optimum-rbln 0.8.2a0__py3-none-any.whl → 0.9.3__py3-none-any.whl

optimum/rbln/transformers/models/swin/modeling_swin.py

@@ -0,0 +1,354 @@
+# Copyright 2025 Rebellions Inc. 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.
+
+import types
+from typing import TYPE_CHECKING, Optional, Tuple, Union
+
+import torch
+import torch.nn.functional as F
+from transformers import SwinConfig
+from transformers.models.swin.modeling_swin import BackboneOutput
+
+from ....configuration_utils import RBLNCompileConfig, RBLNModelConfig
+from ....modeling import RBLNModel
+from ....utils.logging import get_logger
+from .configuration_swin import RBLNSwinBackboneConfig
+
+
+logger = get_logger(__name__)
+
+if TYPE_CHECKING:
+    from transformers import (
+        AutoFeatureExtractor,
+        AutoProcessor,
+        AutoTokenizer,
+        PreTrainedModel,
+        SwinBackbone,
+        SwinEncoder,
+    )
+
+
+def window_partition(input_feature, window_size):
+    """
+    Partitions the given input into windows.
+    """
+    batch_size, height, width, num_channels = input_feature.shape
+    input_feature = input_feature.view(
+        batch_size, height // window_size, window_size, width // window_size, window_size, num_channels
+    )
+    windows = input_feature.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, num_channels)
+    return windows
+
+
+def get_attn_mask(self, height, width, dtype, device):
+    if self.shift_size > 0:
+        # calculate attention mask for SW-MSA
+        img_mask = torch.zeros((1, height, width, 1), dtype=dtype, device=device)
+        height_slices = (
+            slice(0, -self.window_size),
+            slice(-self.window_size, -self.shift_size),
+            slice(-self.shift_size, None),
+        )
+        width_slices = (
+            slice(0, -self.window_size),
+            slice(-self.window_size, -self.shift_size),
+            slice(-self.shift_size, None),
+        )
+        count = torch.zeros(1)
+        for height_slice in height_slices:
+            for width_slice in width_slices:
+                img_mask[:, height_slice, width_slice, :] = count
+                count += 1
+
+        mask_windows = window_partition(img_mask, self.window_size)
+        mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
+        attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
+        attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
+    else:
+        attn_mask = None
+    return attn_mask
+
+
+class _SwinEncoder(torch.nn.Module):
+    def __init__(self, model: "SwinEncoder"):
+        super().__init__()
+        self.layers = model.layers
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        input_dimensions: Tuple[int, int],
+        head_mask: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = False,
+        output_hidden_states: Optional[bool] = False,
+        output_hidden_states_before_downsampling: Optional[bool] = False,
+        always_partition: Optional[bool] = False,
+        return_dict: Optional[bool] = True,
+    ):
+        all_hidden_states = () if output_hidden_states else None
+        all_reshaped_hidden_states = () if output_hidden_states else None
+        all_self_attentions = () if output_attentions else None
+
+        if output_hidden_states:
+            batch_size, _, hidden_size = hidden_states.shape
+            # rearrange b (h w) c -> b c h w
+            reshaped_hidden_state = hidden_states.view(batch_size, *input_dimensions, hidden_size)
+            reshaped_hidden_state = reshaped_hidden_state.permute(0, 3, 1, 2)
+            all_hidden_states += (hidden_states,)
+            all_reshaped_hidden_states += (reshaped_hidden_state,)
+
+        for i, layer_module in enumerate(self.layers):
+            layer_head_mask = head_mask[i] if head_mask is not None else None
+
+            layer_outputs = layer_module(
+                hidden_states, input_dimensions, layer_head_mask, output_attentions, always_partition
+            )
+
+            hidden_states = layer_outputs[0]
+            hidden_states_before_downsampling = layer_outputs[1]
+            output_dimensions = layer_outputs[2]
+
+            input_dimensions = (output_dimensions[-2], output_dimensions[-1])
+
+            if output_hidden_states and output_hidden_states_before_downsampling:
+                batch_size, _, hidden_size = hidden_states_before_downsampling.shape
+                # rearrange b (h w) c -> b c h w
+                # here we use the original (not downsampled) height and width
+                reshaped_hidden_state = hidden_states_before_downsampling.view(
+                    batch_size, *(output_dimensions[0], output_dimensions[1]), hidden_size
+                )
+                reshaped_hidden_state = reshaped_hidden_state.permute(0, 3, 1, 2)
+                all_hidden_states += (hidden_states_before_downsampling,)
