transformers 5.0.0rc0__py3-none-any.whl → 5.0.0rc2__py3-none-any.whl

transformers/models/persimmon/modeling_persimmon.py

@@ -46,6 +46,7 @@ from ...modeling_rope_utils import (
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import Unpack
 from ...utils import auto_docstring, can_return_tuple, is_torch_flex_attn_available, logging
+from ...utils.generic import maybe_autocast
 from .configuration_persimmon import PersimmonConfig
 
 
@@ -76,7 +77,7 @@ class PersimmonRotaryEmbedding(nn.Module):
         inv_freq, self.attention_scaling = rope_init_fn(self.config, device)
 
         self.register_buffer("inv_freq", inv_freq, persistent=False)
-        self.original_inv_freq = inv_freq
+        self.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False)
 
     @staticmethod
     # Ignore copy
@@ -118,7 +119,7 @@ class PersimmonRotaryEmbedding(nn.Module):
         position_ids_expanded = position_ids[:, None, :].float()
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
-        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
+        with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
             freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling

transformers/models/phi/modeling_phi.py

@@ -13,7 +13,7 @@ import torch.nn as nn
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache
 from ...generation import GenerationMixin
-from ...integrations import use_kernel_func_from_hub
+from ...integrations import use_kernel_func_from_hub, use_kernelized_func
 from ...masking_utils import create_causal_mask
 from ...modeling_layers import (
     GenericForSequenceClassification,
@@ -25,7 +25,7 @@ from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import Unpack
 from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, logging
-from ...utils.generic import check_model_inputs
+from ...utils.generic import check_model_inputs, maybe_autocast
 from .configuration_phi import PhiConfig
 
 
@@ -49,7 +49,7 @@ class PhiRotaryEmbedding(nn.Module):
         inv_freq, self.attention_scaling = rope_init_fn(self.config, device)
 
         self.register_buffer("inv_freq", inv_freq, persistent=False)
-        self.original_inv_freq = inv_freq
+        self.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False)
 
     @staticmethod
     def compute_default_rope_parameters(
@@ -90,7 +90,7 @@ class PhiRotaryEmbedding(nn.Module):
         position_ids_expanded = position_ids[:, None, :].float()
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
-        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
+        with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
             freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
@@ -172,6 +172,7 @@ def eager_attention_forward(
     return attn_output, attn_weights
 
 
+@use_kernelized_func(apply_rotary_pos_emb)
 class PhiAttention(nn.Module):
     """Multi-headed attention from 'Attention Is All You Need' paper"""
 
@@ -187,7 +188,6 @@ class PhiAttention(nn.Module):
         self.q_proj = nn.Linear(config.hidden_size, config.num_attention_heads * self.head_dim, bias=True)
         self.k_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * self.head_dim, bias=True)
         self.v_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * self.head_dim, bias=True)
-        self.rotary_fn = apply_rotary_pos_emb
         self.dense = nn.Linear(config.num_attention_heads * self.head_dim, config.hidden_size, bias=True)
         self.rotary_ndims = int(self.head_dim * config.rope_parameters["partial_rotary_factor"])
         self.qk_layernorm = config.qk_layernorm
@@ -206,7 +206,6 @@ class PhiAttention(nn.Module):
         attention_mask: Optional[torch.Tensor],
         past_key_values: Optional[Cache] = None,
         cache_position: Optional[torch.LongTensor] = None,
-        position_ids: Optional[torch.LongTensor] = None,
         **kwargs,
     ) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
         input_shape = hidden_states.shape[:-1]

transformers/models/phi/modular_phi.py

@@ -92,7 +92,6 @@ class PhiAttention(LlamaAttention):
         attention_mask: Optional[torch.Tensor],
         past_key_values: Optional[Cache] = None,
         cache_position: Optional[torch.LongTensor] = None,
-        position_ids: Optional[torch.LongTensor] = None,
         **kwargs,
     ) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
         input_shape = hidden_states.shape[:-1]

transformers/models/phi3/modeling_phi3.py

@@ -44,6 +44,7 @@ from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import Unpack
 from ...utils import TransformersKwargs, auto_docstring, can_return_tuple
+from ...utils.generic import maybe_autocast
 from .configuration_phi3 import Phi3Config
 
 
@@ -82,7 +83,7 @@ class Phi3RotaryEmbedding(nn.Module):
         inv_freq, self.attention_scaling = rope_init_fn(self.config, device)
 
         self.register_buffer("inv_freq", inv_freq, persistent=False)
-        self.original_inv_freq = inv_freq
+        self.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False)
 
