Update ruff to 0.11.2 (#36962)

* update

* update

* update

---------

Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

setup.py

@@ -162,7 +162,7 @@ _deps = [
     "rhoknp>=1.1.0,<1.3.1",
     "rjieba",
     "rouge-score!=0.0.7,!=0.0.8,!=0.1,!=0.1.1",
-    "ruff==0.5.1",
+    "ruff==0.11.2",
     "sacrebleu>=1.4.12,<2.0.0",
     "sacremoses",
     "safetensors>=0.4.3",

src/transformers/agents/tools.py

@@ -167,9 +167,9 @@ class Tool:
                 )
         for input_name, input_content in self.inputs.items():
             assert isinstance(input_content, dict), f"Input '{input_name}' should be a dictionary."
-            assert (
+            assert "type" in input_content and "description" in input_content, (
-                "type" in input_content and "description" in input_content
+                f"Input '{input_name}' should have keys 'type' and 'description', has only {list(input_content.keys())}."
-            ), f"Input '{input_name}' should have keys 'type' and 'description', has only {list(input_content.keys())}."
+            )
             if input_content["type"] not in authorized_types:
                 raise Exception(
                     f"Input '{input_name}': type '{input_content['type']}' is not an authorized value, should be one of {authorized_types}."

src/transformers/commands/add_fast_image_processor.py

@@ -313,11 +313,9 @@ def add_fast_image_processor_to_doc(fast_image_processor_name: str, model_name:
         raise ValueError(f"No doc files found for {model_name}")
 
     base_doc_string = (
-        f"## {fast_image_processor_name[:-4]}\n\n" f"[[autodoc]] {fast_image_processor_name[:-4]}\n" "    - preprocess"
+        f"## {fast_image_processor_name[:-4]}\n\n[[autodoc]] {fast_image_processor_name[:-4]}\n    - preprocess"
-    )
-    fast_doc_string = (
-        f"## {fast_image_processor_name}\n\n" f"[[autodoc]] {fast_image_processor_name}\n" "    - preprocess"
     )
+    fast_doc_string = f"## {fast_image_processor_name}\n\n[[autodoc]] {fast_image_processor_name}\n    - preprocess"
 
     for doc_file in doc_files:
         with open(doc_file, "r", encoding="utf-8") as f:
@@ -385,7 +383,7 @@ def add_fast_image_processor_to_tests(fast_image_processor_name: str, model_name
     # add the fast image processor to the imports
     base_import_string = f"    from transformers import {fast_image_processor_name[:-4]}"
     fast_import_string = (
-        "    if is_torchvision_available():\n" f"        from transformers import {fast_image_processor_name}"
+        f"    if is_torchvision_available():\n        from transformers import {fast_image_processor_name}"
     )
     if fast_import_string not in updated_content:
         updated_content = updated_content.replace(base_import_string, base_import_string + "\n\n" + fast_import_string)
@@ -546,17 +544,17 @@ def add_fast_image_processor_file(
         "    # For an example of a fast image processor requiring more complex augmentations, see `LlavaNextImageProcessorFast`.\n\n"
         "    # Default values should be checked against the slow image processor\n"
         "    # None values left after checking can be removed\n"
-        f'    resample = {default_args_dict.get("resample")}\n'
+        f"    resample = {default_args_dict.get('resample')}\n"
-        f'    image_mean = {default_args_dict.get("image_mean")}\n'
+        f"    image_mean = {default_args_dict.get('image_mean')}\n"
-        f'    image_std = {default_args_dict.get("image_std")}\n'
+        f"    image_std = {default_args_dict.get('image_std')}\n"
-        f'    size = {default_args_dict.get("size")}\n'
+        f"    size = {default_args_dict.get('size')}\n"
-        f'    default_to_square = {default_args_dict.get("default_to_square")}\n'
+        f"    default_to_square = {default_args_dict.get('default_to_square')}\n"
-        f'    crop_size = {default_args_dict.get("crop_size")}\n'
+        f"    crop_size = {default_args_dict.get('crop_size')}\n"
-        f'    do_resize = {default_args_dict.get("do_resize")}\n'
+        f"    do_resize = {default_args_dict.get('do_resize')}\n"
-        f'    do_center_crop = {default_args_dict.get("do_center_crop")}\n'
+        f"    do_center_crop = {default_args_dict.get('do_center_crop')}\n"
-        f'    do_rescale = {default_args_dict.get("do_rescale")}\n'
+        f"    do_rescale = {default_args_dict.get('do_rescale')}\n"
-        f'    do_normalize = {default_args_dict.get("do_normalize")}\n'
+        f"    do_normalize = {default_args_dict.get('do_normalize')}\n"
-        f'    do_convert_rgb = {default_args_dict.get("do_convert_rgb")}\n\n\n'
+        f"    do_convert_rgb = {default_args_dict.get('do_convert_rgb')}\n\n\n"
         f'__all__ = ["{fast_image_processor_name}"]\n'
     )
 

src/transformers/convert_graph_to_onnx.py

@@ -189,7 +189,7 @@ def infer_shapes(nlp: Pipeline, framework: str) -> tuple[list[str], list[str], d
                     raise ValueError(f"Unable to infer tensor axes ({len(tensor.shape)})")
             else:
                 seq_axes = [dim for dim, shape in enumerate(tensor.shape) if shape == seq_len]
-                axes.update({dim: "sequence" for dim in seq_axes})
+                axes.update(dict.fromkeys(seq_axes, "sequence"))
 
         print(f"Found {'input' if is_input else 'output'} {name} with shape: {axes}")
         return axes

src/transformers/data/metrics/squad_metrics.py

@@ -226,7 +226,7 @@ def squad_evaluate(examples, preds, no_answer_probs=None, no_answer_probability_
     no_answer_qids = [qas_id for qas_id, has_answer in qas_id_to_has_answer.items() if not has_answer]
 
     if no_answer_probs is None:
-        no_answer_probs = {k: 0.0 for k in preds}
+        no_answer_probs = dict.fromkeys(preds, 0.0)
 
     exact, f1 = get_raw_scores(examples, preds)
 

src/transformers/data/processors/glue.py

@@ -101,7 +101,7 @@ if is_tf_available():
 
         return tf.data.Dataset.from_generator(
             gen,
-            ({k: tf.int32 for k in input_names}, label_type),
+            (dict.fromkeys(input_names, tf.int32), label_type),
             ({k: tf.TensorShape([None]) for k in input_names}, tf.TensorShape([])),
         )
 

src/transformers/dependency_versions_table.py

@@ -68,7 +68,7 @@ deps = {
     "rhoknp": "rhoknp>=1.1.0,<1.3.1",
     "rjieba": "rjieba",
     "rouge-score": "rouge-score!=0.0.7,!=0.0.8,!=0.1,!=0.1.1",
-    "ruff": "ruff==0.5.1",
+    "ruff": "ruff==0.11.2",
     "sacrebleu": "sacrebleu>=1.4.12,<2.0.0",
     "sacremoses": "sacremoses",
     "safetensors": "safetensors>=0.4.3",

src/transformers/generation/logits_process.py

@@ -2749,9 +2749,7 @@ class SynthIDTextWatermarkLogitsProcessor(LogitsProcessor):
             ngram keys (batch_size, num_ngrams, depth).
         """
         if len(ngrams.shape) != 3:
-            raise ValueError(
+            raise ValueError(f"Ngrams should be of shape (batch_size, num_ngrams, ngram_len), but is {ngrams.shape}")
-                "Ngrams should be of shape (batch_size, num_ngrams, ngram_len), but" f" is {ngrams.shape}"
-            )
         if ngrams.shape[2] != self.ngram_len:
             raise ValueError(
                 "Ngrams should be of shape (batch_size, num_ngrams, ngram_len),"
@@ -2836,7 +2834,7 @@ class SynthIDTextWatermarkLogitsProcessor(LogitsProcessor):
     def _check_input_ids_shape(self, input_ids: torch.LongTensor):
         """Checks the shape of input ids."""
         if len(input_ids.shape) != 2:
-            raise ValueError("Input ids should be of shape (batch_size, input_len), but is" f" {input_ids.shape}")
+            raise ValueError(f"Input ids should be of shape (batch_size, input_len), but is {input_ids.shape}")
 
     def compute_g_values(self, input_ids: torch.LongTensor) -> torch.LongTensor:
         """

src/transformers/generation/utils.py

@@ -1678,7 +1678,7 @@ class GenerationMixin:
         if execution_device_map is None:
             return None
         elif len(execution_device_map) == 1 and "" in execution_device_map:
-            return {idx: execution_device_map[""] for idx in range(num_hidden_layers)}
+            return dict.fromkeys(range(num_hidden_layers), execution_device_map[""])
         layer_device_map = {}
         for layer in execution_device_map:
             for idx in range(num_hidden_layers):

src/transformers/integrations/hqq.py

@@ -106,11 +106,11 @@ def prepare_for_hqq_linear(model, quantization_config=None, modules_to_not_conve
 
     if any(key in linear_tags for key in quant_config.keys()):
         # If the user doesn't specify a key from get_linear_tags, the layer is not quantized via (key, None)
-        patch_params = {key: None for key in linear_tags}
+        patch_params = dict.fromkeys(linear_tags)
         patch_params.update(quant_config)
     else:
         # Same quant_config for all layers
-        patch_params = {k: quant_config for k in linear_tags}
+        patch_params = dict.fromkeys(linear_tags, quant_config)
 
     model, has_been_replaced = _prepare_for_hqq_linear(
         model, patch_params=patch_params, has_been_replaced=has_been_replaced

src/transformers/integrations/tpu.py

@@ -21,9 +21,9 @@ def tpu_spmd_dataloader(dataloader: DataLoader):
     if is_torch_xla_available():
         import torch_xla.distributed.parallel_loader as pl
 
-        assert isinstance(
+        assert isinstance(dataloader, pl.MpDeviceLoader), (
-            dataloader, pl.MpDeviceLoader
+            "The dataloader must be a `torch_xla.distributed.parallel_loader.MpDeviceLoader`."
-        ), "The dataloader must be a `torch_xla.distributed.parallel_loader.MpDeviceLoader`."
+        )
 
         # This is to support PyTorch/XLA FSDP via SPMD.
         # Here we shard the input data's 0th dim across the fsdp axis.

src/transformers/modeling_utils.py

@@ -2509,9 +2509,9 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
             total_decoder_name="",
             total_encoder_name="",
         ):
-            assert isinstance(decoder_pointer, nn.Module) and isinstance(
+            assert isinstance(decoder_pointer, nn.Module) and isinstance(encoder_pointer, nn.Module), (
-                encoder_pointer, nn.Module
+                f"{decoder_pointer} and {encoder_pointer} have to be of type nn.Module"
-            ), f"{decoder_pointer} and {encoder_pointer} have to be of type nn.Module"
+            )
             if hasattr(decoder_pointer, "weight"):
                 assert hasattr(encoder_pointer, "weight")
                 encoder_pointer.weight = decoder_pointer.weight
@@ -2525,9 +2525,9 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
             encoder_modules = encoder_pointer._modules
             decoder_modules = decoder_pointer._modules
             if len(decoder_modules) > 0:
-                assert (
+                assert len(encoder_modules) > 0, (
-                    len(encoder_modules) > 0
+                    f"Encoder module {encoder_pointer} does not match decoder module {decoder_pointer}"
-                ), f"Encoder module {encoder_pointer} does not match decoder module {decoder_pointer}"
+                )
 
                 all_encoder_weights = {module_name + "/" + sub_name for sub_name in encoder_modules.keys()}
                 encoder_layer_pos = 0
@@ -3571,7 +3571,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
                         f"Please upgrade accelerate with `pip install -U accelerate`"
                     )
                 # init state_dict for this shard
-                shard_state_dict = {name: "" for name in shard}
+                shard_state_dict = dict.fromkeys(shard, "")
                 for module_name in shard:
                     # skip to collect this weight again
                     if shard_state_dict.get(module_name) != "":
@@ -4814,7 +4814,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
                 param_device_map = expand_device_map(device_map, checkpoint_keys)
                 str_dtype = str(dtype).replace("torch.", "") if dtype is not None else "float32"
                 if sharded_metadata is None:
-                    weight_map = {p: checkpoint_files[0] for p in checkpoint_keys}
+                    weight_map = dict.fromkeys(checkpoint_keys, checkpoint_files[0])
                 else:
                     folder = os.path.sep.join(checkpoint_files[0].split(os.path.sep)[:-1])
                     # Fix the weight map keys according to the key mapping
@@ -5446,9 +5446,9 @@ class PoolerEndLogits(nn.Module):
         Returns:
             `torch.FloatTensor`: The end logits for SQuAD.
         """
-        assert (
+        assert start_states is not None or start_positions is not None, (
-            start_states is not None or start_positions is not None
+            "One of start_states, start_positions should be not None"
-        ), "One of start_states, start_positions should be not None"
+        )
         if start_positions is not None:
             slen, hsz = hidden_states.shape[-2:]
             start_positions = start_positions[:, None, None].expand(-1, -1, hsz)  # shape (bsz, 1, hsz)
@@ -5514,9 +5514,9 @@ class PoolerAnswerClass(nn.Module):
         """
         # No dependency on end_feature so that we can obtain one single `cls_logits` for each sample.
         hsz = hidden_states.shape[-1]
-        assert (
+        assert start_states is not None or start_positions is not None, (
-            start_states is not None or start_positions is not None
+            "One of start_states, start_positions should be not None"
-        ), "One of start_states, start_positions should be not None"
+        )
         if start_positions is not None:
             start_positions = start_positions[:, None, None].expand(-1, -1, hsz)  # shape (bsz, 1, hsz)
             start_states = hidden_states.gather(-2, start_positions).squeeze(-2)  # shape (bsz, hsz)

src/transformers/models/altclip/modeling_altclip.py

@@ -1058,7 +1058,7 @@ class AltCLIPVisionEmbeddings(nn.Module):
         batch_size, _, height, width = pixel_values.shape
         if not interpolate_pos_encoding and (height != self.image_size or width != self.image_size):
             raise ValueError(
-                f"Input image size ({height}*{width}) doesn't match model" f" ({self.image_size}*{self.image_size})."
+                f"Input image size ({height}*{width}) doesn't match model ({self.image_size}*{self.image_size})."
             )
         target_dtype = self.patch_embedding.weight.dtype
         patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype))  # shape = [*, width, grid, grid]

src/transformers/models/bridgetower/modeling_bridgetower.py

@@ -329,7 +329,7 @@ class BridgeTowerVisionEmbeddings(nn.Module):
         batch_size, _, height, width = pixel_values.shape
         if not interpolate_pos_encoding and (height != self.image_size or width != self.image_size):
             raise ValueError(
-                f"Input image size ({height}*{width}) doesn't match model" f" ({self.image_size}*{self.image_size})."
+                f"Input image size ({height}*{width}) doesn't match model ({self.image_size}*{self.image_size})."
             )
         target_dtype = self.patch_embedding.weight.dtype
         patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype))  # shape = [*, width, grid, grid]

src/transformers/models/chinese_clip/modeling_chinese_clip.py

@@ -234,7 +234,7 @@ class ChineseCLIPVisionEmbeddings(nn.Module):
         batch_size, _, height, width = pixel_values.shape
         if not interpolate_pos_encoding and (height != self.image_size or width != self.image_size):
             raise ValueError(
-                f"Input image size ({height}*{width}) doesn't match model" f" ({self.image_size}*{self.image_size})."
+                f"Input image size ({height}*{width}) doesn't match model ({self.image_size}*{self.image_size})."
             )
         target_dtype = self.patch_embedding.weight.dtype
         patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype))  # shape = [*, width, grid, grid]

src/transformers/models/clip/modeling_clip.py

@@ -242,7 +242,7 @@ class CLIPVisionEmbeddings(nn.Module):
         batch_size, _, height, width = pixel_values.shape
         if not interpolate_pos_encoding and (height != self.image_size or width != self.image_size):
             raise ValueError(
-                f"Input image size ({height}*{width}) doesn't match model" f" ({self.image_size}*{self.image_size})."
+                f"Input image size ({height}*{width}) doesn't match model ({self.image_size}*{self.image_size})."
             )
         target_dtype = self.patch_embedding.weight.dtype
         patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype))  # shape = [*, width, grid, grid]

src/transformers/models/clipseg/modeling_clipseg.py

@@ -209,7 +209,7 @@ class CLIPSegVisionEmbeddings(nn.Module):
         batch_size, _, height, width = pixel_values.shape
         if not interpolate_pos_encoding and (height != self.image_size or width != self.image_size):
             raise ValueError(
-                f"Input image size ({height}*{width}) doesn't match model" f" ({self.image_size}*{self.image_size})."
+                f"Input image size ({height}*{width}) doesn't match model ({self.image_size}*{self.image_size})."
             )
         patch_embeds = self.patch_embedding(pixel_values)  # shape = [*, width, grid, grid]
         patch_embeds = patch_embeds.flatten(2).transpose(1, 2)

src/transformers/models/clvp/configuration_clvp.py

@@ -144,7 +144,7 @@ class ClvpEncoderConfig(PretrainedConfig):
         # this is to make sure that we can load only text or speech configs from the nested ClvpConfig.
         if config_type not in cls.base_config_key:
             raise ValueError(
-                f"We can only load either 'text_config' or 'speech_config' but you are trying to load" f"{config_type}"
+                f"We can only load either 'text_config' or 'speech_config' but you are trying to load{config_type}"
             )
 
         # get the text config dict if we are loading from ClvpConfig

src/transformers/models/data2vec/convert_data2vec_text_original_pytorch_checkpoint_to_pytorch.py

@@ -127,9 +127,9 @@ def convert_data2vec_checkpoint_to_pytorch(
 
         # self-attention output
         self_output: BertSelfOutput = layer.attention.output
-        assert (
+        assert self_output.dense.weight.shape == data2vec_layer.self_attn.out_proj.weight.shape, (
-            self_output.dense.weight.shape == data2vec_layer.self_attn.out_proj.weight.shape
+            f"Shape for self_output.dense.weight should be {data2vec_layer.self_attn.out_proj.weight.shape}"
-        ), f"Shape for self_output.dense.weight should be {data2vec_layer.self_attn.out_proj.weight.shape}"
+        )
         self_output.dense.weight = data2vec_layer.self_attn.out_proj.weight
         self_output.dense.bias = data2vec_layer.self_attn.out_proj.bias
         self_output.LayerNorm.weight = data2vec_layer.self_attn_layer_norm.weight
@@ -137,17 +137,17 @@ def convert_data2vec_checkpoint_to_pytorch(
 
         # intermediate
         intermediate: BertIntermediate = layer.intermediate
-        assert (
+        assert intermediate.dense.weight.shape == data2vec_layer.fc1.weight.shape, (
-            intermediate.dense.weight.shape == data2vec_layer.fc1.weight.shape
+            f"Shape for intermediate.dense.weight should be {data2vec_layer.fc1.weight.shape}"
-        ), f"Shape for intermediate.dense.weight should be {data2vec_layer.fc1.weight.shape}"
+        )
         intermediate.dense.weight = data2vec_layer.fc1.weight
         intermediate.dense.bias = data2vec_layer.fc1.bias
 
         # output
         bert_output: BertOutput = layer.output
-        assert (
+        assert bert_output.dense.weight.shape == data2vec_layer.fc2.weight.shape, (
-            bert_output.dense.weight.shape == data2vec_layer.fc2.weight.shape
+            f"Shape for bert_output.dense.weight should be {data2vec_layer.fc2.weight.shape}"
-        ), f"Shape for bert_output.dense.weight should be {data2vec_layer.fc2.weight.shape}"
+        )
         bert_output.dense.weight = data2vec_layer.fc2.weight
         bert_output.dense.bias = data2vec_layer.fc2.bias
         bert_output.LayerNorm.weight = data2vec_layer.final_layer_norm.weight

src/transformers/models/deprecated/bort/convert_bort_original_gluonnlp_checkpoint_to_pytorch.py

@@ -180,9 +180,9 @@ def convert_bort_checkpoint_to_pytorch(bort_checkpoint_path: str, pytorch_dump_f
         gluon_param = to_torch(params[gluon_param])
         shape_gluon = gluon_param.shape
 
-        assert (
+        assert shape_hf == shape_gluon, (
-            shape_hf == shape_gluon
+            f"The gluon parameter {gluon_param} has shape {shape_gluon}, but expects shape {shape_hf} for Transformers"
-        ), f"The gluon parameter {gluon_param} has shape {shape_gluon}, but expects shape {shape_hf} for Transformers"
+        )
 
         return gluon_param
 

src/transformers/models/deprecated/gptsan_japanese/tokenization_gptsan_japanese.py

@@ -427,7 +427,7 @@ class SubWordJapaneseTokenizer:
             )
         keisen = "─━│┃┄┅┆┇┈┉┊┋┌┍┎┏┐┑┒┓└┕┖┗┘┙┚┛├┝┞┟┠┡┢┣┤┥┦┧┨┩┪┫┬┭┮┯┰┱┲┳┴┵┶┷┸┹┺┻┼┽┾┿╀╁╂╃╄╅╆╇╈╉╊╋╌╍╎╏═║╒╓╔╕╖╗╘╙╚╛╜╝╞╟╠╡╢╣╤╥╦╧╨╩╪╫╬╭╮╯╰╱╲╳╴╵╶╷╸╹╺╻╼╽╾╿"
         blocks = "▀▁▂▃▄▅▆▇█▉▊▋▌▍▎▏▐░▒▓▔▕▖▗▘▙▚▛▜▝▞▟"
-        self.content_trans1 = str.maketrans({k: "<BLOCK>" for k in keisen + blocks})
+        self.content_trans1 = str.maketrans(dict.fromkeys(keisen + blocks, "<BLOCK>"))
 
     def __len__(self):
         return len(self.ids_to_tokens)

src/transformers/models/deprecated/open_llama/configuration_open_llama.py

@@ -154,7 +154,7 @@ class OpenLlamaConfig(PretrainedConfig):
 
         if not isinstance(self.rope_scaling, dict) or len(self.rope_scaling) != 2:
             raise ValueError(
-                "`rope_scaling` must be a dictionary with two fields, `type` and `factor`, " f"got {self.rope_scaling}"
+                f"`rope_scaling` must be a dictionary with two fields, `type` and `factor`, got {self.rope_scaling}"
             )
         rope_scaling_type = self.rope_scaling.get("type", None)
         rope_scaling_factor = self.rope_scaling.get("factor", None)

src/transformers/models/deprecated/realm/modeling_realm.py

@@ -139,9 +139,9 @@ def load_tf_weights_in_realm(model, config, tf_checkpoint_path):
         elif m_name == "kernel":
             array = np.transpose(array)
         try:
-            assert (
+            assert pointer.shape == array.shape, (
-                pointer.shape == array.shape
+                f"Pointer shape {pointer.shape} and array shape {array.shape} mismatched"
-            ), f"Pointer shape {pointer.shape} and array shape {array.shape} mismatched"
+            )
         except AssertionError as e:
             e.args += (pointer.shape, array.shape)
             raise

src/transformers/models/deprecated/speech_to_text_2/modeling_speech_to_text_2.py

@@ -579,7 +579,7 @@ class Speech2Text2Decoder(Speech2Text2PreTrainedModel):
         if self.gradient_checkpointing and self.training:
             if use_cache:
                 logger.warning_once(
-                    "`use_cache = True` is incompatible with gradient checkpointing. Setting `use_cache =" " False`..."
+                    "`use_cache = True` is incompatible with gradient checkpointing. Setting `use_cache = False`..."
                 )
                 use_cache = False
 

src/transformers/models/deprecated/transfo_xl/modeling_tf_transfo_xl.py

@@ -1095,9 +1095,9 @@ class TFTransfoXLForSequenceClassification(TFTransfoXLPreTrainedModel, TFSequenc
                 batch_size, sequence_length = shape_list(input_ids)[:2]
             else:
                 batch_size, sequence_length = shape_list(inputs_embeds)[:2]
-            assert (
+            assert self.config.pad_token_id is not None or batch_size == 1, (
-                self.config.pad_token_id is not None or batch_size == 1
+                "Cannot handle batch sizes > 1 if no padding token is defined."
-            ), "Cannot handle batch sizes > 1 if no padding token is defined."
+            )
 
             if not tf.is_tensor(sequence_lengths):
                 in_logits = logits[0:batch_size, sequence_lengths]

src/transformers/models/deprecated/transfo_xl/modeling_transfo_xl.py

@@ -155,9 +155,9 @@ def load_tf_weights_in_transfo_xl(model, config, tf_path):
                 p_i.data = torch.from_numpy(arr_i)
         else:
             try:
-                assert (
+                assert pointer.shape == array.shape, (
-                    pointer.shape == array.shape
+                    f"Pointer shape {pointer.shape} and array shape {array.shape} mismatched"
-                ), f"Pointer shape {pointer.shape} and array shape {array.shape} mismatched"
+                )
             except AssertionError as e:
                 e.args += (pointer.shape, array.shape)
                 raise
@@ -1238,9 +1238,9 @@ class TransfoXLForSequenceClassification(TransfoXLPreTrainedModel):
         else:
             batch_size, sequence_length = inputs_embeds.shape[:2]
 
-        assert (
+        assert self.config.pad_token_id is not None or batch_size == 1, (
-            self.config.pad_token_id is not None or batch_size == 1
+            "Cannot handle batch sizes > 1 if no padding token is defined."
-        ), "Cannot handle batch sizes > 1 if no padding token is defined."
+        )
         if self.config.pad_token_id is None:
             sequence_lengths = -1
         else:

src/transformers/models/deprecated/xlm_prophetnet/modeling_xlm_prophetnet.py

@@ -588,9 +588,9 @@ class XLMProphetNetPositionalEmbeddings(nn.Embedding):
         super().__init__(config.max_position_embeddings, config.hidden_size, config.pad_token_id)
 
     def forward(self, inputs_shape, device, attention_mask=None, past_key_values=None, position_ids=None):
-        assert (position_ids is None) or (
+        assert (position_ids is None) or (self.padding_idx is None), (
-            self.padding_idx is None
+            "If position_ids is pre-computed then padding_idx should not be set."
-        ), "If position_ids is pre-computed then padding_idx should not be set."
+        )
 
         if position_ids is None:
             if past_key_values is not None:
@@ -784,9 +784,9 @@ class XLMProphetNetNgramSelfAttention(nn.Module):
         self.head_dim = config.hidden_size // self.num_attn_heads
         self.ngram = config.ngram
 
-        assert (
+        assert self.head_dim * self.num_attn_heads == config.hidden_size, (
-            self.head_dim * self.num_attn_heads == config.hidden_size
+            "config.hidden_size must be divisible by num_attn_heads"
-        ), "config.hidden_size must be divisible by num_attn_heads"
+        )
         # key, value, query projection
         self.key_proj = nn.Linear(config.hidden_size, config.hidden_size)
         self.value_proj = nn.Linear(config.hidden_size, config.hidden_size)
@@ -1041,9 +1041,9 @@ class XLMProphetNetNgramSelfAttention(nn.Module):
 
         if predict_relative_position_buckets is None:
             key_sequence_length = attn_weights.shape[-1]
-            assert (
+            assert position_ids[0][0] == key_sequence_length - 1, (
-                position_ids[0][0] == key_sequence_length - 1
+                "`position_ids` are incorrect. They should be of the format 1 2 3 4 5 ... (key_sequence_length - 1)"
-            ), "`position_ids` are incorrect. They should be of the format 1 2 3 4 5 ... (key_sequence_length - 1)"
+            )
             relative_positions = (
                 torch.arange(0, key_sequence_length)
                 .unsqueeze(0)
@@ -1313,9 +1313,9 @@ class XLMProphetNetEncoder(XLMProphetNetPreTrainedModel):
 
         # check if head_mask has a correct number of layers specified if desired
         if head_mask is not None:
-            assert head_mask.size()[0] == (
+            assert head_mask.size()[0] == (len(self.layers)), (
-                len(self.layers)
+                f"The head_mask should be specified for {len(self.layers)} layers, but it is for {head_mask.size()[0]}."
-            ), f"The head_mask should be specified for {len(self.layers)} layers, but it is for {head_mask.size()[0]}."
+            )
         for idx, encoder_layer in enumerate(self.layers):
             if output_hidden_states:
                 encoder_hidden_states = encoder_hidden_states + (hidden_states,)
@@ -1488,9 +1488,9 @@ class XLMProphetNetDecoder(XLMProphetNetPreTrainedModel):
 
         # prepare attention mask
         if past_key_values is not None:
-            assert (
+            assert hidden_states.size(1) == 1, (
-                hidden_states.size(1) == 1
+                "At the moment `use_cache` is only supported for `decoder_input_ids` of length 1"
-            ), "At the moment `use_cache` is only supported for `decoder_input_ids` of length 1"
+            )
 
             ngram_hidden_states = [
                 (ngram_embeddings[ngram - 1] + predicting_stream_pos_embed).repeat(batch_size, 1, 1)

src/transformers/models/depth_pro/configuration_depth_pro.py

@@ -114,7 +114,7 @@ class DepthProConfig(PretrainedConfig):
         # scaled_images_ratios is sorted
         if scaled_images_ratios != sorted(scaled_images_ratios):
             raise ValueError(
-                f"Values in scaled_images_ratios={scaled_images_ratios} " "should be sorted from low to high"
+                f"Values in scaled_images_ratios={scaled_images_ratios} should be sorted from low to high"
             )
 
         # scaled_images_ratios, scaled_images_overlap_ratios, scaled_images_feature_dims should be consistent

src/transformers/models/distilbert/modeling_flax_distilbert.py

@@ -275,9 +275,9 @@ class FlaxTransformerBlock(nn.Module):
     dtype: jnp.dtype = jnp.float32  # the dtype of the computation
 
     def setup(self):
-        assert (
+        assert self.config.dim % self.config.n_heads == 0, (
-            self.config.dim % self.config.n_heads == 0
+            f"Hidden size {self.config.dim} not dividable by number of heads {self.config.n_heads}"
-        ), f"Hidden size {self.config.dim} not dividable by number of heads {self.config.n_heads}"
+        )
 
         self.attention = FlaxMultiHeadSelfAttention(self.config, dtype=self.dtype)
         self.sa_layer_norm = nn.LayerNorm(epsilon=1e-12, dtype=self.dtype)

src/transformers/models/distilbert/modeling_tf_distilbert.py

@@ -269,9 +269,9 @@ class TFTransformerBlock(keras.layers.Layer):
         self.activation = config.activation
         self.output_attentions = config.output_attentions
 
