optimum-rbln 0.8.2a7__py3-none-any.whl → 0.8.3__py3-none-any.whl

optimum/rbln/ops/flash_attn.py

@@ -59,6 +59,47 @@ def paged_flash_attn_decode_fake(
     return torch.empty_like(q)
 
 
+@torch.library.custom_op(
+    "rbln_custom_ops::paged_flash_attn_decode_kv_fp8",
+    mutates_args=(["kcache", "vcache"]),
+)
+def paged_flash_attn_decode_kv_fp8(
+    q: Tensor,
+    k: Tensor,
+    v: Tensor,
+    mask: Tensor,
+    kcache: Tensor,
+    vcache: Tensor,
+    seq: Tensor,
+    scale: Tensor,
+    block_table: Tensor,
+    block_size: int,
+    partition: int,
+    k_scale: Tensor,
+    v_scale: Tensor,
+) -> Tensor:
+    return torch.empty_like(q)
+
+
+@paged_flash_attn_decode_kv_fp8.register_fake
+def paged_flash_attn_decode_kv_fp8_fake(
+    q: Tensor,
+    k: Tensor,
+    v: Tensor,
+    mask: Tensor,
+    kcache: Tensor,
+    vcache: Tensor,
+    seq: Tensor,
+    scale: Tensor,
+    block_table: Tensor,
+    block_size: int,
+    partition: int,
+    k_scale: Tensor,
+    v_scale: Tensor,
+) -> Tensor:
+    return torch.empty_like(q)
+
+
 @torch.library.custom_op(
     "rbln_custom_ops::paged_flash_attn_prefill",
     mutates_args=(["kcache", "vcache"]),
@@ -100,6 +141,47 @@ def paged_flash_attn_prefill_fake(
     return torch.empty_like(q)
 
 
+@torch.library.custom_op(
+    "rbln_custom_ops::paged_flash_attn_prefill_kv_fp8",
+    mutates_args=(["kcache", "vcache"]),
+)
+def paged_flash_attn_prefill_kv_fp8(
+    q: Tensor,
+    k: Tensor,
+    v: Tensor,
+    mask: Tensor,
+    kcache: Tensor,
+    vcache: Tensor,
+    seq: Tensor,
+    scale: Tensor,
+    block_table: Tensor,
+    block_size: int,
+    partition: int,
+    k_scale: Tensor,
+    v_scale: Tensor,
+) -> Tensor:
+    return torch.empty_like(q)
+
+
+@paged_flash_attn_prefill_kv_fp8.register_fake
+def paged_flash_attn_prefill_kv_fp8_fake(
+    q: Tensor,
+    k: Tensor,
+    v: Tensor,
+    mask: Tensor,
+    kcache: Tensor,
+    vcache: Tensor,
+    seq: Tensor,
+    scale: Tensor,
+    block_table: Tensor,
+    block_size: int,
+    partition: int,
+    k_scale: Tensor,
+    v_scale: Tensor,
+) -> Tensor:
+    return torch.empty_like(q)
+
+
 @torch.library.custom_op(
     "rbln_custom_ops::paged_flash_causal_attn_decode",
     mutates_args=(["kcache", "vcache"]),
@@ -141,6 +223,47 @@ def paged_flash_causal_attn_decode_fake(
     return torch.empty_like(q)
 