     @staticmethod
     def compute_default_rope_parameters(
@@ -123,7 +124,7 @@ class Phi3RotaryEmbedding(nn.Module):
         position_ids_expanded = position_ids[:, None, :].float()
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
-        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
+        with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
             freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling

transformers/models/phi4_multimodal/modeling_phi4_multimodal.py

@@ -47,7 +47,7 @@ from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import Unpack
 from ...utils import auto_docstring, can_return_tuple, torch_int
-from ...utils.generic import TransformersKwargs, check_model_inputs
+from ...utils.generic import TransformersKwargs, check_model_inputs, maybe_autocast
 from .configuration_phi4_multimodal import Phi4MultimodalAudioConfig, Phi4MultimodalConfig, Phi4MultimodalVisionConfig
 
 
@@ -602,7 +602,7 @@ class Phi4MultimodalImageEmbedding(nn.Module):
 
         # Temporarily disable autocast to avoid issue on bf16 tensors
         # Ref: https://github.com/pytorch/pytorch/issues/132715
-        with torch.autocast(device_type=inputs_embeds.device.type, enabled=False):
+        with maybe_autocast(device_type=inputs_embeds.device.type, enabled=False):
             image_embeds = inputs_embeds.index_put(
                 indices=positions_tuple, values=merged_img_set_tensor, accumulate=False
             )
@@ -881,6 +881,9 @@ class Phi4MultimodalAudioPreTrainedModel(PreTrainedModel):
         if isinstance(module, Phi4MultimodalAudioGluPointWiseConv):
             init.zeros_(module.b1)
             init.zeros_(module.b2)
+        elif isinstance(module, Phi4MultimodalAudioMeanVarianceNormLayer):
+            init.zeros_(module.global_mean)
+            init.ones_(module.global_invstd)
 
 
 def unfold_tensor(tensor, max_seq_len):
@@ -1014,7 +1017,7 @@ class Phi4MultimodalAudioModel(Phi4MultimodalAudioPreTrainedModel):
         pad_mask = pad_mask & enc_streaming_mask
         return pad_mask
 
-    def forward(self, hidden_states: torch.Tensor, mask: Optional[torch.Tensor]):
+    def forward(self, hidden_states: torch.Tensor, mask: Optional[torch.Tensor], **kwargs):
         hidden_states = self.encoder_embedding(hidden_states)
         hidden_states, hs_mask, mask = self.forward_embeddings(hidden_states, mask)
 
@@ -1116,7 +1119,7 @@ class Phi4MultimodalAudioEmbedding(nn.Module):
         merged_audio_embeds = merged_audio_embeds.to(dtype=inputs_embeds.dtype, device=inputs_embeds.device)
         # Temporarily disable autocast to avoid issue on bf16 tensors
         # Ref: https://github.com/pytorch/pytorch/issues/132715
-        with torch.autocast(device_type=inputs_embeds.device.type, enabled=False):
+        with maybe_autocast(device_type=inputs_embeds.device.type, enabled=False):
             audio_embeds = inputs_embeds.index_put(
                 indices=positions_tuple, values=merged_audio_embeds, accumulate=False
             )
@@ -1459,7 +1462,7 @@ class Phi4MultimodalRotaryEmbedding(nn.Module):
         inv_freq, self.attention_scaling = rope_init_fn(self.config, device)
 
         self.register_buffer("inv_freq", inv_freq, persistent=False)
-        self.original_inv_freq = inv_freq
+        self.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False)
 
     @staticmethod
     def compute_default_rope_parameters(
@@ -1500,7 +1503,7 @@ class Phi4MultimodalRotaryEmbedding(nn.Module):
         position_ids_expanded = position_ids[:, None, :].float()
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
-        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
+        with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
             freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling

transformers/models/phi4_multimodal/modular_phi4_multimodal.py

@@ -37,7 +37,7 @@ from ...modeling_rope_utils import RopeParameters
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import Unpack
 from ...utils import auto_docstring, logging
-from ...utils.generic import TransformersKwargs, check_model_inputs
+from ...utils.generic import TransformersKwargs, check_model_inputs, maybe_autocast
 from ..phi3.configuration_phi3 import Phi3Config
 from ..phi3.modeling_phi3 import (
     Phi3DecoderLayer,
@@ -844,7 +844,7 @@ class Phi4MultimodalImageEmbedding(nn.Module):
 
