API: Update pinning instructions in README

README.md

@@ -9,17 +9,17 @@ or build multiple variants of a conda package which each target one or a subset
 devices.
 
 Similar to how the virtual package `__cuda` constrains the `cuda-version` metapackage to
-represents to conda the CUDA *driver* version available on the system. This plugin creates a
+represents to conda the CUDA *driver* version available on the system. This plugin creates
 the virtual package `__cuda_arch` which constrains the `cuda-arch` metapackage which
 represents to conda the minimum compute capability of all CUDA *devices* on the system.
 
 ## The `cuda-arch` metapackage
 
 Recipes and packages cannot depend on a specific version of `__cuda` or `__cuda_arch`
-directly we need to be able to build multiple variants of a package on the same machine
-without changing the hardware or driver. Creating a wrapper metapackage like `cuda-version`
-or `cuda-arch` which is only `run_constrained` by the corresponding virtual package lets us
-do this.
+directly because we need to be able to build multiple variants of a package on the same
+machine without changing the hardware or driver. Creating a wrapper metapackage like
+`cuda-version` or `cuda-arch` which is only `run_constrained` by the corresponding virtual
+package lets us do this.
 
 > [!IMPORTANT]
 > This plugin does not create the `cuda-arch` metapackage. This package must be created
@@ -32,6 +32,10 @@ package:
   name: cuda-arch
   version: {{ version }}
 
+build:
+  run_exports:
+    - cuda-arch >={{ version }}
+
 requirements:
   run_constrained:
     - __cuda_arch {{ version }}
@@ -41,7 +45,7 @@ One version is published for every known major-minor compute capability. Sub-arc
 such as `100f` are not expressible within this framework but can still be targeted at build
 time.
 
-## Implementing a conda-recipe which depends on `__cuda_arch`
+## Implementing a conda-recipe which depends on `cuda-arch`
 
 Define a `conda_build_config.yaml` to configure conda-build to build the recipe multiple
 times. This file will need variables providing the compiler flags, compute capabilities, and
@@ -127,11 +131,9 @@ requirements:
     - {{ compiler('cuda') }}
     - {{ stdlib('c') }}
   host:
-  # We must pin cuda-arch in the host environment to ensure that dependencies are also
-  # compatible with the desired cuda-arch
+  # We must pin cuda-arch in the host environment to the minimum supported cuda-arch to
+  # ensure that dependencies are also compatible with the minimum supported cuda-arch
     - cuda-arch {{ cuda_arch_min }}
-  run:
-    - cuda-arch >={{ cuda_arch_min }}
 
 ```