Merge main
This commit is contained in:
bubbliiiing
@@ -0,0 +1,215 @@
|
||||
import os
|
||||
import sys
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
from diffusers import FlowMatchEulerDiscreteScheduler
|
||||
from omegaconf import OmegaConf
|
||||
from PIL import Image
|
||||
|
||||
current_file_path = os.path.abspath(__file__)
|
||||
project_roots = [os.path.dirname(current_file_path), os.path.dirname(os.path.dirname(current_file_path)), os.path.dirname(os.path.dirname(os.path.dirname(current_file_path)))]
|
||||
for project_root in project_roots:
|
||||
sys.path.insert(0, project_root) if project_root not in sys.path else None
|
||||
|
||||
from videox_fun.dist import set_multi_gpus_devices, shard_model
|
||||
from videox_fun.models import (AutoencoderKLLongCatVideo, UMT5EncoderModel, AutoTokenizer,
|
||||
LongCatVideoTransformer3DModel)
|
||||
from videox_fun.models.cache_utils import get_teacache_coefficients
|
||||
from videox_fun.pipeline import LongCatVideoPipeline
|
||||
from videox_fun.utils.fp8_optimization import (convert_model_weight_to_float8, replace_parameters_by_name,
|
||||
convert_weight_dtype_wrapper)
|
||||
from videox_fun.utils.lora_utils import merge_lora, unmerge_lora
|
||||
from videox_fun.utils.utils import (filter_kwargs, get_image_to_video_latent,
|
||||
save_videos_grid)
|
||||
from videox_fun.utils.fm_solvers import FlowDPMSolverMultistepScheduler
|
||||
from videox_fun.utils.fm_solvers_unipc import FlowUniPCMultistepScheduler
|
||||
|
||||
# GPU memory mode, which can be chosen in [model_full_load, model_full_load_and_qfloat8, model_cpu_offload, model_cpu_offload_and_qfloat8, sequential_cpu_offload].
|
||||
# model_full_load means that the entire model will be moved to the GPU.
|
||||
#
|
||||
# model_full_load_and_qfloat8 means that the entire model will be moved to the GPU,
|
||||
# and the transformer model has been quantized to float8, which can save more GPU memory.
|
||||
#
|
||||
# model_cpu_offload means that the entire model will be moved to the CPU after use, which can save some GPU memory.
|
||||
#
|
||||
# model_cpu_offload_and_qfloat8 indicates that the entire model will be moved to the CPU after use,
|
||||
# and the transformer model has been quantized to float8, which can save more GPU memory.
|
||||
#
|
||||
# sequential_cpu_offload means that each layer of the model will be moved to the CPU after use,
|
||||
# resulting in slower speeds but saving a large amount of GPU memory.
|
||||
GPU_memory_mode = "sequential_cpu_offload"
|
||||
# Compile will give a speedup in fixed resolution and need a little GPU memory.
|
||||
# The compile_dit is not compatible with the fsdp_dit and sequential_cpu_offload.
|
||||
compile_dit = False
|
||||
|
||||
# model path
|
||||
model_name = "models/Diffusion_Transformer/LongCat-Video"
|
||||
|
||||
# Choose the sampler in "Flow", "Flow_Unipc", "Flow_DPM++"
|
||||
sampler_name = "Flow"
|
||||
|
||||
# Load pretrained model if need
|
||||
transformer_path = None
|
||||
vae_path = None
|
||||
lora_path = None
|
||||
|
||||
# Other params
|
||||
sample_size = [480, 832]
|
||||
video_length = 81
|
||||
fps = 16
|
||||
|
||||
# Use torch.float16 if GPU does not support torch.bfloat16
|
||||
# ome graphics cards, such as v100, 2080ti, do not support torch.bfloat16
|
||||
weight_dtype = torch.bfloat16
|
||||
validation_image_start = "asset/1.png"
|
||||
|
||||
# Prompt
|
||||
prompt = "The dog is shaking head. The video is of high quality, and the view is very clear. High quality, masterpiece, best quality, highres, ultra-detailed, fantastic."
|
||||
negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"
|
||||
guidance_scale = 4.0
|
||||
seed = 43
|
||||
num_inference_steps = 50
|
||||
lora_weight = 0.55
|
||||
save_path = "samples/longcat-videos-i2v"
|
||||
|
||||
device = set_multi_gpus_devices(1, 1)
|
||||
|
||||
transformer = LongCatVideoTransformer3DModel.from_pretrained(
|
||||
os.path.join(model_name, "dit"),
|
||||
low_cpu_mem_usage=True,
|
||||
torch_dtype=weight_dtype, cp_split_hw=[1, 1]
|
||||
)
|
||||
|
||||
if transformer_path is not None:
|
||||
print(f"From checkpoint: {transformer_path}")
|
||||
if transformer_path.endswith("safetensors"):
|
||||
from safetensors.torch import load_file, safe_open
|
||||
state_dict = load_file(transformer_path)
|
||||
else:
|
||||
state_dict = torch.load(transformer_path, map_location="cpu")
|
||||
state_dict = state_dict["state_dict"] if "state_dict" in state_dict else state_dict
|
||||
|
||||
m, u = transformer.load_state_dict(state_dict, strict=False)
|
||||
print(f"missing keys: {len(m)}, unexpected keys: {len(u)}")
|
||||
|
||||
# Get Vae
|
||||
vae = AutoencoderKLLongCatVideo.from_pretrained(
|
||||
os.path.join(model_name, "vae"),
|
||||
).to(weight_dtype)
|
||||
|
||||
if vae_path is not None:
|
||||
print(f"From checkpoint: {vae_path}")
|
||||
if vae_path.endswith("safetensors"):
|
||||
from safetensors.torch import load_file, safe_open
|
||||
state_dict = load_file(vae_path)
|
||||
else:
|
||||
state_dict = torch.load(vae_path, map_location="cpu")
|
||||
state_dict = state_dict["state_dict"] if "state_dict" in state_dict else state_dict
|
||||
|
||||
m, u = vae.load_state_dict(state_dict, strict=False)
|
||||
print(f"missing keys: {len(m)}, unexpected keys: {len(u)}")
|
||||
|
||||
# Get Tokenizer
|
||||
tokenizer = AutoTokenizer.from_pretrained(
|
||||
os.path.join(model_name, "tokenizer"),
|
||||
)
|
||||
|
||||
# Get Text encoder
|
||||
text_encoder = UMT5EncoderModel.from_pretrained(
|
||||
os.path.join(model_name, "text_encoder"),
|
||||
low_cpu_mem_usage=True,
|
||||
torch_dtype=weight_dtype,
|
||||
)
|
||||
|
||||
# Get Scheduler
|
||||
Chosen_Scheduler = scheduler_dict = {
|
||||
"Flow": FlowMatchEulerDiscreteScheduler,
|
||||
"Flow_Unipc": FlowUniPCMultistepScheduler,
|
||||
"Flow_DPM++": FlowDPMSolverMultistepScheduler,
|
||||
}[sampler_name]
|
||||
scheduler = Chosen_Scheduler.from_pretrained(
|
||||
model_name,
|
||||
subfolder="scheduler"
|
||||
)
|
||||
|
||||
# Get Pipeline
|
||||
pipeline = LongCatVideoPipeline(
|
||||
transformer=transformer,
|
||||
vae=vae,
|
||||
tokenizer=tokenizer,
|
||||
text_encoder=text_encoder,
|
||||
scheduler=scheduler,
|
||||
)
|
||||
|
||||
if compile_dit:
|
||||
for i in range(len(pipeline.transformer.blocks)):
|
||||
pipeline.transformer.blocks[i] = torch.compile(pipeline.transformer.blocks[i])
|
||||
print("Add Compile")
|
||||
|
||||
if GPU_memory_mode == "sequential_cpu_offload":
|
||||
replace_parameters_by_name(transformer, ["modulation",], device=device)
|
||||
transformer.freqs = transformer.freqs.to(device=device)
|
||||
pipeline.enable_sequential_cpu_offload(device=device)
|
||||
elif GPU_memory_mode == "model_cpu_offload_and_qfloat8":
|
||||
convert_model_weight_to_float8(transformer, exclude_module_name=["modulation",], device=device)
|
||||
convert_weight_dtype_wrapper(transformer, weight_dtype)
|
||||
pipeline.enable_model_cpu_offload(device=device)
|
||||
elif GPU_memory_mode == "model_cpu_offload":
|
||||
pipeline.enable_model_cpu_offload(device=device)
|
||||
elif GPU_memory_mode == "model_full_load_and_qfloat8":
|
||||
convert_model_weight_to_float8(transformer, exclude_module_name=["modulation",], device=device)
|
||||
convert_weight_dtype_wrapper(transformer, weight_dtype)
|
||||
pipeline.to(device=device)
|
||||
else:
|
||||
pipeline.to(device=device)
|
||||
|
||||
generator = torch.Generator(device=device).manual_seed(seed)
|
||||
|
||||
if lora_path is not None:
|
||||
pipeline = merge_lora(pipeline, lora_path, lora_weight, device=device, dtype=weight_dtype)
|
||||
|
||||
with torch.no_grad():
|
||||
video_length = int((video_length - 1) // vae.scale_factor_temporal * vae.scale_factor_temporal) + 1 if video_length != 1 else 1
|
||||
latent_frames = (video_length - 1) // vae.scale_factor_temporal + 1
|
||||
|
||||
if validation_image_start is not None:
|
||||
input_video, input_video_mask, clip_image = get_image_to_video_latent(validation_image_start, None, video_length=video_length, sample_size=sample_size)
|
||||
else:
|
||||
input_video, input_video_mask, clip_image = None, None, None
|
||||
|
||||
sample = pipeline(
|
||||
prompt,
|
||||
num_frames = video_length,
|
||||
negative_prompt = negative_prompt,
|
||||
height = sample_size[0],
|
||||
width = sample_size[1],
|
||||
generator = generator,
|
||||
guidance_scale = guidance_scale,
|
||||
num_inference_steps = num_inference_steps,
|
||||
video = input_video,
|
||||
mask_video = input_video_mask,
|
||||
).videos
|
||||
|
||||
if lora_path is not None:
|
||||
pipeline = unmerge_lora(pipeline, lora_path, lora_weight, device=device, dtype=weight_dtype)
|
||||
|
||||
def save_results():
|
||||
if not os.path.exists(save_path):
|
||||
os.makedirs(save_path, exist_ok=True)
|
||||
|
||||
index = len([path for path in os.listdir(save_path)]) + 1
|
||||
prefix = str(index).zfill(8)
|
||||
if video_length == 1:
|
||||
video_path = os.path.join(save_path, prefix + ".png")
|
||||
|
||||
image = sample[0, :, 0]
|
||||
image = image.transpose(0, 1).transpose(1, 2)
|
||||
image = (image * 255).numpy().astype(np.uint8)
|
||||
image = Image.fromarray(image)
|
||||
image.save(video_path)
|
||||
else:
|
||||
video_path = os.path.join(save_path, prefix + ".mp4")
|
||||
save_videos_grid(sample, video_path, fps=fps)
|
||||
|
||||
save_results()
|
||||
@@ -0,0 +1,206 @@
|
||||
import os
|
||||
import sys
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
from diffusers import FlowMatchEulerDiscreteScheduler
|
||||
from omegaconf import OmegaConf
|
||||
from PIL import Image
|
||||
|
||||
current_file_path = os.path.abspath(__file__)
|
||||
project_roots = [os.path.dirname(current_file_path), os.path.dirname(os.path.dirname(current_file_path)), os.path.dirname(os.path.dirname(os.path.dirname(current_file_path)))]
|
||||
for project_root in project_roots:
|
||||
sys.path.insert(0, project_root) if project_root not in sys.path else None
|
||||
|
||||
from videox_fun.dist import set_multi_gpus_devices, shard_model
|
||||
from videox_fun.models import (AutoencoderKLLongCatVideo, UMT5EncoderModel, AutoTokenizer,
|
||||
LongCatVideoTransformer3DModel)
|
||||
from videox_fun.models.cache_utils import get_teacache_coefficients
|
||||
from videox_fun.pipeline import LongCatVideoPipeline
|
||||
from videox_fun.utils.fp8_optimization import (convert_model_weight_to_float8, replace_parameters_by_name,
|
||||
convert_weight_dtype_wrapper)
|
||||
from videox_fun.utils.lora_utils import merge_lora, unmerge_lora
|
||||
from videox_fun.utils.utils import (filter_kwargs, get_image_to_video_latent,
|
||||
save_videos_grid)
|
||||
from videox_fun.utils.fm_solvers import FlowDPMSolverMultistepScheduler
|
||||
from videox_fun.utils.fm_solvers_unipc import FlowUniPCMultistepScheduler
|
||||
|
||||
# GPU memory mode, which can be chosen in [model_full_load, model_full_load_and_qfloat8, model_cpu_offload, model_cpu_offload_and_qfloat8, sequential_cpu_offload].
