optimum-rbln 0.9.3.post1__py3-none-any.whl

optimum/rbln/diffusers/models/autoencoders/autoencoder_kl_temporal_decoder.py

@@ -0,0 +1,275 @@
+# 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 TYPE_CHECKING, Dict, List, Tuple, Union
+
+import rebel
+import torch  # noqa: I001
+from diffusers import AutoencoderKLTemporalDecoder
+from diffusers.models.autoencoders.vae import DecoderOutput
+from diffusers.models.modeling_outputs import AutoencoderKLOutput
+from transformers import PretrainedConfig
+
+from ....configuration_utils import RBLNCompileConfig
+from ....modeling import RBLNModel
+from ....utils.logging import get_logger
+from ...configurations import RBLNAutoencoderKLTemporalDecoderConfig
+from ...modeling_diffusers import RBLNDiffusionMixin
+from .vae import (
+    DiagonalGaussianDistribution,
+    RBLNRuntimeVAEDecoder,
+    RBLNRuntimeVAEEncoder,
+    _VAEEncoder,
+    _VAETemporalDecoder,
+)
+
+
+if TYPE_CHECKING:
+    from transformers import AutoFeatureExtractor, AutoProcessor, AutoTokenizer, PretrainedConfig, PreTrainedModel
+
+    from ...modeling_diffusers import RBLNDiffusionMixin, RBLNDiffusionMixinConfig
+
+logger = get_logger(__name__)
+
+
+class RBLNAutoencoderKLTemporalDecoder(RBLNModel):
+    auto_model_class = AutoencoderKLTemporalDecoder
+    hf_library_name = "diffusers"
+    _rbln_config_class = RBLNAutoencoderKLTemporalDecoderConfig
+
+    def __post_init__(self, **kwargs):
+        super().__post_init__(**kwargs)
+
+        if self.rbln_config.uses_encoder:
+            self.encoder = RBLNRuntimeVAEEncoder(runtime=self.model[0], main_input_name="x")
+        self.decoder = RBLNRuntimeVAEDecoder(runtime=self.model[-1], main_input_name="z")
+        self.image_size = self.rbln_config.image_size
+
+    @classmethod
+    def _wrap_model_if_needed(
+        cls, model: torch.nn.Module, rbln_config: RBLNAutoencoderKLTemporalDecoderConfig
+    ) -> torch.nn.Module:
+        decoder_model = _VAETemporalDecoder(model)
+        decoder_model.num_frames = rbln_config.decode_chunk_size
+        decoder_model.eval()
+
+        if rbln_config.uses_encoder:
+            encoder_model = _VAEEncoder(model)
+            encoder_model.eval()
+            return encoder_model, decoder_model
+        else:
+            return decoder_model
+
+    @classmethod
+    def get_compiled_model(
+        cls, model, rbln_config: RBLNAutoencoderKLTemporalDecoderConfig
+    ) -> Dict[str, rebel.RBLNCompiledModel]:
+        compiled_models = {}
+        if rbln_config.uses_encoder:
+            encoder_model, decoder_model = cls._wrap_model_if_needed(model, rbln_config)
+            enc_compiled_model = cls.compile(
+                encoder_model,
+                rbln_compile_config=rbln_config.compile_cfgs[0],
+                create_runtimes=rbln_config.create_runtimes,
+                device=rbln_config.device_map["encoder"],
+            )
+            compiled_models["encoder"] = enc_compiled_model
+        else:
+            decoder_model = cls._wrap_model_if_needed(model, rbln_config)
+        dec_compiled_model = cls.compile(
+            decoder_model,
+            rbln_compile_config=rbln_config.compile_cfgs[-1],
+            create_runtimes=rbln_config.create_runtimes,
+            device=rbln_config.device_map["decoder"],
+        )
+        compiled_models["decoder"] = dec_compiled_model
+
+        return compiled_models
+
+    @classmethod
+    def get_vae_sample_size(
+        cls,
+        pipe: "RBLNDiffusionMixin",
+        rbln_config: RBLNAutoencoderKLTemporalDecoderConfig,
+        return_vae_scale_factor: bool = False,
+    ) -> Tuple[int, int]:
+        sample_size = rbln_config.sample_size
+        if hasattr(pipe, "vae_scale_factor"):
+            vae_scale_factor = pipe.vae_scale_factor
+        else:
+            if hasattr(pipe.vae.config, "block_out_channels"):
+                vae_scale_factor = 2 ** (len(pipe.vae.config.block_out_channels) - 1)
+            else:
+                vae_scale_factor = 8  # vae image processor default value 8 (int)
