optimum-rbln 0.8.3a2__py3-none-any.whl → 0.8.3a4__py3-none-any.whl

optimum/rbln/__init__.py

@@ -81,6 +81,8 @@ _import_structure = {
         "RBLNDistilBertForQuestionAnsweringConfig",
         "RBLNDPTForDepthEstimation",
         "RBLNDPTForDepthEstimationConfig",
+        "RBLNDepthAnythingForDepthEstimationConfig",
+        "RBLNDepthAnythingForDepthEstimation",
         "RBLNExaoneForCausalLM",
         "RBLNExaoneForCausalLMConfig",
         "RBLNGemmaModel",
@@ -147,6 +149,8 @@ _import_structure = {
         "RBLNRobertaForSequenceClassificationConfig",
         "RBLNSiglipVisionModel",
         "RBLNSiglipVisionModelConfig",
+        "RBLNSwinBackbone",
+        "RBLNSwinBackboneConfig",
         "RBLNT5EncoderModel",
         "RBLNT5EncoderModelConfig",
         "RBLNT5ForConditionalGeneration",
@@ -352,6 +356,8 @@ if TYPE_CHECKING:
         RBLNDecoderOnlyModelConfig,
         RBLNDecoderOnlyModelForCausalLM,
         RBLNDecoderOnlyModelForCausalLMConfig,
+        RBLNDepthAnythingForDepthEstimation,
+        RBLNDepthAnythingForDepthEstimationConfig,
         RBLNDistilBertForQuestionAnswering,
         RBLNDistilBertForQuestionAnsweringConfig,
         RBLNDPTForDepthEstimation,
@@ -422,6 +428,8 @@ if TYPE_CHECKING:
         RBLNRobertaForSequenceClassificationConfig,
         RBLNSiglipVisionModel,
         RBLNSiglipVisionModelConfig,
+        RBLNSwinBackbone,
+        RBLNSwinBackboneConfig,
         RBLNT5EncoderModel,
         RBLNT5EncoderModelConfig,
         RBLNT5ForConditionalGeneration,

optimum/rbln/__version__.py

@@ -28,7 +28,7 @@ version_tuple: VERSION_TUPLE
 commit_id: COMMIT_ID
 __commit_id__: COMMIT_ID
 
-__version__ = version = '0.8.3a2'
-__version_tuple__ = version_tuple = (0, 8, 3, 'a2')
+__version__ = version = '0.8.3a4'
+__version_tuple__ = version_tuple = (0, 8, 3, 'a4')
 
 __commit_id__ = commit_id = None

optimum/rbln/configuration_utils.py

@@ -21,6 +21,7 @@ from typing import Any, Dict, List, Optional, Protocol, Tuple, Type, Union, runt
 
 import numpy as np
 import torch
+from packaging.version import Version
 
 from .__version__ import __version__
 from .utils.depreacate_utils import warn_deprecated_npu
@@ -621,6 +622,21 @@ class RBLNModelConfig(RBLNSerializableConfigProtocol):
             self.set_compile_cfgs([RBLNCompileConfig(**cfg) for cfg in self._compile_cfgs])
 
         if len(kwargs) > 0:
+            if optimum_rbln_version is not None:  # loaded from file
+                if Version(__version__) < Version(optimum_rbln_version):
+                    diff = "newer"
+                elif Version(__version__) > Version(optimum_rbln_version):
+                    diff = "older"
+                else:
+                    diff = None
+                if diff is not None:
+                    raise ValueError(
+                        f"Unexpected arguments: {kwargs.keys()}\n"
+                        f"Maybe you are trying to load a model compiled with {diff} version of optimum-rbln. "
+                        "It is recommended to use the same version to compile and load the model.\n"
+                        f"Current version: {__version__}, Loaded version: {optimum_rbln_version}"
+                    )
+
             raise ValueError(f"Unexpected arguments: {kwargs.keys()}")
 
     @property

optimum/rbln/diffusers/pipelines/auto_pipeline.py

@@ -48,7 +48,7 @@ class RBLNAutoPipelineBase:
             rbln_class_name = convert_hf_to_rbln_model_name(hf_model_class.__name__)
         else:
             rbln_class_name = cls.get_rbln_model_cls_name(pretrained_model_name_or_path, **kwargs)
-            if convert_rbln_to_hf_model_name(rbln_class_name) not in cls._model_mapping_names:
+            if convert_rbln_to_hf_model_name(rbln_class_name) not in cls._model_mapping_names.values():
                 raise ValueError(
                     f"The architecture '{rbln_class_name}' is not supported by the `{cls.__name__}.from_pretrained()` method. "
                     "Please use the `from_pretrained()` method of the appropriate class to load this model, "

optimum/rbln/transformers/__init__.py

@@ -69,6 +69,8 @@ _import_structure = {
         "RBLNDistilBertForQuestionAnsweringConfig",
         "RBLNDPTForDepthEstimation",
         "RBLNDPTForDepthEstimationConfig",
+        "RBLNDepthAnythingForDepthEstimation",
+        "RBLNDepthAnythingForDepthEstimationConfig",
         "RBLNExaoneForCausalLM",
         "RBLNExaoneForCausalLMConfig",
         "RBLNGemmaModel",
@@ -137,6 +139,8 @@ _import_structure = {
         "RBLNRobertaForSequenceClassificationConfig",
         "RBLNSiglipVisionModel",
         "RBLNSiglipVisionModelConfig",
+        "RBLNSwinBackbone",
+        "RBLNSwinBackboneConfig",
         "RBLNT5EncoderModel",
         "RBLNT5EncoderModelConfig",
         "RBLNT5ForConditionalGeneration",
@@ -204,6 +208,8 @@ if TYPE_CHECKING:
         RBLNDecoderOnlyModelConfig,
         RBLNDecoderOnlyModelForCausalLM,
         RBLNDecoderOnlyModelForCausalLMConfig,
+        RBLNDepthAnythingForDepthEstimation,
+        RBLNDepthAnythingForDepthEstimationConfig,
         RBLNDistilBertForQuestionAnswering,
         RBLNDistilBertForQuestionAnsweringConfig,
         RBLNDPTForDepthEstimation,
@@ -274,6 +280,8 @@ if TYPE_CHECKING:
         RBLNRobertaForSequenceClassificationConfig,
         RBLNSiglipVisionModel,
         RBLNSiglipVisionModelConfig,
+        RBLNSwinBackbone,
+        RBLNSwinBackboneConfig,
         RBLNT5EncoderModel,
         RBLNT5EncoderModelConfig,
         RBLNT5ForConditionalGeneration,

optimum/rbln/transformers/models/__init__.py

@@ -90,6 +90,7 @@ _import_structure = {
         "RBLNDecoderOnlyModelForCausalLM",
         "RBLNDecoderOnlyModelForCausalLMConfig",
     ],
+    "depth_anything": ["RBLNDepthAnythingForDepthEstimationConfig", "RBLNDepthAnythingForDepthEstimation"],
     "dpt": [
         "RBLNDPTForDepthEstimation",
         "RBLNDPTForDepthEstimationConfig",
@@ -141,6 +142,10 @@ _import_structure = {
         "RBLNSiglipVisionModel",
         "RBLNSiglipVisionModelConfig",
     ],
+    "swin": [
+        "RBLNSwinBackbone",
+        "RBLNSwinBackboneConfig",
+    ],
     "time_series_transformer": [
         "RBLNTimeSeriesTransformerForPrediction",
         "RBLNTimeSeriesTransformerForPredictionConfig",
@@ -219,6 +224,7 @@ if TYPE_CHECKING:
         RBLNDecoderOnlyModelForCausalLM,
         RBLNDecoderOnlyModelForCausalLMConfig,
     )
+    from .depth_anything import RBLNDepthAnythingForDepthEstimation, RBLNDepthAnythingForDepthEstimationConfig
     from .distilbert import RBLNDistilBertForQuestionAnswering, RBLNDistilBertForQuestionAnsweringConfig
     from .dpt import RBLNDPTForDepthEstimation, RBLNDPTForDepthEstimationConfig
     from .exaone import RBLNExaoneForCausalLM, RBLNExaoneForCausalLMConfig
@@ -266,6 +272,7 @@ if TYPE_CHECKING:
         RBLNRobertaForSequenceClassificationConfig,
     )
     from .siglip import RBLNSiglipVisionModel, RBLNSiglipVisionModelConfig
+    from .swin import RBLNSwinBackbone, RBLNSwinBackboneConfig
     from .t5 import (
         RBLNT5EncoderModel,
         RBLNT5EncoderModelConfig,

optimum/rbln/transformers/models/depth_anything/__init__.py

@@ -0,0 +1,16 @@
+# 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.
+
+from .configuration_depth_anything import RBLNDepthAnythingForDepthEstimationConfig
+from .modeling_depth_anything import RBLNDepthAnythingForDepthEstimation

optimum/rbln/transformers/models/depth_anything/configuration_depth_anything.py

@@ -0,0 +1,24 @@
+# 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.
+
+from ...configuration_generic import RBLNModelForDepthEstimationConfig
+
+
+class RBLNDepthAnythingForDepthEstimationConfig(RBLNModelForDepthEstimationConfig):
+    """
+    Configuration class for DepthAnythingForDepthEstimation.
+
+    This configuration class stores the configuration parameters specific to
+    RBLN-optimized Depth Anything V2 Small models for depth estimation tasks.
+    """

optimum/rbln/transformers/models/depth_anything/modeling_depth_anything.py

@@ -0,0 +1,25 @@
+# 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.
+
+
+from ...modeling_generic import RBLNModelForDepthEstimation
+
+
+class RBLNDepthAnythingForDepthEstimation(RBLNModelForDepthEstimation):
+    """
+    RBLN optimized DepthAnythingForDepthEstimation model for depth estimation tasks.
+
+    This class provides hardware-accelerated inference for Depth Anything V2 Small
+    models on RBLN devices, providing the most capable monocular depth estimation (MDE) model.
+    """

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

@@ -0,0 +1,16 @@
+# 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.
+
+from .configuration_swin import RBLNSwinBackboneConfig
+from .modeling_swin import RBLNSwinBackbone

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

@@ -0,0 +1,42 @@
+# 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 typing import Any, Optional, Tuple, Union
+
+from ...configuration_generic import RBLNModelForImageClassificationConfig
+
+
+class RBLNSwinBackboneConfig(RBLNModelForImageClassificationConfig):
+    def __init__(
+        self,
+        image_size: Optional[Union[int, Tuple[int, int]]] = None,
+        batch_size: Optional[int] = None,
+        output_hidden_states: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        **kwargs: Any,
+    ):
+        """
+        Args:
+            batch_size (Optional[int]): The batch size for text processing. Defaults to 1.
+            **kwargs: Additional arguments passed to the parent RBLNModelConfig.
+
+        Raises:
+            ValueError: If batch_size is not a positive integer.
+        """
+        super().__init__(**kwargs)
+        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}")
+
+        self.image_size = image_size
+        self.output_hidden_states = output_hidden_states
+        self.output_attentions = output_attentions

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

@@ -0,0 +1,316 @@
+# 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
+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_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_size[0],
+                    rbln_config.image_size[1],
+                ],
+                "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]:
+        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_size[0] or original_w > self.rbln_config.image_size[1]:
+            raise ValueError(
+                f"Input image size ({original_h}x{original_w}) exceeds the configured maximum size"
+                f" ({self.rbln_config.image_size[0]}x{self.rbln_config.image_size[1]})."
+            )
+
+        pad_h = self.rbln_config.image_size[0] - original_h
+        pad_w = self.rbln_config.image_size[1] - 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-0.8.3a2.dist-info → optimum_rbln-0.8.3a4.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: optimum-rbln
-Version: 0.8.3a2
+Version: 0.8.3a4
 Summary: Optimum RBLN is the interface between the HuggingFace Transformers and Diffusers libraries and RBLN accelerators. It provides a set of tools enabling easy model loading and inference on single and multiple rbln device settings for different downstream tasks.
 Project-URL: Homepage, https://rebellions.ai
 Project-URL: Documentation, https://docs.rbln.ai

{optimum_rbln-0.8.3a2.dist-info → optimum_rbln-0.8.3a4.dist-info}/RECORD

@@ -1,6 +1,6 @@
-optimum/rbln/__init__.py,sha256=YhaBhcyu6BgoJrprUogLGAmiBaHayvg6Tjm6PpfJETw,17382
-optimum/rbln/__version__.py,sha256=LoGi14U0L2os-fSHKgBIGeByegJLodfXKteGMBVsCEc,712
-optimum/rbln/configuration_utils.py,sha256=xneqnRWSUVROqpzbTrBACex42-L9zwo3eSjfHjFuhv4,33072
+optimum/rbln/__init__.py,sha256=1AF2jfrUGYLo4ps_OAPxCUagSnHF60N3W45cTeXlDJE,17698
+optimum/rbln/__version__.py,sha256=vl5bkt2g1rkDYLzC7btKWOmkJwvyed8HxQBkBGIvFzA,712
+optimum/rbln/configuration_utils.py,sha256=fE3HlZblxukKSdS-4VofjuyCAiqwPMX8bqXpOiTZp4g,33926
 optimum/rbln/modeling.py,sha256=0CMQnpVvW9evNrTFHM2XFbNpRY1HkbFzYJ5sRyYFq0o,14293
 optimum/rbln/modeling_base.py,sha256=gHfqIO6lKT8smkUthUuRHnbITpxHpnDeBPT8iTeasCk,24575
 optimum/rbln/diffusers/__init__.py,sha256=1tgU_xWA42BmInqu9bBz_5R_E9TGhhK3mI06YlaiTLg,7232
@@ -36,7 +36,7 @@ optimum/rbln/diffusers/models/transformers/transformer_sd3.py,sha256=yF7sS0Qvawo
 optimum/rbln/diffusers/models/unets/__init__.py,sha256=MaICuK9CWjgzejXy8y2NDrphuEq1rkzanF8u45k6O5I,655
 optimum/rbln/diffusers/models/unets/unet_2d_condition.py,sha256=v3WS9EGKROE_QClXrxC7rmRko1BspAvAbeIfh83LK88,15832
 optimum/rbln/diffusers/pipelines/__init__.py,sha256=r8mu21102cKXdkG1II9tpfpUS6wuyren2oK9y_MptZY,3703
-optimum/rbln/diffusers/pipelines/auto_pipeline.py,sha256=oGZWXfj82w695D2NiYUitgoWiwP2Z4PlgA3q6hoOKww,9502
+optimum/rbln/diffusers/pipelines/auto_pipeline.py,sha256=zFDXbO9Iv0LO7maefV82dmi5Ta6L9oZxY09QFVX6F_Q,9511
 optimum/rbln/diffusers/pipelines/controlnet/__init__.py,sha256=n1Ef22TSeax-kENi_d8K6wGGHSNEo9QkUeygELHgcao,983
 optimum/rbln/diffusers/pipelines/controlnet/multicontrolnet.py,sha256=3S9dogIHW8Bqg5kIlCudhCQG-4g3FcdOPEWhBOf7CJA,4059
 optimum/rbln/diffusers/pipelines/controlnet/pipeline_controlnet.py,sha256=G96bh4D9Cu-w4F9gZBQF6wNzhJQv9kvI34ZFsuEDjSw,35714
@@ -72,13 +72,13 @@ optimum/rbln/ops/flash_attn.py,sha256=yTCdYQVqm_1rHMHWjrMQaIR8WTuG_xA6t033x1IVvT
 optimum/rbln/ops/kv_cache_update.py,sha256=aIvK2Sp7M3EfJzJgNvIvAJv4emoN6QOhmgaWj-VboLs,1440
 optimum/rbln/ops/linear.py,sha256=5K3pcrrUHu_p8LrMIU-jX2TnafksveFjjZSCsYSp_yw,1328
 optimum/rbln/ops/sliding_window_attn.py,sha256=EQrV_yRGc5z6kvwEsAcLP028bJWkQg2UPI3xubt9skU,3487
-optimum/rbln/transformers/__init__.py,sha256=eBPY5FjB_6jul-s3ixero-7irrWMu1D2smuXr_7zDcI,11127
+optimum/rbln/transformers/__init__.py,sha256=uV2rEhw93alpbZ-fnVrAex_6QF1sFHcVM5tsJh64osk,11443
 optimum/rbln/transformers/configuration_generic.py,sha256=95ks6REJYuzI1zLwGlPSlxVV45saVcYOob6ihn-WAAY,5092
 optimum/rbln/transformers/modeling_attention_utils.py,sha256=aLyOaq4me1m-JMmnKbuyNQageDxNU2jjEhGE_ew2P5o,11465
 optimum/rbln/transformers/modeling_generic.py,sha256=2BtroigKuu7z7C98dpLwI875R0EoHN-ceHEVbyPQuYk,12212
 optimum/rbln/transformers/modeling_outputs.py,sha256=cd8ZlhHAGq7S6i5-QK6TJCxgORvoPMnZpqPBlUc_pMY,1177
 optimum/rbln/transformers/modeling_rope_utils.py,sha256=6Zg3r-TeUk4WQAlr95pqfhuoAD_RQ4njT1rbO9uPL0Q,14379
-optimum/rbln/transformers/models/__init__.py,sha256=PfaXIPx5fGroSkcS1xUx7eQ62aZovO3e-f71Tp69cWI,11907
+optimum/rbln/transformers/models/__init__.py,sha256=A9ThjEgBo6RZzqGzoY3tgQucdchkuXWrpgJjuERxjcE,12272
 optimum/rbln/transformers/models/audio_spectrogram_transformer/__init__.py,sha256=I2vL4lrzbT5p4eJcH-EKHzEfcPkj_XVsie7jb9q6yic,775
 optimum/rbln/transformers/models/audio_spectrogram_transformer/configuration_audio_spectrogram_transformer.py,sha256=z7LJiVJPmnlCM3mcyhPJP8AufSrxO_dsPeJ51onq-Nc,833
 optimum/rbln/transformers/models/audio_spectrogram_transformer/modeling_audio_spectrogram_transformer.py,sha256=FIKEVWpIt6-JQX9B_rAfCrAPqdUHtR2i8D_X2k7639E,1498
@@ -108,6 +108,9 @@ optimum/rbln/transformers/models/decoderonly/decoderonly_architecture.py,sha256=
 optimum/rbln/transformers/models/decoderonly/decoderonly_runtime_utils.py,sha256=9acEQxGRzd21YkzxRchkhqxqpX7emQHZigFg60BIulc,19902
 optimum/rbln/transformers/models/decoderonly/generation_decoderonly.py,sha256=4D89IF0yQju_Dp_vLJN_dBkpe2U_LMWaUciYx57D-0M,3379
 optimum/rbln/transformers/models/decoderonly/modeling_decoderonly.py,sha256=hu2eJr0CpLHnRPSLhyBhyyC6DfosKmPu7lPjapcBCkE,33061
+optimum/rbln/transformers/models/depth_anything/__init__.py,sha256=xvPSIriMJWyNeVYoVB1Z7YqB4kkHOIkaHq7loNps-dk,756
+optimum/rbln/transformers/models/depth_anything/configuration_depth_anything.py,sha256=JujBVEUa_zZDXNPr1y-B_PhK5SgFFcY8Ib4EoGjjtmE,989
+optimum/rbln/transformers/models/depth_anything/modeling_depth_anything.py,sha256=ganUtyPKcdKe5QCJ2atQhblzfAstyYkThYDm_DIecU8,1014
 optimum/rbln/transformers/models/distilbert/__init__.py,sha256=zXL78SOEORTnUN_wrdoaDaYpntG8lcFHvPobM6jC0CI,841
 optimum/rbln/transformers/models/distilbert/configuration_distilbert.py,sha256=O3BW9JjyYk9PLyiofvOKEgTdMZ_jpIuPfot281pSsyg,984
 optimum/rbln/transformers/models/distilbert/modeling_distilbert.py,sha256=LUh6zYGa8AR3Yxaj3gtyJRc-czBN3qnHTc-JTAhuqY0,1099
@@ -193,6 +196,9 @@ optimum/rbln/transformers/models/seq2seq/seq2seq_architecture.py,sha256=4MupGjhe
 optimum/rbln/transformers/models/siglip/__init__.py,sha256=X1Fc1GUnJ2EIxFx45nbeoW-T2t0OyP3W73C0HD8Vowo,712
 optimum/rbln/transformers/models/siglip/configuration_siglip.py,sha256=m1h8iDx_X9VmHdJi0sc1a2KsAO3OnpMb4cd9jW2Ic-U,3031
 optimum/rbln/transformers/models/siglip/modeling_siglip.py,sha256=1TyRaxmhp6mg6UfhQTbZhW26013TE3nVnroYG7EROcU,8033
+optimum/rbln/transformers/models/swin/__init__.py,sha256=gUsLDB8ceNxt53Cf69OT32JuZoRdmmIsRfjRdHTLDd0,698
+optimum/rbln/transformers/models/swin/configuration_swin.py,sha256=iVtpT2jXY5LNkUbbr5J08z97unc43KEhArIZ1tBRzEU,1692
+optimum/rbln/transformers/models/swin/modeling_swin.py,sha256=IvBOyIWq233UGZqUG7XhaoYIHk6gjbisYUEr8HDqRbY,12791
 optimum/rbln/transformers/models/t5/__init__.py,sha256=R1Q8Z1vaIdx4rDjeCmm_ZMSgewWaqaI0l93AHwewtew,818
 optimum/rbln/transformers/models/t5/configuration_t5.py,sha256=nqDbibqykeeWn1TlKk6LmCn-DawTVudMMuBn2c2jds8,1362
 optimum/rbln/transformers/models/t5/modeling_t5.py,sha256=pdAWBLVknTzbma0Ij-VQ2Qve-frPjxL-AwMyU-zouPY,5123
@@ -227,7 +233,7 @@ optimum/rbln/utils/model_utils.py,sha256=4k5879Kh75m3x_vS4-qOGfqsOiAvc2kdNFFfvsF
 optimum/rbln/utils/runtime_utils.py,sha256=R6uXDbeJP03-FWdd4vthNe2D4aCra5n12E3WB1ifiGM,7933
 optimum/rbln/utils/save_utils.py,sha256=hG5uOtYmecSXZuGTvCXsTM-SiyZpr5q3InUGCCq_jzQ,3619
 optimum/rbln/utils/submodule.py,sha256=w5mgPgncI740gVKMu3S-69DGNdUSI0bTZxegQGcZ98Y,5011
-optimum_rbln-0.8.3a2.dist-info/METADATA,sha256=KAOx0J5beZebrxsAf9AsklRO43eTWaw222WX1iInnpk,5299
-optimum_rbln-0.8.3a2.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-optimum_rbln-0.8.3a2.dist-info/licenses/LICENSE,sha256=QwcOLU5TJoTeUhuIXzhdCEEDDvorGiC6-3YTOl4TecE,11356
-optimum_rbln-0.8.3a2.dist-info/RECORD,,
+optimum_rbln-0.8.3a4.dist-info/METADATA,sha256=4ikxK2ldRi8PkKZrcSmTH4unPPzl3UZ3JBXlC-fr_MU,5299
+optimum_rbln-0.8.3a4.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+optimum_rbln-0.8.3a4.dist-info/licenses/LICENSE,sha256=QwcOLU5TJoTeUhuIXzhdCEEDDvorGiC6-3YTOl4TecE,11356
+optimum_rbln-0.8.3a4.dist-info/RECORD,,