optimum-rbln 0.7.4a4__py3-none-any.whl → 0.7.4a6__py3-none-any.whl

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

@@ -12,4 +12,5 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from .configuration_dpt import RBLNDPTForDepthEstimationConfig
 from .modeling_dpt import RBLNDPTForDepthEstimation

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

@@ -0,0 +1,19 @@
+# Copyright 2025 Rebellions Inc. All rights reserved.
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at:
+
+#     http://www.apache.org/licenses/LICENSE-2.0
+
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from ...configuration_generic import RBLNModelForDepthEstimationConfig
+
+
+class RBLNDPTForDepthEstimationConfig(RBLNModelForDepthEstimationConfig):
+    pass

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

@@ -12,82 +12,9 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import TYPE_CHECKING, Any, Dict, Iterable, Optional, Union
 
-from transformers import AutoModelForDepthEstimation
-from transformers.modeling_outputs import DepthEstimatorOutput
+from ...modeling_generic import RBLNModelForDepthEstimation
 
-from ....modeling import RBLNModel
-from ....modeling_config import RBLNCompileConfig, RBLNConfig
-from ....utils.logging import get_logger
 
-
-logger = get_logger(__name__)
-
-if TYPE_CHECKING:
-    from transformers import AutoFeatureExtractor, AutoProcessor, AutoTokenizer, PretrainedConfig
-
-
-class RBLNDPTForDepthEstimation(RBLNModel):
-    auto_model_class = AutoModelForDepthEstimation
-    main_input_name = "pixel_values"
-
-    @classmethod
-    def _get_rbln_config(
-        cls,
-        preprocessors: Optional[Union["AutoFeatureExtractor", "AutoProcessor", "AutoTokenizer"]],
-        model_config: Optional["PretrainedConfig"] = None,
-        rbln_kwargs: Dict[str, Any] = {},
-    ) -> RBLNConfig:
-        rbln_image_size = rbln_kwargs.get("image_size", None)
-        rbln_batch_size = rbln_kwargs.get("batch_size", None)
-
-        if rbln_batch_size is None:
-            rbln_batch_size = 1
-
-        if rbln_image_size is None:
-            for processor in preprocessors:
-                image_size = getattr(processor, "size", None)
-
-                if image_size is not None:
-                    if isinstance(image_size, Iterable):
-                        if "shortest_edge" in image_size:
-                            rbln_image_size = image_size["shortest_edge"]
-                            break
-                        elif "height" in image_size and "width" in image_size:
-                            rbln_image_size = image_size["height"], image_size["width"]
-                            break
-                    else:
-                        rbln_image_size = image_size
-
-            if rbln_image_size is None:
-                rbln_image_size = getattr(model_config, "image_size", None)
-
-            if rbln_image_size is None:
-                raise ValueError("`rbln_image_size` should be specified!")
-
-        if isinstance(rbln_image_size, int):
-            rbln_image_size = rbln_image_size, rbln_image_size
-
-        input_info = [("pixel_values", [rbln_batch_size, 3, rbln_image_size[0], rbln_image_size[1]], "float32")]
-
-        rbln_compile_config = RBLNCompileConfig(input_info=input_info)
-
-        rbln_config = RBLNConfig(
-            rbln_cls=cls.__name__,
-            compile_cfgs=[rbln_compile_config],
-            rbln_kwargs=rbln_kwargs,
-        )
-
-        rbln_config.model_cfg.update(
-            {
-                "image_size": rbln_image_size,
-                "batch_size": rbln_batch_size,
-            }
-        )
-
-        return rbln_config
-
-    def forward(self, *args, **kwargs):
-        predicted_depth = super().forward(*args, **kwargs)
-        return DepthEstimatorOutput(predicted_depth=predicted_depth)
+class RBLNDPTForDepthEstimation(RBLNModelForDepthEstimation):
+    pass

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

@@ -20,4 +20,5 @@ this_path = os.path.abspath(__file__)
 local_dir = "/" + os.path.join(*this_path.split("/")[:-1]) + "/hf_hub_cached"
 environ["LOCAL_CACHE_ROOT_CUSTOM_CODE_MIDM"] = local_dir
 
+from .configuration_exaone import RBLNExaoneForCausalLMConfig
 from .modeling_exaone import RBLNExaoneForCausalLM

optimum/rbln/transformers/models/exaone/configuration_exaone.py

@@ -0,0 +1,19 @@
+# Copyright 2025 Rebellions Inc. All rights reserved.
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at:
+
+#     http://www.apache.org/licenses/LICENSE-2.0
+
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from ..decoderonly.configuration_decoderonly import RBLNDecoderOnlyModelForCausalLMConfig
+
+
+class RBLNExaoneForCausalLMConfig(RBLNDecoderOnlyModelForCausalLMConfig):
+    pass

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

@@ -12,4 +12,5 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from .configuration_gemma import RBLNGemmaForCausalLMConfig
 from .modeling_gemma import RBLNGemmaForCausalLM

optimum/rbln/transformers/models/gemma/configuration_gemma.py

@@ -0,0 +1,19 @@
+# Copyright 2025 Rebellions Inc. All rights reserved.
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at:
+
+#     http://www.apache.org/licenses/LICENSE-2.0
+
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from ..decoderonly.configuration_decoderonly import RBLNDecoderOnlyModelForCausalLMConfig
+
+
+class RBLNGemmaForCausalLMConfig(RBLNDecoderOnlyModelForCausalLMConfig):
+    pass

optimum/rbln/transformers/models/gpt2/__init__.py

@@ -12,4 +12,5 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from .configuration_gpt2 import RBLNGPT2LMHeadModelConfig
 from .modeling_gpt2 import RBLNGPT2LMHeadModel

optimum/rbln/transformers/models/gpt2/configuration_gpt2.py

@@ -0,0 +1,19 @@
+# Copyright 2025 Rebellions Inc. All rights reserved.
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at:
+
+#     http://www.apache.org/licenses/LICENSE-2.0
+
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from ..decoderonly.configuration_decoderonly import RBLNDecoderOnlyModelForCausalLMConfig
+
+
+class RBLNGPT2LMHeadModelConfig(RBLNDecoderOnlyModelForCausalLMConfig):
+    pass

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

@@ -12,4 +12,5 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from .configuration_llama import RBLNLlamaForCausalLMConfig
 from .modeling_llama import RBLNLlamaForCausalLM

optimum/rbln/transformers/models/llama/configuration_llama.py

@@ -0,0 +1,19 @@
+# Copyright 2025 Rebellions Inc. All rights reserved.
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at:
+
+#     http://www.apache.org/licenses/LICENSE-2.0
+
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from ..decoderonly.configuration_decoderonly import RBLNDecoderOnlyModelForCausalLMConfig
+
+
+class RBLNLlamaForCausalLMConfig(RBLNDecoderOnlyModelForCausalLMConfig):
+    pass

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

@@ -12,4 +12,5 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from .configuration_llava_next import RBLNLlavaNextForConditionalGenerationConfig
 from .modeling_llava_next import RBLNLlavaNextForConditionalGeneration

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

@@ -0,0 +1,46 @@
+# Copyright 2025 Rebellions Inc. All rights reserved.
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at:
+
+#     http://www.apache.org/licenses/LICENSE-2.0
+
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Optional
+
+from ....configuration_utils import RBLNModelConfig
+
+
+class RBLNLlavaNextForConditionalGenerationConfig(RBLNModelConfig):
+    submodules = ["vision_tower", "language_model"]
+
+    def __init__(
+        self,
+        batch_size: Optional[int] = None,
+        vision_tower: Optional[RBLNModelConfig] = None,
+        language_model: Optional[RBLNModelConfig] = None,
+        **kwargs,
+    ):
+        """
+        Args:
+            batch_size (Optional[int]): The batch size for inference. Defaults to 1.
+            vision_tower (Optional[RBLNModelConfig]): Configuration for the vision encoder component.
+            language_model (Optional[RBLNModelConfig]): Configuration for the language model component.
+            **kwargs: Additional arguments passed to the parent RBLNModelConfig.
+
+        Raises:
+            ValueError: If batch_size is not a positive integer.
+        """
+        super().__init__(**kwargs)
+        self.batch_size = batch_size or 1
+        if not isinstance(self.batch_size, int) or self.batch_size < 0:
+            raise ValueError(f"batch_size must be a positive integer, got {self.batch_size}")
+
+        self.vision_tower = vision_tower
+        self.language_model = language_model

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

@@ -26,8 +26,8 @@ from transformers import (
 )
 from transformers.modeling_outputs import BaseModelOutputWithPooling
 
+from ....configuration_utils import RBLNCompileConfig, RBLNModelConfig
 from ....modeling import RBLNModel
-from ....modeling_config import RBLNCompileConfig, RBLNConfig
 from ....utils.logging import get_logger
 from ..decoderonly.modeling_decoderonly import RBLNDecoderOnlyOutput
 
@@ -134,7 +134,7 @@ class RBLNLlavaNextForConditionalGeneration(RBLNModel):
         model: "LlavaNextForConditionalGeneration",
         save_dir_path: Path,
         subfolder: str,
-        rbln_config: RBLNConfig,
+        rbln_config: RBLNModelConfig,
     ):
         """
         If you are unavoidably running on a CPU rather than an RBLN device,
@@ -161,42 +161,31 @@ class RBLNLlavaNextForConditionalGeneration(RBLNModel):
         return self.language_model.get_input_embeddings()
 
     @classmethod
-    def wrap_model_if_needed(cls, model: "PreTrainedModel", rbln_config: RBLNConfig):
+    def wrap_model_if_needed(cls, model: "PreTrainedModel", rbln_config: RBLNModelConfig):
         return model.multi_modal_projector
 
     @classmethod
-    def _get_rbln_config(
+    def _update_rbln_config(
         cls,
         preprocessors: Optional[Union["AutoFeatureExtractor", "AutoProcessor", "AutoTokenizer"]],
+        model: Optional["PreTrainedModel"] = None,
         model_config: Optional["PretrainedConfig"] = None,
-        rbln_kwargs={},
-    ) -> RBLNConfig:
-        vision_feature_select_strategy = rbln_kwargs.get("vision_feature_select_strategy", None)
-
-        # 1. Multi-modal projection layer
-        batch_size = rbln_kwargs.get("rbln_batch_size", None)
-        if batch_size is None:
-            batch_size = 1
-
+        rbln_config: Optional[RBLNModelConfig] = None,
+    ) -> RBLNModelConfig:
         feature_size = model_config.vision_config.hidden_size
 
-        # See forward function to see more details.
-        vision_feature_select_strategy = (
-            vision_feature_select_strategy
-            if vision_feature_select_strategy is not None
-            else model_config.vision_feature_select_strategy
-        )
-
         # Calculating `num_positions` : See CLIPVisionEmbeddings of transformers for more details.
         num_positions = (model_config.vision_config.image_size // model_config.vision_config.patch_size) ** 2 + 1
-        if vision_feature_select_strategy == "default":
+        if model_config.vision_feature_select_strategy == "default":
             selected_image_feature_dim = num_positions - 1
         else:
             selected_image_feature_dim = num_positions
 
-        input_info = [("image_features", [batch_size, selected_image_feature_dim, feature_size], "float32")]
+        input_info = [
+            ("image_features", [rbln_config.batch_size, selected_image_feature_dim, feature_size], "float32")
+        ]
         rbln_compile_config = RBLNCompileConfig(input_info=input_info)
-        rbln_config = RBLNConfig(rbln_cls=cls.__name__, compile_cfgs=[rbln_compile_config], rbln_kwargs=rbln_kwargs)
+        rbln_config.set_compile_cfgs([rbln_compile_config])
         return rbln_config
 
     def prepare_inputs_for_generation(

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

@@ -20,4 +20,5 @@ this_path = os.path.abspath(__file__)
 local_dir = "/" + os.path.join(*this_path.split("/")[:-1]) + "/hf_hub_cached"
 environ["LOCAL_CACHE_ROOT_CUSTOM_CODE_MIDM"] = local_dir
 
+from .configuration_midm import RBLNMidmLMHeadModelConfig
 from .modeling_midm import RBLNMidmLMHeadModel

optimum/rbln/transformers/models/midm/configuration_midm.py

@@ -0,0 +1,19 @@
+# Copyright 2025 Rebellions Inc. All rights reserved.
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at:
+
+#     http://www.apache.org/licenses/LICENSE-2.0
+
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from ..decoderonly.configuration_decoderonly import RBLNDecoderOnlyModelForCausalLMConfig
+
+
+class RBLNMidmLMHeadModelConfig(RBLNDecoderOnlyModelForCausalLMConfig):
+    pass

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

@@ -12,4 +12,5 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from .configuration_mistral import RBLNMistralForCausalLMConfig
 from .modeling_mistral import RBLNMistralForCausalLM

optimum/rbln/transformers/models/mistral/configuration_mistral.py

@@ -0,0 +1,19 @@
+# Copyright 2025 Rebellions Inc. All rights reserved.
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at:
+
+#     http://www.apache.org/licenses/LICENSE-2.0
+
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from ..decoderonly.configuration_decoderonly import RBLNDecoderOnlyModelForCausalLMConfig
+
+
+class RBLNMistralForCausalLMConfig(RBLNDecoderOnlyModelForCausalLMConfig):
+    pass

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

@@ -12,4 +12,5 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from .configuration_phi import RBLNPhiForCausalLMConfig
 from .modeling_phi import RBLNPhiForCausalLM

optimum/rbln/transformers/models/phi/configuration_phi.py

@@ -0,0 +1,19 @@
+# Copyright 2025 Rebellions Inc. All rights reserved.
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at:
+
+#     http://www.apache.org/licenses/LICENSE-2.0
+
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from ..decoderonly.configuration_decoderonly import RBLNDecoderOnlyModelForCausalLMConfig
+
+
+class RBLNPhiForCausalLMConfig(RBLNDecoderOnlyModelForCausalLMConfig):
+    pass

optimum/rbln/transformers/models/qwen2/__init__.py

@@ -12,4 +12,5 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from .configuration_qwen2 import RBLNQwen2ForCausalLMConfig
 from .modeling_qwen2 import RBLNQwen2ForCausalLM

optimum/rbln/transformers/models/qwen2/configuration_qwen2.py

@@ -0,0 +1,19 @@
+# Copyright 2025 Rebellions Inc. All rights reserved.
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at:
+
+#     http://www.apache.org/licenses/LICENSE-2.0
+
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from ..decoderonly.configuration_decoderonly import RBLNDecoderOnlyModelForCausalLMConfig
+
+
+class RBLNQwen2ForCausalLMConfig(RBLNDecoderOnlyModelForCausalLMConfig):
+    pass

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

@@ -0,0 +1,19 @@
+# Copyright 2025 Rebellions Inc. All rights reserved.
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at:
+
+#     http://www.apache.org/licenses/LICENSE-2.0
+
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .configuration_qwen2_5_vl import (
+    RBLNQwen2_5_VisionTransformerPretrainedModelConfig,
+    RBLNQwen2_5_VLForConditionalGenerationConfig,
+)
+from .modeling_qwen2_5_vl import RBLNQwen2_5_VisionTransformerPretrainedModel, RBLNQwen2_5_VLForConditionalGeneration

optimum/rbln/transformers/models/qwen2_5_vl/configuration_qwen2_5_vl.py

@@ -0,0 +1,68 @@
+# Copyright 2025 Rebellions Inc. All rights reserved.
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at:
+
+#     http://www.apache.org/licenses/LICENSE-2.0
+
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import List, Optional, Union
+
+from ....configuration_utils import RBLNModelConfig
+from ..decoderonly.configuration_decoderonly import RBLNDecoderOnlyModelForCausalLMConfig
+
+
+class RBLNQwen2_5_VLForConditionalGenerationConfig(RBLNDecoderOnlyModelForCausalLMConfig):
+    submodules = ["visual"]
+
+    def __init__(
+        self,
+        visual: Optional[RBLNModelConfig] = None,
+        use_inputs_embeds: bool = True,
+        **kwargs,
+    ):
+        super().__init__(use_inputs_embeds=use_inputs_embeds, **kwargs)
+        if not self.use_inputs_embeds:
+            raise ValueError(
+                "RBLNQwen2_5_VLForConditionalGenerationConfig does not allow `use_inputs_embeds` to be set to False, "
+                "as RBLNQwen2_5_VLForConditionalGeneration accepts only `inputs_embeds` as input."
+            )
+        self.visual = visual
+
+
+class RBLNQwen2_5_VisionTransformerPretrainedModelConfig(RBLNModelConfig):
+    def __init__(self, max_seq_lens: Union[int, List[int]] = None, **kwargs):
+        """
+        Args:
+            max_seq_lens (Optional[Union[int, List[int]]]): Maximum sequence lengths for Vision
+                Transformer attention. Can be an integer or list of integers, each indicating
+                the number of patches in a sequence for an image or video. For example, an image
+                of 224x196 pixels with patch size 14 and window size 112 has its width padded to
+                224, forming a 224x224 image. This yields 256 patches [(224/14) * (224/14)], so
+                `max_seq_len` must be at least 256. For window-based attention, `max_seq_len`
+                must be a multiple of `(window_size / patch_size)^2`, e.g., (112/14)^2 = 64,
+                making 256 (64 * 4) valid. RBLN optimization runs inference per image or video
+                frame, so set `max_seq_len` to match the maximum expected resolution to reduce
+                computation. If not provided, a `ValueError` is raised.
+            **kwargs: Additional arguments passed to the parent RBLNModelConfig.
+
+        Raises:
+            ValueError: If batch_size is not a positive integer.
+        """
+        super().__init__(**kwargs)
+
+        if max_seq_lens is not None:
+            if isinstance(max_seq_lens, int):
+                max_seq_lens = [max_seq_lens]
+            elif isinstance(max_seq_lens, list):
+                max_seq_lens.sort(reverse=True)
+        else:
+            raise ValueError("'max_seq_lens' must be specified.")
+
+        self.max_seq_lens = max_seq_lens