From ac00be1bb47a9874b9d7ed21094544f917ac6e5d Mon Sep 17 00:00:00 2001
From: emme1t <149944796+emme1t@users.noreply.github.com>
Date: Sat, 31 Oct 2026 18:33:42 -0700
Subject: [PATCH] fix: resolve auto models from paths with trailing separators

---
 FlagEmbedding/inference/auto_embedder.py |  5 +-
 FlagEmbedding/inference/auto_reranker.py |  5 +-
 tests/test_auto_model_local_paths.py     | 84 ++++++++++++++++++++++++
 3 files changed, 90 insertions(+), 4 deletions(-)
 create mode 100644 tests/test_auto_model_local_paths.py

diff --git a/FlagEmbedding/inference/auto_embedder.py b/FlagEmbedding/inference/auto_embedder.py
index 6fad0330f..b8e1f09da 100644
--- a/FlagEmbedding/inference/auto_embedder.py
+++ b/FlagEmbedding/inference/auto_embedder.py
@@ -59,9 +59,10 @@ def from_finetuned(
         Returns:
             AbsEmbedder: The model class to load model, which is child class of :class:`AbsEmbedder`.
         """
-        model_name = os.path.basename(model_name_or_path)
+        model_path = os.path.normpath(model_name_or_path)
+        model_name = os.path.basename(model_path)
         if model_name.startswith("checkpoint-"):
-            model_name = os.path.basename(os.path.dirname(model_name_or_path))
+            model_name = os.path.basename(os.path.dirname(model_path))
 
         if model_class is not None:
             _model_class = EMBEDDER_CLASS_MAPPING[EmbedderModelClass(model_class)]
diff --git a/FlagEmbedding/inference/auto_reranker.py b/FlagEmbedding/inference/auto_reranker.py
index a8b2d699e..5157ba0c9 100644
--- a/FlagEmbedding/inference/auto_reranker.py
+++ b/FlagEmbedding/inference/auto_reranker.py
@@ -46,9 +46,10 @@ def from_finetuned(
         Returns:
             AbsReranker: The reranker class to load model, which is child class of :class:`AbsReranker`.
         """
-        model_name = os.path.basename(model_name_or_path)
+        model_path = os.path.normpath(model_name_or_path)
+        model_name = os.path.basename(model_path)
         if model_name.startswith("checkpoint-"):
-            model_name = os.path.basename(os.path.dirname(model_name_or_path))
+            model_name = os.path.basename(os.path.dirname(model_path))
 
         if model_class is not None:
             _model_class = RERANKER_CLASS_MAPPING[RerankerModelClass(model_class)]
diff --git a/tests/test_auto_model_local_paths.py b/tests/test_auto_model_local_paths.py
new file mode 100644
index 000000000..d1b6ff850
--- /dev/null
+++ b/tests/test_auto_model_local_paths.py
@@ -0,0 +1,84 @@
+import os
+
+import numpy as np
+import pytest
+from tokenizers import Tokenizer, models, pre_tokenizers, processors
+from transformers import (
+    BertConfig,
+    BertForSequenceClassification,
+    BertModel,
+    PreTrainedTokenizerFast,
+)
+
+from FlagEmbedding import FlagAutoModel, FlagAutoReranker
+
+
+def run_inference(model, kind):
+    if kind == "embedder":
+        return model.encode(["hello world", "hello"])
+    return np.asarray(model.compute_score([("hello", "world"), ("hello", "hello")]))
+
+
+@pytest.fixture(scope="module", params=["embedder", "reranker"])
+def local_model(request, tmp_path_factory):
+    """Save small, randomly initialized models without downloading checkpoints."""
+    kind = request.param
+    name = "bge-base-en-v1.5" if kind == "embedder" else "bge-reranker-base"
+    model_dir = tmp_path_factory.mktemp(kind) / name
+    checkpoint_dir = model_dir / "checkpoint-10"
+    vocabulary = {
+        token: i
+        for i, token in enumerate(
+            ["[PAD]", "[UNK]", "[CLS]", "[SEP]", "[MASK]", "hello", "world"]
+        )
+    }
+    backend = Tokenizer(models.WordPiece(vocabulary, unk_token="[UNK]"))
+    backend.pre_tokenizer = pre_tokenizers.Whitespace()
+    backend.post_processor = processors.TemplateProcessing(
+        single="[CLS] $A [SEP]",
+        pair="[CLS] $A [SEP] $B:1 [SEP]:1",
+        special_tokens=[("[CLS]", 2), ("[SEP]", 3)],
+    )
+    tokenizer = PreTrainedTokenizerFast(
+        tokenizer_object=backend,
+        unk_token="[UNK]",
+        pad_token="[PAD]",
+        cls_token="[CLS]",
+        sep_token="[SEP]",
+        model_max_length=32,
+    )
+    config = BertConfig(
+        vocab_size=len(vocabulary),
+        hidden_size=8,
+        num_hidden_layers=1,
+        num_attention_heads=2,
+        intermediate_size=16,
+        max_position_embeddings=32,
+        num_labels=1,
+    )
+    model_type = BertModel if kind == "embedder" else BertForSequenceClassification
+    model = model_type(config)
+    for path in (model_dir, checkpoint_dir):
+        model.save_pretrained(path)
+        tokenizer.save_pretrained(path)
+    factory = FlagAutoModel if kind == "embedder" else FlagAutoReranker
+    reference = factory.from_finetuned(str(model_dir), devices="cpu", use_fp16=False)
+    expected = run_inference(reference, kind)
+    assert np.isfinite(expected).all()
+    return kind, factory, model_dir, checkpoint_dir, expected
+
+
+@pytest.mark.parametrize("checkpoint", [False, True])
+@pytest.mark.parametrize("suffix", sorted({"", "/", "//", os.sep, os.sep * 2}))
+def test_auto_model_loads_local_directory_with_trailing_separator(
+    local_model, checkpoint, suffix
+):
+    kind, factory, model_dir, checkpoint_dir, expected = local_model
+    path = str(checkpoint_dir if checkpoint else model_dir) + suffix
+
+    model = factory.from_finetuned(path, devices="cpu", use_fp16=False)
+
+    assert model.model_name_or_path == path
+    np.testing.assert_allclose(
+        run_inference(model, kind), expected, rtol=1e-6, atol=1e-6
+    )