diff --git a/FlagEmbedding/abc/finetune/reranker/AbsDataset.py b/FlagEmbedding/abc/finetune/reranker/AbsDataset.py
index 0150c7e98..812ce6eda 100644
--- a/FlagEmbedding/abc/finetune/reranker/AbsDataset.py
+++ b/FlagEmbedding/abc/finetune/reranker/AbsDataset.py
@@ -16,7 +16,7 @@
 from typing import List
 
 from .AbsArguments import AbsRerankerDataArguments
-from FlagEmbedding.utils.tokenizer_compat import prepare_for_model_compat
+from FlagEmbedding.utils.tokenizer_compat import pad_with_compat, prepare_for_model_compat
 
 logger = logging.getLogger(__name__)
 
@@ -197,7 +197,8 @@ def __call__(self, features) -> List[BatchEncoding]:
         if isinstance(features[0], list):
             features = sum(features, [])
 
-        collated = self.tokenizer.pad(
+        collated = pad_with_compat(
+            self.tokenizer,
             features,
             padding=self.padding,
             max_length=self.query_max_len + self.passage_max_len,
@@ -391,7 +392,8 @@ def __call__(self, features, return_tensors='pt'):
                 else:
                     feature["labels"] = np.concatenate([remainder, feature["labels"]]).astype(np.int64)
 
-        collated = self.tokenizer.pad(
+        collated = pad_with_compat(
+            self.tokenizer,
             features,
             padding=self.padding,
             max_length=self.query_max_len + self.passage_max_len,
diff --git a/FlagEmbedding/inference/embedder/decoder_only/base.py b/FlagEmbedding/inference/embedder/decoder_only/base.py
index 7117c44d6..84375a912 100644
--- a/FlagEmbedding/inference/embedder/decoder_only/base.py
+++ b/FlagEmbedding/inference/embedder/decoder_only/base.py
@@ -6,6 +6,7 @@
 from transformers import AutoModel, AutoTokenizer
 
 from FlagEmbedding.abc.inference import AbsEmbedder
+from FlagEmbedding.utils.tokenizer_compat import pad_with_compat
 
 
 # Pooling function for LLM-based embedding models
@@ -251,7 +252,8 @@ def encode_single_device(
         flag = False
         while flag is False:
             try:
-                inputs_batch = self.tokenizer.pad(
+                inputs_batch = pad_with_compat(
+                    self.tokenizer,
                     all_inputs_sorted[: batch_size],
                     padding=True,
                     return_tensors='pt',
@@ -270,7 +272,8 @@ def encode_single_device(
         for start_index in tqdm(range(0, len(sentences), batch_size), desc="Inference Embeddings",
                                 disable=len(sentences) < batch_size):
             inputs_batch = all_inputs_sorted[start_index:start_index + batch_size]
-            inputs_batch = self.tokenizer.pad(
+            inputs_batch = pad_with_compat(
+                self.tokenizer,
                 inputs_batch,
                 padding=True,
                 return_tensors='pt',
diff --git a/FlagEmbedding/inference/embedder/decoder_only/icl.py b/FlagEmbedding/inference/embedder/decoder_only/icl.py
index bc962e89b..d4d6e087a 100644
--- a/FlagEmbedding/inference/embedder/decoder_only/icl.py
+++ b/FlagEmbedding/inference/embedder/decoder_only/icl.py
@@ -10,6 +10,7 @@
 from transformers import AutoModel, AutoTokenizer
 
 from FlagEmbedding.abc.inference import AbsEmbedder
+from FlagEmbedding.utils.tokenizer_compat import pad_with_compat
 
 
 # Pooling function for LLM-based embedding models
@@ -406,7 +407,8 @@ def encode_queries_single_device(
         flag = False
         while flag is False:
             try:
-                inputs_batch = self.tokenizer.pad(
+                inputs_batch = pad_with_compat(
+                    self.tokenizer,
                     all_inputs_sorted[: batch_size],
                     padding=True,
                     return_tensors='pt',
@@ -425,7 +427,8 @@ def encode_queries_single_device(
         for start_index in tqdm(range(0, len(sentences_sorted), batch_size), desc="Inference Embeddings",
                                 disable=len(sentences_sorted) < batch_size):
             inputs_batch = all_inputs_sorted[start_index:start_index + batch_size]
-            inputs_batch = self.tokenizer.pad(
+            inputs_batch = pad_with_compat(
+                self.tokenizer,
                 inputs_batch,
                 padding=True,
                 return_tensors='pt',
@@ -517,7 +520,8 @@ def encode_single_device(
         flag = False
         while flag is False:
             try:
-                inputs_batch = self.tokenizer.pad(
+                inputs_batch = pad_with_compat(
+                    self.tokenizer,
                     all_inputs_sorted[: batch_size],
                     padding=True,
                     return_tensors='pt',
@@ -536,7 +540,8 @@ def encode_single_device(
         for start_index in tqdm(range(0, len(sentences), batch_size), desc="Inference Embeddings",
                                 disable=len(sentences) < batch_size):
             inputs_batch = all_inputs_sorted[start_index:start_index + batch_size]
-            inputs_batch = self.tokenizer.pad(
+            inputs_batch = pad_with_compat(
+                self.tokenizer,
                 inputs_batch,
                 padding=True,
                 return_tensors='pt',
diff --git a/FlagEmbedding/inference/embedder/decoder_only/pseudo_moe.py b/FlagEmbedding/inference/embedder/decoder_only/pseudo_moe.py
index 5fe0b8e90..83bf070c2 100644
--- a/FlagEmbedding/inference/embedder/decoder_only/pseudo_moe.py
+++ b/FlagEmbedding/inference/embedder/decoder_only/pseudo_moe.py
@@ -4,6 +4,7 @@
 import numpy as np
 
