[DEV] Support Marlin AWQ/GPTQ repack and CUDA graph-safe Add on NVIDIA

[qinyiqun]

Background

Marlin AWQ/GPTQ weight preprocessing needs GPU-side repack support in InfiniCore so InfiniLM can avoid CPU repack during model loading.

CUDA graph replay also exposed an Add issue: the generic elementwise path captured a temporary host vector used for input pointers, which could become invalid during graph replay.

Scope

• Add NVIDIA-only AWQ/GPTQ Marlin repack operators.
• Ensure non-NVIDIA platforms do not compile Marlin CUDA kernels.
• Make NVIDIA Add safe for CUDA graph capture/replay while preserving existing non-NVIDIA backend interfaces.

Expected Outcome

InfiniLM can use GPU Marlin repack on NVIDIA, and Add no longer causes CUDA graph replay failures from captured temporary host input arrays.