Docs: update cuSolverMP documentation for HPC SDK version differences (#6955)

- Add version-specific module loading instructions (HPC-X for SDK < 25.9, NCCL for SDK >= 25.9)
- Add minimum version check (>= 0.4.0) in CMake configuration
- Simplify documentation by removing manual installation option

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

Author: dzzz2001

cmake/SetupCuSolverMp.cmake

@@ -44,6 +44,15 @@ function(abacus_setup_cusolvermp target_name)
     endif()
   endif()
 
+  # Check minimum version requirement (>= 0.4.0)
+  if(CUSOLVERMP_VERSION_STR AND CUSOLVERMP_VERSION_STR VERSION_LESS "0.4.0")
+    message(FATAL_ERROR
+      "cuSolverMp version ${CUSOLVERMP_VERSION_STR} is too old. "
+      "ABACUS requires cuSolverMp >= 0.4.0 (NVIDIA HPC SDK >= 23.5). "
+      "Please upgrade your NVIDIA HPC SDK installation."
+    )
+  endif()
+
   # Auto-select communicator backend by cuSolverMp version.
   # cuSolverMp < 0.7.0 -> CAL, otherwise -> NCCL.
   set(_use_cal OFF)

toolchain/README.md

@@ -292,33 +292,27 @@ cmake -B $BUILD_DIR \
 
 #### Multi-GPU with cuSolverMP
 
-**Option 1: Using NVIDIA HPC SDK (Recommended)**
+cuSolverMP requires NVIDIA HPC SDK. Follow these steps to build with cuSolverMP:
 
 1. Load the NVHPC module:
-```bash
-module load nvhpc
-```
-
-2. Build with cuSolverMP enabled:
-```bash
-cmake -B $BUILD_DIR \
-    -DUSE_CUDA=ON \
-    -DENABLE_CUSOLVERMP=ON \
-    # ... other options
-```
-
-The module file will set `NVHPC_ROOT` automatically, and CMake will find cuSolverMP and NCCL(or CAL) from it.
 
-**Option 2: Manual Installation**
+   - **For NVIDIA HPC SDK version < 25.9**: cuSolverMP relies on HPC-X for communication, so you need to load the `nvhpc-hpcx-cudaxx/xx.x` module. For example, with HPC SDK 25.3:
+     ```bash
+     module use /opt/nvidia/hpc_sdk/modulefiles
+     module load nvhpc-hpcx-cuda12/25.3
+     ```
 
-If you installed cuSolverMP and NCCL separately and not in system default paths, specify the paths explicitly:
+   - **For NVIDIA HPC SDK version >= 25.9**: cuSolverMP uses NCCL for communication instead of HPC-X, so only the base `nvhpc/xx.x` module is needed. For example, with HPC SDK 26.1:
+     ```bash
+     module use /opt/nvidia/hpc_sdk/modulefiles
+     module load nvhpc/26.1
+     ```
 
+2. Build with cuSolverMP enabled:
 ```bash
 cmake -B $BUILD_DIR \
     -DUSE_CUDA=ON \
     -DENABLE_CUSOLVERMP=ON \
-    -DCUSOLVERMP_PATH=/path/to/cusolvermp/ \
-    -DNCCL_PATH=/path/to/nccl/ \
     # ... other options
 ```