|
||||
# model_full_load means that the entire model will be moved to the GPU.
|
||||
#
|
||||
# model_full_load_and_qfloat8 means that the entire model will be moved to the GPU,
|
||||
# and the transformer model has been quantized to float8, which can save more GPU memory.
|
||||
#
|
||||
# model_cpu_offload means that the entire model will be moved to the CPU after use, which can save some GPU memory.
|
||||
#
|
||||
# model_cpu_offload_and_qfloat8 indicates that the entire model will be moved to the CPU after use,
|
||||
# and the transformer model has been quantized to float8, which can save more GPU memory.
|
||||
#
|
||||
# sequential_cpu_offload means that each layer of the model will be moved to the CPU after use,
|
||||
# resulting in slower speeds but saving a large amount of GPU memory.
|
||||
GPU_memory_mode = "sequential_cpu_offload"
|
||||
# Compile will give a speedup in fixed resolution and need a little GPU memory.
|
||||
# The compile_dit is not compatible with the fsdp_dit and sequential_cpu_offload.
|
||||
compile_dit = False
|
||||
|
||||
# model path
|
||||
model_name = "models/Diffusion_Transformer/LongCat-Video"
|
||||
|
||||
# Choose the sampler in "Flow", "Flow_Unipc", "Flow_DPM++"
|
||||
sampler_name = "Flow"
|
||||
|
||||
# Load pretrained model if need
|
||||
transformer_path = None
|
||||
vae_path = None
|
||||
lora_path = None
|
||||
|
||||
# Other params
|
||||
sample_size = [832, 480]
|
||||
video_length = 81
|
||||
fps = 16
|
||||
|
||||
# Use torch.float16 if GPU does not support torch.bfloat16
|
||||
# ome graphics cards, such as v100, 2080ti, do not support torch.bfloat16
|
||||
weight_dtype = torch.bfloat16
|
||||
# Prompt
|
||||
prompt = "1girl, black_hair, brown_eyes, earrings, freckles, grey_background, jewelry, lips, long_hair, looking_at_viewer, nose, piercing, realistic, red_lips, solo, upper_body"
|
||||
negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"
|
||||
guidance_scale = 4.0
|
||||
seed = 43
|
||||
num_inference_steps = 50
|
||||
lora_weight = 0.55
|
||||
save_path = "samples/longcat-videos-t2v"
|
||||
|
||||
device = set_multi_gpus_devices(1, 1)
|
||||
|
||||
transformer = LongCatVideoTransformer3DModel.from_pretrained(
|
||||
os.path.join(model_name, "dit"),
|
||||
low_cpu_mem_usage=True,
|
||||
torch_dtype=weight_dtype, cp_split_hw=[1, 1]
|
||||
)
|
||||
|
||||
if transformer_path is not None:
|
||||
print(f"From checkpoint: {transformer_path}")
|
||||
if transformer_path.endswith("safetensors"):
|
||||
from safetensors.torch import load_file, safe_open
|
||||
state_dict = load_file(transformer_path)
|
||||
else:
|
||||
state_dict = torch.load(transformer_path, map_location="cpu")
|
||||
state_dict = state_dict["state_dict"] if "state_dict" in state_dict else state_dict
|
||||
|
||||
m, u = transformer.load_state_dict(state_dict, strict=False)
|
||||
print(f"missing keys: {len(m)}, unexpected keys: {len(u)}")
|
||||
|
||||
# Get Vae
|
||||
vae = AutoencoderKLLongCatVideo.from_pretrained(
|
||||
os.path.join(model_name, "vae"),
|
||||
).to(weight_dtype)
|
||||
|
||||
if vae_path is not None:
|
||||
print(f"From checkpoint: {vae_path}")
|
||||
if vae_path.endswith("safetensors"):
|
||||
from safetensors.torch import load_file, safe_open
|
||||
state_dict = load_file(vae_path)
|
||||
else:
|
||||
state_dict = torch.load(vae_path, map_location="cpu")
|
||||
state_dict = state_dict["state_dict"] if "state_dict" in state_dict else state_dict
|
||||
|
||||
m, u = vae.load_state_dict(state_dict, strict=False)
|
||||
print(f"missing keys: {len(m)}, unexpected keys: {len(u)}")
|
||||
|
||||
# Get Tokenizer
|
||||
tokenizer = AutoTokenizer.from_pretrained(
|
||||
os.path.join(model_name, "tokenizer"),
|
||||
)
|
||||
|
||||
# Get Text encoder
|
||||
text_encoder = UMT5EncoderModel.from_pretrained(
|
||||
os.path.join(model_name, "text_encoder"),
|
||||
low_cpu_mem_usage=True,
|
||||
torch_dtype=weight_dtype,
|
||||
)
|
||||
|
||||
# Get Scheduler
|
||||
Chosen_Scheduler = scheduler_dict = {
|
||||
"Flow": FlowMatchEulerDiscreteScheduler,
|
||||
"Flow_Unipc": FlowUniPCMultistepScheduler,
|
||||
"Flow_DPM++": FlowDPMSolverMultistepScheduler,
|
||||
}[sampler_name]
|
||||
scheduler = Chosen_Scheduler.from_pretrained(
|
||||
model_name,
|
||||
subfolder="scheduler"
|
||||
)
|
||||
|
||||
# Get Pipeline
|
||||
pipeline = LongCatVideoPipeline(
|
||||
transformer=transformer,
|
||||
vae=vae,
|
||||
tokenizer=tokenizer,
|
||||
text_encoder=text_encoder,
|
||||
scheduler=scheduler,
|
||||
)
|
||||
|
||||
if compile_dit:
|
||||
for i in range(len(pipeline.transformer.blocks)):
|
||||
pipeline.transformer.blocks[i] = torch.compile(pipeline.transformer.blocks[i])
|
||||
print("Add Compile")
|
||||
|
||||
if GPU_memory_mode == "sequential_cpu_offload":
|
||||
replace_parameters_by_name(transformer, ["modulation",], device=device)
|
||||
transformer.freqs = transformer.freqs.to(device=device)
|
||||
pipeline.enable_sequential_cpu_offload(device=device)
|
||||
elif GPU_memory_mode == "model_cpu_offload_and_qfloat8":
|
||||
convert_model_weight_to_float8(transformer, exclude_module_name=["modulation",], device=device)
|
||||
convert_weight_dtype_wrapper(transformer, weight_dtype)
|
||||
pipeline.enable_model_cpu_offload(device=device)
|
||||
elif GPU_memory_mode == "model_cpu_offload":
|
||||
pipeline.enable_model_cpu_offload(device=device)
|
||||
elif GPU_memory_mode == "model_full_load_and_qfloat8":
|
||||
convert_model_weight_to_float8(transformer, exclude_module_name=["modulation",], device=device)
|
||||
convert_weight_dtype_wrapper(transformer, weight_dtype)
|
||||
pipeline.to(device=device)
|
||||
else:
|
||||
pipeline.to(device=device)
|
||||
|
||||
generator = torch.Generator(device=device).manual_seed(seed)
|
||||
|
||||
if lora_path is not None:
|
||||
pipeline = merge_lora(pipeline, lora_path, lora_weight, device=device, dtype=weight_dtype)
|
||||
|
||||
with torch.no_grad():
|
||||
video_length = int((video_length - 1) // vae.scale_factor_temporal * vae.scale_factor_temporal) + 1 if video_length != 1 else 1
|
||||
latent_frames = (video_length - 1) // vae.scale_factor_temporal + 1
|
||||
|
||||
sample = pipeline(
|
||||
prompt,
|
||||
num_frames = video_length,
|
||||
negative_prompt = negative_prompt,
|
||||
height = sample_size[0],
|
||||
width = sample_size[1],
|
||||
generator = generator,
|
||||
guidance_scale = guidance_scale,
|
||||
num_inference_steps = num_inference_steps,
|
||||
).videos
|
||||
|
||||
if lora_path is not None:
|
||||
pipeline = unmerge_lora(pipeline, lora_path, lora_weight, device=device, dtype=weight_dtype)
|
||||
|
||||
def save_results():
|
||||
if not os.path.exists(save_path):
|
||||
os.makedirs(save_path, exist_ok=True)
|
||||
|
||||
index = len([path for path in os.listdir(save_path)]) + 1
|
||||
prefix = str(index).zfill(8)
|
||||
if video_length == 1:
|
||||
video_path = os.path.join(save_path, prefix + ".png")
|
||||
|
||||
image = sample[0, :, 0]
|
||||
image = image.transpose(0, 1).transpose(1, 2)
|
||||
image = (image * 255).numpy().astype(np.uint8)
|
||||
image = Image.fromarray(image)
|
||||
image.save(video_path)
|
||||
else:
|
||||
video_path = os.path.join(save_path, prefix + ".mp4")
|
||||
save_videos_grid(sample, video_path, fps=fps)
|
||||
|
||||
save_results()
|
||||
@@ -0,0 +1,238 @@
|
||||
## Training Code
|
||||
|
||||
The default training commands for the different versions are as follows:
|
||||
|
||||
We can choose whether to use DeepSpeed and FSDP in LongCatVideo, which can save a lot of video memory.
|
||||
|
||||
Some parameters in the sh file can be confusing, and they are explained in this document:
|
||||
|
||||
- `enable_bucket` is used to enable bucket training. When enabled, the model does not crop the images and videos at the center, but instead, it trains the entire images and videos after grouping them into buckets based on resolution.
|
||||
- Sample size Configuration Guide
|
||||
- `video_sample_size` represents the resolution size of videos; when `random_hw_adapt` is True, it represents the minimum value between video and image resolutions.
|
||||
- `image_sample_size` represents the resolution size of images; when `random_hw_adapt` is True, it represents the maximum value between video and image resolutions.
|
||||
- `token_sample_size` represents the resolution corresponding to the maximum token length when `training_with_video_token_length` is True.
|
||||
- Due to potential confusion in configuration, **if you don't require arbitrary resolution for finetuning**, it is recommended to set `video_sample_size`, `image_sample_size`, and `token_sample_size` to the same fixed value, such as **(320, 480, 512, 640, 960)**.
|
||||
- **All set to 320** represents **240P**.
|
||||
- **All set to 480** represents **320P**.
|
||||
- **All set to 640** represents **480P**.
|
||||
- **All set to 960** represents **720P**.
|
||||
- `random_frame_crop` is used for random cropping on video frames to simulate videos with different frame counts.
|
||||
- `random_hw_adapt` is used to enable automatic height and width scaling for images and videos. When `random_hw_adapt` is enabled, the training images will have their height and width set to `image_sample_size` as the maximum and `min(video_sample_size, 512)` as the minimum. For training videos, the height and width will be set to `image_sample_size` as the maximum and `min(video_sample_size, 512)` as the minimum.
|
||||
- For example, when `random_hw_adapt` is enabled, with `video_sample_n_frames=49`, `video_sample_size=1024`, and `image_sample_size=1024`, the resolution of image inputs for training is `512x512` to `1024x1024`, and the resolution of video inputs for training is `512x512x49` to `1024x1024x49`.
|
||||
- For example, when `random_hw_adapt` is enabled, with `video_sample_n_frames=49`, `video_sample_size=256`, and `image_sample_size=1024`, the resolution of image inputs for training is `256x256` to `1024x1024`, and the resolution of video inputs for training is `256x256x49`.
|
||||
- `training_with_video_token_length` specifies training the model according to token length. For training images and videos, the height and width will be set to `image_sample_size` as the maximum and `video_sample_size` as the minimum.
|
||||
- For example, when `training_with_video_token_length` is enabled, with `video_sample_n_frames=49`, `token_sample_size=1024`, `video_sample_size=256`, and `image_sample_size=1024`, the resolution of image inputs for training is `256x256` to `1024x1024`, and the resolution of video inputs for training is `256x256x49` to `1024x1024x49`.
|
||||
- For example, when `training_with_video_token_length` is enabled, with `video_sample_n_frames=49`, `token_sample_size=512`, `video_sample_size=256`, and `image_sample_size=1024`, the resolution of image inputs for training is `256x256` to `1024x1024`, and the resolution of video inputs for training is `256x256x49` to `1024x1024x9`.