+
+        if sample_size is None:
+            sample_size = pipe.unet.config.sample_size
+            if isinstance(sample_size, int):
+                sample_size = (sample_size, sample_size)
+            sample_size = (sample_size[0] * vae_scale_factor, sample_size[1] * vae_scale_factor)
+
+        if return_vae_scale_factor:
+            return sample_size, vae_scale_factor
+        else:
+            return sample_size
+
+    @classmethod
+    def update_rbln_config_using_pipe(
+        cls, pipe: "RBLNDiffusionMixin", rbln_config: "RBLNDiffusionMixinConfig", submodule_name: str
+    ) -> "RBLNDiffusionMixinConfig":
+        rbln_config.vae.sample_size, rbln_config.vae.vae_scale_factor = cls.get_vae_sample_size(
+            pipe, rbln_config.vae, return_vae_scale_factor=True
+        )
+
+        if rbln_config.vae.num_frames is None:
+            if hasattr(pipe.unet.config, "num_frames"):
+                rbln_config.vae.num_frames = pipe.unet.config.num_frames
+            else:
+                raise ValueError("num_frames should be specified in unet config.json")
+
+        if rbln_config.vae.decode_chunk_size is None:
+            rbln_config.vae.decode_chunk_size = rbln_config.vae.num_frames
+
+        def chunk_frame(num_frames, decode_chunk_size):
+            # get closest divisor to num_frames
+            divisors = [i for i in range(1, num_frames) if num_frames % i == 0]
+            closest = min(divisors, key=lambda x: abs(x - decode_chunk_size))
+            if decode_chunk_size != closest:
+                logger.warning(
+                    f"To ensure successful model compilation and prevent device OOM, {decode_chunk_size} is set to {closest}."
+                )
+            return closest
+
+        decode_chunk_size = chunk_frame(rbln_config.vae.num_frames, rbln_config.vae.decode_chunk_size)
+        rbln_config.vae.decode_chunk_size = decode_chunk_size
+        return rbln_config
+
+    @classmethod
+    def _update_rbln_config(
+        cls,
+        preprocessors: Union["AutoFeatureExtractor", "AutoProcessor", "AutoTokenizer"],
+        model: "PreTrainedModel",
+        model_config: "PretrainedConfig",
+        rbln_config: RBLNAutoencoderKLTemporalDecoderConfig,
+    ) -> RBLNAutoencoderKLTemporalDecoderConfig:
+        if rbln_config.sample_size is None:
+            rbln_config.sample_size = model_config.sample_size
+
+        if rbln_config.vae_scale_factor is None:
+            if hasattr(model_config, "block_out_channels"):
+                rbln_config.vae_scale_factor = 2 ** (len(model_config.block_out_channels) - 1)
+            else:
+                # vae image processor default value 8 (int)
+                rbln_config.vae_scale_factor = 8
+
+        compile_cfgs = []
+        if rbln_config.uses_encoder:
+            vae_enc_input_info = [
+                (
+                    "x",
+                    [
+                        rbln_config.batch_size,
+                        model_config.in_channels,
+                        rbln_config.sample_size[0],
+                        rbln_config.sample_size[1],
+                    ],
+                    "float32",
+                )
+            ]
+            compile_cfgs.append(RBLNCompileConfig(compiled_model_name="encoder", input_info=vae_enc_input_info))
+
+        decode_batch_size = rbln_config.batch_size * rbln_config.decode_chunk_size
+        vae_dec_input_info = [
+            (
+                "z",
+                [
+                    decode_batch_size,
+                    model_config.latent_channels,
+                    rbln_config.latent_sample_size[0],
+                    rbln_config.latent_sample_size[1],
+                ],
+                "float32",
+            )
+        ]
+        compile_cfgs.append(RBLNCompileConfig(compiled_model_name="decoder", input_info=vae_dec_input_info))
+
+        rbln_config.set_compile_cfgs(compile_cfgs)
+        return rbln_config
+
+    @classmethod
+    def _create_runtimes(
+        cls,
