optimum-rbln 0.8.1rc0__py3-none-any.whl → 0.8.2__py3-none-any.whl

optimum/rbln/diffusers/modeling_diffusers.py

@@ -70,8 +70,6 @@ class RBLNDiffusionMixin:
     _submodules = []
     _optional_submodules = []
     _prefix = {}
-    _rbln_config_class = None
-    _hf_class = None
 
     @staticmethod
     def _maybe_apply_and_fuse_lora(
@@ -114,14 +112,14 @@ class RBLNDiffusionMixin:
     @classmethod
     def get_rbln_config_class(cls) -> Type[RBLNModelConfig]:
         # Lazily loads and caches the corresponding RBLN model config class.
-        if cls._rbln_config_class is None:
+        if "_rbln_config_class" not in cls.__dict__ or cls._rbln_config_class is None:
             rbln_config_class_name = cls.__name__ + "Config"
             cls._rbln_config_class = get_rbln_config_class(rbln_config_class_name)
         return cls._rbln_config_class
 
     @classmethod
     def get_hf_class(cls):
-        if cls._hf_class is None:
+        if "_hf_class" not in cls.__dict__ or cls._hf_class is None:
             hf_cls_name = cls.__name__[4:]
             library = importlib.import_module("diffusers")
             cls._hf_class = getattr(library, hf_cls_name, None)
@@ -138,7 +136,7 @@ class RBLNDiffusionMixin:
         lora_ids: Optional[Union[str, List[str]]] = None,
         lora_weights_names: Optional[Union[str, List[str]]] = None,
         lora_scales: Optional[Union[float, List[float]]] = None,
-        **kwargs: Dict[str, Any],
+        **kwargs: Any,
     ) -> "RBLNDiffusionMixin":
         """
         Load a pretrained diffusion pipeline from a model checkpoint, with optional compilation for RBLN NPUs.
@@ -230,6 +228,7 @@ class RBLNDiffusionMixin:
             create_runtimes=rbln_config.create_runtimes,
             optimize_host_mem=rbln_config.optimize_host_memory,
             activate_profiler=rbln_config.activate_profiler,
+            timeout=rbln_config.timeout,
         ):
             model = super().from_pretrained(pretrained_model_name_or_path=model_id, **kwargs)
 

optimum/rbln/diffusers/models/__init__.py

@@ -35,20 +35,10 @@ _import_structure = {
 }
 
 if TYPE_CHECKING:
-    from .autoencoders import (
-        RBLNAutoencoderKL,
-        RBLNAutoencoderKLCosmos,
-        RBLNVQModel,
-    )
+    from .autoencoders import RBLNAutoencoderKL, RBLNAutoencoderKLCosmos, RBLNVQModel
     from .controlnet import RBLNControlNetModel
-    from .transformers import (
-        RBLNCosmosTransformer3DModel,
-        RBLNPriorTransformer,
-        RBLNSD3Transformer2DModel,
-    )
-    from .unets import (
-        RBLNUNet2DConditionModel,
-    )
+    from .transformers import RBLNCosmosTransformer3DModel, RBLNPriorTransformer, RBLNSD3Transformer2DModel
+    from .unets import RBLNUNet2DConditionModel
 else:
     import sys
 

optimum/rbln/diffusers/models/autoencoders/autoencoder_kl.py

@@ -209,6 +209,7 @@ class RBLNAutoencoderKL(RBLNModel):
                 tensor_type="pt",
                 device=device_val,
                 activate_profiler=rbln_config.activate_profiler,
+                timeout=rbln_config.timeout,
             )
             for compiled_model, device_val in zip(compiled_models, device_vals)
         ]

optimum/rbln/diffusers/models/autoencoders/autoencoder_kl_cosmos.py

@@ -200,6 +200,7 @@ class RBLNAutoencoderKLCosmos(RBLNModel):
                 tensor_type="pt",
                 device=device_val,
                 activate_profiler=rbln_config.activate_profiler,
+                timeout=rbln_config.timeout,
             )
             for compiled_model, device_val in zip(compiled_models, device_vals)
         ]

optimum/rbln/diffusers/models/autoencoders/vq_model.py

@@ -165,6 +165,7 @@ class RBLNVQModel(RBLNModel):
                 tensor_type="pt",
                 device=device_val,
                 activate_profiler=rbln_config.activate_profiler,
+                timeout=rbln_config.timeout,
             )
             for compiled_model, device_val in zip(compiled_models, device_vals)
         ]

