optimum-rbln 0.1.12__py3-none-any.whl → 0.1.15__py3-none-any.whl

optimum/rbln/diffusers/models/{unet_2d_condition.py → unets/unet_2d_condition.py}

@@ -23,16 +23,15 @@
 
 import logging
 from dataclasses import dataclass
-from pathlib import Path
 from typing import TYPE_CHECKING, Any, Dict, Optional, Tuple, Union
 
 import torch
 from diffusers.models.unets.unet_2d_condition import UNet2DConditionModel
-from optimum.exporters import TasksManager
-from transformers import AutoConfig, AutoModel, PretrainedConfig
+from transformers import PretrainedConfig
 
-from ...modeling_base import RBLNModel
-from ...modeling_config import RBLNCompileConfig, RBLNConfig
+from ....modeling import RBLNModel
+from ....modeling_config import RBLNCompileConfig, RBLNConfig
+from ....modeling_diffusers import RBLNDiffusionMixin
 
 
 if TYPE_CHECKING:
@@ -126,6 +125,9 @@ class _UNet_SDXL(torch.nn.Module):
 
 
 class RBLNUNet2DConditionModel(RBLNModel):
+    hf_library_name = "diffusers"
+    auto_model_class = UNet2DConditionModel
+
     def __post_init__(self, **kwargs):
         super().__post_init__(**kwargs)
         self.in_features = self.rbln_config.model_cfg.get("in_features", None)
@@ -141,33 +143,6 @@ class RBLNUNet2DConditionModel(RBLNModel):
 
             self.add_embedding = ADDEMBEDDING(LINEAR1(self.in_features))
 
-    @classmethod
-    def from_pretrained(cls, *args, **kwargs):
-        def get_model_from_task(
-            task: str,
-            model_name_or_path: Union[str, Path],
-            **kwargs,
-        ):
-            return UNet2DConditionModel.from_pretrained(pretrained_model_name_or_path=model_name_or_path, **kwargs)
-
-        tasktmp = TasksManager.get_model_from_task
-        configtmp = AutoConfig.from_pretrained
-        modeltmp = AutoModel.from_pretrained
-        TasksManager.get_model_from_task = get_model_from_task
-        if kwargs.get("export", None):
-            # This is an ad-hoc to workaround save null values of the config.
-            # if export, pure optimum(not optimum-rbln) loads config using AutoConfig
-            # and diffusers model do not support loading by AutoConfig.
-            AutoConfig.from_pretrained = lambda *args, **kwargs: None
-        else:
-            AutoConfig.from_pretrained = UNet2DConditionModel.load_config
-        AutoModel.from_pretrained = UNet2DConditionModel.from_pretrained
-        rt = super().from_pretrained(*args, **kwargs)
-        AutoConfig.from_pretrained = configtmp
-        AutoModel.from_pretrained = modeltmp
-        TasksManager.get_model_from_task = tasktmp
-        return rt
-
     @classmethod
     def wrap_model_if_needed(cls, model: torch.nn.Module, rbln_config: RBLNConfig) -> torch.nn.Module:
         if model.config.addition_embed_type == "text_time":
@@ -175,6 +150,61 @@ class RBLNUNet2DConditionModel(RBLNModel):
         else:
             return _UNet_SD(model).eval()
 
+    @classmethod
+    def get_unet_sample_size(
+        cls, pipe: RBLNDiffusionMixin, rbln_config: Dict[str, Any]
+    ) -> Union[int, Tuple[int, int]]:
+        image_size = (rbln_config.get("img_height"), rbln_config.get("img_width"))
+        if (image_size[0] is None) != (image_size[1] is None):
+            raise ValueError("Both image height and image width must be given or not given")
+        elif image_size[0] is None and image_size[1] is None:
+            if rbln_config["img2img_pipeline"]:
+                # In case of img2img, sample size of unet is determined by vae encoder.
+                vae_sample_size = pipe.vae.config.sample_size
+                if isinstance(vae_sample_size, int):
+                    sample_size = vae_sample_size // pipe.vae_scale_factor
+                else:
+                    sample_size = (
+                        vae_sample_size[0] // pipe.vae_scale_factor,
+                        vae_sample_size[1] // pipe.vae_scale_factor,
+                    )
+            else:
+                sample_size = pipe.unet.config.sample_size
+        else:
+            sample_size = (image_size[0] // pipe.vae_scale_factor, image_size[1] // pipe.vae_scale_factor)
+
+        return sample_size
+
+    @classmethod
+    def update_rbln_config_using_pipe(cls, pipe: RBLNDiffusionMixin, rbln_config: Dict[str, Any]) -> Dict[str, Any]:
+        text_model_hidden_size = pipe.text_encoder_2.config.hidden_size if hasattr(pipe, "text_encoder_2") else None
+
+        batch_size = rbln_config.get("batch_size")
+        if not batch_size:
+            do_classifier_free_guidance = (
+                rbln_config.get("guidance_scale", 5.0) > 1.0 and pipe.unet.config.time_cond_proj_dim is None
+            )
+            batch_size = 2 if do_classifier_free_guidance else 1
+        else:
+            if rbln_config.get("guidance_scale"):
+                logger.warning(
+                    "guidance_scale is ignored because batch size is explicitly specified. "
+                    "To ensure consistent behavior, consider removing the guidance scale or "
+                    "adjusting the batch size configuration as needed."
+                )
+
+        rbln_config.update(
+            {
+                "max_seq_len": pipe.text_encoder.config.max_position_embeddings,
+                "text_model_hidden_size": text_model_hidden_size,
+                "sample_size": cls.get_unet_sample_size(pipe, rbln_config),
+                "batch_size": batch_size,
+                "is_controlnet": "controlnet" in pipe.config.keys(),
+            }
+        )
+
+        return rbln_config
+
     @classmethod
     def _get_rbln_config(
         cls,
@@ -182,137 +212,68 @@ class RBLNUNet2DConditionModel(RBLNModel):
         model_config: "PretrainedConfig",
         rbln_kwargs: Dict[str, Any] = {},
     ) -> RBLNConfig:
-        rbln_max_seq_len = rbln_kwargs.get("max_seq_len", None)
-        rbln_text_model_hidden_size = rbln_kwargs.get("text_model_hidden_size", None)
-        rbln_batch_size = rbln_kwargs.get("batch_size", None)
-        rbln_in_features = rbln_kwargs.get("in_features", None)
-        rbln_use_encode = rbln_kwargs.get("use_encode", None)
-        rbln_img_width = rbln_kwargs.get("img_width", None)
-        rbln_img_height = rbln_kwargs.get("img_height", None)
-        rbln_vae_scale_factor = rbln_kwargs.get("vae_scale_factor", None)
-        rbln_is_controlnet = rbln_kwargs.get("is_controlnet", None)
-
-        if rbln_max_seq_len is None:
-            rbln_max_seq_len = 77
-        if rbln_batch_size is None:
-            rbln_batch_size = 1
-
-        if rbln_use_encode:
-            if rbln_img_width is None or rbln_img_height is None or rbln_vae_scale_factor is None:
-                raise ValueError(
-                    "rbln_img_width, rbln_img_height, and rbln_vae_scale_factor must be provided when rbln_use_encode is True"
-                )
-            input_width = rbln_img_width // rbln_vae_scale_factor
-            input_height = rbln_img_height // rbln_vae_scale_factor
-        else:
-            input_width, input_height = model_config.sample_size, model_config.sample_size
+        batch_size = rbln_kwargs.get("batch_size")
+        max_seq_len = rbln_kwargs.get("max_seq_len")
+        sample_size = rbln_kwargs.get("sample_size")
+        is_controlnet = rbln_kwargs.get("is_controlnet")
+        rbln_in_features = None
+
+        if batch_size is None:
+            batch_size = 1
+
+        if sample_size is None:
+            sample_size = model_config.sample_size
+
+        if isinstance(sample_size, int):
+            sample_size = (sample_size, sample_size)
+
+        if max_seq_len is None:
+            raise ValueError("`rbln_max_seq_len` (ex. text_encoder's max_position_embeddings) must be specified.")
 
         input_info = [
-            (
-                "sample",
-                [
-                    rbln_batch_size,
-                    model_config.in_channels,
-                    input_height,
-                    input_width,
-                ],
-                "float32",
-            ),
+            ("sample", [batch_size, model_config.in_channels, sample_size[0], sample_size[1]], "float32"),
             ("timestep", [], "float32"),
-            (
-                "encoder_hidden_states",
-                [
-                    rbln_batch_size,
-                    rbln_max_seq_len,
-                    model_config.cross_attention_dim,
-                ],
-                "float32",
-            ),
+            ("encoder_hidden_states", [batch_size, max_seq_len, model_config.cross_attention_dim], "float32"),
         ]
 
-        if rbln_is_controlnet:
-            if len(model_config.block_out_channels) > 0:
-                input_info.extend(
-                    [
-                        (
-                            f"down_block_additional_residuals_{i}",
-                            [rbln_batch_size, model_config.block_out_channels[0], input_height, input_width],
-                            "float32",
-                        )
-                        for i in range(3)
-                    ]
-                )
-            if len(model_config.block_out_channels) > 1:
-                input_info.append(
-                    (
-                        "down_block_additional_residuals_3",
-                        [rbln_batch_size, model_config.block_out_channels[0], input_height // 2, input_width // 2],
-                        "float32",
-                    )
-                )
-                input_info.extend(
-                    [
-                        (
-                            f"down_block_additional_residuals_{i}",
-                            [rbln_batch_size, model_config.block_out_channels[1], input_height // 2, input_width // 2],
-                            "float32",
-                        )
-                        for i in range(4, 6)
-                    ]
-                )
-            if len(model_config.block_out_channels) > 2:
-                input_info.append(
-                    (
-                        f"down_block_additional_residuals_{6}",
-                        [rbln_batch_size, model_config.block_out_channels[1], input_height // 4, input_width // 4],
-                        "float32",
-                    )
-                )
-                input_info.extend(
-                    [
-                        (
-                            f"down_block_additional_residuals_{i}",
-                            [rbln_batch_size, model_config.block_out_channels[2], input_height // 4, input_width // 4],
-                            "float32",
-                        )
-                        for i in range(7, 9)
-                    ]
-                )
-            if len(model_config.block_out_channels) > 3:
-                input_info.extend(
-                    [
-                        (
-                            f"down_block_additional_residuals_{i}",
-                            [rbln_batch_size, model_config.block_out_channels[3], input_height // 8, input_width // 8],
-                            "float32",
-                        )
-                        for i in range(9, 12)
-                    ]
-                )
-            input_info.append(
-                (
-                    "mid_block_additional_residual",
-                    [
-                        rbln_batch_size,
-                        model_config.block_out_channels[-1],
-                        input_height // 2 ** (len(model_config.block_out_channels) - 1),
-                        input_width // 2 ** (len(model_config.block_out_channels) - 1),
-                    ],
-                    "float32",
-                )
-            )
+        if is_controlnet:
+            # down block addtional residuals
+            first_shape = [batch_size, model_config.block_out_channels[0], sample_size[0], sample_size[1]]
+            height, width = sample_size[0], sample_size[1]
+            input_info.append(("down_block_additional_residuals_0", first_shape, "float32"))
+            name_idx = 1
+            for idx, _ in enumerate(model_config.down_block_types):
+                shape = [batch_size, model_config.block_out_channels[idx], height, width]
+                for _ in range(model_config.layers_per_block):
+                    input_info.append((f"down_block_additional_residuals_{name_idx}", shape, "float32"))
+                    name_idx += 1
+                if idx != len(model_config.down_block_types) - 1:
+                    height = height // 2
+                    width = width // 2
+                    shape = [batch_size, model_config.block_out_channels[idx], height, width]
+                    input_info.append((f"down_block_additional_residuals_{name_idx}", shape, "float32"))
+                    name_idx += 1
+
+            # mid block addtional residual
+            num_cross_attn_blocks = model_config.down_block_types.count("CrossAttnDownBlock2D")
+            out_channels = model_config.block_out_channels[-1]
+            shape = [
+                batch_size,
+                out_channels,
+                sample_size[0] // 2**num_cross_attn_blocks,
+                sample_size[1] // 2**num_cross_attn_blocks,
+            ]
+            input_info.append(("mid_block_additional_residual", shape, "float32"))
 
         rbln_compile_config = RBLNCompileConfig(input_info=input_info)
 
         if hasattr(model_config, "addition_embed_type") and model_config.addition_embed_type == "text_time":
-            if rbln_text_model_hidden_size is None:
-                rbln_text_model_hidden_size = 768
-            if rbln_in_features is None:
-                rbln_in_features = model_config.projection_class_embeddings_input_dim
+            rbln_text_model_hidden_size = rbln_kwargs["text_model_hidden_size"]
+            rbln_in_features = model_config.projection_class_embeddings_input_dim
             rbln_compile_config.input_info.append(
-                ("text_embeds", [rbln_batch_size, rbln_text_model_hidden_size], "float32")
+                ("text_embeds", [batch_size, rbln_text_model_hidden_size], "float32")
             )
-            rbln_compile_config.input_info.append(("time_ids", [rbln_batch_size, 6], "float32"))
+            rbln_compile_config.input_info.append(("time_ids", [batch_size, 6], "float32"))
 
         rbln_config = RBLNConfig(
             rbln_cls=cls.__name__,
@@ -320,19 +281,15 @@ class RBLNUNet2DConditionModel(RBLNModel):
             rbln_kwargs=rbln_kwargs,
         )
 
-        rbln_config.model_cfg.update(
-            {
-                "max_seq_len": rbln_max_seq_len,
-                "batch_size": rbln_batch_size,
-                "use_encode": rbln_use_encode,
-            }
-        )
-
         if rbln_in_features is not None:
             rbln_config.model_cfg["in_features"] = rbln_in_features
 
         return rbln_config
 
+    @property
+    def compiled_batch_size(self):
+        return self.rbln_config.compile_cfgs[0].input_info[0][1][0]
+
     def forward(
         self,
         sample: torch.Tensor,
@@ -350,9 +307,18 @@ class RBLNUNet2DConditionModel(RBLNModel):
         return_dict: bool = True,
         **kwargs,
     ):
-        """
-        arg order : latent_model_input, t, prompt_embeds
-        """
+        sample_batch_size = sample.size()[0]
+        compiled_batch_size = self.compiled_batch_size
+        if sample_batch_size != compiled_batch_size and (
+            sample_batch_size * 2 == compiled_batch_size or sample_batch_size == compiled_batch_size * 2
+        ):
+            raise ValueError(
+                f"Mismatch between UNet's runtime batch size ({sample_batch_size}) and compiled batch size ({compiled_batch_size}). "
+                "This may be caused by the 'guidance scale' parameter, which doubles the runtime batch size in Stable Diffusion. "
+                "Adjust the batch size during compilation or modify the 'guidance scale' to match the compiled batch size.\n\n"
+                "For details, see: https://docs.rbln.ai/software/optimum/model_api.html#stable-diffusion"
+            )
+
         added_cond_kwargs = {} if added_cond_kwargs is None else added_cond_kwargs
 
         if down_block_additional_residuals is not None:

optimum/rbln/diffusers/pipelines/__init__.py

@@ -20,16 +20,64 @@
 # are the intellectual property of Rebellions Inc. and may not be
 # copied, modified, or distributed without prior written permission
 # from Rebellions Inc.
+from typing import TYPE_CHECKING
 
-from .controlnet import (
-    RBLNMultiControlNetModel,
-    RBLNStableDiffusionControlNetImg2ImgPipeline,
-    RBLNStableDiffusionControlNetPipeline,
-    RBLNStableDiffusionXLControlNetImg2ImgPipeline,
-    RBLNStableDiffusionXLControlNetPipeline,
-)
-from .stable_diffusion import (
-    RBLNStableDiffusionImg2ImgPipeline,
-    RBLNStableDiffusionPipeline,
-)
-from .stable_diffusion_xl import RBLNStableDiffusionXLImg2ImgPipeline, RBLNStableDiffusionXLPipeline
+from transformers.utils import _LazyModule
+
+
+_import_structure = {
+    "controlnet": [
+        "RBLNMultiControlNetModel",
+        "RBLNStableDiffusionControlNetImg2ImgPipeline",
+        "RBLNStableDiffusionControlNetPipeline",
+        "RBLNStableDiffusionXLControlNetImg2ImgPipeline",
+        "RBLNStableDiffusionXLControlNetPipeline",
+    ],
+    "stable_diffusion": [
+        "RBLNStableDiffusionImg2ImgPipeline",
+        "RBLNStableDiffusionPipeline",
+        "RBLNStableDiffusionInpaintPipeline",
+    ],
+    "stable_diffusion_xl": [
+        "RBLNStableDiffusionXLImg2ImgPipeline",
+        "RBLNStableDiffusionXLPipeline",
+        "RBLNStableDiffusionXLInpaintPipeline",
+    ],
+    "stable_diffusion_3": [
+        "RBLNStableDiffusion3Pipeline",
+        "RBLNStableDiffusion3Img2ImgPipeline",
+        "RBLNStableDiffusion3InpaintPipeline",
+    ],
+}
+if TYPE_CHECKING:
+    from .controlnet import (
+        RBLNMultiControlNetModel,
+        RBLNStableDiffusionControlNetImg2ImgPipeline,
+        RBLNStableDiffusionControlNetPipeline,
+        RBLNStableDiffusionXLControlNetImg2ImgPipeline,
+        RBLNStableDiffusionXLControlNetPipeline,
+    )
+    from .stable_diffusion import (
+        RBLNStableDiffusionImg2ImgPipeline,
+        RBLNStableDiffusionInpaintPipeline,
+        RBLNStableDiffusionPipeline,
+    )
+    from .stable_diffusion_3 import (
+        RBLNStableDiffusion3Img2ImgPipeline,
+        RBLNStableDiffusion3InpaintPipeline,
+        RBLNStableDiffusion3Pipeline,
+    )
+    from .stable_diffusion_xl import (
+        RBLNStableDiffusionXLImg2ImgPipeline,
+        RBLNStableDiffusionXLInpaintPipeline,
+        RBLNStableDiffusionXLPipeline,
+    )
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(
+        __name__,
+        globals()["__file__"],
+        _import_structure,
+        module_spec=__spec__,
+    )

optimum/rbln/diffusers/pipelines/controlnet/multicontrolnet.py

@@ -27,12 +27,9 @@ from pathlib import Path
 from typing import TYPE_CHECKING, Any, Dict, List, Optional, Union
 
 import torch
-from diffusers import ControlNetModel
 from diffusers.pipelines.controlnet.multicontrolnet import MultiControlNetModel
-from optimum.exporters import TasksManager
-from transformers import AutoConfig, AutoModel
 
-from ....modeling_base import RBLNModel
+from ....modeling import RBLNModel
 from ....modeling_config import RBLNConfig
 from ...models.controlnet import RBLNControlNetModel
 
@@ -44,6 +41,9 @@ logger = logging.getLogger(__name__)
 
 
 class RBLNMultiControlNetModel(RBLNModel):
+    hf_library_name = "diffusers"
+    _hf_class = MultiControlNetModel
+
     def __init__(
         self,
         models: List[RBLNControlNetModel],
@@ -52,26 +52,12 @@ class RBLNMultiControlNetModel(RBLNModel):
         self.nets = models
         self.dtype = torch.float32
 
-    @classmethod
-    def from_pretrained(cls, *args, **kwargs):
-        def get_model_from_task(
-            task: str,
-            model_name_or_path: Union[str, Path],
-            **kwargs,
-        ):
-            return MultiControlNetModel.from_pretrained(pretrained_model_name_or_path=model_name_or_path, **kwargs)
-
-        tasktmp = TasksManager.get_model_from_task
-        configtmp = AutoConfig.from_pretrained
-        modeltmp = AutoModel.from_pretrained
-        TasksManager.get_model_from_task = get_model_from_task
-        AutoConfig.from_pretrained = ControlNetModel.load_config
-        AutoModel.from_pretrained = MultiControlNetModel.from_pretrained
-        rt = super().from_pretrained(*args, **kwargs)
-        AutoConfig.from_pretrained = configtmp
-        AutoModel.from_pretrained = modeltmp
-        TasksManager.get_model_from_task = tasktmp
-        return rt
+    @property
+    def compiled_models(self):
+        cm = []
+        for net in self.nets:
+            cm.extend(net.compiled_models)
+        return cm
 
     @classmethod
     def _from_pretrained(
@@ -111,7 +97,7 @@ class RBLNMultiControlNetModel(RBLNModel):
         sample: torch.FloatTensor,
         timestep: Union[torch.Tensor, float, int],
         encoder_hidden_states: torch.Tensor,
-        controlnet_cond: List[torch.tensor],
+        controlnet_cond: List[torch.Tensor],
         conditioning_scale: List[float],
         class_labels: Optional[torch.Tensor] = None,
         timestep_cond: Optional[torch.Tensor] = None,