         # Temporarily disable autocast to avoid issue on bf16 tensors
         # Ref: https://github.com/pytorch/pytorch/issues/132715
-        with torch.autocast(device_type=inputs_embeds.device.type, enabled=False):
+        with maybe_autocast(device_type=inputs_embeds.device.type, enabled=False):
             image_embeds = inputs_embeds.index_put(
                 indices=positions_tuple, values=merged_img_set_tensor, accumulate=False
             )
@@ -1123,6 +1123,9 @@ class Phi4MultimodalAudioPreTrainedModel(PreTrainedModel):
         if isinstance(module, Phi4MultimodalAudioGluPointWiseConv):
             init.zeros_(module.b1)
             init.zeros_(module.b2)
+        elif isinstance(module, Phi4MultimodalAudioMeanVarianceNormLayer):
+            init.zeros_(module.global_mean)
+            init.ones_(module.global_invstd)
 
 
 class Phi4MultimodalAudioModel(Phi4MultimodalAudioPreTrainedModel):
@@ -1205,7 +1208,7 @@ class Phi4MultimodalAudioModel(Phi4MultimodalAudioPreTrainedModel):
         pad_mask = pad_mask & enc_streaming_mask
         return pad_mask
 
-    def forward(self, hidden_states: torch.Tensor, mask: Optional[torch.Tensor]):
+    def forward(self, hidden_states: torch.Tensor, mask: Optional[torch.Tensor], **kwargs):
         hidden_states = self.encoder_embedding(hidden_states)
         hidden_states, hs_mask, mask = self.forward_embeddings(hidden_states, mask)
 
@@ -1358,7 +1361,7 @@ class Phi4MultimodalAudioEmbedding(nn.Module):
         merged_audio_embeds = merged_audio_embeds.to(dtype=inputs_embeds.dtype, device=inputs_embeds.device)
         # Temporarily disable autocast to avoid issue on bf16 tensors
         # Ref: https://github.com/pytorch/pytorch/issues/132715
-        with torch.autocast(device_type=inputs_embeds.device.type, enabled=False):
+        with maybe_autocast(device_type=inputs_embeds.device.type, enabled=False):
             audio_embeds = inputs_embeds.index_put(
                 indices=positions_tuple, values=merged_audio_embeds, accumulate=False
             )

transformers/models/phi4_multimodal/processing_phi4_multimodal.py

@@ -58,8 +58,6 @@ class Phi4MultimodalProcessor(ProcessorMixin):
             The fake audio token pattern.
     """
 
-    audio_processor_class = "Phi4MultimodalFeatureExtractor"
-
     def __init__(
         self,
         image_processor,

transformers/models/phimoe/modeling_phimoe.py

@@ -30,15 +30,20 @@ from ... import initialization as init
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache
 from ...generation import GenerationMixin
-from ...integrations import use_kernel_forward_from_hub, use_kernel_func_from_hub
+from ...integrations import (
+    use_experts_implementation,
+    use_kernel_forward_from_hub,
+    use_kernel_func_from_hub,
+    use_kernelized_func,
+)
 from ...masking_utils import create_causal_mask, create_sliding_window_causal_mask
 from ...modeling_layers import GenericForSequenceClassification, GradientCheckpointingLayer
 from ...modeling_outputs import MoeCausalLMOutputWithPast, MoeModelOutputWithPast
 from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import Unpack
-from ...utils import TransformersKwargs, auto_docstring, can_return_tuple
-from ...utils.generic import OutputRecorder, check_model_inputs
+from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, is_grouped_mm_available
+from ...utils.generic import OutputRecorder, check_model_inputs, maybe_autocast
 from .configuration_phimoe import PhimoeConfig
 
 
@@ -59,7 +64,7 @@ class PhimoeRotaryEmbedding(nn.Module):
         inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
 
         self.register_buffer("inv_freq", inv_freq, persistent=False)
-        self.original_inv_freq = inv_freq
+        self.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False)
 
     @staticmethod
     def compute_default_rope_parameters(
@@ -113,7 +118,7 @@ class PhimoeRotaryEmbedding(nn.Module):
         position_ids_expanded = position_ids[:, None, :].float()
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
-        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
+        with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
             freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * mscale
@@ -194,6 +199,7 @@ def eager_attention_forward(
     return attn_output, attn_weights
 
 
+@use_kernelized_func(apply_rotary_pos_emb)
 class PhimoeAttention(nn.Module):
     """Multi-headed attention from 'Attention Is All You Need' paper"""
 
@@ -219,7 +225,6 @@ class PhimoeAttention(nn.Module):
         self.o_proj = nn.Linear(
             config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
         )
-        self.rotary_fn = apply_rotary_pos_emb
 
     def forward(
         self,
@@ -327,6 +332,7 @@ class PhimoeMultiplier(torch.autograd.Function):
         )
 
 
+@use_experts_implementation
 class PhimoeExperts(nn.Module):
     """Collection of expert weights stored as 3D tensors."""
 
