Remove more unused attributes in config classes (#21543)

* Remove unused decoder_layerdrop

* Update SPECIAL_CASES_TO_ALLOW for MT5Config

* Remove unused position_embedding_init_scale

* Remove unused decoder_max_relative_position

* Use unused decoder_max_relative_position

* Remove unused init_std

* Remove unused forgotten attributes

* Remove unused patch_norm

* Remove unused max_seq_len

* Update SPECIAL_CASES_TO_ALLOW for OneFormerConfig

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

src/transformers/models/deformable_detr/configuration_deformable_detr.py

@@ -80,9 +80,6 @@ class DeformableDetrConfig(PretrainedConfig):
         encoder_layerdrop: (`float`, *optional*, defaults to 0.0):
             The LayerDrop probability for the encoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556)
             for more details.
-        decoder_layerdrop: (`float`, *optional*, defaults to 0.0):
-            The LayerDrop probability for the decoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556)
-            for more details.
         auxiliary_loss (`bool`, *optional*, defaults to `False`):
             Whether auxiliary decoding losses (loss at each decoder layer) are to be used.
         position_embedding_type (`str`, *optional*, defaults to `"sine"`):
@@ -163,7 +160,6 @@ class DeformableDetrConfig(PretrainedConfig):
         decoder_ffn_dim=1024,
         decoder_attention_heads=8,
         encoder_layerdrop=0.0,
-        decoder_layerdrop=0.0,
         is_encoder_decoder=True,
         activation_function="relu",
         d_model=256,
@@ -225,7 +221,6 @@ class DeformableDetrConfig(PretrainedConfig):
         self.init_std = init_std
         self.init_xavier_std = init_xavier_std
         self.encoder_layerdrop = encoder_layerdrop
-        self.decoder_layerdrop = decoder_layerdrop
         self.auxiliary_loss = auxiliary_loss
         self.position_embedding_type = position_embedding_type
         self.backbone = backbone

src/transformers/models/deta/configuration_deta.py

@@ -74,9 +74,6 @@ class DetaConfig(PretrainedConfig):
         encoder_layerdrop: (`float`, *optional*, defaults to 0.0):
             The LayerDrop probability for the encoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556)
             for more details.
-        decoder_layerdrop: (`float`, *optional*, defaults to 0.0):
-            The LayerDrop probability for the decoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556)
-            for more details.
         auxiliary_loss (`bool`, *optional*, defaults to `False`):
             Whether auxiliary decoding losses (loss at each decoder layer) are to be used.
         position_embedding_type (`str`, *optional*, defaults to `"sine"`):
@@ -146,7 +143,6 @@ class DetaConfig(PretrainedConfig):
         decoder_ffn_dim=1024,
         decoder_attention_heads=8,
         encoder_layerdrop=0.0,
-        decoder_layerdrop=0.0,
         is_encoder_decoder=True,
         activation_function="relu",
         d_model=256,
@@ -202,7 +198,6 @@ class DetaConfig(PretrainedConfig):
         self.init_std = init_std
         self.init_xavier_std = init_xavier_std
         self.encoder_layerdrop = encoder_layerdrop
-        self.decoder_layerdrop = decoder_layerdrop
         self.auxiliary_loss = auxiliary_loss
         self.position_embedding_type = position_embedding_type
         # deformable attributes

src/transformers/models/dinat/configuration_dinat.py

@@ -64,8 +64,6 @@ class DinatConfig(PretrainedConfig):
         hidden_act (`str` or `function`, *optional*, defaults to `"gelu"`):
             The non-linear activation function (function or string) in the encoder. If string, `"gelu"`, `"relu"`,
             `"selu"` and `"gelu_new"` are supported.
-        patch_norm (`bool`, *optional*, defaults to `True`):
-            Whether or not to add layer normalization after patch embedding.
         initializer_range (`float`, *optional*, defaults to 0.02):
             The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
         layer_norm_eps (`float`, *optional*, defaults to 1e-12):
@@ -112,7 +110,6 @@ class DinatConfig(PretrainedConfig):
         attention_probs_dropout_prob=0.0,
         drop_path_rate=0.1,
         hidden_act="gelu",
-        patch_norm=True,
         initializer_range=0.02,
         layer_norm_eps=1e-5,
         layer_scale_init_value=0.0,
@@ -135,7 +132,6 @@ class DinatConfig(PretrainedConfig):
         self.attention_probs_dropout_prob = attention_probs_dropout_prob
         self.drop_path_rate = drop_path_rate
         self.hidden_act = hidden_act
-        self.path_norm = patch_norm
         self.layer_norm_eps = layer_norm_eps
         self.initializer_range = initializer_range
         # we set the hidden_size attribute in order to make Dinat work with VisionEncoderDecoderModel

src/transformers/models/donut/configuration_donut_swin.py

@@ -66,8 +66,6 @@ class DonutSwinConfig(PretrainedConfig):
             `"selu"` and `"gelu_new"` are supported.
         use_absolute_embeddings (`bool`, *optional*, defaults to False):
             Whether or not to add absolute position embeddings to the patch embeddings.
-        patch_norm (`bool`, *optional*, defaults to True):
-            Whether or not to add layer normalization after patch embedding.
         initializer_range (`float`, *optional*, defaults to 0.02):
             The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
         layer_norm_eps (`float`, *optional*, defaults to 1e-12):
@@ -110,7 +108,6 @@ class DonutSwinConfig(PretrainedConfig):
         drop_path_rate=0.1,
         hidden_act="gelu",
         use_absolute_embeddings=False,
-        patch_norm=True,
         initializer_range=0.02,
         layer_norm_eps=1e-5,
         **kwargs,
@@ -132,7 +129,6 @@ class DonutSwinConfig(PretrainedConfig):
         self.drop_path_rate = drop_path_rate
         self.hidden_act = hidden_act
         self.use_absolute_embeddings = use_absolute_embeddings
-        self.path_norm = patch_norm
         self.layer_norm_eps = layer_norm_eps
         self.initializer_range = initializer_range
         # we set the hidden_size attribute in order to make Swin work with VisionEncoderDecoderModel

src/transformers/models/jukebox/configuration_jukebox.py

@@ -539,8 +539,6 @@ class JukeboxConfig(PretrainedConfig):
         metadata_conditioning (`bool`, *optional*, defaults to `True`):
             Whether or not to use metadata conditioning, corresponding to the artist, the genre and the min/maximum
             duration.
-        init_std (`float`, *optional*, defaults to 0.2):
-            Standard deviation used to initial the model.
 
     Example:
 
@@ -572,7 +570,6 @@ class JukeboxConfig(PretrainedConfig):
         max_duration=600.0,
         max_nb_genres=5,
         metadata_conditioning=True,
-        init_std=0.2,
         **kwargs,
     ):
         if vqvae_config is None:
@@ -596,7 +593,6 @@ class JukeboxConfig(PretrainedConfig):
 
         self.hop_fraction = self.vqvae_config.hop_fraction
 
-        self.init_std = init_std
         self.nb_priors = nb_priors
 
         # Metadata conditioning

src/transformers/models/maskformer/configuration_maskformer_swin.py

@@ -62,8 +62,6 @@ class MaskFormerSwinConfig(PretrainedConfig):
             `"selu"` and `"gelu_new"` are supported.
         use_absolute_embeddings (`bool`, *optional*, defaults to False):
             Whether or not to add absolute position embeddings to the patch embeddings.
-        patch_norm (`bool`, *optional*, defaults to True):
-            Whether or not to add layer normalization after patch embedding.
         initializer_range (`float`, *optional*, defaults to 0.02):
             The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
         layer_norm_eps (`float`, *optional*, defaults to 1e-12):
@@ -109,7 +107,6 @@ class MaskFormerSwinConfig(PretrainedConfig):
         drop_path_rate=0.1,
         hidden_act="gelu",
         use_absolute_embeddings=False,
-        patch_norm=True,
         initializer_range=0.02,
         layer_norm_eps=1e-5,
         out_features=None,
@@ -132,7 +129,6 @@ class MaskFormerSwinConfig(PretrainedConfig):
         self.drop_path_rate = drop_path_rate
         self.hidden_act = hidden_act
         self.use_absolute_embeddings = use_absolute_embeddings
-        self.path_norm = patch_norm
         self.layer_norm_eps = layer_norm_eps
         self.initializer_range = initializer_range
         # we set the hidden_size attribute in order to make Swin work with VisionEncoderDecoderModel

src/transformers/models/nat/configuration_nat.py

@@ -62,8 +62,6 @@ class NatConfig(PretrainedConfig):
         hidden_act (`str` or `function`, *optional*, defaults to `"gelu"`):
             The non-linear activation function (function or string) in the encoder. If string, `"gelu"`, `"relu"`,
             `"selu"` and `"gelu_new"` are supported.
-        patch_norm (`bool`, *optional*, defaults to `True`):
-            Whether or not to add layer normalization after patch embedding.
         initializer_range (`float`, *optional*, defaults to 0.02):
             The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
         layer_norm_eps (`float`, *optional*, defaults to 1e-12):
@@ -109,7 +107,6 @@ class NatConfig(PretrainedConfig):
         attention_probs_dropout_prob=0.0,
         drop_path_rate=0.1,
         hidden_act="gelu",
-        patch_norm=True,
         initializer_range=0.02,
         layer_norm_eps=1e-5,
         layer_scale_init_value=0.0,
@@ -131,7 +128,6 @@ class NatConfig(PretrainedConfig):
         self.attention_probs_dropout_prob = attention_probs_dropout_prob
         self.drop_path_rate = drop_path_rate
         self.hidden_act = hidden_act
-        self.path_norm = patch_norm
         self.layer_norm_eps = layer_norm_eps
         self.initializer_range = initializer_range
         # we set the hidden_size attribute in order to make Nat work with VisionEncoderDecoderModel

