optimum-rbln 0.8.3a4__py3-none-any.whl → 0.8.4a0__py3-none-any.whl

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/models/swin/modeling_swin.py

@@ -20,7 +20,7 @@ import torch.nn.functional as F
 from transformers import SwinConfig
 from transformers.models.swin.modeling_swin import BackboneOutput
 
-from ....configuration_utils import RBLNCompileConfig
+from ....configuration_utils import RBLNCompileConfig, RBLNModelConfig
 from ....modeling import RBLNModel
 from ....utils.logging import get_logger
 from .configuration_swin import RBLNSwinBackboneConfig
@@ -214,6 +214,31 @@ class RBLNSwinBackbone(RBLNModel):
         }
         return _SwinBackbone(model, **wrapper_cfg).eval()
 
+    @classmethod
+    def _update_submodule_config(
+        cls,
+        model: "PreTrainedModel",
+        rbln_config: RBLNModelConfig,
+        preprocessors: Optional[Union["AutoFeatureExtractor", "AutoProcessor", "AutoTokenizer"]],
+    ):
+        for processor in preprocessors:
+            if rbln_config.image_size is None and hasattr(processor, "image_processor"):
+                if "height" in processor.image_processor.size and "width" in processor.image_processor.size:
+                    rbln_config.image_size = (
+                        processor.image_processor.size["height"],
+                        processor.image_processor.size["width"],
+                    )
+                elif (
+                    "longest_edge" in processor.image_processor.size
+                    and "shortest_edge" in processor.image_processor.size
+                ):
+                    rbln_config.image_size = processor.image_processor.size["longest_edge"]
+                elif "shortest_edge" in processor.image_processor.size:
+                    rbln_config.image_size = processor.image_processor.size["shortest_edge"]
+                break
+
+        return rbln_config
+
     @classmethod
     def _update_rbln_config(
         cls,
@@ -235,8 +260,8 @@ class RBLNSwinBackbone(RBLNModel):
                 [
                     rbln_config.batch_size,
                     3,
-                    rbln_config.image_size[0],
-                    rbln_config.image_size[1],
+                    rbln_config.image_height,
+                    rbln_config.image_width,
                 ],
                 "float32",
             ),
@@ -276,14 +301,14 @@ class RBLNSwinBackbone(RBLNModel):
             )
 
         _, _, original_h, original_w = pixel_values.shape
-        if original_h > self.rbln_config.image_size[0] or original_w > self.rbln_config.image_size[1]:
+        if original_h > self.rbln_config.image_height or original_w > self.rbln_config.image_width:
             raise ValueError(
                 f"Input image size ({original_h}x{original_w}) exceeds the configured maximum size"
-                f" ({self.rbln_config.image_size[0]}x{self.rbln_config.image_size[1]})."
+                f" ({self.rbln_config.image_height}x{self.rbln_config.image_width})."
             )
 
-        pad_h = self.rbln_config.image_size[0] - original_h
-        pad_w = self.rbln_config.image_size[1] - original_w
+        pad_h = self.rbln_config.image_height - original_h
+        pad_w = self.rbln_config.image_width - original_w
         padded_pixel_values = F.pad(pixel_values, (0, pad_w, 0, pad_h))
 
         output = self.model[0](padded_pixel_values)

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/utils/submodule.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 from pathlib import Path
-from typing import TYPE_CHECKING, Any, Dict, List, Type
+from typing import TYPE_CHECKING, Any, Dict, List, Optional, Type, Union
 
 from transformers import PretrainedConfig
 
@@ -22,7 +22,7 @@ from ..utils.model_utils import get_rbln_model_cls
 
 
 if TYPE_CHECKING:
-    from transformers import PreTrainedModel
+    from transformers import AutoFeatureExtractor, AutoProcessor, AutoTokenizer, PreTrainedModel
 
     from ..modeling import RBLNModel
 
@@ -42,7 +42,12 @@ class SubModulesMixin:
             setattr(self, submodule_meta["name"], submodule)
 
     @classmethod
-    def _update_submodule_config(cls, model: "PreTrainedModel", rbln_config: RBLNModelConfig):
+    def _update_submodule_config(
+        cls,
+        model: "PreTrainedModel",
+        rbln_config: RBLNModelConfig,
+        preprocessors: Optional[Union["AutoFeatureExtractor", "AutoProcessor", "AutoTokenizer"]],
+    ):
         return rbln_config
 
     @classmethod
@@ -51,6 +56,7 @@ class SubModulesMixin:
     ) -> List["RBLNModel"]:
         rbln_submodules = []
         submodule_prefix = getattr(cls, "_rbln_submodule_prefix", None)
+        preprocessors = kwargs.pop("preprocessors", [])
 
         for submodule in cls._rbln_submodules:
             submodule_name = submodule["name"]
@@ -69,7 +75,7 @@ class SubModulesMixin:
                 submodule_rbln_config = submodule_rbln_config_class(**submodule_rbln_config)
                 setattr(rbln_config, submodule_name, submodule_rbln_config)
 
-            submodule_rbln_config = submodule_cls._update_submodule_config(model, submodule_rbln_config)
+            submodule_rbln_config = submodule_cls._update_submodule_config(model, submodule_rbln_config, preprocessors)
 
             rbln_submodule = submodule_cls.from_model(
                 model=torch_submodule,

{optimum_rbln-0.8.3a4.dist-info → optimum_rbln-0.8.4a0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: optimum-rbln
-Version: 0.8.3a4
+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-0.8.3a4.dist-info → optimum_rbln-0.8.4a0.dist-info}/RECORD

@@ -1,8 +1,8 @@
-optimum/rbln/__init__.py,sha256=1AF2jfrUGYLo4ps_OAPxCUagSnHF60N3W45cTeXlDJE,17698
-optimum/rbln/__version__.py,sha256=vl5bkt2g1rkDYLzC7btKWOmkJwvyed8HxQBkBGIvFzA,712
-optimum/rbln/configuration_utils.py,sha256=fE3HlZblxukKSdS-4VofjuyCAiqwPMX8bqXpOiTZp4g,33926
-optimum/rbln/modeling.py,sha256=0CMQnpVvW9evNrTFHM2XFbNpRY1HkbFzYJ5sRyYFq0o,14293
-optimum/rbln/modeling_base.py,sha256=gHfqIO6lKT8smkUthUuRHnbITpxHpnDeBPT8iTeasCk,24575
+optimum/rbln/__init__.py,sha256=32ouGKDGus9k5_kD27CxP8jIQOw66zpDTfS0xs1XlfE,18298
+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
@@ -72,29 +72,30 @@ optimum/rbln/ops/flash_attn.py,sha256=yTCdYQVqm_1rHMHWjrMQaIR8WTuG_xA6t033x1IVvT
 optimum/rbln/ops/kv_cache_update.py,sha256=aIvK2Sp7M3EfJzJgNvIvAJv4emoN6QOhmgaWj-VboLs,1440
 optimum/rbln/ops/linear.py,sha256=5K3pcrrUHu_p8LrMIU-jX2TnafksveFjjZSCsYSp_yw,1328
 optimum/rbln/ops/sliding_window_attn.py,sha256=EQrV_yRGc5z6kvwEsAcLP028bJWkQg2UPI3xubt9skU,3487
-optimum/rbln/transformers/__init__.py,sha256=uV2rEhw93alpbZ-fnVrAex_6QF1sFHcVM5tsJh64osk,11443
-optimum/rbln/transformers/configuration_generic.py,sha256=95ks6REJYuzI1zLwGlPSlxVV45saVcYOob6ihn-WAAY,5092
+optimum/rbln/transformers/__init__.py,sha256=6s-VhsqwptqwUuq7vb847bJlfFgBGshOoK3vaN9i_lI,12043
+optimum/rbln/transformers/configuration_generic.py,sha256=jrehv1oONOS-iBTY5gj2TKUfWjDTnukNJt6cZfNMylU,5213
 optimum/rbln/transformers/modeling_attention_utils.py,sha256=aLyOaq4me1m-JMmnKbuyNQageDxNU2jjEhGE_ew2P5o,11465
-optimum/rbln/transformers/modeling_generic.py,sha256=2BtroigKuu7z7C98dpLwI875R0EoHN-ceHEVbyPQuYk,12212
+optimum/rbln/transformers/modeling_generic.py,sha256=82Wi2K6zAp5tjef05lzYIEqbK93h0_OkPDbElB-VMMs,12568
 optimum/rbln/transformers/modeling_outputs.py,sha256=cd8ZlhHAGq7S6i5-QK6TJCxgORvoPMnZpqPBlUc_pMY,1177
 optimum/rbln/transformers/modeling_rope_utils.py,sha256=6Zg3r-TeUk4WQAlr95pqfhuoAD_RQ4njT1rbO9uPL0Q,14379
-optimum/rbln/transformers/models/__init__.py,sha256=A9ThjEgBo6RZzqGzoY3tgQucdchkuXWrpgJjuERxjcE,12272
+optimum/rbln/transformers/models/__init__.py,sha256=V36KWN0fTL0MvfDduUfjIiwXvWmwDKm43G-g5Y773-I,12943
 optimum/rbln/transformers/models/audio_spectrogram_transformer/__init__.py,sha256=I2vL4lrzbT5p4eJcH-EKHzEfcPkj_XVsie7jb9q6yic,775
 optimum/rbln/transformers/models/audio_spectrogram_transformer/configuration_audio_spectrogram_transformer.py,sha256=z7LJiVJPmnlCM3mcyhPJP8AufSrxO_dsPeJ51onq-Nc,833
 optimum/rbln/transformers/models/audio_spectrogram_transformer/modeling_audio_spectrogram_transformer.py,sha256=FIKEVWpIt6-JQX9B_rAfCrAPqdUHtR2i8D_X2k7639E,1498
-optimum/rbln/transformers/models/auto/__init__.py,sha256=cZaOoQBTfdT5qhcfsPbagDvDQ0Vvr6Oej5y0jE9AFbc,1105
+optimum/rbln/transformers/models/auto/__init__.py,sha256=tdYqXkg9xBGNr4fZjH7_O3qRVbHvpEVjrJ6wtNUMMJM,1150
 optimum/rbln/transformers/models/auto/auto_factory.py,sha256=1CA52xV2dS1Uzumcgqe4zobdpoi-Xt2oNjP3uLFtm08,8020
-optimum/rbln/transformers/models/auto/modeling_auto.py,sha256=AbKqxcUxKtuX6Y2T7CdmrQn8I0r5DDkUzLKBnyor-GQ,4470
+optimum/rbln/transformers/models/auto/modeling_auto.py,sha256=SMsWnD8f7VhKmh7h_S2voksEWlNccfF4fQ7AmwLYq6U,4790
 optimum/rbln/transformers/models/bart/__init__.py,sha256=fVo-gZEmJ0yxkIxEX6ciuRAGgXNyuvaXE2s88bhbjAE,830
 optimum/rbln/transformers/models/bart/bart_architecture.py,sha256=mAepjL0paPMK180vGTTCxXQ-hVZ1DD6JR-GvVNGJLqY,6268
 optimum/rbln/transformers/models/bart/configuration_bart.py,sha256=PrRA7OwPTegPamd_mmVnwNygRbNG7pZrsrXdKyfZ6Bo,1351
 optimum/rbln/transformers/models/bart/modeling_bart.py,sha256=H4MmQZbofb9kJq5WKqoFVjmj3HVtgns3t2F3QdSU-QQ,2337
 optimum/rbln/transformers/models/bert/__init__.py,sha256=86FuGRBLw315_Roa9D5OUx6Ku2PM0DqSPZ-YSqbF-io,806
+optimum/rbln/transformers/models/bert/bert_architecture.py,sha256=cZgf-B-FV8qbeJdz2Oa-cHu7crrpwBhr081cEalC-h4,473
 optimum/rbln/transformers/models/bert/configuration_bert.py,sha256=nEZnX6LXpLKWaoPEd4pWSysw9h-PLb2ld0ibC3dcJ7w,1611
-optimum/rbln/transformers/models/bert/modeling_bert.py,sha256=zR0US2laTT0yUkL6yyvrR5STQNJcYqtG98ez4SUYQAY,2040
+optimum/rbln/transformers/models/bert/modeling_bert.py,sha256=7MQZS11k4__oyeni5ek2SzRf-gtD3_hMKl_oOzN7_XQ,2263
 optimum/rbln/transformers/models/blip_2/__init__.py,sha256=L01gPXcUCa8Vg-bcng20vZvBIN_jlqCzwUSFuq0QOag,855
 optimum/rbln/transformers/models/blip_2/configuration_blip_2.py,sha256=JUUp4SahBYwv_o2dsHMsgESbPCJHgrng5m7wwtd7HRQ,3193
-optimum/rbln/transformers/models/blip_2/modeling_blip_2.py,sha256=WrxH-VljAbYEFnxo5tQYE29s7TvoRX0L0IPB2TKN27I,16158
+optimum/rbln/transformers/models/blip_2/modeling_blip_2.py,sha256=pI0HCYIy6SsBc2umWuzXHM6tdu_9e2I5gntoQRoxuhA,16264
 optimum/rbln/transformers/models/clip/__init__.py,sha256=TLeXDqcFK6M6v9x7Xr64kBbqGu3hFHM7p754dQ8UVQc,938
 optimum/rbln/transformers/models/clip/configuration_clip.py,sha256=ishrDbTdJm7_AfOn2MPAdAzCWXMdQldwgx9wR_6GcWU,3808
 optimum/rbln/transformers/models/clip/modeling_clip.py,sha256=knK7gINAluSHcWvg3zaByb3XRLNmSEGw2NcsOGHnIow,12364
@@ -104,13 +105,13 @@ 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=ganUtyPKcdKe5QCJ2atQhblzfAstyYkThYDm_DIecU8,1014
+optimum/rbln/transformers/models/depth_anything/modeling_depth_anything.py,sha256=tTmsVaW9Wb2WD3nKRLwp7swn3hbMvgwUEJwwVIfNYEc,1008
 optimum/rbln/transformers/models/distilbert/__init__.py,sha256=zXL78SOEORTnUN_wrdoaDaYpntG8lcFHvPobM6jC0CI,841
 optimum/rbln/transformers/models/distilbert/configuration_distilbert.py,sha256=O3BW9JjyYk9PLyiofvOKEgTdMZ_jpIuPfot281pSsyg,984
 optimum/rbln/transformers/models/distilbert/modeling_distilbert.py,sha256=LUh6zYGa8AR3Yxaj3gtyJRc-czBN3qnHTc-JTAhuqY0,1099
@@ -129,11 +130,15 @@ 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=2Fg1eyCDQ8mwGWdcB_jaq_i_iSEzXr3UWjWXe1Z_Ie8,24079
+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
 optimum/rbln/transformers/models/gpt2/modeling_gpt2.py,sha256=DhF6hU3oCYGbZ7UijKCsRfTx-VCkTqqqNwqqMSrjqRE,2230
+optimum/rbln/transformers/models/grounding_dino/__init__.py,sha256=DE7DipZGvrKC6b1T77k4I4X3G70ss8mlr-PrZCaohto,307
+optimum/rbln/transformers/models/grounding_dino/configuration_grounding_dino.py,sha256=b6aeAlAMf0aOoTKAqe5nnBfontu_H3zvIHgOiCNMJ1I,3127
+optimum/rbln/transformers/models/grounding_dino/grounding_dino_architecture.py,sha256=A_YBgvPVHwwKgsGLL0z4MyTKb6Hb6r3y6sU3oVIrKiU,22779
+optimum/rbln/transformers/models/grounding_dino/modeling_grounding_dino.py,sha256=bXAOs2QH4sy2UFoFLUSM6u1_VHouUT5COERLQX20F6Y,46897
 optimum/rbln/transformers/models/idefics3/__init__.py,sha256=ulxE7HEfXsNJhd25J9Fvi6vggo9aZH9sLKJjWB6LlzQ,814
 optimum/rbln/transformers/models/idefics3/configuration_idefics3.py,sha256=8BhPLkfE1_ZU0eSm2iTbWQOnVe1q0g99srYHWZM6VJ4,2373
 optimum/rbln/transformers/models/idefics3/modeling_idefics3.py,sha256=UqKUVZ6pZjP2VMfBa3-dJkLNPDqr3H1wHiOo9LPucjs,19636
@@ -177,7 +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
@@ -198,7 +203,7 @@ optimum/rbln/transformers/models/siglip/configuration_siglip.py,sha256=m1h8iDx_X
 optimum/rbln/transformers/models/siglip/modeling_siglip.py,sha256=1TyRaxmhp6mg6UfhQTbZhW26013TE3nVnroYG7EROcU,8033
 optimum/rbln/transformers/models/swin/__init__.py,sha256=gUsLDB8ceNxt53Cf69OT32JuZoRdmmIsRfjRdHTLDd0,698
 optimum/rbln/transformers/models/swin/configuration_swin.py,sha256=iVtpT2jXY5LNkUbbr5J08z97unc43KEhArIZ1tBRzEU,1692
-optimum/rbln/transformers/models/swin/modeling_swin.py,sha256=IvBOyIWq233UGZqUG7XhaoYIHk6gjbisYUEr8HDqRbY,12791
+optimum/rbln/transformers/models/swin/modeling_swin.py,sha256=npQgTCEkonG41HzHzEk-a13NFLJHA-K82HFW2VyR0xc,13968
 optimum/rbln/transformers/models/t5/__init__.py,sha256=R1Q8Z1vaIdx4rDjeCmm_ZMSgewWaqaI0l93AHwewtew,818
 optimum/rbln/transformers/models/t5/configuration_t5.py,sha256=nqDbibqykeeWn1TlKk6LmCn-DawTVudMMuBn2c2jds8,1362
 optimum/rbln/transformers/models/t5/modeling_t5.py,sha256=pdAWBLVknTzbma0Ij-VQ2Qve-frPjxL-AwMyU-zouPY,5123
@@ -222,7 +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
@@ -232,8 +237,8 @@ optimum/rbln/utils/logging.py,sha256=VKKBmlQSdg6iZCGmAXaWYiW67K84jyp1QJhLQSSjPPE
 optimum/rbln/utils/model_utils.py,sha256=4k5879Kh75m3x_vS4-qOGfqsOiAvc2kdNFFfvsFvz3k,1748
 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=w5mgPgncI740gVKMu3S-69DGNdUSI0bTZxegQGcZ98Y,5011
-optimum_rbln-0.8.3a4.dist-info/METADATA,sha256=4ikxK2ldRi8PkKZrcSmTH4unPPzl3UZ3JBXlC-fr_MU,5299
-optimum_rbln-0.8.3a4.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-optimum_rbln-0.8.3a4.dist-info/licenses/LICENSE,sha256=QwcOLU5TJoTeUhuIXzhdCEEDDvorGiC6-3YTOl4TecE,11356
-optimum_rbln-0.8.3a4.dist-info/RECORD,,
+optimum/rbln/utils/submodule.py,sha256=60NGLFvnhjP1DJg1opdb-FVQDsthcLCwWjW_1WQaasU,5280
+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,,