777 lines
31 KiB
Python
777 lines
31 KiB
Python
# SPDX-License-Identifier: Apache-2.0
|
|
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
|
|
|
import inspect
|
|
import json
|
|
import logging
|
|
import math
|
|
import os
|
|
from typing import Any, Optional, Union
|
|
|
|
import numpy as np
|
|
import torch
|
|
from diffusers.image_processor import VaeImageProcessor
|
|
from diffusers.models.autoencoders.autoencoder_kl_qwenimage import (
|
|
AutoencoderKLQwenImage,
|
|
)
|
|
from diffusers.schedulers.scheduling_flow_match_euler_discrete import (
|
|
FlowMatchEulerDiscreteScheduler,
|
|
)
|
|
from diffusers.utils.torch_utils import randn_tensor
|
|
from torch import nn
|
|
from transformers import Qwen2_5_VLForConditionalGeneration, Qwen2Tokenizer
|
|
|
|
from vllm_omni.diffusion.data import DiffusionOutput, OmniDiffusionConfig
|
|
from vllm_omni.diffusion.distributed.utils import get_local_device
|
|
from vllm_omni.diffusion.models.qwen_image.qwen_image_transformer import (
|
|
QwenImageTransformer2DModel,
|
|
)
|
|
from vllm_omni.diffusion.request import OmniDiffusionRequest
|
|
from vllm_omni.model_executor.model_loader.weight_utils import (
|
|
download_weights_from_hf_specific,
|
|
)
|
|
|
|
logger = logging.getLogger(__name__)
|
|
|
|
|
|
def get_qwen_image_post_process_func(
|
|
od_config: OmniDiffusionConfig,
|
|
):
|
|
model_name = od_config.model
|
|
model_path = download_weights_from_hf_specific(model_name, None, ["*"])
|
|
vae_config_path = os.path.join(model_path, "vae/config.json")
|
|
with open(vae_config_path) as f:
|
|
vae_config = json.load(f)
|
|
vae_scale_factor = 2 ** len(vae_config["temporal_downsample"]) if "temporal_downsample" in vae_config else 8
|
|
|
|
image_processor = VaeImageProcessor(vae_scale_factor=vae_scale_factor * 2)
|
|
|
|
def post_process_func(
|
|
images: torch.Tensor,
|
|
):
|
|
return image_processor.postprocess(images)
|
|
|
|
return post_process_func
|
|
|
|
|
|
def calculate_shift(
|
|
image_seq_len,
|
|
base_seq_len: int = 256,
|
|
max_seq_len: int = 4096,
|
|
base_shift: float = 0.5,
|
|
max_shift: float = 1.15,
|
|
):
|
|
m = (max_shift - base_shift) / (max_seq_len - base_seq_len)
|
|
b = base_shift - m * base_seq_len
|
|
mu = image_seq_len * m + b
|
|
return mu
|
|
|
|
|
|
def retrieve_timesteps(
|
|
scheduler,
|
|
num_inference_steps: Optional[int] = None,
|
|
device: Optional[Union[str, torch.device]] = None,
|
|
timesteps: Optional[list[int]] = None,
|
|
sigmas: Optional[list[float]] = None,
|
|
**kwargs,
|
|
) -> tuple[torch.Tensor, int]:
|
|
r"""
|
|
Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
|
|
custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
|
|
|
|
Args:
|
|
scheduler (`SchedulerMixin`):
|
|
The scheduler to get timesteps from.
|
|
num_inference_steps (`int`):
|
|
The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
|
|
must be `None`.
|
|
device (`str` or `torch.device`, *optional*):
|
|
The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
|
|
timesteps (`list[int]`, *optional*):
|
|
Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
|
|
`num_inference_steps` and `sigmas` must be `None`.
|
|
sigmas (`list[float]`, *optional*):
|
|
Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
|
|
`num_inference_steps` and `timesteps` must be `None`.
|
|
|
|
Returns:
|
|
`tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
|
|
second element is the number of inference steps.
|
|
"""
|
|
if timesteps is not None and sigmas is not None:
|
|
raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
|
|
if timesteps is not None:
|
|
accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
|
|
if not accepts_timesteps:
|
|
raise ValueError(
|
|
f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
|
|
f" timestep schedules. Please check whether you are using the correct scheduler."
|
|
)
|
|
scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
|
|
timesteps = scheduler.timesteps
|
|
num_inference_steps = len(timesteps)
|
|
elif sigmas is not None:
|
|
accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
|
|
if not accept_sigmas:
|
|
raise ValueError(
|
|
f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
|
|
f" sigmas schedules. Please check whether you are using the correct scheduler."
|
|
)
|
|
scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
|
|
timesteps = scheduler.timesteps
|
|
num_inference_steps = len(timesteps)
|
|
else:
|
|
scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
|
|
timesteps = scheduler.timesteps
|
|
return timesteps, num_inference_steps
|
|
|
|
|
|
def get_timestep_embedding(
|
|
timesteps: torch.Tensor,
|
|
embedding_dim: int,
|
|
flip_sin_to_cos: bool = False,
|
|
downscale_freq_shift: float = 1,
|
|
scale: float = 1,
|
|
max_period: int = 10000,
|
|
) -> torch.Tensor:
|
|
"""
|
|
This matches the implementation in Denoising Diffusion Probabilistic Models: Create sinusoidal timestep embeddings.
|
|
|
|
Args
|
|
timesteps (torch.Tensor):
|
|
a 1-D Tensor of N indices, one per batch element. These may be fractional.
|
|
embedding_dim (int):
|
|
the dimension of the output.
|
|
flip_sin_to_cos (bool):
|
|
Whether the embedding order should be `cos, sin` (if True) or `sin, cos` (if False)
|
|
downscale_freq_shift (float):
|
|
Controls the delta between frequencies between dimensions
|
|
scale (float):
|
|
Scaling factor applied to the embeddings.
|
|
max_period (int):
|
|
Controls the maximum frequency of the embeddings
|
|
Returns
|
|
torch.Tensor: an [N x dim] Tensor of positional embeddings.
|
|
"""
|
|
assert len(timesteps.shape) == 1, "Timesteps should be a 1d-array"
|
|
|
|
half_dim = embedding_dim // 2
|
|
exponent = -math.log(max_period) * torch.arange(start=0, end=half_dim, dtype=torch.float32, device=timesteps.device)
|
|
exponent = exponent / (half_dim - downscale_freq_shift)
|
|
|
|
emb = torch.exp(exponent).to(timesteps.dtype)
|
|
emb = timesteps[:, None].float() * emb[None, :]
|
|
|
|
# scale embeddings
|
|
emb = scale * emb
|
|
|
|
# concat sine and cosine embeddings
|
|
emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=-1)
|
|
|
|
# flip sine and cosine embeddings
|
|
if flip_sin_to_cos:
|
|
emb = torch.cat([emb[:, half_dim:], emb[:, :half_dim]], dim=-1)
|
|
|
|
# zero pad
|
|
if embedding_dim % 2 == 1:
|
|
emb = torch.nn.functional.pad(emb, (0, 1, 0, 0))
|
|
return emb
|
|
|
|
|
|
def apply_rotary_emb_qwen(
|
|
x: torch.Tensor,
|
|
freqs_cis: Union[torch.Tensor, tuple[torch.Tensor]],
|
|
use_real: bool = True,
|
|
use_real_unbind_dim: int = -1,
|
|
) -> tuple[torch.Tensor, torch.Tensor]:
|
|
"""
|
|
Apply rotary embeddings to input tensors using the given frequency tensor. This function applies rotary embeddings
|
|
to the given query or key 'x' tensors using the provided frequency tensor 'freqs_cis'. The input tensors are
|
|
reshaped as complex numbers, and the frequency tensor is reshaped for broadcasting compatibility. The resulting
|
|
tensors contain rotary embeddings and are returned as real tensors.
|
|
|
|
Args:
|
|
x (`torch.Tensor`):
|
|
Query or key tensor to apply rotary embeddings. [B, S, H, D] xk (torch.Tensor): Key tensor to apply
|
|
freqs_cis (`tuple[torch.Tensor]`): Precomputed frequency tensor for complex exponentials. ([S, D], [S, D],)
|
|
|
|
Returns:
|
|
tuple[torch.Tensor, torch.Tensor]: tuple of modified query tensor and key tensor with rotary embeddings.
|
|
"""
|
|
if use_real:
|
|
cos, sin = freqs_cis # [S, D]
|
|
cos = cos[None, None]
|
|
sin = sin[None, None]
|
|
cos, sin = cos.to(x.device), sin.to(x.device)
|
|
|
|
if use_real_unbind_dim == -1:
|
|
# Used for flux, cogvideox, hunyuan-dit
|
|
x_real, x_imag = x.reshape(*x.shape[:-1], -1, 2).unbind(-1) # [B, S, H, D//2]
|
|
x_rotated = torch.stack([-x_imag, x_real], dim=-1).flatten(3)
|
|
elif use_real_unbind_dim == -2:
|
|
# Used for Stable Audio, OmniGen, CogView4 and Cosmos
|
|
x_real, x_imag = x.reshape(*x.shape[:-1], 2, -1).unbind(-2) # [B, S, H, D//2]
|
|
x_rotated = torch.cat([-x_imag, x_real], dim=-1)
|
|
else:
|
|
raise ValueError(f"`use_real_unbind_dim={use_real_unbind_dim}` but should be -1 or -2.")
|
|
|
|
out = (x.float() * cos + x_rotated.float() * sin).to(x.dtype)
|
|
|
|
return out
|
|
else:
|
|
x_rotated = torch.view_as_complex(x.float().reshape(*x.shape[:-1], -1, 2))
|
|
freqs_cis = freqs_cis.unsqueeze(1)
|
|
x_out = torch.view_as_real(x_rotated * freqs_cis).flatten(3)
|
|
|
|
return x_out.type_as(x)
|
|
|
|
|
|
class QwenImagePipeline(
|
|
nn.Module,
|
|
):
|
|
def __init__(
|
|
self,
|
|
*,
|
|
od_config: OmniDiffusionConfig = None,
|
|
prefix: str = "",
|
|
):
|
|
super().__init__()
|
|
self.device = get_local_device()
|
|
# self.hf_config = vllm_config.model_config.hf_config
|
|
# TODO support user custom folders
|
|
self.scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained("Qwen/Qwen-Image", subfolder="scheduler")
|
|
logger.info("Loaded Qwen-Image scheduler successfully")
|
|
from vllm_omni.diffusion.utils.torch_utils import (
|
|
set_default_torch_dtype,
|
|
) # to avoid ci error
|
|
|
|
with set_default_torch_dtype(torch.bfloat16):
|
|
self.text_encoder = Qwen2_5_VLForConditionalGeneration.from_pretrained(
|
|
"Qwen/Qwen-Image", subfolder="text_encoder"
|
|
)
|
|
logger.info("Loaded Qwen-Image text encoder successfully")
|
|
self.vae = AutoencoderKLQwenImage.from_pretrained("Qwen/Qwen-Image", subfolder="vae").to(self.device)
|
|
logger.info("Loaded Qwen-Image VAE successfully")
|
|
self.transformer = QwenImageTransformer2DModel()
|
|
logger.info("Initialized Qwen-Image transformer successfully.")
|
|
self.tokenizer = Qwen2Tokenizer.from_pretrained("Qwen/Qwen-Image", subfolder="tokenizer")
|
|
logger.info("Loaded Qwen-Image tokenizer successfully.")
|
|
|
|
self.stage = None
|
|
|
|
self.vae_scale_factor = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8
|
|
# QwenImage latents are turned into 2x2 patches and packed.
|
|
# This means the latent width and height has to be divisible
|
|
# by the patch size. So the vae scale factor is multiplied by the patch size to account for this
|
|
# self.image_processor = VaeImageProcessor(
|
|
# vae_scale_factor=self.vae_scale_factor * 2
|
|
# )
|
|
self.tokenizer_max_length = 1024
|
|
self.prompt_template_encode = "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n" # noqa: E501
|
|
self.prompt_template_encode_start_idx = 34
|
|
self.default_sample_size = 128
|
|
|
|
def check_inputs(
|
|
self,
|
|
prompt,
|
|
height,
|
|
width,
|
|
negative_prompt=None,
|
|
prompt_embeds=None,
|
|
negative_prompt_embeds=None,
|
|
prompt_embeds_mask=None,
|
|
negative_prompt_embeds_mask=None,
|
|
callback_on_step_end_tensor_inputs=None,
|
|
max_sequence_length=None,
|
|
):
|
|
if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0:
|
|
logger.warning(
|
|
f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} "
|
|
"but are {height} and {width}. Dimensions will be resized accordingly"
|
|
)
|
|
|
|
# if callback_on_step_end_tensor_inputs is not None and not all(
|
|
# k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
|
|
# ):
|
|
# raise ValueError(
|
|
# f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs},
|
|
# but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
|
|
# )
|
|
|
|
if prompt is not None and prompt_embeds is not None:
|
|
raise ValueError(
|
|
f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
|
|
" only forward one of the two."
|
|
)
|
|
elif prompt is None and prompt_embeds is None:
|
|
raise ValueError(
|
|
"Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
|
|
)
|
|
elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
|
|
raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
|
|
|
|
if negative_prompt is not None and negative_prompt_embeds is not None:
|
|
raise ValueError(
|
|
f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
|
|
f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
|
|
)
|
|
|
|
if prompt_embeds is not None and prompt_embeds_mask is None:
|
|
raise ValueError(
|
|
"If `prompt_embeds` are provided, `prompt_embeds_mask` also have to be passed. "
|
|
"Make sure to generate `prompt_embeds_mask` from the same text encoder "
|
|
"that was used to generate `prompt_embeds`."
|
|
)
|
|
if negative_prompt_embeds is not None and negative_prompt_embeds_mask is None:
|
|
raise ValueError(
|
|
"If `negative_prompt_embeds` are provided, `negative_prompt_embeds_mask` also have to be passed. "
|
|
"Make sure to generate `negative_prompt_embeds_mask` from the same text encoder "
|
|
"that was used to generate `negative_prompt_embeds`."
|
|
)
|
|
|
|
if max_sequence_length is not None and max_sequence_length > 1024:
|
|
raise ValueError(f"`max_sequence_length` cannot be greater than 1024 but is {max_sequence_length}")
|
|
|
|
def _extract_masked_hidden(self, hidden_states: torch.Tensor, mask: torch.Tensor):
|
|
bool_mask = mask.bool()
|
|
valid_lengths = bool_mask.sum(dim=1)
|
|
selected = hidden_states[bool_mask]
|
|
split_result = torch.split(selected, valid_lengths.tolist(), dim=0)
|
|
|
|
return split_result
|
|
|
|
def _get_qwen_prompt_embeds(
|
|
self,
|
|
prompt: Union[str, list[str]] = None,
|
|
dtype: Optional[torch.dtype] = None,
|
|
):
|
|
dtype = dtype or self.text_encoder.dtype
|
|
|
|
prompt = [prompt] if isinstance(prompt, str) else prompt
|
|
|
|
template = self.prompt_template_encode
|
|
drop_idx = self.prompt_template_encode_start_idx
|
|
txt = [template.format(e) for e in prompt]
|
|
txt_tokens = self.tokenizer(
|
|
txt,
|
|
max_length=self.tokenizer_max_length + drop_idx,
|
|
padding=True,
|
|
truncation=True,
|
|
return_tensors="pt",
|
|
).to(self.device)
|
|
# print(f"attention mask: {txt_tokens.attention_mask}")
|
|
encoder_hidden_states = self.text_encoder(
|
|
input_ids=txt_tokens.input_ids,
|
|
attention_mask=txt_tokens.attention_mask,
|
|
output_hidden_states=True,
|
|
)
|
|
hidden_states = encoder_hidden_states.hidden_states[-1]
|
|
split_hidden_states = self._extract_masked_hidden(hidden_states, txt_tokens.attention_mask)
|
|
split_hidden_states = [e[drop_idx:] for e in split_hidden_states]
|
|
attn_mask_list = [torch.ones(e.size(0), dtype=torch.long, device=e.device) for e in split_hidden_states]
|
|
max_seq_len = max([e.size(0) for e in split_hidden_states])
|
|
prompt_embeds = torch.stack(
|
|
[torch.cat([u, u.new_zeros(max_seq_len - u.size(0), u.size(1))]) for u in split_hidden_states]
|
|
)
|
|
encoder_attention_mask = torch.stack(
|
|
[torch.cat([u, u.new_zeros(max_seq_len - u.size(0))]) for u in attn_mask_list]
|
|
)
|
|
|
|
prompt_embeds = prompt_embeds.to(dtype=dtype)
|
|
|
|
return prompt_embeds, encoder_attention_mask
|
|
|
|
def encode_prompt(
|
|
self,
|
|
prompt: Union[str, list[str]],
|
|
num_images_per_prompt: int = 1,
|
|
prompt_embeds: Optional[torch.Tensor] = None,
|
|
prompt_embeds_mask: Optional[torch.Tensor] = None,
|
|
max_sequence_length: int = 1024,
|
|
):
|
|
r"""
|
|
|
|
Args:
|
|
prompt (`str` or `list[str]`, *optional*):
|
|
prompt to be encoded
|
|
device: (`torch.device`):
|
|
torch device
|
|
num_images_per_prompt (`int`):
|
|
number of images that should be generated per prompt
|
|
prompt_embeds (`torch.Tensor`, *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.
|
|
"""
|
|
|
|
prompt = [prompt] if isinstance(prompt, str) else prompt
|
|
batch_size = len(prompt) if prompt_embeds is None else prompt_embeds.shape[0]
|
|
|
|
if prompt_embeds is None:
|
|
prompt_embeds, prompt_embeds_mask = self._get_qwen_prompt_embeds(prompt)
|
|
|
|
prompt_embeds = prompt_embeds[:, :max_sequence_length]
|
|
prompt_embeds_mask = prompt_embeds_mask[:, :max_sequence_length]
|
|
|
|
_, seq_len, _ = prompt_embeds.shape
|
|
prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
|
|
prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
|
|
prompt_embeds_mask = prompt_embeds_mask.repeat(1, num_images_per_prompt, 1)
|
|
prompt_embeds_mask = prompt_embeds_mask.view(batch_size * num_images_per_prompt, seq_len)
|
|
|
|
return prompt_embeds, prompt_embeds_mask
|
|
|
|
@staticmethod
|
|
def _pack_latents(latents, batch_size, num_channels_latents, height, width):
|
|
latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2)
|
|
latents = latents.permute(0, 2, 4, 1, 3, 5)
|
|
latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4)
|
|
|
|
return latents
|
|
|
|
@staticmethod
|
|
def _unpack_latents(latents, height, width, vae_scale_factor):
|
|
batch_size, num_patches, channels = latents.shape
|
|
|
|
# VAE applies 8x compression on images but we must also account for packing which requires
|
|
# latent height and width to be divisible by 2.
|
|
height = 2 * (int(height) // (vae_scale_factor * 2))
|
|
width = 2 * (int(width) // (vae_scale_factor * 2))
|
|
|
|
latents = latents.view(batch_size, height // 2, width // 2, channels // 4, 2, 2)
|
|
latents = latents.permute(0, 3, 1, 4, 2, 5)
|
|
|
|
latents = latents.reshape(batch_size, channels // (2 * 2), 1, height, width)
|
|
|
|
return latents
|
|
|
|
def prepare_latents(
|
|
self,
|
|
batch_size,
|
|
num_channels_latents,
|
|
height,
|
|
width,
|
|
dtype,
|
|
device,
|
|
generator,
|
|
latents=None,
|
|
) -> torch.Tensor:
|
|
# generator=torch.Generator(device="cuda").manual_seed(42)
|
|
# VAE applies 8x compression on images but we must also account for packing which requires
|
|
# latent height and width to be divisible by 2.
|
|
height = 2 * (int(height) // (self.vae_scale_factor * 2))
|
|
width = 2 * (int(width) // (self.vae_scale_factor * 2))
|
|
|
|
shape = (batch_size, 1, num_channels_latents, height, width)
|
|
|
|
if latents is not None:
|
|
return latents.to(device=device, dtype=dtype)
|
|
|
|
if isinstance(generator, list) and len(generator) != batch_size:
|
|
raise ValueError(
|
|
f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
|
|
f" size of {batch_size}. Make sure the batch size matches the length of the generators."
|
|
)
|
|
|
|
latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
|
|
latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width)
|
|
|
|
return latents
|
|
|
|
def prepare_timesteps(self, num_inference_steps, sigmas, image_seq_len):
|
|
sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas
|
|
# image_seq_len = latents.shape[1]
|
|
mu = calculate_shift(
|
|
image_seq_len,
|
|
self.scheduler.config.get("base_image_seq_len", 256),
|
|
self.scheduler.config.get("max_image_seq_len", 4096),
|
|
self.scheduler.config.get("base_shift", 0.5),
|
|
self.scheduler.config.get("max_shift", 1.15),
|
|
)
|
|
timesteps, num_inference_steps = retrieve_timesteps(
|
|
self.scheduler,
|
|
num_inference_steps,
|
|
sigmas=sigmas,
|
|
mu=mu,
|
|
)
|
|
return timesteps, num_inference_steps
|
|
|
|
@property
|
|
def guidance_scale(self):
|
|
return self._guidance_scale
|
|
|
|
@property
|
|
def attention_kwargs(self):
|
|
return self._attention_kwargs
|
|
|
|
@property
|
|
def num_timesteps(self):
|
|
return self._num_timesteps
|
|
|
|
@property
|
|
def current_timestep(self):
|
|
return self._current_timestep
|
|
|
|
@property
|
|
def interrupt(self):
|
|
return self._interrupt
|
|
|
|
def diffuse(
|
|
self,
|
|
prompt_embeds,
|
|
prompt_embeds_mask,
|
|
negative_prompt_embeds,
|
|
negative_prompt_embeds_mask,
|
|
latents,
|
|
img_shapes,
|
|
txt_seq_lens,
|
|
negative_txt_seq_lens,
|
|
timesteps,
|
|
do_true_cfg,
|
|
guidance,
|
|
true_cfg_scale,
|
|
):
|
|
self.scheduler.set_begin_index(0)
|
|
for i, t in enumerate(timesteps):
|
|
if self.interrupt:
|
|
continue
|
|
self._current_timestep = t
|
|
# broadcast to batch dimension and place on same device/dtype as latents
|
|
timestep = t.expand(latents.shape[0]).to(device=latents.device, dtype=latents.dtype)
|
|
noise_pred = self.transformer(
|
|
hidden_states=latents,
|
|
timestep=timestep / 1000,
|
|
guidance=guidance,
|
|
encoder_hidden_states_mask=prompt_embeds_mask,
|
|
encoder_hidden_states=prompt_embeds,
|
|
img_shapes=img_shapes,
|
|
txt_seq_lens=txt_seq_lens,
|
|
attention_kwargs=self.attention_kwargs,
|
|
return_dict=False,
|
|
)[0]
|
|
if do_true_cfg:
|
|
neg_noise_pred = self.transformer(
|
|
hidden_states=latents,
|
|
timestep=timestep / 1000,
|
|
guidance=guidance,
|
|
encoder_hidden_states_mask=negative_prompt_embeds_mask,
|
|
encoder_hidden_states=negative_prompt_embeds,
|
|
img_shapes=img_shapes,
|
|
txt_seq_lens=negative_txt_seq_lens,
|
|
attention_kwargs=self.attention_kwargs,
|
|
return_dict=False,
|
|
)[0]
|
|
comb_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred)
|
|
cond_norm = torch.norm(noise_pred, dim=-1, keepdim=True)
|
|
noise_norm = torch.norm(comb_pred, dim=-1, keepdim=True)
|
|
noise_pred = comb_pred * (cond_norm / noise_norm)
|
|
# compute the previous noisy sample x_t -> x_t-1
|
|
latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
|
|
return latents
|
|
|
|
def forward(
|
|
self,
|
|
req: OmniDiffusionRequest,
|
|
prompt: Union[str, list[str]] = "",
|
|
negative_prompt: Union[str, list[str]] = "",
|
|
true_cfg_scale: float = 4.0,
|
|
height: int | None = None,
|
|
width: int | None = None,
|
|
num_inference_steps: int = 50,
|
|
sigmas: Optional[list[float]] = None,
|
|
guidance_scale: float = 1.0,
|
|
num_images_per_prompt: int = 1,
|
|
generator: Optional[Union[torch.Generator, list[torch.Generator]]] = None,
|
|
latents: Optional[torch.Tensor] = None,
|
|
prompt_embeds: Optional[torch.Tensor] = None,
|
|
prompt_embeds_mask: Optional[torch.Tensor] = None,
|
|
negative_prompt_embeds: Optional[torch.Tensor] = None,
|
|
negative_prompt_embeds_mask: Optional[torch.Tensor] = None,
|
|
output_type: Optional[str] = "pil",
|
|
attention_kwargs: Optional[dict[str, Any]] = None,
|
|
callback_on_step_end_tensor_inputs: list[str] = ["latents"],
|
|
max_sequence_length: int = 512,
|
|
) -> DiffusionOutput:
|
|
# # TODO: only support single prompt now
|
|
# if req.prompt is not None:
|
|
# prompt = req.prompt[0] if isinstance(req.prompt, list) else req.prompt
|
|
prompt = req.prompt if req.prompt is not None else prompt
|
|
negative_prompt = req.negative_prompt if req.negative_prompt is not None else negative_prompt
|
|
height = req.height or self.default_sample_size * self.vae_scale_factor
|
|
width = req.width or self.default_sample_size * self.vae_scale_factor
|
|
num_inference_steps = req.num_inference_steps or num_inference_steps
|
|
generator = req.generator or generator
|
|
true_cfg_scale = req.true_cfg_scale or true_cfg_scale
|
|
# 1. check inputs
|
|
# 2. encode prompts
|
|
# 3. prepare latents and timesteps
|
|
# 4. diffusion process
|
|
# 5. decode latents
|
|
# 6. post-process outputs
|
|
self.check_inputs(
|
|
prompt,
|
|
height,
|
|
width,
|
|
negative_prompt,
|
|
prompt_embeds,
|
|
negative_prompt_embeds,
|
|
prompt_embeds_mask,
|
|
negative_prompt_embeds_mask,
|
|
callback_on_step_end_tensor_inputs,
|
|
max_sequence_length,
|
|
)
|
|
|
|
self._guidance_scale = guidance_scale
|
|
self._attention_kwargs = attention_kwargs
|
|
self._current_timestep = None
|
|
self._interrupt = False
|
|
|
|
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]
|
|
|
|
has_neg_prompt = negative_prompt is not None or (
|
|
negative_prompt_embeds is not None and negative_prompt_embeds_mask is not None
|
|
)
|
|
|
|
do_true_cfg = true_cfg_scale > 1 and has_neg_prompt
|
|
prompt_embeds, prompt_embeds_mask = self.encode_prompt(
|
|
prompt=prompt,
|
|
prompt_embeds=prompt_embeds,
|
|
prompt_embeds_mask=prompt_embeds_mask,
|
|
num_images_per_prompt=num_images_per_prompt,
|
|
max_sequence_length=max_sequence_length,
|
|
)
|
|
if do_true_cfg:
|
|
negative_prompt_embeds, negative_prompt_embeds_mask = self.encode_prompt(
|
|
prompt=negative_prompt,
|
|
prompt_embeds=negative_prompt_embeds,
|
|
prompt_embeds_mask=negative_prompt_embeds_mask,
|
|
num_images_per_prompt=num_images_per_prompt,
|
|
max_sequence_length=max_sequence_length,
|
|
)
|
|
|
|
num_channels_latents = self.transformer.in_channels // 4
|
|
latents = self.prepare_latents(
|
|
batch_size * num_images_per_prompt,
|
|
num_channels_latents,
|
|
height,
|
|
width,
|
|
prompt_embeds.dtype,
|
|
self.device,
|
|
generator,
|
|
latents,
|
|
)
|
|
img_shapes = [
|
|
[
|
|
(
|
|
1,
|
|
height // self.vae_scale_factor // 2,
|
|
width // self.vae_scale_factor // 2,
|
|
)
|
|
]
|
|
] * batch_size
|
|
|
|
timesteps, num_inference_steps = self.prepare_timesteps(num_inference_steps, sigmas, latents.shape[1])
|
|
# num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
|
|
self._num_timesteps = len(timesteps)
|
|
|
|
# handle guidance
|
|
if self.transformer.guidance_embeds:
|
|
guidance = torch.full([1], guidance_scale, dtype=torch.float32)
|
|
guidance = guidance.expand(latents.shape[0])
|
|
else:
|
|
guidance = None
|
|
|
|
if self.attention_kwargs is None:
|
|
self._attention_kwargs = {}
|
|
|
|
txt_seq_lens = prompt_embeds_mask.sum(dim=1).tolist() if prompt_embeds_mask is not None else None
|
|
negative_txt_seq_lens = (
|
|
negative_prompt_embeds_mask.sum(dim=1).tolist() if negative_prompt_embeds_mask is not None else None
|
|
)
|
|
# print inputp params
|
|
|
|
latents = self.diffuse(
|
|
prompt_embeds,
|
|
prompt_embeds_mask,
|
|
negative_prompt_embeds,
|
|
negative_prompt_embeds_mask,
|
|
latents,
|
|
img_shapes,
|
|
txt_seq_lens,
|
|
negative_txt_seq_lens,
|
|
timesteps,
|
|
do_true_cfg,
|
|
guidance,
|
|
true_cfg_scale,
|
|
)
|
|
|
|
self._current_timestep = None
|
|
if output_type == "latent":
|
|
image = latents
|
|
else:
|
|
latents = self._unpack_latents(latents, height, width, self.vae_scale_factor)
|
|
latents = latents.to(self.vae.dtype)
|
|
latents_mean = (
|
|
torch.tensor(self.vae.config.latents_mean)
|
|
.view(1, self.vae.config.z_dim, 1, 1, 1)
|
|
.to(latents.device, latents.dtype)
|
|
)
|
|
latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to(
|
|
latents.device, latents.dtype
|
|
)
|
|
latents = latents / latents_std + latents_mean
|
|
image = self.vae.decode(latents, return_dict=False)[0][:, :, 0]
|
|
# processed_image = self.image_processor.postprocess(image, output_type=output_type)
|
|
|
|
return DiffusionOutput(output=image)
|
|
|
|
def load_weights(self):
|
|
self.load_transformer()
|
|
|
|
# TODO: find a better way to load
|
|
def load_transformer(self):
|
|
import glob
|
|
import os
|
|
|
|
# Define the weight iterator
|
|
def weight_iterator(transformer_path):
|
|
if not os.path.exists(transformer_path):
|
|
logger.warning(f"Path {transformer_path} does not exist.")
|
|
return
|
|
|
|
# Look for safetensors first
|
|
safetensors_files = glob.glob(os.path.join(transformer_path, "*.safetensors"))
|
|
if safetensors_files:
|
|
try:
|
|
from safetensors.torch import load_file
|
|
except ImportError:
|
|
logger.warning("safetensors not installed, cannot load .safetensors files.")
|
|
return
|
|
|
|
for file_path in safetensors_files:
|
|
state_dict = load_file(file_path)
|
|
for name, tensor in state_dict.items():
|
|
yield name, tensor
|
|
else:
|
|
# Fallback to bin
|
|
bin_files = glob.glob(os.path.join(transformer_path, "*.bin"))
|
|
for file_path in bin_files:
|
|
state_dict = torch.load(file_path)
|
|
for name, tensor in state_dict.items():
|
|
yield name, tensor
|
|
|
|
model_name = "Qwen/Qwen-Image"
|
|
try:
|
|
# Use the arguments as verified by the user
|
|
model_path = download_weights_from_hf_specific(model_name, None, ["*"])
|
|
transformer_path = os.path.join(model_path, "transformer")
|
|
self.transformer.load_weights(weight_iterator(transformer_path))
|
|
|
|
except Exception as e:
|
|
print(f"An error occurred: {e}")
|
|
raise e
|