optimum-rbln 0.2.1a2__py3-none-any.whl → 0.2.1a4__py3-none-any.whl

optimum/rbln/__version__.py

@@ -12,5 +12,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.2.1a2'
+__version__ = version = '0.2.1a4'
 __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/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 ....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

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/modeling_decoderonly.py

@@ -38,34 +38,188 @@ from .decoderonly_architecture import (
 logger = get_logger()
 
 if TYPE_CHECKING:
-    from transformers import AutoFeatureExtractor, AutoProcessor, AutoTokenizer, PretrainedConfig
+    from transformers import AutoFeatureExtractor, AutoProcessor, AutoTokenizer
 
 
 class RBLNRuntimeModel(RBLNPytorchRuntime):
     mandatory_members = ["main_input_name", "embed_tokens"]
 
+    def __init__(
+        self,
+        runtime: rebel.Runtime,
+        phase: str,
+        batch_size: int,
+        dec_attn_mask: torch.Tensor,
+        **kwargs: Any,
+    ) -> None:
+        super().__init__(runtime, **kwargs)
+        self.phase = phase
+        self.batch_size = batch_size
+
+        # shared tensor between prefill and decode phase
+        self.dec_attn_mask = dec_attn_mask
+
+        if self.phase == "prefill":
+            vocab_size = kwargs.pop("vocab_size")
+            self.max_seq_len = kwargs.pop("max_seq_len")
+            self.prefill_chunk_size = kwargs.pop("prefill_chunk_size")
+            self.output_size = [1, 1, vocab_size]
+            self.causal_mask = 1 - torch.triu(
+                torch.ones(1, 1, self.prefill_chunk_size, self.prefill_chunk_size), diagonal=1
+            )
+
     def forward(
         self,
-        input_ids: torch.LongTensor,
-        inputs_embeds: torch.Tensor,
-        attention_mask: torch.Tensor,
-        cache_position: torch.Tensor,
-        **kwargs,
+        input_ids: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        cache_position: torch.Tensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        batch_idx: Optional[int] = None,
     ):
+        if input_ids is None and inputs_embeds is None:
+            raise ValueError("Either `input_ids` or `inputs_embeds` must be provided.")
+
         if inputs_embeds is None:
-            inp = input_ids
+            inputs = input_ids
             if self.embed_tokens is not None:
-                inp = self.embed_tokens(inp)
+                inputs = self.embed_tokens(inputs)
         else:
-            inp = inputs_embeds
+            inputs = inputs_embeds
 
-        return super().forward(
-            inp,
-            attention_mask,
+        if self.phase == "decode":
+            return self.decode_forward(
+                inputs,
+                cache_position,
+                attention_mask=attention_mask,
+            )
+        else:
+            return self.prefill_forward(inputs, cache_position, attention_mask, batch_idx)
+
+    def decode_forward(
+        self,
+        inputs: torch.Tensor,
+        cache_position: torch.Tensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+    ) -> torch.FloatTensor:
+        batch_size = inputs.shape[0]
+        if batch_size != self.batch_size:
+            raise RuntimeError(
+                f"Batch size mismatch: got {batch_size}, expected {self.batch_size} (compiled batch size)."
+            )
+
+        if batch_size != cache_position.shape[0]:
+            raise RuntimeError(f"Cache position size mismatch: got {cache_position.shape[0]}, expected {batch_size}.")
+
+        if attention_mask is None:
+            for b_idx in range(batch_size):
+                decoding_step = cache_position[b_idx].item()
+                if not (0 <= decoding_step < self.dec_attn_mask.shape[-1]):
+                    raise ValueError(
+                        f"Decoding step {decoding_step} out of bounds for attention mask with shape {self.dec_attn_mask.shape}."
+                    )
+                self.dec_attn_mask[b_idx, :, :, decoding_step] = 1
+
+        logits = super().forward(
+            inputs,
+            self.dec_attn_mask if attention_mask is None else attention_mask,
             cache_position,
-            **kwargs,
         )
 
+        return logits
+
+    def prefill_forward(
+        self,
+        inputs: torch.Tensor,
+        cache_position: torch.Tensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        batch_idx: int = 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 batch_idx is None or batch_idx >= self.batch_size:
+            raise RuntimeError(
+                f"Invalid batch_idx ({batch_idx}). It must be a non-null value less than the batch size ({self.batch_size})."
+            )
+
+        # Handle continuous batching in a compiled graph by extracting valid inputs
+        # If an attention mask is provided, select only the valid (non-masked) inputs
+        inputs = inputs[:, attention_mask.bool()] if attention_mask is not None else inputs
+
+        query_length = inputs.shape[1]
+        if query_length > self.max_seq_len:
+            raise ValueError(
+                f"Input length ({query_length}) exceeds the maximum allowed sequence length ({self.max_seq_len})."
+            )
+
+        # Initialize attention mask for chunked processing
+        chunked_attention_mask = torch.zeros(1, 1, self.prefill_chunk_size, self.max_seq_len, dtype=torch.float32)
+
+        # Buffer for storing output logits
+        out_buffers = [
+            torch.empty(
+                size=self.output_size,
+                dtype=torch.float32,
+                device="cpu",
+            )
+        ]
+
+        # Process input in chunks of size `prefill_chunk_size`
+        for step in range(0, query_length, self.prefill_chunk_size):
+            # Pad input and cache_position if the last chunk is smaller than `prefill_chunk_size`
+            if (step + self.prefill_chunk_size) > query_length:
+                padding_size = step + self.prefill_chunk_size - query_length
+                # inputs_embeds
+                if inputs.dim() == 3:
+                    inputs = torch.nn.functional.pad(inputs, (0, 0, 0, padding_size))
+                # inputs_ids
+                else:
+                    inputs = torch.nn.functional.pad(inputs, (0, padding_size))
+
+                cache_position = torch.cat(
+                    [
+                        cache_position,
+                        torch.arange(
+                            query_length,
+                            step + self.prefill_chunk_size,
+                            dtype=torch.int32,
+                        ).unsqueeze(0),
+                    ],
+                    dim=-1,
+                )
+
+            # Extract the current chunk of inputs and cache positions
+            input_chunk = inputs[:, step : step + self.prefill_chunk_size]
+            cache_pos_chunk = cache_position[:, step : step + self.prefill_chunk_size]
+
+            # Update attention mask to ensure proper causal behavior
+            if step >= self.prefill_chunk_size:
+                chunked_attention_mask[:, :, :, step - self.prefill_chunk_size : step] = 1
+            chunked_attention_mask[:, :, :, step : step + self.prefill_chunk_size] = self.causal_mask
+
+            # Define batch position and query position
+            batch_position = torch.tensor(batch_idx, dtype=torch.int16)
+            query_position = torch.tensor((query_length - 1) % self.prefill_chunk_size, dtype=torch.int16)
+
+            # Forward pass for the current chunk
+            logits = super().forward(
+                input_chunk,
+                chunked_attention_mask,
+                cache_pos_chunk,
+                batch_position,
+                query_position,
+                out=out_buffers,
+            )
+
+        # Update decoder attention mask with processed KV-cache length from prefill phase
+        self.dec_attn_mask[batch_idx].fill_(0)
+        self.dec_attn_mask[batch_idx, :, :, :query_length] = 1
+
+        return logits
+
 
 @dataclass
 class RBLNDecoderOnlyOutput(ModelOutput):
@@ -103,13 +257,6 @@ class RBLNDecoderOnlyModelForCausalLM(RBLNModel):
         self.max_seq_len = self.rbln_config.model_cfg["max_seq_len"]
         self.prefill_chunk_size = self.rbln_config.model_cfg["prefill_chunk_size"]
 
-        self.prefill_attention_mask = torch.zeros(1, 1, self.prefill_chunk_size, self.max_seq_len, dtype=torch.float32)
-        self.causal_mask = 1 - torch.triu(
-            torch.ones(1, 1, self.prefill_chunk_size, self.prefill_chunk_size), diagonal=1
-        )
-        self.dec_attn_mask_init = torch.zeros(1, 1, 1, self.max_seq_len, dtype=torch.float32)
-        self.dec_attn_mask = torch.zeros(self.batch_size, 1, 1, self.max_seq_len, dtype=torch.float32)
-
         main_input_name = self.main_input_name
         if self.rbln_config.model_cfg["use_inputs_embeds"]:
             main_input_name = "inputs_embeds"
@@ -124,11 +271,25 @@ class RBLNDecoderOnlyModelForCausalLM(RBLNModel):
         else:
             self.embed_tokens = None
 
+        dec_attn_mask = torch.zeros(self.batch_size, 1, 1, self.max_seq_len, dtype=torch.float32)
         self.prefill_decoder = RBLNRuntimeModel(
-            runtime=self.model[0], main_input_name=main_input_name, embed_tokens=self.embed_tokens
+            runtime=self.model[0],
+            main_input_name=main_input_name,
+            embed_tokens=self.embed_tokens,
+            phase="prefill",
+            batch_size=self.batch_size,
+            dec_attn_mask=dec_attn_mask,
+            vocab_size=self.config.vocab_size,
+            max_seq_len=self.max_seq_len,
+            prefill_chunk_size=self.prefill_chunk_size,
         )
         self.decoder = RBLNRuntimeModel(
-            runtime=self.model[1], main_input_name=main_input_name, embed_tokens=self.embed_tokens
+            runtime=self.model[1],
+            main_input_name=main_input_name,
+            embed_tokens=self.embed_tokens,
+            phase="decode",
+            batch_size=self.batch_size,
+            dec_attn_mask=dec_attn_mask,
         )
 
     @classmethod
@@ -155,7 +316,7 @@ class RBLNDecoderOnlyModelForCausalLM(RBLNModel):
     def get_quantized_model(
         cls,
         model_id: str,
-        config: Optional[PretrainedConfig] = None,
+        config: Optional["PretrainedConfig"] = None,
         use_auth_token: Optional[Union[bool, str]] = None,
         revision: Optional[str] = None,
         force_download: bool = False,
@@ -496,32 +657,33 @@ class RBLNDecoderOnlyModelForCausalLM(RBLNModel):
         generate_idx: Optional[torch.Tensor] = None,
         **kwargs,
     ) -> Tuple[torch.FloatTensor]:
-        # prefll
+        """
+        Forward method for the RBLN-optimized model, designed for integration with the HuggingFace generate API.
+        For continuous batching, the prefill stage processes one batch at a time and updates the KV cache using batch_idx.
+        A for-loop ensures synchronization with the HuggingFace generate API.
+        The decoder stage operates as usual, processing inputs in batch mode.
+        """
+        # Prefll
         if cache_position is None:
             logits = []
-            input_tensors = inputs_embeds if inputs_embeds is not None else input_ids
-            batch_size = input_tensors.shape[0]
+            inputs = inputs_embeds if inputs_embeds is not None else input_ids
+            batch_size = inputs.shape[0]
 
             for b_idx in range(batch_size):
-                # Transform inputs as vllm format
-                if attention_mask is not None:
-                    input_tensor = input_tensors[b_idx : b_idx + 1, attention_mask[b_idx].bool()]
-                else:
-                    input_tensor = input_tensors[b_idx : b_idx + 1]
-
                 cache_position = torch.arange(0, generate_idx[b_idx].item(), dtype=torch.int32).unsqueeze(0)
-
-                logit = self._forward_prefill(
-                    input_ids=input_tensor if inputs_embeds is None else None,
-                    inputs_embeds=input_tensor if inputs_embeds is not None else None,
+                logit = self.prefill_decoder(
+                    input_ids=inputs[b_idx : b_idx + 1] if inputs_embeds is None else None,
+                    inputs_embeds=inputs[b_idx : b_idx + 1] if inputs_embeds is not None else None,
+                    attention_mask=attention_mask[b_idx] if attention_mask is not None else None,
                     cache_position=cache_position,
                     batch_idx=b_idx,
                 )
                 logits.append(logit)
+
             logits = torch.cat(logits, dim=0)
-        # decoder
+        # Decoder
         else:
-            logits = self._forward_decoder(
+            logits = self.decoder(
                 input_ids=input_ids,
                 inputs_embeds=inputs_embeds,
                 cache_position=cache_position,
@@ -531,119 +693,3 @@ class RBLNDecoderOnlyModelForCausalLM(RBLNModel):
             logits=logits,
             generate_idx=generate_idx,
         )
-
-    def _forward_prefill(
-        self,
-        input_ids: torch.LongTensor = None,
-        inputs_embeds: torch.Tensor = None,
-        cache_position: torch.Tensor = None,
-        batch_idx: int = None,
-    ) -> torch.FloatTensor:
-        if batch_idx is None or batch_idx >= self.batch_size:
-            raise RuntimeError(
-                f"Invalid batch_idx ({batch_idx}). It must be a non-null value less than the batch size ({self.batch_size})."
-            )
-
-        out_buffers = [
-            torch.empty(
-                size=[
-                    1,
-                    1,
-                    self.config.vocab_size,
-                ],
-                dtype=torch.float32,
-                device="cpu",
-            )
-        ]
-
-        input_tensors = inputs_embeds if inputs_embeds is not None else input_ids
-        query_length = input_tensors.shape[1]
-        if query_length > self.max_seq_len:
-            raise ValueError(
-                f"Input length ({query_length}) exceeds the maximum allowed sequence length ({self.max_seq_len})."
-            )
-
-        _attention_mask = self.prefill_attention_mask.clone()
-
-        for step in range(0, query_length, self.prefill_chunk_size):
-            # pad input_tensors & cache_position for prefill_chunk
-            if (step + self.prefill_chunk_size) > query_length:
-                pad_to_chunk = step + self.prefill_chunk_size - query_length
-                if inputs_embeds is not None:
-                    input_tensors = torch.nn.functional.pad(input_tensors, (0, 0, 0, pad_to_chunk))
-                else:
-                    input_tensors = torch.nn.functional.pad(input_tensors, (0, pad_to_chunk))
-
-                cache_position = torch.cat(
-                    [
-                        cache_position,
-                        torch.arange(
-                            query_length,
-                            step + self.prefill_chunk_size,
-                            dtype=torch.int32,
-                        ).unsqueeze(0),
-                    ],
-                    dim=-1,
-                )
-
-            # slice input_tensor & cache_position with prefill_chunk_size
-            _input_tensors = input_tensors[:, step : step + self.prefill_chunk_size]
-            _cache_position = cache_position[:, step : step + self.prefill_chunk_size]
-
-            # update attention_mask
-            if step >= self.prefill_chunk_size:
-                _attention_mask[:, :, :, step - self.prefill_chunk_size : step] = 1
-            _attention_mask[:, :, :, step : step + self.prefill_chunk_size] = self.causal_mask
-
-            query_position = (query_length - 1) % self.prefill_chunk_size
-
-            logits = self.prefill_decoder(
-                input_ids=_input_tensors.contiguous() if inputs_embeds is None else None,
-                inputs_embeds=_input_tensors.contiguous() if inputs_embeds is not None else None,
-                attention_mask=_attention_mask.contiguous(),
-                cache_position=_cache_position.contiguous(),
-                batch_position=torch.tensor(batch_idx, dtype=torch.int16),
-                query_position=torch.tensor(query_position, dtype=torch.int16),
-                out=out_buffers,
-            )
-
-        # update decoder_attn_mask with preprocessed kv-cache length in prefill phase
-        self.dec_attn_mask[batch_idx] = self.dec_attn_mask_init.clone()
-        self.dec_attn_mask[batch_idx, :, :, :query_length] = 1
-
-        return logits
-
-    def _forward_decoder(
-        self,
-        input_ids: torch.LongTensor = None,
-        inputs_embeds: torch.Tensor = None,
-        cache_position: torch.Tensor = None,
-    ) -> torch.FloatTensor:
-        input_tensors = inputs_embeds if inputs_embeds is not None else input_ids
-        if input_tensors is None:
-            raise ValueError("Either `input_ids` or `inputs_embeds` must be provided.")
-
-        batch_size = input_tensors.shape[0]
-        if batch_size != self.batch_size:
-            raise RuntimeError(
-                f"Batch size mismatch: got {batch_size}, expected {self.batch_size} (compiled batch size)."
-            )
-
-        if batch_size != cache_position.shape[0]:
-            raise RuntimeError(f"Cache position size mismatch: got {cache_position.shape[0]}, expected {batch_size}.")
-
-        for b_idx in range(batch_size):
-            decoding_step = cache_position[b_idx].item()
-            if not (0 <= decoding_step < self.dec_attn_mask.shape[-1]):
-                raise ValueError(
-                    f"Decoding step {decoding_step} out of bounds for attention mask with shape {self.dec_attn_mask.shape}."
-                )
-            self.dec_attn_mask[b_idx, :, :, decoding_step] = 1
-        logits = self.decoder(
-            input_ids=input_tensors.contiguous() if inputs_embeds is None else None,
-            inputs_embeds=input_tensors.contiguous() if inputs_embeds is not None else None,
-            attention_mask=self.dec_attn_mask.contiguous(),
-            cache_position=cache_position.contiguous(),
-        )
-
-        return logits

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
 
@@ -26,14 +25,14 @@ from transformers import (
     PreTrainedModel,
 )
 from transformers.modeling_outputs import BaseModelOutputWithPooling
-from transformers.models.llava_next.modeling_llava_next import LlavaNextCausalLMOutputWithPast
 
 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 (
@@ -337,7 +336,7 @@ class RBLNLlavaNextForConditionalGeneration(RBLNModel):
         generate_idx: Optional[torch.Tensor] = None,
         batch_idx: Optional[int] = None,
         **kwargs,
-    ) -> Union[Tuple, LlavaNextCausalLMOutputWithPast]:
+    ) -> Union[Tuple, RBLNDecoderOnlyOutput]:
         vision_feature_layer = (
             vision_feature_layer if vision_feature_layer is not None else self.config.vision_feature_layer
         )
@@ -378,7 +377,7 @@ class RBLNLlavaNextForConditionalGeneration(RBLNModel):
             inputs_embeds = [inputs_embeds[i : i + 1, attention_mask[i].bool()] for i in range(batch_size)]
             for batch_idx in range(batch_size):
                 generate_idx[batch_idx] = inputs_embeds[batch_idx].shape[-2]
-                logit = self.language_model._forward_prefill(
+                logit = self.language_model.prefill_decoder(
                     inputs_embeds=inputs_embeds[batch_idx],
                     batch_idx=batch_idx,
                     cache_position=torch.arange(
@@ -390,15 +389,13 @@ class RBLNLlavaNextForConditionalGeneration(RBLNModel):
 
                 logits.append(logit)
             logits = torch.cat(logits, dim=0)
-            outputs = RBLNDecoderOnlyOutput(logits=logits, generate_idx=generate_idx)
         else:
-            outputs: RBLNDecoderOnlyOutput = self.language_model(
+            logits = self.language_model.decoder(
                 inputs_embeds=inputs_embeds,
                 cache_position=cache_position,
-                generate_idx=generate_idx,
             )
 
-        return outputs
+        return RBLNDecoderOnlyOutput(logits=logits, generate_idx=generate_idx)
 
     # Almost copied from : https://github.com/huggingface/transformers/blob/6b550462139655d488d4c663086a63e98713c6b9/src/transformers/models/llava_next/modeling_llava_next.py
     def pack_image_features(self, image_features, image_sizes, vision_feature_select_strategy, image_newline=None):

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

@@ -13,30 +13,25 @@
 # limitations under the License.
 
 import inspect
-import logging
 from abc import ABC
 from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple, Union
 
 import rebel
 import torch
 from rebel.compile_context import CompileContext
-from transformers import AutoModelForSeq2SeqLM, GenerationConfig, PretrainedConfig, PreTrainedModel
+from transformers import AutoModelForSeq2SeqLM, PretrainedConfig, PreTrainedModel
 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 (
-        AutoFeatureExtractor,
-        AutoProcessor,
-        AutoTokenizer,
-        PretrainedConfig,
-    )
+    from transformers import AutoFeatureExtractor, AutoProcessor, AutoTokenizer, GenerationConfig, PretrainedConfig
 
 
 class RBLNRuntimeEncoder(RBLNPytorchRuntime):
@@ -50,9 +45,50 @@ class RBLNRuntimeEncoder(RBLNPytorchRuntime):
 class RBLNRuntimeDecoder(RBLNPytorchRuntime):
     mandatory_members = ["main_input_name"]
 
-    def forward(self, *args: List[torch.Tensor], **kwargs: Dict[str, torch.Tensor]):
-        outputs = super().forward(*args, **kwargs)
-        return Seq2SeqLMOutput(logits=outputs)
+    def __init__(
+        self,
+        runtime: rebel.Runtime,
+        batch_size: int,
+        dec_max_seq_len: int,
+        **kwargs: Any,
+    ) -> None:
+        super().__init__(runtime, **kwargs)
+        self.batch_size = batch_size
+        self.dec_max_seq_len = dec_max_seq_len
+
+    def forward(
+        self,
+        decoder_input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        decoder_attention_mask: Optional[torch.BoolTensor] = None,
+        cache_position: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> Tuple[torch.FloatTensor]:
+        batch_size = decoder_input_ids.shape[0]
+        if batch_size != self.batch_size:
+            raise RuntimeError(
+                f"Batch size mismatch: got {batch_size}, expected {self.batch_size} (compiled batch size)."
+            )
+
+        if batch_size != cache_position.shape[0]:
+            raise RuntimeError(f"Cache position size mismatch: got {cache_position.shape[0]}, expected {batch_size}.")
+
+        for b_idx in range(self.batch_size):
+            decoding_step = cache_position[b_idx].item()
+            if not (0 <= decoding_step < self.dec_max_seq_len):
+                raise ValueError(
+                    f"Decoding step {decoding_step} out of bounds for attention mask with shape {self.dec_attn_mask.shape}."
+                )
+            decoder_attention_mask[b_idx, : decoding_step + 1] = 1
+
+        lm_logits = super().forward(
+            decoder_input_ids,
+            decoder_attention_mask,
+            attention_mask,
+            cache_position,
+        )
+
+        return Seq2SeqLMOutput(logits=lm_logits)
 
 
 class RBLNModelForSeq2SeqLM(RBLNModel, ABC):
@@ -72,8 +108,15 @@ class RBLNModelForSeq2SeqLM(RBLNModel, ABC):
     auto_model_class = AutoModelForSeq2SeqLM
 
     def __post_init__(self, **kwargs):
-        self.encoder = RBLNRuntimeEncoder(runtime=self.model[0], main_input_name="input_ids")
-        self.decoder = RBLNRuntimeDecoder(runtime=self.model[1], main_input_name="input_ids")
+        batch_size = self.rbln_config.model_cfg["batch_size"]
+        dec_max_seq_len = self.rbln_config.model_cfg["dec_max_seq_len"]
+        self.encoder = RBLNRuntimeEncoder(
+            runtime=self.model[0],
+            main_input_name="input_ids",
+        )
+        self.decoder = RBLNRuntimeDecoder(
+            runtime=self.model[1], main_input_name="input_ids", batch_size=batch_size, dec_max_seq_len=dec_max_seq_len
+        )
 
     @classmethod
     @torch.inference_mode()
@@ -304,46 +347,24 @@ class RBLNModelForSeq2SeqLM(RBLNModel, ABC):
 
     def forward(
         self,
-        input_ids: torch.LongTensor = None,
+        decoder_input_ids: torch.LongTensor = None,
         cache_position: Union[List[torch.Tensor], torch.Tensor] = None,
         **kwargs,
     ) -> Tuple[torch.FloatTensor]:
         # common decoder
         cache_position = torch.full((self.rbln_config.model_cfg["batch_size"], 1), cache_position, dtype=torch.int32)
-        logits = self._forward_decoder(input_ids=input_ids, cache_position=cache_position, **kwargs).logits
+        logits = self.decoder(decoder_input_ids=decoder_input_ids, cache_position=cache_position, **kwargs).logits
 
         return Seq2SeqLMOutput(
             logits=logits,
         )
 
-    def _forward_decoder(
-        self,
-        attention_mask: Optional[torch.FloatTensor] = None,
-        decoder_input_ids: Optional[torch.LongTensor] = None,
-        decoder_attention_mask: Optional[torch.BoolTensor] = None,
-        cache_position: Optional[torch.Tensor] = None,
-        **kwargs,
-    ) -> Tuple[torch.FloatTensor]:
-        dec_attention_mask = decoder_attention_mask.clone()
-        for b_idx in range(self.rbln_config.model_cfg["batch_size"]):
-            dec_attention_mask[b_idx, : cache_position[b_idx] + 1] = 1
-
-        decoder_output = self.decoder(
-            input_ids=decoder_input_ids,
-            attention_mask=dec_attention_mask,
-            encoder_attention_mask=attention_mask,
-            cache_position=cache_position,
-        )
-        lm_logits = decoder_output.logits
-
-        return Seq2SeqLMOutput(logits=lm_logits)
-
     def _prepare_encoder_decoder_kwargs_for_generation(
         self,
         inputs_tensor: torch.Tensor,
         model_kwargs,
         model_input_name: Optional[str] = None,
-        generation_config: Optional[GenerationConfig] = None,
+        generation_config: Optional["GenerationConfig"] = None,
     ) -> Dict[str, Any]:
         # 1. get encoder
         encoder = self.get_encoder()
@@ -373,6 +394,7 @@ class RBLNModelForSeq2SeqLM(RBLNModel, ABC):
         )
 
         # 3. make sure that encoder returns `ModelOutput`
+        model_input_name = model_input_name if model_input_name is not None else self.main_input_name
         encoder_kwargs["return_dict"] = True
         encoder_kwargs["output_hidden_states"] = False
         encoder_kwargs["output_attentions"] = False

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-0.2.1a2.dist-info → optimum_rbln-0.2.1a4.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: optimum-rbln
-Version: 0.2.1a2
+Version: 0.2.1a4
 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-0.2.1a2.dist-info → optimum_rbln-0.2.1a4.dist-info}/RECORD

@@ -1,23 +1,23 @@
 optimum/rbln/__init__.py,sha256=sLCjJu_MLZEKDOwHIlJP4u4GzGZx-1kqHTYGw5B4xDg,6096
-optimum/rbln/__version__.py,sha256=RUdGjwBqCynJjA53NyAfXHitjfqAoWaGNCXiO_9cNsw,413
-optimum/rbln/modeling.py,sha256=OQGLkzlE3vD3O-ZeE1Z0jK-QCqWy1V46pSCOdmehFTI,8267
-optimum/rbln/modeling_base.py,sha256=sU5Tr3SmhQZPsbKz5xo-FqMU1gC4Xd4m9xZVIx2NY7I,20359
+optimum/rbln/__version__.py,sha256=gOdVXc7MBdh3lFQDUicnX79h1S87OceEEbF0RhUUT8g,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
@@ -49,14 +49,14 @@ optimum/rbln/transformers/models/bart/__init__.py,sha256=32HPe0_GIO0hp9U464Iv6Jd
 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=YVV7k_laU6yJBawZrgjIWjRmIF-Y4oQQHqyf8lsraQs,691
-optimum/rbln/transformers/models/bert/modeling_bert.py,sha256=flzCLWqCaCnQLfWCVSRmQke_CEAXzcG0DOoUM8EAqkg,4649
+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=BjQHwoPZfM-KUQzxm4AU-PdmoMgLxnCG6kfSpGjUvrk,36578
-optimum/rbln/transformers/models/decoderonly/modeling_decoderonly.py,sha256=mAgRRMGVHvTUjJBDlmUOjNhSNjprKSD7tLeFknrx0Rw,25810
+optimum/rbln/transformers/models/decoderonly/modeling_decoderonly.py,sha256=2OO8MEgFgcl1VPrQXxqkvmRJJEuFdexwu8XqbHDbR6Y,27609
 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=w_plsUOzxnhkQBhQeUqW9aJqGCvCvLtsx0XNKYjOprU,26203
 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/modeling_seq2seq.py,sha256=HG_-8ufRWIls67imU1547V0bk9FUWC0haOBL7eyRV6k,16365
 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=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-0.2.1a2.dist-info/METADATA,sha256=hwr_UclFUgkNbtwCanMy_5lF56FPX8WG0GGz3B0cJCc,5300
-optimum_rbln-0.2.1a2.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-optimum_rbln-0.2.1a2.dist-info/licenses/LICENSE,sha256=QwcOLU5TJoTeUhuIXzhdCEEDDvorGiC6-3YTOl4TecE,11356
-optimum_rbln-0.2.1a2.dist-info/RECORD,,
+optimum_rbln-0.2.1a4.dist-info/METADATA,sha256=r7LWvGMq_IPBPFfe8MXAMmI7zxvArSL-rE-y4C73alo,5300
+optimum_rbln-0.2.1a4.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+optimum_rbln-0.2.1a4.dist-info/licenses/LICENSE,sha256=QwcOLU5TJoTeUhuIXzhdCEEDDvorGiC6-3YTOl4TecE,11356
+optimum_rbln-0.2.1a4.dist-info/RECORD,,