optimum-rbln 0.7.4a2__py3-none-any.whl → 0.7.4a4__py3-none-any.whl

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

@@ -12,4 +12,5 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from ....ops import paged_add_softmax_attn_decode
 from .modeling_t5 import RBLNT5EncoderModel, RBLNT5ForConditionalGeneration

optimum/rbln/transformers/models/t5/modeling_t5.py

@@ -36,50 +36,19 @@ from .t5_architecture import T5Wrapper
 logger = get_logger()
 
 if TYPE_CHECKING:
-    from rebel import Runtime
     from transformers import AutoFeatureExtractor, AutoProcessor, AutoTokenizer, PreTrainedModel
 
 
 class RBLNRuntimeModel(RBLNPytorchRuntime):
-    def __init__(
-        self,
-        runtime: "Runtime",
-        max_seq_len: int,
-        **kwargs: Any,
-    ) -> None:
-        super().__init__(runtime, **kwargs)
-        self.max_seq_len = max_seq_len
-
-    def _prepare_inputs(
-        self,
-        input_ids: torch.LongTensor,
-        attention_mask: torch.LongTensor,
-    ):
-        input_len = input_ids.shape[-1]
-        pad_len = None
-        if input_len > self.max_seq_len:
-            raise ValueError(f"Error input_len({input_len}) exceed max_seq_len({self.max_seq_len}).")
-        elif input_len < self.max_seq_len and input_len > 0:
-            pad_len = self.max_seq_len - input_len
-            logger.warning(
-                f"Warning: The input was padded with {pad_len} tokens to meet the compiled model's requirements. "
-                "For optimal performance, consider recompiling with a shorter 'rbln_max_seq_len'."
-            )
-            input_ids = torch.nn.functional.pad(input_ids, (0, pad_len))
-            attention_mask = torch.nn.functional.pad(attention_mask, (0, pad_len), value=0)
-
-        return input_ids, attention_mask, pad_len
-
     def forward(
         self,
         input_ids: torch.LongTensor,
-        attention_mask: torch.LongTensor,
+        attention_mask: torch.FloatTensor,
         head_mask: torch.FloatTensor,
         inputs_embeds: torch.FloatTensor,
         **kwargs,
     ):
-        input_ids, attention_mask, pad_len = self._prepare_inputs(input_ids, attention_mask)
-        logits = super().forward(
+        return super().forward(
             input_ids,
             attention_mask,
             head_mask,
@@ -87,8 +56,6 @@ class RBLNRuntimeModel(RBLNPytorchRuntime):
             **kwargs,
         )
 
-        return logits[:, :-pad_len, :] if pad_len is not None else logits
-
 
 class T5EncoderWrapper(torch.nn.Module):
     def __init__(self, model: "T5EncoderModel") -> None:
@@ -105,8 +72,7 @@ class RBLNT5EncoderModel(RBLNModel):
     rbln_model_input_names = ["input_ids", "attention_mask"]
 
     def __post_init__(self, **kwargs):
-        max_seq_len = self.rbln_config.model_cfg["max_seq_len"]
-        self.model = RBLNRuntimeModel(runtime=self.model[0], max_seq_len=max_seq_len)
+        self.model = RBLNRuntimeModel(runtime=self.model[0])
 
     @classmethod
     def wrap_model_if_needed(self, model: "PreTrainedModel", rbln_config: "RBLNConfig"):

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

@@ -18,7 +18,6 @@ import torch
 from torch import nn
 from transformers.utils import logging
 
-from ....ops import register_rbln_custom_paged_add_softmax_attention
 from ..seq2seq.seq2seq_architecture import (
     Seq2SeqDecoder,
     Seq2SeqDecoderLayer,
@@ -55,7 +54,6 @@ class T5EncoderWrapper(Seq2SeqEncoderWrapper):
 
 class T5DecoderWrapper(Seq2SeqDecoderWrapper):
     def __post_init__(self, model, dec_max_seq_len: int = None):
-        register_rbln_custom_paged_add_softmax_attention()
         self.num_layers = self.config.num_layers
         self.conditional_generation = self.convert_to_rbln_conditional_generation(model, dec_max_seq_len)
 