|
||||
- The token length for a video with dimensions 512x512 and 49 frames is 13,312. We need to set the `token_sample_size = 512`.
|
||||
- At 512x512 resolution, the number of video frames is 49 (~= 512 * 512 * 49 / 512 / 512).
|
||||
- At 768x768 resolution, the number of video frames is 21 (~= 512 * 512 * 49 / 768 / 768).
|
||||
- At 1024x1024 resolution, the number of video frames is 9 (~= 512 * 512 * 49 / 1024 / 1024).
|
||||
- These resolutions combined with their corresponding lengths allow the model to generate videos of different sizes.
|
||||
- `train_mode` is used to specify the training mode, which can be either normal or i2v.
|
||||
- `resume_from_checkpoint` is used to set the training should be resumed from a previous checkpoint. Use a path or `"latest"` to automatically select the last available checkpoint.
|
||||
|
||||
When train model with multi machines, please set the params as follows:
|
||||
```sh
|
||||
export MASTER_ADDR="your master address"
|
||||
export MASTER_PORT=10086
|
||||
export WORLD_SIZE=1 # The number of machines
|
||||
export NUM_PROCESS=8 # The number of processes, such as WORLD_SIZE * 8
|
||||
export RANK=0 # The rank of this machine
|
||||
|
||||
accelerate launch --mixed_precision="bf16" --main_process_ip=$MASTER_ADDR --main_process_port=$MASTER_PORT --num_machines=$WORLD_SIZE --num_processes=$NUM_PROCESS --machine_rank=$RANK scripts/xxx/xxx.py
|
||||
```
|
||||
|
||||
LongCatVideo without deepspeed:
|
||||
|
||||
```sh
|
||||
export MODEL_NAME="models/Diffusion_Transformer/LongCat-Video"
|
||||
export DATASET_NAME="datasets/internal_datasets/"
|
||||
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
|
||||
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
|
||||
# export NCCL_IB_DISABLE=1
|
||||
# export NCCL_P2P_DISABLE=1
|
||||
NCCL_DEBUG=INFO
|
||||
|
||||
accelerate launch --mixed_precision="bf16" scripts/longcatvideo/train.py \
|
||||
--pretrained_model_name_or_path=$MODEL_NAME \
|
||||
--train_data_dir=$DATASET_NAME \
|
||||
--train_data_meta=$DATASET_META_NAME \
|
||||
--image_sample_size=640 \
|
||||
--video_sample_size=640 \
|
||||
--token_sample_size=640 \
|
||||
--video_sample_stride=2 \
|
||||
--video_sample_n_frames=81 \
|
||||
--train_batch_size=1 \
|
||||
--video_repeat=1 \
|
||||
--gradient_accumulation_steps=1 \
|
||||
--dataloader_num_workers=8 \
|
||||
--num_train_epochs=100 \
|
||||
--checkpointing_steps=50 \
|
||||
--learning_rate=2e-05 \
|
||||
--lr_scheduler="constant_with_warmup" \
|
||||
--lr_warmup_steps=100 \
|
||||
--seed=42 \
|
||||
--output_dir="output_dir_longcat_full_finetune" \
|
||||
--gradient_checkpointing \
|
||||
--mixed_precision="bf16" \
|
||||
--adam_weight_decay=3e-2 \
|
||||
--adam_epsilon=1e-10 \
|
||||
--vae_mini_batch=1 \
|
||||
--max_grad_norm=0.05 \
|
||||
--random_hw_adapt \
|
||||
--training_with_video_token_length \
|
||||
--enable_bucket \
|
||||
--uniform_sampling \
|
||||
--low_vram \
|
||||
--train_mode="normal" \
|
||||
--trainable_modules "."
|
||||
|
||||
```
|
||||
|
||||
LongCatVideo with Deepspeed Zero-2:
|
||||
|
||||
```sh
|
||||
export MODEL_NAME="models/Diffusion_Transformer/LongCat-Video"
|
||||
export DATASET_NAME="datasets/internal_datasets/"
|
||||
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
|
||||
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
|
||||
# export NCCL_IB_DISABLE=1
|
||||
# export NCCL_P2P_DISABLE=1
|
||||
NCCL_DEBUG=INFO
|
||||
|
||||
accelerate launch --use_deepspeed --deepspeed_config_file config/zero_stage2_config.json --deepspeed_multinode_launcher standard scripts/longcatvideo/train.py \
|
||||
--pretrained_model_name_or_path=$MODEL_NAME \
|
||||
--train_data_dir=$DATASET_NAME \
|
||||
--train_data_meta=$DATASET_META_NAME \
|
||||
--image_sample_size=640 \
|
||||
--video_sample_size=640 \
|
||||
--token_sample_size=640 \
|
||||
--video_sample_stride=2 \
|
||||
--video_sample_n_frames=81 \
|
||||
--train_batch_size=1 \
|
||||
--video_repeat=1 \
|
||||
--gradient_accumulation_steps=1 \
|
||||
--dataloader_num_workers=8 \
|
||||
--num_train_epochs=100 \
|
||||
--checkpointing_steps=50 \
|
||||
--learning_rate=2e-05 \
|
||||
--lr_scheduler="constant_with_warmup" \
|
||||
--lr_warmup_steps=100 \
|
||||
--seed=42 \
|
||||
--output_dir="output_dir_longcat_full_finetune" \
|
||||
--gradient_checkpointing \
|
||||
--mixed_precision="bf16" \
|
||||
--adam_weight_decay=3e-2 \
|
||||
--adam_epsilon=1e-10 \
|
||||
--vae_mini_batch=1 \
|
||||
--max_grad_norm=0.05 \
|
||||
--random_hw_adapt \
|
||||
--training_with_video_token_length \
|
||||
--enable_bucket \
|
||||
--uniform_sampling \
|
||||
--low_vram \
|
||||
--train_mode="normal" \
|
||||
--trainable_modules "."
|
||||
|
||||
```
|
||||
|
||||
DeepSpeed Zero-3 is not highly recommended at the moment. In this repository, using FSDP has fewer errors and is more stable.
|
||||
|
||||
LongCatVideo with DeepSpeed Zero-3:
|
||||
|
||||
```sh
|
||||
python scripts/zero_to_bf16.py output_dir/checkpoint-{our-num-steps} output_dir/checkpoint-{your-num-steps}-outputs --max_shard_size 80GB --safe_serialization
|
||||
```
|
||||
|
||||
Training shell command is as follows:
|
||||
```sh
|
||||
export MODEL_NAME="models/Diffusion_Transformer/LongCat-Video"
|
||||
export DATASET_NAME="datasets/internal_datasets/"
|
||||
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
|
||||
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
|
||||
# export NCCL_IB_DISABLE=1
|
||||
# export NCCL_P2P_DISABLE=1
|
||||
NCCL_DEBUG=INFO
|
||||
|
||||
accelerate launch --zero_stage 3 --zero3_save_16bit_model true --zero3_init_flag true --use_deepspeed --deepspeed_config_file config/zero_stage3_config.json --deepspeed_multinode_launcher standard scripts/longcatvideo/train.py \
|
||||
--pretrained_model_name_or_path=$MODEL_NAME \
|
||||
--train_data_dir=$DATASET_NAME \
|
||||
--train_data_meta=$DATASET_META_NAME \
|
||||
--image_sample_size=640 \
|
||||
--video_sample_size=640 \
|
||||
--token_sample_size=640 \
|
||||
--video_sample_stride=2 \
|
||||
--video_sample_n_frames=81 \
|
||||
--train_batch_size=1 \
|
||||
--video_repeat=1 \
|
||||
--gradient_accumulation_steps=1 \
|
||||
--dataloader_num_workers=8 \
|
||||
--num_train_epochs=100 \
|
||||
--checkpointing_steps=50 \
|
||||
--learning_rate=2e-05 \
|
||||
--lr_scheduler="constant_with_warmup" \
|
||||
--lr_warmup_steps=100 \
|
||||
--seed=42 \
|
||||
--output_dir="output_dir_longcat_full_finetune" \
|
||||
--gradient_checkpointing \
|
||||
--mixed_precision="bf16" \
|
||||
--adam_weight_decay=3e-2 \
|
||||
--adam_epsilon=1e-10 \
|
||||
--vae_mini_batch=1 \
|
||||
--max_grad_norm=0.05 \
|
||||
--random_hw_adapt \
|
||||
--training_with_video_token_length \
|
||||
--enable_bucket \
|
||||
--uniform_sampling \
|
||||
--low_vram \
|
||||
--train_mode="normal" \
|
||||
--trainable_modules "."
|
||||
|
||||
```
|
||||
|
||||
LongCatVideo with FSDP:
|
||||
|
||||
```sh
|
||||
export MODEL_NAME="models/Diffusion_Transformer/LongCat-Video"
|
||||
export DATASET_NAME="datasets/internal_datasets/"
|
||||
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
|
||||
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
|
||||
# export NCCL_IB_DISABLE=1
|
||||
# export NCCL_P2P_DISABLE=1
|
||||
NCCL_DEBUG=INFO
|
||||
|
||||
accelerate launch --mixed_precision="bf16" --use_fsdp --fsdp_auto_wrap_policy TRANSFORMER_BASED_WRAP --fsdp_transformer_layer_cls_to_wrap=LongCatSingleStreamBlock --fsdp_sharding_strategy "FULL_SHARD" --fsdp_state_dict_type=SHARDED_STATE_DICT --fsdp_backward_prefetch "BACKWARD_PRE" --fsdp_cpu_ram_efficient_loading False scripts/longcatvideo/train.py \
|
||||
--pretrained_model_name_or_path=$MODEL_NAME \
|
||||
--train_data_dir=$DATASET_NAME \
|
||||
--train_data_meta=$DATASET_META_NAME \
|
||||
--image_sample_size=640 \
|
||||
--video_sample_size=640 \
|
||||
--token_sample_size=640 \
|
||||
--video_sample_stride=2 \
|
||||
--video_sample_n_frames=81 \
|
||||
--train_batch_size=1 \
|
||||
--video_repeat=1 \
|
||||
--gradient_accumulation_steps=1 \
|
||||
--dataloader_num_workers=8 \
|
||||
--num_train_epochs=100 \
|
||||
--checkpointing_steps=50 \
|
||||
--learning_rate=2e-05 \
|
||||
--lr_scheduler="constant_with_warmup" \
|
||||
--lr_warmup_steps=100 \
|
||||
--seed=42 \
|
||||
--output_dir="output_dir_longcat_full_finetune" \
|
||||
--gradient_checkpointing \
|
||||
--mixed_precision="bf16" \
|
||||
--adam_weight_decay=3e-2 \
|
||||
--adam_epsilon=1e-10 \
|
||||
--vae_mini_batch=1 \
|
||||
--max_grad_norm=0.05 \
|
||||
--random_hw_adapt \
|
||||
--training_with_video_token_length \
|
||||
--enable_bucket \
|
||||
--uniform_sampling \
|
||||
--low_vram \
|
||||
--train_mode="normal" \
|
||||
--trainable_modules "."
|
||||
|
||||
```
|
||||
@@ -0,0 +1,239 @@
|
||||
## Lora Training Code
|
||||
|
||||
We can choose whether to use DeepSpeed and FSDP in LongCatVideo, which can save a lot of video memory.
|
||||
|
||||
Some parameters in the sh file can be confusing, and they are explained in this document:
|
||||
|
||||
- `enable_bucket` is used to enable bucket training. When enabled, the model does not crop the images and videos at the center, but instead, it trains the entire images and videos after grouping them into buckets based on resolution.
|
||||
- Sample size Configuration Guide
|
||||
- `video_sample_size` represents the resolution size of videos; when `random_hw_adapt` is True, it represents the minimum value between video and image resolutions.
|
||||
- `image_sample_size` represents the resolution size of images; when `random_hw_adapt` is True, it represents the maximum value between video and image resolutions.
|
||||
- `token_sample_size` represents the resolution corresponding to the maximum token length when `training_with_video_token_length` is True.
|
||||
- Due to potential confusion in configuration, **if you don't require arbitrary resolution for finetuning**, it is recommended to set `video_sample_size`, `image_sample_size`, and `token_sample_size` to the same fixed value, such as **(320, 480, 512, 640, 960)**.
|
||||
- **All set to 320** represents **240P**.
|
||||
- **All set to 480** represents **320P**.
|
||||
- **All set to 640** represents **480P**.
|
||||
- **All set to 960** represents **720P**.
