optimum-rbln 0.8.3rc0__py3-none-any.whl → 0.8.4a0__py3-none-any.whl

optimum/rbln/__version__.py

@@ -28,7 +28,7 @@ version_tuple: VERSION_TUPLE
 commit_id: COMMIT_ID
 __commit_id__: COMMIT_ID
 
-__version__ = version = '0.8.3rc0'
-__version_tuple__ = version_tuple = (0, 8, 3, 'rc0')
+__version__ = version = '0.8.4a0'
+__version_tuple__ = version_tuple = (0, 8, 4, 'a0')
 
 __commit_id__ = commit_id = None

optimum/rbln/configuration_utils.py

@@ -476,6 +476,7 @@ class RBLNModelConfig(RBLNSerializableConfigProtocol):
     non_save_attributes = [
         "_frozen",
         "_runtime_options",
+        "torch_dtype",
         "npu",
         "tensor_parallel_size",
         "create_runtimes",
@@ -566,6 +567,7 @@ class RBLNModelConfig(RBLNSerializableConfigProtocol):
         tensor_parallel_size: Optional[int] = None,
         timeout: Optional[int] = None,
         optimum_rbln_version: Optional[str] = None,
+        _torch_dtype: Optional[str] = None,
         _compile_cfgs: List[RBLNCompileConfig] = [],
         **kwargs: Any,
     ):
@@ -583,6 +585,7 @@ class RBLNModelConfig(RBLNSerializableConfigProtocol):
             tensor_parallel_size (Optional[int]): Size for tensor parallelism to distribute the model across devices.
             timeout (Optional[int]): The timeout for the runtime in seconds. If it isn't provided, it will be set to 60 by default.
             optimum_rbln_version (Optional[str]): The optimum-rbln version used for this configuration.
+            _torch_dtype (Optional[str]): The data type to use for the model.
             _compile_cfgs (List[RBLNCompileConfig]): List of compilation configurations for the model.
             **kwargs: Additional keyword arguments.
 
@@ -610,6 +613,7 @@ class RBLNModelConfig(RBLNSerializableConfigProtocol):
         self.npu = npu
         self.tensor_parallel_size = tensor_parallel_size
 
+        self._torch_dtype = _torch_dtype or "float32"
         self.optimum_rbln_version = optimum_rbln_version
         if self.optimum_rbln_version is None:
             self.optimum_rbln_version = __version__
@@ -639,6 +643,17 @@ class RBLNModelConfig(RBLNSerializableConfigProtocol):
 
             raise ValueError(f"Unexpected arguments: {kwargs.keys()}")
 
+    @property
+    def torch_dtype(self):
+        return getattr(torch, self._torch_dtype)
+
+    @torch_dtype.setter
+    def torch_dtype(self, torch_dtype: Union[str, torch.dtype]):
+        if isinstance(torch_dtype, torch.dtype):
+            torch_dtype = RBLNCompileConfig.normalize_dtype(torch_dtype)
+
+        self._torch_dtype = torch_dtype
+
     @property
     def rbln_model_cls_name(self) -> str:
         return self.__class__.__name__[:-6]

optimum/rbln/modeling.py

@@ -19,7 +19,7 @@ from typing import TYPE_CHECKING, Any, Dict, List, Optional, Union, get_args, ge
 import rebel
 import torch
 from huggingface_hub.constants import HUGGINGFACE_HUB_CACHE
-from transformers import AutoConfig, PretrainedConfig
+from transformers import PretrainedConfig
 from transformers.modeling_outputs import BaseModelOutput
 
 from .configuration_utils import DEFAULT_COMPILED_MODEL_NAME, RBLNModelConfig
@@ -119,9 +119,6 @@ class RBLNModel(RBLNBaseModel):
         # Save configs
         if config is None:
             config = model.config
-            # remote_config
-            if hasattr(config, "auto_map") and "AutoConfig" in config.auto_map:
-                config = AutoConfig.from_pretrained(config._name_or_path, **kwargs)
 
         if hasattr(model, "can_generate") and model.can_generate():
             import json

optimum/rbln/modeling_base.py

@@ -34,7 +34,7 @@ from .utils.submodule import SubModulesMixin
 
 
 if TYPE_CHECKING:
-    from transformers import PreTrainedModel
+    from transformers import AutoFeatureExtractor, AutoProcessor, AutoTokenizer, PreTrainedModel
 
 logger = get_logger(__name__)
 
@@ -53,6 +53,7 @@ class RBLNBaseModel(SubModulesMixin, PushToHubMixin, PreTrainedModel):
     config_class = AutoConfig
     config_name = "config.json"
     hf_library_name = "transformers"
+    _supports_non_fp32 = False
 
     def __init__(
         self,
@@ -91,7 +92,7 @@ class RBLNBaseModel(SubModulesMixin, PushToHubMixin, PreTrainedModel):
 
         self.device = torch.device("cpu")
         self.training = False
-        self.dtype = torch.float32
+        self.dtype = rbln_config.torch_dtype
 
         # FIXME :: model_save_dir is not used after initialized. (This can be used when save/load)
         # This attribute is needed to keep one reference on the temporary directory, since garbage collecting it
@@ -400,8 +401,21 @@ class RBLNBaseModel(SubModulesMixin, PushToHubMixin, PreTrainedModel):
         return compiled_model
 
     @classmethod
-    def update_rbln_config(cls, **others) -> RBLNModelConfig:
-        rbln_config = cls._update_rbln_config(**others)
+    def update_rbln_config(
+        cls,
+        preprocessors: Optional[Union["AutoFeatureExtractor", "AutoProcessor", "AutoTokenizer"]],
+        model: "PreTrainedModel",
+        model_config: "PretrainedConfig",
+        rbln_config: RBLNModelConfig,
+    ) -> RBLNModelConfig:
+        rbln_config.torch_dtype = model.dtype
+        if not cls._supports_non_fp32 and rbln_config.torch_dtype != torch.float32:
+            raise NotImplementedError(
+                f"Currently, {cls.__name__} does not support non-fp32 dtype. Please use float32 dtype."
+            )
+        rbln_config = cls._update_rbln_config(
+            preprocessors=preprocessors, model=model, model_config=model_config, rbln_config=rbln_config
+        )
         rbln_config.freeze()
         if rbln_config.rbln_model_cls_name != cls.__name__:
             raise NameError(
@@ -444,12 +458,12 @@ class RBLNBaseModel(SubModulesMixin, PushToHubMixin, PreTrainedModel):
 
         # This method mimics the interface of torch.nn.Module.parameters()
         # specifically for code that uses `next(model.parameters())` to infer
-        # the device or dtype. It yields a single dummy tensor on CPU with float32 dtype.
+        # the device or dtype. It yields a single dummy tensor on CPU with model dtype.
 
         # Warning:
         #     This does NOT yield the actual model parameters used by the RBLN runtime.
         #     Code relying on iterating through all model parameters will not work as expected.
-        yield torch.tensor([1.0], dtype=torch.float32, device=torch.device("cpu"))
+        yield torch.tensor([1.0], dtype=self.dtype, device=torch.device("cpu"))
 
     def __call__(self, *args, **kwargs):
         return self.forward(*args, **kwargs)

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

@@ -1066,7 +1066,7 @@ class RotaryEmbedding(nn.Module):
             rope_type = "default"
 
         inv_freq, attention_scaling = ROPE_INIT_FUNCTIONS[rope_type](config, max_seq_len_cached)
-        cache_position = torch.arange(0, max_seq_len_cached, dtype=torch.float32)
+        cache_position = torch.arange(0, max_seq_len_cached)
         cache_position_expanded = cache_position[:, None]
 
         if rope_type == "dynamic":
@@ -1085,8 +1085,8 @@ class RotaryEmbedding(nn.Module):
 
     def forward(self, x, seq_len):
         return (
-            self._cos_cached[:seq_len].to(dtype=x.dtype),
-            self._sin_cached[:seq_len].to(dtype=x.dtype),
+            self._cos_cached[:seq_len].to(dtype=torch.float32),
+            self._sin_cached[:seq_len].to(dtype=torch.float32),
         )
 
 
@@ -1116,8 +1116,11 @@ def rotate_half(x):
 
 def apply_rotary_pos_emb(q, k, cos, sin):
     """Applies Rotary Position Embedding to the query and key tensors."""
+    dtype = q.dtype
     q_embed = (q * cos) + (rotate_half(q) * sin)
     k_embed = (k * cos) + (rotate_half(k) * sin)
+    q_embed = q_embed.to(dtype)
+    k_embed = k_embed.to(dtype)
     return q_embed, k_embed
 
 

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

@@ -317,7 +317,13 @@ class RBLNRuntimeModel(RBLNPytorchRuntime):
 
         # Initialize attention mask for chunked processing
         chunked_attention_mask = (
-            torch.zeros(1, 1, self.rbln_config.prefill_chunk_size, self.rbln_config.max_seq_len, dtype=torch.float32)
+            torch.zeros(
+                1,
+                1,
+                self.rbln_config.prefill_chunk_size,
+                self.rbln_config.max_seq_len,
+                dtype=self.rbln_config.torch_dtype,
+            )
             if self.rbln_config.use_attention_mask
             else None
         )

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

@@ -19,7 +19,7 @@ from typing import TYPE_CHECKING, Any, Callable, List, Optional, Tuple, Union
 import rebel
 import torch
 from rebel.compile_context import CompileContext
-from transformers import AutoConfig, AutoModel, AutoModelForCausalLM, PretrainedConfig, PreTrainedModel
+from transformers import AutoModel, AutoModelForCausalLM, PretrainedConfig, PreTrainedModel
 from transformers.modeling_outputs import BaseModelOutputWithPast
 from transformers.modeling_utils import no_init_weights
 
@@ -33,7 +33,7 @@ from ...modeling_attention_utils import (
     validate_sliding_window,
 )
 from ...modeling_outputs import RBLNDecoderOnlyOutput
-from ...utils.rbln_quantization import prepare_model_for_quantization
+from ...utils.rbln_quantization import get_quantized_model
 from .configuration_decoderonly import RBLNDecoderOnlyModelConfig, RBLNDecoderOnlyModelForCausalLMConfig
 from .decoderonly_architecture import DecoderOnlyWrapper
 from .decoderonly_runtime_utils import RBLNPageTableManager, RBLNRuntimeModel
@@ -72,6 +72,7 @@ class RBLNDecoderOnlyModel(RBLNModel, RBLNDecoderOnlyFlashAttentionMixin):
     auto_model_class = AutoModel
     _decoder_wrapper_cls = DecoderOnlyWrapper
     _use_rotary_emb = True
+    _supports_non_fp32 = True
 
     def __post_init__(self, **kwargs):
         if self.rbln_config.use_inputs_embeds:
@@ -86,10 +87,8 @@ class RBLNDecoderOnlyModel(RBLNModel, RBLNDecoderOnlyFlashAttentionMixin):
     def setup_runtime(self):
         # Initialize resources to be used across Runtime instances (prefill and decode phases)
         page_table_manager = RBLNPageTableManager(self.rbln_config)
-        dec_attn_mask = torch.zeros(
-            self.rbln_config.batch_size, 1, 1, self.rbln_config.max_seq_len, dtype=torch.float32
-        )
-        out_buffers = [torch.empty(self.prefill_output_size, dtype=torch.float32, device="cpu")]
+        dec_attn_mask = torch.zeros(self.rbln_config.batch_size, 1, 1, self.rbln_config.max_seq_len, dtype=self.dtype)
+        out_buffers = [torch.empty(self.prefill_output_size, dtype=self.dtype)]
 
         common_kwargs = {
             "main_input_name": "inputs_embeds" if self.rbln_config.use_inputs_embeds else "input_ids",
@@ -143,35 +142,17 @@ class RBLNDecoderOnlyModel(RBLNModel, RBLNDecoderOnlyFlashAttentionMixin):
     ):
         kwargs = cls.update_kwargs(kwargs)
 
-        if config is None:
-            config = AutoConfig.from_pretrained(
-                model_id,
-                use_auth_token=use_auth_token,
-                revision=revision,
-                force_download=force_download,
-                cache_dir=cache_dir,
-                trust_remote_code=trust_remote_code,
-                **kwargs,
-            )
-            if config.torch_dtype == torch.bfloat16:
-                # FIXME: bfloat16 is not supported by rebel-compiler
-                config.torch_dtype = torch.float32
-
-        with no_init_weights():
-            model = cls.auto_model_class.from_config(config)
-
-        model = prepare_model_for_quantization(
-            model,
+        return get_quantized_model(
+            cls.auto_model_class,
             model_id,
-            kwargs.get("num_hidden_layers"),
             use_auth_token=use_auth_token,
             revision=revision,
             cache_dir=cache_dir,
             force_download=force_download,
             local_files_only=local_files_only,
             rbln_quantization=rbln_config.quantization,
+            **kwargs,
         )
-        return model
 
     def __getattr__(self, __name: str) -> Any:
         # Special method to delegate attribute access to the original Huggingface LM class.
@@ -365,7 +346,7 @@ class RBLNDecoderOnlyModel(RBLNModel, RBLNDecoderOnlyFlashAttentionMixin):
 
         input_info = []
         if rbln_config.use_inputs_embeds:
-            input_info.append(("inputs_embeds", [batch_size, query_length, hidden_size], "float32"))
+            input_info.append(("inputs_embeds", [batch_size, query_length, hidden_size], rbln_config.torch_dtype))
         else:
             input_info.append(("input_ids", [batch_size, query_length], "int64"))
 
@@ -384,16 +365,16 @@ class RBLNDecoderOnlyModel(RBLNModel, RBLNDecoderOnlyFlashAttentionMixin):
 
         if rbln_config.use_attention_mask:
             if rbln_config.use_position_ids:
-                input_info.append(("attention_mask", [batch_size, rbln_config.max_seq_len], "float32"))
+                input_info.append(("attention_mask", [batch_size, rbln_config.max_seq_len], rbln_config.torch_dtype))
             else:
                 input_info.append(
-                    ("attention_mask", [batch_size, 1, query_length, rbln_config.max_seq_len], "float32")
+                    ("attention_mask", [batch_size, 1, query_length, rbln_config.max_seq_len], rbln_config.torch_dtype)
                 )
 
         if rbln_config.use_position_ids:
             input_info.append(("position_ids", [batch_size, query_length], "int32"))
 
-        kvcache_dtype = "float32"
+        kvcache_dtype = rbln_config.torch_dtype
         if rbln_config.quantization and rbln_config.quantization.kv_caches == "fp8":
             kvcache_dtype = "float8_e4m3fn"
 

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

@@ -345,6 +345,7 @@ class RBLNGemma3ForCausalLM(RBLNDecoderOnlyModelForCausalLM):
     """
 
     _decoder_wrapper_cls = Gemma3ForCausalLMWrapper
+    _supports_non_fp32 = False
 
     def setup_runtime(self):
         # Initialize shared resources to be used across Runtime instances (prefill and decode phases)

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

@@ -372,6 +372,8 @@ class RBLNQwen2_5_VLForConditionalGeneration(RBLNDecoderOnlyModelForCausalLM):
         ```
     """
 
+    _supports_non_fp32 = False
+
     auto_model_class = AutoModelForVision2Seq
     _rbln_submodules = [
         {"name": "visual"},

optimum/rbln/transformers/utils/rbln_quantization.py

@@ -14,18 +14,23 @@
 
 import glob
 import os
-from typing import Any, Dict, Iterable, List, Optional, Tuple, Union
+from typing import TYPE_CHECKING, Any, Dict, Iterable, List, Optional, Tuple, Type, Union
 
 import torch
 from huggingface_hub import hf_hub_download, list_repo_files
 from safetensors.torch import load_file
 from torch.nn import Linear, Parameter
 from torch.nn import functional as F
+from transformers import AutoConfig
+from transformers.modeling_utils import get_state_dict_dtype, no_init_weights
 
 from ...configuration_utils import RBLNSerializableConfigProtocol
 from ...utils.logging import get_logger
 
 
+if TYPE_CHECKING:
+    from transformers.models.auto.modeling_auto import _BaseAutoModelClass
+
 logger = get_logger()
 
 
@@ -138,22 +143,31 @@ class QuantizedLayerFactory:
         return create_fp8linear(layer, self.quantization_config)
 
 
-def prepare_model_for_quantization(
-    model: torch.nn.Module,
+def get_quantized_model(
+    hf_auto_model_class: Type["_BaseAutoModelClass"],
     model_id: str,
-    n_layer: Optional[int] = None,
     use_auth_token: Optional[Union[bool, str]] = None,
     revision: Optional[str] = None,
     cache_dir: Optional[str] = None,
     force_download: bool = False,
     local_files_only: bool = False,
     rbln_quantization: Optional[RBLNQuantizationConfig] = None,
-) -> torch.nn.Module:
+    **kwargs,
+):
     """
-    Prepare the model for quantization by updating specified linear layers to quantized (qlinear) layers.
+    Get a quantized model from a model class and model id.
     """
+    # torch_dtype should not be passed to AutoConfig.from_pretrained
+    # since it doesn't support 'auto'
+    torch_dtype = kwargs.pop("torch_dtype", None)
+    if torch_dtype is not None:
+        logger.warning(
+            "torch_dtype is not supported for quantized models. "
+            "It will be ignored and the dtype of the model will be determined by the weights."
+        )
+        torch_dtype = None
 
-    # 1. Load weight files
+    # get paths of safetensors files in the model repo
     safetensor_files = load_weight_files(
         model_id,
         use_auth_token=use_auth_token,
@@ -163,17 +177,31 @@ def prepare_model_for_quantization(
         local_files_only=local_files_only,
     )
 
-    # 2. Update linear layers based on the quantization config
-    update_layers_to_quantize(model, rbln_quantization)
+    # load safetensors files into memory
+    safetensors = [load_file(safetensor_file) for safetensor_file in safetensor_files]
+
+    # get the dtype of the model from the first safetensor file
+    torch_dtype = get_state_dict_dtype(safetensors[0])
 
-    # 3. Load weights into model parameters
-    load_weights_from_files(
-        model,
-        safetensor_files,
-        n_layer,
-        rbln_quantization=rbln_quantization,
+    config = AutoConfig.from_pretrained(
+        model_id,
+        use_auth_token=use_auth_token,
+        revision=revision,
+        cache_dir=cache_dir,
+        force_download=force_download,
+        local_files_only=local_files_only,
+        **kwargs,
     )
 
+    with no_init_weights():
+        model = hf_auto_model_class.from_config(config, torch_dtype=torch_dtype)
+
+    # Quantize the model
+    update_layers_to_quantize(model, rbln_quantization)
+
+    # Load weights into the model
+    load_weights_from_files(model, safetensors, rbln_quantization)
+
     return model
 
 
@@ -372,32 +400,26 @@ def canonicalize_checkpoint_items(
 
 def load_weights_from_files(
     model: torch.nn.Module,
-    safetensor_files: list[str],
-    n_layer: Optional[int] = None,
+    safetensors: List[Dict[str, torch.Tensor]],
     rbln_quantization: Optional[RBLNQuantizationConfig] = None,
 ):
     """
-    Load safetensor file data directly into the model from provided safetensor files,
-    filtering by layer if n_layer is provided.
+    Load safetensor file data directly into the model from provided safetensor files.
     """
 
     model_params = dict(model.named_parameters(recurse=True))
     model_buffers = dict(model.named_buffers(recurse=True))
 
-    target_layers = list(range(n_layer)) if n_layer is not None else None
-
     unloaded_keys = []
     loaded_input_scale = False
     loaded_kv_scale = False
     loaded_weight_scale = False
 
-    for safetensor_file in safetensor_files:
-        file_data = load_file(safetensor_file)
-
+    for safetensor in safetensors:
         # Normalize all (key, tensor) pairs to the internal schema
         normalized_items = canonicalize_checkpoint_items(
             model=model,
-            items=file_data.items(),
+            items=safetensor.items(),
             rbln_quantization=rbln_quantization,
         )
 
@@ -410,12 +432,6 @@ def load_weights_from_files(
             if key.endswith("k_scale") or key.endswith("v_scale"):
                 loaded_kv_scale = True
 
-            # Filter by layer index if requested
-            if target_layers is not None:
-                parts = key.split(".")
-                if len(parts) > 2 and parts[2].isdigit() and (int(parts[2]) not in target_layers):
-                    continue
-
             # Copy into parameters or buffers
             if key in model_params:
                 # Ensure dtype compatibility

{optimum_rbln-0.8.3rc0.dist-info → optimum_rbln-0.8.4a0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: optimum-rbln
-Version: 0.8.3rc0
+Version: 0.8.4a0
 Summary: Optimum RBLN is the interface between the HuggingFace Transformers and Diffusers libraries and RBLN accelerators. It provides a set of tools enabling easy model loading and inference on single and multiple rbln device settings for different downstream tasks.
 Project-URL: Homepage, https://rebellions.ai
 Project-URL: Documentation, https://docs.rbln.ai

{optimum_rbln-0.8.3rc0.dist-info → optimum_rbln-0.8.4a0.dist-info}/RECORD

@@ -1,8 +1,8 @@
 optimum/rbln/__init__.py,sha256=32ouGKDGus9k5_kD27CxP8jIQOw66zpDTfS0xs1XlfE,18298
-optimum/rbln/__version__.py,sha256=boIaJ8T6HCT9Qh8wBU3n-6ZyjtAKYgztQh0WMaN7BxM,714
-optimum/rbln/configuration_utils.py,sha256=fE3HlZblxukKSdS-4VofjuyCAiqwPMX8bqXpOiTZp4g,33926
-optimum/rbln/modeling.py,sha256=jMiJy9PGjZpXpAmRTFD5fTuj8xEbLCUmncIxGD6XWLk,14338
-optimum/rbln/modeling_base.py,sha256=txBab-zVXcjqnF2gZJBzhrp5ruA3vwt3hjls0Q2S_0w,25492
+optimum/rbln/__version__.py,sha256=YNGYpHnDhFwKFL4ZTx3BIJGtmgon0Pv2G2E10GhWRaY,712
+optimum/rbln/configuration_utils.py,sha256=KtbDM7HnFGiO0PsuvkrCE3R9NF6OJVmV_fyQcQNrmUk,34469
+optimum/rbln/modeling.py,sha256=cAIPWEw5DGzUWeqjCbocRhU6OO3jyhVGW60AmBLh1Nw,14134
+optimum/rbln/modeling_base.py,sha256=kQsBfUoDncNgR5P8_BvyzY6H_4YEXOBzN20lFmOZV_g,26190
 optimum/rbln/diffusers/__init__.py,sha256=1tgU_xWA42BmInqu9bBz_5R_E9TGhhK3mI06YlaiTLg,7232
 optimum/rbln/diffusers/modeling_diffusers.py,sha256=TAuMb7PSMjNwK7mh5ItE_CtAEgYeZKI27XkFFmxjHlQ,19902
 optimum/rbln/diffusers/configurations/__init__.py,sha256=vMRnPY4s-Uju43xP038D2EA18X_mhy2YfsZVpSU-VoA,1322
@@ -105,10 +105,10 @@ optimum/rbln/transformers/models/colpali/configuration_colpali.py,sha256=eDWPVlo
 optimum/rbln/transformers/models/colpali/modeling_colpali.py,sha256=v9rPLmNx-BQZhDFhKnr2kmARElTtKdFZCgFIU4m-HPw,15703
 optimum/rbln/transformers/models/decoderonly/__init__.py,sha256=w3VZOIBYaHXVdnuhK4y0zWAj0IAv7_5LGTJYaz9oYmI,1056
 optimum/rbln/transformers/models/decoderonly/configuration_decoderonly.py,sha256=H2i9Iefy-q5X-0BLWQ-CrxK8ZoT3p9t0lt_3r4TFSCY,15182
-optimum/rbln/transformers/models/decoderonly/decoderonly_architecture.py,sha256=cGkhGc8XlseMWqDCrt13z0Itn9b0emZ2PjHI-1TP0wI,42685
-optimum/rbln/transformers/models/decoderonly/decoderonly_runtime_utils.py,sha256=9acEQxGRzd21YkzxRchkhqxqpX7emQHZigFg60BIulc,19902
+optimum/rbln/transformers/models/decoderonly/decoderonly_architecture.py,sha256=L5LArhjN36fTdiwrUABgn3cnS7hh4SVCF4FMHBbiLZU,42760
+optimum/rbln/transformers/models/decoderonly/decoderonly_runtime_utils.py,sha256=v3mfIlQImQkYYr-rPn7rQR3GYdVUhALRttEduLI7H9c,20012
 optimum/rbln/transformers/models/decoderonly/generation_decoderonly.py,sha256=4D89IF0yQju_Dp_vLJN_dBkpe2U_LMWaUciYx57D-0M,3379
-optimum/rbln/transformers/models/decoderonly/modeling_decoderonly.py,sha256=hu2eJr0CpLHnRPSLhyBhyyC6DfosKmPu7lPjapcBCkE,33061
+optimum/rbln/transformers/models/decoderonly/modeling_decoderonly.py,sha256=dAHV9NgdpXHyTJGT0lieXOB3Pzi_NPlR4rqmRtmAWzM,32412
 optimum/rbln/transformers/models/depth_anything/__init__.py,sha256=xvPSIriMJWyNeVYoVB1Z7YqB4kkHOIkaHq7loNps-dk,756
 optimum/rbln/transformers/models/depth_anything/configuration_depth_anything.py,sha256=JujBVEUa_zZDXNPr1y-B_PhK5SgFFcY8Ib4EoGjjtmE,989
 optimum/rbln/transformers/models/depth_anything/modeling_depth_anything.py,sha256=tTmsVaW9Wb2WD3nKRLwp7swn3hbMvgwUEJwwVIfNYEc,1008
@@ -130,7 +130,7 @@ optimum/rbln/transformers/models/gemma3/__init__.py,sha256=6rugk3615SEt4lh7gduo_
 optimum/rbln/transformers/models/gemma3/configuration_gemma3.py,sha256=rKjKJhyaIM7YoiLR-q8GAZKIQNzDzcb5X7qf_FJE72M,3398
 optimum/rbln/transformers/models/gemma3/gemma3_architecture.py,sha256=fpLDAXCe5paWVsfc0tL59JkRQMRF-WNgIzOIb_QpSLU,6191
 optimum/rbln/transformers/models/gemma3/gemma3_runtime_utils.py,sha256=vYQ9sjRlkfamxZca_hVMQI0ylKeExsV02gOWaYVMjyg,9640
-optimum/rbln/transformers/models/gemma3/modeling_gemma3.py,sha256=zraPjowA8ni9Lb0NrmsiUai2XdOjgYOOpVnIU1n2jGA,24208
+optimum/rbln/transformers/models/gemma3/modeling_gemma3.py,sha256=TxbgkvW2Nv0VGdXNXnN_Beas6E_1D9NAH8f09Fo8t0E,24239
 optimum/rbln/transformers/models/gpt2/__init__.py,sha256=SsawHMStE3wYRtqkH5EvdTFkCdX0LLmp-QSKFhEBrHo,740
 optimum/rbln/transformers/models/gpt2/configuration_gpt2.py,sha256=iGdHfzG7plekZcIz-Z5U8lRE4SB8gbJJNcFQJ9l8Myg,1533
 optimum/rbln/transformers/models/gpt2/gpt2_architecture.py,sha256=MyAWReXmyuHnDpW5HI_TI7psyJZxLujZ9KT5XnNm7nA,2802
@@ -182,7 +182,7 @@ optimum/rbln/transformers/models/qwen2/modeling_qwen2.py,sha256=VOboPJF1rvvSVWkH
 optimum/rbln/transformers/models/qwen2/qwen2_architecture.py,sha256=XlNAMYAcDLohnSAhIFGKOPuCB5XLgzYs5ABWdeQSaZs,720
 optimum/rbln/transformers/models/qwen2_5_vl/__init__.py,sha256=rAW3DKQUzGL6EMwa5r1iLu94yhpiZpk6zfoD7TtYXrc,865
 optimum/rbln/transformers/models/qwen2_5_vl/configuration_qwen2_5_vl.py,sha256=1yyMFxh1SKsKR7rOjuotPvpSneN2_4a89bYfNk42370,4735
-optimum/rbln/transformers/models/qwen2_5_vl/modeling_qwen2_5_vl.py,sha256=BfpALuavpdCqe5RuHaNZNo2IDlLjE4SwsoPAlaictgc,26607
+optimum/rbln/transformers/models/qwen2_5_vl/modeling_qwen2_5_vl.py,sha256=hRvA37sPFC9xH1FqnFbtHS9rQOPwAvLYg4zl4oEyK-w,26639
 optimum/rbln/transformers/models/qwen2_5_vl/qwen2_5_vl_architecture.py,sha256=i_UUWhKoFjJ5CCpgeWicqABM23TxMEKPQ354LoZ6iUU,7445
 optimum/rbln/transformers/models/qwen3/__init__.py,sha256=tI4KwvXpD35dUUaa8aLUXpWoU9gJGcmKXeywOlH14ZE,746
 optimum/rbln/transformers/models/qwen3/configuration_qwen3.py,sha256=BFRPggnH4VlsXlOa19C6KAID-bPgQ8ooQ29dvogh5zk,2102
@@ -227,7 +227,7 @@ optimum/rbln/transformers/models/xlm_roberta/__init__.py,sha256=O3o2KzJ8Li3QhB7G
 optimum/rbln/transformers/models/xlm_roberta/configuration_xlm_roberta.py,sha256=wHRpGTXL9khYqSkKL1IgA7__6_lt9QpOz9tHumjK7fo,1260
 optimum/rbln/transformers/models/xlm_roberta/modeling_xlm_roberta.py,sha256=EZd3flRUEE38DYtdqEnG70LV7fHhkamRZV51xrVyjYI,1093
 optimum/rbln/transformers/utils/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-optimum/rbln/transformers/utils/rbln_quantization.py,sha256=ARngdvRmeVoOphUU3Md9kT6zS5HDrYdEFYljJwaAaio,21020
+optimum/rbln/transformers/utils/rbln_quantization.py,sha256=pORshQUgTInNaibUtd0HL-T8bKW5wuulZs2q0Oshppc,21659
 optimum/rbln/utils/__init__.py,sha256=ieDBT2VFTt2E0M4v_POLBpuGW9LxSydpb_DuPd6PQqc,712
 optimum/rbln/utils/decorator_utils.py,sha256=xu-TrsNi33SRC2a7DBsyoo6-pEQxWKZPZSmM9QlDe2Y,3745
 optimum/rbln/utils/depreacate_utils.py,sha256=uKxl3ENUCNaZXPnaDQvNxrH8hUIWdBWfZH6BM7ZV__4,385
@@ -238,7 +238,7 @@ optimum/rbln/utils/model_utils.py,sha256=4k5879Kh75m3x_vS4-qOGfqsOiAvc2kdNFFfvsF
 optimum/rbln/utils/runtime_utils.py,sha256=R6uXDbeJP03-FWdd4vthNe2D4aCra5n12E3WB1ifiGM,7933
 optimum/rbln/utils/save_utils.py,sha256=hG5uOtYmecSXZuGTvCXsTM-SiyZpr5q3InUGCCq_jzQ,3619
 optimum/rbln/utils/submodule.py,sha256=60NGLFvnhjP1DJg1opdb-FVQDsthcLCwWjW_1WQaasU,5280
-optimum_rbln-0.8.3rc0.dist-info/METADATA,sha256=ls15qV7a7bVTpkphb6aHteuBfil7u1xOzkUuysoRPZg,5300
-optimum_rbln-0.8.3rc0.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-optimum_rbln-0.8.3rc0.dist-info/licenses/LICENSE,sha256=QwcOLU5TJoTeUhuIXzhdCEEDDvorGiC6-3YTOl4TecE,11356
-optimum_rbln-0.8.3rc0.dist-info/RECORD,,
+optimum_rbln-0.8.4a0.dist-info/METADATA,sha256=QqrF_vPDFZO-DiTK0p328Y54qXyk1wApO86SAISpNcc,5299
+optimum_rbln-0.8.4a0.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+optimum_rbln-0.8.4a0.dist-info/licenses/LICENSE,sha256=QwcOLU5TJoTeUhuIXzhdCEEDDvorGiC6-3YTOl4TecE,11356
+optimum_rbln-0.8.4a0.dist-info/RECORD,,