optimum-rbln 0.2.1a1__py3-none-any.whl → 0.2.1a3__py3-none-any.whl

optimum/rbln/__init__.py

@@ -71,6 +71,7 @@ _import_structure = {
         "RBLNRobertaForSequenceClassification",
         "RBLNRobertaForMaskedLM",
         "RBLNViTForImageClassification",
+        "RBLNBertForMaskedLM",
     ],
     "diffusers": [
         "RBLNStableDiffusionPipeline",
@@ -141,6 +142,7 @@ if TYPE_CHECKING:
         RBLNAutoModelForVision2Seq,
         RBLNBartForConditionalGeneration,
         RBLNBartModel,
+        RBLNBertForMaskedLM,
         RBLNBertForQuestionAnswering,
         RBLNBertModel,
         RBLNCLIPTextModel,

optimum/rbln/__version__.py

@@ -12,5 +12,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.2.1a1'
+__version__ = version = '0.2.1a3'
 __version_tuple__ = version_tuple = (0, 2, 1)

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

@@ -12,7 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import logging
 from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple, Union
 
 import rebel
@@ -23,6 +22,7 @@ from transformers import PretrainedConfig
 
 from ....modeling import RBLNModel
 from ....modeling_config import DEFAULT_COMPILED_MODEL_NAME, RBLNCompileConfig, RBLNConfig
+from ....utils.logging import get_logger
 from ...modeling_diffusers import RBLNDiffusionMixin
 from .vae import RBLNRuntimeVAEDecoder, RBLNRuntimeVAEEncoder, _VAEDecoder, _VAEEncoder
 
@@ -31,7 +31,7 @@ if TYPE_CHECKING:
     import torch
     from transformers import AutoFeatureExtractor, AutoProcessor, AutoTokenizer, PretrainedConfig
 
-logger = logging.getLogger(__name__)
+logger = get_logger(__name__)
 
 
 class RBLNAutoencoderKL(RBLNModel):

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

@@ -12,7 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import logging
 from typing import TYPE_CHECKING
 
 import torch  # noqa: I001
@@ -20,13 +19,14 @@ from diffusers import AutoencoderKL
 from diffusers.models.autoencoders.vae import DiagonalGaussianDistribution
 from diffusers.models.modeling_outputs import AutoencoderKLOutput
 
+from ....utils.logging import get_logger
 from ....utils.runtime_utils import RBLNPytorchRuntime
 
 
 if TYPE_CHECKING:
     import torch
 
-logger = logging.getLogger(__name__)
+logger = get_logger(__name__)
 
 
 class RBLNRuntimeVAEEncoder(RBLNPytorchRuntime):

optimum/rbln/diffusers/models/controlnet.py

@@ -13,7 +13,6 @@
 # limitations under the License.
 
 import importlib
-import logging
 from typing import TYPE_CHECKING, Any, Dict, Optional, Union
 
 import torch
@@ -22,6 +21,7 @@ from transformers import PretrainedConfig
 
 from ...modeling import RBLNModel
 from ...modeling_config import RBLNCompileConfig, RBLNConfig
+from ...utils.logging import get_logger
 from ..modeling_diffusers import RBLNDiffusionMixin
 
 
@@ -29,7 +29,7 @@ if TYPE_CHECKING:
     from transformers import AutoFeatureExtractor, AutoProcessor, AutoTokenizer
 
 
-logger = logging.getLogger(__name__)
+logger = get_logger(__name__)
 
 
 class _ControlNetModel(torch.nn.Module):

optimum/rbln/diffusers/models/transformers/transformer_sd3.py

@@ -12,7 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import logging
 from typing import TYPE_CHECKING, Any, Dict, List, Optional, Union
 
 import torch
@@ -22,13 +21,14 @@ from transformers import PretrainedConfig
 
 from ....modeling import RBLNModel
 from ....modeling_config import RBLNCompileConfig, RBLNConfig
+from ....utils.logging import get_logger
 from ...modeling_diffusers import RBLNDiffusionMixin
 
 
 if TYPE_CHECKING:
     from transformers import AutoFeatureExtractor, AutoProcessor, AutoTokenizer
 
-logger = logging.getLogger(__name__)
+logger = get_logger(__name__)
 
 
 class SD3Transformer2DModelWrapper(torch.nn.Module):

optimum/rbln/diffusers/models/unets/unet_2d_condition.py

@@ -12,7 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import logging
 from dataclasses import dataclass
 from typing import TYPE_CHECKING, Any, Dict, Optional, Tuple, Union
 
@@ -22,13 +21,14 @@ from transformers import PretrainedConfig
 
 from ....modeling import RBLNModel
 from ....modeling_config import RBLNCompileConfig, RBLNConfig
+from ....utils.logging import get_logger
 from ...modeling_diffusers import RBLNDiffusionMixin
 
 
 if TYPE_CHECKING:
     from transformers import AutoFeatureExtractor, AutoProcessor, AutoTokenizer
 
-logger = logging.getLogger(__name__)
+logger = get_logger(__name__)
 
 
 class _UNet_SD(torch.nn.Module):

optimum/rbln/diffusers/pipelines/controlnet/multicontrolnet.py

@@ -12,7 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import logging
 import os
 from pathlib import Path
 from typing import TYPE_CHECKING, Any, Dict, List, Optional, Union
@@ -21,13 +20,14 @@ import torch
 from diffusers.pipelines.controlnet.multicontrolnet import MultiControlNetModel
 
 from ....modeling import RBLNModel
+from ....utils.logging import get_logger
 from ...models.controlnet import RBLNControlNetModel
 
 
 if TYPE_CHECKING:
     pass
 
-logger = logging.getLogger(__name__)
+logger = get_logger(__name__)
 
 
 class RBLNMultiControlNetModel(RBLNModel):

optimum/rbln/diffusers/pipelines/controlnet/pipeline_controlnet.py

@@ -34,16 +34,17 @@ from diffusers import StableDiffusionControlNetPipeline
 from diffusers.image_processor import PipelineImageInput
 from diffusers.pipelines.controlnet.pipeline_controlnet import retrieve_timesteps
 from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
-from diffusers.utils import deprecate, logging
+from diffusers.utils import deprecate
 from diffusers.utils.torch_utils import is_compiled_module, is_torch_version
 
 from ....utils.decorator_utils import remove_compile_time_kwargs
+from ....utils.logging import get_logger
 from ...modeling_diffusers import RBLNDiffusionMixin
 from ...models import RBLNControlNetModel
 from ...pipelines.controlnet.multicontrolnet import RBLNMultiControlNetModel
 
 
-logger = logging.get_logger(__name__)
+logger = get_logger(__name__)
 
 
 class RBLNStableDiffusionControlNetPipeline(RBLNDiffusionMixin, StableDiffusionControlNetPipeline):

optimum/rbln/modeling.py

@@ -12,7 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import logging
 from pathlib import Path
 from tempfile import TemporaryDirectory
 from typing import TYPE_CHECKING, Any, Dict, List, Optional, Union
@@ -24,13 +23,14 @@ from transformers import AutoConfig, PretrainedConfig
 
 from .modeling_base import RBLNBaseModel
 from .modeling_config import DEFAULT_COMPILED_MODEL_NAME, RBLNConfig, use_rbln_config
+from .utils.logging import get_logger
 
 
 if TYPE_CHECKING:
     from transformers import PreTrainedModel
 
 
-logger = logging.getLogger(__name__)
+logger = get_logger(__name__)
 
 
 class RBLNModel(RBLNBaseModel):

optimum/rbln/modeling_base.py

@@ -13,7 +13,6 @@
 # limitations under the License.
 
 import importlib
-import logging
 import os
 import shutil
 from abc import ABC, abstractmethod
@@ -32,6 +31,7 @@ from transformers import (
 
 from .modeling_config import RBLNCompileConfig, RBLNConfig, use_rbln_config
 from .utils.hub import PushToHubMixin, pull_compiled_model_from_hub, validate_files
+from .utils.logging import get_logger
 from .utils.runtime_utils import UnavailableRuntime
 from .utils.save_utils import maybe_load_preprocessors
 from .utils.submodule import SubModulesMixin
@@ -40,7 +40,7 @@ from .utils.submodule import SubModulesMixin
 if TYPE_CHECKING:
     from transformers import PreTrainedModel
 
-logger = logging.getLogger(__name__)
+logger = get_logger(__name__)
 
 
 class PreTrainedModel(ABC):  # noqa: F811
@@ -442,27 +442,47 @@ class RBLNBaseModel(SubModulesMixin, PushToHubMixin, PreTrainedModel):
             logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
             return
 
-        os.makedirs(save_directory, exist_ok=True)
-
         real_save_dir = self.model_save_dir / self.subfolder
         save_directory_path = Path(save_directory)
-        if os.path.exists(real_save_dir) and os.path.isdir(real_save_dir):
-            if save_directory_path.absolute() == real_save_dir.absolute():
-                raise FileExistsError(
-                    f"Cannot save model to '{save_directory}'. "
-                    f"This directory already exists and contains the model files."
-                )
-            shutil.copytree(real_save_dir, save_directory, dirs_exist_ok=True)
-            self.config.save_pretrained(save_directory)
-            if self.generation_config is not None:
-                self.generation_config.save_pretrained(save_directory)
-        else:
+
+        if not os.path.exists(real_save_dir) or not os.path.isdir(real_save_dir):
             raise FileNotFoundError(
                 f"Unable to save the model. The model directory '{real_save_dir}' does not exist or is not accessible. "
                 f"Cannot save to the specified destination '{save_directory}'. "
                 f"Please ensure the model directory exists and you have the necessary permissions to access it."
             )
 
+        if save_directory_path.absolute() == real_save_dir.absolute():
+            raise FileExistsError(
+                f"Cannot save model to '{save_directory}'. This directory already exists and contains the model files."
+            )
+
+        # Create a temporary directory next to the target directory
+        tmp_dir = save_directory + ".tmp"
+        try:
+            # Remove temporary directory if it exists from a previous failed attempt
+            if os.path.exists(tmp_dir):
+                shutil.rmtree(tmp_dir)
+
+            # First copy everything to a temporary directory
+            shutil.copytree(real_save_dir, tmp_dir)
+
+            # Save configs to the temporary directory
+            self.config.save_pretrained(tmp_dir)
+            if self.generation_config is not None:
+                self.generation_config.save_pretrained(tmp_dir)
+
+            # If everything succeeded, atomically replace the target directory
+            if os.path.exists(save_directory):
+                shutil.rmtree(save_directory)
+            os.rename(tmp_dir, save_directory)
+
+        except Exception as e:
+            # Clean up the temporary directory if anything fails
+            if os.path.exists(tmp_dir):
+                shutil.rmtree(tmp_dir)
+            raise e  # Re-raise the exception after cleanup
+
         if push_to_hub:
             return super().push_to_hub(save_directory, **kwargs)
 

optimum/rbln/ops/attn.py

@@ -152,16 +152,16 @@ def register_rbln_custom_attention_add_softmax():
         """
         return (
             q,
-            torch.empty(1, *kcache.shape[1:], device=kcache.device),
-            torch.empty(1, *vcache.shape[1:], device=vcache.device),
+            torch.empty(*kcache.shape, device=kcache.device),
+            torch.empty(*vcache.shape, device=vcache.device),
         )
 
     @register_fake("rbln_custom_ops::attn_decode_add_softmax")
     def attn_decode_add_softmax_abstract(q, k, v, m, kcache, vcache, seq, partition):
         return (
             q,
-            torch.empty(1, *kcache.shape[1:], device=kcache.device),
-            torch.empty(1, *vcache.shape[1:], device=vcache.device),
+            torch.empty(*kcache.shape, device=kcache.device),
+            torch.empty(*vcache.shape, device=vcache.device),
         )
 
     torch.library.define(

optimum/rbln/transformers/__init__.py

@@ -35,6 +35,7 @@ _import_structure = {
         "RBLNBartForConditionalGeneration",
         "RBLNBartModel",
         "RBLNBertModel",
+        "RBLNBertForMaskedLM",
         "RBLNBertForQuestionAnswering",
         "RBLNCLIPTextModel",
         "RBLNCLIPTextModelWithProjection",
@@ -92,6 +93,7 @@ if TYPE_CHECKING:
         RBLNAutoModelForVision2Seq,
         RBLNBartForConditionalGeneration,
         RBLNBartModel,
+        RBLNBertForMaskedLM,
         RBLNBertForQuestionAnswering,
         RBLNBertModel,
         RBLNCLIPTextModel,

optimum/rbln/transformers/models/__init__.py

@@ -33,7 +33,7 @@ _import_structure = {
         "RBLNAutoModelForVision2Seq",
     ],
     "bart": ["RBLNBartForConditionalGeneration", "RBLNBartModel"],
-    "bert": ["RBLNBertModel", "RBLNBertForQuestionAnswering"],
+    "bert": ["RBLNBertModel", "RBLNBertForQuestionAnswering", "RBLNBertForMaskedLM"],
     "clip": ["RBLNCLIPTextModel", "RBLNCLIPTextModelWithProjection", "RBLNCLIPVisionModel"],
     "dpt": ["RBLNDPTForDepthEstimation"],
     "exaone": ["RBLNExaoneForCausalLM"],
@@ -67,7 +67,7 @@ if TYPE_CHECKING:
         RBLNAutoModelForVision2Seq,
     )
     from .bart import RBLNBartForConditionalGeneration, RBLNBartModel
-    from .bert import RBLNBertForQuestionAnswering, RBLNBertModel
+    from .bert import RBLNBertForMaskedLM, RBLNBertForQuestionAnswering, RBLNBertModel
     from .clip import RBLNCLIPTextModel, RBLNCLIPTextModelWithProjection, RBLNCLIPVisionModel
     from .dpt import RBLNDPTForDepthEstimation
     from .exaone import RBLNExaoneForCausalLM

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

@@ -12,4 +12,4 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from .modeling_bert import RBLNBertForQuestionAnswering, RBLNBertModel
+from .modeling_bert import RBLNBertForMaskedLM, RBLNBertForQuestionAnswering, RBLNBertModel

optimum/rbln/transformers/models/bert/modeling_bert.py

@@ -13,17 +13,17 @@
 # limitations under the License.
 
 import inspect
-import logging
 from typing import TYPE_CHECKING, Any, Dict, Optional, Union
 
 from transformers import PretrainedConfig
 
 from ....modeling import RBLNModel
 from ....modeling_config import RBLNCompileConfig, RBLNConfig
-from ...modeling_generic import RBLNModelForQuestionAnswering
+from ....utils.logging import get_logger
+from ...modeling_generic import RBLNModelForMaskedLM, RBLNModelForQuestionAnswering
 
 
-logger = logging.getLogger(__name__)
+logger = get_logger(__name__)
 
 if TYPE_CHECKING:
     from transformers import AutoFeatureExtractor, AutoProcessor, AutoTokenizer
@@ -100,5 +100,9 @@ class RBLNBertModel(RBLNModel):
         return rbln_config
 
 
+class RBLNBertForMaskedLM(RBLNModelForMaskedLM):
+    rbln_model_input_names = ["input_ids", "attention_mask", "token_type_ids"]
+
+
 class RBLNBertForQuestionAnswering(RBLNModelForQuestionAnswering):
     rbln_model_input_names = ["input_ids", "attention_mask", "token_type_ids"]

optimum/rbln/transformers/models/clip/modeling_clip.py

@@ -12,7 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import logging
 from typing import TYPE_CHECKING, Any, Dict, Optional, Tuple, Union
 
 import torch
@@ -28,9 +27,10 @@ from transformers.models.clip.modeling_clip import CLIPTextModelOutput
 from ....diffusers.modeling_diffusers import RBLNDiffusionMixin
 from ....modeling import RBLNModel
 from ....modeling_config import RBLNCompileConfig, RBLNConfig
+from ....utils.logging import get_logger
 
 
-logger = logging.getLogger(__name__)
+logger = get_logger(__name__)
 
 if TYPE_CHECKING:
     from transformers import AutoFeatureExtractor, AutoProcessor, AutoTokenizer, CLIPTextModel

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

@@ -544,15 +544,19 @@ class DecoderOnlyAttention(nn.Module):
         super().__init__()
         self._original_mod = self_attn
         self.layer_idx = self_attn.layer_idx
-        self.num_heads = self._original_mod.num_heads
+        self.num_heads = getattr(self._original_mod, "num_heads", None) or getattr(
+            self._original_mod.config, "num_attention_heads"
+        )
         self.head_dim = self._original_mod.head_dim
         self._phase = "prefill"
         self.scale = torch.tensor(self.get_attn_scale())
 
         if hasattr(self._original_mod, "num_key_value_heads"):
             self.num_key_value_heads = self._original_mod.num_key_value_heads
+        elif hasattr(self._original_mod, "config") and hasattr(self._original_mod.config, "num_key_value_heads"):
+            self.num_key_value_heads = self._original_mod.config.num_key_value_heads
         else:
-            self.num_key_value_heads = self._original_mod.num_heads
+            self.num_key_value_heads = self.num_heads
 
         self.attention = self.get_attention()
         self.__post_init__()

optimum/rbln/transformers/models/dpt/modeling_dpt.py

@@ -12,7 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import logging
 from typing import TYPE_CHECKING, Any, Dict, Iterable, Optional, Union
 
 from transformers import AutoModelForDepthEstimation
@@ -20,9 +19,10 @@ from transformers.modeling_outputs import DepthEstimatorOutput
 
 from ....modeling import RBLNModel
 from ....modeling_config import RBLNCompileConfig, RBLNConfig
+from ....utils.logging import get_logger
 
 
-logger = logging.getLogger(__name__)
+logger = get_logger(__name__)
 
 if TYPE_CHECKING:
     from transformers import AutoFeatureExtractor, AutoProcessor, AutoTokenizer, PretrainedConfig

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

@@ -13,7 +13,6 @@
 # limitations under the License.
 
 import inspect
-import logging
 from pathlib import Path
 from typing import TYPE_CHECKING, Any, Callable, Dict, Optional, Tuple, Union
 
@@ -30,10 +29,11 @@ from transformers.models.llava_next.modeling_llava_next import LlavaNextCausalLM
 
 from ....modeling import RBLNModel
 from ....modeling_config import RBLNCompileConfig, RBLNConfig
+from ....utils.logging import get_logger
 from ..decoderonly.modeling_decoderonly import RBLNDecoderOnlyOutput
 
 
-logger = logging.getLogger(__name__)
+logger = get_logger(__name__)
 
 if TYPE_CHECKING:
     from transformers import (

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

@@ -13,7 +13,6 @@
 # limitations under the License.
 
 import inspect
-import logging
 from abc import ABC
 from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple, Union
 
@@ -25,10 +24,11 @@ from transformers.modeling_outputs import BaseModelOutput, Seq2SeqLMOutput
 
 from ....modeling import RBLNModel
 from ....modeling_config import RBLNCompileConfig, RBLNConfig
+from ....utils.logging import get_logger
 from ....utils.runtime_utils import RBLNPytorchRuntime
 
 
-logger = logging.getLogger(__name__)
+logger = get_logger(__name__)
 
 if TYPE_CHECKING:
     from transformers import (

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

@@ -420,7 +420,7 @@ class Seq2SeqSelfAttention(nn.Module):
         pass
 
     def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int) -> torch.Tensor:
-        return tensor.view(bsz, 1, seq_len, 1, self.num_heads, self.head_dim).transpose(2, 4)
+        return tensor.view(bsz, seq_len, 1, self.num_heads, self.head_dim).transpose(1, 3)
 
     def projection(self, hidden_states) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
         """Projects input hidden states into query, key, and value representations.
@@ -450,38 +450,21 @@ class Seq2SeqSelfAttention(nn.Module):
         key_states = self._shape(key_states, -1, bsz)
         value_states = self._shape(value_states, -1, bsz)
 
-        all_key_states = []
-        all_value_states = []
-        all_attn_output = []
-        for b_idx in range(bsz):
-            query_state = query_states[b_idx]
-            key_state = key_states[b_idx]
-            value_state = value_states[b_idx]
-            attn_mask = attention_mask[b_idx].unsqueeze(0).unsqueeze(2)
-            past_key_state = past_key_value[0].view(bsz, self.num_heads, 1, -1, self.head_dim)
-            past_value_state = past_key_value[1].view(bsz, self.num_heads, 1, -1, self.head_dim)
-
-            attn_output, key_state, value_state = self.attn_decode(
-                query_state,
-                key_state,
-                value_state,
-                attn_mask,
-                past_key_state,
-                past_value_state,
-                cache_position[b_idx][0],
-                torch.tensor(1.0, dtype=torch.float32),  # scale
-            )
-
-            attn_output = attn_output.view(1, self.num_heads, -1, self.head_dim).transpose(1, 2)
-            attn_output = attn_output.reshape(1, -1, self.num_heads * self.head_dim)
-
-            all_key_states.append(key_state.squeeze(2))
-            all_value_states.append(value_state.squeeze(2))
-            all_attn_output.append(attn_output)
+        attn_output, key_states, value_states = self.attn_decode(
+            query_states,
+            key_states,
+            value_states,
+            attention_mask.unsqueeze(
+                2
+            ),  # Unsqueeze group axis since CustomKernel expects it for group query attention
+            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.squeeze(1),
+            torch.tensor(1.0, dtype=torch.float32),  # scale
+        )
 
-        key_states = torch.cat(all_key_states, dim=0)
-        value_states = torch.cat(all_value_states, dim=0)
-        attn_output = torch.cat(all_attn_output, dim=0)
+        attn_output = attn_output.view(bsz, self.num_heads, -1, self.head_dim).transpose(1, 2)
+        attn_output = attn_output.reshape(bsz, -1, self.num_heads * self.head_dim)
 
         attn_output = self.out_proj(attn_output)
         present_key_value = (key_states, value_states)

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

@@ -147,6 +147,11 @@ class T5CrossAttention(nn.Module):
     def __init__(self, attn):
         super().__init__()
         self.attn = attn
+        self.q = attn.q
+        self.o = attn.o
+        self.n_heads = attn.n_heads
+        self.key_value_proj_dim = attn.key_value_proj_dim
+        self.inner_dim = attn.inner_dim
 
     def forward(
         self,
@@ -155,9 +160,27 @@ class T5CrossAttention(nn.Module):
         attention_mask: torch.Tensor = None,
         key_value_states: torch.Tensor = None,
     ):
-        return self.attn(
-            hidden_states=hidden_states,
-            past_key_value=past_key_value,
-            position_bias=attention_mask,
-            key_value_states=key_value_states,
-        )
+        batch_size = hidden_states.shape[0]
+
+        query_states = self.q(hidden_states)
+        query_states = query_states.view(batch_size, -1, self.n_heads, self.key_value_proj_dim).transpose(1, 2)
+
+        # reuse k,v, cross_attentions
+        key_states = past_key_value[0]
+        value_states = past_key_value[1]
+
+        # compute scores, equivalent of torch.einsum("bnqd,bnkd->bnqk", query_states, key_states), compatible with onnx op>9
+        scores = torch.matmul(query_states, key_states.transpose(3, 2))
+        scores += attention_mask
+
+        # (batch_size, n_heads, seq_length, key_length)
+        attn_weights = nn.functional.softmax(scores.float(), dim=-1).type_as(scores)
+        attn_output = torch.matmul(attn_weights, value_states)
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.view(batch_size, -1, self.inner_dim)
+        attn_output = self.o(attn_output)
+
+        outputs = (attn_output, past_key_value)
+
+        return outputs

optimum/rbln/transformers/models/wav2vec2/modeling_wav2vec2.py

@@ -12,7 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import logging
 from typing import TYPE_CHECKING, Any, Dict, Union
 
 import torch
@@ -21,9 +20,10 @@ from transformers.modeling_outputs import CausalLMOutput
 
 from ....modeling import RBLNModel
 from ....modeling_config import RBLNCompileConfig, RBLNConfig
+from ....utils.logging import get_logger
 
 
-logger = logging.getLogger(__name__)
+logger = get_logger(__name__)
 
 if TYPE_CHECKING:
     from transformers import (

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

@@ -32,6 +32,8 @@ Modified from `transformers.models.whisper.generation_whisper.py`
 """
 
 import torch
+import transformers
+from packaging import version
 from transformers import GenerationMixin
 from transformers.models.whisper.generation_whisper import WhisperGenerationMixin
 
@@ -47,17 +49,12 @@ class RBLNWhisperGenerationMixin(WhisperGenerationMixin, GenerationMixin):
         self, seek_outputs, decoder_input_ids, return_token_timestamps, generation_config, *args, **kwargs
     ):
         # remove all previously passed decoder input ids
-
-        ################################## rbln_change for 4.40.2###################################
-        # 4.40.2 has no keyword shortform, it has seperate codes from generation_fallback
-        is_shortform = kwargs.get("is_shortform", False)
-        start_idx = decoder_input_ids.shape[-1] if not is_shortform else torch.tensor(0)
+        # should happen only if it is the first generated segment
+        start_idx = decoder_input_ids.shape[-1]
 
         if isinstance(seek_outputs, torch.Tensor):
-            seek_outputs = seek_outputs[:, start_idx:]
-            return seek_outputs, seek_outputs
+            return seek_outputs[:, start_idx:], seek_outputs
 
-        ############## rbln validation#############
         if return_token_timestamps and not self.rbln_token_timestamps:
             raise RuntimeError(
                 "To use .generate() with return_token_timestamps=True, the model must be compiled with rbln_token_timestamps=True. "
@@ -67,11 +64,19 @@ class RBLNWhisperGenerationMixin(WhisperGenerationMixin, GenerationMixin):
 
         if return_token_timestamps and hasattr(generation_config, "alignment_heads"):
             num_frames = getattr(generation_config, "num_frames", None)
-            seek_outputs["token_timestamps"] = self._extract_token_timestamps(
-                seek_outputs, generation_config.alignment_heads, num_frames=num_frames
-            )
-            seek_outputs["token_timestamps"] = seek_outputs["token_timestamps"][:, start_idx:]
-
+            if version.parse(transformers.__version__) >= version.parse("4.46.0"):
+                seek_outputs["token_timestamps"] = self._extract_token_timestamps(
+                    seek_outputs,
+                    generation_config.alignment_heads,
+                    num_frames=num_frames,
+                    num_input_ids=decoder_input_ids.shape[-1],
+                )
+            else:
+                seek_outputs["token_timestamps"] = self._extract_token_timestamps(
+                    seek_outputs,
+                    generation_config.alignment_heads,
+                    num_frames=num_frames,
+                )
         seek_outputs["sequences"] = seek_outputs["sequences"][:, start_idx:]
 
         def split_by_batch_index(values, key, batch_idx):
@@ -87,15 +92,12 @@ class RBLNWhisperGenerationMixin(WhisperGenerationMixin, GenerationMixin):
 
         sequence_tokens = seek_outputs["sequences"]
 
-        ##################################### thkim change #############################################
         valid_seekoutputs = []
         for k, v in seek_outputs.items():
             if v is not None and len(v) > 0 and v[0] is not None:
                 valid_seekoutputs.append((k, v))
         seek_outputs = [
-            {k: split_by_batch_index(v, k, i) for k, v in valid_seekoutputs}
-            # {k: split_by_batch_index(v, k, i, is_shortform) for k, v in seek_outputs.items()}
-            for i in range(sequence_tokens.shape[0])
+            {k: split_by_batch_index(v, k, i) for k, v in valid_seekoutputs} for i in range(sequence_tokens.shape[0])
         ]
 
         return sequence_tokens, seek_outputs

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

@@ -13,7 +13,6 @@
 # limitations under the License.
 
 import inspect
-import logging
 from typing import TYPE_CHECKING, Any, Callable, Dict, List, Optional, Union
 
 import rebel
@@ -30,12 +29,13 @@ from transformers.modeling_outputs import BaseModelOutput, Seq2SeqLMOutput
 
 from ....modeling import RBLNModel
 from ....modeling_config import RBLNCompileConfig, RBLNConfig
+from ....utils.logging import get_logger
 from ....utils.runtime_utils import RBLNPytorchRuntime
 from .generation_whisper import RBLNWhisperGenerationMixin
 from .whisper_architecture import WhisperWrapper
 
 
-logger = logging.getLogger(__name__)
+logger = get_logger(__name__)
 
 if TYPE_CHECKING:
     from transformers import AutoFeatureExtractor, AutoProcessor, PretrainedConfig, PreTrainedModel

optimum/rbln/transformers/models/xlm_roberta/modeling_xlm_roberta.py

@@ -13,7 +13,6 @@
 # limitations under the License.
 
 import inspect
-import logging
 from typing import TYPE_CHECKING, Optional, Union
 
 import torch
@@ -21,9 +20,10 @@ from transformers import PretrainedConfig
 
 from ....modeling import RBLNModel
 from ....modeling_config import RBLNCompileConfig, RBLNConfig
+from ....utils.logging import get_logger
 
 
-logger = logging.getLogger(__name__)
+logger = get_logger(__name__)
 
 if TYPE_CHECKING:
     from transformers import AutoFeatureExtractor, AutoProcessor, AutoTokenizer

optimum/rbln/utils/save_utils.py

@@ -30,14 +30,15 @@
 Refer to huggingface/optimum/blob/4fdeea77d71e79451ba53e0c1f9d8f37e9704268/optimum/utils/save_utils.py
 """
 
-import logging
 from pathlib import Path
 from typing import List, Union
 
 from transformers import AutoFeatureExtractor, AutoProcessor, AutoTokenizer
 
+from .logging import get_logger
 
-logger = logging.getLogger(__name__)
+
+logger = get_logger(__name__)
 
 
 def maybe_load_preprocessors(

{optimum_rbln-0.2.1a1.dist-info → optimum_rbln-0.2.1a3.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: optimum-rbln
-Version: 0.2.1a1
+Version: 0.2.1a3
 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
@@ -28,7 +28,7 @@ Requires-Dist: packaging>=24.1
 Requires-Dist: torch<=2.5.1
 Requires-Dist: torchaudio<=2.5.1
 Requires-Dist: torchvision<=0.20.1
-Requires-Dist: transformers==4.45.2
+Requires-Dist: transformers==4.48.3
 Description-Content-Type: text/markdown
 
 

{optimum_rbln-0.2.1a1.dist-info → optimum_rbln-0.2.1a3.dist-info}/RECORD

@@ -1,23 +1,23 @@
-optimum/rbln/__init__.py,sha256=zTum8bjtVKdT9dq56vyQZy1eG2UUhEbiaGyvtokVDok,6036
-optimum/rbln/__version__.py,sha256=Hh8jxA0Z8Gd3AqkgqwtySq_B5l73iIwQzKQ1Y-52wbw,413
-optimum/rbln/modeling.py,sha256=OQGLkzlE3vD3O-ZeE1Z0jK-QCqWy1V46pSCOdmehFTI,8267
-optimum/rbln/modeling_base.py,sha256=sU5Tr3SmhQZPsbKz5xo-FqMU1gC4Xd4m9xZVIx2NY7I,20359
+optimum/rbln/__init__.py,sha256=sLCjJu_MLZEKDOwHIlJP4u4GzGZx-1kqHTYGw5B4xDg,6096
+optimum/rbln/__version__.py,sha256=Qa8tLTuiehljsgp_ibSY6aee43cZYh5J_fQ5zMTZ6SA,413
+optimum/rbln/modeling.py,sha256=REImAAKO82CqSNABR-9E1jJEsWch9amSOwOOQhFEYLY,8283
+optimum/rbln/modeling_base.py,sha256=_5M8hVySDwCJ6qfeku2_nJAPu_5JLfAUu3HO1bc3ALM,21098
 optimum/rbln/modeling_config.py,sha256=7104bxmrvKW4Q6XTruQayiIGl8GHDFmPkJ3cknMIInE,11335
 optimum/rbln/diffusers/__init__.py,sha256=68FTAMpbbMflm8qiSqfM5J2_gFb3iU3fng6AL0TG47A,2913
 optimum/rbln/diffusers/modeling_diffusers.py,sha256=E1x-iOKEJCUB6ml0RgtFEVPPk6J6pqEF-JTEyOZzOyc,14928
 optimum/rbln/diffusers/models/__init__.py,sha256=aSL5_yd-y8Q6DxNvfQ-yl-BUNyMzI1P6AikjQMKZzpI,1357
-optimum/rbln/diffusers/models/controlnet.py,sha256=AWX_ZFpzyNFEs-B7xHaXnO9grYkNYlLFzFDPdSoKexs,10494
+optimum/rbln/diffusers/models/controlnet.py,sha256=EM_HlzCdaZdnnK0oGpY2fQeigPqHhlwh4NHCzlmoumI,10512
 optimum/rbln/diffusers/models/autoencoders/__init__.py,sha256=nMfnwEwuOje-qKofAw-uOWUWcYV_YvnaN68IGfDdqHg,645
-optimum/rbln/diffusers/models/autoencoders/autoencoder_kl.py,sha256=M4GyaBlUyXetieaGQvfK3pDP5FIvy6hDTMTfOT2ymz8,9201
-optimum/rbln/diffusers/models/autoencoders/vae.py,sha256=Ejazs1JcSr-l0fyYQAtqQlTufbnCmlR_GZmadMkbNn4,2494
+optimum/rbln/diffusers/models/autoencoders/autoencoder_kl.py,sha256=rCbC32bJnfXtsLdVvNVVHpRAkCYy6jeCSwIZ-JSReWk,9220
+optimum/rbln/diffusers/models/autoencoders/vae.py,sha256=A-F2TRJ2vL4gNXiMT_hRGeanIFKWxJ1QaKmYVp41rwI,2513
 optimum/rbln/diffusers/models/transformers/__init__.py,sha256=TEhARgQJx_NUZzI6M8gt3aWbdzmLHnM6FMSQd9M9zCk,654
-optimum/rbln/diffusers/models/transformers/transformer_sd3.py,sha256=pnzKO7EXuEyVWIxvDMl2R0EkT_BfeMI8DwubwkGBvNY,7289
+optimum/rbln/diffusers/models/transformers/transformer_sd3.py,sha256=n_krmMgiRxWrG--567PNpk58EG_X7x7H4gidIkRvwjo,7308
 optimum/rbln/diffusers/models/unets/__init__.py,sha256=MaICuK9CWjgzejXy8y2NDrphuEq1rkzanF8u45k6O5I,655
-optimum/rbln/diffusers/models/unets/unet_2d_condition.py,sha256=tf0ILmn2t-D0omITpF5T1A1UnJJ3VA7sfluVXuQWIks,14002
+optimum/rbln/diffusers/models/unets/unet_2d_condition.py,sha256=Z0-eAZw1Gah24y6uOO5m9-GRruBppCSdV2NQZLNtBaI,14021
 optimum/rbln/diffusers/pipelines/__init__.py,sha256=i8AQJSoV9clLTill7wP5ECci6E7lC2gBaNuqfhYklZk,2469
 optimum/rbln/diffusers/pipelines/controlnet/__init__.py,sha256=n1Ef22TSeax-kENi_d8K6wGGHSNEo9QkUeygELHgcao,983
-optimum/rbln/diffusers/pipelines/controlnet/multicontrolnet.py,sha256=mwQRhKuEsuoFn-49UKobd0a7nNIzPAqRLwZ6ftXCr-s,4094
-optimum/rbln/diffusers/pipelines/controlnet/pipeline_controlnet.py,sha256=NFlE6-gdK1QgfQy5F7IF5ZAxpwDethOv1AaTOVDxtbU,35163
+optimum/rbln/diffusers/pipelines/controlnet/multicontrolnet.py,sha256=JWKtnZYBIfgmbAo0SLFIvHBQCv2BPSFNvpcdjG4GUOY,4113
+optimum/rbln/diffusers/pipelines/controlnet/pipeline_controlnet.py,sha256=dGdw5cwJLS4CLv6IHskk5ZCcPgS7UDuHKbfOZ8ojNUs,35187
 optimum/rbln/diffusers/pipelines/controlnet/pipeline_controlnet_img2img.py,sha256=7xCiXrH4ToCTHohVGFXqO7_f9G8HShYaHgZxoMZARkQ,33664
 optimum/rbln/diffusers/pipelines/controlnet/pipeline_controlnet_sd_xl.py,sha256=Gzt2wg4dgFg0TV3Bu0cs8Xru3wVrxWUxxgciwZ-QKLE,44755
 optimum/rbln/diffusers/pipelines/controlnet/pipeline_controlnet_sd_xl_img2img.py,sha256=RfwxNX_zQWFtvvFQJ5bt3qtHbdYdQV_3XLHm9WYCKOs,46084
@@ -34,29 +34,29 @@ optimum/rbln/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_x
 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=MbAHevg59fXQOFgrnsFFZ97s8-YrgvYCYML_sgKEEfM,816
-optimum/rbln/ops/attn.py,sha256=rB0xYhUxpb1o1JISMNJnPffr7qF8DwnuecOSMWPVGoA,9728
+optimum/rbln/ops/attn.py,sha256=QYvSMg4ps_PenHwpLVhuYRoOAFvHIo19nY0ZEdj4nTE,9700
 optimum/rbln/ops/flash_attn.py,sha256=Zn5nkouY3kk6MBivQpPjgGh4oepjpi8F3tnTrmrNfpg,2304
 optimum/rbln/ops/kv_cache_update.py,sha256=9W4WCO1Dtfy0u5i978JJRa7uLbqrfR2lHuoPynb07fw,3143
-optimum/rbln/transformers/__init__.py,sha256=8NHC8MpmWc8fteSBtFNJ729LON6FRHI2J7zifCAAXZ4,4107
+optimum/rbln/transformers/__init__.py,sha256=SdOjpa4Iufo6aOJPvjQwD_vz28dmmNV9AtF5Cz9ajLM,4167
 optimum/rbln/transformers/modeling_alias.py,sha256=yx7FnZQWAnrWzivaO5hI7T6i-fyLzt2tMIXG2oDNbPo,1657
 optimum/rbln/transformers/modeling_generic.py,sha256=SD7XjpjnCn-ejNAUWgkaaHV6Fv2Y6K-hbXEXXb9W_H4,18177
 optimum/rbln/transformers/modeling_rope_utils.py,sha256=3zwkhYUyTZhxCJUSmwCc88iiY1TppRWEY9ShwUqNB2k,14293
-optimum/rbln/transformers/models/__init__.py,sha256=GBCGLGdo_HMEkGhZSevOgt-M9KMETaKUqRmgEf3WpQE,3639
+optimum/rbln/transformers/models/__init__.py,sha256=wucrA1ybpDfNcrySwdVeK5PZEYl-3ONXJvGpHGTvteo,3683
 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/bart_architecture.py,sha256=dTkgMpNkyh4vT_mZU5tQ5bvH_lRZfRjaJ1gIHvJkmgs,5479
 optimum/rbln/transformers/models/bart/modeling_bart.py,sha256=ADRbE-5N3xJ60AzzjJ4BZs_THmB71qs4XTr9iFqsEqE,5667
-optimum/rbln/transformers/models/bert/__init__.py,sha256=_pEfofte9Ss8y2dgCbwl3FCHr6X2LNSm9VMMMS1vVh0,670
-optimum/rbln/transformers/models/bert/modeling_bert.py,sha256=ikAcgsx5zIeMbmyUWusWTuX35q_gm8ZehEll-xOIJFc,4497
+optimum/rbln/transformers/models/bert/__init__.py,sha256=YVV7k_laU6yJBawZrgjIWjRmIF-Y4oQQHqyf8lsraQs,691
+optimum/rbln/transformers/models/bert/modeling_bert.py,sha256=-nv-sgmHkyHQIoQvF8-lXOJiL4eaa1pq8MpdN4uRi9M,4668
 optimum/rbln/transformers/models/clip/__init__.py,sha256=ssJqlEt318ti2QaEakGh_tO3Ap1VSPCVF-ymUuvjAJs,698
-optimum/rbln/transformers/models/clip/modeling_clip.py,sha256=LGytQ33WGV2qqXnj_1dMiPN63ytL1JlNQlc3aXaG1bA,5705
+optimum/rbln/transformers/models/clip/modeling_clip.py,sha256=E1QfVNq1sTCp7uvuha1ZPfXMwvMTkGV9L4oFdmy1w4g,5724
 optimum/rbln/transformers/models/decoderonly/__init__.py,sha256=pDogsdpJKKB5rqnVFrRjwfhUvOSV-jZ3oARMsqSvOOQ,665
-optimum/rbln/transformers/models/decoderonly/decoderonly_architecture.py,sha256=73sWpSDgm6_ztiGPsGEljJo-a7AisTAqFb7s-UaShVE,36292
+optimum/rbln/transformers/models/decoderonly/decoderonly_architecture.py,sha256=BjQHwoPZfM-KUQzxm4AU-PdmoMgLxnCG6kfSpGjUvrk,36578
 optimum/rbln/transformers/models/decoderonly/modeling_decoderonly.py,sha256=mAgRRMGVHvTUjJBDlmUOjNhSNjprKSD7tLeFknrx0Rw,25810
 optimum/rbln/transformers/models/dpt/__init__.py,sha256=gP1tkR3XMNlHq1GT87ugIVvb2o_1eAUg1JaniXjy1Lw,651
-optimum/rbln/transformers/models/dpt/modeling_dpt.py,sha256=HS_f7bL2CvbWP_7NGMzPYb0GdHRE2xHF0e1DkzlRdRE,3411
+optimum/rbln/transformers/models/dpt/modeling_dpt.py,sha256=ZsS2SOiqcA4azULB-WFEMQZbgIoOyVUKqVKqrw_tWzA,3430
 optimum/rbln/transformers/models/exaone/__init__.py,sha256=zYH_5tVa8-juEdsOIky7I33WSC3Zuhoq1upI0OHYeVw,859
 optimum/rbln/transformers/models/exaone/exaone_architecture.py,sha256=thzWLVz3eUcst4IPiOavta5QeXZw7JQwwfdIzQ_x6Ns,3029
 optimum/rbln/transformers/models/exaone/modeling_exaone.py,sha256=WjyH8PmsMljSea7kJn_Cq1FJ96OXwXAoU7hv2Q8zUnI,1747
@@ -70,7 +70,7 @@ optimum/rbln/transformers/models/llama/__init__.py,sha256=jo_j_eIrHYGNEhR5lb6g3r
 optimum/rbln/transformers/models/llama/llama_architecture.py,sha256=S7MCPfyjG5eUqgaS-QNBB0ApUD6wnb5fR0RHq7k7-pA,728
 optimum/rbln/transformers/models/llama/modeling_llama.py,sha256=Z3iony7icoFhRQ11MAuFx9UF03uJCsvJQZ6bxHXlrgk,1530
 optimum/rbln/transformers/models/llava_next/__init__.py,sha256=VLieyWm-UgvuNxw9B38wrL1Jsa09NBDX_ebABmdpTbs,670
-optimum/rbln/transformers/models/llava_next/modeling_llava_next.py,sha256=shmL1ATB-gfRC-AFW1_GgHSUtbT-ZFQbecMEpVBgn-o,26379
+optimum/rbln/transformers/models/llava_next/modeling_llava_next.py,sha256=_8zKsI-Kj4bbsPLnERJqg-0oC6EyAWrmnxvszsAtRaA,26398
 optimum/rbln/transformers/models/midm/__init__.py,sha256=UJSaErsF-z6dZERIS143WTaygffZyzEGqoQ2ZPDiM-c,855
 optimum/rbln/transformers/models/midm/midm_architecture.py,sha256=mueRmMGX6UplZb0C0RFdUOa9lsNH8YJHV6rYrDLOdlQ,5302
 optimum/rbln/transformers/models/midm/modeling_midm.py,sha256=GG25BozEZriAL-OPFGpzOjyDtSFB-NfeiLJTDAqxe20,1734
@@ -84,19 +84,19 @@ 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=4GHlLf6xm7a8YegYVX-zhIqk4ouwHCzQuj8Z-jXSFJw,15407
-optimum/rbln/transformers/models/seq2seq/seq2seq_architecture.py,sha256=JTcLA-xnSs6KD8xUA6RY0QMWoXChnHvmhyKa_4039UY,19053
+optimum/rbln/transformers/models/seq2seq/modeling_seq2seq.py,sha256=2hkCPvaiyS16zdtUiJKhvpk1qJfsXVLrAQPgAtixCg0,15426
+optimum/rbln/transformers/models/seq2seq/seq2seq_architecture.py,sha256=15yoF-wyhcLcK-Z2MOUmyPlkOMNTVOJ013uBepqtpxA,18387
 optimum/rbln/transformers/models/t5/__init__.py,sha256=1skR1RmnG62WTAP3-F5P1x-V_ReFhMyirH3u56vWwvc,675
 optimum/rbln/transformers/models/t5/modeling_t5.py,sha256=MFs-3yYviV1QqSpsTB2GarTEs9wGH5AYofksLQLMBXg,8043
-optimum/rbln/transformers/models/t5/t5_architecture.py,sha256=_K_IROovNH7zZIj2E1datCLIWbRpAg181p03GDTTtXE,6209
+optimum/rbln/transformers/models/t5/t5_architecture.py,sha256=kkjErS42mW2jv5O_xL7BaKobvvqy7BGmYOowKyHakvI,7189
 optimum/rbln/transformers/models/wav2vec2/__init__.py,sha256=YpgA0K-vyg9veh0eL_jxauosbRpb_kpGKHvvQLBspKM,649
-optimum/rbln/transformers/models/wav2vec2/modeling_wav2vec2.py,sha256=Ws0tw2j9Mp8BREP6nI-Ann_U0rhkqofaQFCKoepDYRA,3837
+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/generation_whisper.py,sha256=eT4pEmahW4VqDKP6VGuuOjiBnoG1j3xkCNR4U2FKXDI,4707
-optimum/rbln/transformers/models/whisper/modeling_whisper.py,sha256=EIiP3DVovFNDMEuE_EaUpX_uFoAlsFVnd5svawp-VGQ,15821
+optimum/rbln/transformers/models/whisper/generation_whisper.py,sha256=GIHTca3b1VtW81kp7BzKQ7f77c2t9OsEsbZetripgDo,4582
+optimum/rbln/transformers/models/whisper/modeling_whisper.py,sha256=0nBADNxE0A1ozBbRutTBvxpo_Y1qkOycT_zronkN-ZU,15840
 optimum/rbln/transformers/models/whisper/whisper_architecture.py,sha256=eP3UgkwCRaaFjc5Jc4ZEiWxr3-L7oJx9KzpJ7eFkwUs,13158
 optimum/rbln/transformers/models/xlm_roberta/__init__.py,sha256=fC7iNcdxBZ_6eOF2snStmf8r2M3c8O_-XcXnQEaHQCE,653
-optimum/rbln/transformers/models/xlm_roberta/modeling_xlm_roberta.py,sha256=jqQXBl0pezmd7eXallgWWvS_a6jXFVHBVA0ily5USPg,4748
+optimum/rbln/transformers/models/xlm_roberta/modeling_xlm_roberta.py,sha256=lKSeL3RUwIyfuca2jZ6SFV4N59EJS4UD59JMUfh3BiA,4767
 optimum/rbln/transformers/utils/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 optimum/rbln/transformers/utils/rbln_quantization.py,sha256=gwBVHf97sQgPNmGa0wq87E8mPyrtXYhMnO4X4sKp3c8,7639
 optimum/rbln/utils/__init__.py,sha256=ieDBT2VFTt2E0M4v_POLBpuGW9LxSydpb_DuPd6PQqc,712
@@ -106,9 +106,9 @@ optimum/rbln/utils/import_utils.py,sha256=ec-tISKIjUPHIfjzj6p-w78NVejWVBohb59f7J
 optimum/rbln/utils/logging.py,sha256=VKKBmlQSdg6iZCGmAXaWYiW67K84jyp1QJhLQSSjPPE,3453
 optimum/rbln/utils/model_utils.py,sha256=DfD_Z2qvZHqcddXqnzTM1AN8khanj3-DXK2lJvVxDvs,1278
 optimum/rbln/utils/runtime_utils.py,sha256=5-DYniyP59nx-mrrbi7AqA77L85b4Cm5oLpaxidSyss,3699
-optimum/rbln/utils/save_utils.py,sha256=W5ON-90xLcz1suFlZwOhmYB5Mf9XSTu00xrfTfyL88U,3608
+optimum/rbln/utils/save_utils.py,sha256=hG5uOtYmecSXZuGTvCXsTM-SiyZpr5q3InUGCCq_jzQ,3619
 optimum/rbln/utils/submodule.py,sha256=oZoGrItB8WqY4i-K9WJPlLlcLohc1YGB9OHB8_XZw3A,4071
-optimum_rbln-0.2.1a1.dist-info/METADATA,sha256=QqxzPcaWJ57Dtaws6oRFO_tCjZPjZkWQldxpWHXpg9w,5300
-optimum_rbln-0.2.1a1.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-optimum_rbln-0.2.1a1.dist-info/licenses/LICENSE,sha256=QwcOLU5TJoTeUhuIXzhdCEEDDvorGiC6-3YTOl4TecE,11356
-optimum_rbln-0.2.1a1.dist-info/RECORD,,
+optimum_rbln-0.2.1a3.dist-info/METADATA,sha256=umGg7JkKhTcNc5AOyzubqzpoPXnGY1WosDi48dfAROw,5300
+optimum_rbln-0.2.1a3.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+optimum_rbln-0.2.1a3.dist-info/licenses/LICENSE,sha256=QwcOLU5TJoTeUhuIXzhdCEEDDvorGiC6-3YTOl4TecE,11356
+optimum_rbln-0.2.1a3.dist-info/RECORD,,