Honor singular device= as an alias for devices= (fixes #907)

FlagEmbedding/abc/inference/AbsEmbedder.py

@@ -69,6 +69,16 @@ def __init__(
         self.use_bf16 = use_bf16
         self.query_instruction_for_retrieval = query_instruction_for_retrieval
         self.query_instruction_format = query_instruction_format
+        # `device` (singular) is a common alias for `devices`. Without this it is silently
+        # swallowed by **kwargs, leaving `devices=None` so the model spans ALL visible GPUs
+        # (out-of-memory). Honor it as `devices` when `devices` is not explicitly set. See #907.
+        if devices is None and "device" in kwargs:
+            devices = kwargs.pop("device")
+            logger.warning(
+                "`device` is not a documented argument for %s; interpreting it as "
+                "`devices=%r`. Pass `devices=` to silence this warning.",
+                type(self).__name__, devices,
+            )
         self.target_devices = self.get_target_devices(devices)
 
         self.batch_size = batch_size

FlagEmbedding/abc/inference/AbsReranker.py

@@ -66,6 +66,16 @@ def __init__(
         self.query_instruction_format = query_instruction_format
         self.passage_instruction_for_rerank = passage_instruction_for_rerank
         self.passage_instruction_format = passage_instruction_format
+        # `device` (singular) is a common alias for `devices`. Without this it is silently
+        # swallowed by **kwargs, leaving `devices=None` so the model spans ALL visible GPUs
+        # (out-of-memory). Honor it as `devices` when `devices` is not explicitly set. See #907.
+        if devices is None and "device" in kwargs:
+            devices = kwargs.pop("device")
+            logger.warning(
+                "`device` is not a documented argument for %s; interpreting it as "
+                "`devices=%r`. Pass `devices=` to silence this warning.",
+                type(self).__name__, devices,
+            )
         self.target_devices = self.get_target_devices(devices)
 
         self.batch_size = batch_size

tests/test_device_alias.py

@@ -0,0 +1,43 @@
+"""Regression test for #907.
+
+`device=` (singular) must be honored as an alias for `devices=` instead of being
+silently swallowed by **kwargs — which left `devices=None`, so the model spanned ALL
+visible GPUs and OOM'd. Model-free: only exercises argument handling in the abstract
+base classes (no weights downloaded, no GPU required).
+"""
+import pytest
+
+from FlagEmbedding.abc.inference.AbsEmbedder import AbsEmbedder
+from FlagEmbedding.abc.inference.AbsReranker import AbsReranker
+
+
+class _DummyEmbedder(AbsEmbedder):
+    def encode_single_device(self, *args, **kwargs):  # abstract stub
+        pass
+
+
+class _DummyReranker(AbsReranker):
+    def compute_score_single_gpu(self, *args, **kwargs):  # abstract stub
+        pass
+
+
+@pytest.mark.parametrize("cls", [_DummyEmbedder, _DummyReranker])
+def test_singular_device_aliases_to_devices(cls):
+    # `device=` should be honored, not dropped into **kwargs.
+    obj = cls("dummy-model-name", device="cuda:5")
+    assert obj.target_devices == ["cuda:5"]
+    # ...and must NOT leak as a stray attribute or into self.kwargs.
+    assert not hasattr(obj, "device")
+    assert "device" not in obj.kwargs
+
+
+@pytest.mark.parametrize("cls", [_DummyEmbedder, _DummyReranker])
+def test_plural_devices_still_works(cls):
+    obj = cls("dummy-model-name", devices="cuda:3")
+    assert obj.target_devices == ["cuda:3"]
+
+
+@pytest.mark.parametrize("cls", [_DummyEmbedder, _DummyReranker])
+def test_explicit_devices_wins_over_device(cls):
+    obj = cls("dummy-model-name", devices="cuda:2", device="cuda:9")
+    assert obj.target_devices == ["cuda:2"]