+        compiled_models: List[rebel.RBLNCompiledModel],
+        rbln_config: RBLNAutoencoderKLTemporalDecoderConfig,
+    ) -> List[rebel.Runtime]:
+        if len(compiled_models) == 1:
+            # decoder
+            expected_models = ["decoder"]
+        else:
+            expected_models = ["encoder", "decoder"]
+
+        if any(model_name not in rbln_config.device_map for model_name in expected_models):
+            cls._raise_missing_compiled_file_error(expected_models)
+
+        device_vals = [rbln_config.device_map[model_name] for model_name in expected_models]
+        return [
+            rebel.Runtime(
+                compiled_model,
+                tensor_type="pt",
+                device=device_val,
+                activate_profiler=rbln_config.activate_profiler,
+                timeout=rbln_config.timeout,
+            )
+            for compiled_model, device_val in zip(compiled_models, device_vals)
+        ]
+
+    def encode(
+        self, x: torch.FloatTensor, return_dict: bool = True
+    ) -> Union[AutoencoderKLOutput, Tuple[DiagonalGaussianDistribution]]:
+        """
+        Encode an input image into a latent representation.
+
+        Args:
+            x: The input image to encode.
+            return_dict:
+                Whether to return output as a dictionary. Defaults to True.
+
+        Returns:
+            The latent representation or AutoencoderKLOutput if return_dict=True
+        """
+        posterior = self.encoder.encode(x)
+
+        if not return_dict:
+            return (posterior,)
+
+        return AutoencoderKLOutput(latent_dist=posterior)
+
+    def decode(self, z: torch.FloatTensor, return_dict: bool = True) -> torch.FloatTensor:
+        """
+        Decode a latent representation into a video.
+
+        Args:
+            z: The latent representation to decode.
+            return_dict:
+                Whether to return output as a dictionary. Defaults to True.
+
+        Returns:
+            The decoded video or DecoderOutput if return_dict=True
+        """
+        decoded = self.decoder.decode(z)
+
+        if not return_dict:
+            return (decoded,)
+
+        return DecoderOutput(sample=decoded)

optimum/rbln/diffusers/models/autoencoders/vae.py

@@ -0,0 +1,178 @@
+# 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 TYPE_CHECKING, List, Union
+
+import torch
+from diffusers.models.autoencoders.vae import DiagonalGaussianDistribution, IdentityDistribution
+
+from ....utils.runtime_utils import RBLNPytorchRuntime
+
+
+if TYPE_CHECKING:
+    from diffusers import AutoencoderKL, AutoencoderKLCosmos, AutoencoderKLTemporalDecoder, VQModel
+
+
+class RBLNRuntimeVAEEncoder(RBLNPytorchRuntime):
+    def encode(self, x: torch.FloatTensor, **kwargs) -> torch.FloatTensor:
+        moments = self.forward(x.contiguous())
+        posterior = DiagonalGaussianDistribution(moments)
+        return posterior
+
+
+class RBLNRuntimeVAEDecoder(RBLNPytorchRuntime):
+    def decode(self, z: torch.FloatTensor, **kwargs) -> torch.FloatTensor:
+        return self.forward(z)
+
+
+class RBLNRuntimeCosmosVAEEncoder(RBLNPytorchRuntime):
+    def encode(self, x: torch.FloatTensor, **kwargs) -> torch.FloatTensor:
+        if self.use_slicing and x.shape[0] > 1:
+            encoded_slices = [self.forward(x_slice) for x_slice in x.split(1)]
+            h = torch.cat(encoded_slices)
+        else:
+            h = self.forward(x)
+        posterior = IdentityDistribution(h)
+        return posterior
+
+
+class RBLNRuntimeCosmosVAEDecoder(RBLNPytorchRuntime):
+    def decode(self, z: torch.FloatTensor, **kwargs) -> torch.FloatTensor:
+        if self.use_slicing and z.shape[0] > 1:
+            decoded_slices = [self.forward(z_slice) for z_slice in z.split(1)]
+            decoded = torch.cat(decoded_slices)
+        else:
+            decoded = self.forward(z)
+        return decoded
+
+
+class _VAEDecoder(torch.nn.Module):
+    def __init__(self, vae: "AutoencoderKL"):
+        super().__init__()
+        self.vae = vae