|
||||
- `random_frame_crop` is used for random cropping on video frames to simulate videos with different frame counts.
|
||||
- `random_hw_adapt` is used to enable automatic height and width scaling for images and videos. When `random_hw_adapt` is enabled, the training images will have their height and width set to `image_sample_size` as the maximum and `min(video_sample_size, 512)` as the minimum. For training videos, the height and width will be set to `image_sample_size` as the maximum and `min(video_sample_size, 512)` as the minimum.
|
||||
- For example, when `random_hw_adapt` is enabled, with `video_sample_n_frames=49`, `video_sample_size=1024`, and `image_sample_size=1024`, the resolution of image inputs for training is `512x512` to `1024x1024`, and the resolution of video inputs for training is `512x512x49` to `1024x1024x49`.
|
||||
- For example, when `random_hw_adapt` is enabled, with `video_sample_n_frames=49`, `video_sample_size=256`, and `image_sample_size=1024`, the resolution of image inputs for training is `256x256` to `1024x1024`, and the resolution of video inputs for training is `256x256x49`.
|
||||
- `training_with_video_token_length` specifies training the model according to token length. For training images and videos, the height and width will be set to `image_sample_size` as the maximum and `video_sample_size` as the minimum.
|
||||
- For example, when `training_with_video_token_length` is enabled, with `video_sample_n_frames=49`, `token_sample_size=1024`, `video_sample_size=256`, and `image_sample_size=1024`, the resolution of image inputs for training is `256x256` to `1024x1024`, and the resolution of video inputs for training is `256x256x49` to `1024x1024x49`.
|
||||
- For example, when `training_with_video_token_length` is enabled, with `video_sample_n_frames=49`, `token_sample_size=512`, `video_sample_size=256`, and `image_sample_size=1024`, the resolution of image inputs for training is `256x256` to `1024x1024`, and the resolution of video inputs for training is `256x256x49` to `1024x1024x9`.
|
||||
- The token length for a video with dimensions 512x512 and 49 frames is 13,312. We need to set the `token_sample_size = 512`.
|
||||
- At 512x512 resolution, the number of video frames is 49 (~= 512 * 512 * 49 / 512 / 512).
|
||||
- At 768x768 resolution, the number of video frames is 21 (~= 512 * 512 * 49 / 768 / 768).
|
||||
- At 1024x1024 resolution, the number of video frames is 9 (~= 512 * 512 * 49 / 1024 / 1024).
|
||||
- These resolutions combined with their corresponding lengths allow the model to generate videos of different sizes.
|
||||
- `train_mode` is used to specify the training mode, which can be either normal or i2v.
|
||||
- `resume_from_checkpoint` is used to set the training should be resumed from a previous checkpoint. Use a path or `"latest"` to automatically select the last available checkpoint.
|
||||
- `target_name` represents the components/modules to which LoRA will be applied, separated by commas.
|
||||
- `use_peft_lora` indicates whether to use the PEFT module for adding LoRA. Using this module will be more memory-efficient.
|
||||
- `rank` means the dimension of the LoRA update matrices.
|
||||
- `network_alpha` means the scale of the LoRA update matrices.
|
||||
|
||||
When train model with multi machines, please set the params as follows:
|
||||
```sh
|
||||
export MASTER_ADDR="your master address"
|
||||
export MASTER_PORT=10086
|
||||
export WORLD_SIZE=1 # The number of machines
|
||||
export NUM_PROCESS=8 # The number of processes, such as WORLD_SIZE * 8
|
||||
export RANK=0 # The rank of this machine
|
||||
|
||||
accelerate launch --mixed_precision="bf16" --main_process_ip=$MASTER_ADDR --main_process_port=$MASTER_PORT --num_machines=$WORLD_SIZE --num_processes=$NUM_PROCESS --machine_rank=$RANK scripts/xxx/xxx.py
|
||||
```
|
||||
|
||||
LongCatVideo without deepspeed:
|
||||
|
||||
```sh
|
||||
export MODEL_NAME="models/Diffusion_Transformer/LongCat-Video"
|
||||
export DATASET_NAME="datasets/internal_datasets/"
|
||||
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
|
||||
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
|
||||
# export NCCL_IB_DISABLE=1
|
||||
# export NCCL_P2P_DISABLE=1
|
||||
NCCL_DEBUG=INFO
|
||||
|
||||
accelerate launch --mixed_precision="bf16" scripts/longcatvideo/train_lora.py \
|
||||
--pretrained_model_name_or_path=$MODEL_NAME \
|
||||
--train_data_dir=$DATASET_NAME \
|
||||
--train_data_meta=$DATASET_META_NAME \
|
||||
--image_sample_size=640 \
|
||||
--video_sample_size=640 \
|
||||
--token_sample_size=640 \
|
||||
--video_sample_stride=2 \
|
||||
--video_sample_n_frames=81 \
|
||||
--train_batch_size=1 \
|
||||
--video_repeat=1 \
|
||||
--gradient_accumulation_steps=1 \
|
||||
--dataloader_num_workers=8 \
|
||||
--num_train_epochs=100 \
|
||||
--checkpointing_steps=50 \
|
||||
--learning_rate=1e-04 \
|
||||
--seed=42 \
|
||||
--output_dir="output_dir_longcat_lora" \
|
||||
--gradient_checkpointing \
|
||||
--mixed_precision="bf16" \
|
||||
--adam_weight_decay=3e-2 \
|
||||
--adam_epsilon=1e-10 \
|
||||
--vae_mini_batch=1 \
|
||||
--max_grad_norm=0.05 \
|
||||
--random_hw_adapt \
|
||||
--training_with_video_token_length \
|
||||
--enable_bucket \
|
||||
--uniform_sampling \
|
||||
--rank=64 \
|
||||
--network_alpha=32 \
|
||||
--target_name="qkv,q_linear,kv_linear,ffn.w1,ffn.w2,ffn.w3" \
|
||||
--use_peft_lora \
|
||||
--train_mode="normal" \
|
||||
--low_vram
|
||||
```
|
||||
|
||||
LongCatVideo with Deepspeed Zero-2:
|
||||
|
||||
```sh
|
||||
export MODEL_NAME="models/Diffusion_Transformer/LongCat-Video"
|
||||
export DATASET_NAME="datasets/internal_datasets/"
|
||||
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
|
||||
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
|
||||
# export NCCL_IB_DISABLE=1
|
||||
# export NCCL_P2P_DISABLE=1
|
||||
NCCL_DEBUG=INFO
|
||||
|
||||
accelerate launch --use_deepspeed --deepspeed_config_file config/zero_stage2_config.json --deepspeed_multinode_launcher standard scripts/longcatvideo/train_lora.py \
|
||||
--pretrained_model_name_or_path=$MODEL_NAME \
|
||||
--train_data_dir=$DATASET_NAME \
|
||||
--train_data_meta=$DATASET_META_NAME \
|
||||
--image_sample_size=640 \
|
||||
--video_sample_size=640 \
|
||||
--token_sample_size=640 \
|
||||
--video_sample_stride=2 \
|
||||
--video_sample_n_frames=81 \
|
||||
--train_batch_size=1 \
|
||||
--video_repeat=1 \
|
||||
--gradient_accumulation_steps=1 \
|
||||
--dataloader_num_workers=8 \
|
||||
--num_train_epochs=100 \
|
||||
--checkpointing_steps=50 \
|
||||
--learning_rate=1e-04 \
|
||||
--seed=42 \
|
||||
--output_dir="output_dir_longcat_lora" \
|
||||
--gradient_checkpointing \
|
||||
--mixed_precision="bf16" \
|
||||
--adam_weight_decay=3e-2 \
|
||||
--adam_epsilon=1e-10 \
|
||||
--vae_mini_batch=1 \
|
||||
--max_grad_norm=0.05 \
|
||||
--random_hw_adapt \
|
||||
--training_with_video_token_length \
|
||||
--enable_bucket \
|
||||
--uniform_sampling \
|
||||
--rank=64 \
|
||||
--network_alpha=32 \
|
||||
--target_name="qkv,q_linear,kv_linear,ffn.w1,ffn.w2,ffn.w3" \
|
||||
--use_peft_lora \
|
||||
--train_mode="normal" \
|
||||
--low_vram
|
||||
```
|
||||
|
||||
DeepSpeed Zero-3 is not highly recommended at the moment. In this repository, using FSDP has fewer errors and is more stable.
|
||||
|
||||
It is known that DeepSpeed Zero-3 is not compatible with PEFT.
|
||||
|
||||
```sh
|
||||
python scripts/zero_to_bf16.py output_dir/checkpoint-{our-num-steps} output_dir/checkpoint-{your-num-steps}-outputs --max_shard_size 80GB --safe_serialization
|
||||
```
|
||||
|
||||
Training shell command is as follows:
|
||||
```sh
|
||||
export MODEL_NAME="models/Diffusion_Transformer/LongCat-Video"
|
||||
export DATASET_NAME="datasets/internal_datasets/"
|
||||
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
|
||||
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
|
||||
# export NCCL_IB_DISABLE=1
|
||||
# export NCCL_P2P_DISABLE=1
|
||||
NCCL_DEBUG=INFO
|
||||
|
||||
accelerate launch --zero_stage 3 --zero3_save_16bit_model true --zero3_init_flag true --use_deepspeed --deepspeed_config_file config/zero_stage3_config.json --deepspeed_multinode_launcher standard scripts/longcatvideo/train_lora.py \
|
||||
--pretrained_model_name_or_path=$MODEL_NAME \
|
||||
--train_data_dir=$DATASET_NAME \
|
||||
--train_data_meta=$DATASET_META_NAME \
|
||||
--image_sample_size=640 \
|
||||
--video_sample_size=640 \
|
||||
--token_sample_size=640 \
|
||||
--video_sample_stride=2 \
|
||||
--video_sample_n_frames=81 \
|
||||
--train_batch_size=1 \
|
||||
--video_repeat=1 \
|
||||
--gradient_accumulation_steps=1 \
|
||||
--dataloader_num_workers=8 \
|
||||
--num_train_epochs=100 \
|
||||
--checkpointing_steps=50 \
|
||||
--learning_rate=1e-04 \
|
||||
--seed=42 \
|
||||
--output_dir="output_dir_longcat_lora" \
|
||||
--gradient_checkpointing \
|
||||
--mixed_precision="bf16" \
|
||||
--adam_weight_decay=3e-2 \
|
||||
--adam_epsilon=1e-10 \
|
||||
--vae_mini_batch=1 \
|
||||
--max_grad_norm=0.05 \
|
||||
--random_hw_adapt \
|
||||
--training_with_video_token_length \
|
||||
--enable_bucket \
|
||||
--uniform_sampling \
|
||||
--rank=64 \
|
||||
--network_alpha=32 \
|
||||
--target_name="qkv,q_linear,kv_linear,ffn.w1,ffn.w2,ffn.w3" \
|
||||
--train_mode="normal" \
|
||||
--low_vram
|
||||
```
|
||||
|
||||
LongCatVideo with FSDP:
|
||||
|
||||
```sh
|
||||
export MODEL_NAME="models/Diffusion_Transformer/LongCat-Video"
|
||||
export DATASET_NAME="datasets/internal_datasets/"
|
||||
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
|
||||
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
|
||||
# export NCCL_IB_DISABLE=1
|
||||
# export NCCL_P2P_DISABLE=1
|
||||
NCCL_DEBUG=INFO
|
||||
|
||||
accelerate launch --mixed_precision="bf16" --use_fsdp --fsdp_auto_wrap_policy TRANSFORMER_BASED_WRAP --fsdp_transformer_layer_cls_to_wrap=LongCatSingleStreamBlock --fsdp_sharding_strategy "FULL_SHARD" --fsdp_state_dict_type=SHARDED_STATE_DICT --fsdp_backward_prefetch "BACKWARD_PRE" --fsdp_cpu_ram_efficient_loading False scripts/longcatvideo/train_lora.py \
|
||||
--pretrained_model_name_or_path=$MODEL_NAME \
|
||||
--train_data_dir=$DATASET_NAME \
|
||||
--train_data_meta=$DATASET_META_NAME \
|
||||
--image_sample_size=640 \
|
||||
--video_sample_size=640 \
|
||||
--token_sample_size=640 \
|
||||
--video_sample_stride=2 \
|
||||
--video_sample_n_frames=81 \
|
||||
--train_batch_size=1 \
|
||||
--video_repeat=1 \
|
||||
--gradient_accumulation_steps=1 \
|
||||
--dataloader_num_workers=8 \
|
||||
--num_train_epochs=100 \
|
||||
--checkpointing_steps=50 \
|
||||
--learning_rate=1e-04 \
|
||||
--seed=42 \
|
||||
--output_dir="output_dir_longcat_lora" \
|
||||
--gradient_checkpointing \
|
||||
--mixed_precision="bf16" \
|
||||
--adam_weight_decay=3e-2 \
|
||||
--adam_epsilon=1e-10 \
|
||||
--vae_mini_batch=1 \
|
||||
--max_grad_norm=0.05 \
|
||||
--random_hw_adapt \
|
||||
--training_with_video_token_length \
|
||||
--enable_bucket \
|
||||
--uniform_sampling \
|
||||
--rank=64 \
|
||||
--network_alpha=32 \
|
||||
--target_name="qkv,q_linear,kv_linear,ffn.w1,ffn.w2,ffn.w3" \
|
||||
--use_peft_lora \
|
||||
--train_mode="normal" \
|
||||
--low_vram
|
||||
```
|
||||
File diff suppressed because it is too large
Load Diff
@@ -0,0 +1,41 @@
|
||||
export MODEL_NAME="models/Diffusion_Transformer/LongCat-Video"
|
||||
export DATASET_NAME="datasets/internal_datasets/"
|
||||
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
|
||||
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
|
||||
# export NCCL_IB_DISABLE=1
|
||||
# export NCCL_P2P_DISABLE=1
|
||||
NCCL_DEBUG=INFO
|
||||
|
||||
accelerate launch --mixed_precision="bf16" scripts/longcatvideo/train.py \
|
||||
--pretrained_model_name_or_path=$MODEL_NAME \
|
||||
--train_data_dir=$DATASET_NAME \
|
||||
--train_data_meta=$DATASET_META_NAME \
|
||||
--image_sample_size=640 \
|
||||
--video_sample_size=640 \
|
||||
--token_sample_size=640 \
|
||||
--video_sample_stride=2 \
|
||||
--video_sample_n_frames=81 \
|
||||
--train_batch_size=1 \
|
||||
--video_repeat=1 \
|
||||
--gradient_accumulation_steps=1 \
|
||||
--dataloader_num_workers=8 \
|
||||
--num_train_epochs=100 \
|
||||
--checkpointing_steps=50 \
|
||||
--learning_rate=2e-05 \
|
||||
--lr_scheduler="constant_with_warmup" \
|
||||
--lr_warmup_steps=100 \
|
||||
--seed=42 \
|
||||
--output_dir="output_dir_longcat_full_finetune" \
|
||||
--gradient_checkpointing \
|
||||
--mixed_precision="bf16" \
|
||||
--adam_weight_decay=3e-2 \
|
||||
--adam_epsilon=1e-10 \
|
||||
--vae_mini_batch=1 \
|
||||
--max_grad_norm=0.05 \
|
||||
--random_hw_adapt \
|
||||
--training_with_video_token_length \
|
||||
--enable_bucket \
|
||||
--uniform_sampling \
|
||||
--low_vram \
|
||||
--train_mode="normal" \
|
||||
--trainable_modules "."
|
||||
File diff suppressed because it is too large
Load Diff
@@ -0,0 +1,42 @@
|
||||
export MODEL_NAME="models/Diffusion_Transformer/LongCat-Video"
|
||||
export DATASET_NAME="datasets/internal_datasets/"
|
||||
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
|
||||
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
|
||||
# export NCCL_IB_DISABLE=1
|
||||
# export NCCL_P2P_DISABLE=1
|
||||
NCCL_DEBUG=INFO
|
||||
|
||||
accelerate launch --mixed_precision="bf16" scripts/longcatvideo/train_lora.py \
|
||||
--pretrained_model_name_or_path=$MODEL_NAME \
|
||||
--train_data_dir=$DATASET_NAME \
|
||||
--train_data_meta=$DATASET_META_NAME \
|
||||
--image_sample_size=640 \
|
||||
--video_sample_size=640 \
|
||||
--token_sample_size=640 \
|
||||
--video_sample_stride=2 \
|
||||
--video_sample_n_frames=81 \
|
||||
--train_batch_size=1 \
|
||||
--video_repeat=1 \
|
||||
--gradient_accumulation_steps=1 \
|
||||
--dataloader_num_workers=8 \
|
||||
--num_train_epochs=100 \
|
||||
--checkpointing_steps=50 \
|
||||
--learning_rate=1e-04 \
|
||||
--seed=42 \
|
||||
--output_dir="output_dir_longcat_lora" \
|
||||
--gradient_checkpointing \
|
||||
--mixed_precision="bf16" \
|
||||
--adam_weight_decay=3e-2 \
|
||||
--adam_epsilon=1e-10 \
|
||||
--vae_mini_batch=1 \
|
||||
--max_grad_norm=0.05 \
|
||||
--random_hw_adapt \
|
||||
--training_with_video_token_length \
|
||||
--enable_bucket \
|
||||
--uniform_sampling \
|
||||
--rank=64 \
|
||||
--network_alpha=32 \
|
||||
--target_name="qkv,q_linear,kv_linear,ffn.w1,ffn.w2,ffn.w3" \
|
||||
--use_peft_lora \
|
||||
--train_mode="normal" \
|
||||
--low_vram
|
||||
@@ -2,12 +2,13 @@ import importlib.util
|
||||
|
||||
from diffusers import AutoencoderKL
|
||||
from transformers import (AutoProcessor, AutoTokenizer, CLIPImageProcessor,
|
||||
CLIPTextModel, CLIPTokenizer, Qwen3Config,
|
||||
CLIPTextModel, CLIPTokenizer,
|
||||
CLIPVisionModelWithProjection, LlamaModel,
|
||||
LlamaTokenizerFast, LlavaForConditionalGeneration,
|
||||
Mistral3ForConditionalGeneration, PixtralProcessor,
|
||||
Qwen3ForCausalLM, T5EncoderModel, T5Tokenizer,
|
||||
Siglip2VisionModel, T5TokenizerFast)
|
||||
Qwen3Config, Qwen3ForCausalLM, Siglip2VisionModel,
|
||||
T5EncoderModel, T5Tokenizer, T5TokenizerFast,
|
||||
UMT5EncoderModel)
|
||||
|
||||
try:
|
||||
from transformers import (Qwen2_5_VLConfig,
|
||||
@@ -29,6 +30,8 @@ from .flux2_vae import AutoencoderKLFlux2
|
||||
from .flux_transformer2d import FluxTransformer2DModel
|
||||
from .hunyuanvideo_transformer3d import HunyuanVideoTransformer3DModel
|
||||
from .hunyuanvideo_vae import AutoencoderKLHunyuanVideo
|
||||
from .longcatvideo_transformer3d import LongCatVideoTransformer3DModel
|
||||
from .longcatvideo_vae import AutoencoderKLLongCatVideo
|
||||
from .qwenimage_transformer2d import QwenImageTransformer2DModel
|
||||
from .qwenimage_vae import AutoencoderKLQwenImage
|
||||
from .wan_audio_encoder import WanAudioEncoder
|
||||
|
||||
@@ -40,6 +40,44 @@ except:
|
||||
sageattn = None
|
||||
SAGE_ATTENTION_AVAILABLE = False
|
||||
|
||||
|
||||
def flash_attention_naive(
|
||||
q,
|
||||
k,
|
||||
v,
|
||||
cu_seqlens_q=None,
|
||||
cu_seqlens_k=None,
|
||||
max_seqlen_q=None,
|
||||
max_seqlen_k=None,
|
||||
):
|
||||
# apply attention
|
||||
if FLASH_ATTN_3_AVAILABLE:
|
||||
# Note: dropout_p, window_size are not supported in FA3 now.
|
||||
x = flash_attn_interface.flash_attn_varlen_func(
|
||||
q=q,
|
||||
k=k,
|
||||
v=v,
|
||||
cu_seqlens_q=cu_seqlens_q,
|
||||
cu_seqlens_k=cu_seqlens_k,
|
||||
max_seqlen_q=max_seqlen_q,
|
||||
max_seqlen_k=max_seqlen_k,
|
||||
)[0]
|
||||
else:
|
||||
assert FLASH_ATTN_2_AVAILABLE
|
||||
x = flash_attn.flash_attn_varlen_func(
|
||||
q=q,
|
||||
k=k,
|
||||
v=v,
|
||||
cu_seqlens_q=cu_seqlens_q,
|
||||
cu_seqlens_k=cu_seqlens_k,
|
||||
max_seqlen_q=max_seqlen_q,
|
||||
max_seqlen_k=max_seqlen_k,
|
||||
)
|
||||
|
||||
# output
|
||||
return x.type(q.dtype)
|
||||
|
||||
|
||||
def flash_attention(
|
||||
q,
|
||||
k,
|
||||
|
||||
@@ -0,0 +1,854 @@
|
||||
import glob
|
||||
import json
|
||||
import math
|
||||
import os
|
||||
import types
|
||||
import warnings
|
||||
from typing import Any, Dict, List, Optional, Tuple, Union
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
import torch.amp as amp
|
||||
import torch.cuda.amp as amp
|
||||
import torch.nn as nn
|
||||
import torch.nn.functional as F
|
||||
from diffusers.configuration_utils import ConfigMixin, register_to_config
|
||||
from diffusers.loaders.single_file_model import FromOriginalModelMixin
|
||||
from diffusers.models.modeling_utils import ModelMixin
|
||||
from diffusers.utils import is_torch_version, logging
|
||||
from einops import rearrange, repeat
|
||||
from torch import nn
|
||||
|
||||
from .attention_utils import attention, flash_attention_naive
|
||||
|
||||
|
||||
def broadcat(tensors, dim=-1):
|
||||
num_tensors = len(tensors)
|
||||
shape_lens = set(list(map(lambda t: len(t.shape), tensors)))
|
||||
assert len(shape_lens) == 1, "tensors must all have the same number of dimensions"
|
||||
shape_len = list(shape_lens)[0]
|
||||
dim = (dim + shape_len) if dim < 0 else dim
|
||||
dims = list(zip(*map(lambda t: list(t.shape), tensors)))
|
||||
expandable_dims = [(i, val) for i, val in enumerate(dims) if i != dim]
|
||||
assert all(
|
||||
[*map(lambda t: len(set(t[1])) <= 2, expandable_dims)]
|
||||
), "invalid dimensions for broadcastable concatentation"
|
||||
max_dims = list(map(lambda t: (t[0], max(t[1])), expandable_dims))
|
||||
expanded_dims = list(map(lambda t: (t[0], (t[1],) * num_tensors), max_dims))
|
||||
expanded_dims.insert(dim, (dim, dims[dim]))
|
||||
expandable_shapes = list(zip(*map(lambda t: t[1], expanded_dims)))
|
||||
tensors = list(map(lambda t: t[0].expand(*t[1]), zip(tensors, expandable_shapes)))
|
||||
return torch.cat(tensors, dim=dim)
|
||||
|
||||
|
||||
def rotate_half(x):
|
||||
x = rearrange(x, "... (d r) -> ... d r", r=2)
|
||||
x1, x2 = x.unbind(dim=-1)
|
||||
x = torch.stack((-x2, x1), dim=-1)
|
||||
return rearrange(x, "... d r -> ... (d r)")
|
||||
|
||||
|
||||
def modulate_fp32(norm_func, x, shift, scale):
|
||||
# Suppose x is (B, N, D), shift is (B, -1, D), scale is (B, -1, D)
|
||||
# ensure the modulation params be fp32
|
||||
assert shift.dtype == torch.float32, scale.dtype == torch.float32
|
||||
dtype = x.dtype
|
||||
x = norm_func(x.to(torch.float32))
|
||||
x = x * (scale + 1) + shift
|
||||
x = x.to(dtype)
|
||||
return x
|
||||
|
||||
|
||||
class PatchEmbed3D(nn.Module):
|
||||
"""Video to Patch Embedding.
|
||||
|
||||
Args:
|
||||
patch_size (int): Patch token size. Default: (2,4,4).
|
||||
in_chans (int): Number of input video channels. Default: 3.
|
||||
embed_dim (int): Number of linear projection output channels. Default: 96.
|
||||
norm_layer (nn.Module, optional): Normalization layer. Default: None
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
patch_size=(2, 4, 4),
|
||||
in_chans=3,
|
||||
embed_dim=96,
|
||||
norm_layer=None,
|
||||
flatten=True,
|
||||
):
|
||||
super().__init__()
|
||||
self.patch_size = patch_size
|
||||
self.flatten = flatten
|
||||
|
||||
self.in_chans = in_chans
|
||||
self.embed_dim = embed_dim
|
||||
|
||||
self.proj = nn.Conv3d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
|
||||
if norm_layer is not None:
|
||||
self.norm = norm_layer(embed_dim)
|
||||
else:
|
||||
self.norm = None
|
||||
|
||||
def forward(self, x):
|
||||
"""Forward function."""
|
||||
# padding
|
||||
_, _, D, H, W = x.size()
|
||||
if W % self.patch_size[2] != 0:
|
||||
x = F.pad(x, (0, self.patch_size[2] - W % self.patch_size[2]))
|
||||
if H % self.patch_size[1] != 0:
|
||||
x = F.pad(x, (0, 0, 0, self.patch_size[1] - H % self.patch_size[1]))
|
||||
if D % self.patch_size[0] != 0:
|
||||
x = F.pad(x, (0, 0, 0, 0, 0, self.patch_size[0] - D % self.patch_size[0]))
|
||||
|
||||
B, C, T, H, W = x.shape
|
||||
x = self.proj(x) # (B C T H W)
|
||||
if self.norm is not None:
|
||||
D, Wh, Ww = x.size(2), x.size(3), x.size(4)
|
||||
x = x.flatten(2).transpose(1, 2)
|
||||
x = self.norm(x)
|
||||
x = x.transpose(1, 2).view(-1, self.embed_dim, D, Wh, Ww)
|
||||
if self.flatten:
|
||||
x = x.flatten(2).transpose(1, 2) # BCTHW -> BNC
|
||||
return x
|
||||
|
||||
|
||||
class TimestepEmbedder(nn.Module):
|
||||
"""
|
||||
Embeds scalar timesteps into vector representations.
|
||||
"""
|
||||
|
||||
def __init__(self, t_embed_dim, frequency_embedding_size=256):
|
||||
super().__init__()
|
||||
self.t_embed_dim = t_embed_dim
|
||||
self.frequency_embedding_size = frequency_embedding_size
|
||||
self.mlp = nn.Sequential(
|
||||
nn.Linear(frequency_embedding_size, t_embed_dim, bias=True),
|
||||
nn.SiLU(),
|
||||
nn.Linear(t_embed_dim, t_embed_dim, bias=True),
|
||||
)
|
||||
|
||||
@staticmethod
|
||||
def timestep_embedding(t, dim, max_period=10000):
|
||||
"""
|
||||
Create sinusoidal timestep embeddings.
|
||||
:param t: a 1-D Tensor of N indices, one per batch element.
|
||||
These may be fractional.
|
||||
:param dim: the dimension of the output.
|
||||
:param max_period: controls the minimum frequency of the embeddings.
|
||||
:return: an (N, D) Tensor of positional embeddings.
|
||||
"""
|
||||
half = dim // 2
|
||||
freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half)
|
||||
freqs = freqs.to(device=t.device)
|
||||
args = t[:, None].float() * freqs[None]
|
||||
embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
|
||||
if dim % 2:
|
||||
embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
|
||||
return embedding
|
||||
|
||||
def forward(self, t, dtype):
|
||||
t_freq = self.timestep_embedding(t, self.frequency_embedding_size)
|
||||
if t_freq.dtype != dtype:
|
||||
t_freq = t_freq.to(dtype)
|
||||
t_emb = self.mlp(t_freq)
|
||||
return t_emb
|
||||
|
||||
|
||||
class CaptionEmbedder(nn.Module):
|
||||
"""
|
||||
Embeds class labels into vector representations.
|
||||
"""
|
||||
|
||||
def __init__(self, in_channels, hidden_size):
|
||||
super().__init__()
|
||||
self.in_channels = in_channels
|
||||
self.hidden_size = hidden_size
|
||||
self.y_proj = nn.Sequential(
|
||||
nn.Linear(in_channels, hidden_size, bias=True),
|
||||
nn.GELU(approximate="tanh"),
|
||||
nn.Linear(hidden_size, hidden_size, bias=True),
|
||||
)
|
||||
|
||||
def forward(self, caption):
|
||||
B, _, N, C = caption.shape
|
||||
caption = self.y_proj(caption)
|
||||
return caption
|
||||
|
||||
|
||||
class RMSNorm_FP32(torch.nn.Module):
|
||||
def __init__(self, dim: int, eps: float):
|
||||
super().__init__()
|
||||
self.eps = eps
|
||||
self.weight = nn.Parameter(torch.ones(dim))
|
||||
|
||||
def _norm(self, x):
|
||||
return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
|
||||
|
||||
def forward(self, x):
|
||||
output = self._norm(x.float()).type_as(x)
|
||||
return output * self.weight
|
||||
|
||||
|
||||
class LayerNorm_FP32(nn.LayerNorm):
|
||||
def __init__(self, dim, eps, elementwise_affine):
|
||||
super().__init__(dim, eps=eps, elementwise_affine=elementwise_affine)
|
||||
|
||||
def forward(self, inputs: torch.Tensor) -> torch.Tensor:
|
||||
origin_dtype = inputs.dtype
|
||||
out = F.layer_norm(
|
||||
inputs.float(),
|
||||
self.normalized_shape,
|
||||
None if self.weight is None else self.weight.float(),
|
||||
None if self.bias is None else self.bias.float() ,
|
||||
self.eps
|
||||
).to(origin_dtype)
|
||||
return out
|
||||
|
||||
|
||||
class RotaryPositionalEmbedding(nn.Module):
|
||||
|
||||
def __init__(self,
|
||||
head_dim,
|
||||
cp_split_hw=None
|
||||
):
|
||||
"""Rotary positional embedding for 3D
|
||||
Reference : https://blog.eleuther.ai/rotary-embeddings/
|
||||
Paper: https://arxiv.org/pdf/2104.09864.pdf
|
||||
Args:
|
||||
dim: Dimension of embedding
|
||||
base: Base value for exponential
|
||||
"""
|
||||
super().__init__()
|
||||
self.head_dim = head_dim
|
||||
assert self.head_dim % 8 == 0, 'Dim must be a multiply of 8 for 3D RoPE.'
|
||||
self.cp_split_hw = cp_split_hw
|
||||
# We take the assumption that the longest side of grid will not larger than 512, i.e, 512 * 8 = 4098 input pixels
|
||||
self.base = 10000
|
||||
self.freqs_dict = {}
|
||||
|
||||
def register_grid_size(self, grid_size):
|
||||
if grid_size not in self.freqs_dict:
|
||||
self.freqs_dict.update({
|
||||
grid_size: self.precompute_freqs_cis_3d(grid_size)
|
||||
})
|
||||
|
||||
def precompute_freqs_cis_3d(self, grid_size):
|
||||
num_frames, height, width = grid_size
|
||||
dim_t = self.head_dim - 4 * (self.head_dim // 6)
|
||||
dim_h = 2 * (self.head_dim // 6)
|
||||
dim_w = 2 * (self.head_dim // 6)
|
||||
freqs_t = 1.0 / (self.base ** (torch.arange(0, dim_t, 2)[: (dim_t // 2)].float() / dim_t))
|
||||
freqs_h = 1.0 / (self.base ** (torch.arange(0, dim_h, 2)[: (dim_h // 2)].float() / dim_h))
|
||||
freqs_w = 1.0 / (self.base ** (torch.arange(0, dim_w, 2)[: (dim_w // 2)].float() / dim_w))
|
||||
grid_t = np.linspace(0, num_frames, num_frames, endpoint=False, dtype=np.float32)
|
||||
grid_h = np.linspace(0, height, height, endpoint=False, dtype=np.float32)
|
||||
grid_w = np.linspace(0, width, width, endpoint=False, dtype=np.float32)
|
||||
grid_t = torch.from_numpy(grid_t).float()
|
||||
grid_h = torch.from_numpy(grid_h).float()
|
||||
grid_w = torch.from_numpy(grid_w).float()
|
||||
freqs_t = torch.einsum("..., f -> ... f", grid_t, freqs_t)
|
||||
freqs_h = torch.einsum("..., f -> ... f", grid_h, freqs_h)
|
||||
freqs_w = torch.einsum("..., f -> ... f", grid_w, freqs_w)
|
||||
freqs_t = repeat(freqs_t, "... n -> ... (n r)", r=2)
|
||||
freqs_h = repeat(freqs_h, "... n -> ... (n r)", r=2)
|
||||
freqs_w = repeat(freqs_w, "... n -> ... (n r)", r=2)
|
||||
freqs = broadcat((freqs_t[:, None, None, :], freqs_h[None, :, None, :], freqs_w[None, None, :, :]), dim=-1)
|
||||
# (T H W D)
|
||||
freqs = rearrange(freqs, "T H W D -> (T H W) D")
|
||||
|
||||
return freqs
|
||||
|
||||
def forward(self, q, k, grid_size):
|
||||
"""3D RoPE.
|
||||
|
||||
Args:
|
||||
query: [B, head, seq, head_dim]
|
||||
key: [B, head, seq, head_dim]
|
||||
Returns:
|
||||
query and key with the same shape as input.
