optimum-rbln 0.9.2a2__py3-none-any.whl → 0.9.2a4__py3-none-any.whl

optimum/rbln/transformers/models/gemma3/gemma3_architecture.py

@@ -63,6 +63,7 @@ class Gemma3TextModel(DecoderOnlyModel):
         rotary_emb: torch.nn.Module = None,
         global_block_tables: Optional[torch.Tensor] = None,
         local_block_tables: Optional[torch.Tensor] = None,
+        lora_int_id: Optional[torch.Tensor] = None,
     ):
         # retrieve input_ids and inputs_embeds
         if (input_ids is None) ^ (inputs_embeds is not None):
@@ -105,6 +106,7 @@ class Gemma3TextModel(DecoderOnlyModel):
                 cos=cos_local if is_sliding else cos_global,
                 sin=sin_local if is_sliding else sin_global,
                 block_tables=local_block_tables if is_sliding else global_block_tables,
+                lora_int_id=lora_int_id,
             )
 
         hidden_states = self.get_last_layernorm()(hidden_states)
@@ -127,12 +129,20 @@ class Gemma3DecoderLayer(DecoderOnlyLayer):
         cos: Optional[torch.Tensor] = None,
         sin: Optional[torch.Tensor] = None,
         block_tables: Optional[torch.Tensor] = None,
+        lora_int_id: Optional[torch.Tensor] = None,
     ):
         residual = hidden_states
         hidden_states = self.get_pre_attention_layernorm()(hidden_states)
 
         hidden_states = self.self_attn(
-            hidden_states, attention_mask, seq_positions, past_key_values, cos, sin, block_tables
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            seq_positions=seq_positions,
+            past_key_values=past_key_values,
+            cos=cos,
+            sin=sin,
+            block_tables=block_tables,
+            lora_int_id=lora_int_id,
         )
         hidden_states = self.get_post_attention_layernorm()(hidden_states)
         hidden_states = residual + hidden_states
@@ -140,7 +150,7 @@ class Gemma3DecoderLayer(DecoderOnlyLayer):
         # Fully Connected
         residual = hidden_states
         hidden_states = self.get_pre_feedforward_layernorm()(hidden_states)
-        hidden_states = self._original_mod.mlp(hidden_states)
+        hidden_states = self.forward_mlp(hidden_states, lora_int_id)
         hidden_states = self.get_post_feedforward_layernorm()(hidden_states)
         hidden_states = residual + hidden_states
 

optimum/rbln/transformers/models/gemma3/gemma3_runtime_utils.py

@@ -17,15 +17,16 @@ import rebel
 import torch
 
 from ...modeling_outputs import RBLNDecoderOnlyOutput, RBLNGemma3ForCausalLMOutput
+from ..decoderonly.decoderonly_runtime_utils import RBLNPytorchRuntime
 from ..decoderonly.modeling_decoderonly import RBLNRuntimeModel
 
 
 class RBLNGemma3RuntimeModel(RBLNRuntimeModel):
     def __init__(self, *args, image_prefill: Optional[rebel.Runtime] = None, **kwargs):
         super().__init__(*args, **kwargs)
-        self.image_prefill = image_prefill  # FIXME(taehoon)
-        self.prefill = self.runtime if self.phase == "prefill" else None  # FIXME
-        self.decode = self.runtime if self.phase == "decode" else None
+        self.image_prefill = RBLNPytorchRuntime(image_prefill)  # FIXME(taehoon)
+        self.prefill = RBLNPytorchRuntime(self.runtime) if self.phase == "prefill" else None  # FIXME
+        self.decode = RBLNPytorchRuntime(self.runtime) if self.phase == "decode" else None
 
