optimum-rbln 0.8.1a1__py3-none-any.whl → 0.8.1a2__py3-none-any.whl

optimum/rbln/__version__.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.8.1a1'
-__version_tuple__ = version_tuple = (0, 8, 1, 'a1')
+__version__ = version = '0.8.1a2'
+__version_tuple__ = version_tuple = (0, 8, 1, 'a2')

optimum/rbln/transformers/modeling_rope_utils.py

@@ -48,10 +48,13 @@ def _compute_default_rope_parameters(
         Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
         post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE).
     """
-
     base = config.rope_theta
     partial_rotary_factor = config.partial_rotary_factor if hasattr(config, "partial_rotary_factor") else 1.0
-    head_dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads
+    head_dim = (
+        config.head_dim
+        if hasattr(config, "head_dim") and config.head_dim is not None
+        else config.hidden_size // config.num_attention_heads
+    )
     dim = int(head_dim * partial_rotary_factor)
 
     attention_factor = 1.0  # Unused in this type of RoPE

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

@@ -167,6 +167,11 @@ class _BaseAutoModelClass:
         rbln_cls = cls.get_rbln_cls(model_id, *args, **kwargs)
         return rbln_cls.from_pretrained(model_id, *args, **kwargs)
 
+    @classmethod
+    def from_model(cls, model, *args, **kwargs):
+        rbln_cls = get_rbln_model_cls(f"RBLN{model.__class__.__name__}")
+        return rbln_cls.from_model(model, *args, **kwargs)
+
     @staticmethod
     def register(rbln_cls: Type[RBLNBaseModel], exist_ok=False):
         """

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

@@ -22,6 +22,7 @@ from transformers.modeling_attn_mask_utils import (
 from transformers.utils import logging
 
 from ..seq2seq.seq2seq_architecture import (
+    Seq2SeqCrossAttention,
     Seq2SeqDecoder,
     Seq2SeqDecoderLayer,
     Seq2SeqDecoderWrapper,
@@ -45,7 +46,8 @@ class BartDecoderWrapper(Seq2SeqDecoderWrapper):
         new_layers = []
         for layer in model.get_decoder().layers:
             self_attn = BartSelfAttention(layer.self_attn, use_attention_mask=self.use_attention_mask)
-            new_layers.append(BartDecoderLayer(layer, self_attn))
+            cross_attn = BartCrossAttention(layer.encoder_attn)
+            new_layers.append(BartDecoderLayer(layer, self_attn, cross_attn))
 
         decoder_model = BartDecoder(model.get_decoder(), new_layers)
         new_model = BartForConditionalGeneration(model, decoder_model)
@@ -153,3 +155,14 @@ class BartSelfAttention(Seq2SeqSelfAttention):
         key_states = self.k_proj(hidden_states)
         value_states = self.v_proj(hidden_states)
         return query_states, key_states, value_states
+
+
+class BartCrossAttention(Seq2SeqCrossAttention):
+    def __post_init__(self):
+        self.q_proj = self._original_mod.q_proj
+        self.k_proj = self._original_mod.k_proj
+        self.v_proj = self._original_mod.v_proj
+        self.out_proj = self._original_mod.out_proj
+        self.num_heads = self._original_mod.num_heads
+        self.head_dim = self._original_mod.embed_dim // self._original_mod.num_heads
+        self.embed_dim = self._original_mod.embed_dim

optimum/rbln/transformers/models/exaone/modeling_exaone.py

@@ -13,7 +13,11 @@
 # limitations under the License.
 
