optimum-rbln 0.8.1a4__py3-none-any.whl → 0.8.1a6__py3-none-any.whl

optimum/rbln/diffusers/pipelines/cosmos/configuration_cosmos_guardrail.py

@@ -0,0 +1,102 @@
+# 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 Any, Dict, Optional, Tuple
+
+from ....configuration_utils import RBLNAutoConfig, RBLNModelConfig
+from ....transformers import RBLNSiglipVisionModelConfig
+
+
+class RBLNVideoSafetyModelConfig(RBLNModelConfig):
+    """
+    Configuration class for RBLN Video Content Safety Filter.
+    """
+
+    def __init__(
+        self,
+        batch_size: Optional[int] = None,
+        input_size: Optional[int] = None,
+        image_size: Optional[Tuple[int, int]] = None,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.batch_size = batch_size or 1
+        self.input_size = input_size or 1152
+
+
+class RBLNRetinaFaceFilterConfig(RBLNModelConfig):
+    """
+    Configuration class for RBLN Retina Face Filter.
+    """
+
+    def __init__(
+        self,
+        batch_size: Optional[int] = None,
+        image_size: Optional[Tuple[int, int]] = None,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.batch_size = batch_size or 1
+        self.image_size = image_size or (704, 1280)
+
+
+class RBLNCosmosSafetyCheckerConfig(RBLNModelConfig):
+    """
+    Configuration class for RBLN Cosmos Safety Checker.
+    """
+
+    submodules = ["aegis", "video_safety_model", "face_blur_filter", "siglip_encoder"]
+
+    def __init__(
+        self,
+        aegis: Optional[RBLNModelConfig] = None,
+        video_safety_model: Optional[RBLNModelConfig] = None,
+        face_blur_filter: Optional[RBLNModelConfig] = None,
+        siglip_encoder: Optional[RBLNSiglipVisionModelConfig] = None,
+        *,
+        batch_size: Optional[int] = None,
+        image_size: Optional[Tuple[int, int]] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        **kwargs: Dict[str, Any],
+    ):
+        super().__init__(**kwargs)
+        if height is not None and width is not None:
+            image_size = (height, width)
+
+        self.aegis = self.init_submodule_config(RBLNModelConfig, aegis)
+        self.siglip_encoder = self.init_submodule_config(
+            RBLNSiglipVisionModelConfig,
+            siglip_encoder,
+            batch_size=batch_size,
+            image_size=(384, 384),
+        )
+
+        self.video_safety_model = self.init_submodule_config(
+            RBLNVideoSafetyModelConfig,
+            video_safety_model,
+            batch_size=batch_size,
+            input_size=1152,
+        )
+        self.face_blur_filter = self.init_submodule_config(
+            RBLNRetinaFaceFilterConfig,
+            face_blur_filter,
+            batch_size=batch_size,
+            image_size=image_size,
+        )
+
+
+RBLNAutoConfig.register(RBLNVideoSafetyModelConfig)
+RBLNAutoConfig.register(RBLNRetinaFaceFilterConfig)
+RBLNAutoConfig.register(RBLNCosmosSafetyCheckerConfig)

optimum/rbln/diffusers/pipelines/cosmos/cosmos_guardrail.py

@@ -0,0 +1,395 @@
+# 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.
+
+import os
+import pathlib
+from functools import partial
+from typing import Any, Dict, Optional, Tuple, Union
+from unittest.mock import patch
+
+import rebel
+import torch
+from diffusers.utils import is_cosmos_guardrail_available
+from huggingface_hub import snapshot_download
+from transformers import AutoTokenizer, SiglipProcessor
+
+from .... import RBLNAutoModelForCausalLM, RBLNSiglipVisionModel
+from ....utils.runtime_utils import RBLNPytorchRuntime
+from .configuration_cosmos_guardrail import RBLNCosmosSafetyCheckerConfig
+
+
+if is_cosmos_guardrail_available():
+    from cosmos_guardrail import CosmosSafetyChecker
+    from cosmos_guardrail.cosmos_guardrail import (
+        COSMOS_GUARDRAIL_CHECKPOINT,
+        Aegis,
+        Blocklist,
+        GuardrailRunner,
+        ModelConfig,
+        RetinaFaceFilter,
+        SafetyClassifier,
+        SigLIPEncoder,
+        VideoContentSafetyFilter,
+        VideoSafetyModel,
+    )
+    from retinaface.data import cfg_re50
+
+    COSMOS_AVAILABLE = True
+else:
+    COSMOS_AVAILABLE = False
+
+    class FailToImportCosmosGuardrail(torch.nn.Module): ...
+
+    class CosmosSafetyChecker(FailToImportCosmosGuardrail): ...
+
+    COSMOS_GUARDRAIL_CHECKPOINT = None
+
+    class Aegis(FailToImportCosmosGuardrail): ...
+
+    class Blocklist(FailToImportCosmosGuardrail): ...
+
+    class GuardrailRunner(FailToImportCosmosGuardrail): ...
+
+    class ModelConfig(FailToImportCosmosGuardrail): ...
+
+    class RetinaFaceFilter(FailToImportCosmosGuardrail): ...
+
+    class SafetyClassifier(FailToImportCosmosGuardrail): ...
+
+    class SigLIPEncoder(FailToImportCosmosGuardrail): ...
+
+    class VideoContentSafetyFilter(FailToImportCosmosGuardrail): ...
+
+    class VideoSafetyModel(FailToImportCosmosGuardrail): ...
+
+    cfg_re50 = None
+
+
+def is_compiled_dir(dir: str) -> bool:
+    # walk directory and check if there is any *.rbln files in that dir.
+    if not os.path.exists(dir):
+        return False
+
+    for root, dirs, files in os.walk(dir):
+        for file in files:
+            if file.endswith(".rbln"):
+                return True
+    return False
+
+
+def get_image_features(
+    self,
+    pixel_values: torch.Tensor,
+    return_dict: bool = True,
+    output_attentions: bool = False,
+    output_hidden_states: bool = False,
+    interpolate_pos_encoding: bool = False,
+):
+    output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+    output_hidden_states = (
+        output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+    )
+    return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+    return self(
+        pixel_values,
+        return_dict=return_dict,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        interpolate_pos_encoding=interpolate_pos_encoding,
+    )[1]
+
+
+class RBLNSigLIPEncoder(SigLIPEncoder):
+    def __init__(
+        self,
+        model_name: str = "google/siglip-so400m-patch14-384",
+        checkpoint_id: str = COSMOS_GUARDRAIL_CHECKPOINT,
+        rbln_config: Optional[RBLNCosmosSafetyCheckerConfig] = None,
+    ):
+        torch.nn.Module.__init__(self)
+        if is_compiled_dir(checkpoint_id):
+            self.checkpoint_dir = (
+                pathlib.Path(checkpoint_id) / "video_content_safety_filter" / "siglip_encoder"
+            ).as_posix()
+            self.processor = SiglipProcessor.from_pretrained(self.checkpoint_dir)
+
+            # We don't use RBLNSiglipModel, but we need to override get_image_features to return pooler_output
+            self.model = RBLNSiglipVisionModel.from_pretrained(
+                self.checkpoint_dir,
+                rbln_device=rbln_config.siglip_encoder.device,
+            )
+        else:
+            super().__init__(model_name, checkpoint_id)
+            model = self.model
+            del self.model
+            self.model = RBLNSiglipVisionModel.from_model(
+                model,
+                rbln_device=rbln_config.siglip_encoder.device,
+                rbln_image_size=rbln_config.siglip_encoder.image_size,
+                rbln_npu=rbln_config.siglip_encoder.npu,
+            )
+        self.rbln_config = rbln_config
+
+        # Override get_image_features to return pooler_output
+        self.model.get_image_features = lambda *args, **kwargs: get_image_features(self.model, *args, **kwargs)
+
+    def save_pretrained(self, checkpoint_id: str):
+        cache_dir = (pathlib.Path(checkpoint_id) / "video_content_safety_filter" / "siglip_encoder").as_posix()
+        self.model.save_pretrained(cache_dir)
+        self.processor.save_pretrained(cache_dir)
+
+
+class RBLNRetinaFaceFilter(RetinaFaceFilter):
+    def __init__(
+        self,
+        checkpoint_id: str = COSMOS_GUARDRAIL_CHECKPOINT,
+        batch_size: int = 1,
+        confidence_threshold: float = 0.7,
+        rbln_config: Optional[RBLNCosmosSafetyCheckerConfig] = None,
+    ):
+        torch.nn.Module.__init__(self)
+        if is_compiled_dir(checkpoint_id):
+            self.compiled_model = rebel.RBLNCompiledModel(
+                pathlib.Path(checkpoint_id) / "face_blur_filter" / "retinaface.rbln"
+            )
+            self.cfg = cfg_re50
+            self.batch_size = batch_size
+            self.confidence_threshold = confidence_threshold
+            self.cfg["pretrain"] = False
+        else:
+            with patch("torch.load", partial(torch.load, weights_only=True, map_location=torch.device("cpu"))):
+                super().__init__(checkpoint_id)
+            net = self.net
+            del self.net
+            self.compiled_model = rebel.compile_from_torch(
+                net,
+                input_info=[
+                    (
+                        "frames",
+                        [
+                            self.batch_size,
+                            3,
+                            rbln_config.face_blur_filter.image_size[0],
+                            rbln_config.face_blur_filter.image_size[1],
+                        ],
+                        "float32",
+                    )
+                ],
+                npu=rbln_config.face_blur_filter.npu,
+            )
+
+        self.rbln_config = rbln_config
+        runtime = rebel.Runtime(self.compiled_model, tensor_type="pt", device=self.rbln_config.face_blur_filter.device)
+        self.net = RBLNPytorchRuntime(runtime)
+
+    def save_pretrained(self, checkpoint_id: str):
+        cache_path = pathlib.Path(checkpoint_id) / "face_blur_filter"
+        cache_path.mkdir(parents=True, exist_ok=True)
+        self.compiled_model.save(cache_path / "retinaface.rbln")
+
+
+class RBLNVideoSafetyModel(VideoSafetyModel):
+    def __init__(
+        self,
+        config: ModelConfig,
+        checkpoint_id: str = COSMOS_GUARDRAIL_CHECKPOINT,
+        rbln_config: Optional["RBLNCosmosSafetyCheckerConfig"] = None,
+    ):
+        torch.nn.Module.__init__(self)
+        self.config = config
+        self.num_classes = config.num_classes
+        self.rbln_config = rbln_config
+
+        if is_compiled_dir(checkpoint_id):
+            self.compiled_model = rebel.RBLNCompiledModel(
+                pathlib.Path(checkpoint_id) / "video_content_safety_filter" / "safety_filter.rbln"
+            )
+        else:
+            # Load model from checkpoint
+            network = SafetyClassifier(
+                input_size=self.rbln_config.video_safety_model.input_size, num_classes=self.num_classes
+            )
+            network.eval()
+
+            checkpoint_dir = snapshot_download(checkpoint_id)
+            checkpoint_dir = (pathlib.Path(checkpoint_dir) / "video_content_safety_filter").as_posix()
+
+            safety_filter_local_path = os.path.join(checkpoint_dir, "safety_filter.pt")
+            checkpoint = torch.load(safety_filter_local_path, weights_only=True)
+            network.load_state_dict({k.replace("network.", ""): v for k, v in checkpoint["model"].items()})
+
+            self.compiled_model = rebel.compile_from_torch(
+                network,
+                input_info=[
+                    (
+                        "data",
+                        [
+                            self.rbln_config.video_safety_model.batch_size,
+                            self.rbln_config.video_safety_model.input_size,
+                        ],
+                        "float32",
+                    )
+                ],
+                npu=self.rbln_config.video_safety_model.npu,
+            )
+
+        runtime = rebel.Runtime(
+            self.compiled_model,
+            tensor_type="pt",
+            device=self.rbln_config.video_safety_model.device,
+        )
+        self.network = RBLNPytorchRuntime(runtime)
+
+    def save_pretrained(self, checkpoint_id: str):
+        cache_path = pathlib.Path(checkpoint_id) / "video_content_safety_filter"
+        cache_path.mkdir(parents=True, exist_ok=True)
+        self.compiled_model.save(cache_path / "safety_filter.rbln")
+
+    def parameters(self):
+        yield torch.tensor([1.0], dtype=torch.float32, device=torch.device("cpu"))
+
+
+class RBLNVideoContentSafetyFilter(VideoContentSafetyFilter):
+    def __init__(
+        self,
+        checkpoint_id: str = COSMOS_GUARDRAIL_CHECKPOINT,
+        rbln_config: Optional["RBLNCosmosSafetyCheckerConfig"] = None,
+    ):
+        torch.nn.Module.__init__(self)
+        self.rbln_config = rbln_config
+        self.encoder = RBLNSigLIPEncoder(checkpoint_id=checkpoint_id, rbln_config=rbln_config)
+
+        model_config = ModelConfig(input_size=1152, num_classes=7)
+        self.model = RBLNVideoSafetyModel(model_config, checkpoint_id=checkpoint_id, rbln_config=rbln_config)
+
+    def save_pretrained(self, checkpoint_id: str):
+        self.model.save_pretrained(checkpoint_id)
+        self.encoder.save_pretrained(checkpoint_id)
+
+
+class RBLNAegis(Aegis):
+    def __init__(
+        self,
+        checkpoint_id: str = COSMOS_GUARDRAIL_CHECKPOINT,
+        base_model_id: str = "meta-llama/LlamaGuard-7b",
+        aegis_adapter: str = "nvidia/Aegis-AI-Content-Safety-LlamaGuard-Defensive-1.0",
+        rbln_config: Optional[RBLNCosmosSafetyCheckerConfig] = None,
+    ) -> None:
+        if is_compiled_dir(checkpoint_id):
+            torch.nn.Module.__init__(self)
+            cache_dir = pathlib.Path(checkpoint_id) / "aegis"
+            self.tokenizer = AutoTokenizer.from_pretrained(cache_dir)
+            self.model = RBLNAutoModelForCausalLM.from_pretrained(cache_dir, rbln_device=rbln_config.aegis.device)
+
+        else:
+            super().__init__(checkpoint_id, base_model_id, aegis_adapter)
+            model = self.model.merge_and_unload()  # peft merge
+            del self.model
+
+            self.model = RBLNAutoModelForCausalLM.from_model(
+                model,
+                rbln_tensor_parallel_size=4,
+                rbln_device=rbln_config.aegis.device,
+                rbln_npu=rbln_config.aegis.npu,
+            )
+
+        self.rbln_config = rbln_config
+        self.dtype = torch.bfloat16
+        self.device = torch.device("cpu")
+
+    def save_pretrained(self, checkpoint_id: str):
+        cache_dir = pathlib.Path(checkpoint_id) / "aegis"
+        self.model.save_pretrained(cache_dir)
+        self.tokenizer.save_pretrained(cache_dir)
+
+
+class RBLNCosmosSafetyChecker(CosmosSafetyChecker):
+    """
+    RBLN-accelerated implementation of Cosmos Safety Checker.
+    """
+
+    def __init__(
+        self,
+        checkpoint_id: str = COSMOS_GUARDRAIL_CHECKPOINT,
+        aegis_model_id: str = "meta-llama/LlamaGuard-7b",
+        aegis_adapter_id: str = "nvidia/Aegis-AI-Content-Safety-LlamaGuard-Defensive-1.0",
+        rbln_config: Optional[RBLNCosmosSafetyCheckerConfig] = None,
+    ) -> None:
+        torch.nn.Module.__init__(self)
+        if not COSMOS_AVAILABLE:
+            raise ImportError(
+                "`cosmos_guardrail` is not installed. Please install it to use the safety checker for Cosmos: `pip install cosmos_guardrail`."
+            )
+
+        if rbln_config is None:
+            rbln_config = RBLNCosmosSafetyCheckerConfig()
+
+        self.text_guardrail = GuardrailRunner(
+            safety_models=[
+                Blocklist(COSMOS_GUARDRAIL_CHECKPOINT),  # Changed since it cannot be saved
+                RBLNAegis(checkpoint_id, aegis_model_id, aegis_adapter_id, rbln_config=rbln_config),
+            ]
+        )
+
+        with patch("torch.load", partial(torch.load, weights_only=True, map_location=torch.device("cpu"))):
+            self.video_guardrail = GuardrailRunner(
+                safety_models=[RBLNVideoContentSafetyFilter(checkpoint_id, rbln_config=rbln_config)],
+                postprocessors=[RBLNRetinaFaceFilter(checkpoint_id, rbln_config=rbln_config)],
+            )
+
+        self.rbln_config = rbln_config
+
+    def save_pretrained(self, save_dir: str):
+        for text_safety_models in self.text_guardrail.safety_models:
+            if isinstance(text_safety_models, RBLNAegis):
+                text_safety_models.save_pretrained(save_dir)
+
+        for video_safety_models in self.video_guardrail.safety_models:
+            if isinstance(video_safety_models, RBLNVideoContentSafetyFilter):
+                video_safety_models.save_pretrained(save_dir)
+
+        for postprocessors in self.video_guardrail.postprocessors:
+            if isinstance(postprocessors, RBLNRetinaFaceFilter):
+                postprocessors.save_pretrained(save_dir)
+
+        self.rbln_config._frozen = True  # Ad-hoc to save config
+        self.rbln_config.save(save_dir)
+
+    @classmethod
+    def from_pretrained(
+        cls,
+        checkpoint_id: str,
+        rbln_config: Optional[RBLNCosmosSafetyCheckerConfig] = None,
+        subfolder: Optional[str] = None,
+        export: Optional[bool] = True,
+        **kwargs,
+    ):
+        rbln_config, kwargs = cls.prepare_rbln_config(rbln_config=rbln_config, **kwargs)
+
+        if len(kwargs) > 0:
+            raise ValueError(f"Unexpected arguments: {kwargs.keys()}")
+
+        if subfolder is not None:
+            checkpoint_id = os.path.join(checkpoint_id, subfolder)
+
+        return cls(checkpoint_id=checkpoint_id, rbln_config=rbln_config)
+
+    @classmethod
+    def prepare_rbln_config(
+        cls, rbln_config: Optional[Union[Dict[str, Any], RBLNCosmosSafetyCheckerConfig]] = None, **kwargs
+    ) -> Tuple[RBLNCosmosSafetyCheckerConfig, Dict[str, Any]]:
+        # Extract rbln-config from kwargs and convert it to RBLNCosmosSafetyCheckerConfig.
+        rbln_config, kwargs = RBLNCosmosSafetyCheckerConfig.initialize_from_kwargs(rbln_config, **kwargs)
+        return rbln_config, kwargs

optimum/rbln/diffusers/pipelines/cosmos/pipeline_cosmos_text2world.py

@@ -0,0 +1,98 @@
+# 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 Any, Dict, Optional
+
+from diffusers import CosmosTextToWorldPipeline
+from diffusers.schedulers import EDMEulerScheduler
+from transformers import T5TokenizerFast
+
+from ....transformers.models.t5.modeling_t5 import RBLNT5EncoderModel
+from ....utils.logging import get_logger
+from ...modeling_diffusers import RBLNDiffusionMixin
+from ...models.autoencoders.autoencoder_kl_cosmos import RBLNAutoencoderKLCosmos
+from ...models.transformers.transformer_cosmos import RBLNCosmosTransformer3DModel
+from .cosmos_guardrail import RBLNCosmosSafetyChecker
+
+
+logger = get_logger(__name__)
+
+
+class RBLNCosmosTextToWorldPipeline(RBLNDiffusionMixin, CosmosTextToWorldPipeline):
+    """
+    RBLN-accelerated implementation of Cosmos Text to World pipeline for text-to-video generation.
+
+    This pipeline compiles Cosmos Text to World models to run efficiently on RBLN NPUs, enabling high-performance
+    inference for generating images with distinctive artistic style and enhanced visual quality.
+    """
+
+    original_class = CosmosTextToWorldPipeline
+    _submodules = ["text_encoder", "transformer", "vae"]
+    _optional_submodules = ["safety_checker"]
+
+    def __init__(
+        self,
+        text_encoder: RBLNT5EncoderModel,
+        tokenizer: T5TokenizerFast,
+        transformer: RBLNCosmosTransformer3DModel,
+        vae: RBLNAutoencoderKLCosmos,
+        scheduler: EDMEulerScheduler,
+        safety_checker: RBLNCosmosSafetyChecker = None,
+    ):
+        if safety_checker is None:
+            safety_checker = RBLNCosmosSafetyChecker()
+
+        super().__init__(
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            transformer=transformer,
+            vae=vae,
+            scheduler=scheduler,
+            safety_checker=safety_checker,
+        )
+
+    def handle_additional_kwargs(self, **kwargs):
+        if "num_frames" in kwargs and kwargs["num_frames"] != self.transformer.rbln_config.num_frames:
+            logger.warning(
+                f"The transformer in this pipeline is compiled with 'num_frames={self.transformer.rbln_config.num_frames}'. 'num_frames' set by the user will be ignored"
+            )
+            kwargs.pop("num_frames")
+        if (
+            "max_sequence_length" in kwargs
+            and kwargs["max_sequence_length"] != self.transformer.rbln_config.max_seq_len
+        ):
+            logger.warning(
+                f"The transformer in this pipeline is compiled with 'max_seq_len={self.transformer.rbln_config.max_seq_len}'. 'max_sequence_length' set by the user will be ignored"
+            )
+            kwargs.pop("max_sequence_length")
+        return kwargs
+
+    @classmethod
+    def from_pretrained(
+        cls,
+        model_id: str,
+        *,
+        export: bool = False,
+        safety_checker: Optional[RBLNCosmosSafetyChecker] = None,
+        rbln_config: Dict[str, Any] = {},
+        **kwargs: Dict[str, Any],
+    ):
+        rbln_config, kwargs = cls.get_rbln_config_class().initialize_from_kwargs(rbln_config, **kwargs)
+        if safety_checker is None and export:
+            safety_checker = RBLNCosmosSafetyChecker(rbln_config=rbln_config.safety_checker)
+
+        return super().from_pretrained(
+            model_id, export=export, safety_checker=safety_checker, rbln_config=rbln_config, **kwargs
+        )

optimum/rbln/diffusers/pipelines/cosmos/pipeline_cosmos_video2world.py

@@ -0,0 +1,98 @@
+# 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 Any, Dict, Optional
+
+from diffusers import CosmosVideoToWorldPipeline
+from diffusers.schedulers import EDMEulerScheduler
+from transformers import T5TokenizerFast
+
+from ....transformers.models.t5.modeling_t5 import RBLNT5EncoderModel
+from ....utils.logging import get_logger
+from ...modeling_diffusers import RBLNDiffusionMixin
+from ...models.autoencoders.autoencoder_kl_cosmos import RBLNAutoencoderKLCosmos
+from ...models.transformers.transformer_cosmos import RBLNCosmosTransformer3DModel
+from .cosmos_guardrail import RBLNCosmosSafetyChecker
+
+
+logger = get_logger(__name__)
+
+
+class RBLNCosmosVideoToWorldPipeline(RBLNDiffusionMixin, CosmosVideoToWorldPipeline):
+    """
+    RBLN-accelerated implementation of Cosmos Video to World pipeline for video-to-video generation.
+
+    This pipeline compiles Cosmos Video to World models to run efficiently on RBLN NPUs, enabling high-performance
+    inference for generating images with distinctive artistic style and enhanced visual quality.
+    """
+
+    original_class = CosmosVideoToWorldPipeline
+    _submodules = ["text_encoder", "transformer", "vae"]
+    _optional_components = ["safety_checker"]
+
+    def __init__(
+        self,
+        text_encoder: RBLNT5EncoderModel,
+        tokenizer: T5TokenizerFast,
+        transformer: RBLNCosmosTransformer3DModel,
+        vae: RBLNAutoencoderKLCosmos,
+        scheduler: EDMEulerScheduler,
+        safety_checker: RBLNCosmosSafetyChecker = None,
+    ):
+        if safety_checker is None:
+            safety_checker = RBLNCosmosSafetyChecker()
+
+        super().__init__(
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            transformer=transformer,
+            vae=vae,
+            scheduler=scheduler,
+            safety_checker=safety_checker,
+        )
+
+    def handle_additional_kwargs(self, **kwargs):
+        if "num_frames" in kwargs and kwargs["num_frames"] != self.transformer.rbln_config.num_frames:
+            logger.warning(
+                f"The transformer in this pipeline is compiled with 'num_frames={self.transformer.rbln_config.num_frames}'. 'num_frames' set by the user will be ignored"
+            )
+            kwargs.pop("num_frames")
+        if (
+            "max_sequence_length" in kwargs
+            and kwargs["max_sequence_length"] != self.transformer.rbln_config.max_seq_len
+        ):
+            logger.warning(
+                f"The transformer in this pipeline is compiled with 'max_seq_len={self.transformer.rbln_config.max_seq_len}'. 'max_sequence_length' set by the user will be ignored"
+            )
+            kwargs.pop("max_sequence_length")
+        return kwargs
+
+    @classmethod
+    def from_pretrained(
+        cls,
+        model_id: str,
+        *,
+        export: bool = False,
+        safety_checker: Optional[RBLNCosmosSafetyChecker] = None,
+        rbln_config: Dict[str, Any] = {},
+        **kwargs: Dict[str, Any],
+    ):
+        rbln_config, kwargs = cls.get_rbln_config_class().initialize_from_kwargs(rbln_config, **kwargs)
+        if safety_checker is None and export:
+            safety_checker = RBLNCosmosSafetyChecker(rbln_config=rbln_config.safety_checker)
+
+        return super().from_pretrained(
+            model_id, export=export, safety_checker=safety_checker, rbln_config=rbln_config, **kwargs
+        )