optimum/rbln/diffusers/models/transformers/transformer_cosmos.py

@@ -279,7 +279,7 @@ class RBLNCosmosTransformer3DModel(RBLNModel):
                 tensor_type="pt",
                 device=rbln_config.device_map[DEFAULT_COMPILED_MODEL_NAME],
                 activate_profiler=rbln_config.activate_profiler,
-                timeout=120,
+                timeout=rbln_config.timeout,
             )
             for compiled_model in compiled_models
         ]

optimum/rbln/diffusers/pipelines/__init__.py

@@ -63,11 +63,7 @@ if TYPE_CHECKING:
         RBLNStableDiffusionXLControlNetImg2ImgPipeline,
         RBLNStableDiffusionXLControlNetPipeline,
     )
-    from .cosmos import (
-        RBLNCosmosSafetyChecker,
-        RBLNCosmosTextToWorldPipeline,
-        RBLNCosmosVideoToWorldPipeline,
-    )
+    from .cosmos import RBLNCosmosSafetyChecker, RBLNCosmosTextToWorldPipeline, RBLNCosmosVideoToWorldPipeline
     from .kandinsky2_2 import (
         RBLNKandinskyV22CombinedPipeline,
         RBLNKandinskyV22Img2ImgCombinedPipeline,

optimum/rbln/diffusers/pipelines/cosmos/configuration_cosmos_guardrail.py

@@ -12,10 +12,10 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import Any, Dict, Optional, Tuple
+from typing import Any, Optional, Tuple
 
 from ....configuration_utils import RBLNAutoConfig, RBLNModelConfig
-from ....transformers import RBLNSiglipVisionModelConfig
+from ....transformers import RBLNLlamaForCausalLMConfig, RBLNSiglipVisionModelConfig
 
 
 class RBLNVideoSafetyModelConfig(RBLNModelConfig):
@@ -69,13 +69,21 @@ class RBLNCosmosSafetyCheckerConfig(RBLNModelConfig):
         image_size: Optional[Tuple[int, int]] = None,
         height: Optional[int] = None,
         width: Optional[int] = None,
-        **kwargs: Dict[str, Any],
+        **kwargs: Any,
     ):
         super().__init__(**kwargs)
         if height is not None and width is not None:
             image_size = (height, width)
 
-        self.aegis = self.init_submodule_config(RBLNModelConfig, aegis)
+        tensor_parallel_size = kwargs.get("tensor_parallel_size")
+
+        self.aegis = self.init_submodule_config(
+            RBLNLlamaForCausalLMConfig,
+            aegis,
+            batch_size=batch_size,
+            tensor_parallel_size=tensor_parallel_size,
+        )
+
         self.siglip_encoder = self.init_submodule_config(
             RBLNSiglipVisionModelConfig,
             siglip_encoder,

optimum/rbln/diffusers/pipelines/cosmos/cosmos_guardrail.py

@@ -127,23 +127,13 @@ class RBLNSigLIPEncoder(SigLIPEncoder):
 
             # We don't use RBLNSiglipModel, but we need to override get_image_features to return pooler_output
             self.model = RBLNSiglipVisionModel.from_pretrained(
-                self.checkpoint_dir,
-                rbln_device=rbln_config.siglip_encoder.device,
-                rbln_create_runtimes=rbln_config.siglip_encoder.create_runtimes,
-                rbln_activate_profiler=rbln_config.aegis.activate_profiler,
+                self.checkpoint_dir, rbln_config=rbln_config.siglip_encoder
             )
         else:
             super().__init__(model_name, checkpoint_id)
             model = self.model
             del self.model
-            self.model = RBLNSiglipVisionModel.from_model(
-                model,
-                rbln_device=rbln_config.siglip_encoder.device,
-                rbln_image_size=rbln_config.siglip_encoder.image_size,
-                rbln_npu=rbln_config.siglip_encoder.npu,
-                rbln_create_runtimes=rbln_config.siglip_encoder.create_runtimes,
-                rbln_activate_profiler=rbln_config.siglip_encoder.activate_profiler,
-            )
+            self.model = RBLNSiglipVisionModel.from_model(model, rbln_config=rbln_config.siglip_encoder)
         self.rbln_config = rbln_config
 
         # Override get_image_features to return pooler_output
@@ -334,26 +324,14 @@ class RBLNAegis(Aegis):
             torch.nn.Module.__init__(self)
             cache_dir = pathlib.Path(checkpoint_id) / "aegis"
             self.tokenizer = AutoTokenizer.from_pretrained(cache_dir)
-            self.model = RBLNAutoModelForCausalLM.from_pretrained(
-                cache_dir,
-                rbln_device=rbln_config.aegis.device,
-                rbln_create_runtimes=rbln_config.aegis.create_runtimes,
-                rbln_activate_profiler=rbln_config.aegis.activate_profiler,
-            )
+            self.model = RBLNAutoModelForCausalLM.from_pretrained(cache_dir, rbln_config=rbln_config.aegis)
 
         else:
             super().__init__(checkpoint_id, base_model_id, aegis_adapter)
             model = self.model.merge_and_unload()  # peft merge
             del self.model
 
-            self.model = RBLNAutoModelForCausalLM.from_model(
-                model,
-                rbln_tensor_parallel_size=4,
-                rbln_device=rbln_config.aegis.device,
-                rbln_create_runtimes=rbln_config.aegis.create_runtimes,
-                rbln_npu=rbln_config.aegis.npu,
-                rbln_activate_profiler=rbln_config.aegis.activate_profiler,
-            )
+            self.model = RBLNAutoModelForCausalLM.from_model(model, rbln_config=rbln_config.aegis)
 
         self.rbln_config = rbln_config
         self.dtype = torch.bfloat16

optimum/rbln/diffusers/pipelines/cosmos/pipeline_cosmos_text2world.py

@@ -35,7 +35,7 @@ class RBLNCosmosTextToWorldPipeline(RBLNDiffusionMixin, CosmosTextToWorldPipelin
     RBLN-accelerated implementation of Cosmos Text to World pipeline for text-to-video generation.
 
     This pipeline compiles Cosmos Text to World models to run efficiently on RBLN NPUs, enabling high-performance
-    inference for generating images with distinctive artistic style and enhanced visual quality.
+    inference for generating videos with distinctive artistic style and enhanced visual quality.
     """
 
     original_class = CosmosTextToWorldPipeline
@@ -87,7 +87,7 @@ class RBLNCosmosTextToWorldPipeline(RBLNDiffusionMixin, CosmosTextToWorldPipelin
         export: bool = False,
         safety_checker: Optional[RBLNCosmosSafetyChecker] = None,
         rbln_config: Dict[str, Any] = {},
-        **kwargs: Dict[str, Any],
+        **kwargs: Any,
     ):
         rbln_config, kwargs = cls.get_rbln_config_class().initialize_from_kwargs(rbln_config, **kwargs)
         if safety_checker is None and export:

optimum/rbln/diffusers/pipelines/cosmos/pipeline_cosmos_video2world.py

@@ -35,7 +35,7 @@ class RBLNCosmosVideoToWorldPipeline(RBLNDiffusionMixin, CosmosVideoToWorldPipel
     RBLN-accelerated implementation of Cosmos Video to World pipeline for video-to-video generation.
 
     This pipeline compiles Cosmos Video to World models to run efficiently on RBLN NPUs, enabling high-performance
-    inference for generating images with distinctive artistic style and enhanced visual quality.
+    inference for generating videos with distinctive artistic style and enhanced visual quality.
     """
 
     original_class = CosmosVideoToWorldPipeline
@@ -87,7 +87,7 @@ class RBLNCosmosVideoToWorldPipeline(RBLNDiffusionMixin, CosmosVideoToWorldPipel
         export: bool = False,
         safety_checker: Optional[RBLNCosmosSafetyChecker] = None,
         rbln_config: Dict[str, Any] = {},
-        **kwargs: Dict[str, Any],
+        **kwargs: Any,
     ):
         rbln_config, kwargs = cls.get_rbln_config_class().initialize_from_kwargs(rbln_config, **kwargs)
         if safety_checker is None and export:

optimum/rbln/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_combined.py

@@ -22,12 +22,7 @@ from diffusers import (
     UNet2DConditionModel,
     VQModel,
 )
-from transformers import (
-    CLIPImageProcessor,
-    CLIPTextModelWithProjection,
-    CLIPTokenizer,
-    CLIPVisionModelWithProjection,
-)
+from transformers import CLIPImageProcessor, CLIPTextModelWithProjection, CLIPTokenizer, CLIPVisionModelWithProjection
 
