed NVRTC linking from static to dynamic libraries and set up NVPTX_LIB appropriately

Comments suppressed due to low confidence (2)

cpp/tensorrt_llm/kernels/decoderMaskedMultiheadAttention/decoderXQAImplJIT/nvrtcWrapper/CMakeLists.txt:20

• Ensure that the addition of ${NVPTX_LIB} integrates well with the rest of the CUDA libraries and that the link ordering honors dependency requirements across different build configurations.

nvrtc_wrapper_src PUBLIC ${NVPTX_LIB} ${NVRTC_LIB} ${NVRTC_BUILTINS_LIB} ${CUDA_DRV_LIB}

cpp/CMakeLists.txt:142

• Review that switching NVRTC and NVRTC_BUILTINS_LIB to dynamic linking (lines 143-144) is consistent with expected usage across all configurations, and that dependent components are compatible with these changes.

set(NVPTX_LIB CUDA::nvptxcompiler_static)

[Funatiq]

@tongyuantongyu could you please take a look at these changes?

[tongyuantongyu]

LGTM.

[Funatiq]

/bot run

[tensorrt-cicd]

PR_Github #8787 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #8787 [ run ] completed with state FAILURE
/LLM/main/L0_MergeRequest_PR pipeline #6383 completed with status: 'FAILURE'

[Funatiq]

/bot run

[tensorrt-cicd]

PR_Github #8803 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #8803 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #6391 completed with status: 'FAILURE'

[lowsfer]

By design, we meant to statically link all three libs, but somehow we missed two of them.

Our intent was to link nvrtc/ptx compilers statically and avoid runtime dependency on these *.so, which can causes complex compatibility problems among nvrtx, ptx compiler, and driver.

Our build script will automatically install the required static libs for building.

[AlessioNetti]

I just implemented a --nvrtc_dynamic_linking option for the build_wheel.py script, which allows to optionally select the dynamic NRTC libraries for linking of the NVRTC wrapper. This is disabled by default, meaning that the static libraries are used unless explicitly requested otherwise.

[Funatiq]

/bot run

[tensorrt-cicd]

PR_Github #9184 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #9184 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #6729 completed with status: 'FAILURE'

[Funatiq]

/bot run

[tensorrt-cicd]

PR_Github #9211 [ run ] triggered by Bot

[Funatiq]

/bot run

[tensorrt-cicd]

PR_Github #9737 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #9737 [ run ] completed with state FAILURE
/LLM/main/L0_MergeRequest_PR pipeline #7171 completed with status: 'FAILURE'

[Funatiq]

/bot run

[tensorrt-cicd]

PR_Github #9820 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #9820 [ run ] completed with state FAILURE
/LLM/main/L0_MergeRequest_PR pipeline #7246 completed with status: 'FAILURE'

[Funatiq]

/bot run

[tensorrt-cicd]

PR_Github #9921 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #9921 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #7324 completed with status: 'SUCCESS'