optimum-rbln 0.9.2a4__py3-none-any.whl → 0.9.2a5__py3-none-any.whl

optimum/rbln/__version__.py

@@ -28,7 +28,7 @@ version_tuple: VERSION_TUPLE
 commit_id: COMMIT_ID
 __commit_id__: COMMIT_ID
 
-__version__ = version = '0.9.2a4'
-__version_tuple__ = version_tuple = (0, 9, 2, 'a4')
+__version__ = version = '0.9.2a5'
+__version_tuple__ = version_tuple = (0, 9, 2, 'a5')
 
 __commit_id__ = commit_id = None

optimum/rbln/modeling.py

@@ -34,6 +34,49 @@ if TYPE_CHECKING:
 logger = get_logger(__name__)
 
 
+def _get_dtype(
+    cls,
+    dtype: Optional[Union[str, torch.dtype, dict]],
+    config: PretrainedConfig,
+) -> tuple[PretrainedConfig, Optional[torch.dtype], Optional[torch.dtype]]:
+    dtype_orig = None
+
+    if dtype is not None:
+        if isinstance(dtype, str):
+            if dtype == "auto":
+                if hasattr(config, "dtype") and config.dtype is not None:
+                    dtype = config.dtype
+                else:
+                    dtype = torch.get_default_dtype()
+            elif hasattr(torch, dtype):
+                dtype = getattr(torch, dtype)
+                config.dtype = dtype
+        elif isinstance(dtype, torch.dtype):
+            config.dtype = dtype
+        elif isinstance(dtype, dict):
+            for key, curr_dtype in dtype.items():
+                if hasattr(config, key):
+                    value = getattr(config, key)
+                    curr_dtype = curr_dtype if not isinstance(curr_dtype, str) else getattr(torch, curr_dtype)
+                    value.dtype = curr_dtype
+            # main torch dtype for modules that aren't part of any sub-config
+            dtype = dtype.get("")
+            dtype = dtype if not isinstance(dtype, str) else getattr(torch, dtype)
+            config.dtype = dtype
+            if dtype is None:
+                dtype = torch.float32
+        else:
+            raise ValueError(f"Invalid dtype: {dtype}")
+
+        dtype_orig = cls._set_default_dtype(dtype)
+    else:
+        # Use default dtype
+        default_dtype = torch.get_default_dtype()
+        config.dtype = default_dtype
+
+    return config, dtype, dtype_orig
+
+
 class RBLNModel(RBLNBaseModel):
     @classmethod
     def update_kwargs(cls, kwargs):
@@ -206,10 +249,37 @@ class RBLNModel(RBLNBaseModel):
         trust_remote_code: bool = False,
         # Some rbln-config should be applied before loading torch module (i.e. quantized llm)
         rbln_config: Optional[RBLNModelConfig] = None,
+        dtype: Optional[Union[str, torch.dtype, dict]] = None,
         **kwargs,
     ) -> "PreTrainedModel":
         kwargs = cls.update_kwargs(kwargs)
-        return cls.get_hf_class().from_pretrained(
+
+        hf_class = cls.get_hf_class()
+
+        if dtype is not None:
+            config = hf_class.config_class.from_pretrained(
+                model_id,
+                subfolder=subfolder,
+                revision=revision,
+                cache_dir=cache_dir,
+                use_auth_token=use_auth_token,
+                local_files_only=local_files_only,
+                force_download=force_download,
+                trust_remote_code=trust_remote_code,
+            )
+
+            config, processed_dtype, dtype_orig = _get_dtype(
+                cls=hf_class,
+                dtype=dtype,
+                config=config,
+            )
+
+            kwargs["torch_dtype"] = processed_dtype
+
+            if dtype_orig is not None:
+                hf_class._set_default_dtype(dtype_orig)
+
+        return hf_class.from_pretrained(
             model_id,
             subfolder=subfolder,
             revision=revision,

optimum/rbln/transformers/modeling_generic.py

@@ -23,6 +23,7 @@ different model architectures.
 import inspect
 from typing import TYPE_CHECKING, Optional, Union
 
+from torch import nn
 from transformers import (
     AutoModel,
     AutoModelForAudioClassification,
@@ -57,6 +58,28 @@ class RBLNTransformerEncoder(RBLNModel):
     rbln_model_input_names = ["input_ids", "attention_mask", "token_type_ids"]
     rbln_dtype = "int64"
 