+                all_reshaped_hidden_states += (reshaped_hidden_state,)
+            elif output_hidden_states and not output_hidden_states_before_downsampling:
+                batch_size, _, hidden_size = hidden_states.shape
+                # rearrange b (h w) c -> b c h w
+                reshaped_hidden_state = hidden_states.view(batch_size, *input_dimensions, hidden_size)
+                reshaped_hidden_state = reshaped_hidden_state.permute(0, 3, 1, 2)
+                all_hidden_states += (hidden_states,)
+                all_reshaped_hidden_states += (reshaped_hidden_state,)
+
+            if output_attentions:
+                all_self_attentions += layer_outputs[3:]
+
+        return tuple(
+            v
+            for v in [hidden_states, all_hidden_states, all_self_attentions, all_reshaped_hidden_states]
+            if v is not None
+        )
+
+
+class _SwinBackbone(torch.nn.Module):
+    def __init__(self, model: "SwinBackbone", output_hidden_states: bool, output_attentions: bool):
+        super().__init__()
+        self.model = model
+        self.embeddings = model.embeddings
+        self.encoder = model.encoder
+        self.stage_names = model.stage_names
+        self.out_features = model.out_features
+        self.hidden_states_norms = model.hidden_states_norms
+        self.output_hidden_states = output_hidden_states
+        self.output_attentions = output_attentions
+
+    def forward(
+        self,
+        pixel_values: torch.Tensor,
+    ):
+        embedding_output, input_dimensions = self.embeddings(pixel_values)
+        outputs = _SwinEncoder(self.encoder)(
+            embedding_output,
+            input_dimensions,
+            head_mask=None,
+            output_attentions=self.output_attentions,
+            output_hidden_states=True,
+            output_hidden_states_before_downsampling=True,
+            always_partition=True,
+            return_dict=False,
+        )
+
+        hidden_states = outputs[-1]
+
+        feature_maps = ()
+        for stage, hidden_state in zip(self.stage_names, hidden_states):
+            if stage in self.out_features:
+                batch_size, num_channels, height, width = hidden_state.shape
+                hidden_state = hidden_state.permute(0, 2, 3, 1).contiguous()
+                hidden_state = hidden_state.view(batch_size, height * width, num_channels)
+                hidden_state = self.hidden_states_norms[stage](hidden_state)
+                hidden_state = hidden_state.view(batch_size, height, width, num_channels)
+                hidden_state = hidden_state.permute(0, 3, 1, 2).contiguous()
+                feature_maps += (hidden_state,)
+
+        output = (feature_maps,)
+
+        if self.output_hidden_states:
+            output += (outputs[1],)
+
+        if self.output_attentions:
+            output += (outputs[2],)
+
+        return output
+
+
+class RBLNSwinBackbone(RBLNModel):
+    @classmethod
+    def _wrap_model_if_needed(cls, model: torch.nn.Module, rbln_config: RBLNSwinBackboneConfig) -> torch.nn.Module:
+        for layer in model.encoder.layers:
+            for block in layer.blocks:
+                block.get_attn_mask = types.MethodType(get_attn_mask, block)
+
+        wrapper_cfg = {
+            "output_hidden_states": rbln_config.output_hidden_states,
+            "output_attentions": rbln_config.output_attentions,
+        }
+        return _SwinBackbone(model, **wrapper_cfg).eval()
+
+    @classmethod
+    def _update_submodule_config(
+        cls,
+        model: "PreTrainedModel",
+        rbln_config: RBLNModelConfig,
+        preprocessors: Optional[Union["AutoFeatureExtractor", "AutoProcessor", "AutoTokenizer"]],
+    ):
+        for processor in preprocessors:
+            if rbln_config.image_size is None and hasattr(processor, "image_processor"):
+                if "height" in processor.image_processor.size and "width" in processor.image_processor.size:
+                    rbln_config.image_size = (
+                        processor.image_processor.size["height"],
+                        processor.image_processor.size["width"],
+                    )
+                elif (
+                    "longest_edge" in processor.image_processor.size
+                    and "shortest_edge" in processor.image_processor.size
+                ):
+                    rbln_config.image_size = processor.image_processor.size["longest_edge"]
+                elif "shortest_edge" in processor.image_processor.size:
+                    rbln_config.image_size = processor.image_processor.size["shortest_edge"]
+                break
+
+        return rbln_config
+
+    @classmethod
+    def _update_rbln_config(
+        cls,
+        preprocessors: Union["AutoFeatureExtractor", "AutoProcessor", "AutoTokenizer"],
+        model: Optional["PreTrainedModel"] = None,
+        model_config: "SwinConfig" = None,
+        rbln_config: Optional[RBLNSwinBackboneConfig] = None,
+    ) -> RBLNSwinBackboneConfig:
+        if rbln_config.image_size is None:
+            for processor in preprocessors:
+                if hasattr(processor, "size"):
+                    if all(required_key in processor.size.keys() for required_key in ["height", "width"]):
+                        rbln_config.image_size = (processor.size["height"], processor.size["width"])
+                    break
+
+        input_info = [
+            (
+                "pixel_values",
+                [
+                    rbln_config.batch_size,
+                    3,
+                    rbln_config.image_height,
+                    rbln_config.image_width,
+                ],
+                "float32",
+            ),
+        ]
+
+        rbln_config.set_compile_cfgs([RBLNCompileConfig(input_info=input_info)])
+        return rbln_config
+
+    def forward(
+        self,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        return_dict: bool = True,
+        output_attentions: bool = None,
+        output_hidden_states: bool = None,
+        **kwargs,
+    ) -> Union[Tuple, BackboneOutput]:
+        """
+        Forward pass for the RBLN-optimized Swin backbone model.
+
+        Args:
+            pixel_values (torch.FloatTensor of shape (batch_size, num_channels, image_size, image_size), optional): The tensors corresponding to the input images. Pixel values can be obtained using ViTImageProcessor. See ViTImageProcessor.call() for details (processor_class uses ViTImageProcessor for processing images).
+            return_dict (bool, optional): Whether or not to return a ModelOutput instead of a plain tuple.
+            output_attentions (bool, optional): Whether or not to return the attentions tensors of all attention layers. See attentions under returned tensors for more detail.
+            output_hidden_states (bool, optional): Whether or not to return the hidden states of all layers. See hidden_states under returned tensors for more detail.
+
+        Returns:
+            The model outputs. If return_dict=False is passed, returns a tuple of tensors. Otherwise, returns a BackboneOutput object.
+        """
+
+        if len(kwargs) > 0 and any(value is not None for value in kwargs.values()):
+            logger.warning(
+                f"Currently, optimum-rbln does not support kwargs {kwargs.keys()} for {self.__class__.__name__}."
+            )
+
+        output_attentions = output_attentions if output_attentions is not None else self.rbln_config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.rbln_config.output_hidden_states
+        )
+
+        if output_attentions != self.rbln_config.output_attentions:
+            raise ValueError(
+                f"Variable output_attentions {output_attentions} is not equal to rbln_config.output_attentions {self.rbln_config.output_attentions} "
+                f"Please compile again with the correct argument."
+            )
+
+        if output_hidden_states != self.rbln_config.output_hidden_states:
+            raise ValueError(
+                f"Variable output_hidden_states {output_hidden_states} is not equal to rbln_config.output_hidden_states {self.rbln_config.output_hidden_states} "
+                f"Please compile again with the correct argument."
+            )
+
+        _, _, original_h, original_w = pixel_values.shape
+        if original_h > self.rbln_config.image_height or original_w > self.rbln_config.image_width:
+            raise ValueError(
+                f"Input image size ({original_h}x{original_w}) exceeds the configured maximum size"
+                f" ({self.rbln_config.image_height}x{self.rbln_config.image_width})."
+            )
+
+        pad_h = self.rbln_config.image_height - original_h
+        pad_w = self.rbln_config.image_width - original_w
+        padded_pixel_values = F.pad(pixel_values, (0, pad_w, 0, pad_h))
+
+        output = self.model[0](padded_pixel_values)
+
+        feature_maps = ()
+        for i in range(len(self.config.out_features)):
+            feature_maps += (output.pop(0),)
+
+        if self.rbln_config.output_hidden_states:
+            hidden_states = ()
+            for i in range(len(self.config.stage_names)):
+                hidden_states += (output.pop(0),)
+        else:
+            hidden_states = None
+
+        if self.rbln_config.output_attentions:
+            attentions = ()
+            for i in range(len(self.config.depths)):
+                attentions += (output.pop(0),)
+        else:
+            attentions = None
+
+        if not return_dict:
+            return tuple(item for item in (feature_maps, hidden_states, attentions) if item is not None)
+        else:
+            return BackboneOutput(
+                feature_maps=feature_maps,
+                hidden_states=hidden_states,
+                attentions=attentions,
+            )

optimum/rbln/transformers/models/t5/configuration_t5.py

@@ -32,3 +32,5 @@ class RBLNT5ForConditionalGenerationConfig(RBLNModelForSeq2SeqLMConfig):
     This configuration class stores the configuration parameters specific to
     RBLN-optimized T5 models for conditional text generation tasks.
     """
+
+    support_paged_attention = False

optimum/rbln/transformers/models/t5/modeling_t5.py

@@ -68,7 +68,7 @@ class RBLNT5EncoderModel(RBLNTransformerEncoderForFeatureExtraction):
     output_class = BaseModelOutputWithPastAndCrossAttentions
 
     @classmethod
-    def wrap_model_if_needed(self, model: "PreTrainedModel", rbln_config: RBLNT5EncoderModelConfig):
+    def _wrap_model_if_needed(self, model: "PreTrainedModel", rbln_config: RBLNT5EncoderModelConfig):
         return T5EncoderWrapper(model)
 
     @classmethod
@@ -113,7 +113,7 @@ class RBLNT5ForConditionalGeneration(RBLNModelForSeq2SeqLM):
     support_causal_attn = False
 
     @classmethod
-    def wrap_model_if_needed(self, model: "PreTrainedModel", rbln_config: RBLNT5ForConditionalGenerationConfig):
+    def _wrap_model_if_needed(self, model: "PreTrainedModel", rbln_config: RBLNT5ForConditionalGenerationConfig):
         return T5Wrapper(
             model, enc_max_seq_len=rbln_config.enc_max_seq_len, dec_max_seq_len=rbln_config.dec_max_seq_len
         )

optimum/rbln/transformers/models/t5/t5_architecture.py

@@ -126,7 +126,14 @@ class T5Decoder(Seq2SeqDecoder):
         b_size = attention_mask.shape[0]
         batch_decoder_position_bias = []
         for i in range(b_size):
-            batch_position_bias = self._dec_position_bias[:, :, cache_position[i][0]].unsqueeze(2)
+            if torch.compiler.is_exporting():
+                cache_pos = cache_position[i][0].item()
+                torch._check_is_size(cache_pos)
+                torch._check(cache_pos >= 0)
+                torch._check(cache_pos < self._dec_position_bias.shape[2])
+            else:
+                cache_pos = cache_position[i][0]
+            batch_position_bias = torch.select(self._dec_position_bias, dim=2, index=cache_pos).unsqueeze(2)
             batch_decoder_position_bias.append(batch_position_bias)
         position_bias = torch.cat(batch_decoder_position_bias, dim=0)
 

optimum/rbln/transformers/models/time_series_transformer/configuration_time_series_transformer.py

@@ -1,4 +1,4 @@
-from typing import Any, Dict, Optional
+from typing import Any, Optional
 
 from ....configuration_utils import RBLNModelConfig
 
@@ -17,7 +17,7 @@ class RBLNTimeSeriesTransformerForPredictionConfig(RBLNModelConfig):
         enc_max_seq_len: Optional[int] = None,
         dec_max_seq_len: Optional[int] = None,
         num_parallel_samples: Optional[int] = None,
-        **kwargs: Dict[str, Any],
+        **kwargs: Any,
     ):
         """
         Args:
@@ -25,7 +25,7 @@ class RBLNTimeSeriesTransformerForPredictionConfig(RBLNModelConfig):
             enc_max_seq_len (Optional[int]): Maximum sequence length for the encoder.
             dec_max_seq_len (Optional[int]): Maximum sequence length for the decoder.
             num_parallel_samples (Optional[int]): Number of samples to generate in parallel during prediction.
-            **kwargs: Additional arguments passed to the parent RBLNModelConfig.
+            kwargs: Additional arguments passed to the parent RBLNModelConfig.
 
         Raises:
             ValueError: If batch_size is not a positive integer.

optimum/rbln/transformers/models/time_series_transformer/modeling_time_series_transformer.py

@@ -23,24 +23,20 @@
 
 import inspect
 import logging
-from dataclasses import dataclass
 from pathlib import Path
-from typing import TYPE_CHECKING, Any, Callable, List, Optional, Tuple, Union
+from typing import TYPE_CHECKING, Any, Callable, List, Optional, Union
 
 import rebel
 import torch
 from rebel.compile_context import CompileContext
-from transformers import (
-    PretrainedConfig,
-    TimeSeriesTransformerForPrediction,
-    TimeSeriesTransformerModel,
-)
-from transformers.modeling_outputs import ModelOutput, SampleTSPredictionOutput, Seq2SeqTSModelOutput
+from transformers import PretrainedConfig, TimeSeriesTransformerForPrediction, TimeSeriesTransformerModel
+from transformers.modeling_outputs import SampleTSPredictionOutput, Seq2SeqTSModelOutput
 from transformers.modeling_utils import no_init_weights
 
 from ....configuration_utils import RBLNCompileConfig
 from ....modeling import RBLNModel
 from ....utils.runtime_utils import RBLNPytorchRuntime
+from ...modeling_outputs import RBLNSeq2SeqTSDecoderOutput
 from .configuration_time_series_transformer import RBLNTimeSeriesTransformerForPredictionConfig
 from .time_series_transformers_architecture import TimeSeriesTransformersWrapper
 
@@ -113,12 +109,6 @@ class RBLNRuntimeDecoder(RBLNPytorchRuntime):
         )
 
 
-@dataclass
-class RBLNSeq2SeqTSDecoderOutput(ModelOutput):
-    last_hidden_states: torch.FloatTensor = None
-    params: Tuple[torch.FloatTensor] = None
-
-
 class RBLNTimeSeriesTransformerForPrediction(RBLNModel):
     """
     The Time Series Transformer Model with a distribution head on top for time-series forecasting. e.g., for datasets like M4, NN5, or other time series forecasting benchmarks.
@@ -163,7 +153,7 @@ class RBLNTimeSeriesTransformerForPrediction(RBLNModel):
             return redirect(val)
 
     @classmethod
-    def wrap_model_if_needed(
+    def _wrap_model_if_needed(
         self, model: "PreTrainedModel", rbln_config: RBLNTimeSeriesTransformerForPredictionConfig
     ):
         return TimeSeriesTransformersWrapper(model, rbln_config.num_parallel_samples)
@@ -171,7 +161,7 @@ class RBLNTimeSeriesTransformerForPrediction(RBLNModel):
     @classmethod
     @torch.inference_mode()
     def get_compiled_model(cls, model, rbln_config: RBLNTimeSeriesTransformerForPredictionConfig):
-        wrapped_model = cls.wrap_model_if_needed(model, rbln_config)
+        wrapped_model = cls._wrap_model_if_needed(model, rbln_config)
 
         enc_compile_config = rbln_config.compile_cfgs[0]
         dec_compile_config = rbln_config.compile_cfgs[1]
@@ -331,12 +321,14 @@ class RBLNTimeSeriesTransformerForPrediction(RBLNModel):
                 tensor_type="pt",
                 device=rbln_config.device_map["encoder"],
                 activate_profiler=rbln_config.activate_profiler,
+                timeout=rbln_config.timeout,
             ),
             rebel.Runtime(
                 compiled_models[1],
                 tensor_type="pt",
                 device=rbln_config.device_map["decoder"],
                 activate_profiler=rbln_config.activate_profiler,
+                timeout=rbln_config.timeout,
             ),
         ]
 
@@ -361,6 +353,20 @@ class RBLNTimeSeriesTransformerForPrediction(RBLNModel):
         static_real_features: Optional[torch.Tensor] = None,
         **kwargs,
     ) -> SampleTSPredictionOutput:
+        """
+        Generate pass for the RBLN-optimized Time Series Transformer model for time series forecasting.
+
+        Args:
+            past_values (torch.FloatTensor of shape (batch_size, sequence_length) or (batch_size, sequence_length, input_size)): Past values of the time series, that serve as context in order to predict the future.
+            past_time_features (torch.FloatTensor of shape (batch_size, sequence_length, num_features)): Required time features, which the model internally will add to past_values.
+            future_time_features (torch.FloatTensor of shape (batch_size, prediction_length, num_features)): Required time features for the prediction window, which the model internally will add to future_values.
+            past_observed_mask (torch.BoolTensor of shape (batch_size, sequence_length) or (batch_size, sequence_length, input_size), optional): Boolean mask to indicate which past_values were observed and which were missing.
+            static_categorical_features (torch.LongTensor of shape (batch_size, number of static categorical features), optional): Optional static categorical features for which the model will learn an embedding, which it will add to the values of the time series.
+            static_real_features (torch.FloatTensor of shape (batch_size, number of static real features), optional): Optional static real features which the model will add to the values of the time series.
+
+        Returns:
+            The model outputs. If return_dict=False is passed, returns a tuple of tensors. Otherwise, returns a SampleTSPredictionOutput object.
+        """
         self.validate_batch_size(**{k: v for k, v in locals().items() if isinstance(v, torch.Tensor)})
 
         outputs = self.encoder(

optimum/rbln/transformers/models/time_series_transformer/time_series_transformers_architecture.py

@@ -162,7 +162,13 @@ class TimeSeriesTransformersDecoder(nn.Module):
         attention_mask = _prepare_4d_causal_attention_mask(attention_mask, input_shape, inputs_embeds, cache_position)
 
         hidden_states = self.value_embedding(inputs_embeds)
-        embed_pos = self.embed_positions.weight[cache_position + self.config.context_length]
+        embed_idx = cache_position + self.config.context_length
+        if torch.compiler.is_exporting():
+            embed_idx = embed_idx.item()
+            torch._check_is_size(embed_idx)
+            torch._check(embed_idx >= 0)
+            torch._check(embed_idx < len(self.embed_positions.weight))
+        embed_pos = self.embed_positions.weight[embed_idx]
         hidden_states = self.layernorm_embedding(hidden_states + embed_pos)
 
         # iterate decoder_layer

optimum/rbln/transformers/models/vit/modeling_vit.py

@@ -12,6 +12,11 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from typing import Tuple, Union
+
+import torch
+from transformers.modeling_outputs import ImageClassifierOutput
+
 from ...modeling_generic import RBLNModelForImageClassification
 
 
@@ -23,3 +28,17 @@ class RBLNViTForImageClassification(RBLNModelForImageClassification):
     on RBLN devices, supporting image classification with transformer-based architectures
     that process images as sequences of patches.
     """
+
+    def forward(self, pixel_values: torch.Tensor, **kwargs) -> Union[ImageClassifierOutput, Tuple]:
+        """
+        Forward pass for the RBLN-optimized Vision Transformer model for image classification.
+
+        Args:
+            pixel_values (torch.FloatTensor of shape (batch_size, channels, height, width)):
+                The tensors corresponding to the input images.
+
+        Returns:
+            The model outputs. If return_dict=False is passed, returns a tuple of tensors. Otherwise, returns an ImageClassifierOutput object.
+
+        """
+        return super().forward(pixel_values, **kwargs)

optimum/rbln/transformers/models/wav2vec2/configuration_wav2vec2.py

@@ -12,10 +12,12 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ...configuration_generic import RBLNModelForMaskedLMConfig
+from typing import Any, Optional
 
+from ....configuration_utils import RBLNModelConfig
 
-class RBLNWav2Vec2ForCTCConfig(RBLNModelForMaskedLMConfig):
+
+class RBLNWav2Vec2ForCTCConfig(RBLNModelConfig):
     """
     Configuration class for RBLNWav2Vec2ForCTC.
 
@@ -23,4 +25,14 @@ class RBLNWav2Vec2ForCTCConfig(RBLNModelForMaskedLMConfig):
     RBLN-optimized Wav2Vec2 models for Connectionist Temporal Classification (CTC) tasks.
     """
 
-    rbln_model_input_names = ["input_values"]
+    def __init__(
+        self,
+        max_seq_len: Optional[int] = None,
+        batch_size: Optional[int] = None,
+        **kwargs: Any,
+    ):
+        super().__init__(**kwargs)
+        self.max_seq_len = max_seq_len
+        self.batch_size = batch_size or 1
+        if not isinstance(self.batch_size, int) or self.batch_size < 0:
+            raise ValueError(f"batch_size must be a positive integer, got {self.batch_size}")