src/transformers/models/oneformer/configuration_oneformer.py

@@ -83,8 +83,6 @@ class OneFormerConfig(PretrainedConfig):
             List containing the strides for feature maps in the encoder.
         task_seq_len (`int`, *optional*, defaults to 77)
             Sequence length for tokenizing text list input.
-        max_seq_len (`int`, *optional*, defaults to 77)
-            Sequence length for tokenizing task input.
         text_encoder_width (`int`, *optional*, defaults to 256)
             Hidden size for text encoder.
         text_encoder_context_length (`int`, *optional*, defaults to 77):
@@ -165,7 +163,6 @@ class OneFormerConfig(PretrainedConfig):
         output_auxiliary_logits: bool = True,
         strides: Optional[list] = [4, 8, 16, 32],
         task_seq_len: int = 77,
-        max_seq_len: int = 77,
         text_encoder_width: int = 256,
         text_encoder_context_length: int = 77,
         text_encoder_num_layers: int = 6,
@@ -229,7 +226,6 @@ class OneFormerConfig(PretrainedConfig):
         self.output_auxiliary_logits = output_auxiliary_logits
         self.strides = strides
         self.task_seq_len = task_seq_len
-        self.max_seq_len = max_seq_len
         self.text_encoder_width = text_encoder_width
         self.text_encoder_context_length = text_encoder_context_length
         self.text_encoder_num_layers = text_encoder_num_layers

src/transformers/models/perceiver/configuration_perceiver.py

@@ -133,7 +133,6 @@ class PerceiverConfig(PretrainedConfig):
         cross_attention_widening_factor=1,
         hidden_act="gelu",
         attention_probs_dropout_prob=0.1,
-        position_embedding_init_scale=0.02,
         initializer_range=0.02,
         layer_norm_eps=1e-12,
         use_query_residual=True,

src/transformers/models/speecht5/configuration_speecht5.py

@@ -168,8 +168,6 @@ class SpeechT5Config(PretrainedConfig):
             The maximum sequence length of text features that this model might ever be used with.
         encoder_max_relative_position (`int`, *optional*, defaults to 160):
             Maximum distance for relative position embedding in the encoder.
-        decoder_max_relative_position (`int`, *optional*, defaults to 160):
-            Maximum distance for relative position embedding in the dencoder.
         use_cache (`bool`, *optional*, defaults to `True`):
             Whether or not the model should return the last key/values attentions (not used by all models).
 
@@ -243,7 +241,6 @@ class SpeechT5Config(PretrainedConfig):
         max_speech_positions=4000,
         max_text_positions=450,
         encoder_max_relative_position=160,
-        decoder_max_relative_position=160,
         use_cache=True,
         is_encoder_decoder=True,
         **kwargs,
@@ -314,7 +311,6 @@ class SpeechT5Config(PretrainedConfig):
         self.max_speech_positions = max_speech_positions
         self.max_text_positions = max_text_positions
         self.encoder_max_relative_position = encoder_max_relative_position
-        self.decoder_max_relative_position = decoder_max_relative_position
         self.use_cache = use_cache
         self.is_encoder_decoder = is_encoder_decoder
 

src/transformers/models/swin/configuration_swin.py

@@ -75,8 +75,6 @@ class SwinConfig(PretrainedConfig):
             `"selu"` and `"gelu_new"` are supported.
         use_absolute_embeddings (`bool`, *optional*, defaults to False):
             Whether or not to add absolute position embeddings to the patch embeddings.
-        patch_norm (`bool`, *optional*, defaults to True):
-            Whether or not to add layer normalization after patch embedding.
         initializer_range (`float`, *optional*, defaults to 0.02):
             The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
         layer_norm_eps (`float`, *optional*, defaults to 1e-12):
@@ -124,7 +122,6 @@ class SwinConfig(PretrainedConfig):
         drop_path_rate=0.1,
         hidden_act="gelu",
         use_absolute_embeddings=False,
-        patch_norm=True,
         initializer_range=0.02,
         layer_norm_eps=1e-5,
         encoder_stride=32,
@@ -148,7 +145,6 @@ class SwinConfig(PretrainedConfig):
         self.drop_path_rate = drop_path_rate
         self.hidden_act = hidden_act
         self.use_absolute_embeddings = use_absolute_embeddings
-        self.path_norm = patch_norm
         self.layer_norm_eps = layer_norm_eps
         self.initializer_range = initializer_range
         self.encoder_stride = encoder_stride

