optimum-rbln 0.7.4a6__tar.gz → 0.7.4a7__tar.gz

{optimum_rbln-0.7.4a6 → optimum_rbln-0.7.4a7}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: optimum-rbln
-Version: 0.7.4a6
+Version: 0.7.4a7
 Summary: Optimum RBLN is the interface between the Hugging Face Transformers and Diffusers libraries and RBLN accelerators. It provides a set of tools enabling easy model loading and inference on single and multiple rbln device settings for different downstream tasks.
 Project-URL: Homepage, https://rebellions.ai
 Project-URL: Documentation, https://docs.rbln.ai

optimum_rbln-0.7.4a7/examples/image-to-text/run_idefics3.py

@@ -0,0 +1,67 @@
+import os
+import typing
+
+import fire
+from datasets import load_dataset
+from transformers import AutoProcessor
+
+from optimum.rbln import RBLNIdefics3ForConditionalGeneration
+
+
+def main(
+    model_id: str = "HuggingFaceM4/Idefics3-8B-Llama3",
+    batch_size: int = 1,
+    from_transformers: bool = False,
+    prompt: typing.Optional[str] = None,
+    max_seq_len: typing.Optional[int] = None,
+    tensor_parallel_size: typing.Optional[int] = 4,
+):
+    processor = AutoProcessor.from_pretrained(model_id)
+
+    if from_transformers:
+        model = RBLNIdefics3ForConditionalGeneration.from_pretrained(
+            model_id,
+            export=True,
+            rbln_config={
+                "text_model": {
+                    "attn_impl": "flash_attn",
+                    "max_seq_len": max_seq_len,
+                    "use_inputs_embeds": True,
+                    "tensor_parallel_size": tensor_parallel_size,
+                    "batch_size": batch_size,
+                }
+            },
+        )
+        model.save_pretrained(os.path.basename(model_id))
+    else:
+        model = RBLNIdefics3ForConditionalGeneration.from_pretrained(
+            os.path.basename(model_id),
+            export=False,
+        )
+
+    ds = load_dataset("HuggingFaceM4/the_cauldron", "ai2d", split="train")
+    samples = ds.select(range(batch_size))
+    images = []
+    prompts = []
+
+    for sample in samples:
+        img = sample["images"]
+        images.append(img)
+
+        message = [{"role": "user", "content": [{"type": "image"}, {"type": "text", "text": "Describe this image."}]}]
+        prompt = processor.apply_chat_template(message, add_generation_prompt=True)
+        prompts.append(prompt)
+
+    inputs = processor(text=prompts, images=images, return_tensors="pt", padding=True)
+    inputs = dict(inputs)
+    # Generate
+
+    generated_ids = model.generate(**inputs, max_new_tokens=500)
+    generated_texts = processor.batch_decode(generated_ids, skip_special_tokens=True)
+
+    for i, text in enumerate(generated_texts):
+        print(f"Sample {i + 1} generate:\n{text}\n")
+
+
+if __name__ == "__main__":
+    fire.Fire(main)

{optimum_rbln-0.7.4a6 → optimum_rbln-0.7.4a7}/src/optimum/rbln/__init__.py

@@ -74,6 +74,10 @@ _import_structure = {
         "RBLNGemmaForCausalLMConfig",
         "RBLNGPT2LMHeadModel",
         "RBLNGPT2LMHeadModelConfig",
+        "RBLNIdefics3VisionTransformer",
+        "RBLNIdefics3ForConditionalGeneration",
+        "RBLNIdefics3ForConditionalGenerationConfig",
+        "RBLNIdefics3VisionTransformerConfig",
         "RBLNLlamaForCausalLM",
         "RBLNLlamaForCausalLMConfig",
         "RBLNLlavaNextForConditionalGeneration",
@@ -281,6 +285,10 @@ if TYPE_CHECKING:
         RBLNGemmaForCausalLMConfig,
         RBLNGPT2LMHeadModel,
         RBLNGPT2LMHeadModelConfig,
+        RBLNIdefics3ForConditionalGeneration,
+        RBLNIdefics3ForConditionalGenerationConfig,
+        RBLNIdefics3VisionTransformer,
+        RBLNIdefics3VisionTransformerConfig,
         RBLNLlamaForCausalLM,
         RBLNLlamaForCausalLMConfig,
         RBLNLlavaNextForConditionalGeneration,

{optimum_rbln-0.7.4a6 → optimum_rbln-0.7.4a7}/src/optimum/rbln/__version__.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.7.4a6'
-__version_tuple__ = version_tuple = (0, 7, 4, 'a6')
+__version__ = version = '0.7.4a7'
+__version_tuple__ = version_tuple = (0, 7, 4, 'a7')

{optimum_rbln-0.7.4a6 → optimum_rbln-0.7.4a7}/src/optimum/rbln/transformers/__init__.py

@@ -68,6 +68,10 @@ _import_structure = {
         "RBLNGemmaForCausalLMConfig",
         "RBLNGPT2LMHeadModel",
         "RBLNGPT2LMHeadModelConfig",
+        "RBLNIdefics3VisionTransformer",
+        "RBLNIdefics3ForConditionalGeneration",
+        "RBLNIdefics3ForConditionalGenerationConfig",
+        "RBLNIdefics3VisionTransformerConfig",
         "RBLNLlamaForCausalLM",
         "RBLNLlamaForCausalLMConfig",
         "RBLNLlavaNextForConditionalGeneration",
@@ -169,6 +173,10 @@ if TYPE_CHECKING:
         RBLNGemmaForCausalLMConfig,
         RBLNGPT2LMHeadModel,
         RBLNGPT2LMHeadModelConfig,
+        RBLNIdefics3ForConditionalGeneration,
+        RBLNIdefics3ForConditionalGenerationConfig,
+        RBLNIdefics3VisionTransformer,
+        RBLNIdefics3VisionTransformerConfig,
         RBLNLlamaForCausalLM,
         RBLNLlamaForCausalLMConfig,
         RBLNLlavaNextForConditionalGeneration,

{optimum_rbln-0.7.4a6 → optimum_rbln-0.7.4a7}/src/optimum/rbln/transformers/models/__init__.py

@@ -73,6 +73,12 @@ _import_structure = {
     "exaone": ["RBLNExaoneForCausalLM", "RBLNExaoneForCausalLMConfig"],
     "gemma": ["RBLNGemmaForCausalLM", "RBLNGemmaForCausalLMConfig"],
     "gpt2": ["RBLNGPT2LMHeadModel", "RBLNGPT2LMHeadModelConfig"],
+    "idefics3": [
+        "RBLNIdefics3VisionTransformer",
+        "RBLNIdefics3ForConditionalGeneration",
+        "RBLNIdefics3ForConditionalGenerationConfig",
+        "RBLNIdefics3VisionTransformerConfig",
+    ],
     "llama": ["RBLNLlamaForCausalLM", "RBLNLlamaForCausalLMConfig"],
     "llava_next": ["RBLNLlavaNextForConditionalGeneration", "RBLNLlavaNextForConditionalGenerationConfig"],
     "midm": ["RBLNMidmLMHeadModel", "RBLNMidmLMHeadModelConfig"],
@@ -144,6 +150,12 @@ if TYPE_CHECKING:
     from .exaone import RBLNExaoneForCausalLM, RBLNExaoneForCausalLMConfig
     from .gemma import RBLNGemmaForCausalLM, RBLNGemmaForCausalLMConfig
     from .gpt2 import RBLNGPT2LMHeadModel, RBLNGPT2LMHeadModelConfig
+    from .idefics3 import (
+        RBLNIdefics3ForConditionalGeneration,
+        RBLNIdefics3ForConditionalGenerationConfig,
+        RBLNIdefics3VisionTransformer,
+        RBLNIdefics3VisionTransformerConfig,
+    )
     from .llama import RBLNLlamaForCausalLM, RBLNLlamaForCausalLMConfig
     from .llava_next import RBLNLlavaNextForConditionalGeneration, RBLNLlavaNextForConditionalGenerationConfig
     from .midm import RBLNMidmLMHeadModel, RBLNMidmLMHeadModelConfig

{optimum_rbln-0.7.4a6 → optimum_rbln-0.7.4a7}/src/optimum/rbln/transformers/models/decoderonly/decoderonly_architecture.py

@@ -184,6 +184,7 @@ class DecoderOnlyWrapper(nn.Module):
 
     def convert_to_rbln_causal_lm(self, causal_lm: PreTrainedModel, max_seq_len: int):
         new_layers = []
+
         for layer in causal_lm.model.layers:
             if self.attn_impl == "eager":
                 new_self_attn = DecoderOnlyAttention(
@@ -201,6 +202,7 @@ class DecoderOnlyWrapper(nn.Module):
 
             new_layer = DecoderOnlyLayer(layer, new_self_attn)
             new_layers.append(new_layer)
+
         new_model = DecoderOnlyModel(
             causal_lm.model,
             new_layers,

{optimum_rbln-0.7.4a6 → optimum_rbln-0.7.4a7}/src/optimum/rbln/transformers/models/decoderonly/modeling_decoderonly.py

@@ -451,6 +451,12 @@ class RBLNDecoderOnlyModelForCausalLM(RBLNModel):
     def get_input_embeddings(self):
         return self.embed_tokens
 
+    def get_attn_impl(self) -> str:
+        return self.rbln_config.attn_impl
+
+    def get_kvcache_num_blocks(self) -> int:
+        return self.rbln_config.kvcache_num_blocks
+
     @classmethod
     def get_quantized_model(
         cls,

optimum_rbln-0.7.4a7/src/optimum/rbln/transformers/models/idefics3/__init__.py

@@ -0,0 +1,16 @@
+# 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_idefics3 import RBLNIdefics3ForConditionalGenerationConfig, RBLNIdefics3VisionTransformerConfig
+from .modeling_idefics3 import RBLNIdefics3ForConditionalGeneration, RBLNIdefics3VisionTransformer

optimum_rbln-0.7.4a7/src/optimum/rbln/transformers/models/idefics3/configuration_idefics3.py

@@ -0,0 +1,51 @@
+# 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 RBLNIdefics3VisionTransformerConfig(RBLNModelConfig):
+    pass
+
+
+class RBLNIdefics3ForConditionalGenerationConfig(RBLNModelConfig):
+    submodules = ["vision_model", "text_model"]
+
+    def __init__(
+        self,
+        batch_size: Optional[int] = None,
+        vision_model: Optional[RBLNModelConfig] = None,
+        text_model: Optional[RBLNModelConfig] = None,
+        **kwargs,
+    ):
+        """
+        Args:
+            batch_size (Optional[int]): The batch size for inference. Defaults to 1.
+            vision_model (Optional[RBLNModelConfig]): Configuration for the vision transformer component.
+            text_model (Optional[RBLNModelConfig]): Configuration for the text 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_model = vision_model
+        self.text_model = text_model