     def _prepare_prefill_inputs(self, *args, **kwargs):
         (
@@ -73,12 +74,24 @@ class RBLNGemma3RuntimeModel(RBLNRuntimeModel):
         position_embed: Optional[torch.Tensor] = None,
         token_type_ids: Optional[torch.Tensor] = None,
         local_block_tables: Optional[torch.Tensor] = None,
+        lora_int_ids: Optional[torch.Tensor] = None,
     ) -> torch.FloatTensor:
         """
         Performs chunked prefill for efficient KV-cache updates and memory optimization.
         Instead of processing the entire sequence at once, the input is divided into chunks of size `prefill_chunk_size`,
         and each chunk is processed sequentially. This allows for better memory utilization and compatibility with continuous batching.
         """
+        if self.rbln_config.use_lora and lora_int_ids is None:
+            if self.lora_int_ids is None:
+                raise ValueError(
+                    "lora_int_id is required when using LoRA. "
+                    "You should call set_lora_int_ids() before forward() or pass lora_int_id to forward()."
+                )
+            if batch_idx is not None:
+                lora_int_ids = self.lora_int_ids[batch_idx : batch_idx + 1].clone()
+            else:
+                lora_int_ids = self.lora_int_ids.clone()
+
         (
             inputs,
             cache_position,
@@ -141,6 +154,7 @@ class RBLNGemma3RuntimeModel(RBLNRuntimeModel):
                     query_position,
                     chunked_attention_mask,
                     position_ids_chunk,
+                    lora_int_ids if self.rbln_config.use_lora else None,
                 )
             else:
                 logits = self.prefill(
@@ -151,6 +165,7 @@ class RBLNGemma3RuntimeModel(RBLNRuntimeModel):
                     query_position,
                     chunked_attention_mask,
                     position_ids_chunk,
+                    lora_int_ids if self.rbln_config.use_lora else None,
                 )
 
             padded_cache_lengths += current_padded_cache_lengths
@@ -173,7 +188,20 @@ class RBLNGemma3RuntimeModel(RBLNRuntimeModel):
         position_embed: Optional[torch.Tensor] = None,
         position_ids: Optional[torch.Tensor] = None,
         local_block_tables: Optional[torch.Tensor] = None,
+        lora_int_ids: Optional[torch.Tensor] = None,
     ) -> torch.FloatTensor:
+        if self.rbln_config.use_lora and lora_int_ids is None:
+            if self.lora_int_ids is None:
+                raise ValueError(
+                    "lora_int_id is required when using LoRA. "
+                    "You should call set_lora_int_ids() before forward() or pass lora_int_id to forward()."
+                )
+
+            lora_int_ids = self.lora_int_ids
+
+        if lora_int_ids is not None and lora_int_ids.shape[0] != self.batch_size:
+            raise ValueError(f"lora_int_ids size mismatch: got {lora_int_ids.shape[0]}, expected {self.batch_size}.")
+
         batch_size = inputs.shape[0]
         if batch_size != self.batch_size:
             raise RuntimeError(

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

@@ -408,6 +408,13 @@ class RBLNGemma3ForCausalLM(RBLNDecoderOnlyModelForCausalLM):
     def _update_sliding_window_config(cls, model_config: PretrainedConfig, rbln_config: RBLNGemma3ForCausalLMConfig):
         sliding_window = getattr(model_config, "sliding_window", None)
         sliding_window_pattern = getattr(model_config, "sliding_window_pattern", None)
+        if sliding_window_pattern is None:
+            if hasattr(model_config, "layer_types"):
+                first_full_attention_index = model_config.layer_types.index("full_attention")
+                sliding_window_pattern = first_full_attention_index + 1
+            else:
+                raise ValueError("Cannot determine sliding_window_pattern from model_config")
+
         if sliding_window_pattern <= model_config.num_hidden_layers:
             rbln_config.cache_impl = "hybrid"
             rbln_config.sliding_window = sliding_window

optimum/rbln/transformers/models/gpt2/gpt2_architecture.py

@@ -75,7 +75,10 @@ class GPT2Attention(DecoderOnlyAttention):
         self.o_proj = self._original_mod.c_proj
         self.split_size = self._original_mod.split_size
 
-    def projection(self, hidden_states) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+    def projection(self, hidden_states, lora_int_id) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        if lora_int_id is not None:
+            raise NotImplementedError("LoRA is not supported for GPT2Attention")
+
         query_states, key_states, value_states = self.c_attn(hidden_states).split(self.split_size, dim=2)
         return query_states, key_states, value_states
 

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

@@ -123,7 +123,10 @@ class MidmAttention(DecoderOnlyAttention):
         self.split_size = self._original_mod.split_size
         self.num_key_value_heads = self._original_mod.num_heads
 
-    def projection(self, hidden_states) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+    def projection(self, hidden_states, lora_int_id) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        if lora_int_id is not None:
+            raise NotImplementedError("LoRA is not supported for MidmAttention")
+
         query_states, key_states, value_states = self.c_attn(hidden_states).split(self.split_size, dim=2)
         return query_states, key_states, value_states
 

optimum/rbln/transformers/models/phi/phi_architecture.py

@@ -56,7 +56,10 @@ class PhiAttention(DecoderOnlyAttention):
         self.qk_layernorm = self._original_mod.qk_layernorm
         self.rotary_ndims = self._original_mod.rotary_ndims
 