|
||||
"""
|
||||
|
||||
if grid_size not in self.freqs_dict:
|
||||
self.register_grid_size(grid_size)
|
||||
|
||||
freqs_cis = self.freqs_dict[grid_size].to(q.device)
|
||||
q_, k_ = q.float(), k.float()
|
||||
freqs_cis = freqs_cis.float().to(q.device)
|
||||
cos, sin = freqs_cis.cos(), freqs_cis.sin()
|
||||
cos, sin = rearrange(cos, 'n d -> 1 1 n d'), rearrange(sin, 'n d -> 1 1 n d')
|
||||
q_ = (q_ * cos) + (rotate_half(q_) * sin)
|
||||
k_ = (k_ * cos) + (rotate_half(k_) * sin)
|
||||
|
||||
return q_.type_as(q), k_.type_as(k)
|
||||
|
||||
|
||||
class Attention(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
dim: int,
|
||||
num_heads: int,
|
||||
enable_flashattn3: bool = False,
|
||||
enable_flashattn2: bool = False,
|
||||
enable_xformers: bool = False,
|
||||
enable_bsa: bool = False,
|
||||
bsa_params: dict = None,
|
||||
cp_split_hw: Optional[List[int]] = None
|
||||
) -> None:
|
||||
super().__init__()
|
||||
assert dim % num_heads == 0, "dim should be divisible by num_heads"
|
||||
self.dim = dim
|
||||
self.num_heads = num_heads
|
||||
self.head_dim = dim // num_heads
|
||||
self.scale = self.head_dim**-0.5
|
||||
self.enable_flashattn3 = enable_flashattn3
|
||||
self.enable_flashattn2 = enable_flashattn2
|
||||
self.enable_xformers = enable_xformers
|
||||
self.enable_bsa = enable_bsa
|
||||
self.bsa_params = bsa_params
|
||||
self.cp_split_hw = cp_split_hw
|
||||
|
||||
self.qkv = nn.Linear(dim, dim * 3, bias=True)
|
||||
self.q_norm = RMSNorm_FP32(self.head_dim, eps=1e-6)
|
||||
self.k_norm = RMSNorm_FP32(self.head_dim, eps=1e-6)
|
||||
self.proj = nn.Linear(dim, dim)
|
||||
|
||||
self.rope_3d = RotaryPositionalEmbedding(
|
||||
self.head_dim,
|
||||
cp_split_hw=cp_split_hw
|
||||
)
|
||||
|
||||
def _process_attn(self, q, k, v, shape):
|
||||
q = rearrange(q, "B H S D -> B S H D")
|
||||
k = rearrange(k, "B H S D -> B S H D")
|
||||
v = rearrange(v, "B H S D -> B S H D")
|
||||
x = attention(q, k, v)
|
||||
x = rearrange(x, "B S H D -> B H S D")
|
||||
return x
|
||||
|
||||
def forward(self, x: torch.Tensor, shape=None, num_cond_latents=None, return_kv=False) -> torch.Tensor:
|
||||
"""
|
||||
"""
|
||||
B, N, C = x.shape
|
||||
qkv = self.qkv(x)
|
||||
|
||||
qkv_shape = (B, N, 3, self.num_heads, self.head_dim)
|
||||
qkv = qkv.view(qkv_shape).permute((2, 0, 3, 1, 4)) # [3, B, H, N, D]
|
||||
q, k, v = qkv.unbind(0)
|
||||
q, k = self.q_norm(q), self.k_norm(k)
|
||||
|
||||
if return_kv:
|
||||
k_cache, v_cache = k.clone(), v.clone()
|
||||
|
||||
q, k = self.rope_3d(q, k, shape)
|
||||
|
||||
# cond mode
|
||||
if num_cond_latents is not None and num_cond_latents > 0:
|
||||
num_cond_latents_thw = num_cond_latents * (N // shape[0])
|
||||
# process the condition tokens
|
||||
q_cond = q[:, :, :num_cond_latents_thw].contiguous()
|
||||
k_cond = k[:, :, :num_cond_latents_thw].contiguous()
|
||||
v_cond = v[:, :, :num_cond_latents_thw].contiguous()
|
||||
x_cond = self._process_attn(q_cond, k_cond, v_cond, shape)
|
||||
# process the noise tokens
|
||||
q_noise = q[:, :, num_cond_latents_thw:].contiguous()
|
||||
x_noise = self._process_attn(q_noise, k, v, shape)
|
||||
# merge x_cond and x_noise
|
||||
x = torch.cat([x_cond, x_noise], dim=2).contiguous()
|
||||
else:
|
||||
x = self._process_attn(q, k, v, shape)
|
||||
|
||||
x_output_shape = (B, N, C)
|
||||
x = x.transpose(1, 2) # [B, H, N, D] --> [B, N, H, D]
|
||||
x = x.reshape(x_output_shape) # [B, N, H, D] --> [B, N, C]
|
||||
x = self.proj(x)
|
||||
|
||||
if return_kv:
|
||||
return x, (k_cache, v_cache)
|
||||
else:
|
||||
return x
|
||||
|
||||
def forward_with_kv_cache(self, x: torch.Tensor, shape=None, num_cond_latents=None, kv_cache=None) -> torch.Tensor:
|
||||
"""
|
||||
"""
|
||||
B, N, C = x.shape
|
||||
qkv = self.qkv(x)
|
||||
|
||||
qkv_shape = (B, N, 3, self.num_heads, self.head_dim)
|
||||
qkv = qkv.view(qkv_shape).permute((2, 0, 3, 1, 4)) # [3, B, H, N, D]
|
||||
q, k, v = qkv.unbind(0)
|
||||
q, k = self.q_norm(q), self.k_norm(k)
|
||||
|
||||
T, H, W = shape
|
||||
k_cache, v_cache = kv_cache
|
||||
assert k_cache.shape[0] == v_cache.shape[0] and k_cache.shape[0] in [1, B]
|
||||
if k_cache.shape[0] == 1:
|
||||
k_cache = k_cache.repeat(B, 1, 1, 1)
|
||||
v_cache = v_cache.repeat(B, 1, 1, 1)
|
||||
|
||||
if num_cond_latents is not None and num_cond_latents > 0:
|
||||
k_full = torch.cat([k_cache, k], dim=2).contiguous()
|
||||
v_full = torch.cat([v_cache, v], dim=2).contiguous()
|
||||
q_padding = torch.cat([torch.empty_like(k_cache), q], dim=2).contiguous()
|
||||
q_padding, k_full = self.rope_3d(q_padding, k_full, (T + num_cond_latents, H, W))
|
||||
q = q_padding[:, :, -N:].contiguous()
|
||||
|
||||
x = self._process_attn(q, k_full, v_full, shape)
|
||||
|
||||
x_output_shape = (B, N, C)
|
||||
x = x.transpose(1, 2) # [B, H, N, D] --> [B, N, H, D]
|
||||
x = x.reshape(x_output_shape) # [B, N, H, D] --> [B, N, C]
|
||||
x = self.proj(x)
|
||||
|
||||
return x
|
||||
|
||||
|
||||
class MultiHeadCrossAttention(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
dim,
|
||||
num_heads,
|
||||
enable_flashattn3=False,
|
||||
enable_flashattn2=False,
|
||||
enable_xformers=False,
|
||||
):
|
||||
super(MultiHeadCrossAttention, self).__init__()
|
||||
assert dim % num_heads == 0, "d_model must be divisible by num_heads"
|
||||
|
||||
self.dim = dim
|
||||
self.num_heads = num_heads
|
||||
self.head_dim = dim // num_heads
|
||||
|
||||
self.q_linear = nn.Linear(dim, dim)
|
||||
self.kv_linear = nn.Linear(dim, dim * 2)
|
||||
self.proj = nn.Linear(dim, dim)
|
||||
|
||||
self.q_norm = RMSNorm_FP32(self.head_dim, eps=1e-6)
|
||||
self.k_norm = RMSNorm_FP32(self.head_dim, eps=1e-6)
|
||||
|
||||
self.enable_flashattn3 = enable_flashattn3
|
||||
self.enable_flashattn2 = enable_flashattn2
|
||||
self.enable_xformers = enable_xformers
|
||||
|
||||
def _process_cross_attn(self, x, cond, kv_seqlen):
|
||||
B, N, C = x.shape
|
||||
assert C == self.dim and cond.shape[2] == self.dim
|
||||
|
||||
q = self.q_linear(x).view(1, -1, self.num_heads, self.head_dim)
|
||||
kv = self.kv_linear(cond).view(1, -1, 2, self.num_heads, self.head_dim)
|
||||
k, v = kv.unbind(2)
|
||||
|
||||
q, k = self.q_norm(q), self.k_norm(k)
|
||||
x = flash_attention_naive(
|
||||
q=q[0],
|
||||
k=k[0],
|
||||
v=v[0],
|
||||
cu_seqlens_q=torch.tensor([0] + [N] * B, device=q.device).cumsum(0).to(torch.int32),
|
||||
cu_seqlens_k=torch.tensor([0] + kv_seqlen, device=q.device).cumsum(0).to(torch.int32),
|
||||
max_seqlen_q=N,
|
||||
max_seqlen_k=max(kv_seqlen),
|
||||
)
|
||||
|
||||
x = x.view(B, -1, C)
|
||||
x = self.proj(x)
|
||||
return x
|
||||
|
||||
def forward(self, x, cond, kv_seqlen, num_cond_latents=None, shape=None):
|
||||
"""
|
||||
x: [B, N, C]
|
||||
cond: [B, M, C]
|
||||
"""
|
||||
if num_cond_latents is None or num_cond_latents == 0:
|
||||
return self._process_cross_attn(x, cond, kv_seqlen)
|
||||
else:
|
||||
B, N, C = x.shape
|
||||
if num_cond_latents is not None and num_cond_latents > 0:
|
||||
assert shape is not None, "SHOULD pass in the shape"
|
||||
num_cond_latents_thw = num_cond_latents * (N // shape[0])
|
||||
x_noise = x[:, num_cond_latents_thw:] # [B, N_noise, C]
|
||||
output_noise = self._process_cross_attn(x_noise, cond, kv_seqlen) # [B, N_noise, C]
|
||||
output = torch.cat([
|
||||
torch.zeros((B, num_cond_latents_thw, C), dtype=output_noise.dtype, device=output_noise.device),
|
||||
output_noise
|
||||
], dim=1).contiguous()
|
||||
else:
|
||||
raise NotImplementedError
|
||||
|
||||
return output
|
||||
|
||||
|
||||
class FeedForwardSwiGLU(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
dim: int,
|
||||
hidden_dim: int,
|
||||
multiple_of: int = 256,
|
||||
ffn_dim_multiplier: Optional[float] = None,
|
||||
):
|
||||
super().__init__()
|
||||
hidden_dim = int(2 * hidden_dim / 3)
|
||||
# custom dim factor multiplier
|
||||
if ffn_dim_multiplier is not None:
|
||||
hidden_dim = int(ffn_dim_multiplier * hidden_dim)
|
||||
hidden_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
|
||||
|
||||
self.dim = dim
|
||||
self.hidden_dim = hidden_dim
|
||||
self.w1 = nn.Linear(dim, hidden_dim, bias=False)
|
||||
self.w2 = nn.Linear(hidden_dim, dim, bias=False)
|
||||
self.w3 = nn.Linear(dim, hidden_dim, bias=False)
|
||||
|
||||
def forward(self, x):
|
||||
return self.w2(F.silu(self.w1(x)) * self.w3(x))
|
||||
|
||||
|
||||
class LongCatSingleStreamBlock(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
hidden_size: int,
|
||||
num_heads: int,
|
||||
mlp_ratio: int,
|
||||
adaln_tembed_dim: int,
|
||||
enable_flashattn3: bool = False,
|
||||
enable_flashattn2: bool = False,
|
||||
enable_xformers: bool = False,
|
||||
enable_bsa: bool = False,
|
||||
bsa_params=None,
|
||||
cp_split_hw=None
|
||||
):
|
||||
super().__init__()
|
||||
|
||||
self.hidden_size = hidden_size
|
||||
|
||||
# scale and gate modulation
|
||||
self.adaLN_modulation = nn.Sequential(
|
||||
nn.SiLU(),
|
||||
nn.Linear(adaln_tembed_dim, 6 * hidden_size, bias=True)
|
||||
)
|
||||
|
||||
self.mod_norm_attn = LayerNorm_FP32(hidden_size, eps=1e-6, elementwise_affine=False)
|
||||
self.mod_norm_ffn = LayerNorm_FP32(hidden_size, eps=1e-6, elementwise_affine=False)
|
||||
self.pre_crs_attn_norm = LayerNorm_FP32(hidden_size, eps=1e-6, elementwise_affine=True)
|
||||
|
||||
self.attn = Attention(
|
||||
dim=hidden_size,
|
||||
num_heads=num_heads,
|
||||
enable_flashattn3=enable_flashattn3,
|
||||
enable_flashattn2=enable_flashattn2,
|
||||
enable_xformers=enable_xformers,
|
||||
enable_bsa=enable_bsa,
|
||||
bsa_params=bsa_params,
|
||||
cp_split_hw=cp_split_hw
|
||||
)
|
||||
self.cross_attn = MultiHeadCrossAttention(
|
||||
dim=hidden_size,
|
||||
num_heads=num_heads,
|
||||
enable_flashattn3=enable_flashattn3,
|
||||
enable_flashattn2=enable_flashattn2,
|
||||
enable_xformers=enable_xformers,
|
||||
)
|
||||
self.ffn = FeedForwardSwiGLU(dim=hidden_size, hidden_dim=int(hidden_size * mlp_ratio))
|
||||
|
||||
def forward(self, x, y, t, y_seqlen, latent_shape, num_cond_latents=None, return_kv=False, kv_cache=None, skip_crs_attn=False):
|
||||
"""
|
||||
x: [B, N, C]
|
||||
y: [1, N_valid_tokens, C]
|
||||
t: [B, T, C_t]
|
||||
y_seqlen: [B]; type of a list
|
||||
latent_shape: latent shape of a single item
|
||||
"""
|
||||
x_dtype = x.dtype
|
||||
|
||||
B, N, C = x.shape
|
||||
T, _, _ = latent_shape # S != T*H*W in case of CP split on H*W.
|
||||
|
||||
# compute modulation params in fp32
|
||||
with amp.autocast(dtype=torch.float32):
|
||||
shift_msa, scale_msa, gate_msa, \
|
||||
shift_mlp, scale_mlp, gate_mlp = \
|
||||
self.adaLN_modulation(t).unsqueeze(2).chunk(6, dim=-1) # [B, T, 1, C]
|
||||
|
||||
# self attn with modulation
|
||||
x_m = modulate_fp32(self.mod_norm_attn, x.view(B, T, -1, C), shift_msa, scale_msa).view(B, N, C)
|
||||
|
||||
if kv_cache is not None:
|
||||
kv_cache = (kv_cache[0].to(x.device), kv_cache[1].to(x.device))
|
||||
attn_outputs = self.attn.forward_with_kv_cache(x_m, shape=latent_shape, num_cond_latents=num_cond_latents, kv_cache=kv_cache)
|
||||
else:
|
||||
attn_outputs = self.attn(x_m, shape=latent_shape, num_cond_latents=num_cond_latents, return_kv=return_kv)
|
||||
|
||||
if return_kv:
|
||||
x_s, kv_cache = attn_outputs
|
||||
else:
|
||||
x_s = attn_outputs
|
||||
|
||||
with amp.autocast(dtype=torch.float32):
|
||||
x = x + (gate_msa * x_s.view(B, -1, N//T, C)).view(B, -1, C) # [B, N, C]
|
||||
x = x.to(x_dtype)
|
||||
|
||||
# cross attn
|
||||
if not skip_crs_attn:
|
||||
if kv_cache is not None:
|
||||
num_cond_latents = None
|
||||
x = x + self.cross_attn(self.pre_crs_attn_norm(x), y, y_seqlen, num_cond_latents=num_cond_latents, shape=latent_shape)
|
||||
|
||||
# ffn with modulation
|
||||
x_m = modulate_fp32(self.mod_norm_ffn, x.view(B, -1, N//T, C), shift_mlp, scale_mlp).view(B, -1, C)
|
||||
x_s = self.ffn(x_m)
|
||||
with amp.autocast(dtype=torch.float32):
|
||||
x = x + (gate_mlp * x_s.view(B, -1, N//T, C)).view(B, -1, C) # [B, N, C]
|
||||
x = x.to(x_dtype)
|
||||
|
||||
if return_kv:
|
||||
return x, kv_cache
|
||||
else:
|
||||
return x
|
||||
|
||||
|
||||
class FinalLayer_FP32(nn.Module):
|
||||
"""
|
||||
The final layer of DiT.