-        assert (
+        assert config.dim % config.n_heads == 0, (
-            config.dim % config.n_heads == 0
+            f"Hidden size {config.dim} not dividable by number of heads {config.n_heads}"
-        ), f"Hidden size {config.dim} not dividable by number of heads {config.n_heads}"
+        )
 
         self.attention = TFMultiHeadSelfAttention(config, name="attention")
         self.sa_layer_norm = keras.layers.LayerNormalization(epsilon=1e-12, name="sa_layer_norm")

src/transformers/models/dpr/convert_dpr_original_checkpoint_to_pytorch.py

@@ -137,7 +137,7 @@ if __name__ == "__main__":
     dest_dir = f"converted-{src_file.name}" if args.dest is None else args.dest
     dest_dir = Path(dest_dir)
     assert src_file.exists()
-    assert (
+    assert args.type is not None, (
-        args.type is not None
+        "Please specify the component type of the DPR model to convert: 'ctx_encoder', 'question_encoder' or 'reader'."
-    ), "Please specify the component type of the DPR model to convert: 'ctx_encoder', 'question_encoder' or 'reader'."
+    )
     convert(args.type, src_file, dest_dir)

src/transformers/models/dpr/tokenization_dpr_fast.py

@@ -170,9 +170,9 @@ class CustomDPRReaderTokenizerMixin:
         texts = texts if not isinstance(texts, str) else [texts]
         n_passages = len(titles)
         questions = questions if not isinstance(questions, str) else [questions] * n_passages
-        assert len(titles) == len(
+        assert len(titles) == len(texts), (
-            texts
+            f"There should be as many titles than texts but got {len(titles)} titles and {len(texts)} texts."
-        ), f"There should be as many titles than texts but got {len(titles)} titles and {len(texts)} texts."
+        )
         encoded_question_and_titles = super().__call__(questions, titles, padding=False, truncation=False)["input_ids"]
         encoded_texts = super().__call__(texts, add_special_tokens=False, padding=False, truncation=False)["input_ids"]
         encoded_inputs = {

src/transformers/models/encodec/modeling_encodec.py

@@ -617,8 +617,7 @@ class EncodecModel(EncodecPreTrainedModel):
             bandwidth = self.config.target_bandwidths[0]
         if bandwidth not in self.config.target_bandwidths:
             raise ValueError(
-                f"This model doesn't support the bandwidth {bandwidth}. "
+                f"This model doesn't support the bandwidth {bandwidth}. Select one of {self.config.target_bandwidths}."
-                f"Select one of {self.config.target_bandwidths}."
             )
 
         _, channels, input_length = input_values.shape

src/transformers/models/esm/openfold_utils/residue_constants.py

@@ -399,13 +399,11 @@ def map_structure_with_atom_order(in_list: list, first_call: bool = True) -> lis
 
 
 @functools.lru_cache(maxsize=None)
-def load_stereo_chemical_props() -> (
+def load_stereo_chemical_props() -> Tuple[
-    Tuple[
     Mapping[str, List[Bond]],
     Mapping[str, List[Bond]],
     Mapping[str, List[BondAngle]],
-    ]
+]:
-):
     """Load stereo_chemical_props.txt into a nice structure.
 
     Load literature values for bond lengths and bond angles and translate bond angles into the length of the opposite

src/transformers/models/fsmt/modeling_fsmt.py

@@ -539,9 +539,9 @@ class FSMTEncoder(nn.Module):
         all_attentions = () if output_attentions else None
         # check if head_mask has a correct number of layers specified if desired
         if head_mask is not None:
-            assert head_mask.size()[0] == (
+            assert head_mask.size()[0] == (len(self.layers)), (
-                len(self.layers)
+                f"The head_mask should be specified for {len(self.layers)} layers, but it is for {head_mask.size()[0]}."
-            ), f"The head_mask should be specified for {len(self.layers)} layers, but it is for {head_mask.size()[0]}."
+            )
         for idx, encoder_layer in enumerate(self.layers):
             if output_hidden_states:
                 x = x.transpose(0, 1)  # T x B x C -> B x T x C
@@ -960,9 +960,9 @@ class Attention(nn.Module):
         attn_weights = nn.functional.softmax(attn_weights, dim=-1)
 
         if layer_head_mask is not None:
-            assert layer_head_mask.size() == (
+            assert layer_head_mask.size() == (self.num_heads,), (
-                self.num_heads,
+                f"Head mask for a single layer should be of size {(self.num_heads,)}, but is {layer_head_mask.size()}"
-            ), f"Head mask for a single layer should be of size {(self.num_heads,)}, but is {layer_head_mask.size()}"
+            )
             attn_weights = layer_head_mask.view(1, -1, 1, 1) * attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
             attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
 

src/transformers/models/funnel/configuration_funnel.py

@@ -113,9 +113,9 @@ class FunnelConfig(PretrainedConfig):
         self.vocab_size = vocab_size
         self.block_sizes = block_sizes
         self.block_repeats = [1] * len(block_sizes) if block_repeats is None else block_repeats
-        assert len(block_sizes) == len(
+        assert len(block_sizes) == len(self.block_repeats), (
-            self.block_repeats
+            "`block_sizes` and `block_repeats` should have the same length."
-        ), "`block_sizes` and `block_repeats` should have the same length."
+        )
         self.num_decoder_layers = num_decoder_layers
         self.d_model = d_model
         self.n_head = n_head

src/transformers/models/fuyu/configuration_fuyu.py

@@ -195,7 +195,7 @@ class FuyuConfig(PretrainedConfig):
 
         if not isinstance(self.rope_scaling, dict) or len(self.rope_scaling) != 2:
             raise ValueError(
-                "`rope_scaling` must be a dictionary with two fields, `type` and `factor`, " f"got {self.rope_scaling}"
+                f"`rope_scaling` must be a dictionary with two fields, `type` and `factor`, got {self.rope_scaling}"
             )
         rope_scaling_type = self.rope_scaling.get("type", None)
         rope_scaling_factor = self.rope_scaling.get("factor", None)

src/transformers/models/git/modeling_git.py

@@ -683,7 +683,7 @@ class GitVisionEmbeddings(nn.Module):
         batch_size, _, height, width = pixel_values.shape
         if not interpolate_pos_encoding and (height != self.image_size or width != self.image_size):
             raise ValueError(
-                f"Input image size ({height}*{width}) doesn't match model" f" ({self.image_size}*{self.image_size})."
+                f"Input image size ({height}*{width}) doesn't match model ({self.image_size}*{self.image_size})."
             )
         target_dtype = self.patch_embedding.weight.dtype
         patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype))  # shape = [*, width, grid, grid]

src/transformers/models/gpt_neox_japanese/tokenization_gpt_neox_japanese.py

@@ -250,7 +250,7 @@ class SubWordJapaneseTokenizer:
             )
         keisen = "─━│┃┄┅┆┇┈┉┊┋┌┍┎┏┐┑┒┓└┕┖┗┘┙┚┛├┝┞┟┠┡┢┣┤┥┦┧┨┩┪┫┬┭┮┯┰┱┲┳┴┵┶┷┸┹┺┻┼┽┾┿╀╁╂╃╄╅╆╇╈╉╊╋╌╍╎╏═║╒╓╔╕╖╗╘╙╚╛╜╝╞╟╠╡╢╣╤╥╦╧╨╩╪╫╬╭╮╯╰╱╲╳╴╵╶╷╸╹╺╻╼╽╾╿"
         blocks = "▀▁▂▃▄▅▆▇█▉▊▋▌▍▎▏▐░▒▓▔▕▖▗▘▙▚▛▜▝▞▟"
-        self.content_trans1 = str.maketrans({k: "<BLOCK>" for k in keisen + blocks})
+        self.content_trans1 = str.maketrans(dict.fromkeys(keisen + blocks, "<BLOCK>"))
 
     def __len__(self):
         return len(self.ids_to_tokens)

src/transformers/models/ibert/quant_modules.py

@@ -171,9 +171,9 @@ class QuantAct(nn.Module):
             x_min = x_act.data.min()
             x_max = x_act.data.max()
 
-            assert (
+            assert x_max.isnan().sum() == 0 and x_min.isnan().sum() == 0, (
-                x_max.isnan().sum() == 0 and x_min.isnan().sum() == 0
+                "NaN detected when computing min/max of the activation"
-            ), "NaN detected when computing min/max of the activation"
+            )
 
             # Initialization
             if self.x_min.min() > -1.1e-5 and self.x_max.max() < 1.1e-5:

src/transformers/models/kosmos2/modeling_kosmos2.py

@@ -451,7 +451,7 @@ class Kosmos2VisionEmbeddings(nn.Module):
         batch_size, _, height, width = pixel_values.shape
         if not interpolate_pos_encoding and (height != self.image_size or width != self.image_size):
             raise ValueError(
-                f"Input image size ({height}*{width}) doesn't match model" f" ({self.image_size}*{self.image_size})."
+                f"Input image size ({height}*{width}) doesn't match model ({self.image_size}*{self.image_size})."
             )
         target_dtype = self.patch_embedding.weight.dtype
         patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype))  # shape = [*, width, grid, grid]

src/transformers/models/layoutxlm/processing_layoutxlm.py

@@ -101,8 +101,7 @@ class LayoutXLMProcessor(ProcessorMixin):
         # verify input
         if self.image_processor.apply_ocr and (boxes is not None):
             raise ValueError(
-                "You cannot provide bounding boxes "
+                "You cannot provide bounding boxes if you initialized the image processor with apply_ocr set to True."
-                "if you initialized the image processor with apply_ocr set to True."
             )
 
         if self.image_processor.apply_ocr and (word_labels is not None):

src/transformers/models/led/modeling_led.py

@@ -130,12 +130,12 @@ class LEDEncoderSelfAttention(nn.Module):
 
         self.layer_id = layer_id
         attention_window = config.attention_window[self.layer_id]
-        assert (
+        assert attention_window % 2 == 0, (
-            attention_window % 2 == 0
+            f"`attention_window` for layer {self.layer_id} has to be an even value. Given {attention_window}"
-        ), f"`attention_window` for layer {self.layer_id} has to be an even value. Given {attention_window}"
+        )
-        assert (
+        assert attention_window > 0, (
-            attention_window > 0
+            f"`attention_window` for layer {self.layer_id} has to be positive. Given {attention_window}"
-        ), f"`attention_window` for layer {self.layer_id} has to be positive. Given {attention_window}"
+        )
 
         self.one_sided_attn_window_size = attention_window // 2
 
@@ -169,9 +169,9 @@ class LEDEncoderSelfAttention(nn.Module):
         value_vectors = self.value(hidden_states)
 
         seq_len, batch_size, embed_dim = hidden_states.size()
-        assert (
+        assert embed_dim == self.embed_dim, (
-            embed_dim == self.embed_dim
+            f"hidden_states should have embed_dim = {self.embed_dim}, but has {embed_dim}"
-        ), f"hidden_states should have embed_dim = {self.embed_dim}, but has {embed_dim}"
+        )
 
         # normalize query
         query_vectors /= math.sqrt(self.head_dim)
@@ -239,9 +239,9 @@ class LEDEncoderSelfAttention(nn.Module):
         )  # use fp32 for numerical stability
 
         if layer_head_mask is not None:
-            assert layer_head_mask.size() == (
+            assert layer_head_mask.size() == (self.num_heads,), (
-                self.num_heads,
+                f"Head mask for a single layer should be of size {(self.num_heads,)}, but is {layer_head_mask.size()}"
-            ), f"Head mask for a single layer should be of size {(self.num_heads,)}, but is {layer_head_mask.size()}"
+            )
             attn_probs = layer_head_mask.view(1, 1, -1, 1) * attn_probs
 
         # softmax sometimes inserts NaN if all positions are masked, replace them with 0
@@ -433,9 +433,9 @@ class LEDEncoderSelfAttention(nn.Module):
         overlap of size window_overlap
         """
         batch_size, seq_len, num_heads, head_dim = query.size()
-        assert (
+        assert seq_len % (window_overlap * 2) == 0, (
-            seq_len % (window_overlap * 2) == 0
+            f"Sequence length should be multiple of {window_overlap * 2}. Given {seq_len}"
-        ), f"Sequence length should be multiple of {window_overlap * 2}. Given {seq_len}"
+        )
         assert query.size() == key.size()
 
         chunks_count = torch.div(seq_len, window_overlap, rounding_mode="trunc") - 1
@@ -706,9 +706,9 @@ class LEDEncoderSelfAttention(nn.Module):
 
         # apply layer head masking
         if layer_head_mask is not None:
-            assert layer_head_mask.size() == (
+            assert layer_head_mask.size() == (self.num_heads,), (
-                self.num_heads,
+                f"Head mask for a single layer should be of size {(self.num_heads,)}, but is {layer_head_mask.size()}"
-            ), f"Head mask for a single layer should be of size {(self.num_heads,)}, but is {layer_head_mask.size()}"
+            )
             global_attn_probs_float = layer_head_mask.view(1, -1, 1, 1) * global_attn_probs_float.view(
                 batch_size, self.num_heads, max_num_global_attn_indices, seq_len
             )

src/transformers/models/led/modeling_tf_led.py

@@ -182,12 +182,12 @@ class TFLEDEncoderSelfAttention(keras.layers.Layer):
         self.layer_id = layer_id
         attention_window = config.attention_window[self.layer_id]
 
-        assert (
+        assert attention_window % 2 == 0, (
-            attention_window % 2 == 0
+            f"`attention_window` for layer {self.layer_id} has to be an even value. Given {attention_window}"
-        ), f"`attention_window` for layer {self.layer_id} has to be an even value. Given {attention_window}"
+        )
-        assert (
+        assert attention_window > 0, (
-            attention_window > 0
+            f"`attention_window` for layer {self.layer_id} has to be positive. Given {attention_window}"
-        ), f"`attention_window` for layer {self.layer_id} has to be positive. Given {attention_window}"
+        )
 
         self.one_sided_attn_window_size = attention_window // 2
 

src/transformers/models/longformer/modeling_longformer.py

@@ -510,12 +510,12 @@ class LongformerSelfAttention(nn.Module):
 
         self.layer_id = layer_id
         attention_window = config.attention_window[self.layer_id]
-        assert (
+        assert attention_window % 2 == 0, (
-            attention_window % 2 == 0
+            f"`attention_window` for layer {self.layer_id} has to be an even value. Given {attention_window}"
-        ), f"`attention_window` for layer {self.layer_id} has to be an even value. Given {attention_window}"
+        )
-        assert (
+        assert attention_window > 0, (
-            attention_window > 0
+            f"`attention_window` for layer {self.layer_id} has to be positive. Given {attention_window}"
-        ), f"`attention_window` for layer {self.layer_id} has to be positive. Given {attention_window}"
+        )
 
         self.one_sided_attn_window_size = attention_window // 2
 
@@ -549,9 +549,9 @@ class LongformerSelfAttention(nn.Module):
         value_vectors = self.value(hidden_states)
 
         seq_len, batch_size, embed_dim = hidden_states.size()
-        assert (
+        assert embed_dim == self.embed_dim, (
-            embed_dim == self.embed_dim
+            f"hidden_states should have embed_dim = {self.embed_dim}, but has {embed_dim}"
-        ), f"hidden_states should have embed_dim = {self.embed_dim}, but has {embed_dim}"
+        )
 
         # normalize query
         query_vectors /= math.sqrt(self.head_dim)
@@ -619,9 +619,9 @@ class LongformerSelfAttention(nn.Module):
         )  # use fp32 for numerical stability
 
         if layer_head_mask is not None:
-            assert layer_head_mask.size() == (
+            assert layer_head_mask.size() == (self.num_heads,), (
-                self.num_heads,
+                f"Head mask for a single layer should be of size {(self.num_heads,)}, but is {layer_head_mask.size()}"
-            ), f"Head mask for a single layer should be of size {(self.num_heads,)}, but is {layer_head_mask.size()}"
+            )
             attn_probs = layer_head_mask.view(1, 1, -1, 1) * attn_probs
 
         # softmax sometimes inserts NaN if all positions are masked, replace them with 0
@@ -813,9 +813,9 @@ class LongformerSelfAttention(nn.Module):
         overlap of size window_overlap
         """
         batch_size, seq_len, num_heads, head_dim = query.size()
-        assert (
+        assert seq_len % (window_overlap * 2) == 0, (
-            seq_len % (window_overlap * 2) == 0
+            f"Sequence length should be multiple of {window_overlap * 2}. Given {seq_len}"
-        ), f"Sequence length should be multiple of {window_overlap * 2}. Given {seq_len}"
+        )
         assert query.size() == key.size()
 
         chunks_count = torch.div(seq_len, window_overlap, rounding_mode="trunc") - 1
@@ -1086,9 +1086,9 @@ class LongformerSelfAttention(nn.Module):
 
         # apply layer head masking
         if layer_head_mask is not None:
-            assert layer_head_mask.size() == (
+            assert layer_head_mask.size() == (self.num_heads,), (
-                self.num_heads,
+                f"Head mask for a single layer should be of size {(self.num_heads,)}, but is {layer_head_mask.size()}"
-            ), f"Head mask for a single layer should be of size {(self.num_heads,)}, but is {layer_head_mask.size()}"
+            )
             global_attn_probs_float = layer_head_mask.view(1, -1, 1, 1) * global_attn_probs_float.view(
                 batch_size, self.num_heads, max_num_global_attn_indices, seq_len
             )
@@ -1287,9 +1287,9 @@ class LongformerEncoder(nn.Module):
 
         # check if head_mask has a correct number of layers specified if desired
         if head_mask is not None:
-            assert head_mask.size()[0] == (
+            assert head_mask.size()[0] == (len(self.layer)), (
-                len(self.layer)
+                f"The head_mask should be specified for {len(self.layer)} layers, but it is for {head_mask.size()[0]}."
-            ), f"The head_mask should be specified for {len(self.layer)} layers, but it is for {head_mask.size()[0]}."
+            )
         for idx, layer_module in enumerate(self.layer):
             if output_hidden_states:
                 all_hidden_states = all_hidden_states + (hidden_states,)
@@ -1590,8 +1590,7 @@ class LongformerModel(LongformerPreTrainedModel):
         # this path should be recorded in the ONNX export, it is fine with padding_len == 0 as well
         if padding_len > 0:
             logger.warning_once(
-                f"Input ids are automatically padded to be a multiple of "
+                f"Input ids are automatically padded to be a multiple of `config.attention_window`: {attention_window}"
-                f"`config.attention_window`: {attention_window}"
             )
             if input_ids is not None:
                 input_ids = nn.functional.pad(input_ids, (0, padding_len), value=pad_token_id)

src/transformers/models/longformer/modeling_tf_longformer.py

@@ -746,12 +746,12 @@ class TFLongformerSelfAttention(keras.layers.Layer):
         self.layer_id = layer_id
         attention_window = config.attention_window[self.layer_id]
 
-        assert (
+        assert attention_window % 2 == 0, (
-            attention_window % 2 == 0
+            f"`attention_window` for layer {self.layer_id} has to be an even value. Given {attention_window}"
-        ), f"`attention_window` for layer {self.layer_id} has to be an even value. Given {attention_window}"
+        )
-        assert (
+        assert attention_window > 0, (
-            attention_window > 0
+            f"`attention_window` for layer {self.layer_id} has to be positive. Given {attention_window}"
-        ), f"`attention_window` for layer {self.layer_id} has to be positive. Given {attention_window}"
+        )
 
         self.one_sided_attn_window_size = attention_window // 2
 

src/transformers/models/m2m_100/modeling_m2m_100.py

@@ -1294,7 +1294,7 @@ class M2M100Decoder(M2M100PreTrainedModel):
         if self.gradient_checkpointing and self.training:
             if use_cache:
                 logger.warning_once(
-                    "`use_cache=True` is incompatible with gradient checkpointing. Setting" " `use_cache=False`..."
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                 )
                 use_cache = False
 

src/transformers/models/marian/convert_marian_tatoeba_to_pytorch.py

@@ -228,7 +228,7 @@ class TatoebaConverter:
         # combine with Tatoeba markdown
         readme_url = f"{TATOEBA_MODELS_URL}/{model_dict['_name']}/README.md"
         extra_markdown = f"""
-### {model_dict['_name']}
+### {model_dict["_name"]}
 
 * source language name: {self.tag2name[a3_src]}
 * target language name: {self.tag2name[a3_tgt]}
@@ -237,12 +237,12 @@ class TatoebaConverter:
 
         content = (
             f"""
-* model: {model_dict['modeltype']}
+* model: {model_dict["modeltype"]}
-* source language code{src_multilingual*'s'}: {', '.join(a2_src_tags)}
+* source language code{src_multilingual * "s"}: {", ".join(a2_src_tags)}
-* target language code{tgt_multilingual*'s'}: {', '.join(a2_tgt_tags)}
+* target language code{tgt_multilingual * "s"}: {", ".join(a2_tgt_tags)}
 * dataset: opus {backtranslated_data}
-* release date: {model_dict['release-date']}
+* release date: {model_dict["release-date"]}
-* pre-processing: {model_dict['pre-processing']}
+* pre-processing: {model_dict["pre-processing"]}
 """
             + multilingual_data
             + tuned

src/transformers/models/marian/modeling_marian.py

@@ -741,9 +741,9 @@ class MarianEncoder(MarianPreTrainedModel):
 
         # check if head_mask has a correct number of layers specified if desired
         if head_mask is not None:
-            assert head_mask.size()[0] == (
+            assert head_mask.size()[0] == (len(self.layers)), (
-                len(self.layers)
+                f"The head_mask should be specified for {len(self.layers)} layers, but it is for {head_mask.size()[0]}."
-            ), f"The head_mask should be specified for {len(self.layers)} layers, but it is for {head_mask.size()[0]}."
+            )
         for idx, encoder_layer in enumerate(self.layers):
             if output_hidden_states:
                 encoder_states = encoder_states + (hidden_states,)

src/transformers/models/marian/tokenization_marian.py

@@ -339,7 +339,7 @@ class MarianTokenizer(PreTrainedTokenizer):
     def __getstate__(self) -> Dict:
         state = self.__dict__.copy()
         state.update(
-            {k: None for k in ["spm_source", "spm_target", "current_spm", "punc_normalizer", "target_vocab_file"]}
+            dict.fromkeys(["spm_source", "spm_target", "current_spm", "punc_normalizer", "target_vocab_file"])
         )
         return state
 

src/transformers/models/mask2former/convert_mask2former_original_pytorch_checkpoint_to_pytorch.py

@@ -863,9 +863,9 @@ def test(
         for original_model_feature, our_model_feature in zip(
             original_model_backbone_features.values(), our_model_output.encoder_hidden_states
         ):
-            assert torch.allclose(
+            assert torch.allclose(original_model_feature, our_model_feature, atol=tolerance), (
-                original_model_feature, our_model_feature, atol=tolerance
+                "The backbone features are not the same."
-            ), "The backbone features are not the same."
+            )
 
         # Test pixel decoder
         mask_features, _, multi_scale_features = original_model.sem_seg_head.pixel_decoder.forward_features(
@@ -875,9 +875,9 @@ def test(
         for original_model_feature, our_model_feature in zip(
             multi_scale_features, our_model_output.pixel_decoder_hidden_states
         ):
-            assert torch.allclose(
+            assert torch.allclose(original_model_feature, our_model_feature, atol=tolerance), (
-                original_model_feature, our_model_feature, atol=tolerance
+                "The pixel decoder feature are not the same"
-            ), "The pixel decoder feature are not the same"
+            )
 
         # Let's test the full model
         tr_complete = T.Compose(
@@ -894,12 +894,12 @@ def test(
 
         assert original_mask_logits.shape == our_mask_logits.shape, "Output masks shapes are not matching."
         assert original_class_logits.shape == our_class_logits.shape, "Output class logits shapes are not matching."
-        assert torch.allclose(
+        assert torch.allclose(original_class_logits, our_class_logits, atol=tolerance), (
-            original_class_logits, our_class_logits, atol=tolerance
+            "The class logits are not the same."
-        ), "The class logits are not the same."
+        )
-        assert torch.allclose(
+        assert torch.allclose(original_mask_logits, our_mask_logits, atol=tolerance), (
-            original_mask_logits, our_mask_logits, atol=tolerance
+            "The predicted masks are not the same."
-        ), "The predicted masks are not the same."
+        )
 
         logger.info("✅ Test passed!")
 

src/transformers/models/maskformer/convert_maskformer_original_pytorch_checkpoint_to_pytorch.py

@@ -581,9 +581,9 @@ def test(original_model, our_model: MaskFormerForInstanceSegmentation, image_pro
         for original_model_feature, our_model_feature in zip(
             original_model_backbone_features.values(), our_model_output.encoder_hidden_states
         ):
-            assert torch.allclose(
+            assert torch.allclose(original_model_feature, our_model_feature, atol=1e-3), (
-                original_model_feature, our_model_feature, atol=1e-3
+                "The backbone features are not the same."
-            ), "The backbone features are not the same."
+            )
 
         original_model_pixel_out = original_model.sem_seg_head.pixel_decoder.forward_features(
             original_model_backbone_features
@@ -602,9 +602,9 @@ def test(original_model, our_model: MaskFormerForInstanceSegmentation, image_pro
 
         our_segmentation = image_processor.post_process_segmentation(our_model_out, target_size=(384, 384))
 
-        assert torch.allclose(
+        assert torch.allclose(original_segmentation, our_segmentation, atol=1e-3), (
-            original_segmentation, our_segmentation, atol=1e-3
+            "The segmentation image is not the same."
-        ), "The segmentation image is not the same."
+        )
 
         logger.info("✅ Test passed!")
 

src/transformers/models/mobilebert/modeling_mobilebert.py

@@ -144,9 +144,9 @@ def load_tf_weights_in_mobilebert(model, config, tf_checkpoint_path):
         elif m_name == "kernel":
             array = np.transpose(array)
         try:
-            assert (
+            assert pointer.shape == array.shape, (
-                pointer.shape == array.shape
+                f"Pointer shape {pointer.shape} and array shape {array.shape} mismatched"
-            ), f"Pointer shape {pointer.shape} and array shape {array.shape} mismatched"
+            )
         except AssertionError as e:
             e.args += (pointer.shape, array.shape)
             raise

src/transformers/models/mobilevitv2/convert_mlcvnets_to_pytorch.py

@@ -99,9 +99,9 @@ def get_mobilevitv2_config(task_name, orig_cfg_file):
     orig_config = load_orig_config_file(orig_cfg_file)
     assert getattr(orig_config, "model.classification.name", -1) == "mobilevit_v2", "Invalid model"
     config.width_multiplier = getattr(orig_config, "model.classification.mitv2.width_multiplier", 1.0)
-    assert (
+    assert getattr(orig_config, "model.classification.mitv2.attn_norm_layer", -1) == "layer_norm_2d", (
-        getattr(orig_config, "model.classification.mitv2.attn_norm_layer", -1) == "layer_norm_2d"
+        "Norm layers other than layer_norm_2d is not supported"
-    ), "Norm layers other than layer_norm_2d is not supported"
+    )
     config.hidden_act = getattr(orig_config, "model.classification.activation.name", "swish")
     # config.image_size == getattr(orig_config,  'sampler.bs.crop_size_width', 256)
 
@@ -184,7 +184,9 @@ def create_rename_keys(state_dict, base_model=False):
                 )
 
             if f"layer_{i}.1.conv_proj." in k:
-                k_new = k_new.replace(f"layer_{i}.1.conv_proj.", f"{model_prefix}encoder.layer.{i-1}.conv_projection.")
+                k_new = k_new.replace(
+                    f"layer_{i}.1.conv_proj.", f"{model_prefix}encoder.layer.{i - 1}.conv_projection."
+                )
 
         if "pre_norm_attn.0." in k:
             k_new = k_new.replace("pre_norm_attn.0.", "layernorm_before.")

src/transformers/models/moonshine/convert_usefulsensors_to_hf.py

@@ -56,7 +56,7 @@ def _read_h5_weights(group, current_key="", weights={}):
 def _convert_layer_names(name, gated_mlp=False):
     name = re.sub(
         r"layers\.functional(?:_(\d+))?\.layers",
-        lambda m: f'layers.{m.group(1) if m.group(1) else "0"}',
+        lambda m: f"layers.{m.group(1) if m.group(1) else '0'}",
         name,
         count=1,
     )

src/transformers/models/mt5/modeling_mt5.py

@@ -719,9 +719,9 @@ def load_tf_weights_in_mt5(model, config, tf_checkpoint_path):
             logger.info(f"Transposing numpy weight of shape {array.shape} for {name}")
             array = np.transpose(array)
         try:
-            assert (
+            assert pointer.shape == array.shape, (
-                pointer.shape == array.shape
+                f"Pointer shape {pointer.shape} and array shape {array.shape} mismatched"
-            ), f"Pointer shape {pointer.shape} and array shape {array.shape} mismatched"
+            )
         except AssertionError as e:
             e.args += (pointer.shape, array.shape)
             raise

src/transformers/models/nemotron/convert_nemotron_nemo_to_hf.py

@@ -65,13 +65,13 @@ def get_args():
         "--hf_input_path",
         type=str,
         default=None,
-        help="A HF model path, " "e.g. a folder containing https://huggingface.co/nvidia/Minitron-8B-Base",
+        help="A HF model path, e.g. a folder containing https://huggingface.co/nvidia/Minitron-8B-Base",
     )
     parser.add_argument(
         "--hf_output_path",
         type=str,
         default=None,
-        help="Output HF model path, " "with the same format as above but user's own weights",
+        help="Output HF model path, with the same format as above but user's own weights",
     )
     parser.add_argument(
         "--precision",

src/transformers/models/nllb_moe/convert_nllb_moe_sharded_original_checkpoint_to_pytorch.py

@@ -91,7 +91,9 @@ def shard_on_the_fly(switch_checkpoint_path, dump_path, num_experts, dtype, weig
             )
 