-    def projection(self, hidden_states) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+    def projection(self, hidden_states, lora_int_id) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        if lora_int_id is not None:
+            raise NotImplementedError("LoRA is not supported for PhiAttention")
+
         query_states = self.q_proj(hidden_states)
         key_states = self.k_proj(hidden_states)
         value_states = self.v_proj(hidden_states)
@@ -84,6 +87,7 @@ class PhiLayer(DecoderOnlyLayer):
         cos: Optional[torch.Tensor] = None,
         sin: Optional[torch.Tensor] = None,
         block_tables: Optional[torch.Tensor] = None,
+        lora_int_id: Optional[torch.Tensor] = None,
     ):
         residual = hidden_states
 

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

@@ -165,6 +165,7 @@ class Qwen2_5_VL_LanguageModelWrapper(DecoderOnlyWrapper):
         position_embeds = args.pop(0)
         query_position = args.pop(0) if self.phase == "prefill" else None
         position_ids = None
+        lora_int_id = None
         attention_mask = args.pop(0) if self.rbln_config.use_attention_mask else None
         past_key_values = args
 
@@ -192,6 +193,7 @@ class Qwen2_5_VL_LanguageModelWrapper(DecoderOnlyWrapper):
             query_position,
             attention_mask,
             position_ids,
+            lora_int_id,
             past_key_values,
             position_embeds,
         )

optimum/rbln/transformers/models/qwen2_vl/qwen2_vl_architecture.py

@@ -111,6 +111,7 @@ class Qwen2VL_LanguageModelWrapper(DecoderOnlyWrapper):
         query_position = args.pop(0) if self.phase == "prefill" else None
         position_ids = None
         attention_mask = args.pop(0) if self.rbln_config.use_attention_mask else None
+        lora_int_id = args.pop(0) if self.rbln_config.lora_config else None
         past_key_values = args
 
         if len(past_key_values) != 2 * self.num_hidden_layers:
@@ -137,6 +138,7 @@ class Qwen2VL_LanguageModelWrapper(DecoderOnlyWrapper):
             query_position,
             attention_mask,
             position_ids,
+            lora_int_id,
             past_key_values,
             position_embeds,
         )

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

@@ -39,11 +39,21 @@ from transformers.models.whisper.generation_whisper import WhisperGenerationMixi
 
 
 class RBLNWhisperGenerationMixin(WhisperGenerationMixin, GenerationMixin):
-    """
-    This class is based on transformers version 4.44.2.
-    It uses the same generate() method, so it's crucial to maintain the inheritance order.
-    Ensure WhisperGenerationMixin is listed before GenerationMixin.
-    """
+    def generate(self, *args, generation_config=None, **kwargs):
+        num_beams = kwargs.get(
+            "num_beams",
+            generation_config.num_beams
+            if hasattr(generation_config, "num_beams") and generation_config.num_beams is not None
+            else 1,
+        )
+        if num_beams > 1:
+            raise ValueError(
+                f"Beam search is not supported in RBLNWhisperGenerationMixin. "
+                f"Received num_beams={num_beams}, but only num_beams=1 is allowed. "
+                f"Please set num_beams=1 for greedy search or adjust your configuration."
+            )
+
+        return super().generate(*args, **kwargs)
 
     def _postprocess_outputs(
         self, seek_outputs, decoder_input_ids, return_token_timestamps, generation_config, *args, **kwargs

{optimum_rbln-0.9.2a2.dist-info → optimum_rbln-0.9.2a4.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: optimum-rbln
-Version: 0.9.2a2
+Version: 0.9.2a4
 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

{optimum_rbln-0.9.2a2.dist-info → optimum_rbln-0.9.2a4.dist-info}/RECORD

@@ -1,6 +1,6 @@
-optimum/rbln/__init__.py,sha256=DAJM5PWAYFiWVlyxVXUvj3CaFOEhX1yhEfhIt1LxL-A,18714
-optimum/rbln/__version__.py,sha256=6deWC5N0w1gBosV4k0OKRvgXRVzID3r8w84ih1uRdkY,712
-optimum/rbln/configuration_utils.py,sha256=nAzDwa_2-X32bikEJ2zbRZIk68s50hAP7gTmKN07Ig0,38080
+optimum/rbln/__init__.py,sha256=AZ-7X3ZCjMNcz4mkC_98y-HWRw38ijh8gETJet9tiyM,18828
+optimum/rbln/__version__.py,sha256=z2uQrmRnC9pwi3WkC72y8UjBbBRNzVjlrzh6d2XytNE,712
+optimum/rbln/configuration_utils.py,sha256=KFibQ8IYcapw3M1GpgNLgQ1ZolRIip0_bOlwfU0OYac,38193
 optimum/rbln/modeling.py,sha256=IZ8loagxm--2BcqTl16KRHUR3hkccpeaY2grOWOtwqk,14473
 optimum/rbln/modeling_base.py,sha256=poXfHZCAlFd28MY9dvMi7tC2RytLx77Lee2XGS_KeZg,27684
 optimum/rbln/diffusers/__init__.py,sha256=1tgU_xWA42BmInqu9bBz_5R_E9TGhhK3mI06YlaiTLg,7232
@@ -72,13 +72,13 @@ optimum/rbln/ops/flash_attn.py,sha256=yTCdYQVqm_1rHMHWjrMQaIR8WTuG_xA6t033x1IVvT
 optimum/rbln/ops/kv_cache_update.py,sha256=aIvK2Sp7M3EfJzJgNvIvAJv4emoN6QOhmgaWj-VboLs,1440
 optimum/rbln/ops/linear.py,sha256=5K3pcrrUHu_p8LrMIU-jX2TnafksveFjjZSCsYSp_yw,1328
 optimum/rbln/ops/sliding_window_attn.py,sha256=EQrV_yRGc5z6kvwEsAcLP028bJWkQg2UPI3xubt9skU,3487
-optimum/rbln/transformers/__init__.py,sha256=g5G6Eqk80NzS0tMmwghFI2DMKgPaOpoafv1m0Euhw6A,12459
+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_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=9gAXrYeYPdLbQH8KlRG4eSOFQ8h4kyWzXPM1grHvpDQ,13418
+optimum/rbln/transformers/models/__init__.py,sha256=yzcjyHCHH4-Mi26N34HzNs7Tl5HjjT1rrwQ8f_W2_nc,13532
 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
@@ -102,13 +102,15 @@ optimum/rbln/transformers/models/clip/modeling_clip.py,sha256=BLAYJAtv_2ZnKOlZ8i
 optimum/rbln/transformers/models/colpali/__init__.py,sha256=n3rueXT_oC0N8myoZiic0YkVK24CW5hZBPa-0L8so6Y,119
 optimum/rbln/transformers/models/colpali/colpali_architecture.py,sha256=TCOW3v5l9fIt1uIFtWa8ZAxq1cdCER8gXWjmbLQD20M,8079
 optimum/rbln/transformers/models/colpali/configuration_colpali.py,sha256=_HuZBVV-ponml95UapkYpRhffZy53-9jSZknx7hID7o,3348
-optimum/rbln/transformers/models/colpali/modeling_colpali.py,sha256=YnAUMTbcX2idRLMkZ3Qiv-GZFodfCun5-3sDfiIz77s,17955
-optimum/rbln/transformers/models/decoderonly/__init__.py,sha256=w3VZOIBYaHXVdnuhK4y0zWAj0IAv7_5LGTJYaz9oYmI,1056
-optimum/rbln/transformers/models/decoderonly/configuration_decoderonly.py,sha256=yUhB9yMFVbq7B9WtAM6m8cQ4K0U3a2N-TU34N5zvjRM,15180
-optimum/rbln/transformers/models/decoderonly/decoderonly_architecture.py,sha256=ySsiU0Acj5enJW-SqtFMAfBeH0HeqlhCd78QlpKJNQw,42780
-optimum/rbln/transformers/models/decoderonly/decoderonly_runtime_utils.py,sha256=v3mfIlQImQkYYr-rPn7rQR3GYdVUhALRttEduLI7H9c,20012
+optimum/rbln/transformers/models/colpali/modeling_colpali.py,sha256=2lHxvtrK3x2GOv7r-5nZelmjezm3ehe6Qf28cMdNmoQ,17961
+optimum/rbln/transformers/models/decoderonly/__init__.py,sha256=pKBXAtE3y_6nnwYfQJjdPmWqUwxuJ0lr8rrqkgyH07M,1126
+optimum/rbln/transformers/models/decoderonly/configuration_decoderonly.py,sha256=GX-IwTe6ywM9hmyquIu66y0YgIVZS5JNIz8LKAb4Ow8,17003
+optimum/rbln/transformers/models/decoderonly/configuration_lora.py,sha256=5DuTs2vy7jF7MLy161QD_KvCTaNW-5Mok7hBH0yK44U,17356
+optimum/rbln/transformers/models/decoderonly/decoderonly_architecture.py,sha256=h1n9vSHgQc7D0ds1C9SAzWxmIdTaqnDL7auDU_VJNXg,46813
+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/modeling_decoderonly.py,sha256=1umcuZg_ifIyVBfp_6oHQJvZsp_y2UCiv45-rv89_VA,32434
+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/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
@@ -128,12 +130,12 @@ optimum/rbln/transformers/models/gemma/gemma_architecture.py,sha256=4Ry2pFfWg0sV
 optimum/rbln/transformers/models/gemma/modeling_gemma.py,sha256=AsuFnrwZcRCKtF39BpHHNea0S34N2lNWKV4qZimmY8I,4170
 optimum/rbln/transformers/models/gemma3/__init__.py,sha256=6rugk3615SEt4lh7gduo_J9VyGiSReuEIvL0Uno0eaI,790
 optimum/rbln/transformers/models/gemma3/configuration_gemma3.py,sha256=NJJfarzbWJc3pm0XvICN7D0FFF9nqidagIEoOvYLixQ,4696
-optimum/rbln/transformers/models/gemma3/gemma3_architecture.py,sha256=fpLDAXCe5paWVsfc0tL59JkRQMRF-WNgIzOIb_QpSLU,6191
-optimum/rbln/transformers/models/gemma3/gemma3_runtime_utils.py,sha256=vYQ9sjRlkfamxZca_hVMQI0ylKeExsV02gOWaYVMjyg,9640
-optimum/rbln/transformers/models/gemma3/modeling_gemma3.py,sha256=WQ5qRZcM2EYLNqogjy-I04abwZXLZilCGFM4vO-MH4c,25381
+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/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=MyAWReXmyuHnDpW5HI_TI7psyJZxLujZ9KT5XnNm7nA,2802
+optimum/rbln/transformers/models/gpt2/gpt2_architecture.py,sha256=ul87zvaLkqsuNJirvl6QtGXM147taNEbnb9qPulR1Ps,2933
 optimum/rbln/transformers/models/gpt2/modeling_gpt2.py,sha256=DhF6hU3oCYGbZ7UijKCsRfTx-VCkTqqqNwqqMSrjqRE,2230
 optimum/rbln/transformers/models/grounding_dino/__init__.py,sha256=DE7DipZGvrKC6b1T77k4I4X3G70ss8mlr-PrZCaohto,307
 optimum/rbln/transformers/models/grounding_dino/configuration_grounding_dino.py,sha256=s-5MjEEle0zDBhskeYZQiPbbNsFvpTNcqcz21-kl6Gk,3820
@@ -154,7 +156,7 @@ optimum/rbln/transformers/models/llava_next/configuration_llava_next.py,sha256=S
 optimum/rbln/transformers/models/llava_next/modeling_llava_next.py,sha256=ZGVTOaAOpUOvyVLfSsHpfje4W8FDR_PV6MhS9QNj-Uk,21230
 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=RlkmNhaWE5h_awt9aTtR8VZfshNTah0IoUfD2Z9vfxI,5055
+optimum/rbln/transformers/models/midm/midm_architecture.py,sha256=f9IwLLyYErliWJhkRj880QByMEYs_XVwm2Yh6r-Y_ik,5186
 optimum/rbln/transformers/models/midm/modeling_midm.py,sha256=EXTBFaOWco2h3NL8jUACeDmPFJuAjFrtQcgO_BTMuCk,6004
 optimum/rbln/transformers/models/mistral/__init__.py,sha256=bYPqrkmqXmhNDqRgKFaL9iH7piGLSHKzsVrGl_0qs1Q,758
 optimum/rbln/transformers/models/mistral/configuration_mistral.py,sha256=mIfz8J8GZV9ojCMuNj9Zeky_PNu1Ir34DQ7FDZrGkP8,1595
@@ -171,7 +173,7 @@ optimum/rbln/transformers/models/pegasus/pegasus_architecture.py,sha256=Hk4N7LDA
 optimum/rbln/transformers/models/phi/__init__.py,sha256=M5Sh4AtIhJYegl-yAKPggAU3DtJtQOa8MrIQypZ6N7U,734
 optimum/rbln/transformers/models/phi/configuration_phi.py,sha256=CXHIG3xlBdr628oDu_u4OGsu_QZLx5EUSqu3zfmfEnk,1553
 optimum/rbln/transformers/models/phi/modeling_phi.py,sha256=r7B0NlqwIGjm-MmE-h5_xeRJPzs4O2OotgbjI-FYA2o,3403
-optimum/rbln/transformers/models/phi/phi_architecture.py,sha256=ygJIJvn20bnxE9nHKo4CBW9_1FJsz7MEVolB5asTmI0,3684
+optimum/rbln/transformers/models/phi/phi_architecture.py,sha256=bzK7Qhd1FAC049fdDhzzPYK7HtlHTjBqVmuFAhTX80Q,3866
 optimum/rbln/transformers/models/pixtral/__init__.py,sha256=fhclVAWnIDsfMfC-TW6mYrJXxgyehlLaadK64LOShH4,716
 optimum/rbln/transformers/models/pixtral/configuration_pixtral.py,sha256=b79zkJB1jzHx4S1wTe-Ju_Yel_PS5Q8bfmlQPzkchKU,1677
 optimum/rbln/transformers/models/pixtral/modeling_pixtral.py,sha256=P1lzi6JOTB43nBfCOonUDYhIXoMq6DnQpcvGfOO7ZP8,12252
@@ -183,11 +185,11 @@ optimum/rbln/transformers/models/qwen2/qwen2_architecture.py,sha256=XlNAMYAcDLoh
 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/qwen2_5_vl_architecture.py,sha256=y0W5qxrke7JbFNkPTkEOC8TvKnAYZP9bNsQK_IYvxnA,8340
+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
 optimum/rbln/transformers/models/qwen2_vl/modeling_qwen2_vl.py,sha256=YO8cKBEb7dU9D--gidYsPyhS2arOwgVqDe3tLlGHdx4,20424
-optimum/rbln/transformers/models/qwen2_vl/qwen2_vl_architecture.py,sha256=_5Erjk7udq7YFVVtx4XeWaIzpDlEZKJPu2bBSVBGfKE,5728
+optimum/rbln/transformers/models/qwen2_vl/qwen2_vl_architecture.py,sha256=xjp52RXqKA_BiyZ5CqwFAosav7ysvOJxeRxbPnTVIjM,5829
 optimum/rbln/transformers/models/qwen3/__init__.py,sha256=tI4KwvXpD35dUUaa8aLUXpWoU9gJGcmKXeywOlH14ZE,746
 optimum/rbln/transformers/models/qwen3/configuration_qwen3.py,sha256=BFRPggnH4VlsXlOa19C6KAID-bPgQ8ooQ29dvogh5zk,2102
 optimum/rbln/transformers/models/qwen3/modeling_qwen3.py,sha256=S05efusxjXJhMMYztstGes6ZbqkSr5I4fHFaLSYVG8c,5760
@@ -224,7 +226,7 @@ optimum/rbln/transformers/models/wav2vec2/configuration_wav2vec2.py,sha256=24sXi
 optimum/rbln/transformers/models/wav2vec2/modeling_wav2vec2.py,sha256=bMKHdUDHgzu1pXH0yrrOFCiA_T9xqb9B19kljCQ9yUU,1945
 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=GIHTca3b1VtW81kp7BzKQ7f77c2t9OsEsbZetripgDo,4582
+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/whisper_architecture.py,sha256=fKUbAMIl20o6EBMVcLg9TDSsJ1FDp8NKcl4jT9RWCEM,13981
 optimum/rbln/transformers/models/xlm_roberta/__init__.py,sha256=O3o2KzJ8Li3QhB7GHdRQASc93SYO2jz00Rx4pxYRuDg,982
@@ -243,7 +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-0.9.2a2.dist-info/METADATA,sha256=jcahTsRNj4iwUrHeEijhiyOGcPw-_YI0BpTNqcaSPYc,5350
-optimum_rbln-0.9.2a2.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-optimum_rbln-0.9.2a2.dist-info/licenses/LICENSE,sha256=QwcOLU5TJoTeUhuIXzhdCEEDDvorGiC6-3YTOl4TecE,11356
-optimum_rbln-0.9.2a2.dist-info/RECORD,,
+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,,