 from ...configurations import RBLNKandinskyV22CombinedPipelineConfig
 from ...modeling_diffusers import RBLNDiffusionMixin

optimum/rbln/modeling.py

@@ -35,8 +35,6 @@ logger = get_logger(__name__)
 
 
 class RBLNModel(RBLNBaseModel):
-    _output_class = None
-
     @classmethod
     def update_kwargs(cls, kwargs):
         # Update user-given kwargs to get proper pytorch model.
@@ -80,7 +78,7 @@ class RBLNModel(RBLNBaseModel):
         rbln_config: Optional[Union[RBLNModelConfig, Dict]] = None,
         model_save_dir: Optional[Union[str, Path, TemporaryDirectory]] = None,
         subfolder: str = "",
-        **kwargs: Dict[str, Any],
+        **kwargs: Any,
     ) -> "RBLNModel":
         """
         Converts and compiles a pre-trained HuggingFace library model into a RBLN model.
@@ -238,11 +236,12 @@ class RBLNModel(RBLNBaseModel):
                 tensor_type="pt",
                 device=rbln_config.device_map[DEFAULT_COMPILED_MODEL_NAME],
                 activate_profiler=rbln_config.activate_profiler,
+                timeout=rbln_config.timeout,
             )
             for compiled_model in compiled_models
         ]
 
-    def forward(self, *args: Any, return_dict: Optional[bool] = None, **kwargs: Dict[str, Any]) -> Any:
+    def forward(self, *args: Any, return_dict: Optional[bool] = None, **kwargs: Any) -> Any:
         """
         Defines the forward pass of the RBLN model, providing a drop-in replacement for HuggingFace PreTrainedModel.
 
@@ -288,7 +287,7 @@ class RBLNModel(RBLNBaseModel):
     @classmethod
     def get_hf_output_class(cls):
         # Dynamically gets the output class from the corresponding HuggingFace model class.
-        if cls._output_class:
+        if "_output_class" in cls.__dict__ and cls._output_class is not None:
             return cls._output_class
 
         hf_class = cls.get_hf_class()

optimum/rbln/modeling_base.py

@@ -23,9 +23,10 @@ from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple, Type, Union
 import rebel
 import torch
 from transformers import AutoConfig, AutoModel, GenerationConfig, PretrainedConfig
+from transformers.utils.hub import PushToHubMixin
 
 from .configuration_utils import RBLNAutoConfig, RBLNCompileConfig, RBLNModelConfig, get_rbln_config_class
-from .utils.hub import PushToHubMixin, pull_compiled_model_from_hub, validate_files
+from .utils.hub import pull_compiled_model_from_hub, validate_files
 from .utils.logging import get_logger
 from .utils.runtime_utils import UnavailableRuntime, tp_and_devices_are_ok
 from .utils.save_utils import maybe_load_preprocessors
@@ -50,11 +51,8 @@ class RBLNBaseModel(SubModulesMixin, PushToHubMixin, PreTrainedModel):
     model_type = "rbln_model"
     auto_model_class = AutoModel
     config_class = AutoConfig
-
     config_name = "config.json"
     hf_library_name = "transformers"
-    _hf_class = None
-    _rbln_config_class = None
 
     def __init__(
         self,
@@ -115,7 +113,7 @@ class RBLNBaseModel(SubModulesMixin, PushToHubMixin, PreTrainedModel):
     def _load_compiled_model_dir(
         cls,
         model_id: Union[str, Path],
-        use_auth_token: Optional[Union[bool, str]] = None,
+        token: Optional[Union[bool, str]] = None,
         revision: Optional[str] = None,
         force_download: bool = False,
         cache_dir: Optional[str] = None,
@@ -134,7 +132,7 @@ class RBLNBaseModel(SubModulesMixin, PushToHubMixin, PreTrainedModel):
             model_path = pull_compiled_model_from_hub(
                 model_id=model_id,
                 subfolder=subfolder,
-                use_auth_token=use_auth_token,
+                token=token,
                 revision=revision,
                 cache_dir=cache_dir,
                 force_download=force_download,
@@ -172,7 +170,7 @@ class RBLNBaseModel(SubModulesMixin, PushToHubMixin, PreTrainedModel):
         cls,
         model_id: Union[str, Path],
         config: Optional["PretrainedConfig"] = None,
-        use_auth_token: Optional[Union[bool, str]] = None,
+        token: Optional[Union[bool, str]] = None,
         revision: Optional[str] = None,
         force_download: bool = False,
         cache_dir: Optional[str] = None,
@@ -189,7 +187,7 @@ class RBLNBaseModel(SubModulesMixin, PushToHubMixin, PreTrainedModel):
         if rbln_compiled_models is None:
             model_path_subfolder = cls._load_compiled_model_dir(
                 model_id=model_id,
-                use_auth_token=use_auth_token,
+                token=token,
                 revision=revision,
                 force_download=force_download,
                 cache_dir=cache_dir,
@@ -232,7 +230,7 @@ class RBLNBaseModel(SubModulesMixin, PushToHubMixin, PreTrainedModel):
                         cache_dir=cache_dir,
                         force_download=force_download,
                         revision=revision,
-                        token=use_auth_token,
+                        token=token,
                         trust_remote_code=trust_remote_code,
                     )
                 elif cls.hf_library_name == "diffusers":
@@ -250,7 +248,7 @@ class RBLNBaseModel(SubModulesMixin, PushToHubMixin, PreTrainedModel):
                         force_download=force_download,
                         local_files_only=local_files_only,
                         revision=revision,
-                        token=use_auth_token,
+                        token=token,
                         subfolder=subfolder,
                     )
                     config = PretrainedConfig(**config)
@@ -350,7 +348,7 @@ class RBLNBaseModel(SubModulesMixin, PushToHubMixin, PreTrainedModel):
         model_id: Union[str, Path],
         export: bool = False,
         rbln_config: Optional[Union[Dict, RBLNModelConfig]] = None,
-        **kwargs: Dict[str, Any],
+        **kwargs: Any,
     ) -> "RBLNBaseModel":
         """
         The `from_pretrained()` function is utilized in its standard form as in the HuggingFace transformers library.
@@ -421,7 +419,7 @@ class RBLNBaseModel(SubModulesMixin, PushToHubMixin, PreTrainedModel):
 
         # Returns:
         #     type: The original HuggingFace model class
-        if cls._hf_class is None:
+        if "_hf_class" not in cls.__dict__ or cls._hf_class is None:
             hf_cls_name = cls.__name__[4:]
             library = importlib.import_module(cls.hf_library_name)
             cls._hf_class = getattr(library, hf_cls_name, None)
@@ -430,7 +428,7 @@ class RBLNBaseModel(SubModulesMixin, PushToHubMixin, PreTrainedModel):
     @classmethod
     def get_rbln_config_class(cls) -> Type[RBLNModelConfig]:
         # Lazily loads and caches the corresponding RBLN model config class.
-        if cls._rbln_config_class is None:
+        if "_rbln_config_class" not in cls.__dict__ or cls._rbln_config_class is None:
             rbln_config_class_name = cls.__name__ + "Config"
             cls._rbln_config_class = get_rbln_config_class(rbln_config_class_name)
         return cls._rbln_config_class
@@ -507,6 +505,9 @@ class RBLNBaseModel(SubModulesMixin, PushToHubMixin, PreTrainedModel):
                 f"Please ensure the model directory exists and you have the necessary permissions to access it."
             )
 
+        if isinstance(self.config, PretrainedConfig):
+            self.config.save_pretrained(real_save_dir)
+
         if save_directory_path == real_save_dir:
             raise FileExistsError(
                 f"Cannot save model to '{save_directory}'. This directory already exists and contains the model files."
@@ -534,7 +535,10 @@ class RBLNBaseModel(SubModulesMixin, PushToHubMixin, PreTrainedModel):
             raise e  # Re-raise the exception after cleanup
 
         if push_to_hub:
-            return super().push_to_hub(str(save_directory_path), **kwargs)
+            repo_id = kwargs.pop("repo_id", None)
+            if repo_id is None:
+                raise ValueError("`repo_id` must be provided to push the model to the HuggingFace model hub.")
+            return super().push_to_hub(repo_id=repo_id, **kwargs)
 
     @staticmethod
     def _raise_missing_compiled_file_error(missing_files: List[str]):

optimum/rbln/ops/kv_cache_update.py

@@ -22,3 +22,8 @@ def rbln_cache_update(cache: Tensor, state: Tensor, position: Tensor, axis: Tens
     # This operation is designed to perform in-place updates directly on the device without needing to transfer the cache back to the host.
     # The `position` parameter specifies the start index for the update along the specified axis, allowing flexible updates to any part of the cache tensor.
     return torch.empty_like(cache)
+
+
+@rbln_cache_update.register_fake
+def rbln_cache_update_fake(cache: Tensor, state: Tensor, position: Tensor, axis: Tensor) -> Tensor:
+    return torch.empty_like(cache)

optimum/rbln/ops/linear.py

@@ -23,3 +23,10 @@ def linear(input: Tensor, weight: Tensor, bias: Optional[Tensor] = None) -> Tens
     output_shape = list(input.shape[:-1])
     output_shape += [weight.shape[0]]
     return torch.empty(size=output_shape, dtype=input.dtype, device=input.device, requires_grad=input.requires_grad)
+
+
+@linear.register_fake
+def linear_fake(input: Tensor, weight: Tensor, bias: Optional[Tensor] = None) -> Tensor:
+    output_shape = list(input.shape[:-1])
+    output_shape += [weight.shape[0]]
+    return torch.empty(size=output_shape, dtype=input.dtype, device=input.device, requires_grad=input.requires_grad)

optimum/rbln/transformers/__init__.py

@@ -62,12 +62,16 @@ _import_structure = {
         "RBLNCLIPVisionModelWithProjectionConfig",
         "RBLNDecoderOnlyModelForCausalLM",
         "RBLNDecoderOnlyModelForCausalLMConfig",
+        "RBLNDecoderOnlyModelConfig",
+        "RBLNDecoderOnlyModel",
         "RBLNDistilBertForQuestionAnswering",
         "RBLNDistilBertForQuestionAnsweringConfig",
         "RBLNDPTForDepthEstimation",
         "RBLNDPTForDepthEstimationConfig",
         "RBLNExaoneForCausalLM",
         "RBLNExaoneForCausalLMConfig",
+        "RBLNGemmaModel",
+        "RBLNGemmaModelConfig",
         "RBLNGemma3ForCausalLM",
         "RBLNGemma3ForCausalLMConfig",
         "RBLNGemma3ForConditionalGeneration",
@@ -76,28 +80,54 @@ _import_structure = {
         "RBLNGemmaForCausalLMConfig",
         "RBLNGPT2LMHeadModel",
         "RBLNGPT2LMHeadModelConfig",
+        "RBLNGPT2Model",
+        "RBLNGPT2ModelConfig",
         "RBLNIdefics3ForConditionalGeneration",
         "RBLNIdefics3ForConditionalGenerationConfig",
         "RBLNIdefics3VisionTransformer",
         "RBLNIdefics3VisionTransformerConfig",
         "RBLNLlamaForCausalLM",
         "RBLNLlamaForCausalLMConfig",
+        "RBLNLlavaForConditionalGeneration",
+        "RBLNLlavaForConditionalGenerationConfig",
+        "RBLNLlamaModel",
+        "RBLNLlamaModelConfig",
+        "RBLNOPTForCausalLM",
+        "RBLNOPTForCausalLMConfig",
+        "RBLNPegasusForConditionalGeneration",
+        "RBLNPegasusForConditionalGenerationConfig",
+        "RBLNPegasusModel",
+        "RBLNPegasusModelConfig",
         "RBLNLlavaNextForConditionalGeneration",
         "RBLNLlavaNextForConditionalGenerationConfig",
         "RBLNMidmLMHeadModel",
         "RBLNMidmLMHeadModelConfig",
         "RBLNMistralForCausalLM",
         "RBLNMistralForCausalLMConfig",
+        "RBLNMistralModel",
+        "RBLNMistralModelConfig",
         "RBLNOPTForCausalLM",
         "RBLNOPTForCausalLMConfig",
+        "RBLNOPTModel",
+        "RBLNOPTModelConfig",
         "RBLNPhiForCausalLM",
         "RBLNPhiForCausalLMConfig",
+        "RBLNPixtralVisionModelConfig",
+        "RBLNPixtralVisionModel",
+        "RBLNPhiModel",
+        "RBLNPhiModelConfig",
         "RBLNQwen2_5_VisionTransformerPretrainedModel",
         "RBLNQwen2_5_VisionTransformerPretrainedModelConfig",
         "RBLNQwen2_5_VLForConditionalGeneration",
         "RBLNQwen2_5_VLForConditionalGenerationConfig",
+        "RBLNQwen2Model",
+        "RBLNQwen2ModelConfig",
         "RBLNQwen2ForCausalLM",
         "RBLNQwen2ForCausalLMConfig",
+        "RBLNQwen3ForCausalLM",
+        "RBLNQwen3ForCausalLMConfig",
+        "RBLNQwen3Model",
+        "RBLNQwen3ModelConfig",
         "RBLNResNetForImageClassification",
         "RBLNResNetForImageClassificationConfig",
         "RBLNRobertaForMaskedLM",
@@ -166,6 +196,10 @@ if TYPE_CHECKING:
         RBLNCLIPVisionModelConfig,
         RBLNCLIPVisionModelWithProjection,
         RBLNCLIPVisionModelWithProjectionConfig,
+        RBLNColPaliForRetrieval,
+        RBLNColPaliForRetrievalConfig,
+        RBLNDecoderOnlyModel,
+        RBLNDecoderOnlyModelConfig,
         RBLNDecoderOnlyModelForCausalLM,
         RBLNDecoderOnlyModelForCausalLMConfig,
         RBLNDistilBertForQuestionAnswering,
@@ -180,30 +214,56 @@ if TYPE_CHECKING:
         RBLNGemma3ForConditionalGenerationConfig,
         RBLNGemmaForCausalLM,
         RBLNGemmaForCausalLMConfig,
+        RBLNGemmaModel,
+        RBLNGemmaModelConfig,
         RBLNGPT2LMHeadModel,
         RBLNGPT2LMHeadModelConfig,
+        RBLNGPT2Model,
+        RBLNGPT2ModelConfig,
         RBLNIdefics3ForConditionalGeneration,
         RBLNIdefics3ForConditionalGenerationConfig,
         RBLNIdefics3VisionTransformer,
         RBLNIdefics3VisionTransformerConfig,
         RBLNLlamaForCausalLM,
         RBLNLlamaForCausalLMConfig,
+        RBLNLlamaModel,
+        RBLNLlamaModelConfig,
+        RBLNLlavaForConditionalGeneration,
+        RBLNLlavaForConditionalGenerationConfig,
         RBLNLlavaNextForConditionalGeneration,
         RBLNLlavaNextForConditionalGenerationConfig,
         RBLNMidmLMHeadModel,
         RBLNMidmLMHeadModelConfig,
         RBLNMistralForCausalLM,
         RBLNMistralForCausalLMConfig,
+        RBLNMistralModel,
+        RBLNMistralModelConfig,
         RBLNOPTForCausalLM,
         RBLNOPTForCausalLMConfig,
+        RBLNOPTModel,
+        RBLNOPTModelConfig,
+        RBLNPegasusForConditionalGeneration,
+        RBLNPegasusForConditionalGenerationConfig,
+        RBLNPegasusModel,
+        RBLNPegasusModelConfig,
         RBLNPhiForCausalLM,
         RBLNPhiForCausalLMConfig,
+        RBLNPhiModel,
+        RBLNPhiModelConfig,
+        RBLNPixtralVisionModel,
+        RBLNPixtralVisionModelConfig,
         RBLNQwen2_5_VisionTransformerPretrainedModel,
         RBLNQwen2_5_VisionTransformerPretrainedModelConfig,
         RBLNQwen2_5_VLForConditionalGeneration,
         RBLNQwen2_5_VLForConditionalGenerationConfig,
         RBLNQwen2ForCausalLM,
         RBLNQwen2ForCausalLMConfig,
+        RBLNQwen2Model,
+        RBLNQwen2ModelConfig,
+        RBLNQwen3ForCausalLM,
+        RBLNQwen3ForCausalLMConfig,
+        RBLNQwen3Model,
+        RBLNQwen3ModelConfig,
         RBLNResNetForImageClassification,
         RBLNResNetForImageClassificationConfig,
         RBLNRobertaForMaskedLM,

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 +25,7 @@ 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],
+        **kwargs: Any,
     ):
         """
         Args:
@@ -52,7 +52,7 @@ 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 +124,7 @@ 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: