optimum-rbln 0.7.5rc0__py3-none-any.whl → 0.7.5rc1__py3-none-any.whl

optimum/rbln/__version__.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.7.5rc0'
-__version_tuple__ = version_tuple = (0, 7, 5, 'rc0')
+__version__ = version = '0.7.5rc1'
+__version_tuple__ = version_tuple = (0, 7, 5, 'rc1')

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

@@ -58,8 +58,7 @@ class _PriorTransformer(torch.nn.Module):
 class RBLNPriorTransformer(RBLNModel):
     hf_library_name = "diffusers"
     auto_model_class = PriorTransformer
-    output_class = PriorTransformerOutput
-    output_key = "predicted_image_embedding"
+    _output_class = PriorTransformerOutput
 
     def __post_init__(self, **kwargs):
         super().__post_init__(**kwargs)

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

@@ -61,8 +61,7 @@ class SD3Transformer2DModelWrapper(torch.nn.Module):
 class RBLNSD3Transformer2DModel(RBLNModel):
     hf_library_name = "diffusers"
     auto_model_class = SD3Transformer2DModel
-    output_class = Transformer2DModelOutput
-    output_key = "sample"
+    _output_class = Transformer2DModelOutput
 
     def __post_init__(self, **kwargs):
         super().__post_init__(**kwargs)

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

@@ -143,8 +143,7 @@ class RBLNUNet2DConditionModel(RBLNModel):
     hf_library_name = "diffusers"
     auto_model_class = UNet2DConditionModel
     _rbln_config_class = RBLNUNet2DConditionModelConfig
-    output_class = UNet2DConditionOutput
-    output_key = "sample"
+    _output_class = UNet2DConditionOutput
 
     def __post_init__(self, **kwargs):
         super().__post_init__(**kwargs)

optimum/rbln/modeling.py

@@ -14,7 +14,7 @@
 
 from pathlib import Path
 from tempfile import TemporaryDirectory
-from typing import TYPE_CHECKING, Dict, List, Optional, Union
+from typing import TYPE_CHECKING, Dict, List, Optional, Union, get_args, get_origin, get_type_hints
 
 import rebel
 import torch
@@ -49,8 +49,7 @@ class RBLNModel(RBLNBaseModel):
         ```
     """
 
-    output_class = None
-    output_key = "last_hidden_state"
+    _output_class = None
 
     @classmethod
     def update_kwargs(cls, kwargs):
@@ -245,16 +244,61 @@ class RBLNModel(RBLNBaseModel):
         # Format output according to task requirements
         return self._prepare_output(output, return_dict)
 
+    @classmethod
+    def get_hf_output_class(cls):
+        """
+        Dynamically gets the output class from the corresponding HuggingFace model class.
+
+        Returns:
+            type: The appropriate output class from transformers or diffusers
+        """
+        if cls._output_class:
+            return cls._output_class
+
+        hf_class = cls.get_hf_class()
+        if hf_class is None:
+            raise ValueError(f"No HuggingFace model class found for {cls.__name__}")
+
+        hints = get_type_hints(hf_class.forward) if hasattr(hf_class, "forward") else {}
+        ret = hints.get("return")
+
+        if ret is not None:
+            candidates = get_args(ret) if get_origin(ret) is Union else (ret,)
+
+            for t in candidates:
+                if t is type(None):  # Skip NoneType in Union
+                    continue
+                mod = getattr(t, "__module__", "")
+                if "transformers" in mod or "diffusers" in mod:
+                    cls._output_class = t
+                    return t
+
+        # Fallback to BaseModelOutput
+        cls._output_class = BaseModelOutput
+        return BaseModelOutput
+
     def _prepare_output(self, output, return_dict):
         """
         Prepare model output based on return_dict flag.
         This method can be overridden by subclasses to provide task-specific output handling.
         """
+        tuple_output = (output,) if not isinstance(output, (tuple, list)) else output
         if not return_dict:
-            return (output,) if not isinstance(output, (tuple, list)) else output
+            return tuple_output
         else:
-            if self.output_class is None:
-                return BaseModelOutput(last_hidden_state=output)
-
-            # Create output with the appropriate class and key
-            return self.output_class(**{self.output_key: output})
+            output_class = self.get_hf_output_class()
+            if hasattr(output_class, "loss"):
+                tuple_output = (None,) + tuple_output
+
+            # Truncate if we have too many outputs, otherwise use as is
+            if hasattr(output_class, "__annotations__"):
+                num_fields = len(output_class.__annotations__)
+                if len(tuple_output) > num_fields:
+                    tuple_output = tuple_output[:num_fields]
+                    logger.warning(
+                        f"Truncating output to {num_fields} fields for {output_class.__name__}. "
+                        f"Expected {num_fields} fields, but got {len(tuple_output)} fields."
+                        "This is unexpected. Please report this issue to the developers."
+                    )
+
+            return output_class(*tuple_output)

optimum/rbln/modeling_base.py

@@ -178,9 +178,27 @@ class RBLNBaseModel(SubModulesMixin, PushToHubMixin, PreTrainedModel):
         return str(model_path)
 
     @classmethod
-    def _load_compiled_models(cls, model_path: str):
+    def _load_compiled_models(cls, model_path: str, expected_compiled_model_names: List[str]):
         compiled_models = Path(model_path).glob("*.rbln")
-        rbln_compiled_models = {cm.stem: rebel.RBLNCompiledModel(cm) for cm in compiled_models}
+        expected_compiled_models = [
+            Path(model_path) / f"{compiled_model_name}.rbln" for compiled_model_name in expected_compiled_model_names
+        ]
+        unexpected_compiled_models = [cm for cm in compiled_models if cm not in expected_compiled_models]
+        if unexpected_compiled_models:
+            # TODO(jongho): fix after May release. raise error if unexpected compiled models are found
+            logger.warning(
+                f"Unexpected compiled models found: {[cm.name for cm in unexpected_compiled_models]}. "
+                f"Please check the model path: {model_path}"
+            )
+
+        rbln_compiled_models = {}
+        for compiled_model in expected_compiled_models:
+            if not compiled_model.exists():
+                raise FileNotFoundError(
+                    f"Expected RBLN compiled model '{compiled_model.name}' not found at '{model_path}'. "
+                    "Please ensure all models specified in `rbln_config` are present."
+                )
+            rbln_compiled_models[compiled_model.stem] = rebel.RBLNCompiledModel(compiled_model)
         return rbln_compiled_models
 
     @classmethod
@@ -271,7 +289,8 @@ class RBLNBaseModel(SubModulesMixin, PushToHubMixin, PreTrainedModel):
                     )
                     config = PretrainedConfig(**config)
 
-            rbln_compiled_models = cls._load_compiled_models(model_path_subfolder)
+            compiled_model_names = [cfg.compiled_model_name for cfg in rbln_config.compile_cfgs]
+            rbln_compiled_models = cls._load_compiled_models(model_path_subfolder, compiled_model_names)
 
             if subfolder != "":
                 model_save_dir = Path(model_path_subfolder).absolute().parent

optimum/rbln/transformers/modeling_generic.py

@@ -36,11 +36,7 @@ from transformers import (
 )
 from transformers.modeling_outputs import (
     BaseModelOutput,
-    DepthEstimatorOutput,
-    ImageClassifierOutput,
-    MaskedLMOutput,
     QuestionAnsweringModelOutput,
-    SequenceClassifierOutput,
 )
 
 from ..configuration_utils import RBLNCompileConfig
@@ -63,8 +59,6 @@ class _RBLNTransformerEncoder(RBLNModel):
     auto_model_class = AutoModel
     rbln_model_input_names = ["input_ids", "attention_mask", "token_type_ids"]
     rbln_dtype = "int64"
-    output_class = BaseModelOutput
-    output_key = "last_hidden_state"
 
     @classmethod
     def _update_rbln_config(
@@ -149,7 +143,6 @@ class _RBLNImageModel(RBLNModel):
     auto_model_class = AutoModel
     main_input_name = "pixel_values"
     output_class = BaseModelOutput
-    output_key = "last_hidden_state"
 
     @classmethod
     def _update_rbln_config(
@@ -223,15 +216,11 @@ class RBLNModelForQuestionAnswering(_RBLNTransformerEncoder):
 class RBLNModelForSequenceClassification(_RBLNTransformerEncoder):
     auto_model_class = AutoModelForSequenceClassification
     rbln_model_input_names = ["input_ids", "attention_mask"]
-    output_class = SequenceClassifierOutput
-    output_key = "logits"
 
 
 class RBLNModelForMaskedLM(_RBLNTransformerEncoder):
     auto_model_class = AutoModelForMaskedLM
     rbln_model_input_names = ["input_ids", "attention_mask"]
-    output_class = MaskedLMOutput
-    output_key = "logits"
 
 
 class RBLNModelForTextEncoding(_RBLNTransformerEncoder):
@@ -243,20 +232,14 @@ class RBLNTransformerEncoderForFeatureExtraction(_RBLNTransformerEncoder):
     # TODO: RBLNModel is also for feature extraction.
     auto_model_class = AutoModel
     rbln_model_input_names = ["input_ids", "attention_mask"]
-    output_class = BaseModelOutput
-    output_key = "last_hidden_state"
 
 
 class RBLNModelForImageClassification(_RBLNImageModel):
     auto_model_class = AutoModelForImageClassification
-    output_class = ImageClassifierOutput
-    output_key = "logits"
 
 
 class RBLNModelForDepthEstimation(_RBLNImageModel):
     auto_model_class = AutoModelForDepthEstimation
-    output_class = DepthEstimatorOutput
-    output_key = "predicted_depth"
 
 
 class RBLNModelForAudioClassification(RBLNModel):
@@ -273,8 +256,6 @@ class RBLNModelForAudioClassification(RBLNModel):
     """
 
     auto_model_class = AutoModelForAudioClassification
-    output_class = SequenceClassifierOutput
-    output_key = "logits"
 
     @classmethod
     def _update_rbln_config(

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

@@ -15,7 +15,6 @@
 
 import torch
 from transformers import AutoModelForMaskedLM, Wav2Vec2ForCTC
-from transformers.modeling_outputs import CausalLMOutput
 
 from ...modeling_generic import RBLNModelForMaskedLM
 from .configuration_wav2vec import RBLNWav2Vec2ForCTCConfig
@@ -46,8 +45,6 @@ class RBLNWav2Vec2ForCTC(RBLNModelForMaskedLM):
     main_input_name = "input_values"
     auto_model_class = AutoModelForMaskedLM
     rbln_dtype = "float32"
-    output_class = CausalLMOutput
-    output_key = "logits"
 
     @classmethod
     def wrap_model_if_needed(cls, model: torch.nn.Module, rbln_config: RBLNWav2Vec2ForCTCConfig) -> torch.nn.Module:

{optimum_rbln-0.7.5rc0.dist-info → optimum_rbln-0.7.5rc1.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: optimum-rbln
-Version: 0.7.5rc0
+Version: 0.7.5rc1
 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-0.7.5rc0.dist-info → optimum_rbln-0.7.5rc1.dist-info}/RECORD

@@ -1,8 +1,8 @@
 optimum/rbln/__init__.py,sha256=oAnsJSMrPYwBGEttUt3CMXTIESVNe15ftTWRTShwhZI,14386
-optimum/rbln/__version__.py,sha256=34rdkaF19JfWW5k_S6Q9ZJaeOOAsCoPR3_vN57l-b28,521
+optimum/rbln/__version__.py,sha256=6UGbTME6xZZ1ojJoRpul_clf4TsgGIZHt3214_8maxM,521
 optimum/rbln/configuration_utils.py,sha256=gvAjRFEGw5rnSoH0IoyuLrE4fkxtk3DN1pikqrN_Rpk,31277
-optimum/rbln/modeling.py,sha256=4Xwi3ovWDHOOqxUDH_ZgsgTuea8Kyg25D9s81zVYpr0,9669
-optimum/rbln/modeling_base.py,sha256=iQKw2IORu1cN6sOK0xeBVrhatt-ZPeinT_v6l2FnGRw,24173
+optimum/rbln/modeling.py,sha256=CWYpOLQOu1RUQrHvoX3FoidiP2XltDzC9gWIzznUTFo,11455
+optimum/rbln/modeling_base.py,sha256=HQgscr5jpUEtuXU1ACJHSLIntX-kq6Ef0SQ_W2-rp5A,25341
 optimum/rbln/diffusers/__init__.py,sha256=XL6oKPHbPCV6IVCw3fu0-M9mD2KO_x6unx5kJdAtpVY,6180
 optimum/rbln/diffusers/modeling_diffusers.py,sha256=bPyP5RMbOFLb2DfEAuLVp7hTuQWJvWid7El72wGmFrY,19535
 optimum/rbln/diffusers/configurations/__init__.py,sha256=Sk_sQVTuTl01RVgYViWknQSLmulxKaISS0w-oPdNoBQ,1164
@@ -26,10 +26,10 @@ optimum/rbln/diffusers/models/autoencoders/autoencoder_kl.py,sha256=qjReFNmuQEjn
 optimum/rbln/diffusers/models/autoencoders/vae.py,sha256=_fyFco2697uT1zo_P_fGML-_zqZw2sUQp3tRRjA5pg4,4172
 optimum/rbln/diffusers/models/autoencoders/vq_model.py,sha256=DC8Nee8_BabGhagJgpCUDhA-oaTpZMg-lCVzXJ6dNEw,6134
 optimum/rbln/diffusers/models/transformers/__init__.py,sha256=V8rSR7WzHs-i8Cwb_MNxhY2NFbwPgxu24vGtkwl-6tk,706
-optimum/rbln/diffusers/models/transformers/prior_transformer.py,sha256=d7CYmm88lozepqXjmrFr4qsQ-lRE_10wQRwnenMSflU,4989
-optimum/rbln/diffusers/models/transformers/transformer_sd3.py,sha256=ywWiRAYQ0wfKMMZBrJ9C34GBwIu92F5QXSG8qi7Cr6s,6579
+optimum/rbln/diffusers/models/transformers/prior_transformer.py,sha256=XaIICLeMdGyqm9B3f2A3vqh1haJpqfT3GJ3ZM0DKcaY,4945
+optimum/rbln/diffusers/models/transformers/transformer_sd3.py,sha256=H1dsDOnAK4Dp0ixCVIt_4_4KJ5ZcTygfG7sFFdpOvrI,6554
 optimum/rbln/diffusers/models/unets/__init__.py,sha256=MaICuK9CWjgzejXy8y2NDrphuEq1rkzanF8u45k6O5I,655
-optimum/rbln/diffusers/models/unets/unet_2d_condition.py,sha256=dyrez3vS0_YSxTVwcjlSijDJhe6XchdsLsCFv74HiFQ,15555
+optimum/rbln/diffusers/models/unets/unet_2d_condition.py,sha256=3dzqJQPiklkgoyxRHysOE7q9hrhaT4K0_SNiCflFvLg,15530
 optimum/rbln/diffusers/pipelines/__init__.py,sha256=5KLZ5LrpMzBya2e_3_PvEoPwG24U8JMexfw_ygZREKc,3140
 optimum/rbln/diffusers/pipelines/controlnet/__init__.py,sha256=n1Ef22TSeax-kENi_d8K6wGGHSNEo9QkUeygELHgcao,983
 optimum/rbln/diffusers/pipelines/controlnet/multicontrolnet.py,sha256=Cv__E0Boc6TSOIv8TdXVE821zIiPG4MVI_lnaGSqquk,4102
@@ -65,7 +65,7 @@ optimum/rbln/transformers/__init__.py,sha256=LW6abfb0W0jHziE8dIEDBeyb4Cj-aq8dUld
 optimum/rbln/transformers/configuration_alias.py,sha256=qFVfg6ohsR7a6b-CBgxjBUPDrk9MyiJwtO8AQah_RTU,1505
 optimum/rbln/transformers/configuration_generic.py,sha256=XIiZ1-5p1CMHhG7Sr2qR4SLYKcYw9aph7eGlga3Opx0,5056
 optimum/rbln/transformers/modeling_alias.py,sha256=yx7FnZQWAnrWzivaO5hI7T6i-fyLzt2tMIXG2oDNbPo,1657
-optimum/rbln/transformers/modeling_generic.py,sha256=nT_lytAILkYtwBVJKxXg0dxmh0UpjGYO6zOdLoMs1uU,12891
+optimum/rbln/transformers/modeling_generic.py,sha256=L5ndJJzKhXa4de1YAA8uxNzMKWOHsAHPoJrANxWYWjE,12265
 optimum/rbln/transformers/modeling_rope_utils.py,sha256=3zwkhYUyTZhxCJUSmwCc88iiY1TppRWEY9ShwUqNB2k,14293
 optimum/rbln/transformers/models/__init__.py,sha256=qNh_d7bBKxhxBbUImXJ66n0Vo0NW1m7tMIU5M2ZxGmw,8510
 optimum/rbln/transformers/models/auto/__init__.py,sha256=34Xghf1ogG4u-jhBMlj134nHdgnR3JEHSeZTPuy3MpY,1071
@@ -158,7 +158,7 @@ optimum/rbln/transformers/models/time_series_transformers/modeling_time_series_t
 optimum/rbln/transformers/models/time_series_transformers/time_series_transformers_architecture.py,sha256=XJDjQGbWXUq4ZimNojlcbm3mTDpxUMCl6tkFSzfYFl4,13769
 optimum/rbln/transformers/models/wav2vec2/__init__.py,sha256=dzXqyf_uiI45hPJGbnF1v780Izi2TigsbAo3hxFmhy0,709
 optimum/rbln/transformers/models/wav2vec2/configuration_wav2vec.py,sha256=hXsN_pc_gb_xcQdXXnvpp-o0dk5lNepXnt9O5HB-3g4,771
-optimum/rbln/transformers/models/wav2vec2/modeling_wav2vec2.py,sha256=Sw44ZJVQpXfBiq34oyKxSY2SslCeF3QT_yBysWiTyHY,2060
+optimum/rbln/transformers/models/wav2vec2/modeling_wav2vec2.py,sha256=Lmm39NhvJIQtCkBa5BynkJYNqWOF7GaWsV5qYX-4L94,1943
 optimum/rbln/transformers/models/whisper/__init__.py,sha256=ErquiUlYycSYPsDcq9IwwmbZXoYLn1MVZ8VikWY5gQo,792
 optimum/rbln/transformers/models/whisper/configuration_whisper.py,sha256=-Su7pbkg3gkYTf-ECRJyxkpD3JtUJX4y5Mfml8tJJBI,2612
 optimum/rbln/transformers/models/whisper/generation_whisper.py,sha256=GIHTca3b1VtW81kp7BzKQ7f77c2t9OsEsbZetripgDo,4582
@@ -178,7 +178,7 @@ optimum/rbln/utils/model_utils.py,sha256=V2kFpUe2aqVzLwbpztD8JOVFQqRHncvIWwJbgnU
 optimum/rbln/utils/runtime_utils.py,sha256=LoKNK3AQNV_BSScstIZWjICkJf265MnUgy360BOocVI,5454
 optimum/rbln/utils/save_utils.py,sha256=hG5uOtYmecSXZuGTvCXsTM-SiyZpr5q3InUGCCq_jzQ,3619
 optimum/rbln/utils/submodule.py,sha256=ZfI7e3YzbjbbBW4Yjfucj8NygEsukfIkaJi3PtwHrhc,5105
-optimum_rbln-0.7.5rc0.dist-info/METADATA,sha256=aXeccsNinGR5xXxBOKIMhxfeHyF-wQE5DxWtOrt2WyI,5300
-optimum_rbln-0.7.5rc0.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-optimum_rbln-0.7.5rc0.dist-info/licenses/LICENSE,sha256=QwcOLU5TJoTeUhuIXzhdCEEDDvorGiC6-3YTOl4TecE,11356
-optimum_rbln-0.7.5rc0.dist-info/RECORD,,
+optimum_rbln-0.7.5rc1.dist-info/METADATA,sha256=RUPCGW8cEzu6extEsTB9xYDgOb8hAqgEKG0tG3K5feA,5300
+optimum_rbln-0.7.5rc1.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+optimum_rbln-0.7.5rc1.dist-info/licenses/LICENSE,sha256=QwcOLU5TJoTeUhuIXzhdCEEDDvorGiC6-3YTOl4TecE,11356
+optimum_rbln-0.7.5rc1.dist-info/RECORD,,