Merge branch 'master' into dependabot-pip-requirements-ipython-all--lt-9.5.0

Author: Borda

requirements/pytorch/test.txt

@@ -7,7 +7,7 @@ pytest-random-order ==1.2.0
 
 # needed in tests
 cloudpickle >=1.3, <3.2.0
-scikit-learn >0.22.1, <1.7.0
+scikit-learn >0.22.1, <1.8.0
 numpy >=1.17.2, <1.27.0
 onnx >=1.12.0, <1.19.0
 onnxruntime >=1.12.0, <1.23.0

requirements/typing.txt

@@ -1,5 +1,5 @@
 mypy==1.17.1
-torch==2.7.1 # todo: update typing in separate PR
+torch==2.8.0
 
 types-Markdown
 types-PyYAML

src/lightning/fabric/plugins/collectives/torch_collective.py

@@ -24,6 +24,8 @@ class TorchCollective(Collective):
     """
 
     manages_default_group = False
+    addr_key = "MASTER_ADDR"
+    port_key = "MASTER_PORT"
 
     def __init__(self) -> None:
         if not dist.is_available():
@@ -136,26 +138,21 @@ def setup(self, main_address: Optional[str] = None, main_port: Optional[str] = N
         if self.is_initialized():
             return self
         # maybe set addr
-        set_addr = False
-        addr_key = "MASTER_ADDR"
-        if main_address is not None and addr_key not in os.environ:
-            os.environ[addr_key] = main_address
-            set_addr = True
+        setting_env = []
+        if main_address is not None and self.addr_key not in os.environ:
+            os.environ[self.addr_key] = main_address
+            setting_env.append(self.addr_key)
         # maybe set port
-        set_port = False
-        port_key = "MASTER_PORT"
-        if main_port is not None and port_key not in os.environ:
-            os.environ[port_key] = str(main_port)
-            set_port = True
+        if main_port is not None and self.port_key not in os.environ:
+            os.environ[self.port_key] = str(main_port)
+            setting_env.append(self.port_key)
         # this will `init_group`
         super().setup(**kwargs)
         # set as a class attribute so any instance can know whether we initialized the default process group
         TorchCollective.manages_default_group = True
         # cleanup
-        if set_addr:
-            os.environ.pop("MASTER_ADDR", None)
-        if set_port:
-            os.environ.pop("MASTER_PORT", None)
+        for kenv in setting_env:
+            os.environ.pop(kenv, None)
         return self
 
     @override

src/lightning/fabric/utilities/distributed.py

@@ -319,11 +319,7 @@ def _destroy_dist_connection() -> None:
 
 
 def _get_default_process_group_backend_for_device(device: torch.device) -> str:
-    """Return corresponding distributed backend for a given device."""
-    device_backend_map = torch.distributed.Backend.default_device_backend_map
-    if device.type in device_backend_map:
-        return device_backend_map[device.type]
-    return "gloo"
+    return "nccl" if device.type == "cuda" else "gloo"
 
 
 class _DatasetSamplerWrapper(Dataset):

tests/tests_fabric/strategies/launchers/test_multiprocessing_integration.py

@@ -46,7 +46,8 @@ def test_memory_sharing_disabled(strategy):
 
 def _test_memory_sharing_disabled(fabric, tensor, model):
     is_spawn = fabric.strategy.launcher._start_method == "spawn"
-    assert not is_spawn or tensor.is_shared()
+    if is_spawn:
+        assert tensor.is_shared()
     assert not model.layer.weight.is_shared()
     assert not model.tied_layer.weight.is_shared()
     assert not model.buffer.is_shared()

tests/tests_fabric/utilities/test_distributed.py

@@ -17,7 +17,6 @@
 from lightning.fabric.utilities.distributed import (
     _destroy_dist_connection,
     _gather_all_tensors,
-    _get_default_process_group_backend_for_device,
     _InfiniteBarrier,
     _init_dist_connection,
     _is_dtensor,
@@ -244,27 +243,6 @@ def test_init_dist_connection_registers_destruction_handler(_, atexit_mock):
     atexit_mock.register.assert_not_called()
 
 
-def test_get_default_process_group_backend_for_device():
-    """Test that each device type maps to its correct default process group backend."""
-    # register a custom backend for test
-    torch.utils.rename_privateuse1_backend("pcu")
-
-    def mock_backend(store, group_rank, group_size, timeout):
-        pass
-
-    torch.distributed.Backend.register_backend(
-        "pccl",
-        lambda store, group_rank, group_size, timeout: mock_backend(store, group_rank, group_size, timeout),
-        devices=["pcu"],
-    )
-
-    # test that the default backend is correctly set for each device
-    devices = [torch.device("cpu"), torch.device("cuda:0"), torch.device("pcu:0")]
-    backends = ["gloo", "nccl", "pccl"]
-    for device, backend in zip(devices, backends):
-        assert _get_default_process_group_backend_for_device(device) == backend
-
-
 @RunIf(min_torch="2.4")
 def test_is_dtensor(monkeypatch):
     from torch.distributed._tensor import DTensor