dylanyunlong/project_6
1
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

feat(CRITICAL): 从 GitHub 扫描搬运 ixformer SDK + xllm 完整 GDN/MoE 代码

Browse Source 
来源:
  1. Chranos/ixformer (GitHub) → ixformer_sdk/ (230 files, 70K lines)
     - inference/functions/vllm.py: vllm_moe_topk_softmax 完整实现 (2033 lines)
     - inference/functions/moe.py: MoE ops 完整实现 (1380 lines)
     - contrib/vllm_flash_attn/: FA2 Python 接口 (1018 lines)
     - contrib/tgi/fused_moe.py: TGI fused MoE (429 lines)
     - csrc/include/ixformer/: C++ kernel headers + cmake

  2. Deep-Spark/xllm (GitHub) → upstream_ref/xllm_latest/ (+15 files)
     - npu_torch/qwen3_5_decoder_layer_impl.cpp/.h
     - npu_torch/qwen3_5_gated_delta_net.cpp/.h
     - npu_torch/qwen3_next_*.cpp/.h (6 files)
     - npu_torch/attention.cpp/.h + fused_moe.cpp/.h + CMakeLists.txt
     - models/llm/qwen3_5.h + qwen3_5_mtp.h + qwen3_next.h
     - models/vlm/qwen3_5.h

调用链完整性:
  ixformer_sdk/inference/functions/vllm.py
    → ops.infer.moe_topk_softmax() (C++ 层)
    → 这就是 base 镜像 libixformer.so 里的实现

  upstream_ref/xllm_latest/core/layers/ilu/fused_moe.cpp
    → ixformer::infer::topk_softmax() (直接 C++ 调用)
    → ixformer::infer::group_gemm() → 完整 7-step MoE pipeline
This commit is contained in:
project6-dev
2026-08-11 02:31:56 +00:00
parent a8b16da5da
commit 87a19d2d00
250 changed files with 76690 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

0
ixformer_sdk/contrib/DeepCache/sd/__init__.py Normal file
View File

812
ixformer_sdk/contrib/DeepCache/sd/pipeline_stable_diffusion.py Normal file
View File

@@ -0,0 +1,812 @@
# Copyright 2023 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 time
import inspect
from typing import Any, Callable, Dict, List, Optional, Union
import torch
import numpy as np
from packaging import version
from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
from diffusers.configuration_utils import FrozenDict
from diffusers.image_processor import VaeImageProcessor
from diffusers.loaders import FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin
from diffusers.models import AutoencoderKL
from diffusers.models.lora import adjust_lora_scale_text_encoder
from diffusers.schedulers import KarrasDiffusionSchedulers
from diffusers.utils import (
deprecate,
logging,
replace_example_docstring,
)
from diffusers.utils.torch_utils import randn_tensor
from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
from .unet_2d_condition import UNet2DConditionModel
from .pipeline_utils import DiffusionPipeline
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
EXAMPLE_DOC_STRING = """
Examples:
```py
>>> import torch
>>> from diffusers import StableDiffusionPipeline
>>> pipe = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16)
>>> pipe = pipe.to("cuda")
>>> prompt = "a photo of an astronaut riding a horse on mars"
>>> image = pipe(prompt).images[0]
```
"""
def sample_gaussian_centered(n=1000, sample_size=100, std_dev=100):
samples = []
while len(samples) < sample_size:
# Sample from a Gaussian centered at n/2
sample = int(np.random.normal(loc=n/2, scale=std_dev))
# Check if the sample is in bounds
if 1 <= sample < n and sample not in samples:
samples.append(sample)
return samples
def sample_from_quad(total_numbers, n_samples, pow=1.2):
while pow > 1:
# Generate linearly spaced values between 0 and a max value
x_values = np.linspace(0, total_numbers**(1/pow), n_samples+1)
# Raise these values to the power of 1.5 to get a non-linear distribution
indices = np.unique(np.int32(x_values**pow))[:-1]
if len(indices) == n_samples:
break
pow -=0.02
if pow <= 1:
raise ValueError("Cannot find suitable pow. Please adjust n_samples or decrease center.")
return indices, pow
def sample_from_quad_center(total_numbers, n_samples, center, pow=1.2):
while pow > 1:
# Generate linearly spaced values between 0 and a max value
x_values = np.linspace((-center)**(1/pow), (total_numbers-center)**(1/pow), n_samples+1)
indices = [0] + [x+center for x in np.unique(np.int32(x_values**pow))[1:-1]]
if len(indices) == n_samples:
break
pow -=0.02
if pow <= 1:
raise ValueError("Cannot find suitable pow. Please adjust n_samples or decrease center.")
return indices, pow
def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
"""
Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4
"""
std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
# rescale the results from guidance (fixes overexposure)
noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
# mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
return noise_cfg
class StableDiffusionPipeline(DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, FromSingleFileMixin):
r"""
Pipeline for text-to-image generation using Stable Diffusion.
This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
implemented for all pipelines (downloading, saving, running on a particular device, etc.).
The pipeline also inherits the following loading methods:
- [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
- [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
- [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
- [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
Args:
vae ([`AutoencoderKL`]):
Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
text_encoder ([`~transformers.CLIPTextModel`]):
Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
tokenizer ([`~transformers.CLIPTokenizer`]):
A `CLIPTokenizer` to tokenize text.
unet ([`UNet2DConditionModel`]):
A `UNet2DConditionModel` to denoise the encoded image latents.
scheduler ([`SchedulerMixin`]):
A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
[`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
safety_checker ([`StableDiffusionSafetyChecker`]):
Classification module that estimates whether generated images could be considered offensive or harmful.
Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
about a model's potential harms.
feature_extractor ([`~transformers.CLIPImageProcessor`]):
A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
"""
model_cpu_offload_seq = "text_encoder->unet->vae"
_optional_components = ["safety_checker", "feature_extractor"]
_exclude_from_cpu_offload = ["safety_checker"]
def __init__(
self,
vae: AutoencoderKL,
text_encoder: CLIPTextModel,
tokenizer: CLIPTokenizer,
unet: UNet2DConditionModel,
scheduler: KarrasDiffusionSchedulers,
safety_checker: StableDiffusionSafetyChecker,
feature_extractor: CLIPImageProcessor,
requires_safety_checker: bool = True,
):
super().__init__()
if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
deprecation_message = (
f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
"to update the config accordingly as leaving `steps_offset` might led to incorrect results"
" in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
" it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
" file"
)
deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
new_config = dict(scheduler.config)
new_config["steps_offset"] = 1
scheduler._internal_dict = FrozenDict(new_config)
if hasattr(scheduler.config, "clip_sample") and scheduler.config.clip_sample is True:
deprecation_message = (
f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`."
" `clip_sample` should be set to False in the configuration file. Please make sure to update the"
" config accordingly as not setting `clip_sample` in the config might lead to incorrect results in"
" future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very"
" nice if you could open a Pull request for the `scheduler/scheduler_config.json` file"
)
deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False)
new_config = dict(scheduler.config)
new_config["clip_sample"] = False
scheduler._internal_dict = FrozenDict(new_config)
if safety_checker is None and requires_safety_checker:
logger.warning(
f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
" that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
" results in services or applications open to the public. Both the diffusers team and Hugging Face"
" strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
" it only for use-cases that involve analyzing network behavior or auditing its results. For more"
" information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
)
if safety_checker is not None and feature_extractor is None:
raise ValueError(
"Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
" checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
)
is_unet_version_less_0_9_0 = hasattr(unet.config, "_diffusers_version") and version.parse(
version.parse(unet.config._diffusers_version).base_version
) < version.parse("0.9.0.dev0")
is_unet_sample_size_less_64 = hasattr(unet.config, "sample_size") and unet.config.sample_size < 64
if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64:
deprecation_message = (
"The configuration file of the unet has set the default `sample_size` to smaller than"
" 64 which seems highly unlikely. If your checkpoint is a fine-tuned version of any of the"
" following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-"
" CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- runwayml/stable-diffusion-v1-5"
" \n- runwayml/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the"
" configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`"
" in the config might lead to incorrect results in future versions. If you have downloaded this"
" checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for"
" the `unet/config.json` file"
)
deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False)
new_config = dict(unet.config)
new_config["sample_size"] = 64
unet._internal_dict = FrozenDict(new_config)
self.register_modules(
vae=vae,
text_encoder=text_encoder,
tokenizer=tokenizer,
unet=unet,
scheduler=scheduler,
safety_checker=safety_checker,
feature_extractor=feature_extractor,
)
self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
self.register_to_config(requires_safety_checker=requires_safety_checker)
def enable_vae_slicing(self):
r"""
Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
"""
self.vae.enable_slicing()
def disable_vae_slicing(self):
r"""
Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
computing decoding in one step.
"""
self.vae.disable_slicing()
def enable_vae_tiling(self):
r"""
Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
processing larger images.
"""
self.vae.enable_tiling()
def disable_vae_tiling(self):
r"""
Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
computing decoding in one step.
"""
self.vae.disable_tiling()
def _encode_prompt(
self,
prompt,
device,
num_images_per_prompt,
do_classifier_free_guidance,
negative_prompt=None,
prompt_embeds: Optional[torch.FloatTensor] = None,
negative_prompt_embeds: Optional[torch.FloatTensor] = None,
lora_scale: Optional[float] = None,
):
deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
prompt_embeds_tuple = self.encode_prompt(
prompt=prompt,
device=device,
num_images_per_prompt=num_images_per_prompt,
do_classifier_free_guidance=do_classifier_free_guidance,
negative_prompt=negative_prompt,
prompt_embeds=prompt_embeds,
negative_prompt_embeds=negative_prompt_embeds,
lora_scale=lora_scale,
)
# concatenate for backwards comp
prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
return prompt_embeds
def encode_prompt(
self,
prompt,
device,
num_images_per_prompt,
do_classifier_free_guidance,
negative_prompt=None,
prompt_embeds: Optional[torch.FloatTensor] = None,
negative_prompt_embeds: Optional[torch.FloatTensor] = None,
lora_scale: Optional[float] = None,
):
r"""
Encodes the prompt into text encoder hidden states.
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
do_classifier_free_guidance (`bool`):
whether to use classifier free guidance or not
negative_prompt (`str` or `List[str]`, *optional*):
The prompt or prompts not to guide the image generation. If not defined, one has to pass
`negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
less than `1`).
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.
negative_prompt_embeds (`torch.FloatTensor`, *optional*):
Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
argument.
lora_scale (`float`, *optional*):
A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
"""
# set lora scale so that monkey patched LoRA
# function of text encoder can correctly access it
if lora_scale is not None and isinstance(self, LoraLoaderMixin):
self._lora_scale = lora_scale
# dynamically adjust the LoRA scale
adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
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]
if prompt_embeds is None:
# textual inversion: procecss multi-vector tokens if necessary
if isinstance(self, TextualInversionLoaderMixin):
prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
text_inputs = self.tokenizer(
prompt,
padding="max_length",
max_length=self.tokenizer.model_max_length,
truncation=True,
return_tensors="pt",
)
text_input_ids = text_inputs.input_ids
untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
text_input_ids, untruncated_ids
):
removed_text = self.tokenizer.batch_decode(
untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
)
logger.warning(
"The following part of your input was truncated because CLIP can only handle sequences up to"
f" {self.tokenizer.model_max_length} tokens: {removed_text}"
)
if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
attention_mask = text_inputs.attention_mask.to(device)
else:
attention_mask = None
prompt_embeds = self.text_encoder(
text_input_ids.to(device),
attention_mask=attention_mask,
)
prompt_embeds = prompt_embeds[0]
if self.text_encoder is not None:
prompt_embeds_dtype = self.text_encoder.dtype
elif self.unet is not None:
prompt_embeds_dtype = self.unet.dtype
else:
prompt_embeds_dtype = prompt_embeds.dtype
prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
bs_embed, seq_len, _ = prompt_embeds.shape
# duplicate text embeddings for each generation per prompt, using mps friendly method
prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
# get unconditional embeddings for classifier free guidance
if do_classifier_free_guidance and negative_prompt_embeds is None:
uncond_tokens: List[str]
if negative_prompt is None:
uncond_tokens = [""] * batch_size
elif prompt is not None and type(prompt) is not type(negative_prompt):
raise TypeError(
f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
f" {type(prompt)}."
)
elif isinstance(negative_prompt, str):
uncond_tokens = [negative_prompt]
elif batch_size != len(negative_prompt):
raise ValueError(
f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
" the batch size of `prompt`."
)
else:
uncond_tokens = negative_prompt
# textual inversion: procecss multi-vector tokens if necessary
if isinstance(self, TextualInversionLoaderMixin):
uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
max_length = prompt_embeds.shape[1]
uncond_input = self.tokenizer(
uncond_tokens,
padding="max_length",
max_length=max_length,
truncation=True,
return_tensors="pt",
)
if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
attention_mask = uncond_input.attention_mask.to(device)
else:
attention_mask = None
negative_prompt_embeds = self.text_encoder(
uncond_input.input_ids.to(device),
attention_mask=attention_mask,
)
negative_prompt_embeds = negative_prompt_embeds[0]
if do_classifier_free_guidance:
# duplicate unconditional embeddings for each generation per prompt, using mps friendly method
seq_len = negative_prompt_embeds.shape[1]
negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
return prompt_embeds, negative_prompt_embeds
def run_safety_checker(self, image, device, dtype):
if self.safety_checker is None:
has_nsfw_concept = None
else:
if torch.is_tensor(image):
feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
else:
feature_extractor_input = self.image_processor.numpy_to_pil(image)
safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
image, has_nsfw_concept = self.safety_checker(
images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
)
return image, has_nsfw_concept
def decode_latents(self, latents):
deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
latents = 1 / self.vae.config.scaling_factor * latents
image = self.vae.decode(latents, return_dict=False)[0]
image = (image / 2 + 0.5).clamp(0, 1)
# we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
image = image.cpu().permute(0, 2, 3, 1).float().numpy()
return image
def prepare_extra_step_kwargs(self, generator, eta):
# prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
# eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
# eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
# and should be between [0, 1]
accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
extra_step_kwargs = {}
if accepts_eta:
extra_step_kwargs["eta"] = eta
# check if the scheduler accepts generator
accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
if accepts_generator:
extra_step_kwargs["generator"] = generator
return extra_step_kwargs
def check_inputs(
self,
prompt,
height,
width,
callback_steps,
negative_prompt=None,
prompt_embeds=None,
negative_prompt_embeds=None,
):
if height % 8 != 0 or width % 8 != 0:
raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
if (callback_steps is None) or (
callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
):
raise ValueError(
f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
f" {type(callback_steps)}."
)
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 negative_prompt_embeds is not None:
if prompt_embeds.shape != negative_prompt_embeds.shape:
raise ValueError(
"`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
f" {negative_prompt_embeds.shape}."
)
def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
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."
)
if latents is None:
latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
else:
latents = latents.to(device)
# scale the initial noise by the standard deviation required by the scheduler
latents = latents * self.scheduler.init_noise_sigma
return latents
@torch.no_grad()
@replace_example_docstring(EXAMPLE_DOC_STRING)
def __call__(
self,
prompt: Union[str, List[str]] = None,
height: Optional[int] = None,
width: Optional[int] = None,
num_inference_steps: int = 50,
guidance_scale: float = 7.5,
negative_prompt: Optional[Union[str, List[str]]] = None,
num_images_per_prompt: Optional[int] = 1,
eta: float = 0.0,
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
latents: Optional[torch.FloatTensor] = None,
prompt_embeds: Optional[torch.FloatTensor] = None,
negative_prompt_embeds: Optional[torch.FloatTensor] = None,
output_type: Optional[str] = "pil",
return_dict: bool = True,
callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
callback_steps: int = 1,
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
guidance_rescale: float = 0.0,
cache_interval: int = 1,
cache_layer_id: int = None,
cache_block_id: int = None,
uniform: bool = True,
pow: float = None,
center: int = None,
output_all_sequence: bool = False,
):
r"""
The call function to the pipeline for generation.
Args:
prompt (`str` or `List[str]`, *optional*):
The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
The height in pixels of the generated image.
width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
The width in pixels of the generated image.
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.
guidance_scale (`float`, *optional*, defaults to 7.5):
A higher guidance scale value encourages the model to generate images closely linked to the text
`prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
negative_prompt (`str` or `List[str]`, *optional*):
The prompt or prompts to guide what to not include in image generation. If not defined, you need to
pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
num_images_per_prompt (`int`, *optional*, defaults to 1):
The number of images to generate per prompt.
eta (`float`, *optional*, defaults to 0.0):
Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
A [`torch.Generator`](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 is 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 (prompt weighting). If not
provided, text embeddings are generated from the `prompt` input argument.
negative_prompt_embeds (`torch.FloatTensor`, *optional*):
Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
output_type (`str`, *optional*, defaults to `"pil"`):
The output format of the generated image. Choose between `PIL.Image` or `np.array`.
return_dict (`bool`, *optional*, defaults to `True`):
Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
plain tuple.
callback (`Callable`, *optional*):
A function that calls every `callback_steps` steps during inference. The function is called with the
following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
callback_steps (`int`, *optional*, defaults to 1):
The frequency at which the `callback` function is called. If not specified, the callback is called at
every step.
cross_attention_kwargs (`dict`, *optional*):
A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
[`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
guidance_rescale (`float`, *optional*, defaults to 0.7):
Guidance rescale factor from [Common Diffusion Noise Schedules and Sample Steps are
Flawed](https://arxiv.org/pdf/2305.08891.pdf). Guidance rescale factor should fix overexposure when
using zero terminal SNR.
Examples:
Returns:
[`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
otherwise a `tuple` is returned where the first element is a list with the generated images and the
second element is a list of `bool`s indicating whether the corresponding generated image contains
"not-safe-for-work" (nsfw) content.
"""
# 0. Default height and width to unet
height = height or self.unet.config.sample_size * self.vae_scale_factor
width = width or self.unet.config.sample_size * self.vae_scale_factor
# 1. Check inputs. Raise error if not correct
self.check_inputs(
prompt, height, width, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds
)
# 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
# here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
# of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
# corresponds to doing no classifier free guidance.
do_classifier_free_guidance = guidance_scale > 1.0
# 3. Encode input prompt
text_encoder_lora_scale = (
cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
)
prompt_embeds, negative_prompt_embeds = self.encode_prompt(
prompt,
device,
num_images_per_prompt,
do_classifier_free_guidance,
negative_prompt,
prompt_embeds=prompt_embeds,
negative_prompt_embeds=negative_prompt_embeds,
lora_scale=text_encoder_lora_scale,
)
# For classifier free guidance, we need to do two forward passes.
# Here we concatenate the unconditional and text embeddings into a single batch
# to avoid doing two forward passes
if do_classifier_free_guidance:
prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
# 4. Prepare timesteps
self.scheduler.set_timesteps(num_inference_steps, device=device)
timesteps = self.scheduler.timesteps
# 5. Prepare latent variables
num_channels_latents = self.unet.config.in_channels
latents = self.prepare_latents(
batch_size * num_images_per_prompt,
num_channels_latents,
height,
width,
prompt_embeds.dtype,
device,
generator,
latents,
)
# 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
# 7. Denoising loop
num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
prv_features = None
latents_list = [latents]
if cache_interval == 1:
interval_seq = list(range(num_inference_steps))
else:
if uniform:
interval_seq = list(range(0, num_inference_steps, cache_interval))
else:
num_slow_step = num_inference_steps//cache_interval
if num_inference_steps%cache_interval != 0:
num_slow_step += 1
interval_seq, pow = sample_from_quad_center(num_inference_steps, num_slow_step, center=center, pow=pow)#[0, 3, 6, 9, 12, 16, 22, 28, 35, 43,]
#interval_seq, pow = sample_from_quad(num_inference_steps, num_inference_steps//cache_interval, pow=pow)#[0, 3, 6, 9, 12, 16, 22, 28, 35, 43,]
interval_seq = sorted(interval_seq)
#print(interval_seq, len(interval_seq), pow)
with self.progress_bar(total=num_inference_steps) as progress_bar:
#print("[INFO] Update Feature Interval = {}, Update Layer Number = {}, Update Block Number = {}".format(cache_interval, cache_layer_id, cache_block_id))
for i, t in enumerate(timesteps):
# expand the latents if we are doing classifier free guidance
latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
if i in interval_seq:
prv_features = None
# predict the noise residual
noise_pred, prv_features = self.unet(
latent_model_input,
t,
encoder_hidden_states=prompt_embeds,
cross_attention_kwargs=cross_attention_kwargs,
replicate_prv_feature=prv_features,
quick_replicate= cache_interval>1,
cache_layer_id=cache_layer_id,
cache_block_id=cache_block_id,
return_dict=False,
)
# perform guidance
if do_classifier_free_guidance:
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
if do_classifier_free_guidance and guidance_rescale > 0.0:
# Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale)
# compute the previous noisy sample x_t -> x_t-1
latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
latents_list.append(latents)
# call the callback, if provided
if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
progress_bar.update()
if callback is not None and i % callback_steps == 0:
callback(i, t, latents)
if not output_type == "latent":
if output_all_sequence:
image = [self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] for latents in latents_list]
has_nsfw_concept = None #self.run_safety_checker(images[0], device, prompt_embeds.dtype)
num_img = len(image)
else:
image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
has_nsfw_concept = None
num_img = image.shape[0]
else:
image = latents
has_nsfw_concept = None
if has_nsfw_concept is None:
do_denormalize = [True] * num_img
else:
do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
if output_all_sequence:
image = [self.image_processor.postprocess(img, output_type=output_type, do_denormalize=do_denormalize) for img in image]
else:
image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
# Offload all models
self.maybe_free_model_hooks()
if not return_dict:
return (image, has_nsfw_concept,)
return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)

741
ixformer_sdk/contrib/DeepCache/sd/pipeline_text_to_video_zero.py Normal file
View File

@@ -0,0 +1,741 @@
import copy
from dataclasses import dataclass
from typing import Callable, List, Optional, Union
import numpy as np
import PIL.Image
import torch
import torch.nn.functional as F
from torch.nn.functional import grid_sample
from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
from diffusers.models import AutoencoderKL
from .unet_2d_condition import UNet2DConditionModel
from .pipeline_stable_diffusion import StableDiffusionPipeline, StableDiffusionSafetyChecker
from diffusers.schedulers import KarrasDiffusionSchedulers
from diffusers.utils import BaseOutput
from diffusers.utils.torch_utils import randn_tensor
def sample_gaussian_centered(n=1000, sample_size=100, std_dev=100):
samples = []
while len(samples) < sample_size:
# Sample from a Gaussian centered at n/2
sample = int(np.random.normal(loc=n/2, scale=std_dev))
# Check if the sample is in bounds
if 1 <= sample < n and sample not in samples:
samples.append(sample)
return samples
def sample_from_quad(total_numbers, n_samples, pow=1.2):
while pow > 1:
# Generate linearly spaced values between 0 and a max value
x_values = np.linspace(0, total_numbers**(1/pow), n_samples+1)
# Raise these values to the power of 1.5 to get a non-linear distribution
indices = np.unique(np.int32(x_values**pow))[:-1]
if len(indices) == n_samples:
break
pow -=0.02
if pow <= 1:
raise ValueError("Cannot find suitable pow. Please adjust n_samples or decrease center.")
return indices, pow
def sample_from_quad_center(total_numbers, n_samples, center, pow=1.2):
while pow > 1:
# Generate linearly spaced values between 0 and a max value
x_values = np.linspace((-center)**(1/pow), (total_numbers-center)**(1/pow), n_samples+1)
indices = [0] + [x+center for x in np.unique(np.int32(x_values**pow))[1:-1]]
if len(indices) == n_samples:
break
pow -=0.02
if pow <= 1:
raise ValueError("Cannot find suitable pow. Please adjust n_samples or decrease center.")
return indices, pow
def rearrange_0(tensor, f):
F, C, H, W = tensor.size()
tensor = torch.permute(torch.reshape(tensor, (F // f, f, C, H, W)), (0, 2, 1, 3, 4))
return tensor
def rearrange_1(tensor):
B, C, F, H, W = tensor.size()
return torch.reshape(torch.permute(tensor, (0, 2, 1, 3, 4)), (B * F, C, H, W))
def rearrange_3(tensor, f):
F, D, C = tensor.size()
return torch.reshape(tensor, (F // f, f, D, C))
def rearrange_4(tensor):
B, F, D, C = tensor.size()
return torch.reshape(tensor, (B * F, D, C))
class CrossFrameAttnProcessor:
"""
Cross frame attention processor. Each frame attends the first frame.
Args:
batch_size: The number that represents actual batch size, other than the frames.
For example, calling unet with a single prompt and num_images_per_prompt=1, batch_size should be equal to
2, due to classifier-free guidance.
"""
def __init__(self, batch_size=2):
self.batch_size = batch_size
def __call__(self, attn, hidden_states, encoder_hidden_states=None, attention_mask=None):
batch_size, sequence_length, _ = hidden_states.shape
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
query = attn.to_q(hidden_states)
is_cross_attention = encoder_hidden_states is not None
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
elif attn.norm_cross:
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
key = attn.to_k(encoder_hidden_states)
value = attn.to_v(encoder_hidden_states)
# Cross Frame Attention
if not is_cross_attention:
video_length = key.size()[0] // self.batch_size
first_frame_index = [0] * video_length
# rearrange keys to have batch and frames in the 1st and 2nd dims respectively
key = rearrange_3(key, video_length)
key = key[:, first_frame_index]
# rearrange values to have batch and frames in the 1st and 2nd dims respectively
value = rearrange_3(value, video_length)
value = value[:, first_frame_index]
# rearrange back to original shape
key = rearrange_4(key)
value = rearrange_4(value)
query = attn.head_to_batch_dim(query)
key = attn.head_to_batch_dim(key)
value = attn.head_to_batch_dim(value)
attention_probs = attn.get_attention_scores(query, key, attention_mask)
hidden_states = torch.bmm(attention_probs, value)
hidden_states = attn.batch_to_head_dim(hidden_states)
# linear proj
hidden_states = attn.to_out[0](hidden_states)
# dropout
hidden_states = attn.to_out[1](hidden_states)
return hidden_states
class CrossFrameAttnProcessor2_0:
"""
Cross frame attention processor with scaled_dot_product attention of Pytorch 2.0.
Args:
batch_size: The number that represents actual batch size, other than the frames.
For example, calling unet with a single prompt and num_images_per_prompt=1, batch_size should be equal to
2, due to classifier-free guidance.
"""
def __init__(self, batch_size=2):
if not hasattr(F, "scaled_dot_product_attention"):
raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
self.batch_size = batch_size
def __call__(self, attn, hidden_states, encoder_hidden_states=None, attention_mask=None):
batch_size, sequence_length, _ = (
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
)
inner_dim = hidden_states.shape[-1]
if attention_mask is not None:
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
# scaled_dot_product_attention expects attention_mask shape to be
# (batch, heads, source_length, target_length)
attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
query = attn.to_q(hidden_states)
is_cross_attention = encoder_hidden_states is not None
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
elif attn.norm_cross:
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
key = attn.to_k(encoder_hidden_states)
value = attn.to_v(encoder_hidden_states)
# Cross Frame Attention
if not is_cross_attention:
video_length = max(1, key.size()[0] // self.batch_size)
first_frame_index = [0] * video_length
# rearrange keys to have batch and frames in the 1st and 2nd dims respectively
key = rearrange_3(key, video_length)
key = key[:, first_frame_index]
# rearrange values to have batch and frames in the 1st and 2nd dims respectively
value = rearrange_3(value, video_length)
value = value[:, first_frame_index]
# rearrange back to original shape
key = rearrange_4(key)
value = rearrange_4(value)
head_dim = inner_dim // attn.heads
query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# the output of sdp = (batch, num_heads, seq_len, head_dim)
# TODO: add support for attn.scale when we move to Torch 2.1
hidden_states = F.scaled_dot_product_attention(
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
)
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
hidden_states = hidden_states.to(query.dtype)
# linear proj
hidden_states = attn.to_out[0](hidden_states)
# dropout
hidden_states = attn.to_out[1](hidden_states)
return hidden_states
@dataclass
class TextToVideoPipelineOutput(BaseOutput):
r"""
Output class for zero-shot text-to-video pipeline.
Args:
images (`[List[PIL.Image.Image]`, `np.ndarray`]):
List of denoised PIL images of length `batch_size` or NumPy array of shape `(batch_size, height, width,
num_channels)`.
nsfw_content_detected (`[List[bool]]`):
List indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content or
`None` if safety checking could not be performed.
"""
images: Union[List[PIL.Image.Image], np.ndarray]
nsfw_content_detected: Optional[List[bool]]
def coords_grid(batch, ht, wd, device):
# Adapted from https://github.com/princeton-vl/RAFT/blob/master/core/utils/utils.py
coords = torch.meshgrid(torch.arange(ht, device=device), torch.arange(wd, device=device))
coords = torch.stack(coords[::-1], dim=0).float()
return coords[None].repeat(batch, 1, 1, 1)
def warp_single_latent(latent, reference_flow):
"""
Warp latent of a single frame with given flow
Args:
latent: latent code of a single frame
reference_flow: flow which to warp the latent with
Returns:
warped: warped latent
"""
_, _, H, W = reference_flow.size()
_, _, h, w = latent.size()
coords0 = coords_grid(1, H, W, device=latent.device).to(latent.dtype)
coords_t0 = coords0 + reference_flow
coords_t0[:, 0] /= W
coords_t0[:, 1] /= H
coords_t0 = coords_t0 * 2.0 - 1.0
coords_t0 = F.interpolate(coords_t0, size=(h, w), mode="bilinear")
coords_t0 = torch.permute(coords_t0, (0, 2, 3, 1))
warped = grid_sample(latent, coords_t0, mode="nearest", padding_mode="reflection")
return warped
def create_motion_field(motion_field_strength_x, motion_field_strength_y, frame_ids, device, dtype):
"""
Create translation motion field
Args:
motion_field_strength_x: motion strength along x-axis
motion_field_strength_y: motion strength along y-axis
frame_ids: indexes of the frames the latents of which are being processed.
This is needed when we perform chunk-by-chunk inference
device: device
dtype: dtype
Returns:
"""
seq_length = len(frame_ids)
reference_flow = torch.zeros((seq_length, 2, 512, 512), device=device, dtype=dtype)
for fr_idx in range(seq_length):
reference_flow[fr_idx, 0, :, :] = motion_field_strength_x * (frame_ids[fr_idx])
reference_flow[fr_idx, 1, :, :] = motion_field_strength_y * (frame_ids[fr_idx])
return reference_flow
def create_motion_field_and_warp_latents(motion_field_strength_x, motion_field_strength_y, frame_ids, latents):
"""
Creates translation motion and warps the latents accordingly
Args:
motion_field_strength_x: motion strength along x-axis
motion_field_strength_y: motion strength along y-axis
frame_ids: indexes of the frames the latents of which are being processed.
This is needed when we perform chunk-by-chunk inference
latents: latent codes of frames
Returns:
warped_latents: warped latents
"""
motion_field = create_motion_field(
motion_field_strength_x=motion_field_strength_x,
motion_field_strength_y=motion_field_strength_y,
frame_ids=frame_ids,
device=latents.device,
dtype=latents.dtype,
)
warped_latents = latents.clone().detach()
for i in range(len(warped_latents)):
warped_latents[i] = warp_single_latent(latents[i][None], motion_field[i][None])
return warped_latents
class TextToVideoZeroPipeline(StableDiffusionPipeline):
r"""
Pipeline for zero-shot text-to-video generation using Stable Diffusion.
This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
implemented for all pipelines (downloading, saving, running on a particular device, etc.).
Args:
vae ([`AutoencoderKL`]):
Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
text_encoder ([`CLIPTextModel`]):
Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
tokenizer (`CLIPTokenizer`):
A [`~transformers.CLIPTokenizer`] to tokenize text.
unet ([`UNet2DConditionModel`]):
A [`UNet3DConditionModel`] to denoise the encoded video latents.
scheduler ([`SchedulerMixin`]):
A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
[`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
safety_checker ([`StableDiffusionSafetyChecker`]):
Classification module that estimates whether generated images could be considered offensive or harmful.
Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
about a model's potential harms.
feature_extractor ([`CLIPImageProcessor`]):
A [`CLIPImageProcessor`] to extract features from generated images; used as inputs to the `safety_checker`.
"""
def __init__(
self,
vae: AutoencoderKL,
text_encoder: CLIPTextModel,
tokenizer: CLIPTokenizer,
unet: UNet2DConditionModel,
scheduler: KarrasDiffusionSchedulers,
safety_checker: StableDiffusionSafetyChecker,
feature_extractor: CLIPImageProcessor,
requires_safety_checker: bool = True,
):
super().__init__(
vae, text_encoder, tokenizer, unet, scheduler, safety_checker, feature_extractor, requires_safety_checker
)
processor = (
CrossFrameAttnProcessor2_0(batch_size=2)
if hasattr(F, "scaled_dot_product_attention")
else CrossFrameAttnProcessor(batch_size=2)
)
self.unet.set_attn_processor(processor)
def forward_loop(self, x_t0, t0, t1, generator):
"""
Perform DDPM forward process from time t0 to t1. This is the same as adding noise with corresponding variance.
Args:
x_t0:
Latent code at time t0.
t0:
Timestep at t0.
t1:
Timestamp at t1.
generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
generation deterministic.
Returns:
x_t1:
Forward process applied to x_t0 from time t0 to t1.
"""
eps = randn_tensor(x_t0.size(), generator=generator, dtype=x_t0.dtype, device=x_t0.device)
alpha_vec = torch.prod(self.scheduler.alphas[t0:t1])
x_t1 = torch.sqrt(alpha_vec) * x_t0 + torch.sqrt(1 - alpha_vec) * eps
return x_t1
def backward_loop(
self,
latents,
timesteps,
prompt_embeds,
guidance_scale,
callback,
callback_steps,
num_warmup_steps,
extra_step_kwargs,
prv_features,
interval_seq,
cache_interval,
cache_block_id,
cache_layer_id,
cross_attention_kwargs=None,
):
"""
Perform backward process given list of time steps.
Args:
latents:
Latents at time timesteps[0].
timesteps:
Time steps along which to perform backward process.
prompt_embeds:
Pre-generated text embeddings.
guidance_scale:
A higher guidance scale value encourages the model to generate images closely linked to the text
`prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
callback (`Callable`, *optional*):
A function that calls every `callback_steps` steps during inference. The function is called with the
following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
callback_steps (`int`, *optional*, defaults to 1):
The frequency at which the `callback` function is called. If not specified, the callback is called at
every step.
extra_step_kwargs:
Extra_step_kwargs.
cross_attention_kwargs:
A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
[`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
num_warmup_steps:
number of warmup steps.
Returns:
latents:
Latents of backward process output at time timesteps[-1].
"""
do_classifier_free_guidance = guidance_scale > 1.0
num_steps = (len(timesteps) - num_warmup_steps) // self.scheduler.order
with self.progress_bar(total=num_steps) as progress_bar:
for i, t in enumerate(timesteps):
# expand the latents if we are doing classifier free guidance
latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
########
if i in interval_seq:
prv_features = None
# predict the noise residual
noise_pred, prv_features = self.unet(
latent_model_input,
t,
encoder_hidden_states=prompt_embeds,
cross_attention_kwargs=cross_attention_kwargs,
replicate_prv_feature=prv_features,
quick_replicate= cache_interval>1,
cache_layer_id=cache_layer_id,
cache_block_id=cache_block_id,
return_dict=False,
)
########
# perform guidance
if do_classifier_free_guidance:
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
# compute the previous noisy sample x_t -> x_t-1
latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
# call the callback, if provided
if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
progress_bar.update()
if callback is not None and i % callback_steps == 0:
step_idx = i // getattr(self.scheduler, "order", 1)
callback(step_idx, t, latents)
return latents.clone().detach()
@torch.no_grad()
def __call__(
self,
prompt: Union[str, List[str]],
video_length: Optional[int] = 8,
height: Optional[int] = None,
width: Optional[int] = None,
num_inference_steps: int = 50,
guidance_scale: float = 7.5,
negative_prompt: Optional[Union[str, List[str]]] = None,
num_videos_per_prompt: Optional[int] = 1,
eta: float = 0.0,
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
latents: Optional[torch.FloatTensor] = None,
motion_field_strength_x: float = 12,
motion_field_strength_y: float = 12,
output_type: Optional[str] = "tensor",
return_dict: bool = True,
callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
callback_steps: Optional[int] = 1,
t0: int = 44,
t1: int = 47,
frame_ids: Optional[List[int]] = None,
########
cache_interval: int = 1,
cache_layer_id: int = None,
cache_block_id: int = None,
uniform: bool = True,
pow: float = None,
center: int = None,
output_all_sequence: bool = False,
########
):
"""
The call function to the pipeline for generation.
Args:
prompt (`str` or `List[str]`, *optional*):
The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
video_length (`int`, *optional*, defaults to 8):
The number of generated video frames.
height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
The height in pixels of the generated image.
width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
The width in pixels of the generated image.
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.
guidance_scale (`float`, *optional*, defaults to 7.5):
A higher guidance scale value encourages the model to generate images closely linked to the text
`prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
negative_prompt (`str` or `List[str]`, *optional*):
The prompt or prompts to guide what to not include in video generation. If not defined, you need to
pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
num_videos_per_prompt (`int`, *optional*, defaults to 1):
The number of videos to generate per prompt.
eta (`float`, *optional*, defaults to 0.0):
Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
A [`torch.Generator`](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 video
generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
tensor is generated by sampling using the supplied random `generator`.
output_type (`str`, *optional*, defaults to `"numpy"`):
The output format of the generated video. Choose between `"latent"` and `"numpy"`.
return_dict (`bool`, *optional*, defaults to `True`):
Whether or not to return a
[`~pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.TextToVideoPipelineOutput`] instead of
a plain tuple.
callback (`Callable`, *optional*):
A function that calls every `callback_steps` steps during inference. The function is called with the
following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
callback_steps (`int`, *optional*, defaults to 1):
The frequency at which the `callback` function is called. If not specified, the callback is called at
every step.
motion_field_strength_x (`float`, *optional*, defaults to 12):
Strength of motion in generated video along x-axis. See the [paper](https://arxiv.org/abs/2303.13439),
Sect. 3.3.1.
motion_field_strength_y (`float`, *optional*, defaults to 12):
Strength of motion in generated video along y-axis. See the [paper](https://arxiv.org/abs/2303.13439),
Sect. 3.3.1.
t0 (`int`, *optional*, defaults to 44):
Timestep t0. Should be in the range [0, num_inference_steps - 1]. See the
[paper](https://arxiv.org/abs/2303.13439), Sect. 3.3.1.
t1 (`int`, *optional*, defaults to 47):
Timestep t0. Should be in the range [t0 + 1, num_inference_steps - 1]. See the
[paper](https://arxiv.org/abs/2303.13439), Sect. 3.3.1.
frame_ids (`List[int]`, *optional*):
Indexes of the frames that are being generated. This is used when generating longer videos
chunk-by-chunk.
Returns:
[`~pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.TextToVideoPipelineOutput`]:
The output contains a `ndarray` of the generated video, when `output_type` != `"latent"`, otherwise a
latent code of generated videos and a list of `bool`s indicating whether the corresponding generated
video contains "not-safe-for-work" (nsfw) content..
"""
assert video_length > 0
if frame_ids is None:
frame_ids = list(range(video_length))
assert len(frame_ids) == video_length
assert num_videos_per_prompt == 1
if isinstance(prompt, str):
prompt = [prompt]
if isinstance(negative_prompt, str):
negative_prompt = [negative_prompt]
# Default height and width to unet
height = height or self.unet.config.sample_size * self.vae_scale_factor
width = width or self.unet.config.sample_size * self.vae_scale_factor
# Check inputs. Raise error if not correct
self.check_inputs(prompt, height, width, callback_steps)
# Define call parameters
batch_size = 1 if isinstance(prompt, str) else len(prompt)
device = self._execution_device
# here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
# of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
# corresponds to doing no classifier free guidance.
do_classifier_free_guidance = guidance_scale > 1.0
# Encode input prompt
prompt_embeds = self._encode_prompt(
prompt, device, num_videos_per_prompt, do_classifier_free_guidance, negative_prompt
)
# Prepare timesteps
self.scheduler.set_timesteps(num_inference_steps, device=device)
timesteps = self.scheduler.timesteps
# Prepare latent variables
num_channels_latents = self.unet.config.in_channels
latents = self.prepare_latents(
batch_size * num_videos_per_prompt,
num_channels_latents,
height,
width,
prompt_embeds.dtype,
device,
generator,
latents,
)
# Prepare extra step kwargs.
extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
prv_features = None #record cache feature ****
latents_list = [latents]
if cache_interval == 1:
interval_seq = list(range(num_inference_steps))
else:
if uniform:
interval_seq = list(range(0, num_inference_steps, cache_interval))
else:
num_slow_step = num_inference_steps//cache_interval
if num_inference_steps%cache_interval != 0:
num_slow_step += 1
interval_seq, pow = sample_from_quad_center(num_inference_steps, num_slow_step, center=center, pow=pow)#[0, 3, 6, 9, 12, 16, 22, 28, 35, 43,]
#interval_seq, pow = sample_from_quad(num_inference_steps, num_inference_steps//cache_interval, pow=pow)#[0, 3, 6, 9, 12, 16, 22, 28, 35, 43,]
interval_seq = sorted(interval_seq)
# Perform the first backward process up to time T_1
x_1_t1 = self.backward_loop(
timesteps=timesteps[: -t1 - 1],
prompt_embeds=prompt_embeds,
latents=latents,
guidance_scale=guidance_scale,
callback=callback,
callback_steps=callback_steps,
extra_step_kwargs=extra_step_kwargs,
num_warmup_steps=num_warmup_steps,
prv_features=prv_features,
interval_seq=interval_seq,
cache_interval=cache_interval,
cache_block_id=cache_block_id,
cache_layer_id=cache_layer_id,
)
scheduler_copy = copy.deepcopy(self.scheduler)
# Perform the second backward process up to time T_0
x_1_t0 = self.backward_loop(
timesteps=timesteps[-t1 - 1 : -t0 - 1],
prompt_embeds=prompt_embeds,
latents=x_1_t1,
guidance_scale=guidance_scale,
callback=callback,
callback_steps=callback_steps,
extra_step_kwargs=extra_step_kwargs,
num_warmup_steps=0,
prv_features=prv_features,
interval_seq=interval_seq,
cache_interval=cache_interval,
cache_block_id=cache_block_id,
cache_layer_id=cache_layer_id,
)
# Propagate first frame latents at time T_0 to remaining frames
x_2k_t0 = x_1_t0.repeat(video_length - 1, 1, 1, 1)
# Add motion in latents at time T_0
x_2k_t0 = create_motion_field_and_warp_latents(
motion_field_strength_x=motion_field_strength_x,
motion_field_strength_y=motion_field_strength_y,
latents=x_2k_t0,
frame_ids=frame_ids[1:],
)
# Perform forward process up to time T_1
x_2k_t1 = self.forward_loop(
x_t0=x_2k_t0,
t0=timesteps[-t0 - 1].item(),
t1=timesteps[-t1 - 1].item(),
generator=generator,
)
# Perform backward process from time T_1 to 0
x_1k_t1 = torch.cat([x_1_t1, x_2k_t1])
b, l, d = prompt_embeds.size()
prompt_embeds = prompt_embeds[:, None].repeat(1, video_length, 1, 1).reshape(b * video_length, l, d)
self.scheduler = scheduler_copy
x_1k_0 = self.backward_loop(
timesteps=timesteps[-t1 - 1 :],
prompt_embeds=prompt_embeds,
latents=x_1k_t1,
guidance_scale=guidance_scale,
callback=callback,
callback_steps=callback_steps,
extra_step_kwargs=extra_step_kwargs,
num_warmup_steps=0,
prv_features=prv_features,
interval_seq=interval_seq,
cache_interval=cache_interval,
cache_block_id=cache_block_id,
cache_layer_id=cache_layer_id,
)
latents = x_1k_0
# manually for max memory savings
if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
self.unet.to("cpu")
torch.cuda.empty_cache()
if output_type == "latent":
image = latents
has_nsfw_concept = None
else:
image = self.decode_latents(latents)
# Run safety checker
image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
# Offload all models
self.maybe_free_model_hooks()
if not return_dict:
return (image, has_nsfw_concept)
return TextToVideoPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)

1839
ixformer_sdk/contrib/DeepCache/sd/pipeline_utils.py Normal file
View File

File diff suppressed because it is too large Load Diff

3296
ixformer_sdk/contrib/DeepCache/sd/unet_2d_blocks.py Normal file
View File

File diff suppressed because it is too large Load Diff

1257
ixformer_sdk/contrib/DeepCache/sd/unet_2d_condition.py Normal file
View File

File diff suppressed because it is too large Load Diff