     # Add the last block
-    save_path = os.path.join(dump_path, weights_name.replace(".bin", f"-{len(sharded_state_dicts)+1:05d}-of-???.bin"))
+    save_path = os.path.join(
+        dump_path, weights_name.replace(".bin", f"-{len(sharded_state_dicts) + 1:05d}-of-???.bin")
+    )
     shared_weights = torch.load(switch_checkpoint_path + "-shared.pt")["model"]
     remove_ignore_keys_(shared_weights)
     shared_weights = rename_fairseq_keys(shared_weights, None)

src/transformers/models/nllb_moe/modeling_nllb_moe.py

@@ -1352,7 +1352,7 @@ class NllbMoeDecoder(NllbMoePreTrainedModel):
         if self.gradient_checkpointing and self.training:
             if use_cache:
                 logger.warning_once(
-                    "`use_cache=True` is incompatible with gradient checkpointing. Setting" " `use_cache=False`..."
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                 )
                 use_cache = False
 

src/transformers/models/olmo2/configuration_olmo2.py

@@ -5,7 +5,6 @@
 #                          modular_olmo2.py file directly. One of our CI enforces this.
 #                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
 
-
 from ...configuration_utils import PretrainedConfig
 
 

src/transformers/models/olmo2/modular_olmo2.py

@@ -1,7 +1,7 @@
 from typing import Callable, Optional, Tuple
 
 import torch
-from torch import nn
+import torch.nn as nn
 
 from ...cache_utils import Cache
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS

src/transformers/models/oneformer/convert_to_hf_oneformer.py

@@ -1010,9 +1010,9 @@ def test(
         for original_model_feature, our_model_feature in zip(
             original_model_backbone_features.values(), our_model_output.encoder_hidden_states
         ):
-            assert torch.allclose(
+            assert torch.allclose(original_model_feature, our_model_feature, atol=3e-3), (
-                original_model_feature, our_model_feature, atol=3e-3
+                "The backbone features are not the same."
-            ), "The backbone features are not the same."
+            )
         mask_features, _, multi_scale_features, _, _ = original_model.sem_seg_head.pixel_decoder.forward_features(
             original_model_backbone_features
         )
@@ -1025,9 +1025,9 @@ def test(
         for original_model_feature, our_model_feature in zip(
             original_pixel_decoder_features, our_model_output.pixel_decoder_hidden_states
         ):
-            assert torch.allclose(
+            assert torch.allclose(original_model_feature, our_model_feature, atol=3e-4), (
-                original_model_feature, our_model_feature, atol=3e-4
+                "The pixel decoder feature are not the same"
-            ), "The pixel decoder feature are not the same"
+            )
 
         tr_complete = T.Compose(
             [
@@ -1049,9 +1049,9 @@ def test(
 
         our_segmentation = post_process_sem_seg_output(our_model_out, target_size=(640, 640))[0]
 
-        assert torch.allclose(
+        assert torch.allclose(original_segmentation, our_segmentation, atol=1e-3), (
-            original_segmentation, our_segmentation, atol=1e-3
+            "The segmentation image is not the same."
-        ), "The segmentation image is not the same."
+        )
 
         logger.info("✅ Test passed!")
 

src/transformers/models/openai/modeling_tf_openai.py

@@ -62,9 +62,9 @@ class TFAttention(keras.layers.Layer):
 
         n_state = nx  # in Attention: n_state=768 (nx=n_embd)
         # [switch nx => n_state from Block to Attention to keep identical to TF implementation]
-        assert (
+        assert n_state % config.n_head == 0, (
-            n_state % config.n_head == 0
+            f"Hidden dimension {n_state} not dividable by number of heads {config.n_head}"
-        ), f"Hidden dimension {n_state} not dividable by number of heads {config.n_head}"
+        )
         self.n_head = config.n_head
         self.split_size = n_state
         self.scale = scale

src/transformers/models/owlv2/image_processing_owlv2.py

@@ -173,7 +173,7 @@ def _preprocess_resize_output_shape(image, output_shape):
         # multichannel case: append shape of last axis
         output_shape = output_shape + (image.shape[-1],)
     elif output_ndim < image.ndim:
-        raise ValueError("output_shape length cannot be smaller than the " "image number of dimensions")
+        raise ValueError("output_shape length cannot be smaller than the image number of dimensions")
 
     return image, output_shape
 
@@ -345,10 +345,10 @@ class Owlv2ImageProcessor(BaseImageProcessor):
             else:
                 anti_aliasing_sigma = np.atleast_1d(anti_aliasing_sigma) * np.ones_like(factors)
                 if np.any(anti_aliasing_sigma < 0):
-                    raise ValueError("Anti-aliasing standard deviation must be " "greater than or equal to zero")
+                    raise ValueError("Anti-aliasing standard deviation must be greater than or equal to zero")
                 elif np.any((anti_aliasing_sigma > 0) & (factors <= 1)):
                     warnings.warn(
-                        "Anti-aliasing standard deviation greater than zero but " "not down-sampling along all axes"
+                        "Anti-aliasing standard deviation greater than zero but not down-sampling along all axes"
                     )
             filtered = ndi.gaussian_filter(image, anti_aliasing_sigma, cval=cval, mode=ndi_mode)
         else:

src/transformers/models/prophetnet/convert_prophetnet_original_pytorch_checkpoint_to_pytorch.py

@@ -118,9 +118,9 @@ def convert_prophetnet_checkpoint_to_pytorch(prophetnet_checkpoint_path: str, py
                 is_key_init = True
                 break
             elif attribute == "position_embeddings":
-                assert (
+                assert model.position_embeddings.weight.shape[-1] == old_model.embed_positions.weight.shape[-1], (
-                    model.position_embeddings.weight.shape[-1] == old_model.embed_positions.weight.shape[-1]
+                    "Hidden size has to match"
-                ), "Hidden size has to match"
+                )
                 assert model.position_embeddings.weight.shape[0] == 512, "We want 512 position_embeddings."
                 model.position_embeddings.weight = nn.Parameter(old_model.embed_positions.weight[:512, :])
                 is_key_init = True

src/transformers/models/prophetnet/modeling_prophetnet.py

@@ -588,9 +588,9 @@ class ProphetNetPositionalEmbeddings(nn.Embedding):
         super().__init__(config.max_position_embeddings, config.hidden_size, config.pad_token_id)
 
     def forward(self, inputs_shape, device, attention_mask=None, past_key_values=None, position_ids=None):
-        assert (position_ids is None) or (
+        assert (position_ids is None) or (self.padding_idx is None), (
-            self.padding_idx is None
+            "If position_ids is pre-computed then padding_idx should not be set."
-        ), "If position_ids is pre-computed then padding_idx should not be set."
+        )
 
         if position_ids is None:
             if past_key_values is not None:
@@ -784,9 +784,9 @@ class ProphetNetNgramSelfAttention(nn.Module):
         self.head_dim = config.hidden_size // self.num_attn_heads
         self.ngram = config.ngram
 
-        assert (
+        assert self.head_dim * self.num_attn_heads == config.hidden_size, (
-            self.head_dim * self.num_attn_heads == config.hidden_size
+            "config.hidden_size must be divisible by num_attn_heads"
-        ), "config.hidden_size must be divisible by num_attn_heads"
+        )
         # key, value, query projection
         self.key_proj = nn.Linear(config.hidden_size, config.hidden_size)
         self.value_proj = nn.Linear(config.hidden_size, config.hidden_size)
@@ -1041,9 +1041,9 @@ class ProphetNetNgramSelfAttention(nn.Module):
 
         if predict_relative_position_buckets is None:
             key_sequence_length = attn_weights.shape[-1]
-            assert (
+            assert position_ids[0][0] == key_sequence_length - 1, (
-                position_ids[0][0] == key_sequence_length - 1
+                "`position_ids` are incorrect. They should be of the format 1 2 3 4 5 ... (key_sequence_length - 1)"
-            ), "`position_ids` are incorrect. They should be of the format 1 2 3 4 5 ... (key_sequence_length - 1)"
+            )
             relative_positions = (
                 torch.arange(0, key_sequence_length)
                 .unsqueeze(0)
@@ -1313,9 +1313,9 @@ class ProphetNetEncoder(ProphetNetPreTrainedModel):
 
         # check if head_mask has a correct number of layers specified if desired
         if head_mask is not None:
-            assert head_mask.size()[0] == (
+            assert head_mask.size()[0] == (len(self.layers)), (
-                len(self.layers)
+                f"The head_mask should be specified for {len(self.layers)} layers, but it is for {head_mask.size()[0]}."
-            ), f"The head_mask should be specified for {len(self.layers)} layers, but it is for {head_mask.size()[0]}."
+            )
         for idx, encoder_layer in enumerate(self.layers):
             if output_hidden_states:
                 encoder_hidden_states = encoder_hidden_states + (hidden_states,)
@@ -1488,9 +1488,9 @@ class ProphetNetDecoder(ProphetNetPreTrainedModel):
 
         # prepare attention mask
         if past_key_values is not None:
-            assert (
+            assert hidden_states.size(1) == 1, (
-                hidden_states.size(1) == 1
+                "At the moment `use_cache` is only supported for `decoder_input_ids` of length 1"
-            ), "At the moment `use_cache` is only supported for `decoder_input_ids` of length 1"
+            )
 
             ngram_hidden_states = [
                 (ngram_embeddings[ngram - 1] + predicting_stream_pos_embed).repeat(batch_size, 1, 1)

src/transformers/models/pvt/convert_pvt_to_pytorch.py

@@ -162,7 +162,7 @@ def convert_pvt_checkpoint(pvt_size, pvt_checkpoint, pytorch_dump_folder_path):
     elif pvt_size == "large":
         config_path = "Zetatech/pvt-large-224"
     else:
-        raise ValueError(f"Available model's size: 'tiny', 'small', 'medium', 'large', but " f"'{pvt_size}' was given")
+        raise ValueError(f"Available model's size: 'tiny', 'small', 'medium', 'large', but '{pvt_size}' was given")
     config = PvtConfig(name_or_path=config_path)
     # load original model from https://github.com/whai362/PVT
     state_dict = torch.load(pvt_checkpoint, map_location="cpu")
@@ -192,7 +192,7 @@ def convert_pvt_checkpoint(pvt_size, pvt_checkpoint, pytorch_dump_folder_path):
     elif pvt_size == "large":
         expected_slice_logits = torch.tensor([0.3740, -0.7739, -0.4214])
     else:
-        raise ValueError(f"Available model's size: 'tiny', 'small', 'medium', 'large', but " f"'{pvt_size}' was given")
+        raise ValueError(f"Available model's size: 'tiny', 'small', 'medium', 'large', but '{pvt_size}' was given")
 
     assert torch.allclose(logits[0, :3], expected_slice_logits, atol=1e-4)
 

src/transformers/models/pvt_v2/convert_pvt_v2_to_pytorch.py

@@ -203,8 +203,7 @@ def convert_pvt_v2_checkpoint(pvt_v2_size, pvt_v2_checkpoint, pytorch_dump_folde
         config_path = "OpenGVLab/pvt_v2_b5"
     else:
         raise ValueError(
-            f"Available model sizes: 'b0', 'b1', 'b2', 'b2-linear', 'b3', 'b4', 'b5', but "
+            f"Available model sizes: 'b0', 'b1', 'b2', 'b2-linear', 'b3', 'b4', 'b5', but '{pvt_v2_size}' was given"
-            f"'{pvt_v2_size}' was given"
         )
     config = PvtV2Config.from_pretrained(config_path)
     # load original model from https://github.com/whai362/PVT
@@ -248,9 +247,9 @@ def convert_pvt_v2_checkpoint(pvt_v2_size, pvt_v2_checkpoint, pytorch_dump_folde
                 f"'{pvt_v2_size}' was given"
             )
 
-        assert torch.allclose(
+        assert torch.allclose(logits[0, :3], expected_slice_logits, atol=1e-4), (
-            logits[0, :3], expected_slice_logits, atol=1e-4
+            "ImageNet weights not converted successfully."
-        ), "ImageNet weights not converted successfully."
+        )
 
         print("ImageNet weights verified, conversion successful.")
 

src/transformers/models/qwen2_audio/modeling_qwen2_audio.py

@@ -623,9 +623,9 @@ class Qwen2AudioEncoder(Qwen2AudioPreTrainedModel):
 
         # check if head_mask has a correct number of layers specified if desired
         if head_mask is not None:
-            assert head_mask.size()[0] == (
+            assert head_mask.size()[0] == (len(self.layers)), (
-                len(self.layers)
+                f"The head_mask should be specified for {len(self.layers)} layers, but it is for {head_mask.size()[0]}."
-            ), f"The head_mask should be specified for {len(self.layers)} layers, but it is for {head_mask.size()[0]}."
+            )
 
         for idx, encoder_layer in enumerate(self.layers):
             if output_hidden_states:

src/transformers/models/rag/modeling_rag.py

@@ -494,9 +494,9 @@ class RagModel(RagPreTrainedModel):
         retriever: Optional[RagRetriever] = None,  # or maybe just use a `set_retriever(...)` method
         **kwargs,
     ):
-        assert config is not None or (
+        assert config is not None or (question_encoder is not None and generator is not None), (
-            question_encoder is not None and generator is not None
+            "Either a configuration or an question_encoder and a generator has to be provided."
-        ), "Either a configuration or an question_encoder and a generator has to be provided."
+        )
 
         if config is None:
             config = RagConfig.from_question_encoder_generator_configs(
@@ -517,9 +517,9 @@ class RagModel(RagPreTrainedModel):
 
         self.retriever = retriever
         if self.retriever is not None:
-            assert isinstance(
+            assert isinstance(retriever, RagRetriever), (
-                retriever, RagRetriever
+                f"`self.retriever` is of type {type(self.retriever)}, but should be of type `RagRetriever`"
-            ), f"`self.retriever` is of type {type(self.retriever)}, but should be of type `RagRetriever`"
+            )
             self.retriever = retriever
 
         self.question_encoder = question_encoder
@@ -660,9 +660,9 @@ class RagModel(RagPreTrainedModel):
                     " retriever using the `set_retriever(...)` function."
                 )
 
-        assert (
+        assert doc_scores is not None, (
-            doc_scores is not None
+            "Make sure that `doc_scores` are passed when passing `encoder_outputs` to the forward function."
-        ), "Make sure that `doc_scores` are passed when passing `encoder_outputs` to the forward function."
+        )
 
         assert (doc_scores.shape[1] % n_docs) == 0, (
             f" The first dimension of `context_input_ids` should be a multiple of `n_docs`={n_docs}, but is"
@@ -740,9 +740,9 @@ class RagSequenceForGeneration(RagPreTrainedModel):
         retriever: Optional[RagRetriever] = None,
         **kwargs,
     ):
-        assert config is not None or (
+        assert config is not None or (question_encoder is not None and generator is not None), (
-            question_encoder is not None and generator is not None
+            "Either a configuration or an encoder and a generator has to be provided."
-        ), "Either a configuration or an encoder and a generator has to be provided."
+        )
 
         if config is None:
             config = RagConfig.from_question_encoder_generator_configs(
@@ -973,9 +973,9 @@ class RagSequenceForGeneration(RagPreTrainedModel):
         )
         num_beams = num_beams if num_beams is not None else self.config.num_beams
 
-        assert (
+        assert input_ids is not None or context_input_ids is not None, (
-            input_ids is not None or context_input_ids is not None
+            " At least one of input_ids or context_input_ids must be given"
-        ), " At least one of input_ids or context_input_ids must be given"
+        )
 
         if self.retriever is not None and context_input_ids is None:
             question_hidden_states = self.question_encoder(input_ids, attention_mask=attention_mask)[0]
@@ -1138,9 +1138,9 @@ class RagTokenForGeneration(RagPreTrainedModel):
         retriever: Optional[RagRetriever] = None,
         **kwargs,
     ):
-        assert config is not None or (
+        assert config is not None or (question_encoder is not None and generator is not None), (
-            question_encoder is not None and generator is not None
+            "Either a configuration or an encoder and a generator has to be provided."
-        ), "Either a configuration or an encoder and a generator has to be provided."
+        )
 
         if config is None:
             config = RagConfig.from_question_encoder_generator_configs(

src/transformers/models/rag/modeling_tf_rag.py

@@ -506,9 +506,9 @@ class TFRagModel(TFRagPreTrainedModel):
         load_weight_prefix: Optional[str] = None,
         **kwargs,
     ):
-        assert config is not None or (
+        assert config is not None or (question_encoder is not None and generator is not None), (
-            question_encoder is not None and generator is not None
+            "Either a configuration or an question_encoder and a generator has to be provided."
-        ), "Either a configuration or an question_encoder and a generator has to be provided."
+        )
 
         if config is None:
             config = RagConfig.from_question_encoder_generator_configs(
@@ -533,9 +533,9 @@ class TFRagModel(TFRagPreTrainedModel):
 
         self.retriever = retriever
         if self.retriever is not None:
-            assert isinstance(
+            assert isinstance(retriever, RagRetriever), (
-                retriever, RagRetriever
+                f"`self.retriever` is of type {type(self.retriever)}, but should be of type `RagRetriever`"
-            ), f"`self.retriever` is of type {type(self.retriever)}, but should be of type `RagRetriever`"
+            )
             self.retriever = retriever
 
         self.question_encoder = question_encoder
@@ -589,9 +589,9 @@ class TFRagModel(TFRagPreTrainedModel):
         >>> input_ids = input_dict["input_ids"]
         >>> outputs = model(input_ids)
         ```"""
-        assert (
+        assert "decoder_cached_states" not in kwargs, (
-            "decoder_cached_states" not in kwargs
+            "Please use past_key_values to cache intermediate outputs"
-        ), "Please use past_key_values to cache intermediate outputs"  # from modeling_tf_bart.py
+        )  # from modeling_tf_bart.py
 
         # aliasing to minimize code changing
         n_docs = n_docs if n_docs is not None else self.config.n_docs
@@ -657,9 +657,9 @@ class TFRagModel(TFRagPreTrainedModel):
                     " retriever using the `set_retriever(...)` function."
                 )
 
-        assert (
+        assert doc_scores is not None, (
-            doc_scores is not None
+            "Make sure that `doc_scores` are passed when passing `encoder_outputs` to the forward function."
-        ), "Make sure that `doc_scores` are passed when passing `encoder_outputs` to the forward function."
+        )
 
         assert (doc_scores.shape[1] % n_docs) == 0, (
             f" The first dimension of `context_input_ids` should be a multiple of `n_docs`={n_docs}, but is"
@@ -747,9 +747,9 @@ class TFRagTokenForGeneration(TFRagPreTrainedModel, TFCausalLanguageModelingLoss
         retriever: Optional[RagRetriever] = None,
         **kwargs,
     ):
-        assert config is not None or (
+        assert config is not None or (question_encoder is not None and generator is not None), (
-            question_encoder is not None and generator is not None
+            "Either a configuration or an encoder and a generator has to be provided."
-        ), "Either a configuration or an encoder and a generator has to be provided."
+        )
 
         if config is None:
             config = RagConfig.from_question_encoder_generator_configs(
@@ -939,9 +939,9 @@ class TFRagTokenForGeneration(TFRagPreTrainedModel, TFCausalLanguageModelingLoss
         >>> generated_string = tokenizer.batch_decode(generated, skip_special_tokens=True)
         ```"""
 
-        assert (
+        assert "decoder_cached_states" not in kwargs, (
-            "decoder_cached_states" not in kwargs
+            "Please use past_key_values to cache intermediate outputs"
-        ), "Please use past_key_values to cache intermediate outputs"  # from modeling_tf_bart.py
+        )  # from modeling_tf_bart.py
 
         do_marginalize = do_marginalize if do_marginalize else self.config.do_marginalize
         reduce_loss = reduce_loss if reduce_loss else self.config.reduce_loss
@@ -1327,9 +1327,9 @@ class TFRagSequenceForGeneration(TFRagPreTrainedModel, TFCausalLanguageModelingL
         retriever: Optional[RagRetriever] = None,
         **kwargs,
     ):
-        assert config is not None or (
+        assert config is not None or (question_encoder is not None and generator is not None), (
-            question_encoder is not None and generator is not None
+            "Either a configuration or an encoder and a generator has to be provided."
-        ), "Either a configuration or an encoder and a generator has to be provided."
+        )
 
         if config is None:
             config = RagConfig.from_question_encoder_generator_configs(
@@ -1454,9 +1454,9 @@ class TFRagSequenceForGeneration(TFRagPreTrainedModel, TFCausalLanguageModelingL
         >>> generated_string = tokenizer.batch_decode(generated, skip_special_tokens=True)
         ```"""
 
-        assert (
+        assert "decoder_cached_states" not in kwargs, (
-            "decoder_cached_states" not in kwargs
+            "Please use past_key_values to cache intermediate outputs"
-        ), "Please use past_key_values to cache intermediate outputs"  # from modeling_tf_bart.py
+        )  # from modeling_tf_bart.py
 
         exclude_bos_score = exclude_bos_score if exclude_bos_score else self.config.exclude_bos_score
         reduce_loss = reduce_loss if reduce_loss else self.config.reduce_loss
@@ -1663,9 +1663,9 @@ class TFRagSequenceForGeneration(TFRagPreTrainedModel, TFCausalLanguageModelingL
         )
         num_beams = num_beams if num_beams is not None else self.config.num_beams
 
-        assert (
+        assert input_ids is not None or context_input_ids is not None, (
-            input_ids is not None or context_input_ids is not None
+            " At least one of input_ids or context_input_ids must be given"
-        ), " At least one of input_ids or context_input_ids must be given"
+        )
 
         if self.retriever is not None and context_input_ids is None:
             question_hidden_states = self.question_encoder(input_ids, attention_mask=attention_mask)[0]

src/transformers/models/rag/retrieval_rag.py

@@ -156,9 +156,9 @@ class LegacyIndex(Index):
             )
         with open(resolved_meta_path, "rb") as metadata_file:
             self.index_id_to_db_id = pickle.load(metadata_file)
-        assert (
+        assert len(self.index_id_to_db_id) == self.index.ntotal, (
-            len(self.index_id_to_db_id) == self.index.ntotal
+            "Deserialized index_id_to_db_id should match faiss index size"
-        ), "Deserialized index_id_to_db_id should match faiss index size"
+        )
 
     def is_initialized(self):
         return self._index_initialized

src/transformers/models/reformer/convert_reformer_trax_checkpoint_to_pytorch.py

@@ -150,15 +150,15 @@ def set_model_weights_in_torch(weights, torch_model, hidden_size):
         position_embeddings = torch_model_reformer.embeddings.position_embeddings
         for emb_idx in range(len(position_embeddings.weights)):
             emb_weights = np.asarray(weights[3][emb_idx][0])
-            assert (
+            assert position_embeddings.weights[emb_idx].shape == emb_weights.shape, (
-                position_embeddings.weights[emb_idx].shape == emb_weights.shape
+                f"{position_embeddings[emb_idx]} emb does not match"
-            ), f"{position_embeddings[emb_idx]} emb does not match"
+            )
             position_embeddings.weights[emb_idx] = nn.Parameter(torch.tensor(emb_weights))
 
     trax_layer_weights = weights[5]
-    assert len(torch_model_reformer.encoder.layers) * 4 == len(
+    assert len(torch_model_reformer.encoder.layers) * 4 == len(trax_layer_weights), (
-        trax_layer_weights
+        "HF and trax model do not have the same number of layers"
-    ), "HF and trax model do not have the same number of layers"
+    )
     for layer_idx, layer in enumerate(torch_model_reformer.encoder.layers):
         block_weights = trax_layer_weights[4 * layer_idx : 4 * (layer_idx + 1)]
         set_block_weights_in_torch(block_weights, layer, hidden_size)

src/transformers/models/reformer/modeling_reformer.py

@@ -446,12 +446,12 @@ class LSHSelfAttention(nn.Module, EfficientAttentionMixin):
         # free memory
         del hidden_states
 
-        assert (
+        assert query_key_vectors.shape[-1] == self.attention_head_size, (
-            query_key_vectors.shape[-1] == self.attention_head_size
+            f"last dim of query_key_vectors is {query_key_vectors.shape[-1]} but should be {self.attention_head_size}."
-        ), f"last dim of query_key_vectors is {query_key_vectors.shape[-1]} but should be {self.attention_head_size}."
+        )
-        assert (
+        assert value_vectors.shape[-1] == self.attention_head_size, (
-            value_vectors.shape[-1] == self.attention_head_size
+            f"last dim of value_vectors is {value_vectors.shape[-1]} but should be {self.attention_head_size}."
-        ), f"last dim of value_vectors is {value_vectors.shape[-1]} but should be {self.attention_head_size}."
+        )
 
         do_standard_self_attention = (sequence_length <= self.chunk_length) or (
             use_cache and past_buckets_states[1] is not None
@@ -470,9 +470,9 @@ class LSHSelfAttention(nn.Module, EfficientAttentionMixin):
                 # make sure buckets has correct shape for LSH attention
                 buckets = buckets.view(batch_size, self.num_attention_heads, num_hashes * sequence_length)
 
-            assert (
+            assert int(buckets.shape[-1]) == num_hashes * sequence_length, (
-                int(buckets.shape[-1]) == num_hashes * sequence_length
+                f"last dim of buckets is {buckets.shape[-1]}, but should be {num_hashes * sequence_length}"
-            ), f"last dim of buckets is {buckets.shape[-1]}, but should be {num_hashes * sequence_length}"
+            )
 
             sorted_bucket_idx, undo_sorted_bucket_idx = self._get_sorted_bucket_idx_and_undo_sorted_bucket_idx(
                 sequence_length, buckets, num_hashes
@@ -612,18 +612,18 @@ class LSHSelfAttention(nn.Module, EfficientAttentionMixin):
         # We sample a different random rotation for each round of hashing to
         # decrease the probability of hash misses.
         if isinstance(self.num_buckets, int):
-            assert (
+            assert self.num_buckets % 2 == 0, (
-                self.num_buckets % 2 == 0
+                f"There should be an even number of buckets, but `self.num_buckets`: {self.num_buckets}"
-            ), f"There should be an even number of buckets, but `self.num_buckets`: {self.num_buckets}"
+            )
             rotation_size = self.num_buckets
             num_buckets = self.num_buckets
         else:
             # Factorize the hash if self.num_buckets is a list or tuple
             rotation_size, num_buckets = 0, 1
             for bucket_factor in self.num_buckets:
-                assert (
+                assert bucket_factor % 2 == 0, (
-                    bucket_factor % 2 == 0
+                    f"The number of buckets should be even, but `num_bucket`: {bucket_factor}"
-                ), f"The number of buckets should be even, but `num_bucket`: {bucket_factor}"
+                )
                 rotation_size = rotation_size + bucket_factor
                 num_buckets = num_buckets * bucket_factor
 
@@ -1090,15 +1090,15 @@ class LocalSelfAttention(nn.Module, EfficientAttentionMixin):
         key_vectors = self._split_hidden_size_dim(key_vectors, self.num_attention_heads, self.attention_head_size)
         value_vectors = self._split_hidden_size_dim(value_vectors, self.num_attention_heads, self.attention_head_size)
 
-        assert (
+        assert query_vectors.shape[-1] == self.attention_head_size, (
-            query_vectors.shape[-1] == self.attention_head_size
+            f"last dim of query_key_vectors is {query_vectors.shape[-1]} but should be {self.attention_head_size}."
-        ), f"last dim of query_key_vectors is {query_vectors.shape[-1]} but should be {self.attention_head_size}."
+        )
-        assert (
+        assert key_vectors.shape[-1] == self.attention_head_size, (
-            key_vectors.shape[-1] == self.attention_head_size
+            f"last dim of query_key_vectors is {key_vectors.shape[-1]} but should be {self.attention_head_size}."
-        ), f"last dim of query_key_vectors is {key_vectors.shape[-1]} but should be {self.attention_head_size}."
+        )
-        assert (
+        assert value_vectors.shape[-1] == self.attention_head_size, (
-            value_vectors.shape[-1] == self.attention_head_size
+            f"last dim of query_key_vectors is {value_vectors.shape[-1]} but should be {self.attention_head_size}."
-        ), f"last dim of query_key_vectors is {value_vectors.shape[-1]} but should be {self.attention_head_size}."
+        )
 
         if self.chunk_length is None:
             assert self.num_chunks_before == 0 and self.num_chunks_after == 0, (
@@ -1976,9 +1976,9 @@ class ReformerModel(ReformerPreTrainedModel):
     def __init__(self, config):
         super().__init__(config)
         self.config = config
-        assert (
+        assert self.config.num_hidden_layers > 0, (
-            self.config.num_hidden_layers > 0
+            "`config.attn_layers` is empty. Select at least one attn layer form ['lsh', 'local']"
-        ), "`config.attn_layers` is empty. Select at least one attn layer form ['lsh', 'local']"
+        )
 
         self.embeddings = ReformerEmbeddings(config)
         self.encoder = ReformerEncoder(config)
@@ -2039,9 +2039,9 @@ class ReformerModel(ReformerPreTrainedModel):
         else:
             raise ValueError("You have to specify either input_ids or inputs_embeds")
 
-        assert (
+        assert len(input_shape) == 2, (
-            len(input_shape) == 2
+            f"`input_ids` have be of shape `[batch_size, sequence_length]`, but got shape: {input_shape}"
-        ), f"`input_ids` have be of shape `[batch_size, sequence_length]`, but got shape: {input_shape}"
+        )
 
         if past_buckets_states is not None:
             assert not self.training, "`past_buckets_states` can only be used for inference, not for training`."

src/transformers/models/seamless_m4t/modeling_seamless_m4t.py

@@ -2869,7 +2869,7 @@ class SeamlessM4TForTextToText(SeamlessM4TPreTrainedModel, GenerationMixin):
                 if tgt_lang not in self.generation_config.text_decoder_lang_to_code_id:
                     raise ValueError(
                         f"""`tgt_lang={tgt_lang}` is not supported by this model. Please specify a `tgt_lang` in
-                        {', '.join(self.generation_config.text_decoder_lang_to_code_id.keys())}"""
+                        {", ".join(self.generation_config.text_decoder_lang_to_code_id.keys())}"""
                     )
                 # tgt_lang gets priority over decoder input ids
                 text_tgt_lang_id = self.generation_config.text_decoder_lang_to_code_id.get(tgt_lang)
@@ -3140,7 +3140,7 @@ class SeamlessM4TForSpeechToText(SeamlessM4TPreTrainedModel, GenerationMixin):
                 if tgt_lang not in self.generation_config.text_decoder_lang_to_code_id:
                     raise ValueError(
                         f"""`tgt_lang={tgt_lang}` is not supported by this model. Please specify a `tgt_lang` in
-                        {', '.join(self.generation_config.text_decoder_lang_to_code_id.keys())}"""
+                        {", ".join(self.generation_config.text_decoder_lang_to_code_id.keys())}"""
                     )
                 # tgt_lang gets priority over decoder input ids
                 text_tgt_lang_id = self.generation_config.text_decoder_lang_to_code_id.get(tgt_lang)
@@ -3407,7 +3407,7 @@ class SeamlessM4TForTextToSpeech(SeamlessM4TPreTrainedModel, GenerationMixin):
                 elif tgt_lang not in lang_code_to_id:
                     raise ValueError(
                         f"""`tgt_lang={tgt_lang}` is not supported by this model.
-                    Please specify a `tgt_lang` in {','.join(lang_code_to_id.keys())}. Note that SeamlessM4T supports
+                    Please specify a `tgt_lang` in {",".join(lang_code_to_id.keys())}. Note that SeamlessM4T supports
                     more languages for text translation than for speech synthesis."""
                     )
 
@@ -3736,7 +3736,7 @@ class SeamlessM4TForSpeechToSpeech(SeamlessM4TPreTrainedModel, GenerationMixin):
                 elif tgt_lang not in lang_code_to_id:
                     raise ValueError(
                         f"""`tgt_lang={tgt_lang}` is not supported by this model.
-                    Please specify a `tgt_lang` in {','.join(lang_code_to_id.keys())}. Note that SeamlessM4T supports
+                    Please specify a `tgt_lang` in {",".join(lang_code_to_id.keys())}. Note that SeamlessM4T supports
                     more languages for text translation than for speech synthesis."""
                     )
 
@@ -4151,7 +4151,7 @@ class SeamlessM4TModel(SeamlessM4TPreTrainedModel, GenerationMixin):
                 elif tgt_lang not in lang_code_to_id:
                     raise ValueError(
                         f"""`tgt_lang={tgt_lang}` is not supported by this model.
-                    Please specify a `tgt_lang` in {','.join(lang_code_to_id.keys())}. Note that SeamlessM4T supports
+                    Please specify a `tgt_lang` in {",".join(lang_code_to_id.keys())}. Note that SeamlessM4T supports
                     more languages for text translation than for speech synthesis."""
                     )
 

src/transformers/models/seamless_m4t_v2/modeling_seamless_m4t_v2.py

@@ -3149,7 +3149,7 @@ class SeamlessM4Tv2ForTextToText(SeamlessM4Tv2PreTrainedModel, GenerationMixin):
                 if tgt_lang not in self.generation_config.text_decoder_lang_to_code_id:
                     raise ValueError(
                         f"""`tgt_lang={tgt_lang}` is not supported by this model. Please specify a `tgt_lang` in
-                        {', '.join(self.generation_config.text_decoder_lang_to_code_id.keys())}"""
+                        {", ".join(self.generation_config.text_decoder_lang_to_code_id.keys())}"""
                     )
                 # tgt_lang gets priority over decoder input ids
                 text_tgt_lang_id = self.generation_config.text_decoder_lang_to_code_id.get(tgt_lang)
@@ -3430,7 +3430,7 @@ class SeamlessM4Tv2ForSpeechToText(SeamlessM4Tv2PreTrainedModel, GenerationMixin
                 if tgt_lang not in self.generation_config.text_decoder_lang_to_code_id:
                     raise ValueError(
                         f"""`tgt_lang={tgt_lang}` is not supported by this model. Please specify a `tgt_lang` in
-                        {', '.join(self.generation_config.text_decoder_lang_to_code_id.keys())}"""
+                        {", ".join(self.generation_config.text_decoder_lang_to_code_id.keys())}"""
                     )
                 # tgt_lang gets priority over decoder input ids
                 text_tgt_lang_id = self.generation_config.text_decoder_lang_to_code_id.get(tgt_lang)
@@ -3707,7 +3707,7 @@ class SeamlessM4Tv2ForTextToSpeech(SeamlessM4Tv2PreTrainedModel, GenerationMixin
                 elif tgt_lang not in lang_code_to_id:
                     raise ValueError(
                         f"""`tgt_lang={tgt_lang}` is not supported by this model.
-                    Please specify a `tgt_lang` in {','.join(lang_code_to_id.keys())}. Note that SeamlessM4Tv2 supports
+                    Please specify a `tgt_lang` in {",".join(lang_code_to_id.keys())}. Note that SeamlessM4Tv2 supports
                     more languages for text translation than for speech synthesis."""
                     )
 
@@ -4078,7 +4078,7 @@ class SeamlessM4Tv2ForSpeechToSpeech(SeamlessM4Tv2PreTrainedModel, GenerationMix
                 elif tgt_lang not in lang_code_to_id:
                     raise ValueError(
                         f"""`tgt_lang={tgt_lang}` is not supported by this model.
-                    Please specify a `tgt_lang` in {','.join(lang_code_to_id.keys())}. Note that SeamlessM4Tv2 supports
+                    Please specify a `tgt_lang` in {",".join(lang_code_to_id.keys())}. Note that SeamlessM4Tv2 supports
                     more languages for text translation than for speech synthesis."""
                     )
 
@@ -4539,7 +4539,7 @@ class SeamlessM4Tv2Model(SeamlessM4Tv2PreTrainedModel, GenerationMixin):
                 elif tgt_lang not in lang_code_to_id:
                     raise ValueError(
                         f"""`tgt_lang={tgt_lang}` is not supported by this model.
-                    Please specify a `tgt_lang` in {','.join(lang_code_to_id.keys())}. Note that SeamlessM4Tv2 supports
+                    Please specify a `tgt_lang` in {",".join(lang_code_to_id.keys())}. Note that SeamlessM4Tv2 supports
                     more languages for text translation than for speech synthesis."""
                     )
 

src/transformers/models/speech_encoder_decoder/convert_mbart_wav2vec2_seq2seq_original_to_pytorch.py

@@ -192,41 +192,41 @@ def load_adapter(full_name, value, adapter, unused_weights):
         if "proj_ln" in full_name:
             # has to be layer norm
             if "bias" in name:
-                assert (
+                assert value.shape == adapter.proj_layer_norm.bias.data.shape, (
-                    value.shape == adapter.proj_layer_norm.bias.data.shape
+                    f"{full_name} has size {value.shape}, but {adapter.proj_layer_norm.bias.data.shape} was found."
-                ), f"{full_name} has size {value.shape}, but {adapter.proj_layer_norm.bias.data.shape} was found."
+                )
                 adapter.proj_layer_norm.bias.data = value
                 logger.info(f"Adapter proj layer norm bias was initialized from {full_name}.")
             if "weight" in name:
-                assert (
+                assert value.shape == adapter.proj_layer_norm.weight.data.shape, (
-                    value.shape == adapter.proj_layer_norm.weight.data.shape
+                    f"{full_name} has size {value.shape}, but {adapter.proj_layer_norm.weight.data.shape} was found."
-                ), f"{full_name} has size {value.shape}, but {adapter.proj_layer_norm.weight.data.shape} was found."
+                )
                 adapter.proj_layer_norm.weight.data = value
         else:
             # has to be projection layer
             if "bias" in name:
-                assert (
+                assert value.shape == adapter.proj.bias.data.shape, (
-                    value.shape == adapter.proj.bias.data.shape
+                    f"{full_name} has size {value.shape}, but {adapter.proj.bias.data.shape} was found."
-                ), f"{full_name} has size {value.shape}, but {adapter.proj.bias.data.shape} was found."
+                )
                 adapter.proj.bias.data = value
                 logger.info(f"Adapter proj layer bias was initialized from {full_name}.")
             if "weight" in name:
-                assert (
+                assert value.shape == adapter.proj.weight.data.shape, (
-                    value.shape == adapter.proj.weight.data.shape
+                    f"{full_name} has size {value.shape}, but {adapter.proj.weight.data.shape} was found."
-                ), f"{full_name} has size {value.shape}, but {adapter.proj.weight.data.shape} was found."
+                )
                 adapter.proj.weight.data = value
                 logger.info(f"Adapter proj layer weight was initialized from {full_name}.")
     elif isinstance(layer_id, int):
         if "bias" in name:
-            assert (
+            assert value.shape == adapter.layers[layer_id].conv.bias.data.shape, (
-                value.shape == adapter.layers[layer_id].conv.bias.data.shape
+                f"{full_name} has size {value.shape}, but {adapter.layers[layer_id].conv.bias.data.shape} was found."
-            ), f"{full_name} has size {value.shape}, but {adapter.layers[layer_id].conv.bias.data.shape} was found."
+            )
             adapter.layers[layer_id].conv.bias.data = value
             logger.info(f"Adapter layer {layer_id} bias was initialized from {full_name}.")
         elif "weight" in name:
-            assert (
+            assert value.shape == adapter.layers[layer_id].conv.weight.data.shape, (
-                value.shape == adapter.layers[layer_id].conv.weight.data.shape
+                f"{full_name} has size {value.shape}, but {adapter.layers[layer_id].conv.weight.data.shape} was found."
-            ), f"{full_name} has size {value.shape}, but {adapter.layers[layer_id].conv.weight.data.shape} was found."
+            )
             adapter.layers[layer_id].conv.weight.data = value
             logger.info(f"Adapter layer {layer_id} bias was initialized from {full_name}.")
     else:

src/transformers/models/speech_to_text/modeling_speech_to_text.py

@@ -774,9 +774,9 @@ class Speech2TextEncoder(Speech2TextPreTrainedModel):
 
         # check if head_mask has a correct number of layers specified if desired
         if head_mask is not None:
-            assert head_mask.size()[0] == (
+            assert head_mask.size()[0] == (len(self.layers)), (
-                len(self.layers)
+                f"The head_mask should be specified for {len(self.layers)} layers, but it is for {head_mask.size()[0]}."
-            ), f"The head_mask should be specified for {len(self.layers)} layers, but it is for {head_mask.size()[0]}."
+            )
 
         for idx, encoder_layer in enumerate(self.layers):
             if output_hidden_states:

src/transformers/models/swin2sr/convert_swin2sr_original_to_pytorch.py

@@ -224,9 +224,9 @@ def convert_swin2sr_checkpoint(checkpoint_url, pytorch_dump_folder_path, push_to
             [[-0.5238, -0.5557, -0.6321], [-0.6016, -0.5903, -0.6391], [-0.6244, -0.6334, -0.6889]]
         )
 
-    assert (
+    assert outputs.reconstruction.shape == expected_shape, (
-        outputs.reconstruction.shape == expected_shape
+        f"Shape of reconstruction should be {expected_shape}, but is {outputs.reconstruction.shape}"
-    ), f"Shape of reconstruction should be {expected_shape}, but is {outputs.reconstruction.shape}"
+    )
     assert torch.allclose(outputs.reconstruction[0, 0, :3, :3], expected_slice, atol=1e-3)
     print("Looks ok!")
 

src/transformers/models/switch_transformers/convert_big_switch.py

@@ -116,7 +116,9 @@ def shard_on_the_fly(switch_checkpoint_path, dump_path, max_shard_size, dtype, w
         total_size += weight_size
 
     # Add the last block
-    save_path = os.path.join(dump_path, weights_name.replace(".bin", f"-{len(sharded_state_dicts)+1:05d}-of-???.bin"))
+    save_path = os.path.join(
+        dump_path, weights_name.replace(".bin", f"-{len(sharded_state_dicts) + 1:05d}-of-???.bin")
+    )
     rename_and_save_block(current_block, save_path)
     sharded_state_dicts.append(current_block.keys())
 

src/transformers/models/t5/modeling_tf_t5.py

@@ -363,9 +363,9 @@ class TFT5Attention(keras.layers.Layer):
         real_seq_length = seq_length
 
         if past_key_value is not None:
-            assert (
+            assert len(past_key_value) == 2, (
-                len(past_key_value) == 2
+                f"past_key_value should have 2 past states: keys and values. Got {len(past_key_value)} past states"
-            ), f"past_key_value should have 2 past states: keys and values. Got {len(past_key_value)} past states"
+            )
             real_seq_length += shape_list(past_key_value[0])[2] if query_length is None else query_length
 
         key_length = real_seq_length if key_value_states is None else shape_list(key_value_states)[1]

src/transformers/models/tapas/modeling_tapas.py

@@ -1284,9 +1284,9 @@ class TapasForQuestionAnswering(TapasPreTrainedModel):
                 aggregate_mask = None
             else:
                 if float_answer is not None:
-                    assert (
+                    assert labels.shape[0] == float_answer.shape[0], (
-                        labels.shape[0] == float_answer.shape[0]
+                        "Make sure the answers are a FloatTensor of shape (batch_size,)"
-                    ), "Make sure the answers are a FloatTensor of shape (batch_size,)"
+                    )
                     # <float32>[batch_size]
                     aggregate_mask = _calculate_aggregate_mask(
                         float_answer,
@@ -1336,9 +1336,9 @@ class TapasForQuestionAnswering(TapasPreTrainedModel):
                 if is_supervised:
                     # Note that `aggregate_mask` is None if the setting is supervised.
                     if aggregation_labels is not None:
-                        assert (
+                        assert labels.shape[0] == aggregation_labels.shape[0], (
-                            labels.shape[0] == aggregation_labels.shape[0]
+                            "Make sure the aggregation labels are a LongTensor of shape (batch_size,)"
-                        ), "Make sure the aggregation labels are a LongTensor of shape (batch_size,)"
+                        )
                         per_example_additional_loss = _calculate_aggregation_loss(
                             logits_aggregation,
                             aggregate_mask,

src/transformers/models/tapas/modeling_tf_tapas.py

@@ -1562,9 +1562,9 @@ class TFTapasForQuestionAnswering(TFTapasPreTrainedModel):
                 aggregate_mask = None
             else:
                 if float_answer is not None:
-                    assert (
+                    assert shape_list(labels)[0] == shape_list(float_answer)[0], (
-                        shape_list(labels)[0] == shape_list(float_answer)[0]
+                        "Make sure the answers are a FloatTensor of shape (batch_size,)"
-                    ), "Make sure the answers are a FloatTensor of shape (batch_size,)"
+                    )
                     # <float32>[batch_size]
                     aggregate_mask = _calculate_aggregate_mask(
                         float_answer,
@@ -1615,9 +1615,9 @@ class TFTapasForQuestionAnswering(TFTapasPreTrainedModel):
                 if is_supervised:
                     # Note that `aggregate_mask` is None if the setting is supervised.
                     if aggregation_labels is not None:
-                        assert (
+                        assert shape_list(labels)[0] == shape_list(aggregation_labels)[0], (
-                            shape_list(labels)[0] == shape_list(aggregation_labels)[0]
+                            "Make sure the aggregation labels are a LongTensor of shape (batch_size,)"
-                        ), "Make sure the aggregation labels are a LongTensor of shape (batch_size,)"
+                        )
                         per_example_additional_loss = _calculate_aggregation_loss(
                             logits_aggregation,
                             aggregate_mask,

src/transformers/models/tvp/modeling_tvp.py

@@ -773,7 +773,7 @@ TVP_PROMPTER_CLASSES_MAPPING = {
 
 
 @add_start_docstrings(
-    "The bare Tvp Model transformer outputting BaseModelOutputWithPooling object without any specific head on" " top.",
+    "The bare Tvp Model transformer outputting BaseModelOutputWithPooling object without any specific head on top.",
     TVP_START_DOCSTRING,
 )
 class TvpModel(TvpPreTrainedModel):

src/transformers/models/udop/modeling_udop.py

@@ -407,8 +407,7 @@ class UdopPatchEmbeddings(nn.Module):
         batch_size, num_channels, height, width = pixel_values.shape
         if height != self.image_size[0] or width != self.image_size[1]:
             raise ValueError(
-                f"Input image size ({height}*{width}) doesn't match model"
+                f"Input image size ({height}*{width}) doesn't match model ({self.image_size[0]}*{self.image_size[1]})."
-                f" ({self.image_size[0]}*{self.image_size[1]})."
             )
         embeddings = self.proj(pixel_values)
         embeddings = embeddings.flatten(2).transpose(1, 2)

src/transformers/models/visual_bert/convert_visual_bert_original_pytorch_checkpoint_to_pytorch.py

@@ -84,9 +84,9 @@ def convert_visual_bert_checkpoint(checkpoint_path, pytorch_dump_folder_path):
     Copy/paste/tweak model's weights to our VisualBERT structure.
     """
 
-    assert (
+    assert checkpoint_path.split("/")[-1] in ACCEPTABLE_CHECKPOINTS, (
-        checkpoint_path.split("/")[-1] in ACCEPTABLE_CHECKPOINTS
+        f"The checkpoint provided must be in {ACCEPTABLE_CHECKPOINTS}."
-    ), f"The checkpoint provided must be in {ACCEPTABLE_CHECKPOINTS}."
+    )
 
     # Get Config
     if "pre" in checkpoint_path:

src/transformers/models/vivit/modeling_vivit.py

@@ -72,8 +72,7 @@ class VivitTubeletEmbeddings(nn.Module):
         batch_size, num_frames, num_channels, height, width = pixel_values.shape
         if not interpolate_pos_encoding and (height != self.image_size or width != self.image_size):
             raise ValueError(
-                f"Image image size ({height}*{width}) doesn't match model"
+                f"Image image size ({height}*{width}) doesn't match model ({self.image_size[0]}*{self.image_size[1]})."
-                f" ({self.image_size[0]}*{self.image_size[1]})."
             )
 
         # permute to (batch_size, num_channels, num_frames, height, width)

src/transformers/models/whisper/modeling_whisper.py

@@ -1043,9 +1043,9 @@ class WhisperEncoder(WhisperPreTrainedModel):
 
         # check if head_mask has a correct number of layers specified if desired
         if head_mask is not None:
-            assert head_mask.size()[0] == (
+            assert head_mask.size()[0] == (len(self.layers)), (
-                len(self.layers)
+                f"The head_mask should be specified for {len(self.layers)} layers, but it is for {head_mask.size()[0]}."
-            ), f"The head_mask should be specified for {len(self.layers)} layers, but it is for {head_mask.size()[0]}."
+            )
 
         for idx, encoder_layer in enumerate(self.layers):
             if output_hidden_states:

src/transformers/models/x_clip/modeling_x_clip.py

@@ -167,7 +167,7 @@ class XCLIPVisionEmbeddings(nn.Module):
         batch_size, _, height, width = pixel_values.shape
         if not interpolate_pos_encoding and (height != self.image_size or width != self.image_size):
             raise ValueError(
-                f"Input image size ({height}*{width}) doesn't match model" f" ({self.image_size}*{self.image_size})."
+                f"Input image size ({height}*{width}) doesn't match model ({self.image_size}*{self.image_size})."
             )
         target_dtype = self.patch_embedding.weight.dtype
         patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype))  # shape = [*, width, grid, grid]

src/transformers/models/xglm/modeling_xglm.py

@@ -601,8 +601,7 @@ class XGLMModel(XGLMPreTrainedModel):
         if self.gradient_checkpointing and self.training:
             if use_cache:
                 logger.warning_once(
-                    "`use_cache = True` is incompatible with gradient checkpointing`. Setting `use_cache ="
+                    "`use_cache = True` is incompatible with gradient checkpointing`. Setting `use_cache = False`..."
-                    " False`..."
                 )
                 use_cache = False
 

src/transformers/models/xlnet/modeling_xlnet.py

@@ -164,15 +164,15 @@ def load_tf_weights_in_xlnet(model, config, tf_path):
             array = np.transpose(array)
         if isinstance(pointer, list):
             # Here we will split the TF weights
-            assert (
+            assert len(pointer) == array.shape[0], (
-                len(pointer) == array.shape[0]
+                f"Pointer length {len(pointer)} and array length {array.shape[0]} mismatched"
-            ), f"Pointer length {len(pointer)} and array length {array.shape[0]} mismatched"
+            )
             for i, p_i in enumerate(pointer):
                 arr_i = array[i, ...]
                 try:
-                    assert (
+                    assert p_i.shape == arr_i.shape, (
-                        p_i.shape == arr_i.shape
+                        f"Pointer shape {p_i.shape} and array shape {arr_i.shape} mismatched"
-                    ), f"Pointer shape {p_i.shape} and array shape {arr_i.shape} mismatched"
+                    )
                 except AssertionError as e:
                     e.args += (p_i.shape, arr_i.shape)
                     raise
@@ -180,9 +180,9 @@ def load_tf_weights_in_xlnet(model, config, tf_path):
                 p_i.data = torch.from_numpy(arr_i)
         else:
             try:
-                assert (
+                assert pointer.shape == array.shape, (
-                    pointer.shape == array.shape
+                    f"Pointer shape {pointer.shape} and array shape {array.shape} mismatched"
-                ), f"Pointer shape {pointer.shape} and array shape {array.shape} mismatched"
+                )
             except AssertionError as e:
                 e.args += (pointer.shape, array.shape)
                 raise

src/transformers/models/zamba/configuration_zamba.py

@@ -203,9 +203,9 @@ class ZambaConfig(PretrainedConfig):
 
         self.layers_block_type = self._layers_block_type(num_hidden_layers, attn_layer_period, attn_layer_offset)
 
-        assert (
+        assert (self.mamba_expand * self.hidden_size) % self.n_mamba_heads == 0, (
-            self.mamba_expand * self.hidden_size
+            "`intermediate_size` should be divisible by `n_mamba_heads`."
-        ) % self.n_mamba_heads == 0, "`intermediate_size` should be divisible by `n_mamba_heads`."
+        )
 
         super().__init__(
             pad_token_id=pad_token_id,

src/transformers/pipelines/document_question_answering.py

@@ -339,7 +339,7 @@ class DocumentQuestionAnsweringPipeline(ChunkPipeline):
             )
 
         if self.model_type == ModelType.VisionEncoderDecoder:
-            task_prompt = f'<s_docvqa><s_question>{input["question"]}</s_question><s_answer>'
+            task_prompt = f"<s_docvqa><s_question>{input['question']}</s_question><s_answer>"
             # Adapted from https://huggingface.co/spaces/nielsr/donut-docvqa/blob/main/app.py
             encoding = {
                 "inputs": image_features["pixel_values"],

src/transformers/quantizers/quantizer_fbgemm_fp8.py

@@ -104,8 +104,7 @@ class FbgemmFp8HfQuantizer(HfQuantizer):
             )
         elif torch_dtype == torch.float16:
             raise ValueError(
-                "You cannot use FP8 with torch_dtype=torch.float16."
+                "You cannot use FP8 with torch_dtype=torch.float16.We recommend you passing torch_dtype=torch.bfloat16"
-                "We recommend you passing torch_dtype=torch.bfloat16"
             )
         return torch_dtype
 

src/transformers/quantizers/quantizer_torchao.py

@@ -257,8 +257,7 @@ class TorchAoHfQuantizer(HfQuantizer):
     def is_serializable(self, safe_serialization=None) -> bool:
         if safe_serialization:
             logger.warning(
-                "torchao quantized model does not support safe serialization, "
+                "torchao quantized model does not support safe serialization, please set `safe_serialization` to False"
-                "please set `safe_serialization` to False"
             )
             return False
         _is_torchao_serializable = version.parse(importlib.metadata.version("huggingface_hub")) >= version.parse(

src/transformers/tokenization_utils_base.py

@@ -868,7 +868,7 @@ class SpecialTokensMixin:
     def __init__(self, verbose=False, **kwargs):
         self._pad_token_type_id = 0
         self.verbose = verbose
-        self._special_tokens_map = {attr: None for attr in self.SPECIAL_TOKENS_ATTRIBUTES}
+        self._special_tokens_map = dict.fromkeys(self.SPECIAL_TOKENS_ATTRIBUTES)
         self._special_tokens_map["additional_special_tokens"] = []  # for BC where it defaults to empty list
 
         # We directly set the hidden value to allow initialization with special tokens
@@ -881,9 +881,9 @@ class SpecialTokensMixin:
             if key in self.SPECIAL_TOKENS_ATTRIBUTES:
                 if key == "additional_special_tokens":
                     assert isinstance(value, (list, tuple)), f"Value {value} is not a list or tuple"
-                    assert all(
+                    assert all(isinstance(t, (str, AddedToken)) for t in value), (
-                        isinstance(t, (str, AddedToken)) for t in value
+                        "One of the tokens is not a string or an AddedToken"
-                    ), "One of the tokens is not a string or an AddedToken"
+                    )
                     setattr(self, key, value)
                 elif isinstance(value, (str, AddedToken)):
                     setattr(self, key, value)
@@ -967,9 +967,9 @@ class SpecialTokensMixin:
                 logger.info(f"Assigning {value} to the {key} key of the tokenizer")
 
             if key == "additional_special_tokens":
-                assert isinstance(value, (list, tuple)) and all(
+                assert isinstance(value, (list, tuple)) and all(isinstance(t, (str, AddedToken)) for t in value), (
-                    isinstance(t, (str, AddedToken)) for t in value
+                    f"Tokens {value} for key {key} should all be str or AddedToken instances"
-                ), f"Tokens {value} for key {key} should all be str or AddedToken instances"
+                )
 
                 to_add = []
                 for token in value:
@@ -3379,9 +3379,9 @@ class PreTrainedTokenizerBase(SpecialTokensMixin, PushToHubMixin):
             return BatchEncoding(encoded_inputs, tensor_type=return_tensors)
 
         batch_size = len(required_input)
-        assert all(
+        assert all(len(v) == batch_size for v in encoded_inputs.values()), (
-            len(v) == batch_size for v in encoded_inputs.values()
+            "Some items in the output dictionary have a different batch size than others."
-        ), "Some items in the output dictionary have a different batch size than others."
+        )
 
         if padding_strategy == PaddingStrategy.LONGEST:
             max_length = max(len(inputs) for inputs in required_input)

src/transformers/trainer_callback.py

@@ -749,12 +749,12 @@ class EarlyStoppingCallback(TrainerCallback, ExportableState):
                 "Using EarlyStoppingCallback without load_best_model_at_end=True. "
                 "Once training is finished, the best model will not be loaded automatically."
             )
-        assert (
+        assert args.metric_for_best_model is not None, (
-            args.metric_for_best_model is not None
+            "EarlyStoppingCallback requires metric_for_best_model to be defined"
-        ), "EarlyStoppingCallback requires metric_for_best_model to be defined"
+        )
-        assert (
+        assert args.eval_strategy != IntervalStrategy.NO, (
-            args.eval_strategy != IntervalStrategy.NO
+            "EarlyStoppingCallback requires IntervalStrategy of steps or epoch"
-        ), "EarlyStoppingCallback requires IntervalStrategy of steps or epoch"
+        )
 
     def on_evaluate(self, args, state, control, metrics, **kwargs):
         metric_to_check = args.metric_for_best_model

src/transformers/trainer_pt_utils.py

@@ -121,9 +121,9 @@ def nested_concat(tensors, new_tensors, padding_index=-100):
     nested list/tuples/dict of tensors.
     """
     if not (isinstance(tensors, torch.Tensor) and isinstance(new_tensors, torch.Tensor)):
-        assert (
+        assert type(tensors) is type(new_tensors), (
-            type(tensors) is type(new_tensors)
+            f"Expected `tensors` and `new_tensors` to have the same type but found {type(tensors)} and {type(new_tensors)}."
-        ), f"Expected `tensors` and `new_tensors` to have the same type but found {type(tensors)} and {type(new_tensors)}."
+        )
     if isinstance(tensors, (list, tuple)):
         return type(tensors)(nested_concat(t, n, padding_index=padding_index) for t, n in zip(tensors, new_tensors))
     elif isinstance(tensors, torch.Tensor):
@@ -381,15 +381,15 @@ class SequentialDistributedSampler(Sampler):
 
         # add extra samples to make it evenly divisible
         indices += indices[: (self.total_size - len(indices))]
-        assert (
+        assert len(indices) == self.total_size, (
-            len(indices) == self.total_size
+            f"Indices length {len(indices)} and total size {self.total_size} mismatched"
-        ), f"Indices length {len(indices)} and total size {self.total_size} mismatched"
+        )
 
         # subsample
         indices = indices[self.rank * self.num_samples : (self.rank + 1) * self.num_samples]
-        assert (
+        assert len(indices) == self.num_samples, (
-            len(indices) == self.num_samples
+            f"Indices length {len(indices)} and sample number {self.num_samples} mismatched"
-        ), f"Indices length {len(indices)} and sample number {self.num_samples} mismatched"
+        )
 
         return iter(indices)
 
@@ -506,9 +506,9 @@ class DistributedTensorGatherer:
         if isinstance(arrays, (list, tuple)):
             result = [self._nested_set_tensors(x, y) for x, y in zip(storage, arrays)]
             return result[0][0], type(arrays)(r[1] for r in result)
-        assert (
+        assert arrays.shape[0] % self.world_size == 0, (
-            arrays.shape[0] % self.world_size == 0
+            f"Arrays passed should all have a first dimension multiple of {self.world_size}, found {arrays.shape[0]}."
-        ), f"Arrays passed should all have a first dimension multiple of {self.world_size}, found {arrays.shape[0]}."
+        )
 
         slice_len = arrays.shape[0] // self.world_size
         for i in range(self.world_size):

src/transformers/utils/chat_template_utils.py

@@ -412,7 +412,7 @@ def _compile_jinja_template(chat_template):
 
     if version.parse(jinja2.__version__) < version.parse("3.1.0"):
         raise ImportError(
-            "apply_chat_template requires jinja2>=3.1.0 to be installed. Your version is " f"{jinja2.__version__}."
+            f"apply_chat_template requires jinja2>=3.1.0 to be installed. Your version is {jinja2.__version__}."
         )
 
     def raise_exception(message):

src/transformers/utils/hub.py

@@ -513,7 +513,9 @@ def cached_files(
             return None
         # Now we raise for missing entries
         revision_ = "main" if revision is None else revision
-        msg = f"a file named {missing_entries[0]}" if len(missing_entries) == 1 else f"files named {*missing_entries,}"
+        msg = (
+            f"a file named {missing_entries[0]}" if len(missing_entries) == 1 else f"files named {(*missing_entries,)}"
+        )
         raise EnvironmentError(
             f"{path_or_repo_id} does not appear to have {msg}. Checkout 'https://huggingface.co/{path_or_repo_id}/tree/{revision_}'"
             "for available files."