 
+@torch.library.custom_op(
+    "rbln_custom_ops::paged_flash_causal_attn_decode_kv_fp8",
+    mutates_args=(["kcache", "vcache"]),
+)
+def paged_flash_causal_attn_decode_kv_fp8(
+    q: Tensor,
+    k: Tensor,
+    v: Tensor,
+    kcache: Tensor,
+    vcache: Tensor,
+    seq: Tensor,
+    scale: Tensor,
+    block_table: Tensor,
+    block_size: int,
+    partition: int,
+    k_scale: Tensor,
+    v_scale: Tensor,
+    mask: Optional[Tensor] = None,
+) -> Tensor:
+    return torch.empty_like(q)
+
+
+@paged_flash_causal_attn_decode_kv_fp8.register_fake
+def paged_flash_causal_attn_decode_kv_fp8_fake(
+    q: Tensor,
+    k: Tensor,
+    v: Tensor,
+    kcache: Tensor,
+    vcache: Tensor,
+    seq: Tensor,
+    scale: Tensor,
+    block_table: Tensor,
+    block_size: int,
+    partition: int,
+    k_scale: Tensor,
+    v_scale: Tensor,
+    mask: Optional[Tensor] = None,
+) -> Tensor:
+    return torch.empty_like(q)
+
+
 @torch.library.custom_op(
     "rbln_custom_ops::paged_flash_causal_attn_prefill",
     mutates_args=(["kcache", "vcache"]),
@@ -182,3 +305,46 @@ def paged_flash_causal_attn_prefill_fake(
     mask: Optional[Tensor] = None,
 ) -> Tensor:
     return torch.empty_like(q)
+
+
+@torch.library.custom_op(
+    "rbln_custom_ops::paged_flash_causal_attn_prefill_kv_fp8",
+    mutates_args=(["kcache", "vcache"]),
+)
+def paged_flash_causal_attn_prefill_kv_fp8(
+    q: Tensor,
+    k: Tensor,
+    v: Tensor,
+    kcache: Tensor,
+    vcache: Tensor,
+    seq: Tensor,
+    scale: Tensor,
+    block_table: Tensor,
+    block_size: int,
+    partition: int,
+    is_bidirectional: bool,
+    k_scale: Tensor,
+    v_scale: Tensor,
+    mask: Optional[Tensor] = None,
+) -> Tensor:
+    return torch.empty_like(q)
+
+
+@paged_flash_causal_attn_prefill_kv_fp8.register_fake
+def paged_flash_causal_attn_prefill_kv_fp8_fake(
+    q: Tensor,
+    k: Tensor,
+    v: Tensor,
+    kcache: Tensor,
+    vcache: Tensor,
+    seq: Tensor,
+    scale: Tensor,
+    block_table: Tensor,
+    block_size: int,
+    partition: int,
+    is_bidirectional: bool,
+    k_scale: Tensor,
+    v_scale: Tensor,
+    mask: Optional[Tensor] = None,
+) -> Tensor:
+    return torch.empty_like(q)

optimum/rbln/transformers/__init__.py

@@ -34,6 +34,8 @@ _import_structure = {
         "RBLNAutoModelForSequenceClassification",
         "RBLNAutoModelForSpeechSeq2Seq",
         "RBLNAutoModelForVision2Seq",
+        "RBLNAutoModelForTextEncoding",
+        "RBLNAutoModelForZeroShotObjectDetection",
         "RBLNBartForConditionalGeneration",
         "RBLNBartForConditionalGenerationConfig",
         "RBLNBartModel",
@@ -62,10 +64,14 @@ _import_structure = {
         "RBLNCLIPVisionModelWithProjectionConfig",
         "RBLNDecoderOnlyModelForCausalLM",
         "RBLNDecoderOnlyModelForCausalLMConfig",
+        "RBLNDecoderOnlyModelConfig",
+        "RBLNDecoderOnlyModel",
         "RBLNDistilBertForQuestionAnswering",
         "RBLNDistilBertForQuestionAnsweringConfig",
         "RBLNDPTForDepthEstimation",
         "RBLNDPTForDepthEstimationConfig",
+        "RBLNDepthAnythingForDepthEstimation",
+        "RBLNDepthAnythingForDepthEstimationConfig",
         "RBLNExaoneForCausalLM",
         "RBLNExaoneForCausalLMConfig",
         "RBLNGemmaModel",
@@ -80,6 +86,12 @@ _import_structure = {
         "RBLNGPT2LMHeadModelConfig",
         "RBLNGPT2Model",
         "RBLNGPT2ModelConfig",
+        "RBLNGroundingDinoDecoder",
+        "RBLNGroundingDinoDecoderConfig",
+        "RBLNGroundingDinoForObjectDetection",
+        "RBLNGroundingDinoForObjectDetectionConfig",
+        "RBLNGroundingDinoEncoder",
+        "RBLNGroundingDinoEncoderConfig",
         "RBLNIdefics3ForConditionalGeneration",
         "RBLNIdefics3ForConditionalGenerationConfig",
         "RBLNIdefics3VisionTransformer",
@@ -134,6 +146,8 @@ _import_structure = {
         "RBLNRobertaForSequenceClassificationConfig",
         "RBLNSiglipVisionModel",
         "RBLNSiglipVisionModelConfig",
+        "RBLNSwinBackbone",
+        "RBLNSwinBackboneConfig",
         "RBLNT5EncoderModel",
         "RBLNT5EncoderModelConfig",
         "RBLNT5ForConditionalGeneration",
@@ -169,7 +183,9 @@ if TYPE_CHECKING:
         RBLNAutoModelForSeq2SeqLM,
         RBLNAutoModelForSequenceClassification,
         RBLNAutoModelForSpeechSeq2Seq,
+        RBLNAutoModelForTextEncoding,
         RBLNAutoModelForVision2Seq,
+        RBLNAutoModelForZeroShotObjectDetection,
         RBLNBartForConditionalGeneration,
         RBLNBartForConditionalGenerationConfig,
         RBLNBartModel,
@@ -196,8 +212,12 @@ if TYPE_CHECKING:
         RBLNCLIPVisionModelWithProjectionConfig,
         RBLNColPaliForRetrieval,
         RBLNColPaliForRetrievalConfig,
+        RBLNDecoderOnlyModel,
+        RBLNDecoderOnlyModelConfig,
         RBLNDecoderOnlyModelForCausalLM,
         RBLNDecoderOnlyModelForCausalLMConfig,
+        RBLNDepthAnythingForDepthEstimation,
+        RBLNDepthAnythingForDepthEstimationConfig,
         RBLNDistilBertForQuestionAnswering,
         RBLNDistilBertForQuestionAnsweringConfig,
         RBLNDPTForDepthEstimation,
@@ -216,6 +236,12 @@ if TYPE_CHECKING:
         RBLNGPT2LMHeadModelConfig,
         RBLNGPT2Model,
         RBLNGPT2ModelConfig,
+        RBLNGroundingDinoDecoder,
+        RBLNGroundingDinoDecoderConfig,
+        RBLNGroundingDinoEncoder,
+        RBLNGroundingDinoEncoderConfig,
+        RBLNGroundingDinoForObjectDetection,
+        RBLNGroundingDinoForObjectDetectionConfig,
         RBLNIdefics3ForConditionalGeneration,
         RBLNIdefics3ForConditionalGenerationConfig,
         RBLNIdefics3VisionTransformer,
@@ -268,6 +294,8 @@ if TYPE_CHECKING:
         RBLNRobertaForSequenceClassificationConfig,
         RBLNSiglipVisionModel,
         RBLNSiglipVisionModelConfig,
+        RBLNSwinBackbone,
+        RBLNSwinBackboneConfig,
         RBLNT5EncoderModel,
         RBLNT5EncoderModelConfig,
         RBLNT5ForConditionalGeneration,

optimum/rbln/transformers/configuration_generic.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import Any, Dict, List, Optional, Tuple, Union
+from typing import Any, List, Optional, Tuple, Union
 
 from ..configuration_utils import RBLNModelConfig
 
@@ -25,7 +25,8 @@ class RBLNTransformerEncoderConfig(RBLNModelConfig):
         max_seq_len: Optional[int] = None,
         batch_size: Optional[int] = None,
         model_input_names: Optional[List[str]] = None,
-        **kwargs: Dict[str, Any],
+        model_input_shapes: Optional[List[Tuple[int, int]]] = None,
+        **kwargs: Any,
     ):
         """
         Args:
@@ -45,6 +46,7 @@ class RBLNTransformerEncoderConfig(RBLNModelConfig):
             raise ValueError(f"batch_size must be a positive integer, got {self.batch_size}")
 
         self.model_input_names = model_input_names or self.rbln_model_input_names
+        self.model_input_shapes = model_input_shapes
 
 
 class RBLNImageModelConfig(RBLNModelConfig):
@@ -52,7 +54,7 @@ class RBLNImageModelConfig(RBLNModelConfig):
         self,
         image_size: Optional[Union[int, Tuple[int, int]]] = None,
         batch_size: Optional[int] = None,
-        **kwargs: Dict[str, Any],
+        **kwargs: Any,
     ):
         """
         Args:
@@ -124,7 +126,7 @@ class RBLNModelForAudioClassificationConfig(RBLNModelConfig):
         batch_size: Optional[int] = None,
         max_length: Optional[int] = None,
         num_mel_bins: Optional[int] = None,
-        **kwargs: Dict[str, Any],
+        **kwargs: Any,
     ):
         """
         Args:

optimum/rbln/transformers/modeling_generic.py

@@ -34,10 +34,7 @@ from transformers import (
     AutoModelForTextEncoding,
     PretrainedConfig,
 )
-from transformers.modeling_outputs import (
-    BaseModelOutput,
-    QuestionAnsweringModelOutput,
-)
+from transformers.modeling_outputs import BaseModelOutput, QuestionAnsweringModelOutput
 
 from ..configuration_utils import RBLNCompileConfig
 from ..modeling import RBLNModel
@@ -130,10 +127,18 @@ class RBLNTransformerEncoder(RBLNModel):
                 "This is an internal error. Please report it to the developers."
             )
 
-        input_info = [
-            (model_input_name, [rbln_config.batch_size, rbln_config.max_seq_len], cls.rbln_dtype)
-            for model_input_name in rbln_config.model_input_names
-        ]
+        if rbln_config.model_input_shapes is None:
+            input_info = [
+                (model_input_name, [rbln_config.batch_size, rbln_config.max_seq_len], cls.rbln_dtype)
+                for model_input_name in rbln_config.model_input_names
+            ]
+        else:
+            input_info = [
+                (model_input_name, model_input_shape, cls.rbln_dtype)
+                for model_input_name, model_input_shape in zip(
+                    rbln_config.model_input_names, rbln_config.model_input_shapes
+                )
+            ]
 
         rbln_config.set_compile_cfgs([RBLNCompileConfig(input_info=input_info)])
         return rbln_config

optimum/rbln/transformers/modeling_outputs.py

@@ -0,0 +1,37 @@
+# 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 dataclasses import dataclass
+from typing import Optional, Tuple
+
+import torch
+from transformers.modeling_outputs import ModelOutput
+
+
+@dataclass
+class RBLNDecoderOnlyOutput(ModelOutput):
+    logits: torch.FloatTensor = None
+    generate_idx: torch.Tensor = None
+    padded_cache_lengths: int = None
+
+
+@dataclass
+class RBLNGemma3ForCausalLMOutput(RBLNDecoderOnlyOutput):
+    attention_mask: Optional[torch.Tensor] = None
+
+
+@dataclass
+class RBLNSeq2SeqTSDecoderOutput(ModelOutput):
+    last_hidden_states: torch.FloatTensor = None
+    params: Tuple[torch.FloatTensor] = None

optimum/rbln/transformers/models/__init__.py

@@ -36,6 +36,8 @@ _import_structure = {
         "RBLNAutoModelForSpeechSeq2Seq",
         "RBLNAutoModelForVision2Seq",
         "RBLNAutoModelForImageTextToText",
+        "RBLNAutoModelForTextEncoding",
+        "RBLNAutoModelForZeroShotObjectDetection",
     ],
     "bart": [
         "RBLNBartForConditionalGeneration",
@@ -84,9 +86,12 @@ _import_structure = {
         "RBLNQwen2_5_VLForConditionalGenerationConfig",
     ],
     "decoderonly": [
+        "RBLNDecoderOnlyModelConfig",
+        "RBLNDecoderOnlyModel",
         "RBLNDecoderOnlyModelForCausalLM",
         "RBLNDecoderOnlyModelForCausalLMConfig",
     ],
+    "depth_anything": ["RBLNDepthAnythingForDepthEstimationConfig", "RBLNDepthAnythingForDepthEstimation"],
     "dpt": [
         "RBLNDPTForDepthEstimation",
         "RBLNDPTForDepthEstimationConfig",
@@ -138,6 +143,10 @@ _import_structure = {
         "RBLNSiglipVisionModel",
         "RBLNSiglipVisionModelConfig",
     ],
+    "swin": [
+        "RBLNSwinBackbone",
+        "RBLNSwinBackboneConfig",
+    ],
     "time_series_transformer": [
         "RBLNTimeSeriesTransformerForPrediction",
         "RBLNTimeSeriesTransformerForPredictionConfig",
@@ -157,13 +166,18 @@ _import_structure = {
         "RBLNXLMRobertaForSequenceClassification",
         "RBLNXLMRobertaForSequenceClassificationConfig",
     ],
+    "grounding_dino": [
+        "RBLNGroundingDinoForObjectDetection",
+        "RBLNGroundingDinoForObjectDetectionConfig",
+        "RBLNGroundingDinoEncoder",
+        "RBLNGroundingDinoEncoderConfig",
+        "RBLNGroundingDinoDecoder",
+        "RBLNGroundingDinoDecoderConfig",
+    ],
 }
 
 if TYPE_CHECKING:
-    from .audio_spectrogram_transformer import (
-        RBLNASTForAudioClassification,
-        RBLNASTForAudioClassificationConfig,
-    )
+    from .audio_spectrogram_transformer import RBLNASTForAudioClassification, RBLNASTForAudioClassificationConfig
     from .auto import (
         RBLNAutoModel,
         RBLNAutoModelForAudioClassification,
@@ -177,7 +191,9 @@ if TYPE_CHECKING:
         RBLNAutoModelForSeq2SeqLM,
         RBLNAutoModelForSequenceClassification,
         RBLNAutoModelForSpeechSeq2Seq,
+        RBLNAutoModelForTextEncoding,
         RBLNAutoModelForVision2Seq,
+        RBLNAutoModelForZeroShotObjectDetection,
     )
     from .bart import (
         RBLNBartForConditionalGeneration,
@@ -211,22 +227,16 @@ if TYPE_CHECKING:
         RBLNCLIPVisionModelWithProjection,
         RBLNCLIPVisionModelWithProjectionConfig,
     )
-    from .colpali import (
-        RBLNColPaliForRetrieval,
-        RBLNColPaliForRetrievalConfig,
-    )
+    from .colpali import RBLNColPaliForRetrieval, RBLNColPaliForRetrievalConfig
     from .decoderonly import (
+        RBLNDecoderOnlyModel,
+        RBLNDecoderOnlyModelConfig,
         RBLNDecoderOnlyModelForCausalLM,
         RBLNDecoderOnlyModelForCausalLMConfig,
     )
-    from .distilbert import (
-        RBLNDistilBertForQuestionAnswering,
-        RBLNDistilBertForQuestionAnsweringConfig,
-    )
-    from .dpt import (
-        RBLNDPTForDepthEstimation,
-        RBLNDPTForDepthEstimationConfig,
-    )
+    from .depth_anything import RBLNDepthAnythingForDepthEstimation, RBLNDepthAnythingForDepthEstimationConfig
+    from .distilbert import RBLNDistilBertForQuestionAnswering, RBLNDistilBertForQuestionAnsweringConfig
+    from .dpt import RBLNDPTForDepthEstimation, RBLNDPTForDepthEstimationConfig
     from .exaone import RBLNExaoneForCausalLM, RBLNExaoneForCausalLMConfig
     from .gemma import RBLNGemmaForCausalLM, RBLNGemmaForCausalLMConfig, RBLNGemmaModel, RBLNGemmaModelConfig
     from .gemma3 import (
@@ -236,6 +246,14 @@ if TYPE_CHECKING:
         RBLNGemma3ForConditionalGenerationConfig,
     )
     from .gpt2 import RBLNGPT2LMHeadModel, RBLNGPT2LMHeadModelConfig, RBLNGPT2Model, RBLNGPT2ModelConfig
+    from .grounding_dino import (
+        RBLNGroundingDinoDecoder,
+        RBLNGroundingDinoDecoderConfig,
+        RBLNGroundingDinoEncoder,
+        RBLNGroundingDinoEncoderConfig,
+        RBLNGroundingDinoForObjectDetection,
+        RBLNGroundingDinoForObjectDetectionConfig,
+    )
     from .idefics3 import (
         RBLNIdefics3ForConditionalGeneration,
         RBLNIdefics3ForConditionalGenerationConfig,
@@ -272,6 +290,7 @@ if TYPE_CHECKING:
         RBLNRobertaForSequenceClassificationConfig,
     )
     from .siglip import RBLNSiglipVisionModel, RBLNSiglipVisionModelConfig
+    from .swin import RBLNSwinBackbone, RBLNSwinBackboneConfig
     from .t5 import (
         RBLNT5EncoderModel,
         RBLNT5EncoderModelConfig,

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

@@ -25,5 +25,7 @@ from .modeling_auto import (
     RBLNAutoModelForSeq2SeqLM,
     RBLNAutoModelForSequenceClassification,
     RBLNAutoModelForSpeechSeq2Seq,
+    RBLNAutoModelForTextEncoding,
     RBLNAutoModelForVision2Seq,
+    RBLNAutoModelForZeroShotObjectDetection,
 )

optimum/rbln/transformers/models/auto/modeling_auto.py

@@ -35,8 +35,12 @@ from transformers.models.auto.modeling_auto import (
     MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES,
     MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING,
     MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING_NAMES,
+    MODEL_FOR_TEXT_ENCODING_MAPPING,
+    MODEL_FOR_TEXT_ENCODING_MAPPING_NAMES,
     MODEL_FOR_VISION_2_SEQ_MAPPING,
     MODEL_FOR_VISION_2_SEQ_MAPPING_NAMES,
+    MODEL_FOR_ZERO_SHOT_OBJECT_DETECTION_MAPPING,
+    MODEL_FOR_ZERO_SHOT_OBJECT_DETECTION_MAPPING_NAMES,
     MODEL_MAPPING,
     MODEL_MAPPING_NAMES,
 )
@@ -115,3 +119,13 @@ class RBLNAutoModelForImageClassification(_BaseAutoModelClass):
 class RBLNAutoModelForQuestionAnswering(_BaseAutoModelClass):
     _model_mapping = MODEL_FOR_QUESTION_ANSWERING_MAPPING
     _model_mapping_names = MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES
+
+
+class RBLNAutoModelForTextEncoding(_BaseAutoModelClass):
+    _model_mapping = MODEL_FOR_TEXT_ENCODING_MAPPING
+    _model_mapping_names = MODEL_FOR_TEXT_ENCODING_MAPPING_NAMES
+
+
+class RBLNAutoModelForZeroShotObjectDetection(_BaseAutoModelClass):
+    _model_mapping = MODEL_FOR_ZERO_SHOT_OBJECT_DETECTION_MAPPING
+    _model_mapping_names = MODEL_FOR_ZERO_SHOT_OBJECT_DETECTION_MAPPING_NAMES

optimum/rbln/transformers/models/bart/bart_architecture.py

@@ -16,9 +16,7 @@ from typing import Tuple
 
 import torch
 from torch import nn
-from transformers.modeling_attn_mask_utils import (
-    _prepare_4d_attention_mask,
-)
+from transformers.modeling_attn_mask_utils import _prepare_4d_attention_mask
 from transformers.utils import logging
 
 from ..seq2seq.seq2seq_architecture import (

optimum/rbln/transformers/models/bart/configuration_bart.py

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

optimum/rbln/transformers/models/bert/bert_architecture.py

@@ -0,0 +1,16 @@
+import torch
+
+
+class BertModelWrapper(torch.nn.Module):
+    def __init__(self, model, rbln_config):
+        super().__init__()
+        self.model = model
+        self.rbln_config = rbln_config
+
+    def forward(self, *args, **kwargs):
+        output = self.model(*args, **kwargs)
+        if isinstance(output, torch.Tensor):
+            return output
+        elif isinstance(output, tuple):
+            return tuple(x for x in output if x is not None)
+        return output

optimum/rbln/transformers/models/bert/modeling_bert.py

@@ -12,15 +12,15 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ....utils.logging import get_logger
+import torch
+
 from ...modeling_generic import (
     RBLNModelForMaskedLM,
     RBLNModelForQuestionAnswering,
     RBLNTransformerEncoderForFeatureExtraction,
 )
-
-
-logger = get_logger(__name__)
+from .bert_architecture import BertModelWrapper
+from .configuration_bert import RBLNBertModelConfig
 
 
 class RBLNBertModel(RBLNTransformerEncoderForFeatureExtraction):
@@ -34,6 +34,10 @@ class RBLNBertModel(RBLNTransformerEncoderForFeatureExtraction):
 
     rbln_model_input_names = ["input_ids", "attention_mask"]
 
+    @classmethod
+    def wrap_model_if_needed(cls, model: torch.nn.Module, rbln_config: RBLNBertModelConfig) -> torch.nn.Module:
+        return BertModelWrapper(model, rbln_config)
+
 
 class RBLNBertForMaskedLM(RBLNModelForMaskedLM):
     """

optimum/rbln/transformers/models/blip_2/configuration_blip_2.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import Any, Dict, Optional
+from typing import Any, Optional
 
 from ....configuration_utils import RBLNModelConfig
 
@@ -62,7 +62,7 @@ class RBLNBlip2ForConditionalGenerationConfig(RBLNModelConfig):
         vision_model: Optional[RBLNModelConfig] = None,
         qformer: Optional[RBLNModelConfig] = None,
         language_model: Optional[RBLNModelConfig] = None,
-        **kwargs: Dict[str, Any],
+        **kwargs: Any,
     ):
         """
         Args:

optimum/rbln/transformers/models/blip_2/modeling_blip_2.py

@@ -35,11 +35,7 @@ from ....modeling import RBLNModel
 logger = logging.get_logger(__name__)
 
 if TYPE_CHECKING:
-    from transformers import (
-        AutoFeatureExtractor,
-        AutoProcessor,
-        AutoTokenizer,
-    )
+    from transformers import AutoFeatureExtractor, AutoProcessor, AutoTokenizer
 
 
 class LoopProjector:
@@ -178,7 +174,12 @@ class RBLNBlip2QFormerModel(RBLNModel):
         return Blip2QFormerModelWrapper(model).eval()
 
     @classmethod
-    def _update_submodule_config(cls, model: "PreTrainedModel", rbln_config: "RBLNModelConfig") -> "RBLNModelConfig":
+    def _update_submodule_config(
+        cls,
+        model: "PreTrainedModel",
+        rbln_config: RBLNModelConfig,
+        preprocessors: Optional[Union["AutoFeatureExtractor", "AutoProcessor", "AutoTokenizer"]],
+    ):
         if rbln_config.num_query_tokens is None:
             rbln_config.num_query_tokens = model.config.num_query_tokens
 

optimum/rbln/transformers/models/clip/configuration_clip.py

@@ -12,13 +12,13 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import Any, Dict, Optional
+from typing import Any, Optional
 
 from ....configuration_utils import RBLNModelConfig
 
 
 class RBLNCLIPTextModelConfig(RBLNModelConfig):
-    def __init__(self, batch_size: Optional[int] = None, **kwargs: Dict[str, Any]):
+    def __init__(self, batch_size: Optional[int] = None, **kwargs: Any):
         """
         Args:
             batch_size (Optional[int]): The batch size for text processing. Defaults to 1.
@@ -50,7 +50,7 @@ class RBLNCLIPVisionModelConfig(RBLNModelConfig):
         interpolate_pos_encoding: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         output_attentions: Optional[bool] = None,
-        **kwargs: Dict[str, Any],
+        **kwargs: Any,
     ):
         """
         Args: