ci(cuda): build only for CI runner GPU arch and raise parallelism

[MrLi000001]

The CUDA CI built every .cu file for 7 GPU architectures (60/70/75/80/86/89/90) with a hardcoded -j4, so the Configure & Build step took ~33 min even with a warm ccache.

• Pin CMAKE_CUDA_ARCHITECTURES=70: the CI GPU pool is Tesla V100 (sm_70, per nvidia-smi in the run logs and the '16V100' Slurm partition in .ci/slurm/config.ini). This cuts nvcc work by ~7x.
• Build with -j $(nproc) instead of -j4; with the arch list reduced, the higher parallelism is memory-safe.

Expected: Configure & Build ~33 min -> ~10 min on a cache-cold run.

Reminder

• I have read AGENTS.md and docs/developers_guide/agent_governance.md.
• I have linked an issue or explained why this PR does not need one.
• I have added adequate unit tests and/or case tests, or explained why not.
• I have listed the exact verification commands run and their results.
• I have described user-visible behavior changes, including INPUT parameter changes.
• I have explained core-module impact for ESolver, HSolver, ElecState, Hamilt, Operator, Psi, or other source/ changes.
• I have requested any needed governance exception below.

Linked Issue

Fix #

Unit Tests and/or Case Tests for my changes

• Commands run:
• Result summary:
• Checks not run, with reason:

What's changed?

• Example: brief summary of the user-visible or developer-facing change.

Governance Notes

• INPUT/docs changes:
• Core module impact:
• Exceptions requested:

[Stardust0831]

This is an effective direction: Configure & Build dropped from about 33 minutes to 18m01s in the actual run. Before merging, please confirm that the entire gpu runner pool guarantees sm_70 and update the ~10-minute estimate with the measured result; otherwise, a heterogeneous runner may not be able to run the binary.

[Critsium-xy]

The auto-detection block in CMakeLists.txt is dead code — it never runs, and even if it did, its result would be discarded. Two independent defects, checked against the head commit 167904f:

1. if(COMMAND nvidia-smi) is always false.

if(NOT DEFINED CMAKE_CUDA_ARCHITECTURES AND USE_CUDA AND COMMAND nvidia-smi)

if(COMMAND <name>) tests whether a CMake command, macro or function of that name exists — it has nothing to do with executables on PATH. nvidia-smi is not a CMake command, so the condition is unconditionally false and the execute_process never executes. The external-program test is find_program:

find_program(NVIDIA_SMI_EXECUTABLE nvidia-smi)
if(NVIDIA_SMI_EXECUTABLE)
  execute_process(COMMAND ${NVIDIA_SMI_EXECUTABLE} --query-gpu=compute_cap ...)

(AND USE_CUDA is also redundant — this is already inside the if(USE_CUDA) at line 419.)

2. Even with that fixed, the detected value is immediately overwritten.

The detection block and the historical default list run in sequence, with no else() and no early exit:

if(NOT DEFINED CMAKE_CUDA_ARCHITECTURES)          # outer guard
    if(... nvidia-smi ...)
      set(CMAKE_CUDA_ARCHITECTURES "70" CACHE STRING ...)   # cache entry
    endif()

    if(CUDAToolkit_VERSION VERSION_LESS "13.0")
      set(CMAKE_CUDA_ARCHITECTURES 60 70)         # <-- normal variable, shadows the cache entry
    ...
    list(APPEND CMAKE_CUDA_ARCHITECTURES 75)      # 80 / 86 / 89 / 90 follow
endif()

set(... CACHE STRING) creates a cache entry; the very next set() creates a same-named normal variable that shadows it, and the list is appended as usual. The result is still all 7 architectures. Since the cache write makes DEFINED true, the minimal fix is to give the historical default its own guard:

if(NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
  # detection (via find_program)
endif()

if(NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
  # historical multi-arch default
endif()

Consequence: the 33 min → 18m01s that @Stardust0831 measured comes entirely from -j4 → -j $(nproc) (~2x, which matches), not from any reduction in nvcc work. Worth re-measuring once the logic actually runs.

---

3. Design: this belongs in the CI workflow, not in CMakeLists.txt.

The title and description say "build only for CI runner GPU arch", but the diff changes the default for every user. On a cluster, people routinely configure on a login node whose GPU differs from — or is absent on — the compute node. Silent auto-detection would then produce a binary that dies with no kernel image is available for execution on the device, with no warning at configure time. That is a rough default for a code that gets built once and run on heterogeneous nodes.

Adding -DCMAKE_CUDA_ARCHITECTURES=70 to cuda.yml achieves the stated goal in one line with zero regression risk — the if(NOT DEFINED CMAKE_CUDA_ARCHITECTURES) guard at line 434 exists precisely for this. If auto-detection is still wanted, it should sit behind an option(...) defaulting to OFF.

This also settles @Stardust0831's question about whether the whole gpu pool is sm_70: an explicit pin is strictly better than detection, because on a heterogeneous pool detection makes each node emit different -gencode flags, which fragments the ccache key — and that works directly against the cache warmer in #7756.