+    @classmethod
+    def wrap_model_if_needed(cls, model: "PreTrainedModel", rbln_config: RBLNTransformerEncoderConfig) -> nn.Module:
+        class TransformerEncoderWrapper(nn.Module):
+            # Parameters to disable for RBLN compilation
+            DISABLED_PARAMS = {"return_dict", "use_cache"}
+
+            def __init__(self, model: "PreTrainedModel", rbln_config: RBLNTransformerEncoderConfig):
+                super().__init__()
+                self.model = model
+                self.rbln_config = rbln_config
+                self._forward_signature = inspect.signature(model.forward)
+
+            def forward(self, *args, **kwargs):
+                # Disable parameters that are not compatible with RBLN compilation
+                for param_name in self.DISABLED_PARAMS:
+                    if param_name in self._forward_signature.parameters:
+                        kwargs[param_name] = False
+
+                return self.model(*args, **kwargs)
+
+        return TransformerEncoderWrapper(model, rbln_config).eval()
+
     @classmethod
     def _update_rbln_config(
         cls,
@@ -208,7 +231,6 @@ class RBLNModelForQuestionAnswering(RBLNTransformerEncoder):
 
     def _prepare_output(self, output, return_dict):
         # Prepare QuestionAnswering specific output format.
-
         start_logits, end_logits = output
 
         if not return_dict:

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

@@ -31,6 +31,7 @@ from transformers.utils import logging
 from ....configuration_utils import RBLNCompileConfig, RBLNModelConfig
 from ....modeling import RBLNModel
 from ...utils.rbln_runtime_wrapper import LoopProcessor
+from ..decoderonly.generation_decoderonly import RBLNDecoderOnlyGenerationMixin
 
 
 logger = logging.get_logger(__name__)
@@ -265,7 +266,7 @@ class RBLNBlip2QFormerModel(RBLNModel):
         )
 
 
-class RBLNBlip2ForConditionalGeneration(RBLNModel):
+class RBLNBlip2ForConditionalGeneration(RBLNModel, RBLNDecoderOnlyGenerationMixin):
     """
     RBLNBlip2ForConditionalGeneration is a multi-modal model that integrates vision and language processing capabilities,
     optimized for RBLN NPUs. It is designed for conditional generation tasks that involve both image and text inputs.
@@ -433,3 +434,66 @@ class RBLNBlip2ForConditionalGeneration(RBLNModel):
             )
 
         return inputs_embeds
+
+    @torch.no_grad()
+    def generate(
+        self,
+        pixel_values: torch.FloatTensor,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        interpolate_pos_encoding: bool = False,
+        **generate_kwargs,
+    ) -> torch.LongTensor:
+        batch_size = pixel_values.shape[0]
+        image_embeds = self.vision_model(
+            pixel_values,
+            return_dict=True,
+            interpolate_pos_encoding=interpolate_pos_encoding,
+        ).last_hidden_state
+        image_attention_mask = torch.ones(image_embeds.size()[:-1], dtype=torch.long, device=image_embeds.device)
+
+        query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)
+        query_outputs = self.qformer(
+            query_embeds=query_tokens,
+            encoder_hidden_states=image_embeds,
+            encoder_attention_mask=image_attention_mask,
+            return_dict=True,
+        )
+        query_output = query_outputs.last_hidden_state
+
+        if query_output.dtype != image_embeds.dtype:
+            query_output = query_output.to(image_embeds.dtype)
+
+        language_model_inputs = self.language_projection(query_output)
+
+        if inputs_embeds is None:
+            if input_ids is None:
+                image_tokens = [self.config.image_token_index] * self.config.num_query_tokens
+                start_tokens = image_tokens + [self.config.text_config.bos_token_id]
+                input_ids = torch.tensor([start_tokens], dtype=torch.long, device=image_embeds.device)
+                input_ids = input_ids.repeat(batch_size, 1)
+            inputs_embeds = self.get_input_embeddings()(input_ids)
+
+        if attention_mask is None:
+            attention_mask = torch.ones_like(input_ids)
+
+        if input_ids is None:
+            special_image_mask = inputs_embeds == self.get_input_embeddings()(
+                torch.tensor(self.config.image_token_id, dtype=torch.long, device=inputs_embeds.device)
+            )
+            special_image_mask = special_image_mask.all(-1)
+        else:
+            special_image_mask = input_ids == self.config.image_token_id
+
+        special_image_mask = special_image_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
+        language_model_inputs = language_model_inputs.to(inputs_embeds.device, inputs_embeds.dtype)
+        inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, language_model_inputs)
+
+        inputs = {"inputs_embeds": inputs_embeds, "attention_mask": attention_mask}
+        if not self.language_model.config.is_encoder_decoder:
+            inputs["input_ids"] = input_ids
+
+        outputs = self.language_model.generate(**inputs, **generate_kwargs)
+
+        return outputs

optimum/rbln/transformers/models/decoderonly/modeling_decoderonly.py

@@ -19,7 +19,7 @@ from typing import TYPE_CHECKING, Any, Callable, List, Optional, Tuple, Union
 import rebel
 import torch
 from rebel.compile_context import CompileContext
-from transformers import AutoModel, AutoModelForCausalLM, PretrainedConfig, PreTrainedModel
+from transformers import AutoConfig, AutoModel, AutoModelForCausalLM, PretrainedConfig, PreTrainedModel
 from transformers.modeling_outputs import BaseModelOutputWithPast
 from transformers.modeling_utils import no_init_weights
 
@@ -317,12 +317,27 @@ class RBLNDecoderOnlyModel(RBLNModel, RBLNDecoderOnlyFlashAttentionMixin):
 
     @classmethod
     def get_pytorch_model(
-        cls, *args, rbln_config: Optional[RBLNDecoderOnlyModelConfig] = None, **kwargs
+        cls,
+        model_id: str,
+        *args,
+        rbln_config: Optional[RBLNDecoderOnlyModelConfig] = None,
+        num_hidden_layers: Optional[int] = None,
+        **kwargs,
     ) -> PreTrainedModel:
         if rbln_config and rbln_config.quantization:
-            model = cls.get_quantized_model(*args, rbln_config=rbln_config, **kwargs)
+            model = cls.get_quantized_model(model_id, *args, rbln_config=rbln_config, **kwargs)
         else:
-            model = super().get_pytorch_model(*args, **kwargs)
+            if num_hidden_layers is not None:
+                trust_remote_code = kwargs.get("trust_remote_code", None)
+                config, kwargs = AutoConfig.from_pretrained(
+                    model_id, return_unused_kwargs=True, num_hidden_layers=num_hidden_layers, **kwargs
+                )
+                if hasattr(config, "layer_types"):
+                    config.layer_types = config.layer_types[:num_hidden_layers]
+                kwargs["config"] = config
+                kwargs["trust_remote_code"] = trust_remote_code
+
+            model = super().get_pytorch_model(model_id, *args, **kwargs)
 
         return model
 

optimum/rbln/transformers/models/gemma3/modeling_gemma3.py

@@ -28,6 +28,7 @@ from ....modeling import RBLNModel
 from ...modeling_outputs import RBLNDecoderOnlyOutput
 from ...utils.rbln_runtime_wrapper import LoopProcessor
 from ..decoderonly.decoderonly_runtime_utils import RBLNPageTableManager
+from ..decoderonly.generation_decoderonly import RBLNDecoderOnlyGenerationMixin
 from ..decoderonly.modeling_decoderonly import (
     RBLNDecoderOnlyModelForCausalLM,
 )
@@ -77,7 +78,7 @@ class LoopProjector(LoopProcessor):
         return output[0]
 
 
-class RBLNGemma3ForConditionalGeneration(RBLNModel):
+class RBLNGemma3ForConditionalGeneration(RBLNModel, RBLNDecoderOnlyGenerationMixin):
     auto_model_class = AutoModelForImageTextToText
     _rbln_submodules = [
         {"name": "vision_tower"},

optimum/rbln/transformers/models/idefics3/modeling_idefics3.py

@@ -35,6 +35,7 @@ from ....configuration_utils import RBLNCompileConfig, RBLNModelConfig
 from ....modeling import RBLNModel
 from ....utils.runtime_utils import RBLNPytorchRuntime
 from ...modeling_outputs import RBLNDecoderOnlyOutput
+from ..decoderonly.generation_decoderonly import RBLNDecoderOnlyGenerationMixin
 
 
 if TYPE_CHECKING:
@@ -120,9 +121,6 @@ class RBLNIdefics3VisionTransformer(RBLNModel):
                 encoder_outputs = self.encoder(
                     inputs_embeds=hidden_states,
                     attention_mask=patch_attention_mask,
-                    output_attentions=None,
-                    output_hidden_states=None,
-                    return_dict=False,
                 )
                 last_hidden_state = encoder_outputs[0]
                 last_hidden_state = self.post_layernorm(last_hidden_state)
@@ -185,7 +183,7 @@ class RBLNIdefics3VisionTransformer(RBLNModel):
             return BaseModelOutput(last_hidden_state=last_hidden_state)
 
 
-class RBLNIdefics3ForConditionalGeneration(RBLNModel):
+class RBLNIdefics3ForConditionalGeneration(RBLNModel, RBLNDecoderOnlyGenerationMixin):
     """
     RBLNIdefics3ForConditionalGeneration is a multi-modal model that integrates vision and language processing capabilities,
     optimized for RBLN NPUs. It is designed for conditional generation tasks that involve both image and text inputs.