 
+import inspect
+from typing import Any, Callable
+
 from transformers import AutoModelForCausalLM
+from transformers.generation.utils import GenerationMixin
 
 from ....utils import logging
 from ..decoderonly import RBLNDecoderOnlyModelForCausalLM
@@ -85,8 +89,19 @@ class RBLNExaoneForCausalLM(RBLNDecoderOnlyModelForCausalLM):
 
     _decoder_wrapper_cls = ExaoneForCausalLMWrapper
     _hf_class = AutoModelForCausalLM
+    _supports_cache_class = True
 
     @classmethod
     def from_pretrained(cls, *args, **kwargs):
         kwargs.setdefault("trust_remote_code", True)
         return super().from_pretrained(*args, **kwargs)
+
+    def __getattr__(self, __name: str) -> Any:
+        def redirect(func):
+            return lambda *pargs, **kwargs: func(self, *pargs, **kwargs)
+
+        val = getattr(GenerationMixin, __name)
+
+        if isinstance(val, Callable) and "self" in set(inspect.signature(val).parameters):
+            return redirect(val)
+        return val

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

@@ -11,6 +11,7 @@
 # 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 importlib
 import inspect
 from collections import deque
 from dataclasses import dataclass
@@ -123,6 +124,23 @@ class RBLNGemma3ForConditionalGeneration(RBLNModel):
     def can_generate(self):
         return True
 
+    @classmethod
+    def get_pytorch_model(cls, *args, **kwargs):
+        model = super().get_pytorch_model(*args, **kwargs)
+
+        with no_init_weights():
+            model_cls_name = model.model.language_model.__class__.__name__
+            causal_model_cls_name = model_cls_name.replace("TextModel", "ForCausalLM")
+            causal_model_cls = getattr(importlib.import_module("transformers"), causal_model_cls_name)
+            new_language_model = causal_model_cls(model.model.language_model.config)
+
+        new_language_model.lm_head = model.lm_head
+        new_language_model.model = model.model.language_model
+        model.model.language_model = new_language_model
+        model.lm_head = None
+        del model.lm_head
+        return model
+
     def __post_init__(self, **kwargs):
         self.vision_tower = LoopVisionTower(self.rbln_submodules[0])
         self.language_model = self.rbln_submodules[1]

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

@@ -168,7 +168,6 @@ class RBLNLlavaNextForConditionalGeneration(RBLNModel):
     ):
         # If you are unavoidably running on a CPU rather than an RBLN device,
         # store the torch tensor, weight, etc. in this function.
-
         save_dict = {}
         save_dict["image_newline"] = model.image_newline
         torch.save(save_dict, save_dir_path / subfolder / "torch_artifacts.pth")

optimum/rbln/transformers/models/midm/modeling_midm.py

@@ -12,7 +12,11 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import inspect
+from typing import Any, Callable
+
 from transformers import AutoModelForCausalLM
+from transformers.generation.utils import GenerationMixin
 
 from ....utils import logging
 from ..decoderonly import RBLNDecoderOnlyModelForCausalLM
@@ -84,8 +88,19 @@ class RBLNMidmLMHeadModel(RBLNDecoderOnlyModelForCausalLM):
 
     _decoder_wrapper_cls = MidmLMHeadModelWrapper
     _hf_class = AutoModelForCausalLM
+    _supports_cache_class = True
 
     @classmethod
     def from_pretrained(cls, *args, **kwargs):
         kwargs.setdefault("trust_remote_code", True)
         return super().from_pretrained(*args, **kwargs)
+
+    def __getattr__(self, __name: str) -> Any:
+        def redirect(func):
+            return lambda *pargs, **kwargs: func(self, *pargs, **kwargs)
+
+        val = getattr(GenerationMixin, __name)
+
+        if isinstance(val, Callable) and "self" in set(inspect.signature(val).parameters):
+            return redirect(val)
+        return val

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

@@ -28,6 +28,7 @@ from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import (
     Qwen2_5_VisionPatchEmbed,
     Qwen2_5_VisionRotaryEmbedding,
     Qwen2_5_VisionTransformerPretrainedModel,
+    Qwen2_5_VLModel,
     Qwen2_5_VLRotaryEmbedding,
 )
 
@@ -390,6 +391,14 @@ class RBLNQwen2_5_VLForConditionalGeneration(RBLNDecoderOnlyModelForCausalLM):
     def can_generate(self):
         return True
 
+    @classmethod
+    def get_pytorch_model(cls, *args, **kwargs):
+        model = super().get_pytorch_model(*args, **kwargs)
+        model.model.lm_head = model.lm_head
+        model.lm_head = None
+        del model.lm_head
+        return model
+
     @classmethod
     def update_kwargs(cls, kwargs):
         kwargs.update(
@@ -531,7 +540,8 @@ class RBLNQwen2_5_VLForConditionalGeneration(RBLNDecoderOnlyModelForCausalLM):
             vision_tokens = input_id[0][vision_start_indices + 1]
             image_nums = (vision_tokens == image_token_id).sum()
             video_nums = (vision_tokens == video_token_id).sum()
-            position_ids, rope_deltas = self.get_rope_index(
+            position_ids, rope_deltas = Qwen2_5_VLModel.get_rope_index(
+                self,
                 input_id,
                 image_grid_thw[image_idx : image_idx + image_nums] if image_grid_thw is not None else None,
                 video_grid_thw[video_idx : video_idx + video_nums] if video_grid_thw is not None else None,

optimum/rbln/transformers/models/qwen2_5_vl/qwen2_5_vl_architecture.py

@@ -3,8 +3,14 @@ from typing import Tuple
 
 import torch
 import torch.nn as nn
+from transformers import PreTrainedModel
 
 from ..decoderonly.decoderonly_architecture import (
+    DecoderOnlyAttention,
+    DecoderOnlyFlashAttention,
+    DecoderOnlyForCausalLM,
+    DecoderOnlyLayer,
+    DecoderOnlyModel,
     DecoderOnlyWrapper,
     apply_rotary_pos_emb,
 )
@@ -197,3 +203,40 @@ class Qwen2_5_VL_LanguageModelWrapper(DecoderOnlyWrapper):
             past_key_values,
             position_embeds,
         )
+
+    def convert_to_rbln_causal_lm(self, causal_lm: PreTrainedModel, max_seq_len: int):
+        new_layers = []
+
+        for layer in causal_lm.model.language_model.layers:
+            if self.attn_impl == "eager":
+                new_self_attn = DecoderOnlyAttention(
+                    layer.self_attn,
+                    self.use_attention_mask,
+                    self.use_position_ids,
+                    kvcache_block_size=self.kvcache_block_size,
+                )
+            elif self.attn_impl == "flash_attn":
+                new_self_attn = DecoderOnlyFlashAttention(
+                    layer.self_attn,
+                    kvcache_partition_len=self.kvcache_partition_len,
+                    kvcache_block_size=self.kvcache_block_size,
+                    use_attention_mask=self.use_attention_mask,
+                    use_position_ids=self.use_position_ids,
+                )
+            else:
+                raise NotImplementedError(f"Unknwon attn : {self.attn_impl}")
+
+            new_layer = DecoderOnlyLayer(layer, new_self_attn)
+            new_layers.append(new_layer)
+
+        new_model = DecoderOnlyModel(
+            causal_lm.model.language_model,
+            new_layers,
+            partition_len=self.kvcache_partition_len,
+            max_seq_len=max_seq_len,
+            kvcache_block_size=self.kvcache_block_size,
+            use_learned_pos_emb=self.use_learned_pos_emb,
+            sliding_window_layers=self.sliding_window_layers,
+        )
+        new_causal_lm = DecoderOnlyForCausalLM(causal_lm.model, new_model)
+        return new_causal_lm

optimum/rbln/transformers/models/seq2seq/seq2seq_architecture.py

@@ -148,7 +148,8 @@ class Seq2SeqDecoderWrapper(nn.Module):
         new_layers = []
         for layer in model.get_decoder().layers:
             self_attn = Seq2SeqSelfAttention(layer.self_attn)
-            new_layers.append(Seq2SeqDecoderLayer(layer, self_attn))
+            cross_attn = Seq2SeqCrossAttention(layer.encoder_attn)
+            new_layers.append(Seq2SeqDecoderLayer(layer, self_attn, cross_attn))
 
         decoder_model = Seq2SeqDecoder(model.get_decoder(), new_layers)
         new_model = Seq2SeqForConditionalGeneration(model, decoder_model)
@@ -341,10 +342,11 @@ class Seq2SeqDecoderLayer(torch.nn.Module):
         self_attn (Seq2SeqSelfAttention): Modified self-attention layer optimized for RBLN
     """
 
-    def __init__(self, decoder_layer, self_attn):
+    def __init__(self, decoder_layer, self_attn, cross_attn):
         super().__init__()
         self._original_mod = decoder_layer
         self.self_attn = self_attn
+        self.cross_attn = cross_attn
         self.__post_init__()
 
     def __post_init__(self, **kwargs):
@@ -402,7 +404,8 @@ class Seq2SeqDecoderLayer(torch.nn.Module):
         # Cross-Attention Block
         residual = hidden_states
         hidden_states = self.pre_cross_attn_layer_norm(hidden_states)
-        cross_attn_output = self.encoder_attn(
+
+        cross_attn_output = self.cross_attn(
             hidden_states=hidden_states,
             past_key_value=cross_past_key_value,
             attention_mask=encoder_attention_mask,
@@ -487,3 +490,38 @@ class Seq2SeqSelfAttention(nn.Module):
         attn_output = self.out_proj(attn_output)
 
         return attn_output
+
+
+class Seq2SeqCrossAttention(nn.Module):
+    def __init__(self, attn, **kwargs):
+        super().__init__()
+        self._original_mod = attn
+        self.__post_init__(**kwargs)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        key_value_states: torch.Tensor = None,
+        past_key_value: Optional[object] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+    ):
+        bsz, tgt_len, _ = hidden_states.size()
+        query_states = self.q_proj(hidden_states).view(bsz, -1, self.num_heads, self.head_dim).transpose(1, 2)
+
+        is_cross_attention = key_value_states is not None
+        if is_cross_attention:
+            key_states = past_key_value[0]
+            value_states = past_key_value[1]
+
+        attn_output = torch.nn.functional.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=attention_mask,
+        )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.view(bsz, tgt_len, self.embed_dim)
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, None, past_key_value

optimum/rbln/transformers/models/t5/t5_architecture.py

@@ -136,10 +136,14 @@ class T5Decoder(Seq2SeqDecoder):
 
 
 class T5Block(Seq2SeqDecoderLayer):
+    def __init__(self, decoder_layer, self_attn):
+        super().__init__(decoder_layer, self_attn, cross_attn=None)
+        self.__post_init__()
+
     def __post_init__(self):
         self.self_attn_layer_norm = self._original_mod.layer[0].layer_norm
         self.encoder_attn_layer_norm = self._original_mod.layer[1].layer_norm
-        self.encoder_attn = T5CrossAttention(self._original_mod.layer[1].EncDecAttention)
+        self.cross_attn = T5CrossAttention(self._original_mod.layer[1].EncDecAttention)
         self.ff_layer = self._original_mod.layer[2]
 
     def pre_self_attn_layer_norm(self, hidden_states):

{optimum_rbln-0.8.1a1.dist-info → optimum_rbln-0.8.1a2.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: optimum-rbln
-Version: 0.8.1a1
+Version: 0.8.1a2
 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
@@ -28,7 +28,7 @@ Requires-Dist: packaging>=24.1
 Requires-Dist: torch==2.6.0
 Requires-Dist: torchaudio<=2.6.0
 Requires-Dist: torchvision<=0.21.0
-Requires-Dist: transformers==4.50.3
+Requires-Dist: transformers==4.51.3
 Description-Content-Type: text/markdown
 
 

{optimum_rbln-0.8.1a1.dist-info → optimum_rbln-0.8.1a2.dist-info}/RECORD

@@ -1,5 +1,5 @@
 optimum/rbln/__init__.py,sha256=qJJTumXhoFnawXGpeGJbAm4J4A9FFwD1SQ2MqcKDXoM,14436
-optimum/rbln/__version__.py,sha256=Cr1VHmCmATBTdzggODEP9AOXlbrVmrnrzDpRkmuoeYw,519
+optimum/rbln/__version__.py,sha256=shHuZmlC1NNfixc5X-chXIonN-Rp-nB4KxJDhJNxa3k,519
 optimum/rbln/configuration_utils.py,sha256=o5oer7fBdE-MHLGNXoP35FjmuQbMmjEIDv0QE_k3kpo,32336
 optimum/rbln/modeling.py,sha256=ZlJ_tOCWiFjDIlwJ_B_HOCO0kBduWrBAbW9VSEVIAFg,12088
 optimum/rbln/modeling_base.py,sha256=5fUb1FaxfjApzJIkT8-SrPhuygGo_1Uc0i7UedawOeE,23393
@@ -64,16 +64,16 @@ optimum/rbln/ops/sliding_window_attn.py,sha256=EQrV_yRGc5z6kvwEsAcLP028bJWkQg2UP
 optimum/rbln/transformers/__init__.py,sha256=fE-kzDnWj0ueAG-xDrIKdBX59wCE__8m86uBMBOEb9g,9031
 optimum/rbln/transformers/configuration_generic.py,sha256=kNhPWtzF0IovUnrsXfxXdXITqgpfCAAedjfB6jSAhEg,5131
 optimum/rbln/transformers/modeling_generic.py,sha256=u1JzjWcPsQgH_rqBzRVr582NARqOk7XVKgY4CdEfXe8,12228
-optimum/rbln/transformers/modeling_rope_utils.py,sha256=t2tufkqFdG_itmyv-sIPYgvXpYzKeJ1F3ciyDFhvlPg,14301
+optimum/rbln/transformers/modeling_rope_utils.py,sha256=6Zg3r-TeUk4WQAlr95pqfhuoAD_RQ4njT1rbO9uPL0Q,14379
 optimum/rbln/transformers/models/__init__.py,sha256=-rc_00p4d58cdM2ylmgURxoAGKgIRF7X7r6z1w6h3mo,10061
 optimum/rbln/transformers/models/audio_spectrogram_transformer/__init__.py,sha256=I2vL4lrzbT5p4eJcH-EKHzEfcPkj_XVsie7jb9q6yic,775
 optimum/rbln/transformers/models/audio_spectrogram_transformer/configuration_audio_spectrogram_transformer.py,sha256=z7LJiVJPmnlCM3mcyhPJP8AufSrxO_dsPeJ51onq-Nc,833
 optimum/rbln/transformers/models/audio_spectrogram_transformer/modeling_audio_spectrogram_transformer.py,sha256=FIKEVWpIt6-JQX9B_rAfCrAPqdUHtR2i8D_X2k7639E,1498
 optimum/rbln/transformers/models/auto/__init__.py,sha256=34Xghf1ogG4u-jhBMlj134nHdgnR3JEHSeZTPuy3MpY,1071
-optimum/rbln/transformers/models/auto/auto_factory.py,sha256=GOcx4bDX41P08dbwRYSpuQqu2-9blQf6Z1tLZkkAPyw,7821
+optimum/rbln/transformers/models/auto/auto_factory.py,sha256=1CA52xV2dS1Uzumcgqe4zobdpoi-Xt2oNjP3uLFtm08,8020
 optimum/rbln/transformers/models/auto/modeling_auto.py,sha256=Ex2ARRRIt3LtKhazr0UWy67R6WFAwSKRoi5n6B8TvoI,4213
 optimum/rbln/transformers/models/bart/__init__.py,sha256=fVo-gZEmJ0yxkIxEX6ciuRAGgXNyuvaXE2s88bhbjAE,830
-optimum/rbln/transformers/models/bart/bart_architecture.py,sha256=Oo-Cdne7igKEex8wwP-gztKJHgs5GLHQjK1oc3IZIDE,5801
+optimum/rbln/transformers/models/bart/bart_architecture.py,sha256=oxb3l0lw0qrnbeQxT-R3c0OH4IEp7UzEp2Iz_jhZe_Q,6373
 optimum/rbln/transformers/models/bart/configuration_bart.py,sha256=iIuotDRU0IZteTNUY97jnsecIEYokaJ0BF5Sv_MJk00,909
 optimum/rbln/transformers/models/bart/modeling_bart.py,sha256=q42gQ42Ldkfrw_rv0__7dFQEM8UYfKEMCYd6Z-vg9KQ,1918
 optimum/rbln/transformers/models/bert/__init__.py,sha256=86FuGRBLw315_Roa9D5OUx6Ku2PM0DqSPZ-YSqbF-io,806
@@ -98,7 +98,7 @@ optimum/rbln/transformers/models/dpt/modeling_dpt.py,sha256=oKLX7MQZvfk1QB8wOtcd
 optimum/rbln/transformers/models/exaone/__init__.py,sha256=eUL0mq3yGVzCQfjLlOtVF2MecIN3DQWm07EmXubGSTs,921
 optimum/rbln/transformers/models/exaone/configuration_exaone.py,sha256=S4s4kJemPbmn-otYv-XNHE40DJaEYY6cmzaWV6MTGsY,1388
 optimum/rbln/transformers/models/exaone/exaone_architecture.py,sha256=-gCUDIJ1SJqNgIALJejJ75XPtAkd83oyomBke9xGRsc,3610
-optimum/rbln/transformers/models/exaone/modeling_exaone.py,sha256=V2OIrvZcJi8tpmwpQYFWYaMhZtCNr8LJ18ZpSzIle-A,3448
+optimum/rbln/transformers/models/exaone/modeling_exaone.py,sha256=sr_ICK-rw_fYmLY5r0IOc-vDtSZEcSwFIQp3Gn92zqE,3929
 optimum/rbln/transformers/models/gemma/__init__.py,sha256=VqPIlokw3kjn_ZoLXINCLXw3vaysQFo5oPGGy6bnt4Q,708
 optimum/rbln/transformers/models/gemma/configuration_gemma.py,sha256=3hAxl7LL9vFpCHrs-g3BwVDdVjnnJ-fzSO88wdfyGDQ,1361
 optimum/rbln/transformers/models/gemma/gemma_architecture.py,sha256=I9EyRIEtMw9u7HhSa8PqUco4uNe3gl6_lx0r1uDT8hA,2546
@@ -106,7 +106,7 @@ optimum/rbln/transformers/models/gemma/modeling_gemma.py,sha256=Ojvum34EhDHWfMB4
 optimum/rbln/transformers/models/gemma3/__init__.py,sha256=6rugk3615SEt4lh7gduo_J9VyGiSReuEIvL0Uno0eaI,790
 optimum/rbln/transformers/models/gemma3/configuration_gemma3.py,sha256=eupMGTHJGJNNrAZ3GE6M6GQBAQzBb7KFJvalyDmbM-A,3063
 optimum/rbln/transformers/models/gemma3/gemma3_architecture.py,sha256=sgFQQbvEr15tb2Sxk_tgcgQFcjhKGbNSW6fm2u7-Vck,8609
-optimum/rbln/transformers/models/gemma3/modeling_gemma3.py,sha256=tfaXdTheZFiwRT-lI16GfCwxcHk-mD2F4523EWKGQ-g,38711
+optimum/rbln/transformers/models/gemma3/modeling_gemma3.py,sha256=BjICf2xmiUKGkraoGmkTRMIpgqjHrBCiJxc4XgeQw9I,39483
 optimum/rbln/transformers/models/gpt2/__init__.py,sha256=socBMIBZSiLbrVN12rQ4nL9gFeT0axMgz6SWaCaD4Ac,704
 optimum/rbln/transformers/models/gpt2/configuration_gpt2.py,sha256=9sS6-EGapmow3rG9ViejK9qwrqy_X86VBxQ7u9x0Yqk,923
 optimum/rbln/transformers/models/gpt2/gpt2_architecture.py,sha256=pnGgixjgjW7HULbs5211cC2guw_4e4-MlS69vdCRMMg,3206
@@ -120,11 +120,11 @@ optimum/rbln/transformers/models/llama/llama_architecture.py,sha256=S7MCPfyjG5eU
 optimum/rbln/transformers/models/llama/modeling_llama.py,sha256=G91Yd8BfhiD8GT5bF46ZIHuc-ajGC-PO-mOQN3BhE1A,3122
 optimum/rbln/transformers/models/llava_next/__init__.py,sha256=kDXKr7wMkp1XqE__DER2B8kQF_NYMxhzsQS5ytGg56I,752
 optimum/rbln/transformers/models/llava_next/configuration_llava_next.py,sha256=a_Hg4_AN9rCbKvg53CpTPKIHY1IZuVl23b0GDQMgL3I,1865
-optimum/rbln/transformers/models/llava_next/modeling_llava_next.py,sha256=5xoySMVL5Zg7AN-USENHDW0kUkjoFg8CpSX9L4-MNvg,27422
+optimum/rbln/transformers/models/llava_next/modeling_llava_next.py,sha256=paYtCk58--FSZp8xjVrfZAxkJxO02X-jxaVPqL-l7ZU,27421
 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=XXY_uDGkXeVQnKpmSWrgljgxtSdTgLLFLfMqtZdRJdM,5642
-optimum/rbln/transformers/models/midm/modeling_midm.py,sha256=6reJLwXkxni1QCa2kombDsDg9IcKios0MmVOfbwAb_I,3395
+optimum/rbln/transformers/models/midm/modeling_midm.py,sha256=zbziYZ3f_dX_MOLwORTfJn22psZ1g3FFeQffM_TIh7A,3876
 optimum/rbln/transformers/models/mistral/__init__.py,sha256=9FE64bCYfSIyrBkRcwlqF8QyacSJFWvwEufHFi1ZIrM,716
 optimum/rbln/transformers/models/mistral/configuration_mistral.py,sha256=pMYJSwqmtx0uD2uExHx4S-JXal9rqQ5A2ulT2IoglTg,1383
 optimum/rbln/transformers/models/mistral/mistral_architecture.py,sha256=_aU8TE_tdvfo0K7QpgTlz_d0qwk4O82dl9268lPL16E,733
@@ -143,8 +143,8 @@ optimum/rbln/transformers/models/qwen2/modeling_qwen2.py,sha256=OKd7SXQLLtzPVolr
 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=U3ngIfkA58itqQZqTf-gbISMPoV7ipDttI7V2uwK_18,4155
-optimum/rbln/transformers/models/qwen2_5_vl/modeling_qwen2_5_vl.py,sha256=Q4U-avMkby-CunNXEERqvRZx9duC5i-6UmfF1376ciU,26336
-optimum/rbln/transformers/models/qwen2_5_vl/qwen2_5_vl_architecture.py,sha256=PAQz__9o_f5phlozhhXAB8JErBlS1jc4FYZkZkSYJuI,7312
+optimum/rbln/transformers/models/qwen2_5_vl/modeling_qwen2_5_vl.py,sha256=x14mGDbCGVFNTw3yXoRJ70W6U5wAlIe_lxqk190P3z4,26636
+optimum/rbln/transformers/models/qwen2_5_vl/qwen2_5_vl_architecture.py,sha256=G7zAMC2FSqz1dYisBUmBJF5hkyueaOrJSI2gD3nLvCQ,9068
 optimum/rbln/transformers/models/resnet/__init__.py,sha256=0QqtEQF1IMYgEmmfXMGarCDS8kJB5tzODfwTEzDVZRg,837
 optimum/rbln/transformers/models/resnet/configuration_resnet.py,sha256=KQd887jgNOl_Am3b407P2OvKtzkkeBS1cEhCfiN0tJg,769
 optimum/rbln/transformers/models/resnet/modeling_resnet.py,sha256=E8vg3Rw_KsHt6vaOg0ungZD7sXe0T4OMP0X8NFG1EXI,816
@@ -154,14 +154,14 @@ optimum/rbln/transformers/models/roberta/modeling_roberta.py,sha256=j2bTGwFFQhzp
 optimum/rbln/transformers/models/seq2seq/__init__.py,sha256=HiSyWFcKeZ8okfo-s-_Mf_upyvAoZwraUIJyGNLNurY,714
 optimum/rbln/transformers/models/seq2seq/configuration_seq2seq.py,sha256=lOLCGd9sv4N2fYsJmbf4eFPMpI5NFtA656qCEjpfulE,2651
 optimum/rbln/transformers/models/seq2seq/modeling_seq2seq.py,sha256=hxdsZMmsWFa10WoMleMvp9F0VjlFSwazWL5aKwM7_Xg,16979
-optimum/rbln/transformers/models/seq2seq/seq2seq_architecture.py,sha256=i2YXcusNvWjqYmS3ZvrWJiilovKIlfozB92fqDrtwq8,18729
+optimum/rbln/transformers/models/seq2seq/seq2seq_architecture.py,sha256=4MupGjhe_DpBLl43VVYY72WWiVdAr0mRPp8cHalQh2w,20028
 optimum/rbln/transformers/models/siglip/__init__.py,sha256=39MdhvWLZU0_9julQtJYVEiWI4csPrryS9krMauEA3s,730
 optimum/rbln/transformers/models/siglip/configuration_siglip.py,sha256=fXL4g985oeoH_aGHKXt_6YTEj8aAG_aGiGewZBDsAY0,2805
 optimum/rbln/transformers/models/siglip/modeling_siglip.py,sha256=se74bZTSTjqcV5UoBowSHVUTrE6CapoEcqgzqMs7Q9Y,7827
 optimum/rbln/transformers/models/t5/__init__.py,sha256=R1Q8Z1vaIdx4rDjeCmm_ZMSgewWaqaI0l93AHwewtew,818
 optimum/rbln/transformers/models/t5/configuration_t5.py,sha256=cOa2NAAocNSSECTtLpOsVVLfn4-7l7K7jz3nf6CSNCs,912
 optimum/rbln/transformers/models/t5/modeling_t5.py,sha256=SasIteIdpM-7TG6iPiRVlchd_3jaG7phFvdQJ_3VXjo,4924
-optimum/rbln/transformers/models/t5/t5_architecture.py,sha256=YdFOr8LfBhEXuPsLZDfYjmUSI-a3Otrzye1bOf8NUyw,9678
+optimum/rbln/transformers/models/t5/t5_architecture.py,sha256=X_9X4QRhkiiMrwFHv3mzER3yGmF9oQ2U-HdH6jbwVmw,9824
 optimum/rbln/transformers/models/time_series_transformer/__init__.py,sha256=xJaFWQawlwtv4H5tVFcY1pxLYzjHtMAlLq6nXysdkN8,1243
 optimum/rbln/transformers/models/time_series_transformer/configuration_time_series_transformer.py,sha256=FCzESVMNtpJ4Me-Vxr4i-naUfiR3S7YXWH_szLuoYeA,1390
 optimum/rbln/transformers/models/time_series_transformer/modeling_time_series_transformer.py,sha256=SJI1mgAWPfDWqptGBSddFuuVtWfONbIUo2_akGS7POA,17058
@@ -191,7 +191,7 @@ optimum/rbln/utils/model_utils.py,sha256=4k5879Kh75m3x_vS4-qOGfqsOiAvc2kdNFFfvsF
 optimum/rbln/utils/runtime_utils.py,sha256=LoKNK3AQNV_BSScstIZWjICkJf265MnUgy360BOocVI,5454
 optimum/rbln/utils/save_utils.py,sha256=hG5uOtYmecSXZuGTvCXsTM-SiyZpr5q3InUGCCq_jzQ,3619
 optimum/rbln/utils/submodule.py,sha256=w5mgPgncI740gVKMu3S-69DGNdUSI0bTZxegQGcZ98Y,5011
-optimum_rbln-0.8.1a1.dist-info/METADATA,sha256=HQbfacbOwIQ-KC327DSNRdT5-NYQoUqxHik2DnjuHyc,5299
-optimum_rbln-0.8.1a1.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-optimum_rbln-0.8.1a1.dist-info/licenses/LICENSE,sha256=QwcOLU5TJoTeUhuIXzhdCEEDDvorGiC6-3YTOl4TecE,11356
-optimum_rbln-0.8.1a1.dist-info/RECORD,,
+optimum_rbln-0.8.1a2.dist-info/METADATA,sha256=qiUSuqA0Rledv8xJhhmMT4DgtLU-wPUuvDXnyGv78CY,5299
+optimum_rbln-0.8.1a2.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+optimum_rbln-0.8.1a2.dist-info/licenses/LICENSE,sha256=QwcOLU5TJoTeUhuIXzhdCEEDDvorGiC6-3YTOl4TecE,11356
+optimum_rbln-0.8.1a2.dist-info/RECORD,,