@@ -617,7 +623,9 @@ class PhimoePreTrainedModel(PreTrainedModel):
     _supports_flash_attn = True
     _supports_sdpa = True
     _supports_flex_attn = True
-    _can_compile_fullgraph = False  # MoE models don't work with torch.compile (`torch.where(condition)` not supported)
+    _can_compile_fullgraph = (
+        is_grouped_mm_available()
+    )  # https://huggingface.co/docs/transformers/experts_interface#torchcompile
     _supports_attention_backend = True
     _can_record_outputs = {
         "router_logits": OutputRecorder(PhimoeTopKRouter, layer_name="mlp.router", index=0),

transformers/models/phimoe/modular_phimoe.py

@@ -24,7 +24,7 @@ from ...modeling_layers import (
     GenericForSequenceClassification,
 )
 from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS
-from ...utils.generic import OutputRecorder
+from ...utils.generic import OutputRecorder, maybe_autocast
 from ..llama.modeling_llama import LlamaAttention
 from ..mixtral.modeling_mixtral import (
     MixtralDecoderLayer,
@@ -52,7 +52,7 @@ class PhimoeRotaryEmbedding(MixtralRotaryEmbedding):
         inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
 
         self.register_buffer("inv_freq", inv_freq, persistent=False)
-        self.original_inv_freq = inv_freq
+        self.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False)
 
     def forward(self, x, position_ids=None, layer_type=None):
         if layer_type is not None:
@@ -74,7 +74,7 @@ class PhimoeRotaryEmbedding(MixtralRotaryEmbedding):
         position_ids_expanded = position_ids[:, None, :].float()
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
-        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
+        with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
             freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * mscale

transformers/models/pix2struct/modeling_pix2struct.py

@@ -481,6 +481,7 @@ class Pix2StructVisionModel(Pix2StructPreTrainedModel):
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        **kwargs,
     ) -> Union[tuple, BaseModelOutputWithPooling]:
         r"""
         flattened_patches (`torch.FloatTensor` of shape `(batch_size, sequence_length, num_channels x patch_height x patch_width)`):
@@ -1359,6 +1360,7 @@ class Pix2StructForConditionalGeneration(Pix2StructPreTrainedModel, GenerationMi
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         cache_position: Optional[torch.LongTensor] = None,
+        **kwargs,
     ) -> Union[tuple[torch.FloatTensor], Seq2SeqModelOutput]:
         r"""
         flattened_patches (`torch.FloatTensor` of shape `(batch_size, seq_length, hidden_size)`):

transformers/models/pix2struct/processing_pix2struct.py

@@ -61,10 +61,6 @@ class Pix2StructProcessor(ProcessorMixin):
             An instance of ['T5Tokenizer`]. The tokenizer is a required input.
     """
 
-    attributes = ["image_processor", "tokenizer"]
-    image_processor_class = "Pix2StructImageProcessor"
-    tokenizer_class = ("T5Tokenizer",)
-
     def __init__(self, image_processor, tokenizer):
         tokenizer.return_token_type_ids = False
         super().__init__(image_processor, tokenizer)

transformers/models/pixio/__init__.py

@@ -0,0 +1,30 @@
+# coding=utf-8
+# Copyright 2025 Meta AI and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Pixio model configuration"""
+
+from typing import TYPE_CHECKING
+
+from ...utils import _LazyModule
+from ...utils.import_utils import define_import_structure
+
+
+if TYPE_CHECKING:
+    from .configuration_pixio import *
+    from .modeling_pixio import *
+else:
+    import sys
+
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

transformers/models/pixio/configuration_pixio.py

@@ -0,0 +1,151 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/pixio/modular_pixio.py.
+#               Do NOT edit this file manually as any edits will be overwritten by the generation of
+#             the file from the modular. If any change should be done, please apply the change to the
+#                          modular_pixio.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+# coding=utf-8
+# Copyright 2025 Meta AI and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from ...configuration_utils import PreTrainedConfig
+from ...utils.backbone_utils import BackboneConfigMixin, get_aligned_output_features_output_indices
+
+
+class PixioConfig(BackboneConfigMixin, PreTrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`PixioModel`]. It is used to instantiate a
+    Pixio model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of the ViT
+    [facebook/pixio-huge](https://huggingface.co/facebook/pixio-huge) architecture.
+
+    Configuration objects inherit from [`PreTrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PreTrainedConfig`] for more information.
+
+    Args:
+        hidden_size (`int`, *optional*, defaults to 1280):
+            Dimensionality of the encoder layers and the pooler layer.
+        num_hidden_layers (`int`, *optional*, defaults to 32):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        mlp_ratio (`int`, *optional*, defaults to 4):
+            Ratio of the hidden size of the MLPs relative to the `hidden_size`.
+        n_cls_tokens (`int`, *optional*, defaults to 8):
+            Number of class tokens in the Transformer encoder.
+        hidden_act (`str` or `function`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"selu"` and `"gelu_new"` are supported.
+        hidden_dropout_prob (`float`, *optional*, defaults to 0.0):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_probs_dropout_prob (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        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-06):
+            The epsilon used by the layer normalization layers.
+        image_size (`int`, *optional*, defaults to 256):
+            The size (resolution) of each image.
+        patch_size (`int`, *optional*, defaults to 16):
+            The size (resolution) of each patch.
+        num_channels (`int`, *optional*, defaults to 3):
+            The number of input channels.
+        qkv_bias (`bool`, *optional*, defaults to `True`):
+            Whether to add a bias to the queries, keys and values.
+        drop_path_rate (`float`, *optional*, defaults to 0.0):
+            Stochastic depth rate per sample (when applied in the main path of residual layers).
+        out_features (`list[str]`, *optional*):
+            If used as backbone, list of features to output. Can be any of `"stem"`, `"stage1"`, `"stage2"`, etc.
+            (depending on how many stages the model has). If unset and `out_indices` is set, will default to the
+            corresponding stages. If unset and `out_indices` is unset, will default to the last stage. Must be in the
+            same order as defined in the `stage_names` attribute.
+        out_indices (`list[int]`, *optional*):
+            If used as backbone, list of indices of features to output. Can be any of 0, 1, 2, etc. (depending on how
+            many stages the model has). If unset and `out_features` is set, will default to the corresponding stages.
+            If unset and `out_features` is unset, will default to the last stage. Must be in the
+            same order as defined in the `stage_names` attribute.
+        apply_layernorm (`bool`, *optional*, defaults to `True`):
+            Whether to apply layer normalization to the feature maps in case the model is used as backbone.
+        reshape_hidden_states (`bool`, *optional*, defaults to `True`):
+            Whether to reshape the feature maps to 4D tensors of shape `(batch_size, hidden_size, height, width)` in
+            case the model is used as backbone. If `False`, the feature maps will be 3D tensors of shape `(batch_size,
+            seq_len, hidden_size)`.
+
+    Example:
+
+    ```python
+    >>> from transformers import PixioConfig, PixioModel
+
+    >>> # Initializing a Pixio pixio-huge style configuration
+    >>> configuration = PixioConfig()
+
+    >>> # Initializing a model (with random weights) from the pixio-huge style configuration
+    >>> model = PixioModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "pixio"
+
+    def __init__(
+        self,
+        hidden_size=1280,
+        num_hidden_layers=32,
+        num_attention_heads=16,
+        mlp_ratio=4,
+        n_cls_tokens=8,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.0,
+        attention_probs_dropout_prob=0.0,
+        initializer_range=0.02,
+        layer_norm_eps=1e-6,
+        image_size=256,
+        patch_size=16,
+        num_channels=3,
+        qkv_bias=True,
+        drop_path_rate=0.0,
+        out_features=None,
+        out_indices=None,
+        apply_layernorm=True,
+        reshape_hidden_states=True,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.mlp_ratio = mlp_ratio
+        self.hidden_act = hidden_act
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.num_channels = num_channels
+        self.qkv_bias = qkv_bias
+        self.drop_path_rate = drop_path_rate
+        self.stage_names = ["stem"] + [f"stage{idx}" for idx in range(1, num_hidden_layers + 1)]
+        self._out_features, self._out_indices = get_aligned_output_features_output_indices(
+            out_features=out_features, out_indices=out_indices, stage_names=self.stage_names
+        )
+        self.apply_layernorm = apply_layernorm
+        self.reshape_hidden_states = reshape_hidden_states
+
+        self.n_cls_tokens = n_cls_tokens
+
+
+__all__ = ["PixioConfig"]