optimum/rbln/transformers/models/llava/modeling_llava.py

@@ -27,6 +27,7 @@ from ....modeling import RBLNModel
 from ....utils.logging import get_logger
 from ...modeling_outputs import RBLNDecoderOnlyOutput
 from ...utils.rbln_runtime_wrapper import LoopProcessor
+from ..decoderonly.generation_decoderonly import RBLNDecoderOnlyGenerationMixin
 
 
 logger = get_logger(__name__)
@@ -103,7 +104,7 @@ class LoopProjector(LoopProcessor):
         return output[0]
 
 
-class RBLNLlavaForConditionalGeneration(RBLNModel):
+class RBLNLlavaForConditionalGeneration(RBLNModel, RBLNDecoderOnlyGenerationMixin):
     """
     RBLNLlavaForConditionalGeneration is a multi-modal model that combines vision and language processing capabilities,
     optimized for RBLN NPUs. It is designed for conditional generation tasks that involve both image and text inputs.

optimum/rbln/transformers/models/llava_next/modeling_llava_next.py

@@ -32,6 +32,7 @@ from ....configuration_utils import RBLNCompileConfig, RBLNModelConfig
 from ....modeling import RBLNModel
 from ....utils.logging import get_logger
 from ...utils.rbln_runtime_wrapper import LoopProcessor
+from ..decoderonly.generation_decoderonly import RBLNDecoderOnlyGenerationMixin
 from ..decoderonly.modeling_decoderonly import RBLNDecoderOnlyOutput
 
 
@@ -87,7 +88,7 @@ class LoopProjector(LoopProcessor):
         return output[0]
 
 
-class RBLNLlavaNextForConditionalGeneration(RBLNModel):
+class RBLNLlavaNextForConditionalGeneration(RBLNModel, RBLNDecoderOnlyGenerationMixin):
     """
     RBLNLlavaNextForConditionalGeneration is a multi-modal model that combines vision and language processing capabilities,
     optimized for RBLN NPUs. It is designed for conditional generation tasks that involve both image and text inputs.

optimum/rbln/transformers/models/qwen2_5_vl/modeling_qwen2_5_vl.py

@@ -400,15 +400,6 @@ class RBLNQwen2_5_VLForConditionalGeneration(RBLNDecoderOnlyModelForCausalLM):
         del model.lm_head
         return model
 
-    @classmethod
-    def update_kwargs(cls, kwargs):
-        kwargs.update(
-            {
-                "_attn_implementation": "eager",
-            }
-        )
-        return super().update_kwargs(kwargs)
-
     @classmethod
     def get_input_info(
         cls,

optimum/rbln/transformers/models/seq2seq/modeling_seq2seq.py

@@ -20,6 +20,7 @@ import rebel
 import torch
 from rebel.compile_context import CompileContext
 from transformers import AutoModelForSeq2SeqLM, PretrainedConfig, PreTrainedModel
+from transformers.generation.utils import GenerationMixin
 from transformers.modeling_outputs import BaseModelOutput, Seq2SeqLMOutput
 
 from ....configuration_utils import RBLNCompileConfig
@@ -101,7 +102,7 @@ class RBLNRuntimeDecoder(RBLNPytorchRuntime):
         return Seq2SeqLMOutput(logits=lm_logits)
 
 
-class RBLNModelForSeq2SeqLM(RBLNModel, ABC):
+class RBLNModelForSeq2SeqLM(RBLNModel, GenerationMixin, ABC):
     """
     This is a generic model class that will be instantiated as one of the model classes of the library (with a sequence-to-sequence language modeling head) when created with the from_pretrained() class method.
     This model inherits from [`RBLNModel`]. Check the superclass documentation for the generic methods the library implements for all its models.
@@ -117,6 +118,7 @@ class RBLNModelForSeq2SeqLM(RBLNModel, ABC):
     main_input_name = "input_ids"
     auto_model_class = AutoModelForSeq2SeqLM
     support_causal_attn = None
+    _is_stateful = False
 
     def __post_init__(self, **kwargs):
         batch_size = self.rbln_config.batch_size

optimum/rbln/transformers/models/whisper/modeling_whisper.py

@@ -150,7 +150,8 @@ class RBLNWhisperForConditionalGeneration(RBLNModel, RBLNWhisperGenerationMixin)
     """
 
     auto_model_class = AutoModelForSpeechSeq2Seq
-    main_input_name = "input_ids"
+    main_input_name = "input_features"
+    _is_stateful = False
 
     def __post_init__(self, **kwargs):
         super().__post_init__(**kwargs)

{optimum_rbln-0.9.2a4.dist-info → optimum_rbln-0.9.2a5.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: optimum-rbln
-Version: 0.9.2a4
+Version: 0.9.2a5
 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
@@ -26,10 +26,10 @@ Requires-Python: <3.14,>=3.9
 Requires-Dist: accelerate>=1.0.1
 Requires-Dist: diffusers==0.35.1
 Requires-Dist: packaging>=24.1
-Requires-Dist: torch==2.7.0
-Requires-Dist: torchaudio<=2.7.0
-Requires-Dist: torchvision<=0.22.0
-Requires-Dist: transformers==4.53.1
+Requires-Dist: torch==2.8.0
+Requires-Dist: torchaudio<=2.8.0
+Requires-Dist: torchvision<=0.23.0
+Requires-Dist: transformers==4.57.1
 Description-Content-Type: text/markdown
 
 

{optimum_rbln-0.9.2a4.dist-info → optimum_rbln-0.9.2a5.dist-info}/RECORD

@@ -1,7 +1,7 @@
 optimum/rbln/__init__.py,sha256=AZ-7X3ZCjMNcz4mkC_98y-HWRw38ijh8gETJet9tiyM,18828
-optimum/rbln/__version__.py,sha256=z2uQrmRnC9pwi3WkC72y8UjBbBRNzVjlrzh6d2XytNE,712
+optimum/rbln/__version__.py,sha256=SknjPrWnnK8RN1n8X1WB74WmhEFBtAnbp5Zb4GrhDlA,712
 optimum/rbln/configuration_utils.py,sha256=KFibQ8IYcapw3M1GpgNLgQ1ZolRIip0_bOlwfU0OYac,38193
-optimum/rbln/modeling.py,sha256=IZ8loagxm--2BcqTl16KRHUR3hkccpeaY2grOWOtwqk,14473
+optimum/rbln/modeling.py,sha256=jfYPOl0FmYq6B3et3rmoCbiWV4939LlVgs-PrZ7IWIk,16928
 optimum/rbln/modeling_base.py,sha256=poXfHZCAlFd28MY9dvMi7tC2RytLx77Lee2XGS_KeZg,27684
 optimum/rbln/diffusers/__init__.py,sha256=1tgU_xWA42BmInqu9bBz_5R_E9TGhhK3mI06YlaiTLg,7232
 optimum/rbln/diffusers/modeling_diffusers.py,sha256=egx137ECmNA0bK2JPRUOpxCl1Wb3qOpE-xSaPaJOs-g,20549
@@ -75,7 +75,7 @@ optimum/rbln/ops/sliding_window_attn.py,sha256=EQrV_yRGc5z6kvwEsAcLP028bJWkQg2UP
 optimum/rbln/transformers/__init__.py,sha256=hyv53b_d_IJ9KYsDogTmKHDNuXIGNCFkrJI21RHE5ak,12573
 optimum/rbln/transformers/configuration_generic.py,sha256=rM4XY1a_UlRf3ZCZkCav59JKRuvqiEEUUgnqNlgdcv8,5207
 optimum/rbln/transformers/modeling_attention_utils.py,sha256=aLyOaq4me1m-JMmnKbuyNQageDxNU2jjEhGE_ew2P5o,11465
-optimum/rbln/transformers/modeling_generic.py,sha256=QXFCD35NSelRTMkiYkWnN3qfkkepQ29G0sZ4IOFZmpE,12569
+optimum/rbln/transformers/modeling_generic.py,sha256=csViS2KrCr4qaID2jZ1vdHn_W-3wdPFEAtsheEB_LFc,13619
 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=yzcjyHCHH4-Mi26N34HzNs7Tl5HjjT1rrwQ8f_W2_nc,13532
@@ -95,7 +95,7 @@ optimum/rbln/transformers/models/bert/configuration_bert.py,sha256=nEZnX6LXpLKWa
 optimum/rbln/transformers/models/bert/modeling_bert.py,sha256=7MQZS11k4__oyeni5ek2SzRf-gtD3_hMKl_oOzN7_XQ,2263
 optimum/rbln/transformers/models/blip_2/__init__.py,sha256=L01gPXcUCa8Vg-bcng20vZvBIN_jlqCzwUSFuq0QOag,855
 optimum/rbln/transformers/models/blip_2/configuration_blip_2.py,sha256=8eSilBwcPWQhBg-oilCmDPo-DN6V5lpLMlTB7WPknII,4630
-optimum/rbln/transformers/models/blip_2/modeling_blip_2.py,sha256=_IGDGS7VJfw5Qi5AJYCjduRVlWHGADhSAiTY42Jppys,16389
+optimum/rbln/transformers/models/blip_2/modeling_blip_2.py,sha256=MUDwSboH8gdIaJxbPUJsBPuhQf8ViNbVAViU2DASm1g,19308
 optimum/rbln/transformers/models/clip/__init__.py,sha256=TLeXDqcFK6M6v9x7Xr64kBbqGu3hFHM7p754dQ8UVQc,938
 optimum/rbln/transformers/models/clip/configuration_clip.py,sha256=Ea8TCVmMayydfw9p4kTP3UdtvoaPWf4Z4claB61JuE4,4175
 optimum/rbln/transformers/models/clip/modeling_clip.py,sha256=BLAYJAtv_2ZnKOlZ8iDBr2Su3bKM_eMWeUSK9MOaj7I,13198
@@ -110,7 +110,7 @@ optimum/rbln/transformers/models/decoderonly/decoderonly_architecture.py,sha256=
 optimum/rbln/transformers/models/decoderonly/decoderonly_runtime_utils.py,sha256=2yGU1JfxCepG37DbyY8oWq3gMIKhMnPyBkLH8VniKfA,21530
 optimum/rbln/transformers/models/decoderonly/generation_decoderonly.py,sha256=zabSgQd2VzHhkpbhUFW5Z-CjYB1JvSJOb5yXKjXCQV0,4326
 optimum/rbln/transformers/models/decoderonly/lora_architecture.py,sha256=jo-jYy95JhdvOsX1UTCXeYTNer37wBbtY578C0QQpZo,8306
-optimum/rbln/transformers/models/decoderonly/modeling_decoderonly.py,sha256=MqZUKffOvpVDPtVYSlm8hQEF4Tw1GV2ZPuevbjsNv7Y,35238
+optimum/rbln/transformers/models/decoderonly/modeling_decoderonly.py,sha256=myXdyYLivoPVJVGa2vmeaGLFxmjw-vUyYYsNTAX41yk,35907
 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=tTmsVaW9Wb2WD3nKRLwp7swn3hbMvgwUEJwwVIfNYEc,1008
@@ -132,7 +132,7 @@ optimum/rbln/transformers/models/gemma3/__init__.py,sha256=6rugk3615SEt4lh7gduo_
 optimum/rbln/transformers/models/gemma3/configuration_gemma3.py,sha256=NJJfarzbWJc3pm0XvICN7D0FFF9nqidagIEoOvYLixQ,4696
 optimum/rbln/transformers/models/gemma3/gemma3_architecture.py,sha256=TkGt2g313hXbB8vPFz8-oDBEsuR3HJI6LjSFgqec_Sc,6533
 optimum/rbln/transformers/models/gemma3/gemma3_runtime_utils.py,sha256=ZhWgecT4v4Ewd1hmrlJH47QUZuQweVB1qAaK-Qw24-Q,11127
-optimum/rbln/transformers/models/gemma3/modeling_gemma3.py,sha256=5kdARfkjkZ9a4LIeSQnNR5RM7ZQSYgavsphz19Vb7bY,25756
+optimum/rbln/transformers/models/gemma3/modeling_gemma3.py,sha256=Yx1rUxTgGvaCsNldggL3rFc2zxsndDKkQovjCmmNf28,25868
 optimum/rbln/transformers/models/gpt2/__init__.py,sha256=SsawHMStE3wYRtqkH5EvdTFkCdX0LLmp-QSKFhEBrHo,740
 optimum/rbln/transformers/models/gpt2/configuration_gpt2.py,sha256=iGdHfzG7plekZcIz-Z5U8lRE4SB8gbJJNcFQJ9l8Myg,1533
 optimum/rbln/transformers/models/gpt2/gpt2_architecture.py,sha256=ul87zvaLkqsuNJirvl6QtGXM147taNEbnb9qPulR1Ps,2933
@@ -143,17 +143,17 @@ optimum/rbln/transformers/models/grounding_dino/grounding_dino_architecture.py,s
 optimum/rbln/transformers/models/grounding_dino/modeling_grounding_dino.py,sha256=bXAOs2QH4sy2UFoFLUSM6u1_VHouUT5COERLQX20F6Y,46897
 optimum/rbln/transformers/models/idefics3/__init__.py,sha256=ulxE7HEfXsNJhd25J9Fvi6vggo9aZH9sLKJjWB6LlzQ,814
 optimum/rbln/transformers/models/idefics3/configuration_idefics3.py,sha256=7IENNxflZL8ZH3YRqtCXfYdKs-RdUeGiPzq-C03te_s,3679
-optimum/rbln/transformers/models/idefics3/modeling_idefics3.py,sha256=ajgLsQsU4n5NwcCqZvG7gswXhjPTHeyTjwjhje5Z7xI,19936
+optimum/rbln/transformers/models/idefics3/modeling_idefics3.py,sha256=LEFJu9JsoiS3ZJoG2J3QkwDAyTa75xJQmMtvpomYwsw,19918
 optimum/rbln/transformers/models/llama/__init__.py,sha256=6tgx9-qlM5r9ouoeZEouVRNLs3r6Sku-cuXNkyfeFHc,746
 optimum/rbln/transformers/models/llama/configuration_llama.py,sha256=_uxfH5kaGbeJTMJfESYn0Vg3OEkINS2ShGtVQTeOcs4,1578
 optimum/rbln/transformers/models/llama/llama_architecture.py,sha256=S7MCPfyjG5eUqgaS-QNBB0ApUD6wnb5fR0RHq7k7-pA,728
 optimum/rbln/transformers/models/llama/modeling_llama.py,sha256=uRxEXYhHOuEwPjBo_Ps3eFU1uwScasla6P8HwsQgAu0,4214
 optimum/rbln/transformers/models/llava/__init__.py,sha256=FaVLgBIqKGjT_nvwYO9k9BVqrzH_Ym3DfjGRCSUhG2s,734
 optimum/rbln/transformers/models/llava/configuration_llava.py,sha256=c1rie8LCypxlsT7SNjZJE07_xCLAasV4EBs97o1757Q,2998
-optimum/rbln/transformers/models/llava/modeling_llava.py,sha256=R_wDTJDNVweivDIsZ0eDd5Z21NEcH9O3xdTV9z8mpjQ,20239
+optimum/rbln/transformers/models/llava/modeling_llava.py,sha256=MaszTboXRr-PTWZKxhbw5w5rBZ95ES2_fHEW280T2LU,20351
 optimum/rbln/transformers/models/llava_next/__init__.py,sha256=kDXKr7wMkp1XqE__DER2B8kQF_NYMxhzsQS5ytGg56I,752
 optimum/rbln/transformers/models/llava_next/configuration_llava_next.py,sha256=Sz8L8p_23T7xw7pkUmW5pyK_wZclph1p_kQYbslc8m8,2708
-optimum/rbln/transformers/models/llava_next/modeling_llava_next.py,sha256=ZGVTOaAOpUOvyVLfSsHpfje4W8FDR_PV6MhS9QNj-Uk,21230
+optimum/rbln/transformers/models/llava_next/modeling_llava_next.py,sha256=0aooyMG7ElVIa52MB0ysVKqB4Pdxyl4tbeD1QdehiZk,21342
 optimum/rbln/transformers/models/midm/__init__.py,sha256=IC3FETwgYinbp3wDj7tp4zIHJhbqM-c6GfTRdYcMNj8,913
 optimum/rbln/transformers/models/midm/configuration_midm.py,sha256=DxhcSJlApxfi00XxYmSkKZ6bY9vfLXT0zh-oMKkZot0,1365
 optimum/rbln/transformers/models/midm/midm_architecture.py,sha256=f9IwLLyYErliWJhkRj880QByMEYs_XVwm2Yh6r-Y_ik,5186
@@ -184,7 +184,7 @@ optimum/rbln/transformers/models/qwen2/modeling_qwen2.py,sha256=VOboPJF1rvvSVWkH
 optimum/rbln/transformers/models/qwen2/qwen2_architecture.py,sha256=XlNAMYAcDLohnSAhIFGKOPuCB5XLgzYs5ABWdeQSaZs,720
 optimum/rbln/transformers/models/qwen2_5_vl/__init__.py,sha256=rAW3DKQUzGL6EMwa5r1iLu94yhpiZpk6zfoD7TtYXrc,865
 optimum/rbln/transformers/models/qwen2_5_vl/configuration_qwen2_5_vl.py,sha256=WHLH72i7Pe16Ee1waMixMsR3eD6TsMGN08QD82qdVvw,6162
-optimum/rbln/transformers/models/qwen2_5_vl/modeling_qwen2_5_vl.py,sha256=pI1TNDMO-ZiqXtoHboxLlqfplZbRh22lT1gxhqy6Jtg,26939
+optimum/rbln/transformers/models/qwen2_5_vl/modeling_qwen2_5_vl.py,sha256=ey6uZqf9ULr5LOugf-KrHGKMkdPyZ5XOPt8I-tBBXOc,26730
 optimum/rbln/transformers/models/qwen2_5_vl/qwen2_5_vl_architecture.py,sha256=hlx9Tt9n9m-fL4m21QFKgsN719CDhwhgfOMjnhde4RE,8392
 optimum/rbln/transformers/models/qwen2_vl/__init__.py,sha256=O3t6zKda92CnZDzEnz_dcisMOQ71-OOJxElXzKCH5e0,849
 optimum/rbln/transformers/models/qwen2_vl/configuration_qwen2_vl.py,sha256=mi5CqSKZ77G5Fib3g8a86_4CEB6lb-qJOhDnSqslvNk,4714
