[chore] Linking fixes to 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 changes linking of the NVRTC wrapper to use the dynamic NVRTC libraries.

[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'