pipeline_flux.py

# Copyright 2024 Black Forest Labs and The HuggingFace Team. 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
from typing import Any, Dict, List, Tuple, Callable, Optional, Union

import numpy as np
import torch
import torch.distributed
from diffusers import FluxPipeline
from diffusers.utils import is_torch_xla_available
from diffusers.pipelines.flux.pipeline_output import FluxPipelineOutput
from diffusers.pipelines.flux.pipeline_flux import retrieve_timesteps, calculate_shift

from xfuser.config import EngineConfig, InputConfig
from xfuser.core.distributed import (
    get_pipeline_parallel_world_size,
    get_runtime_state,
    get_pp_group,
    get_sequence_parallel_world_size,
    get_sequence_parallel_rank,
    get_sp_group,
    is_pipeline_first_stage,
    is_pipeline_last_stage,
    is_dp_last_group,
)
from .base_pipeline import xFuserPipelineBaseWrapper
from .register import xFuserPipelineWrapperRegister

if is_torch_xla_available():
    import torch_xla.core.xla_model as xm

    XLA_AVAILABLE = True
else:
    XLA_AVAILABLE = False


@xFuserPipelineWrapperRegister.register(FluxPipeline)
class xFuserFluxPipeline(xFuserPipelineBaseWrapper):

    @classmethod
    def from_pretrained(
        cls,
        pretrained_model_name_or_path: Optional[Union[str, os.PathLike]],
        engine_config: EngineConfig,
        **kwargs,
    ):
        pipeline = FluxPipeline.from_pretrained(pretrained_model_name_or_path, **kwargs)
        return cls(pipeline, engine_config)

    def prepare_run(
        self,
        input_config: InputConfig,
        steps: int = 3,
        sync_steps: int = 1,
    ):
        prompt = [""] * input_config.batch_size if input_config.batch_size > 1 else ""
        warmup_steps = get_runtime_state().runtime_config.warmup_steps
        get_runtime_state().runtime_config.warmup_steps = sync_steps
        self.__call__(
            height=input_config.height,
            width=input_config.width,
            prompt=prompt,
            num_inference_steps=steps,
            max_sequence_length=input_config.max_sequence_length,
            generator=torch.Generator(device="cuda").manual_seed(42),
            output_type=input_config.output_type,
        )
        get_runtime_state().runtime_config.warmup_steps = warmup_steps

    @property
    def guidance_scale(self):
        return self._guidance_scale

    @property
    def joint_attention_kwargs(self):
        return self._joint_attention_kwargs

    @property
    def num_timesteps(self):
        return self._num_timesteps

    @property
    def interrupt(self):
        return self._interrupt

    @torch.no_grad()
    @xFuserPipelineBaseWrapper.check_model_parallel_state(cfg_parallel_available=False)
    @xFuserPipelineBaseWrapper.enable_data_parallel
    @xFuserPipelineBaseWrapper.check_to_use_naive_forward
    def __call__(
        self,
        prompt: Union[str, List[str]] = None,
        prompt_2: Optional[Union[str, List[str]]] = None,
        height: Optional[int] = None,
        width: Optional[int] = None,
        num_inference_steps: int = 28,
        timesteps: List[int] = None,
        guidance_scale: float = 7.0,
        num_images_per_prompt: Optional[int] = 1,
        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
        latents: Optional[torch.FloatTensor] = None,
        prompt_embeds: Optional[torch.FloatTensor] = None,
        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
        output_type: Optional[str] = "pil",
        return_dict: bool = True,
        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
        max_sequence_length: int = 512,
        **kwargs,
    ):
        r"""
        Function invoked when calling the pipeline for generation.

        Args:
            prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
                instead.
            prompt_2 (`str` or `List[str]`, *optional*):
                The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
                will be used instead
            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                The height in pixels of the generated image. This is set to 1024 by default for the best results.
            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                The width in pixels of the generated image. This is set to 1024 by default for the best results.
            num_inference_steps (`int`, *optional*, defaults to 50):
                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
                expense of slower inference.
            timesteps (`List[int]`, *optional*):
                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
                passed will be used. Must be in descending order.
            guidance_scale (`float`, *optional*, defaults to 7.0):
                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
                `guidance_scale` is defined as `w` of equation 2. of [Imagen
                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
                usually at the expense of lower image quality.
            num_images_per_prompt (`int`, *optional*, defaults to 1):
                The number of images to generate per prompt.
            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
                to make generation deterministic.
            latents (`torch.FloatTensor`, *optional*):
                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
                tensor will ge generated by sampling using the supplied random `generator`.
            prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
                provided, text embeddings will be generated from `prompt` input argument.
            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
                If not provided, pooled text embeddings will be generated from `prompt` input argument.
            output_type (`str`, *optional*, defaults to `"pil"`):
                The output format of the generate image. Choose between
                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`~pipelines.flux.FluxPipelineOutput`] instead of a plain tuple.
            joint_attention_kwargs (`dict`, *optional*):
                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
                `self.processor` in
                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
            callback_on_step_end (`Callable`, *optional*):
                A function that calls at the end of each denoising steps during the inference. The function is called
                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
                `callback_on_step_end_tensor_inputs`.
            callback_on_step_end_tensor_inputs (`List`, *optional*):
                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
                `._callback_tensor_inputs` attribute of your pipeline class.
            max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`.

        Examples:

        Returns:
            [`~pipelines.flux.FluxPipelineOutput`] or `tuple`: [`~pipelines.flux.FluxPipelineOutput`] if `return_dict`
            is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated
            images.
        """

        height = height or self.default_sample_size * self.vae_scale_factor
        width = width or self.default_sample_size * self.vae_scale_factor

        # 1. Check inputs. Raise error if not correct
        self.check_inputs(
            prompt,
            prompt_2,
            height,
            width,
            prompt_embeds=prompt_embeds,
            pooled_prompt_embeds=pooled_prompt_embeds,
            callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
            max_sequence_length=max_sequence_length,
        )

        self._guidance_scale = guidance_scale
        self._joint_attention_kwargs = joint_attention_kwargs
        self._interrupt = False

        # 2. Define call parameters
        if prompt is not None and isinstance(prompt, str):
            batch_size = 1
        elif prompt is not None and isinstance(prompt, list):
            batch_size = len(prompt)
        else:
            batch_size = prompt_embeds.shape[0]
        device = self._execution_device

        #! ---------------------------------------- ADDED BELOW ----------------------------------------
        # * set runtime state input parameters
        get_runtime_state().set_input_parameters(
            height=height,
            width=width,
            batch_size=batch_size,
            num_inference_steps=num_inference_steps,
            max_condition_sequence_length=max_sequence_length,
            split_text_embed_in_sp=get_pipeline_parallel_world_size() == 1,
        )
        #! ---------------------------------------- ADDED ABOVE ----------------------------------------

        lora_scale = (
            self.joint_attention_kwargs.get("scale", None)
            if self.joint_attention_kwargs is not None
            else None
        )
        (
            prompt_embeds,
            pooled_prompt_embeds,
            text_ids,
        ) = self.encode_prompt(
            prompt=prompt,
            prompt_2=prompt_2,
            prompt_embeds=prompt_embeds,
            pooled_prompt_embeds=pooled_prompt_embeds,
            device=device,
            num_images_per_prompt=num_images_per_prompt,
            max_sequence_length=max_sequence_length,
            lora_scale=lora_scale,
        )

        # 4. Prepare latent variables
        num_channels_latents = self.transformer.config.in_channels // 4
        latents, latent_image_ids = self.prepare_latents(
            batch_size * num_images_per_prompt,
            num_channels_latents,
            height,
            width,
            prompt_embeds.dtype,
            device,
            generator,
            latents,
        )

        # 5. Prepare timesteps
        sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps)
        image_seq_len = latents.shape[1]
        mu = calculate_shift(
            image_seq_len,
            self.scheduler.config.base_image_seq_len,
            self.scheduler.config.max_image_seq_len,
            self.scheduler.config.base_shift,
            self.scheduler.config.max_shift,
        )
        timesteps, num_inference_steps = retrieve_timesteps(
            self.scheduler,
            num_inference_steps,
            device,
            timesteps,
            sigmas,
            mu=mu,
        )
        num_warmup_steps = max(
            len(timesteps) - num_inference_steps * self.scheduler.order, 0
        )
        self._num_timesteps = len(timesteps)

        # handle guidance
        if self.transformer.config.guidance_embeds:
            guidance = torch.full(
                [1], guidance_scale, device=device, dtype=torch.float32
            )
            guidance = guidance.expand(latents.shape[0])
        else:
            guidance = None

        num_pipeline_warmup_steps = get_runtime_state().runtime_config.warmup_steps
        # 6. Denoising loop
        with self.progress_bar(total=num_inference_steps) as progress_bar:
            if (
                get_pipeline_parallel_world_size() > 1
                and len(timesteps) > num_pipeline_warmup_steps
            ):
                # raise RuntimeError("Async pipeline not supported in flux")
                latents = self._sync_pipeline(
                    latents=latents,
                    prompt_embeds=prompt_embeds,
                    pooled_prompt_embeds=pooled_prompt_embeds,
                    text_ids=text_ids,
                    latent_image_ids=latent_image_ids,
                    guidance=guidance,
                    timesteps=timesteps[:num_pipeline_warmup_steps],
                    num_warmup_steps=num_warmup_steps,
                    progress_bar=progress_bar,
                    callback_on_step_end=callback_on_step_end,
                    callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
                )
                latents = self._async_pipeline(
                    latents=latents,
                    prompt_embeds=prompt_embeds,
                    pooled_prompt_embeds=pooled_prompt_embeds,
                    text_ids=text_ids,
                    latent_image_ids=latent_image_ids,
                    guidance=guidance,
                    timesteps=timesteps[num_pipeline_warmup_steps:],
                    num_warmup_steps=num_warmup_steps,
                    progress_bar=progress_bar,
                    callback_on_step_end=callback_on_step_end,
                    callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
                )
            else:
                latents = self._sync_pipeline(
                    latents=latents,
                    prompt_embeds=prompt_embeds,
                    pooled_prompt_embeds=pooled_prompt_embeds,
                    text_ids=text_ids,
                    latent_image_ids=latent_image_ids,
                    guidance=guidance,
                    timesteps=timesteps,
                    num_warmup_steps=num_warmup_steps,
                    progress_bar=progress_bar,
                    callback_on_step_end=callback_on_step_end,
                    callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
                    sync_only=True,
                )

        def vae_decode(latents):
            latents = self._unpack_latents(
                latents, height, width, self.vae_scale_factor
            )
            latents = (
                latents / self.vae.config.scaling_factor
            ) + self.vae.config.shift_factor

            image = self.vae.decode(latents, return_dict=False)[0]
            return image

        if not output_type == "latent":
            if get_runtime_state().runtime_config.use_parallel_vae:
                latents = self.gather_broadcast_latents(latents)
                image = vae_decode(latents)
            else:
                if is_dp_last_group():
                    image = vae_decode(latents)

        if self.is_dp_last_group():
            if output_type == "latent":
                image = latents

            else:
                image = self.image_processor.postprocess(image, output_type=output_type)

            # Offload all models
            self.maybe_free_model_hooks()

            if not return_dict:
                return (image,)

            return FluxPipelineOutput(images=image)
        else:
            return None

    def _init_sync_pipeline(
        self, latents: torch.Tensor, latent_image_ids: torch.Tensor, 
        prompt_embeds: torch.Tensor, text_ids: torch.Tensor
    ):
        get_runtime_state().set_patched_mode(patch_mode=False)

        latents_list = [
            latents[:, start_idx:end_idx, :]
            for start_idx, end_idx in get_runtime_state().pp_patches_token_start_end_idx_global
        ]
        latents = torch.cat(latents_list, dim=-2)
        latent_image_ids_list = [
            latent_image_ids[start_idx:end_idx]
            for start_idx, end_idx in get_runtime_state().pp_patches_token_start_end_idx_global
        ]
        latent_image_ids = torch.cat(latent_image_ids_list, dim=-2)

        if get_runtime_state().split_text_embed_in_sp:
            if prompt_embeds.shape[-2] % get_sequence_parallel_world_size() == 0:
                prompt_embeds = torch.chunk(prompt_embeds, get_sequence_parallel_world_size(), dim=-2)[get_sequence_parallel_rank()]
            else:
                get_runtime_state().split_text_embed_in_sp = False                

        if get_runtime_state().split_text_embed_in_sp:
            if text_ids.shape[-2] % get_sequence_parallel_world_size() == 0:
                text_ids = torch.chunk(text_ids, get_sequence_parallel_world_size(), dim=-2)[get_sequence_parallel_rank()]
            else:
                get_runtime_state().split_text_embed_in_sp = False                

        return latents, latent_image_ids, prompt_embeds, text_ids

    # synchronized compute the whole feature map in each pp stage
    def _sync_pipeline(
        self,
        latents: torch.Tensor,
        prompt_embeds: torch.Tensor,
        pooled_prompt_embeds: torch.Tensor,
        text_ids: torch.Tensor,
        latent_image_ids: torch.Tensor,
        guidance,
        timesteps: List[int],
        num_warmup_steps: int,
        progress_bar,
        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
        sync_only: bool = False,
    ):
        latents, latent_image_ids, prompt_embeds, text_ids = self._init_sync_pipeline(latents, latent_image_ids, prompt_embeds, text_ids)
        for i, t in enumerate(timesteps):
            if self.interrupt:
                continue
            if is_pipeline_last_stage():
                last_timestep_latents = latents

            # when there is only one pp stage, no need to recv
            if get_pipeline_parallel_world_size() == 1:
                pass
            # all ranks should recv the latent from the previous rank except
            #   the first rank in the first pipeline forward which should use
            #   the input latent
            elif is_pipeline_first_stage() and i == 0:
                pass
            else:
                latents = get_pp_group().pipeline_recv()
                if not is_pipeline_first_stage():
                    encoder_hidden_state = get_pp_group().pipeline_recv(
                        0, "encoder_hidden_state"
                    )

            # # handle guidance
            # if self.transformer.config.guidance_embeds:
            #     guidance = torch.tensor([guidance_scale], device=self._execution_device)
            #     guidance = guidance.expand(latents.shape[0])
            # else:
            #     guidance = None

            latents, encoder_hidden_state = self._backbone_forward(
                latents=latents,
                encoder_hidden_states=(
                    prompt_embeds if is_pipeline_first_stage() else encoder_hidden_state
                ),
                pooled_prompt_embeds=pooled_prompt_embeds,
                text_ids=text_ids,
                latent_image_ids=latent_image_ids,
                guidance=guidance,
                t=t,
            )

            if is_pipeline_last_stage():
                latents_dtype = latents.dtype
                latents = self._scheduler_step(latents, last_timestep_latents, t)

                if latents.dtype != latents_dtype:
                    if torch.backends.mps.is_available():
                        # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
                        latents = latents.to(latents_dtype)

                if callback_on_step_end is not None:
                    callback_kwargs = {}
                    for k in callback_on_step_end_tensor_inputs:
                        callback_kwargs[k] = locals()[k]
                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)

                    latents = callback_outputs.pop("latents", latents)
                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)

            if i == len(timesteps) - 1 or (
                (i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0
            ):
                progress_bar.update()

            if XLA_AVAILABLE:
                xm.mark_step()

            if sync_only and is_pipeline_last_stage() and i == len(timesteps) - 1:
                pass
            elif get_pipeline_parallel_world_size() > 1:
                get_pp_group().pipeline_send(latents)
                if not is_pipeline_last_stage():
                    get_pp_group().pipeline_send(
                        encoder_hidden_state, name="encoder_hidden_state"
                    )

        if (
            sync_only
            and get_sequence_parallel_world_size() > 1
            and is_pipeline_last_stage()
        ):
            sp_degree = get_sequence_parallel_world_size()
            sp_latents_list = get_sp_group().all_gather(latents, separate_tensors=True)
            latents_list = []
            for pp_patch_idx in range(get_runtime_state().num_pipeline_patch):
                latents_list += [
                    sp_latents_list[sp_patch_idx][
                        :,
                        get_runtime_state()
                        .pp_patches_token_start_idx_local[pp_patch_idx] : get_runtime_state()
                        .pp_patches_token_start_idx_local[pp_patch_idx + 1],
                        :,
                    ]
                    for sp_patch_idx in range(sp_degree)
                ]
            latents = torch.cat(latents_list, dim=-2)

        return latents

    def _async_pipeline(
        self,
        latents: torch.Tensor,
        prompt_embeds: torch.Tensor,
        pooled_prompt_embeds: torch.Tensor,
        text_ids: torch.Tensor,
        latent_image_ids: torch.Tensor,
        guidance,
        timesteps: List[int],
        num_warmup_steps: int,
        progress_bar,
        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
    ):
        if len(timesteps) == 0:
            return latents
        num_pipeline_patch = get_runtime_state().num_pipeline_patch
        num_pipeline_warmup_steps = get_runtime_state().runtime_config.warmup_steps
        patch_latents, patch_latent_image_ids = self._init_async_pipeline(
            num_timesteps=len(timesteps),
            latents=latents,
            num_pipeline_warmup_steps=num_pipeline_warmup_steps,
            latent_image_ids=latent_image_ids,
        )
        last_patch_latents = (
            [None for _ in range(num_pipeline_patch)]
            if (is_pipeline_last_stage())
            else None
        )

        first_async_recv = True
        for i, t in enumerate(timesteps):
            if self.interrupt:
                continue
            for patch_idx in range(num_pipeline_patch):
                if is_pipeline_last_stage():
                    last_patch_latents[patch_idx] = patch_latents[patch_idx]

                if is_pipeline_first_stage() and i == 0:
                    pass
                else:
                    if first_async_recv:
                        if not is_pipeline_first_stage() and patch_idx == 0:
                            get_pp_group().recv_next()
                        get_pp_group().recv_next()
                        first_async_recv = False

                    if not is_pipeline_first_stage() and patch_idx == 0:
                        last_encoder_hidden_states = (
                            get_pp_group().get_pipeline_recv_data(
                                idx=patch_idx, name="encoder_hidden_states"
                            )
                        )
                    patch_latents[patch_idx] = get_pp_group().get_pipeline_recv_data(
                        idx=patch_idx
                    )

                patch_latents[patch_idx], next_encoder_hidden_states = (
                    self._backbone_forward(
                        latents=patch_latents[patch_idx],
                        encoder_hidden_states=(
                            prompt_embeds
                            if is_pipeline_first_stage()
                            else last_encoder_hidden_states
                        ),
                        pooled_prompt_embeds=pooled_prompt_embeds,
                        text_ids=text_ids,
                        latent_image_ids=patch_latent_image_ids[patch_idx],
                        guidance=guidance,
                        t=t,
                    )
                )
                if is_pipeline_last_stage():
                    latents_dtype = patch_latents[patch_idx].dtype
                    patch_latents[patch_idx] = self._scheduler_step(
                        patch_latents[patch_idx],
                        last_patch_latents[patch_idx],
                        t,
                    )

                    if latents.dtype != latents_dtype:
                        if torch.backends.mps.is_available():
                            # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
                            latents = latents.to(latents_dtype)

                    if callback_on_step_end is not None:
                        callback_kwargs = {}
                        for k in callback_on_step_end_tensor_inputs:
                            callback_kwargs[k] = locals()[k]
                        callback_outputs = callback_on_step_end(
                            self, i, t, callback_kwargs
                        )

                        latents = callback_outputs.pop("latents", latents)
                        prompt_embeds = callback_outputs.pop(
                            "prompt_embeds", prompt_embeds
                        )
                        negative_prompt_embeds = callback_outputs.pop(
                            "negative_prompt_embeds", negative_prompt_embeds
                        )
                        negative_pooled_prompt_embeds = callback_outputs.pop(
                            "negative_pooled_prompt_embeds",
                            negative_pooled_prompt_embeds,
                        )

                    if i != len(timesteps) - 1:
                        get_pp_group().pipeline_isend(
                            patch_latents[patch_idx], segment_idx=patch_idx
                        )
                else:
                    if patch_idx == 0:
                        get_pp_group().pipeline_isend(
                            next_encoder_hidden_states, name="encoder_hidden_states"
                        )
                    get_pp_group().pipeline_isend(
                        patch_latents[patch_idx], segment_idx=patch_idx
                    )

                if is_pipeline_first_stage() and i == 0:
                    pass
                else:
                    if i == len(timesteps) - 1 and patch_idx == num_pipeline_patch - 1:
                        pass
                    elif is_pipeline_first_stage():
                        get_pp_group().recv_next()
                    else:
                        # recv encoder_hidden_state
                        if patch_idx == num_pipeline_patch - 1:
                            get_pp_group().recv_next()
                        # recv latents
                        get_pp_group().recv_next()

                get_runtime_state().next_patch()

            if i == len(timesteps) - 1 or (
                (i + num_pipeline_warmup_steps + 1) > num_warmup_steps
                and (i + num_pipeline_warmup_steps + 1) % self.scheduler.order == 0
            ):
                progress_bar.update()

            if XLA_AVAILABLE:
                xm.mark_step()

        latents = None
        if is_pipeline_last_stage():
            latents = torch.cat(patch_latents, dim=-2)
            if get_sequence_parallel_world_size() > 1:
                sp_degree = get_sequence_parallel_world_size()
                sp_latents_list = get_sp_group().all_gather(
                    latents, separate_tensors=True
                )
                latents_list = []
                for pp_patch_idx in range(get_runtime_state().num_pipeline_patch):
                    latents_list += [
                        sp_latents_list[sp_patch_idx][
                            ...,
                            get_runtime_state()
                            .pp_patches_token_start_idx_local[
                                pp_patch_idx
                            ] : get_runtime_state()
                            .pp_patches_token_start_idx_local[pp_patch_idx + 1],
                            :,
                        ]
                        for sp_patch_idx in range(sp_degree)
                    ]
                latents = torch.cat(latents_list, dim=-2)
        return latents

    def _init_async_pipeline(
        self,
        num_timesteps: int,
        latents: torch.Tensor,
        num_pipeline_warmup_steps: int,
        latent_image_ids: torch.Tensor,
    ):
        get_runtime_state().set_patched_mode(patch_mode=True)

        if is_pipeline_first_stage():
            # get latents computed in warmup stage
            # ignore latents after the last timestep
            latents = (
                get_pp_group().pipeline_recv()
                if num_pipeline_warmup_steps > 0
                else latents
            )
            patch_latents = list(
                latents.split(get_runtime_state().pp_patches_token_num, dim=-2)
            )
        elif is_pipeline_last_stage():
            patch_latents = list(
                latents.split(get_runtime_state().pp_patches_token_num, dim=-2)
            )
        else:
            patch_latents = [
                None for _ in range(get_runtime_state().num_pipeline_patch)
            ]

        patch_latent_image_ids = list(
            latent_image_ids[start_idx:end_idx]
            for start_idx, end_idx in get_runtime_state().pp_patches_token_start_end_idx_global
        )

        recv_timesteps = (
            num_timesteps - 1 if is_pipeline_first_stage() else num_timesteps
        )

        if is_pipeline_first_stage():
            for _ in range(recv_timesteps):
                for patch_idx in range(get_runtime_state().num_pipeline_patch):
                    get_pp_group().add_pipeline_recv_task(patch_idx)
        else:
            for _ in range(recv_timesteps):
                get_pp_group().add_pipeline_recv_task(0, "encoder_hidden_states")
                for patch_idx in range(get_runtime_state().num_pipeline_patch):
                    get_pp_group().add_pipeline_recv_task(patch_idx)

        return patch_latents, patch_latent_image_ids

    def _backbone_forward(
        self,
        latents: torch.Tensor,
        encoder_hidden_states: torch.Tensor,
        pooled_prompt_embeds: torch.Tensor,
        text_ids,
        latent_image_ids,
        guidance,
        t: Union[float, torch.Tensor],
    ):
        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
        timestep = t.expand(latents.shape[0]).to(latents.dtype)

        noise_pred, encoder_hidden_states = self.transformer(
            hidden_states=latents,
            timestep=timestep / 1000,
            guidance=guidance,
            pooled_projections=pooled_prompt_embeds,
            encoder_hidden_states=encoder_hidden_states,
            txt_ids=text_ids,
            img_ids=latent_image_ids,
            joint_attention_kwargs=self.joint_attention_kwargs,
            return_dict=False,
        )[0]

        return noise_pred, encoder_hidden_states

    def _scheduler_step(
        self,
        noise_pred: torch.Tensor,
        latents: torch.Tensor,
        t: Union[float, torch.Tensor],
    ):
        return self.scheduler.step(
            noise_pred,
            t,
            latents,
            return_dict=False,
        )[0]