Drop CUDA::cuda_driver link so mbpt.exe launches on CPU-only nodes

[gauravharsha]

Problem

When green-mbpt's mbpt.exe is built with green-gpu as a CUSTOM_KERNEL, the resulting binary fails to even launch on CPU-only nodes with:

libcuda.so.1: cannot open shared object file: No such file or directory

This happens even for pure-CPU runs that never touch the GPU, which prevents shipping a single binary usable on both CPU-only and GPU nodes.

Root cause

src/CMakeLists.txt links CUDA::cuda_driver (i.e. libcuda.so.1, the NVIDIA driver library, which only exists on nodes with the GPU driver installed). But green-gpu makes no CUDA driver API calls — the only cu* symbols anywhere in the source are the inline cuComplex.h helpers (cuCadd/cuCmul/cuCreal/…), which come from a header, not from libcuda. Everything else is CUDA runtime API + cuBLAS + cuSolver.

Because accel-lib/gpu are static libraries absorbed into mbpt.exe, that spurious link put a hard DT_NEEDED on libcuda.so.1 into the final executable, so the dynamic loader demanded the driver at startup.

Fix

Remove CUDA::cuda_driver from both link lines. The shared CUDA runtime loads the driver lazily on the first CUDA call, so the binary starts on CPU-only nodes and only needs the driver if GPU code actually runs.

Verification

readelf -d ./mbpt.exe | grep -i libcuda.so.1   # now empty

Confirmed by the reporter that a CUDA-enabled mbpt.exe built against this change launches on CPU-only nodes.

Note: if future code adds real driver-API calls (cuInit/cuCtxCreate/cuLaunchKernel/…), the link should be restored — or better, the kernel loaded as a dlopen plugin — rather than reintroducing a hard startup dependency.

🤖 Generated with Claude Code

[gauravharsha]

Good catch on cuDeviceGetAttribute — that driver-API call (added in 2a4bc34 for CUDA 13, after cudaDeviceProp::memoryClockRate was removed) was the one remaining thing forcing a libcuda link, which would have defeated this PR.

Fixed in 27b64b3 by switching to the runtime-API equivalent cudaDeviceGetAttribute(&memClockKHz, cudaDevAttrMemoryClockRate, i). This is NVIDIA's documented CUDA 13 replacement for the removed memoryClockRate field (see the cuda-samples CHANGELOG), returns the same kHz value, is available across CUDA versions, and links only against libcudart. No driver-API symbols remain, so mbpt.exe needs no libcuda at startup.

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 96.71%. Comparing base (23c530c) to head (27b64b3).

Additional details and impacted files

@@           Coverage Diff           @@
##             main      #18   +/-   ##
=======================================
  Coverage   96.71%   96.71%           
=======================================
  Files          13       13           
  Lines         852      852           
=======================================
  Hits          824      824           
  Misses         28       28           

Flag	Coverage Δ
unittests	96.71% <100.00%> (ø)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Copilot]

Pull request overview

This PR removes an unnecessary hard link dependency on the CUDA driver library (libcuda.so.1) so CUDA-enabled binaries that embed green-gpu can still launch on CPU-only nodes (and only require the NVIDIA driver if GPU code actually executes).

Changes:

• Drop CUDA::cuda_driver from accel-lib and gpu link lines to avoid a DT_NEEDED entry for libcuda.so.1 at program start.
• Replace the one CUDA driver API usage (cuDeviceGetAttribute) with the CUDA runtime API equivalent (cudaDeviceGetAttribute) in device-info printing code.

Reviewed changes

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

File	Description
src/cuda_check.cpp	Removes the lone CUDA driver API call by switching to the runtime attribute query API.
src/CMakeLists.txt	Removes CUDA::cuda_driver from library linkage to prevent startup-time dependency on libcuda.so.1.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[egull]

Looks good and is actually quite a useful convenience feature. I'm approving and merging.