@@ -202,7 +202,7 @@ optimum/rbln/transformers/models/roberta/configuration_roberta.py,sha256=6KhO-xB
 optimum/rbln/transformers/models/roberta/modeling_roberta.py,sha256=74Pswb5JJNtctvrQHlo2zYocKZN0npWhjAaKMUDVBUU,1535
 optimum/rbln/transformers/models/seq2seq/__init__.py,sha256=HiSyWFcKeZ8okfo-s-_Mf_upyvAoZwraUIJyGNLNurY,714
 optimum/rbln/transformers/models/seq2seq/configuration_seq2seq.py,sha256=pFnt382URDduIpeNb7z_xmONCSOt_2mKssro5xe8y3E,3121
-optimum/rbln/transformers/models/seq2seq/modeling_seq2seq.py,sha256=G7Rkx4paSDlSE2pa1dtXTOGN9L9uy7ovVZjSrS8c5Kk,18295
+optimum/rbln/transformers/models/seq2seq/modeling_seq2seq.py,sha256=nb511JHbi1wLCr5dOTClItuScx1fb-PCaNxKXjFTVQs,18395
 optimum/rbln/transformers/models/seq2seq/seq2seq_architecture.py,sha256=jmBgj7BkUS_S-T-9DI53rE3KXUHSCoIofr7k5JDVPrU,20024
 optimum/rbln/transformers/models/siglip/__init__.py,sha256=X1Fc1GUnJ2EIxFx45nbeoW-T2t0OyP3W73C0HD8Vowo,712
 optimum/rbln/transformers/models/siglip/configuration_siglip.py,sha256=Fy-ANF91bQno_QVd4ZpyRs-uNgC_XRyBRScBg2uKM6w,3029
@@ -227,7 +227,7 @@ optimum/rbln/transformers/models/wav2vec2/modeling_wav2vec2.py,sha256=bMKHdUDHgz
 optimum/rbln/transformers/models/whisper/__init__.py,sha256=ErquiUlYycSYPsDcq9IwwmbZXoYLn1MVZ8VikWY5gQo,792
 optimum/rbln/transformers/models/whisper/configuration_whisper.py,sha256=bSwDN7VLuk1aVXvfrQIgb9SLdFBDhO5q8ZFaPQPJal0,3077
 optimum/rbln/transformers/models/whisper/generation_whisper.py,sha256=0MYzMTZwTHFcJV_ZEtCm2AZbKN9RHgAr9jefuOAouVI,5017
-optimum/rbln/transformers/models/whisper/modeling_whisper.py,sha256=R7o-0Ota8geUNLoyzkm7smP5Pd_cZKhC-kcwiMaVG8U,19293
+optimum/rbln/transformers/models/whisper/modeling_whisper.py,sha256=MFKYAqO1ep3teYumMY5E_jjyCU4552GKZacSNFyjVQM,19323
 optimum/rbln/transformers/models/whisper/whisper_architecture.py,sha256=fKUbAMIl20o6EBMVcLg9TDSsJ1FDp8NKcl4jT9RWCEM,13981
 optimum/rbln/transformers/models/xlm_roberta/__init__.py,sha256=O3o2KzJ8Li3QhB7GHdRQASc93SYO2jz00Rx4pxYRuDg,982
 optimum/rbln/transformers/models/xlm_roberta/configuration_xlm_roberta.py,sha256=wHRpGTXL9khYqSkKL1IgA7__6_lt9QpOz9tHumjK7fo,1260
@@ -245,7 +245,7 @@ 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=SKLnM3KsX8_rv3HauO4oB2-JSjzuadQjRwo_BhMUzLI,6362
-optimum_rbln-0.9.2a4.dist-info/METADATA,sha256=FZjCn9neDPlv0v6jWtx8gJCuUn0Cw-r5DWaimql02oo,5350
-optimum_rbln-0.9.2a4.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-optimum_rbln-0.9.2a4.dist-info/licenses/LICENSE,sha256=QwcOLU5TJoTeUhuIXzhdCEEDDvorGiC6-3YTOl4TecE,11356
-optimum_rbln-0.9.2a4.dist-info/RECORD,,
+optimum_rbln-0.9.2a5.dist-info/METADATA,sha256=8gm204ZrCL7yjPNG5vEGsSE5RhIDtPk2H-eWGFxX8yw,5350
+optimum_rbln-0.9.2a5.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+optimum_rbln-0.9.2a5.dist-info/licenses/LICENSE,sha256=QwcOLU5TJoTeUhuIXzhdCEEDDvorGiC6-3YTOl4TecE,11356
+optimum_rbln-0.9.2a5.dist-info/RECORD,,