@@ -78,11 +76,12 @@ class T5DecoderWrapper(Seq2SeqDecoderWrapper):
         encoder_attention_mask,
         cache_position,
         block_tables,
-        cross_kv_cache,
-        *self_kv_cache,
+        *kv_cache,
     ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor]]:
         self_past_key_values = ()
         cross_past_key_values = ()
+        self_kv_cache = kv_cache[self.num_layers * 2 :]
+        cross_kv_cache = kv_cache[: self.num_layers * 2]
 
         for i in range(0, self.num_layers * 2, 2):
             self_past_key_values = self_past_key_values + ((self_kv_cache[i], self_kv_cache[i + 1]),)

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

@@ -21,4 +21,5 @@
 # copied, modified, or distributed without prior written permission
 # from Rebellions Inc.
 
+from ....ops import paged_add_softmax_attn_decode, rbln_cache_update
 from .modeling_time_series_transformers import RBLNTimeSeriesTransformerForPrediction

optimum/rbln/transformers/models/time_series_transformers/time_series_transformers_architecture.py

@@ -25,25 +25,16 @@ from typing import Optional, Tuple, Union
 
 import torch
 from torch import nn
-from transformers.modeling_attn_mask_utils import (
-    _prepare_4d_causal_attention_mask,
-)
-from transformers.modeling_outputs import (
-    BaseModelOutput,
-    Seq2SeqLMOutput,
-)
+from transformers.modeling_attn_mask_utils import _prepare_4d_causal_attention_mask
+from transformers.modeling_outputs import BaseModelOutput, Seq2SeqLMOutput
 from transformers.utils import logging
 
-from ....ops import register_rbln_custom_cache_update, register_rbln_custom_paged_add_softmax_attention
-
 
 logger = logging.get_logger(__name__)
 
 
 class TimeSeriesTransformersWrapper:
     def __init__(self, model, num_parallel_samples):
-        register_rbln_custom_cache_update()
-        register_rbln_custom_paged_add_softmax_attention()
         self.encoder = TimeSeriesTransformersEncoderWrapper(model)
         self.decoder = TimeSeriesTransformersDecoderWrapper(model, num_parallel_samples)
 
@@ -181,7 +172,6 @@ class TimeSeriesTransformersDecoder(nn.Module):
             hidden_states = decoder_layer(
                 hidden_states,
                 attention_mask=attention_mask,
-                # encoder_attention_mask=encoder_attention_mask,
                 self_past_key_value=self_past_key_value,
                 cross_past_key_value=cross_past_key_value,
                 cache_position=cache_position,
@@ -285,16 +275,16 @@ class TimeSeriesTransformersSelfAttention(TimeSeriesTransformersAttention):
 
         block_size = past_key_value[0].shape[-2]
         attn_output = torch.ops.rbln_custom_ops.paged_add_softmax_attn_decode(
-            query_states,
-            key_states,
-            value_states,
-            attention_mask.unsqueeze(2),
-            past_key_value[0].view(1, bsz * self.num_heads, 1, -1, self.head_dim),
-            past_key_value[1].view(1, bsz * self.num_heads, 1, -1, self.head_dim),
-            cache_position.expand(bsz, 1),
-            torch.tensor(1.0, dtype=torch.float32),  # scale
-            block_tables,
-            block_size,
+            q=query_states,
+            k=key_states,
+            v=value_states,
+            mask=attention_mask.unsqueeze(2),
+            kcache=past_key_value[0].view(1, bsz * self.num_heads, 1, -1, self.head_dim),
+            vcache=past_key_value[1].view(1, bsz * self.num_heads, 1, -1, self.head_dim),
+            seq=cache_position.expand(bsz, 1),
+            scale=torch.tensor(1.0, dtype=torch.float32),  # scale
+            block_table=block_tables,
+            block_size=block_size,
         )
 
         attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)

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

@@ -12,4 +12,5 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from ....ops import paged_add_softmax_attn_decode
 from .modeling_whisper import RBLNWhisperForConditionalGeneration

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

@@ -140,7 +140,6 @@ class RBLNWhisperForConditionalGeneration(RBLNModel, RBLNWhisperGenerationMixin)
         #     input_stride = self.model.encoder.conv1.stride[0] * self.model.encoder.conv2.stride[0]
         self.model = WhisperModel(self.config)
         self.pad_token_id = self.config.pad_token_id
-        self.generation_config.forced_decoder_ids = None
 
     def can_generate(self):
         return True

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

@@ -16,25 +16,15 @@ from typing import Optional, Tuple, Union
 
 import torch
 from torch import nn
-from transformers.modeling_outputs import (
-    BaseModelOutput,
-    Seq2SeqLMOutput,
-)
+from transformers.modeling_outputs import BaseModelOutput, Seq2SeqLMOutput
 from transformers.utils import logging
 
-from ....ops import (
-    register_rbln_custom_cache_update,
-    register_rbln_custom_paged_attention,
-    register_rbln_custom_paged_causal_attention,
-)
-
 
 logger = logging.get_logger(__name__)
 
 
 class WhisperWrapper:
     def __init__(self, model, use_attention_mask, rbln_token_timestamps):
-        register_rbln_custom_cache_update()
         self.encoder = WhisperEncoderWrapper(model)
         self.decoder = WhisperDecoderWrapper(
             model, use_attention_mask=use_attention_mask, output_attentions=rbln_token_timestamps
@@ -80,9 +70,11 @@ class WhisperEncoderWrapper(torch.nn.Module):
 
         # 3. update cross_attention's past_key_value to the device-dram for optimization.
         batch_axis = torch.tensor(1, dtype=torch.int16)
-        enc_output = torch.ops.rbln_custom_ops.rbln_cache_update(cross_key_values, cross_kv, b_idx[0], batch_axis)
+        cross_key_values = torch.ops.rbln_custom_ops.rbln_cache_update(
+            cross_key_values, cross_kv, b_idx[0], batch_axis
+        )
 
-        return enc_output
+        return cross_key_values
 
 
 class WhisperDecoderWrapper(torch.nn.Module):
@@ -100,11 +92,6 @@ class WhisperDecoderWrapper(torch.nn.Module):
         It is inspired by the BART architecture, but it is designed to be flexible and can be overridden
         by subclasses to modify or add custom attributes as necessary.
         """
-        if self.use_attention_mask:
-            register_rbln_custom_paged_attention()
-        else:
-            register_rbln_custom_paged_causal_attention()
-
         self.num_layers = self.config.decoder_layers
         self.decoder = self.convert_to_rbln_conditional_generation(model)
 
@@ -310,22 +297,23 @@ class WhisperSelfAttention(WhisperAttention):
         value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
         block_size = past_key_value[0].shape[-2]
 
-        args = [
-            query_states,
-            key_states,
-            value_states,
-            past_key_value[0].view(bsz, self.num_heads, 1, -1, self.head_dim),
-            past_key_value[1].view(bsz, self.num_heads, 1, -1, self.head_dim),
-            cache_position,
-            torch.tensor(1.0, dtype=torch.float32),  # scale
-            block_tables,
-            block_size,
-        ]
+        args = {
+            "q": query_states,
+            "k": key_states,
+            "v": value_states,
+            "kcache": past_key_value[0].view(bsz, self.num_heads, 1, -1, self.head_dim),
+            "vcache": past_key_value[1].view(bsz, self.num_heads, 1, -1, self.head_dim),
+            "seq": cache_position.expand(bsz, 1),
+            "scale": torch.tensor(1.0, dtype=torch.float32),
+            "block_table": block_tables,
+            "block_size": block_size,
+        }
+
         if attention_mask is not None:
-            args.insert(3, attention_mask.unsqueeze(2))
-            attn_output = torch.ops.rbln_custom_ops.paged_attn_decode(*args)
+            args["mask"] = attention_mask.unsqueeze(2)
+            attn_output = torch.ops.rbln_custom_ops.paged_attn_decode(**args)
         else:
-            attn_output = torch.ops.rbln_custom_ops.paged_causal_attn_decode(*args)
+            attn_output = torch.ops.rbln_custom_ops.paged_causal_attn_decode(**args)
 
         attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
         attn_output = attn_output.transpose(1, 2)

{optimum_rbln-0.7.4a2.dist-info → optimum_rbln-0.7.4a4.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: optimum-rbln
-Version: 0.7.4a2
+Version: 0.7.4a4
 Summary: Optimum RBLN is the interface between the Hugging Face 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-0.7.4a2.dist-info → optimum_rbln-0.7.4a4.dist-info}/RECORD

@@ -1,7 +1,7 @@
 optimum/rbln/__init__.py,sha256=qW45z47BiNLTDtRFEhVEzr4THNFX0ygqCbdNKqI0biI,6992
-optimum/rbln/__version__.py,sha256=7nOn__ePBcQxiENj-KnOIjfCYYWSP4QlkMW59HLjtRs,513
-optimum/rbln/modeling.py,sha256=nJsAs5zs--VVOYGFjYNpqfxYIemJIK4Lr0WEzlDLdP0,8390
-optimum/rbln/modeling_base.py,sha256=dNCL-BhrWCpuOVkZaj8-MW567Tf4lLo3p3Z3ldjWJfU,21779
+optimum/rbln/__version__.py,sha256=d6hw8hJyAPi7Z3yrW8lAIxLNilvYyOIhPO-OPRYoduA,513
+optimum/rbln/modeling.py,sha256=0Hh75PVNDchAx9xd5dDSwrhV9PkT1lDuiJy-Wm2hchA,8797
+optimum/rbln/modeling_base.py,sha256=FA5K6z4K3aFQ4UyMfYEGF6doDlTig82JfcSSGHNak2U,21560
 optimum/rbln/modeling_config.py,sha256=7104bxmrvKW4Q6XTruQayiIGl8GHDFmPkJ3cknMIInE,11335
 optimum/rbln/diffusers/__init__.py,sha256=Hq87CbtiCy85YmK2SB-OmUyfv77oe3j4bsTenTRnu6w,3623
 optimum/rbln/diffusers/modeling_diffusers.py,sha256=IS6Mlgexofap7f9Lefk5cKFP7ejSG_oWN3v2PX9_IDQ,20118
@@ -41,10 +41,10 @@ optimum/rbln/diffusers/pipelines/stable_diffusion_xl/__init__.py,sha256=9iIMZYvp
 optimum/rbln/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py,sha256=OvB5bxX6HUiqJeIc3uukuEmUXYEx1pTqGNOtdG2l1m8,902
 optimum/rbln/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_img2img.py,sha256=3aB1Rw-OgKytQOHwOaShbEvq_XVHPOGvsGm8pstEmKU,930
 optimum/rbln/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_inpaint.py,sha256=MzVP1wscaO1sUIiBIPJqG6zuGyez9VUbA42-JSIm-mk,930
-optimum/rbln/ops/__init__.py,sha256=LmTIX9yTfRiMDcalmb52yz5LhLRWqq3H5S94r0VDYDw,974
-optimum/rbln/ops/attn.py,sha256=OSgPoEgCwvR7HdjbnaVkFVMBcJ5RpRWcE6OCg2lVyGk,10634
-optimum/rbln/ops/flash_attn.py,sha256=wfyiCxDGf034IngzwRU160R7_DlKYpd-uWT0BDEGFks,3408
-optimum/rbln/ops/kv_cache_update.py,sha256=pxf8kAptPaQF5xE8qItvmlFOq_sgim6ZERD7AVaOtec,3221
+optimum/rbln/ops/__init__.py,sha256=rSz6mfC0aGbNYjMaNSsOZSPYxPRenW8DWbNpAkjTfAc,703
+optimum/rbln/ops/attn.py,sha256=x02yFLk7FcONFqfow0ROmVy9fmxo5Pw0SPCiDY3AZNg,9012
+optimum/rbln/ops/flash_attn.py,sha256=NmCqUdMTzgJ4sbYGj8IWXJEsLWvbuCMponR01w5DK6w,4121
+optimum/rbln/ops/kv_cache_update.py,sha256=HjnHBR-oFrJQibsVnkYb0P5_-wEma8jl0mkjkylwakU,1270
 optimum/rbln/ops/linear.py,sha256=1_7Hg-9wXxhu97fqPobotLQx17k7VPeSSL91_9Z7EDg,1018
 optimum/rbln/transformers/__init__.py,sha256=rW2wEgNpkcBwrrib2tui5sEpw04s1YUDHB50m2L7Os8,4353
 optimum/rbln/transformers/modeling_alias.py,sha256=yx7FnZQWAnrWzivaO5hI7T6i-fyLzt2tMIXG2oDNbPo,1657
@@ -54,15 +54,15 @@ optimum/rbln/transformers/models/__init__.py,sha256=Qyt9E61FDpnyAXTmRKDbv7CTtn-m
 optimum/rbln/transformers/models/auto/__init__.py,sha256=GvGbb3ZpMv-h6euXeZ42jSizoOfrL2O1uvpAnfKxYEo,1034
 optimum/rbln/transformers/models/auto/auto_factory.py,sha256=IK9jFrJ3EEzYQa9_aKpcp2TO68M5YGkA-HcfBVpA2QU,7027
 optimum/rbln/transformers/models/auto/modeling_auto.py,sha256=Un9qoqdy3dO8JBza_bTJF_6_fRVNM9QisihSgTRFI-o,3933
-optimum/rbln/transformers/models/bart/__init__.py,sha256=32HPe0_GIO0hp9U464Iv6Jd7M-1nop9g8hA1UZMHhyw,674
+optimum/rbln/transformers/models/bart/__init__.py,sha256=MqOlNX0QJvvLzTkTPd3SF-lq4PLeztXGfLHyJVMN6uU,738
 optimum/rbln/transformers/models/bart/bart_architecture.py,sha256=Oo-Cdne7igKEex8wwP-gztKJHgs5GLHQjK1oc3IZIDE,5801
 optimum/rbln/transformers/models/bart/modeling_bart.py,sha256=naFpsOSjNRG8s5QPjeAsYCk2oJCxnn0Au0aYnMKZOBY,5679
 optimum/rbln/transformers/models/bert/__init__.py,sha256=YVV7k_laU6yJBawZrgjIWjRmIF-Y4oQQHqyf8lsraQs,691
 optimum/rbln/transformers/models/bert/modeling_bert.py,sha256=p3utRqf3dv9_RkHwaMCa1EfXttNJkqCJUIZo3CeZ9YY,4674
 optimum/rbln/transformers/models/clip/__init__.py,sha256=H9vuBwrmFO0-CqZhXUrKF-uQL6igCqMlqrT1X_ELaAI,754
 optimum/rbln/transformers/models/clip/modeling_clip.py,sha256=NiSm7bHs4SReHDUr53BBWSX0Y8bkKOeUSpsBDrp8YDw,6628
-optimum/rbln/transformers/models/decoderonly/__init__.py,sha256=pDogsdpJKKB5rqnVFrRjwfhUvOSV-jZ3oARMsqSvOOQ,665
-optimum/rbln/transformers/models/decoderonly/decoderonly_architecture.py,sha256=m93-qKN7NMw3i0XDmFmttmRIRK4np_fWtLFlBb2RFgU,41351
+optimum/rbln/transformers/models/decoderonly/__init__.py,sha256=vFJvOsmUuvxLkhZELb5JKok-I1Bh_xCn1vb3pjMw1Ug,929
+optimum/rbln/transformers/models/decoderonly/decoderonly_architecture.py,sha256=gIkXCtjbcUbd4KcB_MFoPdhXM4MdrZNeKXnxK5OrT4c,41255
 optimum/rbln/transformers/models/decoderonly/modeling_decoderonly.py,sha256=qNnWSe3p2LDwNJ6utrilsqid-rQ8YLloqYkSOZamvhs,39918
 optimum/rbln/transformers/models/dpt/__init__.py,sha256=gP1tkR3XMNlHq1GT87ugIVvb2o_1eAUg1JaniXjy1Lw,651
 optimum/rbln/transformers/models/dpt/modeling_dpt.py,sha256=ZsS2SOiqcA4azULB-WFEMQZbgIoOyVUKqVKqrw_tWzA,3430
@@ -93,20 +93,20 @@ optimum/rbln/transformers/models/qwen2/__init__.py,sha256=RAMWc21W_2I6DH9xBjeNxP
 optimum/rbln/transformers/models/qwen2/modeling_qwen2.py,sha256=9-aFDvjMzPNUyGOz0qo33RE18bUFGYZ3Wt_68zb5uJY,1530
 optimum/rbln/transformers/models/qwen2/qwen2_architecture.py,sha256=XlNAMYAcDLohnSAhIFGKOPuCB5XLgzYs5ABWdeQSaZs,720
 optimum/rbln/transformers/models/seq2seq/__init__.py,sha256=EmEMV4rOYqKyruX85d0fR73-b8N6BSD6CPcbpYdBuVk,651
-optimum/rbln/transformers/models/seq2seq/modeling_seq2seq.py,sha256=XcZb57v42wju1qOJ1AKqmtJXcmz6MEWaJZ8jyzaEiTw,17701
-optimum/rbln/transformers/models/seq2seq/seq2seq_architecture.py,sha256=tvzacIZam1sIr_1BvvZ_fDr8u5dXAiYiynFdX9tArtY,18877
-optimum/rbln/transformers/models/t5/__init__.py,sha256=1skR1RmnG62WTAP3-F5P1x-V_ReFhMyirH3u56vWwvc,675
-optimum/rbln/transformers/models/t5/modeling_t5.py,sha256=Gyq5aAfkl4hBbLiR0114nDxLBs5P6YTw7hCnyuDyRrM,9494
-optimum/rbln/transformers/models/t5/t5_architecture.py,sha256=Ups6drBbYe4wEAiBLcBIyO9wqrIQbvOPFR_ybbAgR8c,9722
-optimum/rbln/transformers/models/time_series_transformers/__init__.py,sha256=RL4SO8tKEd4wQrzyU4Nv4-hhITKPhblUsBd3anXNkA8,1079
+optimum/rbln/transformers/models/seq2seq/modeling_seq2seq.py,sha256=PQQ9dZyFToNGKYp1GeKXYHKaLuAVyeZwgdCfylmWY6Y,17837
+optimum/rbln/transformers/models/seq2seq/seq2seq_architecture.py,sha256=w5pMnWbJhgYmpl5NT_vJEzcb4RfEbHfalJ371IL3wp8,18685
+optimum/rbln/transformers/models/t5/__init__.py,sha256=r6RJKPbbAoz_jCWC_yTyC9jBGYYze-vauPt1ahWax0E,725
+optimum/rbln/transformers/models/t5/modeling_t5.py,sha256=-fG-h0wwsfjZ3par0QHbXKA7hbvw_lPJOIf8iXQDOfM,8082
+optimum/rbln/transformers/models/t5/t5_architecture.py,sha256=YdFOr8LfBhEXuPsLZDfYjmUSI-a3Otrzye1bOf8NUyw,9678
+optimum/rbln/transformers/models/time_series_transformers/__init__.py,sha256=KIeTMfC-8Q4-QdTdzAybVrpPf0nVI7AuGOGAxq8C59M,1148
 optimum/rbln/transformers/models/time_series_transformers/modeling_time_series_transformers.py,sha256=1Ippt0Rmt2TxJ5X4-4tlALQOkKmOfMaTrbOLWIUIKWw,16614
-optimum/rbln/transformers/models/time_series_transformers/time_series_transformers_architecture.py,sha256=ohoP4sAxyQZwrQ6euGfRx9w_pPWAh6KT9nKC8Y9taes,14006
+optimum/rbln/transformers/models/time_series_transformers/time_series_transformers_architecture.py,sha256=XJDjQGbWXUq4ZimNojlcbm3mTDpxUMCl6tkFSzfYFl4,13769
 optimum/rbln/transformers/models/wav2vec2/__init__.py,sha256=YpgA0K-vyg9veh0eL_jxauosbRpb_kpGKHvvQLBspKM,649
 optimum/rbln/transformers/models/wav2vec2/modeling_wav2vec2.py,sha256=JYJmV52j6cBwim4RanVJryfKnV80V96ol0A-oR6o7cg,3856
-optimum/rbln/transformers/models/whisper/__init__.py,sha256=ktnNe5ri3ycCWZ_W_voFB9y9-vgGgxS1X9s8LBRZmWc,665
+optimum/rbln/transformers/models/whisper/__init__.py,sha256=Do4rVonghYiIa9zgHh4A0XtF-GQNy6UmCetO_KmIUuc,715
 optimum/rbln/transformers/models/whisper/generation_whisper.py,sha256=GIHTca3b1VtW81kp7BzKQ7f77c2t9OsEsbZetripgDo,4582
-optimum/rbln/transformers/models/whisper/modeling_whisper.py,sha256=GegyAi3a8fF0psdYsffTQ1pC4KAUqE7WYLj4ZqObWXI,18184
-optimum/rbln/transformers/models/whisper/whisper_architecture.py,sha256=DS9AQYhNkaR7sUz_loee-fFtCCYy1BUsx7_dX_o1Le8,14199
+optimum/rbln/transformers/models/whisper/modeling_whisper.py,sha256=HB7BF-SG_QIqeid-PN8Tn9qVTeiv897RUOyYzapa8zA,18127
+optimum/rbln/transformers/models/whisper/whisper_architecture.py,sha256=zEwfn8DDTbt2TN7lHKMZG9JXZc5WdW9Cp8mH4OVfo3s,13949
 optimum/rbln/transformers/models/xlm_roberta/__init__.py,sha256=fC7iNcdxBZ_6eOF2snStmf8r2M3c8O_-XcXnQEaHQCE,653
 optimum/rbln/transformers/models/xlm_roberta/modeling_xlm_roberta.py,sha256=8YNLz0bc5ze-QuU8rN-QhUfGzlSUs3iMJiWTxO3o6AM,4366
 optimum/rbln/transformers/utils/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
@@ -120,7 +120,7 @@ optimum/rbln/utils/model_utils.py,sha256=DfD_Z2qvZHqcddXqnzTM1AN8khanj3-DXK2lJvV
 optimum/rbln/utils/runtime_utils.py,sha256=5-DYniyP59nx-mrrbi7AqA77L85b4Cm5oLpaxidSyss,3699
 optimum/rbln/utils/save_utils.py,sha256=hG5uOtYmecSXZuGTvCXsTM-SiyZpr5q3InUGCCq_jzQ,3619
 optimum/rbln/utils/submodule.py,sha256=oZoGrItB8WqY4i-K9WJPlLlcLohc1YGB9OHB8_XZw3A,4071
-optimum_rbln-0.7.4a2.dist-info/METADATA,sha256=Pl6SOVN73gxS7Po-R5hMdTI_2izOrjWqyb9FhaDnr-A,5300
-optimum_rbln-0.7.4a2.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-optimum_rbln-0.7.4a2.dist-info/licenses/LICENSE,sha256=QwcOLU5TJoTeUhuIXzhdCEEDDvorGiC6-3YTOl4TecE,11356
-optimum_rbln-0.7.4a2.dist-info/RECORD,,
+optimum_rbln-0.7.4a4.dist-info/METADATA,sha256=rfxJZYj0d3KgHgaSfXvk3dV391X8kgWuAptjNfLZGvo,5300
+optimum_rbln-0.7.4a4.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+optimum_rbln-0.7.4a4.dist-info/licenses/LICENSE,sha256=QwcOLU5TJoTeUhuIXzhdCEEDDvorGiC6-3YTOl4TecE,11356
+optimum_rbln-0.7.4a4.dist-info/RECORD,,