src/transformers/models/swin2sr/configuration_swin2sr.py

@@ -67,8 +67,6 @@ class Swin2SRConfig(PretrainedConfig):
             `"selu"` and `"gelu_new"` are supported.
         use_absolute_embeddings (`bool`, *optional*, defaults to `False`):
             Whether or not to add absolute position embeddings to the patch embeddings.
-        patch_norm (`bool`, *optional*, defaults to `True`):
-            Whether or not to add layer normalization after patch embedding.
         initializer_range (`float`, *optional*, defaults to 0.02):
             The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
         layer_norm_eps (`float`, *optional*, defaults to 1e-12):
@@ -121,7 +119,6 @@ class Swin2SRConfig(PretrainedConfig):
         drop_path_rate=0.1,
         hidden_act="gelu",
         use_absolute_embeddings=False,
-        patch_norm=True,
         initializer_range=0.02,
         layer_norm_eps=1e-5,
         upscale=2,
@@ -147,7 +144,6 @@ class Swin2SRConfig(PretrainedConfig):
         self.drop_path_rate = drop_path_rate
         self.hidden_act = hidden_act
         self.use_absolute_embeddings = use_absolute_embeddings
-        self.path_norm = patch_norm
         self.layer_norm_eps = layer_norm_eps
         self.initializer_range = initializer_range
         self.upscale = upscale

src/transformers/models/swinv2/configuration_swinv2.py

@@ -68,8 +68,6 @@ class Swinv2Config(PretrainedConfig):
             `"selu"` and `"gelu_new"` are supported.
         use_absolute_embeddings (`bool`, *optional*, defaults to `False`):
             Whether or not to add absolute position embeddings to the patch embeddings.
-        patch_norm (`bool`, *optional*, defaults to `True`):
-            Whether or not to add layer normalization after patch embedding.
         initializer_range (`float`, *optional*, defaults to 0.02):
             The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
         layer_norm_eps (`float`, *optional*, defaults to 1e-12):
@@ -114,7 +112,6 @@ class Swinv2Config(PretrainedConfig):
         drop_path_rate=0.1,
         hidden_act="gelu",
         use_absolute_embeddings=False,
-        patch_norm=True,
         initializer_range=0.02,
         layer_norm_eps=1e-5,
         encoder_stride=32,
@@ -137,7 +134,6 @@ class Swinv2Config(PretrainedConfig):
         self.drop_path_rate = drop_path_rate
         self.hidden_act = hidden_act
         self.use_absolute_embeddings = use_absolute_embeddings
-        self.path_norm = patch_norm
         self.layer_norm_eps = layer_norm_eps
         self.initializer_range = initializer_range
         self.encoder_stride = encoder_stride

src/transformers/models/van/modeling_van.py

@@ -374,7 +374,7 @@ class VanPreTrainedModel(PreTrainedModel):
     def _init_weights(self, module):
         """Initialize the weights"""
         if isinstance(module, nn.Linear):
-            nn.init.trunc_normal_(module.weight, std=0.02)
+            nn.init.trunc_normal_(module.weight, std=self.config.initializer_range)
             if isinstance(module, nn.Linear) and module.bias is not None:
                 nn.init.constant_(module.bias, 0)
         elif isinstance(module, nn.LayerNorm):

utils/check_config_attributes.py

@@ -45,13 +45,17 @@ SPECIAL_CASES_TO_ALLOW = {
     "EsmConfig": ["is_folding_model"],
     # used during training (despite we don't have training script for these models yet)
     "Mask2FormerConfig": ["ignore_value"],
-    # used during training (despite we don't have training script for these models yet)
-    "OneFormerConfig": ["ignore_value"],
+    # `ignore_value` used during training (despite we don't have training script for these models yet)
+    # `norm` used in conversion script (despite not using in the modeling file)
+    "OneFormerConfig": ["ignore_value", "norm"],
     # used during preprocessing and collation, see `collating_graphormer.py`
     "GraphormerConfig": ["spatial_pos_max"],
     # used internally in the configuration class file
     "T5Config": ["feed_forward_proj"],
     # used internally in the configuration class file
+    # `tokenizer_class` get default value `T5Tokenizer` intentionally
+    "MT5Config": ["feed_forward_proj", "tokenizer_class"],
+    # used internally in the configuration class file
     "LongT5Config": ["feed_forward_proj"],
     # used internally in the configuration class file
     "SwitchTransformersConfig": ["feed_forward_proj"],