 from .base import BaseLLMEmbedder, last_token_pool
+from FlagEmbedding.utils.tokenizer_compat import pad_with_compat
 
 
 class PseudoMoELLMEmbedder(BaseLLMEmbedder):
@@ -139,7 +140,8 @@ def encode_single_device(
         flag = False
         while flag is False:
             try:
-                inputs_batch = self.tokenizer.pad(
+                inputs_batch = pad_with_compat(
+                    self.tokenizer,
                     all_inputs_sorted[: batch_size],
                     padding=True,
                     return_tensors='pt',
@@ -160,7 +162,8 @@ def encode_single_device(
         all_embeddings = []
         for start_index in range(0, len(sentences), batch_size):
             inputs_batch = all_inputs_sorted[start_index:start_index + batch_size]
-            inputs_batch = self.tokenizer.pad(
+            inputs_batch = pad_with_compat(
+                self.tokenizer,
                 inputs_batch,
                 padding=True,
                 return_tensors='pt',
diff --git a/FlagEmbedding/inference/embedder/encoder_only/base.py b/FlagEmbedding/inference/embedder/encoder_only/base.py
index d01eb885a..e3f8ae11b 100644
--- a/FlagEmbedding/inference/embedder/encoder_only/base.py
+++ b/FlagEmbedding/inference/embedder/encoder_only/base.py
@@ -6,6 +6,7 @@
 from transformers import AutoModel, AutoTokenizer
 
 from FlagEmbedding.abc.inference import AbsEmbedder
+from FlagEmbedding.utils.tokenizer_compat import pad_with_compat
 
 
 class BaseEmbedder(AbsEmbedder):
@@ -232,7 +233,8 @@ def encode_single_device(
         flag = False
         while flag is False:
             try:
-                inputs_batch = self.tokenizer.pad(
+                inputs_batch = pad_with_compat(
+                    self.tokenizer,
                     all_inputs_sorted[: batch_size],
                     padding=True,
                     return_tensors='pt',
@@ -251,7 +253,8 @@ def encode_single_device(
         for start_index in tqdm(range(0, len(sentences), batch_size), desc="Inference Embeddings",
                                 disable=len(sentences) < batch_size):
             inputs_batch = all_inputs_sorted[start_index:start_index + batch_size]
-            inputs_batch = self.tokenizer.pad(
+            inputs_batch = pad_with_compat(
+                self.tokenizer,
                 inputs_batch,
                 padding=True,
                 return_tensors='pt',
diff --git a/FlagEmbedding/inference/embedder/encoder_only/m3.py b/FlagEmbedding/inference/embedder/encoder_only/m3.py
index b98efcab3..dc533f554 100644
--- a/FlagEmbedding/inference/embedder/encoder_only/m3.py
+++ b/FlagEmbedding/inference/embedder/encoder_only/m3.py
@@ -10,6 +10,7 @@
 from typing import Any, List, Union, Dict, Literal, Tuple, Optional
 
 from FlagEmbedding.abc.inference import AbsEmbedder
+from FlagEmbedding.utils.tokenizer_compat import pad_with_compat
 from FlagEmbedding.finetune.embedder.encoder_only.m3 import (
     EncoderOnlyEmbedderM3ModelForInference, EncoderOnlyEmbedderM3Runner
 )
@@ -396,7 +397,8 @@ def _process_colbert_vecs(colbert_vecs: np.ndarray, attention_mask: list):
         flag = False
         while flag is False:
             try:
-                inputs_batch = self.tokenizer.pad(
+                inputs_batch = pad_with_compat(
+                    self.tokenizer,
                     all_inputs_sorted[: batch_size],
                     padding=True,
                     return_tensors='pt',
@@ -419,7 +421,8 @@ def _process_colbert_vecs(colbert_vecs: np.ndarray, attention_mask: list):
         for start_index in tqdm(range(0, len(sentences), batch_size), desc="Inference Embeddings",
                                 disable=len(sentences) < batch_size):
             inputs_batch = all_inputs_sorted[start_index:start_index + batch_size]
-            inputs_batch = self.tokenizer.pad(
+            inputs_batch = pad_with_compat(
+                self.tokenizer,
                 inputs_batch,
                 padding=True,
                 return_tensors='pt',
diff --git a/FlagEmbedding/inference/reranker/decoder_only/base.py b/FlagEmbedding/inference/reranker/decoder_only/base.py
index b4aa8f50d..574b48bf3 100644
--- a/FlagEmbedding/inference/reranker/decoder_only/base.py
+++ b/FlagEmbedding/inference/reranker/decoder_only/base.py
@@ -10,7 +10,7 @@
 
 from FlagEmbedding.abc.inference import AbsReranker
 from FlagEmbedding.inference.reranker.encoder_only.base import sigmoid
-from FlagEmbedding.utils.tokenizer_compat import prepare_for_model_compat
+from FlagEmbedding.utils.tokenizer_compat import pad_with_compat, prepare_for_model_compat
 
 
 def last_logit_pool(logits: Tensor,
@@ -163,7 +163,8 @@ def __call__(self, data):
                 else:
                     feature["labels"] = np.concatenate([remainder, feature["labels"]]).astype(np.int64)
 
-        return self.tokenizer.pad(
+        return pad_with_compat(
+            self.tokenizer,
             data,
             padding=True,
             pad_to_multiple_of=8,
diff --git a/FlagEmbedding/inference/reranker/decoder_only/lightweight.py b/FlagEmbedding/inference/reranker/decoder_only/lightweight.py
index 44422c523..6bbac3269 100644
--- a/FlagEmbedding/inference/reranker/decoder_only/lightweight.py
+++ b/FlagEmbedding/inference/reranker/decoder_only/lightweight.py
@@ -10,7 +10,7 @@
 
 from FlagEmbedding.abc.inference import AbsReranker
 from FlagEmbedding.inference.reranker.encoder_only.base import sigmoid
-from FlagEmbedding.utils.tokenizer_compat import prepare_for_model_compat
+from FlagEmbedding.utils.tokenizer_compat import pad_with_compat, prepare_for_model_compat
 
 
 def last_logit_pool_lightweight(logits: Tensor,
@@ -78,7 +78,8 @@ def __call__(self, data):
                 else:
                     feature["labels"] = np.concatenate([remainder, feature["labels"]]).astype(np.int64)
 
-        collected = self.tokenizer.pad(
+        collected = pad_with_compat(
+            self.tokenizer,
             features,
             padding=True,
             pad_to_multiple_of=8,
diff --git a/FlagEmbedding/inference/reranker/encoder_only/base.py b/FlagEmbedding/inference/reranker/encoder_only/base.py
index 15686dcca..69b9c3505 100644
--- a/FlagEmbedding/inference/reranker/encoder_only/base.py
+++ b/FlagEmbedding/inference/reranker/encoder_only/base.py
@@ -5,7 +5,7 @@
 from transformers import AutoModelForSequenceClassification, AutoTokenizer
 
 from FlagEmbedding.abc.inference import AbsReranker
-from FlagEmbedding.utils.tokenizer_compat import prepare_for_model_compat
+from FlagEmbedding.utils.tokenizer_compat import pad_with_compat, prepare_for_model_compat
 
 
 def sigmoid(x):
@@ -162,7 +162,8 @@ def compute_score_single_gpu(
         flag = False
         while flag is False:
             try:
-                test_inputs_batch = self.tokenizer.pad(
+                test_inputs_batch = pad_with_compat(
+                    self.tokenizer,
                     all_inputs_sorted[:min(len(all_inputs_sorted), batch_size)],
                     padding=True,
                     return_tensors='pt',
@@ -179,7 +180,8 @@ def compute_score_single_gpu(
         for start_index in tqdm(range(0, len(all_inputs_sorted), batch_size), desc="Compute Scores",
                                 disable=len(all_inputs_sorted) < batch_size):
             sentences_batch = all_inputs_sorted[start_index:start_index + batch_size]
-            inputs = self.tokenizer.pad(
+            inputs = pad_with_compat(
+                self.tokenizer,
                 sentences_batch,
                 padding=True,
                 return_tensors='pt',
diff --git a/FlagEmbedding/utils/tokenizer_compat.py b/FlagEmbedding/utils/tokenizer_compat.py
index 955c43083..cdd6b61fe 100644
--- a/FlagEmbedding/utils/tokenizer_compat.py
+++ b/FlagEmbedding/utils/tokenizer_compat.py
@@ -1,8 +1,34 @@
 """Tokenizer compatibility helpers for supported Transformers versions."""
 
+from collections.abc import Mapping
 from typing import Any, List, Optional, Sequence
 
 
+def pad_with_compat(tokenizer: Any, encoded_inputs: Any, **kwargs: Any) -> Any:
+    """Pad tokenized examples across Transformers v4 and v5.
+
+    Several FlagEmbedding inference paths sort individually tokenized
+    examples before padding, producing a list of mappings.  Transformers
+    releases do not all handle that representation consistently, so normalize
+    it to the equivalent mapping-of-lists representation before calling
+    ``tokenizer.pad``.
+
+    Inputs that are already mappings, empty inputs, and other supported
+    tokenizer inputs are passed through unchanged.
+    """
+    if (
+        isinstance(encoded_inputs, (list, tuple))
+        and encoded_inputs
+        and isinstance(encoded_inputs[0], Mapping)
+    ):
+        encoded_inputs = {
+            key: [example[key] for example in encoded_inputs]
+            for key in encoded_inputs[0].keys()
+        }
+
+    return tokenizer.pad(encoded_inputs, **kwargs)
+
+
 def _decode_token_ids(tokenizer: Any, token_ids: Sequence[int]) -> str:
     """Decode token ids without dropping unknown or other special tokens."""
     return tokenizer.decode(
diff --git a/examples/finetune/ds_stage0.json b/examples/finetune/ds_stage0.json
index 495d5321c..9d03271ee 100644
--- a/examples/finetune/ds_stage0.json
+++ b/examples/finetune/ds_stage0.json
@@ -22,7 +22,8 @@
             "lr": "auto",
             "betas": "auto",
             "eps": "auto",
-            "weight_decay": "auto"
+            "weight_decay": "auto",
+            "torch_adam": true
         }
     },
 
diff --git a/examples/finetune/ds_stage1.json b/examples/finetune/ds_stage1.json
index 580d29772..4bb81a3e0 100644
--- a/examples/finetune/ds_stage1.json
+++ b/examples/finetune/ds_stage1.json
@@ -13,12 +13,7 @@
         "min_loss_scale": 1
     },
     "bf16": {
-        "enabled": "auto",
-        "loss_scale": 0,
-        "initial_scale_power": 10,
-        "loss_scale_window": 1000,
-        "hysteresis": 2,
-        "min_loss_scale": 1
+        "enabled": "auto"
     },
     "optimizer": {
         "type": "AdamW",
diff --git a/examples/finetune/embedder/README.md b/examples/finetune/embedder/README.md
index b1197ee5e..f3b508928 100644
--- a/examples/finetune/embedder/README.md
+++ b/examples/finetune/embedder/README.md
@@ -197,8 +197,9 @@ torchrun --nproc_per_node 2 \
     --temperature 0.02 \
     --sentence_pooling_method cls \
     --normalize_embeddings True \
-    --kd_loss_type kl_div
-# For transformers<=4.57.3, add --warmup_ratio 0.1; for transformers>=5.0.0, add --warmup_steps 0.1.
+    --kd_loss_type kl_div \
+    --warmup_steps 0.1
+# For transformers <= 4.57.3, replace --warmup_steps 0.1 with --warmup_ratio 0.1.
 ```
 
 ### (2) bge-m3
@@ -239,8 +240,9 @@ torchrun --nproc_per_node 2 \
     --unified_finetuning True \
     --use_self_distill True \
     --fix_encoder False \
-    --self_distill_start_step 0
-# For transformers<=4.57.3, add --warmup_ratio 0.1; for transformers>=5.0.0, add --warmup_steps 0.1.
+    --self_distill_start_step 0 \
+    --warmup_steps 0.1
+# For transformers <= 4.57.3, replace --warmup_steps 0.1 with --warmup_ratio 0.1.
 ```
 
 Here are some new arguments:
@@ -293,8 +295,9 @@ torchrun --nproc_per_node 2 \
     --temperature 0.02 \
     --sentence_pooling_method last_token \
     --normalize_embeddings True \
-    --kd_loss_type m3_kd_loss
-# For transformers<=4.57.3, add --warmup_ratio 0.1; for transformers>=5.0.0, add --warmup_steps 0.1.
+    --kd_loss_type m3_kd_loss \
+    --warmup_steps 0.1
+# For transformers <= 4.57.3, replace --warmup_steps 0.1 with --warmup_ratio 0.1.
 ```
 
 Here are some new arguments:
@@ -356,8 +359,9 @@ torchrun --nproc_per_node 2 \
     --temperature 0.02 \
     --sentence_pooling_method last_token \
     --normalize_embeddings True \
-    --kd_loss_type kl_div
-# For transformers<=4.57.3, add --warmup_ratio 0.1; for transformers>=5.0.0, add --warmup_steps 0.1.
+    --kd_loss_type kl_div \
+    --warmup_steps 0.1
+# For transformers <= 4.57.3, replace --warmup_steps 0.1 with --warmup_ratio 0.1.
 ```
 
 Here are some new arguments:
diff --git a/examples/finetune/embedder/decoder_only/base.sh b/examples/finetune/embedder/decoder_only/base.sh
index 8268a2228..73b6bb44f 100644
--- a/examples/finetune/embedder/decoder_only/base.sh
+++ b/examples/finetune/embedder/decoder_only/base.sh
@@ -56,9 +56,10 @@ training_args="\
     --sentence_pooling_method last_token \
     --normalize_embeddings True \
     --kd_loss_type m3_kd_loss \
+    --warmup_steps 0.1 \
 "
 
-# For transformers<=4.57.3, add --warmup_ratio 0.1; for transformers>=5.0.0, add --warmup_steps 0.1.
+# For transformers <= 4.57.3, replace --warmup_steps 0.1 with --warmup_ratio 0.1.
 
 cmd="torchrun --nproc_per_node $num_gpus \
     -m FlagEmbedding.finetune.embedder.decoder_only.base \
diff --git a/examples/finetune/embedder/decoder_only/base_same_dataset.sh b/examples/finetune/embedder/decoder_only/base_same_dataset.sh
index 197e52888..5ad9a5396 100644
--- a/examples/finetune/embedder/decoder_only/base_same_dataset.sh
+++ b/examples/finetune/embedder/decoder_only/base_same_dataset.sh
@@ -59,9 +59,10 @@ training_args="\
     --sentence_pooling_method last_token \
     --normalize_embeddings True \
     --kd_loss_type m3_kd_loss \
+    --warmup_steps 0.1 \
 "
 
-# For transformers<=4.57.3, add --warmup_ratio 0.1; for transformers>=5.0.0, add --warmup_steps 0.1.
+# For transformers <= 4.57.3, replace --warmup_steps 0.1 with --warmup_ratio 0.1.
 
 cmd="torchrun --nproc_per_node $num_gpus \
     -m FlagEmbedding.finetune.embedder.decoder_only.base \
diff --git a/examples/finetune/embedder/decoder_only/icl_same_dataset.sh b/examples/finetune/embedder/decoder_only/icl_same_dataset.sh
index 2e080f04c..9b7843a0e 100644
--- a/examples/finetune/embedder/decoder_only/icl_same_dataset.sh
+++ b/examples/finetune/embedder/decoder_only/icl_same_dataset.sh
@@ -63,9 +63,10 @@ training_args="\
     --sentence_pooling_method last_token \
     --normalize_embeddings True \
     --kd_loss_type kl_div \
+    --warmup_steps 0.1 \
 "
 
-# For transformers<=4.57.3, add --warmup_ratio 0.1; for transformers>=5.0.0, add --warmup_steps 0.1.
+# For transformers <= 4.57.3, replace --warmup_steps 0.1 with --warmup_ratio 0.1.
 
 cmd="torchrun --nproc_per_node $num_gpus \
     -m FlagEmbedding.finetune.embedder.decoder_only.icl \
diff --git a/examples/finetune/embedder/encoder_only/base.sh b/examples/finetune/embedder/encoder_only/base.sh
index c345fa8df..b8fadf6ea 100644
--- a/examples/finetune/embedder/encoder_only/base.sh
+++ b/examples/finetune/embedder/encoder_only/base.sh
@@ -50,9 +50,10 @@ training_args="\
     --sentence_pooling_method cls \
     --normalize_embeddings True \
     --kd_loss_type kl_div \
+    --warmup_steps 0.1 \
 "
 
-# For transformers<=4.57.3, add --warmup_ratio 0.1; for transformers>=5.0.0, add --warmup_steps 0.1.
+# For transformers <= 4.57.3, replace --warmup_steps 0.1 with --warmup_ratio 0.1.
 
 cmd="torchrun --nproc_per_node $num_gpus \
     -m FlagEmbedding.finetune.embedder.encoder_only.base \
diff --git a/examples/finetune/embedder/encoder_only/base_same_dataset.sh b/examples/finetune/embedder/encoder_only/base_same_dataset.sh
index 944e463e5..3aa4e3cba 100644
--- a/examples/finetune/embedder/encoder_only/base_same_dataset.sh
+++ b/examples/finetune/embedder/encoder_only/base_same_dataset.sh
@@ -53,9 +53,10 @@ training_args="\
     --sentence_pooling_method cls \
     --normalize_embeddings True \
     --kd_loss_type kl_div \
+    --warmup_steps 0.1 \
 "
 
-# For transformers<=4.57.3, add --warmup_ratio 0.1; for transformers>=5.0.0, add --warmup_steps 0.1.
+# For transformers <= 4.57.3, replace --warmup_steps 0.1 with --warmup_ratio 0.1.
 
 cmd="torchrun --nproc_per_node $num_gpus \
     -m FlagEmbedding.finetune.embedder.encoder_only.base \
diff --git a/examples/finetune/embedder/encoder_only/m3.sh b/examples/finetune/embedder/encoder_only/m3.sh
index 201cebb89..0222e1af8 100644
--- a/examples/finetune/embedder/encoder_only/m3.sh
+++ b/examples/finetune/embedder/encoder_only/m3.sh
@@ -52,9 +52,10 @@ training_args="\
     --use_self_distill True \
     --fix_encoder False \
     --self_distill_start_step 0 \
+    --warmup_steps 0.1 \
 "
 
-# For transformers<=4.57.3, add --warmup_ratio 0.1; for transformers>=5.0.0, add --warmup_steps 0.1.
+# For transformers <= 4.57.3, replace --warmup_steps 0.1 with --warmup_ratio 0.1.
 
 cmd="torchrun --nproc_per_node $num_gpus \
     -m FlagEmbedding.finetune.embedder.encoder_only.m3 \
diff --git a/examples/finetune/embedder/encoder_only/m3_same_dataset.sh b/examples/finetune/embedder/encoder_only/m3_same_dataset.sh
index 3b7d1296a..1a3fa8bbb 100644
--- a/examples/finetune/embedder/encoder_only/m3_same_dataset.sh
+++ b/examples/finetune/embedder/encoder_only/m3_same_dataset.sh
@@ -55,9 +55,10 @@ training_args="\
     --use_self_distill True \
     --fix_encoder False \
     --self_distill_start_step 0 \
+    --warmup_steps 0.1 \
 "
 
-# For transformers<=4.57.3, add --warmup_ratio 0.1; for transformers>=5.0.0, add --warmup_steps 0.1.
+# For transformers <= 4.57.3, replace --warmup_steps 0.1 with --warmup_ratio 0.1.
 
 cmd="torchrun --nproc_per_node $num_gpus \
     -m FlagEmbedding.finetune.embedder.encoder_only.m3 \
diff --git a/examples/finetune/reranker/README.md b/examples/finetune/reranker/README.md
index 4fde383d5..209bdba62 100644
--- a/examples/finetune/reranker/README.md
+++ b/examples/finetune/reranker/README.md
@@ -168,8 +168,9 @@ torchrun --nproc_per_node 2 \
     --weight_decay 0.01 \
     --deepspeed ../ds_stage0.json \
     --logging_steps 1 \
-    --save_steps 1000
-# For transformers<=4.57.3, add --warmup_ratio 0.1; for transformers>=5.0.0, add --warmup_steps 0.1.
+    --save_steps 1000 \
+    --warmup_steps 0.1
+# For transformers <= 4.57.3, replace --warmup_steps 0.1 with --warmup_ratio 0.1.
 ```
 
 ### (2) bge-reranker-v2-gemma
@@ -208,8 +209,9 @@ torchrun --nproc_per_node 2 \
     --weight_decay 0.01 \
     --deepspeed ../ds_stage0.json \
     --logging_steps 1 \
-    --save_steps 1000
-# For transformers<=4.57.3, add --warmup_ratio 0.1; for transformers>=5.0.0, add --warmup_steps 0.1.
+    --save_steps 1000 \
+    --warmup_steps 0.1
+# For transformers <= 4.57.3, replace --warmup_steps 0.1 with --warmup_ratio 0.1.
 ```
 
 Here are some new arguments:
@@ -266,8 +268,9 @@ torchrun --nproc_per_node 2 \
     --weight_decay 0.01 \
     --deepspeed ../ds_stage0.json \
     --logging_steps 1 \
-    --save_steps 1000
-# For transformers<=4.57.3, add --warmup_ratio 0.1; for transformers>=5.0.0, add --warmup_steps 0.1.
+    --save_steps 1000 \
+    --warmup_steps 0.1
+# For transformers <= 4.57.3, replace --warmup_steps 0.1 with --warmup_ratio 0.1.
 ```
 
 Here are some new arguments:
diff --git a/examples/finetune/reranker/decoder_only/base.sh b/examples/finetune/reranker/decoder_only/base.sh
index 2230edfa3..59b363b55 100644
--- a/examples/finetune/reranker/decoder_only/base.sh
+++ b/examples/finetune/reranker/decoder_only/base.sh
@@ -55,9 +55,10 @@ training_args="\
     --deepspeed ../../ds_stage0.json \
     --logging_steps 1 \
     --save_steps 1000 \
+    --warmup_steps 0.1 \
 "
 
-# For transformers<=4.57.3, add --warmup_ratio 0.1; for transformers>=5.0.0, add --warmup_steps 0.1.
+# For transformers <= 4.57.3, replace --warmup_steps 0.1 with --warmup_ratio 0.1.
 
 cmd="torchrun --nproc_per_node $num_gpus \
     -m FlagEmbedding.finetune.reranker.decoder_only.base \
diff --git a/examples/finetune/reranker/decoder_only/layerwise.sh b/examples/finetune/reranker/decoder_only/layerwise.sh
index 8d2b6848f..4762fb0d7 100644
--- a/examples/finetune/reranker/decoder_only/layerwise.sh
+++ b/examples/finetune/reranker/decoder_only/layerwise.sh
@@ -60,9 +60,10 @@ training_args="\
     --deepspeed ../../ds_stage0.json \
     --logging_steps 1 \
     --save_steps 1000 \
+    --warmup_steps 0.1 \
 "
 
-# For transformers<=4.57.3, add --warmup_ratio 0.1; for transformers>=5.0.0, add --warmup_steps 0.1.
+# For transformers <= 4.57.3, replace --warmup_steps 0.1 with --warmup_ratio 0.1.
 
 cmd="torchrun --nproc_per_node $num_gpus \
     -m FlagEmbedding.finetune.reranker.decoder_only.layerwise \
diff --git a/examples/finetune/reranker/encoder_only/base.sh b/examples/finetune/reranker/encoder_only/base.sh
index d0cbea6ee..4229967ce 100644
--- a/examples/finetune/reranker/encoder_only/base.sh
+++ b/examples/finetune/reranker/encoder_only/base.sh
@@ -44,9 +44,10 @@ training_args="\
     --deepspeed ../../ds_stage0.json \
     --logging_steps 1 \
     --save_steps 1000 \
+    --warmup_steps 0.1 \
 "
 
-# For transformers<=4.57.3, add --warmup_ratio 0.1; for transformers>=5.0.0, add --warmup_steps 0.1.
+# For transformers <= 4.57.3, replace --warmup_steps 0.1 with --warmup_ratio 0.1.
 
 cmd="torchrun --nproc_per_node $num_gpus \
     -m FlagEmbedding.finetune.reranker.encoder_only.base \
diff --git a/tests/README.md b/tests/README.md
index fcf55e37f..9ab0f3638 100644
--- a/tests/README.md
+++ b/tests/README.md
@@ -7,6 +7,7 @@ This directory contains tests for the FlagEmbedding library, including compatibi
 - `test_imports_v5.py`: Tests that imports work with Transformers v5, particularly the compatibility layer for `is_torch_fx_available`.
 - `test_finetune_trainer_compat.py`: Tests the Transformers Trainer API migration, including `processing_class`/legacy `tokenizer` construction, reranker runner arguments, and processor checkpoint saving.
 - `test_reranker_tokenizer_compat.py`: Tests reranker pair preparation across Transformers v4/v5, including the v5 removal of `prepare_for_model`.
+- `test_tokenizer_pad_compat.py`: Tests normalization of sorted per-example tokenized records before `tokenizer.pad`, including the local BGE reranker tokenizer when available.
 - `test_infer_embedder_basic.py`: Tests basic functionality of BGE embedder models with a small public checkpoint.
 - `test_infer_reranker_basic.py`: Tests basic functionality of reranker models.
 
@@ -32,6 +33,9 @@ pytest tests/test_finetune_trainer_compat.py
 # Run the reranker tokenizer compatibility tests
 pytest tests/test_reranker_tokenizer_compat.py
 
+# Run the tokenizer padding compatibility tests
+pytest tests/test_tokenizer_pad_compat.py
+
 # Run with verbose output
 pytest -v tests/
 ```
diff --git a/tests/test_reranker_tokenizer_compat.py b/tests/test_reranker_tokenizer_compat.py
index 41f96819c..75decf1cf 100644
--- a/tests/test_reranker_tokenizer_compat.py
+++ b/tests/test_reranker_tokenizer_compat.py
@@ -1,16 +1,15 @@
 import os
 from types import SimpleNamespace
 
-import pytest
 from transformers import AutoTokenizer
 
 from FlagEmbedding.abc.finetune.reranker.AbsDataset import AbsRerankerTrainDataset
 from FlagEmbedding.utils.tokenizer_compat import prepare_for_model_compat
 
 
-RERANKER_MODEL_PATH = os.environ.get(
+RERANKER_MODEL_NAME = os.environ.get(
     "FLAGEMBEDDING_RERANKER_MODEL",
-    "/share/project/Search-SWE/model/bge-reranker-base",
+    "BAAI/bge-reranker-base",
 )
 
 
@@ -26,14 +25,9 @@ def prepare_for_model(self, first_ids, second_ids, **kwargs):
         return {"input_ids": list(first_ids) + list(second_ids)}
 
 
-@pytest.mark.skipif(
-    not os.path.isdir(RERANKER_MODEL_PATH),
-    reason="local reranker model is not available",
-)
 def test_encoder_pair_preserves_v5_tokenizer_input_ids():
     tokenizer = AutoTokenizer.from_pretrained(
-        RERANKER_MODEL_PATH,
-        local_files_only=True,
+        RERANKER_MODEL_NAME,
     )
 
     query = "😀🧪"
@@ -62,14 +56,9 @@ def test_encoder_pair_preserves_v5_tokenizer_input_ids():
     assert actual["attention_mask"] == expected["attention_mask"]
 
 
-@pytest.mark.skipif(
-    not os.path.isdir(RERANKER_MODEL_PATH),
-    reason="local reranker model is not available",
-)
 def test_encoder_training_example_works_without_prepare_for_model():
     tokenizer = AutoTokenizer.from_pretrained(
-        RERANKER_MODEL_PATH,
-        local_files_only=True,
+        RERANKER_MODEL_NAME,
     )
     dataset = AbsRerankerTrainDataset.__new__(AbsRerankerTrainDataset)
     dataset.tokenizer = tokenizer
diff --git a/tests/test_tokenizer_pad_compat.py b/tests/test_tokenizer_pad_compat.py
new file mode 100644
index 000000000..aaf948a56
--- /dev/null
+++ b/tests/test_tokenizer_pad_compat.py
@@ -0,0 +1,115 @@
+import os
+
+import torch
+from transformers import AutoTokenizer
+
+from FlagEmbedding.utils.tokenizer_compat import pad_with_compat
+
+
+RERANKER_MODEL_NAME = os.environ.get(
+    "FLAGEMBEDDING_RERANKER_MODEL",
+    "BAAI/bge-reranker-base",
+)
+ENCODER_MODEL_NAME = os.environ.get(
+    "FLAGEMBEDDING_ENCODER_MODEL",
+    "BAAI/bge-base-en-v1.5",
+)
+
+
+class MappingOnlyTokenizer:
+    """Minimal tokenizer double that rejects list-of-mapping inputs."""
+
+    def __init__(self):
+        self.received = None
+        self.kwargs = None
+
+    def pad(self, encoded_inputs, **kwargs):
+        if not isinstance(encoded_inputs, dict):
+            raise AttributeError("'list' object has no attribute 'keys'")
+        self.received = encoded_inputs
+        self.kwargs = kwargs
+        return encoded_inputs
+
+
+def test_pad_with_compat_normalizes_list_of_mappings():
+    tokenizer = MappingOnlyTokenizer()
+    records = [
+        {"input_ids": [101, 10, 102], "attention_mask": [1, 1, 1]},
+        {"input_ids": [101, 20, 30, 102], "attention_mask": [1, 1, 1, 1]},
+    ]
+
+    result = pad_with_compat(tokenizer, records, padding=True, return_tensors="pt")
+
+    assert result == {
+        "input_ids": [[101, 10, 102], [101, 20, 30, 102]],
+        "attention_mask": [[1, 1, 1], [1, 1, 1, 1]],
+    }
+    assert tokenizer.received == result
+    assert tokenizer.kwargs == {"padding": True, "return_tensors": "pt"}
+
+
+def test_pad_with_compat_passes_mapping_inputs_through():
+    tokenizer = MappingOnlyTokenizer()
+    encoded_inputs = {
+        "input_ids": [[101, 10, 102]],
+        "attention_mask": [[1, 1, 1]],
+    }
+
+    result = pad_with_compat(tokenizer, encoded_inputs, padding=True)
+
+    assert result is encoded_inputs
+    assert tokenizer.received is encoded_inputs
+
+
+def test_local_bge_reranker_tokenizer_padding_compatibility():
+    tokenizer = AutoTokenizer.from_pretrained(
+        RERANKER_MODEL_NAME,
+    )
+    encoded = tokenizer(
+        ["What is AI?", "Explain artificial intelligence in one sentence."],
+        truncation=True,
+        padding=False,
+        return_tensors=None,
+    )
+    records = [
+        {key: encoded[key][index] for key in encoded.keys()}
+        for index in range(2)
+    ]
+
+    padded = pad_with_compat(
+        tokenizer,
+        records,
+        padding=True,
+        return_tensors="pt",
+    )
+
+    assert padded["input_ids"].shape[0] == 2
+    assert padded["attention_mask"].shape == padded["input_ids"].shape
+    assert torch.all(padded["attention_mask"][:, 0] == 1)
+
+
+def test_local_bge_encoder_token_type_ids_are_preserved():
+    tokenizer = AutoTokenizer.from_pretrained(
+        ENCODER_MODEL_NAME,
+    )
+    encoded = tokenizer(
+        ["short text", "a longer text for padding"],
+        truncation=True,
+        padding=False,
+        return_tensors=None,
+    )
+    records = [
+        {key: encoded[key][index] for key in encoded.keys()}
+        for index in range(2)
+    ]
+
+    padded = pad_with_compat(
+        tokenizer,
+        records,
+        padding=True,
+        return_tensors="pt",
+    )
+
+    assert "token_type_ids" in padded
+    assert padded["token_type_ids"].shape == padded["input_ids"].shape
+    assert padded["attention_mask"].shape == padded["input_ids"].shape