+
+    def forward(self, z):
+        vae_out = self.vae.decode(z, return_dict=False)
+        return vae_out
+
+
+class _VAETemporalDecoder(torch.nn.Module):
+    def __init__(self, vae: "AutoencoderKLTemporalDecoder"):
+        super().__init__()
+        self.vae = vae
+        self.num_frames = None
+
+    def forward(self, z):
+        vae_out = self.vae.decode(z, num_frames=self.num_frames, return_dict=False)
+        return vae_out
+
+
+class _VAEEncoder(torch.nn.Module):
+    def __init__(self, vae: Union["AutoencoderKL", "AutoencoderKLTemporalDecoder"]):
+        super().__init__()
+        self.vae = vae
+
+    def encode(self, x: torch.FloatTensor, return_dict: bool = True):
+        if hasattr(self, "use_tiling") and hasattr(self, "use_slicing"):
+            if self.use_tiling and (
+                x.shape[-1] > self.tile_sample_min_size or x.shape[-2] > self.tile_sample_min_size
+            ):
+                return self.tiled_encode(x, return_dict=return_dict)
+
+            if self.use_slicing and x.shape[0] > 1:
+                encoded_slices = [self.encoder(x_slice) for x_slice in x.split(1)]
+                h = torch.cat(encoded_slices)
+            else:
+                h = self.encoder(x)
+                if self.quant_conv is not None:
+                    h = self.quant_conv(h)
+
+        else:
+            h = self.encoder(x)
+            if self.quant_conv is not None:
+                h = self.quant_conv(h)
+        return h
+
+    def forward(self, x):
+        vae_out = _VAEEncoder.encode(self.vae, x, return_dict=False)
+        return vae_out
+
+
+class _VAECosmosEncoder(torch.nn.Module):
+    def __init__(self, vae: "AutoencoderKLCosmos"):
+        super().__init__()
+        self.vae = vae
+
+    def forward(self, x):
+        vae_out = self.vae._encode(x)
+        return vae_out
+
+
+class _VAECosmosDecoder(torch.nn.Module):
+    def __init__(self, vae: "AutoencoderKLCosmos"):
+        super().__init__()
+        self.vae = vae
+
+    def forward(self, z):
+        vae_out = self.vae._decode(z, return_dict=False)
+        return vae_out
+
+
+class RBLNRuntimeVQEncoder(RBLNPytorchRuntime):
+    def encode(self, x: torch.FloatTensor, **kwargs) -> torch.FloatTensor:
+        h = self.forward(x.contiguous())
+        return h
+
+
+class RBLNRuntimeVQDecoder(RBLNPytorchRuntime):
+    def decode(self, h: torch.Tensor, force_not_quantize: bool = False, shape=None, **kwargs) -> List[torch.Tensor]:
+        if not (force_not_quantize and not self.lookup_from_codebook):
+            raise ValueError(
+                "Currently, the `decode` method of the class `RBLNVQModel` is executed successfully only if `force_not_quantize` is True and `config.lookup_from_codebook` is False"
+            )
+        commit_loss = torch.zeros((h.shape[0])).to(h.device, dtype=h.dtype)
+        dec = self.forward(h.contiguous())
+        return dec, commit_loss
+
+
+class _VQEncoder(torch.nn.Module):
+    def __init__(self, vq_model: "VQModel"):
+        super().__init__()
+        self.vq_model = vq_model
+
+    def encode(self, x: torch.Tensor, return_dict: bool = True):
+        h = self.vq_model.encoder(x)
+        h = self.vq_model.quant_conv(h)
+        return h
+
+    def forward(self, x: torch.Tensor):
+        vq_out = self.encode(x)
+        return vq_out
+
+
+class _VQDecoder(torch.nn.Module):
+    def __init__(self, vq_model: "VQModel"):
+        super().__init__()
+        self.vq_model = vq_model
+
+    def decode(self, h: torch.Tensor, force_not_quantize: bool = False, return_dict: bool = True, shape=None):
+        quant = h
+        quant2 = self.vq_model.post_quant_conv(quant)
+        quant = quant if self.vq_model.config.norm_type == "spatial" else None
+        dec = self.vq_model.decoder(quant2, quant)
+        return dec
+
+    def forward(self, h: torch.Tensor):
+        vq_out = self.decode(h)
+        return vq_out

optimum/rbln/diffusers/models/autoencoders/vq_model.py

@@ -0,0 +1,211 @@
+# Copyright 2024 Rebellions Inc.
+
+# 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 TYPE_CHECKING, Any, List, Union
+
+import rebel
+import torch
+from diffusers import VQModel
+from diffusers.models.autoencoders.vae import DecoderOutput
+from diffusers.models.autoencoders.vq_model import VQEncoderOutput
+
+from ....configuration_utils import RBLNCompileConfig, RBLNModelConfig
+from ....modeling import RBLNModel
+from ....utils.logging import get_logger
+from ...configurations.models.configuration_vq_model import RBLNVQModelConfig
+from ...modeling_diffusers import RBLNDiffusionMixin, RBLNDiffusionMixinConfig
+from .vae import RBLNRuntimeVQDecoder, RBLNRuntimeVQEncoder, _VQDecoder, _VQEncoder
+
+
+if TYPE_CHECKING:
+    from transformers import AutoFeatureExtractor, AutoProcessor, AutoTokenizer, PretrainedConfig, PreTrainedModel
+
+logger = get_logger(__name__)
+
+
+class RBLNVQModel(RBLNModel):
+    """
+    RBLN implementation of VQModel for diffusion models.
+
+    This model is used to accelerate VQModel models from diffusers library on RBLN NPUs.
+    It can be configured to include both encoder and decoder, or just the decoder part for latent-to-image
+    conversion.
+
+    This class inherits from [`RBLNModel`]. Check the superclass documentation for the generic methods
+    the library implements for all its models.
+    """
+
+    auto_model_class = VQModel
+    config_name = "config.json"
+    hf_library_name = "diffusers"
+
+    def __post_init__(self, **kwargs):
+        super().__post_init__(**kwargs)
+
+        if self.rbln_config.uses_encoder:
+            self.encoder = RBLNRuntimeVQEncoder(runtime=self.model[0], main_input_name="x")
+        else:
+            self.encoder = None
+
+        self.decoder = RBLNRuntimeVQDecoder(runtime=self.model[-1], main_input_name="z")
+        self.decoder.lookup_from_codebook = self.config.lookup_from_codebook
+        self.image_size = self.rbln_config.image_size
+
+    @classmethod
+    def get_compiled_model(cls, model, rbln_config: RBLNModelConfig):
+        if rbln_config.uses_encoder:
+            expected_models = ["encoder", "decoder"]
+        else:
+            expected_models = ["decoder"]
+
+        compiled_models = {}
+        for i, model_name in enumerate(expected_models):
+            if model_name == "encoder":
+                wrapped_model = _VQEncoder(model)
+            else:
+                wrapped_model = _VQDecoder(model)
+
+            wrapped_model.eval()
+
+            compiled_models[model_name] = cls.compile(
+                wrapped_model,
+                rbln_compile_config=rbln_config.compile_cfgs[i],
+                create_runtimes=rbln_config.create_runtimes,
+                device=rbln_config.device_map[model_name],
+            )
+
+        return compiled_models
+
+    @classmethod
+    def update_rbln_config_using_pipe(
+        cls, pipe: RBLNDiffusionMixin, rbln_config: "RBLNDiffusionMixinConfig", submodule_name: str
+    ) -> "RBLNDiffusionMixinConfig":
+        return rbln_config
+
+    @classmethod
+    def _update_rbln_config(
+        cls,
+        preprocessors: Union["AutoFeatureExtractor", "AutoProcessor", "AutoTokenizer"],
+        model: "PreTrainedModel",
+        model_config: "PretrainedConfig",
+        rbln_config: RBLNVQModelConfig,
+    ) -> RBLNVQModelConfig:
+        if hasattr(model_config, "block_out_channels"):
+            rbln_config.vqmodel_scale_factor = 2 ** (len(model_config.block_out_channels) - 1)
+        else:
+            # image processor default value 8 (int)
+            rbln_config.vqmodel_scale_factor = 8
+
+        compile_cfgs = []
+        if rbln_config.uses_encoder:
+            enc_input_info = [
+                (
+                    "x",
+                    [
+                        rbln_config.batch_size,
+                        model_config.in_channels,
+                        rbln_config.sample_size[0],
+                        rbln_config.sample_size[1],
+                    ],
+                    "float32",
+                )
+            ]
+            enc_rbln_compile_config = RBLNCompileConfig(compiled_model_name="encoder", input_info=enc_input_info)
+            compile_cfgs.append(enc_rbln_compile_config)
+
+        dec_input_info = [
+            (
+                "h",
+                [
+                    rbln_config.batch_size,
+                    model_config.latent_channels,
+                    rbln_config.latent_sample_size[0],
+                    rbln_config.latent_sample_size[1],
+                ],
+                "float32",
+            )
+        ]
+        dec_rbln_compile_config = RBLNCompileConfig(compiled_model_name="decoder", input_info=dec_input_info)
+        compile_cfgs.append(dec_rbln_compile_config)
+
+        rbln_config.set_compile_cfgs(compile_cfgs)
+        return rbln_config
+
+    @classmethod
+    def _create_runtimes(
+        cls,
+        compiled_models: List[rebel.RBLNCompiledModel],
+        rbln_config: RBLNVQModelConfig,
+    ) -> List[rebel.Runtime]:
+        if len(compiled_models) == 1:
+            # decoder
+            expected_models = ["decoder"]
+        else:
+            # encoder, decoder
+            expected_models = ["encoder", "decoder"]
+
+        if any(model_name not in rbln_config.device_map for model_name in expected_models):
+            cls._raise_missing_compiled_file_error(expected_models)
+
+        device_vals = [rbln_config.device_map[model_name] for model_name in expected_models]
+        return [
+            rebel.Runtime(
+                compiled_model,
+                tensor_type="pt",
+                device=device_val,
+                activate_profiler=rbln_config.activate_profiler,
+                timeout=rbln_config.timeout,
+            )
+            for compiled_model, device_val in zip(compiled_models, device_vals)
+        ]
+
+    def encode(
+        self, x: torch.FloatTensor, return_dict: bool = True, **kwargs: Any
+    ) -> Union[torch.FloatTensor, VQEncoderOutput]:
+        """
+        Encode an input image into a quantized latent representation.
+
+        Args:
+            x: The input image to encode.
+            return_dict:
+                Whether to return output as a dictionary. Defaults to True.
+            kwargs: Additional arguments to pass to the encoder/quantizer.
+
+        Returns:
+            The quantized latent representation or a specific output object.
+        """
+        posterior = self.encoder.encode(x)
+        if not return_dict:
+            return (posterior,)
+        return VQEncoderOutput(latents=posterior)
+
+    def decode(
+        self, h: torch.FloatTensor, return_dict: bool = True, **kwargs: Any
+    ) -> Union[torch.FloatTensor, DecoderOutput]:
+        """
+        Decode a quantized latent representation back into an image.
+
+        Args:
+            h: The quantized latent representation to decode.
+            return_dict:
+                Whether to return output as a dictionary. Defaults to True.
+            kwargs: Additional arguments to pass to the decoder.
+
+        Returns:
+            The decoded image or a DecoderOutput object.
+        """
+        dec, commit_loss = self.decoder.decode(h, **kwargs)
+        if not return_dict:
+            return (dec, commit_loss)
+        return DecoderOutput(sample=dec, commit_loss=commit_loss)