diff --git a/pkgs/development/cuda-modules/README.md b/pkgs/development/cuda-modules/README.md
index 7f89552ea13c..e4cf737e9f9c 100644
--- a/pkgs/development/cuda-modules/README.md
+++ b/pkgs/development/cuda-modules/README.md
@@ -38,8 +38,6 @@ not straightforward to include. These packages are:
   - `mft_oem`: unsure of purpose; contains FHS paths.
   - `mft`: unsure of purpose; contains FHS paths.
   - `nvidia_driver`: we don't use the drivers from the CUDA releases; irrelevant.
-- `cublasmp`:
-  - `libcublasmp`: `nvshmem` isn't packaged.
 - `cudnn`:
   - `cudnn_samples`: requires FreeImage, which is abandoned and not packaged.
 
diff --git a/pkgs/development/cuda-modules/packages/libcublasmp.nix b/pkgs/development/cuda-modules/packages/libcublasmp.nix
index 6e51b76cd48c..fb171c29d039 100644
--- a/pkgs/development/cuda-modules/packages/libcublasmp.nix
+++ b/pkgs/development/cuda-modules/packages/libcublasmp.nix
@@ -1,43 +1,80 @@
 {
-  lib,
+  backendStdenv,
   buildRedist,
+  cudaAtLeast,
+  cudaMajorVersion,
+  lib,
   libcublas,
   nccl,
 }:
-buildRedist {
-  redistName = "cublasmp";
-  pname = "libcublasmp";
+buildRedist (
+  finalAttrs:
+  let
+    inherit (backendStdenv) cudaCapabilities;
+    cublasmpAtLeast = lib.versionAtLeast finalAttrs.version;
+  in
+  {
+    redistName = "cublasmp";
+    pname = "libcublasmp";
 
-  outputs = [
-    "out"
-    "dev"
-    "include"
-    "lib"
-  ];
+    outputs = [
+      "out"
+      "dev"
+      "include"
+      "lib"
+    ];
 
-  # TODO: Looks like the minimum supported capability is 7.0 as of the latest:
-  # https://docs.nvidia.com/cuda/cublasmp/getting_started/index.html
-  # NOTE: cuBLASMp has not depended on NVSHMEM since 0.8.0:
-  # https://docs.nvidia.com/cuda/cublasmp/release_notes/index.html
-  buildInputs = [
-    libcublas
-    nccl
-  ];
+    # NOTE: cuBLASMp has not depended on NVSHMEM since 0.8.0:
+    # https://docs.nvidia.com/cuda/cublasmp/release_notes/index.html
+    buildInputs = [
+      libcublas
+      nccl
+    ];
 
-  autoPatchelfIgnoreMissingDeps = [
-    "libcuda.so.1"
-  ];
+    autoPatchelfIgnoreMissingDeps = [
+      "libcuda.so.1"
+    ];
 
-  meta = {
-    description = "High-performance, multi-process, GPU-accelerated library for distributed basic dense linear algebra";
-    longDescription = ''
-      NVIDIA cuBLASMp is a high-performance, multi-process, GPU-accelerated library for distributed basic dense linear
-      algebra.
+    # https://docs.nvidia.com/cuda/cublasmp/getting_started/index.html#hardware-and-software-requirements
+    platformAssertions =
+      let
+        minCudaCapability = if cudaAtLeast "13.0" then "7.5" else "7.0";
+      in
+      [
+        {
+          message =
+            "cuBLASMp for CUDA ${cudaMajorVersion} supports CUDA compute capabilities ${minCudaCapability} and newer"
+            + " (found ${builtins.toJSON cudaCapabilities})";
+          assertion = lib.all (lib.flip lib.versionAtLeast minCudaCapability) cudaCapabilities;
+        }
+        {
+          message = "cuBLASMp requires NCCL, which is unavailable for this platform";
+          assertion = nccl.meta.available;
+        }
+      ];
 
-      cuBLASMp is compatible with 2D block-cyclic data layout and provides PBLAS-like C APIs.
-    '';
-    homepage = "https://docs.nvidia.com/cuda/cublasmp";
-    changelog = "https://docs.nvidia.com/cuda/cublasmp/release_notes";
-    license = lib.licenses.nvidiaMath_sdk_sla;
-  };
-}
+    # NOTE: The requirement was raised to NCCL 2.29.2 when cuBLASMp 0.8.0 replaced NVSHMEM with NCCL symmetric memory:
+    # https://web.archive.org/web/20260215220513/https://docs.nvidia.com/cuda/cublasmp/getting_started/index.html
+    brokenAssertions = [
+      {
+        message =
+          "cuBLASMp releases since 0.8.0 (found ${finalAttrs.version})"
+          + " require NCCL 2.29.2 or newer (found ${nccl.version})";
+        assertion = cublasmpAtLeast "0.8" -> lib.versionAtLeast nccl.version "2.29.2";
+      }
+    ];
+
+    meta = {
+      description = "High-performance, multi-process, GPU-accelerated library for distributed basic dense linear algebra";
+      longDescription = ''
+        NVIDIA cuBLASMp is a high-performance, multi-process, GPU-accelerated library for distributed basic dense linear
+        algebra.
+
+        cuBLASMp is compatible with 2D block-cyclic data layout and provides PBLAS-like C APIs.
+      '';
+      homepage = "https://docs.nvidia.com/cuda/cublasmp";
+      changelog = "https://docs.nvidia.com/cuda/cublasmp/release_notes";
+      license = lib.licenses.nvidiaMath_sdk_sla;
+    };
+  }
+)