From cf516f8788bf37f543820ca8a465a581002f2eff Mon Sep 17 00:00:00 2001
From: Jiacheng Xu <13862180016@163.com>
Date: Wed, 22 Jul 2026 14:16:52 +0800
Subject: [PATCH] cmake: require cuSolverMp 0.9.0

---
 cmake/modules/SetupCuSolverMp.cmake | 15 +++++++++++----
 docs/advanced/acceleration/cuda.md  |  8 ++++++++
 2 files changed, 19 insertions(+), 4 deletions(-)

diff --git a/cmake/modules/SetupCuSolverMp.cmake b/cmake/modules/SetupCuSolverMp.cmake
index 7ad1af4789..6ce2be1899 100644
--- a/cmake/modules/SetupCuSolverMp.cmake
+++ b/cmake/modules/SetupCuSolverMp.cmake
@@ -44,12 +44,19 @@ function(abacus_setup_cusolvermp)
     endif()
   endif()
 
-  # Check minimum version requirement (>= 0.4.0)
-  if(CUSOLVERMP_VERSION_STR AND CUSOLVERMP_VERSION_STR VERSION_LESS "0.4.0")
+  # Check minimum version requirement (>= 0.9.0)
+  if(NOT CUSOLVERMP_VERSION_STR)
+    message(FATAL_ERROR
+      "Unable to detect the cuSOLVERMp version from ${CUSOLVERMP_VERSION_HEADER}. "
+      "ABACUS requires cuSOLVERMp >= 0.9.0."
+    )
+  elseif(CUSOLVERMP_VERSION_STR VERSION_LESS "0.9.0")
     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."
+      "NVIDIA documents an STEDC defect in cuSOLVERMp 0.4.2 through 0.8.0 that affects "
+      "non-power-of-two block sizes and certain 2D process grids; Syevd and "
+      "Sygvd use STEDC internally. "
+      "ABACUS requires cuSOLVERMp >= 0.9.0. Please upgrade cuSOLVERMp."
     )
   endif()
 
diff --git a/docs/advanced/acceleration/cuda.md b/docs/advanced/acceleration/cuda.md
index 42c09859d7..7be2c2a243 100644
--- a/docs/advanced/acceleration/cuda.md
+++ b/docs/advanced/acceleration/cuda.md
@@ -24,6 +24,14 @@ To compile and use ABACUS in CUDA mode, you currently need to have an NVIDIA GPU
 
 - Install a driver and toolkit appropriate for your system (SDK is not necessary)
 
+NVIDIA reports that cuSOLVERMp 0.4.2 through 0.8.0 contain an STEDC defect
+affecting non-power-of-two block sizes and certain 2D process grids. Because
+`Syevd` and `Sygvd` use STEDC internally, affected versions may fail or hang
+during distributed diagonalization. ABACUS therefore requires cuSOLVERMp 0.9.0
+or newer when `ENABLE_CUSOLVERMP=ON`. The recommended stack is cuSOLVERMp 0.9.0
+with cuBLASMp 0.9.1. See the
+[cuSOLVERMp 0.9.0 release notes](https://docs.nvidia.com/cuda/cusolvermp/release_notes/index.html#cusolvermp-v0-9-0)
+for the upstream fix details.
 
 ## Building ABACUS with the GPU support: