optimum-rbln 0.7.5a0__py3-none-any.whl → 0.7.5a1__py3-none-any.whl

optimum/rbln/__init__.py

@@ -54,6 +54,12 @@ _import_structure = {
         "RBLNBertForQuestionAnsweringConfig",
         "RBLNBertModel",
         "RBLNBertModelConfig",
+        "RBLNBlip2VisionModelConfig",
+        "RBLNBlip2VisionModel",
+        "RBLNBlip2QFormerModel",
+        "RBLNBlip2QFormerModelConfig",
+        "RBLNBlip2ForConditionalGeneration",
+        "RBLNBlip2ForConditionalGenerationConfig",
         "RBLNCLIPTextModel",
         "RBLNCLIPTextModelConfig",
         "RBLNCLIPTextModelWithProjection",
@@ -80,6 +86,8 @@ _import_structure = {
         "RBLNIdefics3VisionTransformerConfig",
         "RBLNLlamaForCausalLM",
         "RBLNLlamaForCausalLMConfig",
+        "RBLNOPTForCausalLM",
+        "RBLNOPTForCausalLMConfig",
         "RBLNLlavaNextForConditionalGeneration",
         "RBLNLlavaNextForConditionalGenerationConfig",
         "RBLNMidmLMHeadModel",
@@ -100,6 +108,8 @@ _import_structure = {
         "RBLNRobertaForMaskedLMConfig",
         "RBLNRobertaForSequenceClassification",
         "RBLNRobertaForSequenceClassificationConfig",
+        "RBLNSiglipVisionModel",
+        "RBLNSiglipVisionModelConfig",
         "RBLNT5EncoderModel",
         "RBLNT5EncoderModelConfig",
         "RBLNT5ForConditionalGeneration",
@@ -265,6 +275,12 @@ if TYPE_CHECKING:
         RBLNBertForQuestionAnsweringConfig,
         RBLNBertModel,
         RBLNBertModelConfig,
+        RBLNBlip2ForConditionalGeneration,
+        RBLNBlip2ForConditionalGenerationConfig,
+        RBLNBlip2QFormerModel,
+        RBLNBlip2QFormerModelConfig,
+        RBLNBlip2VisionModel,
+        RBLNBlip2VisionModelConfig,
         RBLNCLIPTextModel,
         RBLNCLIPTextModelConfig,
         RBLNCLIPTextModelWithProjection,
@@ -297,6 +313,8 @@ if TYPE_CHECKING:
         RBLNMidmLMHeadModelConfig,
         RBLNMistralForCausalLM,
         RBLNMistralForCausalLMConfig,
+        RBLNOPTForCausalLM,
+        RBLNOPTForCausalLMConfig,
         RBLNPhiForCausalLM,
         RBLNPhiForCausalLMConfig,
         RBLNQwen2_5_VisionTransformerPretrainedModel,
@@ -311,6 +329,8 @@ if TYPE_CHECKING:
         RBLNRobertaForMaskedLMConfig,
         RBLNRobertaForSequenceClassification,
         RBLNRobertaForSequenceClassificationConfig,
+        RBLNSiglipVisionModel,
+        RBLNSiglipVisionModelConfig,
         RBLNT5EncoderModel,
         RBLNT5EncoderModelConfig,
         RBLNT5ForConditionalGeneration,

optimum/rbln/__version__.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.7.5a0'
-__version_tuple__ = version_tuple = (0, 7, 5, 'a0')
+__version__ = version = '0.7.5a1'
+__version_tuple__ = version_tuple = (0, 7, 5, 'a1')

optimum/rbln/configuration_utils.py

@@ -17,7 +17,7 @@ import inspect
 import json
 from dataclasses import asdict, dataclass
 from pathlib import Path
-from typing import Any, Dict, List, Optional, Tuple, Type, Union
+from typing import Any, Dict, List, Optional, Protocol, Tuple, Type, Union, runtime_checkable
 
 import torch
 
@@ -34,6 +34,11 @@ DEFAULT_MOD_NAME = "default"
 TypeInputInfo = List[Tuple[str, Tuple[int], str]]
 
 
+@runtime_checkable
+class RBLNSerializableConfigProtocol(Protocol):
+    def _prepare_for_serialization(self) -> Dict[str, Any]: ...
+
+
 @dataclass
 class RBLNCompileConfig:
     """
@@ -234,7 +239,7 @@ class RBLNAutoConfig:
             return cls(**config_file)
 
 
-class RBLNModelConfig:
+class RBLNModelConfig(RBLNSerializableConfigProtocol):
     """Base configuration class for RBLN models that handles compilation settings, runtime options, and submodules.
 
     This class provides functionality for:
@@ -594,14 +599,14 @@ class RBLNModelConfig:
             )
         return rbln_model_cls
 
-    def _prepare_for_serialization(self):
+    def _prepare_for_serialization(self) -> Dict[str, Any]:
         """
         Prepare the attributes map for serialization by converting nested RBLNModelConfig
         objects to their serializable form.
         """
         serializable_map = {}
         for key, value in self._attributes_map.items():
-            if isinstance(value, RBLNModelConfig):
+            if isinstance(value, RBLNSerializableConfigProtocol):
                 # Convert nested RBLNModelConfig to its serializable form
                 serializable_map[key] = value._prepare_for_serialization()
             elif key == "_compile_cfgs":

optimum/rbln/modeling.py

@@ -56,11 +56,7 @@ class RBLNModel(RBLNBaseModel):
     def update_kwargs(cls, kwargs):
         """
         Update user-given kwargs to get proper pytorch model.
-
-        For example, `torchscript`=True should be set because torch.jit
-        does not support `transformers` output instances as module output;
         """
-        kwargs.update({"torchscript": True})
         return kwargs
 
     @classmethod
@@ -133,7 +129,6 @@ class RBLNModel(RBLNBaseModel):
 
         if not isinstance(config, PretrainedConfig):  # diffusers config
             config = PretrainedConfig(**config)
-        config.save_pretrained(save_dir_path / subfolder)
 
         # Save preprocessor
         for preprocessor in preprocessors:
@@ -155,6 +150,10 @@ class RBLNModel(RBLNBaseModel):
             preprocessors=preprocessors, model=model, model_config=config, rbln_config=rbln_config
         )
 
+        # torchscript should be True for jit to work
+        torchscript_backup = config.torchscript
+        config.torchscript = True
+
         compiled_model: Union[rebel.RBLNCompiledModel, Dict[str, rebel.RBLNCompiledModel]] = cls.get_compiled_model(
             model, rbln_config=rbln_config
         )
@@ -169,6 +168,9 @@ class RBLNModel(RBLNBaseModel):
             cm.save(save_dir_path / subfolder / f"{compiled_model_name}.rbln")
         rbln_config.save(save_dir_path / subfolder)
 
+        config.torchscript = torchscript_backup
+        config.save_pretrained(save_dir_path / subfolder)
+
         # Save torch artifacts (e.g. embedding matrix if needed.)
         cls.save_torch_artifacts(model, save_dir_path=save_dir_path, subfolder=subfolder, rbln_config=rbln_config)
 

optimum/rbln/ops/__init__.py

@@ -16,3 +16,4 @@ from .attn import *
 from .flash_attn import *
 from .kv_cache_update import *
 from .linear import linear
+from .sliding_window_attn import *

optimum/rbln/ops/attn.py

@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from typing import Optional
 
 import torch
 from torch import Tensor
@@ -125,6 +126,7 @@ def paged_causal_attn_decode(
     scale: Tensor,
     block_table: Tensor,
     block_size: int,
+    mask: Optional[Tensor] = None,
 ) -> Tensor:
     """Defines the computation pattern for fused attention with KV cache updates.
 
@@ -147,6 +149,7 @@ def paged_causal_attn_decode(
     - scale: [] - Attention scale factor
     - block_table: [batch_size, max_seq_len // block_size] - Block indices for KV cache management
     - block_size: [] - Number of tokens per block
+    - mask: [batch=1, max_seq_len] - attention mask when use position_ids
 
     Returns:
         Tensor: attn_output: [batch=1, n_heads, n_groups, 1, head_dim] - Attention output
@@ -165,6 +168,7 @@ def paged_causal_attn_decode_fake(
     scale: Tensor,
     block_table: Tensor,
     block_size: int,
+    mask: Optional[Tensor] = None,
 ) -> Tensor:
     return torch.empty_like(q)
 
@@ -183,6 +187,8 @@ def paged_causal_attn_prefill(
     scale: Tensor,
     block_table: Tensor,
     block_size: int,
+    is_bidirectional: bool,
+    mask: Optional[Tensor] = None,
 ) -> Tensor:
     """Defines the computation pattern for prefill phase attention with KV cache updates.
 
@@ -204,6 +210,8 @@ def paged_causal_attn_prefill(
     - scale: [] - Attention scale factor
     - block_table: [batch_size, max_seq_len // block_size] - Block indices for KV cache management
     - block_size: [] - Number of tokens per block
+    - is_bidirectional: [] - Whether the attention is bidirectional at current sequence position
+    - mask: [batch=1, max_seq_len] - attention mask when use position_ids
 
     Returns:
         Tensor: attn_output: [batch=1, n_heads, n_groups, seq_len, head_dim] - Attention output
@@ -222,6 +230,8 @@ def paged_causal_attn_prefill_fake(
     scale: Tensor,
     block_table: Tensor,
     block_size: int,
+    is_bidirectional: bool,
+    mask: Optional[Tensor] = None,
 ) -> Tensor:
     return torch.empty_like(q)
 

optimum/rbln/ops/flash_attn.py

@@ -12,6 +12,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from typing import Optional
+
 import torch
 from torch import Tensor
 
@@ -113,6 +115,7 @@ def paged_flash_causal_attn_decode(
     block_table: Tensor,
     block_size: int,
     partition: int,
+    mask: Optional[Tensor] = None,
 ) -> Tensor:
     """Defines the computation pattern for fused causal flash attention with KV cache for decoding.
 
@@ -133,6 +136,7 @@ def paged_flash_causal_attn_decode_fake(
     block_table: Tensor,
     block_size: int,
     partition: int,
+    mask: Optional[Tensor] = None,
 ) -> Tensor:
     return torch.empty_like(q)
 
@@ -152,6 +156,8 @@ def paged_flash_causal_attn_prefill(
     block_table: Tensor,
     block_size: int,
     partition: int,
+    is_bidirectional: bool,
+    mask: Optional[Tensor] = None,
 ) -> Tensor:
     """Defines the computation pattern for fused causal flash attention with KV cache for prefill.
 
@@ -172,5 +178,7 @@ def paged_flash_causal_attn_prefill_fake(
     block_table: Tensor,
     block_size: int,
     partition: int,
+    is_bidirectional: bool,
+    mask: Optional[Tensor] = None,
 ) -> Tensor:
     return torch.empty_like(q)

optimum/rbln/ops/sliding_window_attn.py

@@ -0,0 +1,111 @@
+# 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 torch
+from torch import Tensor
+
+
+@torch.library.custom_op(
+    "rbln_custom_ops::paged_sliding_window_attn_prefill",
+    mutates_args=(["kcache", "vcache"]),
+)
+def paged_sliding_window_attn_prefill(
+    q: Tensor,
+    k: Tensor,
+    v: Tensor,
+    kcache: Tensor,
+    vcache: Tensor,
+    cache_seq_len: Tensor,
+    cache_offset: Tensor,
+    scale: Tensor,
+    block_table: Tensor,
+    block_size: int,
+    is_bidirectional: bool,
+) -> Tensor:
+    """Defines the computation pattern for prefill phase attention with KV cache updates.
+
+    IMPORTANT: This op serves as a pattern definition for the RBLN compiler to generate
+    a single optimized NPU operation. It is NOT meant for CPU execution.
+
+    Key differences from decode pattern:
+    - Handles prefill phase with multiple input tokens
+    - Takes explicit batch index for continuous batching
+
+    Expected tensor shapes:
+    - q: [batch=1, n_heads, n_groups, seq_len, head_dim] - Query states for multiple tokens
+    - k: [batch=1, n_heads, 1, seq_len, head_dim] - Key states for current input
+    - v: [batch=1, n_heads, 1, seq_len, head_dim] - Value states for current input
+    - kcache: [batch_size, n_heads, 1, max_seq_len, head_dim] - Key cache
+    - vcache: [batch_size, n_heads, 1, max_seq_len, head_dim] - Value cache
+    - cache_seq_len: [] - the sequence length of the cached states that were seen by the model
+    - cache_offset: [] - The valid length in the combined sequence of the KV cache and the current projected key states.
+    - scale: [] - Attention scale factor
+    - is_bidirectional: [] - Whether the attention is bidirectional
+    Returns:
+        Tensor: attn_output: [batch=1, n_heads, n_groups, seq_len, head_dim] - Attention output
+    """
+    return torch.empty_like(q)
+
+
+@paged_sliding_window_attn_prefill.register_fake
+def paged_sliding_window_attn_prefill_fake(
+    q: Tensor,
+    k: Tensor,
+    v: Tensor,
+    kcache: Tensor,
+    vcache: Tensor,
+    cache_seq_len: Tensor,
+    cache_offset: Tensor,
+    scale: Tensor,
+    block_table: Tensor,
+    block_size: int,
+    is_bidirectional: bool,
+) -> Tensor:
+    return torch.empty_like(q)
+
+
+@torch.library.custom_op(
+    "rbln_custom_ops::paged_sliding_window_attn_decode",
+    mutates_args=(["kcache", "vcache"]),
+)
+def paged_sliding_window_attn_decode(
+    q: Tensor,
+    k: Tensor,
+    v: Tensor,
+    kcache: Tensor,
+    vcache: Tensor,
+    cache_seq_len: Tensor,
+    cache_offset: Tensor,
+    scale: Tensor,
+    block_table: Tensor,
+    block_size: int,
+) -> Tensor:
+    return torch.empty_like(q)
+
+
+@paged_sliding_window_attn_decode.register_fake
+def paged_sliding_window_attn_decode_fake(
+    q: Tensor,
+    k: Tensor,
+    v: Tensor,
+    kcache: Tensor,
+    vcache: Tensor,
+    cache_seq_len: Tensor,
+    cache_offset: Tensor,
+    scale: Tensor,
+    block_table: Tensor,
+    block_size: int,
+) -> Tensor:
+    return torch.empty_like(q)

optimum/rbln/transformers/__init__.py

@@ -50,6 +50,12 @@ _import_structure = {
         "RBLNBertForQuestionAnsweringConfig",
         "RBLNBertModel",
         "RBLNBertModelConfig",
+        "RBLNBlip2VisionModelConfig",
+        "RBLNBlip2VisionModel",
+        "RBLNBlip2QFormerModel",
+        "RBLNBlip2QFormerModelConfig",
+        "RBLNBlip2ForConditionalGeneration",
+        "RBLNBlip2ForConditionalGenerationConfig",
         "RBLNCLIPTextModel",
         "RBLNCLIPTextModelConfig",
         "RBLNCLIPTextModelWithProjection",
@@ -74,6 +80,8 @@ _import_structure = {
         "RBLNIdefics3VisionTransformerConfig",
         "RBLNLlamaForCausalLM",
         "RBLNLlamaForCausalLMConfig",
+        "RBLNOPTForCausalLM",
+        "RBLNOPTForCausalLMConfig",
         "RBLNLlavaNextForConditionalGeneration",
         "RBLNLlavaNextForConditionalGenerationConfig",
         "RBLNMidmLMHeadModel",
@@ -88,17 +96,18 @@ _import_structure = {
         "RBLNQwen2_5_VisionTransformerPretrainedModelConfig",
         "RBLNQwen2_5_VLForConditionalGeneration",
         "RBLNQwen2_5_VLForConditionalGenerationConfig",
+        "RBLNSiglipVisionModel",
+        "RBLNSiglipVisionModelConfig",
         "RBLNT5EncoderModel",
         "RBLNT5EncoderModelConfig",
         "RBLNT5ForConditionalGeneration",
         "RBLNT5ForConditionalGenerationConfig",
+        "RBLNTimeSeriesTransformerForPrediction",
+        "RBLNTimeSeriesTransformerForPredictionConfig",
         "RBLNWav2Vec2ForCTC",
         "RBLNWav2Vec2ForCTCConfig",
         "RBLNWhisperForConditionalGeneration",
         "RBLNWhisperForConditionalGenerationConfig",
-        "RBLNTimeSeriesTransformerForPrediction",
-        "RBLNTimeSeriesTransformerForPredictionConfig",
-        "RBLNLlavaNextForConditionalGeneration",
         "RBLNXLMRobertaModel",
         "RBLNXLMRobertaModelConfig",
     ],
@@ -155,6 +164,12 @@ if TYPE_CHECKING:
         RBLNBertForQuestionAnsweringConfig,
         RBLNBertModel,
         RBLNBertModelConfig,
+        RBLNBlip2ForConditionalGeneration,
+        RBLNBlip2ForConditionalGenerationConfig,
+        RBLNBlip2QFormerModel,
+        RBLNBlip2QFormerModelConfig,
+        RBLNBlip2VisionModel,
+        RBLNBlip2VisionModelConfig,
         RBLNCLIPTextModel,
         RBLNCLIPTextModelConfig,
         RBLNCLIPTextModelWithProjection,
@@ -185,6 +200,8 @@ if TYPE_CHECKING:
         RBLNMidmLMHeadModelConfig,
         RBLNMistralForCausalLM,
         RBLNMistralForCausalLMConfig,
+        RBLNOPTForCausalLM,
+        RBLNOPTForCausalLMConfig,
         RBLNPhiForCausalLM,
         RBLNPhiForCausalLMConfig,
         RBLNQwen2_5_VisionTransformerPretrainedModel,
@@ -193,6 +210,8 @@ if TYPE_CHECKING:
         RBLNQwen2_5_VLForConditionalGenerationConfig,
         RBLNQwen2ForCausalLM,
         RBLNQwen2ForCausalLMConfig,
+        RBLNSiglipVisionModel,
+        RBLNSiglipVisionModelConfig,
         RBLNT5EncoderModel,
         RBLNT5EncoderModelConfig,
         RBLNT5ForConditionalGeneration,

optimum/rbln/transformers/models/__init__.py

@@ -46,6 +46,14 @@ _import_structure = {
         "RBLNBertForMaskedLM",
         "RBLNBertForMaskedLMConfig",
     ],
+    "blip_2": [
+        "RBLNBlip2VisionModelConfig",
+        "RBLNBlip2VisionModel",
+        "RBLNBlip2ForConditionalGeneration",
+        "RBLNBlip2ForConditionalGenerationConfig",
+        "RBLNBlip2QFormerModel",
+        "RBLNBlip2QFormerModelConfig",
+    ],
     "clip": [
         "RBLNCLIPTextModel",
         "RBLNCLIPTextModelConfig",
@@ -80,11 +88,16 @@ _import_structure = {
         "RBLNIdefics3VisionTransformerConfig",
     ],
     "llama": ["RBLNLlamaForCausalLM", "RBLNLlamaForCausalLMConfig"],
+    "opt": ["RBLNOPTForCausalLM", "RBLNOPTForCausalLMConfig"],
     "llava_next": ["RBLNLlavaNextForConditionalGeneration", "RBLNLlavaNextForConditionalGenerationConfig"],
     "midm": ["RBLNMidmLMHeadModel", "RBLNMidmLMHeadModelConfig"],
     "mistral": ["RBLNMistralForCausalLM", "RBLNMistralForCausalLMConfig"],
     "phi": ["RBLNPhiForCausalLM", "RBLNPhiForCausalLMConfig"],
     "qwen2": ["RBLNQwen2ForCausalLM", "RBLNQwen2ForCausalLMConfig"],
+    "siglip": [
+        "RBLNSiglipVisionModel",
+        "RBLNSiglipVisionModelConfig",
+    ],
     "time_series_transformers": [
         "RBLNTimeSeriesTransformerForPrediction",
         "RBLNTimeSeriesTransformerForPredictionConfig",
@@ -129,6 +142,14 @@ if TYPE_CHECKING:
         RBLNBertModel,
         RBLNBertModelConfig,
     )
+    from .blip_2 import (
+        RBLNBlip2ForConditionalGeneration,
+        RBLNBlip2ForConditionalGenerationConfig,
+        RBLNBlip2QFormerModel,
+        RBLNBlip2QFormerModelConfig,
+        RBLNBlip2VisionModel,
+        RBLNBlip2VisionModelConfig,
+    )
     from .clip import (
         RBLNCLIPTextModel,
         RBLNCLIPTextModelConfig,
@@ -160,6 +181,7 @@ if TYPE_CHECKING:
     from .llava_next import RBLNLlavaNextForConditionalGeneration, RBLNLlavaNextForConditionalGenerationConfig
     from .midm import RBLNMidmLMHeadModel, RBLNMidmLMHeadModelConfig
     from .mistral import RBLNMistralForCausalLM, RBLNMistralForCausalLMConfig
+    from .opt import RBLNOPTForCausalLM, RBLNOPTForCausalLMConfig
     from .phi import RBLNPhiForCausalLM, RBLNPhiForCausalLMConfig
     from .qwen2 import RBLNQwen2ForCausalLM, RBLNQwen2ForCausalLMConfig
     from .qwen2_5_vl import (
@@ -168,6 +190,7 @@ if TYPE_CHECKING:
         RBLNQwen2_5_VLForConditionalGeneration,
         RBLNQwen2_5_VLForConditionalGenerationConfig,
     )
+    from .siglip import RBLNSiglipVisionModel, RBLNSiglipVisionModelConfig
     from .t5 import (
         RBLNT5EncoderModel,
         RBLNT5EncoderModelConfig,

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

@@ -0,0 +1,20 @@
+# 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_blip_2 import (
+    RBLNBlip2ForConditionalGenerationConfig,
+    RBLNBlip2QFormerModelConfig,
+    RBLNBlip2VisionModelConfig,
+)
+from .modeling_blip_2 import RBLNBlip2ForConditionalGeneration, RBLNBlip2QFormerModel, RBLNBlip2VisionModel

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

@@ -0,0 +1,93 @@
+# 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 typing import Optional
+
+from ....configuration_utils import RBLNModelConfig
+
+
+class RBLNBlip2VisionModelConfig(RBLNModelConfig):
+    def __init__(
+        self,
+        batch_size: Optional[int] = None,
+        **kwargs,
+    ):
+        """
+        Args:
+            batch_size (Optional[int]): The batch size for inference. 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}")
+
+
+class RBLNBlip2QFormerModelConfig(RBLNModelConfig):
+    def __init__(
+        self,
+        batch_size: Optional[int] = None,
+        num_query_tokens: Optional[int] = None,
+        image_text_hidden_size: Optional[int] = None,
+        **kwargs,
+    ):
+        """
+        Args:
+            batch_size (Optional[int]): The batch size for inference. 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.num_query_tokens = num_query_tokens
+        self.image_text_hidden_size = image_text_hidden_size
+
+
+class RBLNBlip2ForConditionalGenerationConfig(RBLNModelConfig):
+    submodules = ["vision_model", "qformer", "language_model"]
+
+    def __init__(
+        self,
+        batch_size: Optional[int] = None,
+        vision_model: Optional[RBLNModelConfig] = None,
+        qformer: Optional[RBLNModelConfig] = None,
+        language_model: Optional[RBLNModelConfig] = None,
+        **kwargs,
+    ):
+        """
+        Args:
+            batch_size (Optional[int]): The batch size for inference. Defaults to 1.
+            vision_model (Optional[RBLNModelConfig]): Configuration for the vision encoder component.
+            language_model (Optional[RBLNModelConfig]): Configuration for the language model component.
+            **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.vision_model = self.init_submodule_config(RBLNBlip2VisionModelConfig, vision_model, batch_size=batch_size)
+        self.language_model = language_model
+        self.qformer = self.init_submodule_config(RBLNBlip2QFormerModelConfig, qformer, batch_size=batch_size)