Improve performance of available_accelerators by returning a set (#20672)

perf[registry]: update available_accelerators to return a set

Author: allrob23

src/lightning/fabric/accelerators/registry.py

@@ -107,9 +107,9 @@ def remove(self, name: str) -> None:
         """Removes the registered accelerator by name."""
         self.pop(name)
 
-    def available_accelerators(self) -> list[str]:
-        """Returns a list of registered accelerators."""
-        return list(self.keys())
+    def available_accelerators(self) -> set[str]:
+        """Returns a set of registered accelerators."""
+        return set(self.keys())
 
     def __str__(self) -> str:
         return "Registered Accelerators: {}".format(", ".join(self.available_accelerators()))

tests/tests_fabric/accelerators/test_registry.py

@@ -70,4 +70,4 @@ def is_available():
 
 
 def test_available_accelerators_in_registry():
-    assert ACCELERATOR_REGISTRY.available_accelerators() == ["cpu", "cuda", "mps", "tpu"]
+    assert ACCELERATOR_REGISTRY.available_accelerators() == {"cpu", "cuda", "mps", "tpu"}

tests/tests_pytorch/accelerators/test_registry.py

@@ -16,7 +16,7 @@
 
 def test_available_accelerators_in_registry():
     """Tests the accelerators available by default, not including external, third-party accelerators."""
-    available = set(AcceleratorRegistry.available_accelerators())
+    available = AcceleratorRegistry.available_accelerators()
     expected = {"cpu", "cuda", "mps", "tpu"}
     # Note: the registry is global, other tests may register new strategies as a side effect
     assert expected.issubset(available)