|
||||
"""
|
||||
|
||||
def __init__(self, hidden_size, num_patch, out_channels, adaln_tembed_dim):
|
||||
super().__init__()
|
||||
self.hidden_size = hidden_size
|
||||
self.num_patch = num_patch
|
||||
self.out_channels = out_channels
|
||||
self.adaln_tembed_dim = adaln_tembed_dim
|
||||
|
||||
self.norm_final = LayerNorm_FP32(hidden_size, elementwise_affine=False, eps=1e-6)
|
||||
self.linear = nn.Linear(hidden_size, num_patch * out_channels, bias=True)
|
||||
self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(adaln_tembed_dim, 2 * hidden_size, bias=True))
|
||||
|
||||
def forward(self, x, t, latent_shape):
|
||||
# timestep shape: [B, T, C]
|
||||
assert t.dtype == torch.float32
|
||||
B, N, C = x.shape
|
||||
T, _, _ = latent_shape
|
||||
|
||||
with amp.autocast(dtype=torch.float32):
|
||||
shift, scale = self.adaLN_modulation(t).unsqueeze(2).chunk(2, dim=-1) # [B, T, 1, C]
|
||||
x = modulate_fp32(self.norm_final, x.view(B, T, -1, C), shift, scale).view(B, N, C)
|
||||
x = self.linear(x)
|
||||
return x
|
||||
|
||||
|
||||
class LongCatVideoTransformer3DModel(ModelMixin, ConfigMixin, FromOriginalModelMixin):
|
||||
r"""
|
||||
Wan diffusion backbone supporting both text-to-video and image-to-video.
|
||||
"""
|
||||
|
||||
# _no_split_modules = ['LongCatSingleStreamBlock']
|
||||
_supports_gradient_checkpointing = True
|
||||
|
||||
@register_to_config
|
||||
def __init__(
|
||||
self,
|
||||
in_channels: int = 16,
|
||||
out_channels: int = 16,
|
||||
hidden_size: int = 4096,
|
||||
depth: int = 48,
|
||||
num_heads: int = 32,
|
||||
caption_channels: int = 4096,
|
||||
mlp_ratio: int = 4,
|
||||
adaln_tembed_dim: int = 512,
|
||||
frequency_embedding_size: int = 256,
|
||||
# default params
|
||||
patch_size: Tuple[int] = (1, 2, 2),
|
||||
# attention config
|
||||
enable_flashattn3: bool = False,
|
||||
enable_flashattn2: bool = True,
|
||||
enable_xformers: bool = False,
|
||||
enable_bsa: bool = False,
|
||||
bsa_params: dict = {'sparsity': 0.9375, 'chunk_3d_shape_q': [4, 4, 4], 'chunk_3d_shape_k': [4, 4, 4]},
|
||||
cp_split_hw: Optional[List[int]] = [1, 1],
|
||||
text_tokens_zero_pad: bool = True,
|
||||
) -> None:
|
||||
super().__init__()
|
||||
|
||||
self.patch_size = patch_size
|
||||
self.in_channels = in_channels
|
||||
self.out_channels = out_channels
|
||||
self.cp_split_hw = cp_split_hw
|
||||
|
||||
self.x_embedder = PatchEmbed3D(patch_size, in_channels, hidden_size)
|
||||
self.t_embedder = TimestepEmbedder(t_embed_dim=adaln_tembed_dim, frequency_embedding_size=frequency_embedding_size)
|
||||
self.y_embedder = CaptionEmbedder(
|
||||
in_channels=caption_channels,
|
||||
hidden_size=hidden_size,
|
||||
)
|
||||
|
||||
self.blocks = nn.ModuleList(
|
||||
[
|
||||
LongCatSingleStreamBlock(
|
||||
hidden_size=hidden_size,
|
||||
num_heads=num_heads,
|
||||
mlp_ratio=mlp_ratio,
|
||||
adaln_tembed_dim=adaln_tembed_dim,
|
||||
enable_flashattn3=enable_flashattn3,
|
||||
enable_flashattn2=enable_flashattn2,
|
||||
enable_xformers=enable_xformers,
|
||||
enable_bsa=enable_bsa,
|
||||
bsa_params=bsa_params,
|
||||
cp_split_hw=cp_split_hw
|
||||
)
|
||||
for i in range(depth)
|
||||
]
|
||||
)
|
||||
|
||||
self.final_layer = FinalLayer_FP32(
|
||||
hidden_size,
|
||||
np.prod(self.patch_size),
|
||||
out_channels,
|
||||
adaln_tembed_dim,
|
||||
)
|
||||
|
||||
self.gradient_checkpointing = False
|
||||
self.text_tokens_zero_pad = text_tokens_zero_pad
|
||||
|
||||
self.lora_dict = {}
|
||||
self.active_loras = []
|
||||
|
||||
def _set_gradient_checkpointing(self, *args, **kwargs):
|
||||
if "value" in kwargs:
|
||||
self.gradient_checkpointing = kwargs["value"]
|
||||
if hasattr(self, "motioner") and hasattr(self.motioner, "gradient_checkpointing"):
|
||||
self.motioner.gradient_checkpointing = kwargs["value"]
|
||||
elif "enable" in kwargs:
|
||||
self.gradient_checkpointing = kwargs["enable"]
|
||||
if hasattr(self, "motioner") and hasattr(self.motioner, "gradient_checkpointing"):
|
||||
self.motioner.gradient_checkpointing = kwargs["enable"]
|
||||
else:
|
||||
raise ValueError("Invalid set gradient checkpointing")
|
||||
|
||||
def _gradient_checkpointing_func(module, *args):
|
||||
ckpt_kwargs = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
|
||||
return torch.utils.checkpoint.checkpoint(
|
||||
module.__call__,
|
||||
*args,
|
||||
**ckpt_kwargs,
|
||||
)
|
||||
self._gradient_checkpointing_func = _gradient_checkpointing_func
|
||||
|
||||
def enable_bsa(self,):
|
||||
for block in self.blocks:
|
||||
block.attn.enable_bsa = True
|
||||
|
||||
def disable_bsa(self,):
|
||||
for block in self.blocks:
|
||||
block.attn.enable_bsa = False
|
||||
|
||||
def forward(
|
||||
self,
|
||||
hidden_states,
|
||||
timestep,
|
||||
encoder_hidden_states,
|
||||
encoder_attention_mask=None,
|
||||
num_cond_latents=0,
|
||||
return_kv=False,
|
||||
kv_cache_dict={},
|
||||
skip_crs_attn=False,
|
||||
offload_kv_cache=False,
|
||||
use_gradient_checkpointing=False,
|
||||
use_gradient_checkpointing_offload=False,
|
||||
):
|
||||
|
||||
B, _, T, H, W = hidden_states.shape
|
||||
|
||||
N_t = T // self.patch_size[0]
|
||||
N_h = H // self.patch_size[1]
|
||||
N_w = W // self.patch_size[2]
|
||||
|
||||
assert self.patch_size[0]==1, "Currently, 3D x_embedder should not compress the temporal dimension."
|
||||
|
||||
# expand the shape of timestep from [B] to [B, T]
|
||||
if len(timestep.shape) == 1:
|
||||
timestep = timestep.unsqueeze(1).expand(-1, N_t).clone() # [B, T]
|
||||
timestep[:, :num_cond_latents] = 0
|
||||
|
||||
dtype = hidden_states.dtype
|
||||
hidden_states = hidden_states.to(dtype)
|
||||
timestep = timestep.to(dtype)
|
||||
encoder_hidden_states = encoder_hidden_states.to(dtype)
|
||||
|
||||
hidden_states = self.x_embedder(hidden_states) # [B, N, C]
|
||||
|
||||
with amp.autocast(dtype=torch.float32):
|
||||
t = self.t_embedder(timestep.float().flatten(), dtype=torch.float32).reshape(B, N_t, -1) # [B, T, C_t]
|
||||
|
||||
encoder_hidden_states = self.y_embedder(encoder_hidden_states) # [B, 1, N_token, C]
|
||||
|
||||
if self.text_tokens_zero_pad and encoder_attention_mask is not None:
|
||||
encoder_hidden_states = encoder_hidden_states * encoder_attention_mask[:, None, :, None]
|
||||
encoder_attention_mask = (encoder_attention_mask * 0 + 1).to(encoder_attention_mask.dtype)
|
||||
|
||||
if encoder_attention_mask is not None:
|
||||
encoder_attention_mask = encoder_attention_mask.squeeze(1).squeeze(1)
|
||||
encoder_hidden_states = encoder_hidden_states.squeeze(1).masked_select(encoder_attention_mask.unsqueeze(-1) != 0).view(1, -1, hidden_states.shape[-1]) # [1, N_valid_tokens, C]
|
||||
y_seqlens = encoder_attention_mask.sum(dim=1).tolist() # [B]
|
||||
else:
|
||||
y_seqlens = [encoder_hidden_states.shape[2]] * encoder_hidden_states.shape[0]
|
||||
encoder_hidden_states = encoder_hidden_states.squeeze(1).view(1, -1, hidden_states.shape[-1])
|
||||
|
||||
# blocks
|
||||
kv_cache_dict_ret = {}
|
||||
for i, block in enumerate(self.blocks):
|
||||
if torch.is_grad_enabled() and self.gradient_checkpointing:
|
||||
block_outputs = self._gradient_checkpointing_func(
|
||||
block, hidden_states, encoder_hidden_states, t, y_seqlens,
|
||||
(N_t, N_h, N_w), num_cond_latents, return_kv, kv_cache_dict.get(i, None), skip_crs_attn
|
||||
)
|
||||
else:
|
||||
block_outputs = block(
|
||||
hidden_states, encoder_hidden_states, t, y_seqlens,
|
||||
(N_t, N_h, N_w), num_cond_latents, return_kv, kv_cache_dict.get(i, None), skip_crs_attn
|
||||
)
|
||||
|
||||
if return_kv:
|
||||
hidden_states, kv_cache = block_outputs
|
||||
if offload_kv_cache:
|
||||
kv_cache_dict_ret[i] = (kv_cache[0].cpu(), kv_cache[1].cpu())
|
||||
else:
|
||||
kv_cache_dict_ret[i] = (kv_cache[0].contiguous(), kv_cache[1].contiguous())
|
||||
else:
|
||||
hidden_states = block_outputs
|
||||
|
||||
hidden_states = self.final_layer(hidden_states, t, (N_t, N_h, N_w)) # [B, N, C=T_p*H_p*W_p*C_out]
|
||||
|
||||
hidden_states = self.unpatchify(hidden_states, N_t, N_h, N_w) # [B, C_out, H, W]
|
||||
|
||||
# cast to float32 for better accuracy
|
||||
hidden_states = hidden_states.to(torch.float32)
|
||||
|
||||
if return_kv:
|
||||
return hidden_states, kv_cache_dict_ret
|
||||
else:
|
||||
return hidden_states
|
||||
|
||||
|
||||
def unpatchify(self, x, N_t, N_h, N_w):
|
||||
"""
|
||||
Args:
|
||||
x (torch.Tensor): of shape [B, N, C]
|
||||
|
||||
Return:
|
||||
x (torch.Tensor): of shape [B, C_out, T, H, W]
|
||||
"""
|
||||
T_p, H_p, W_p = self.patch_size
|
||||
x = rearrange(
|
||||
x,
|
||||
"B (N_t N_h N_w) (T_p H_p W_p C_out) -> B C_out (N_t T_p) (N_h H_p) (N_w W_p)",
|
||||
N_t=N_t,
|
||||
N_h=N_h,
|
||||
N_w=N_w,
|
||||
T_p=T_p,
|
||||
H_p=H_p,
|
||||
W_p=W_p,
|
||||
C_out=self.out_channels,
|
||||
)
|
||||
return x
|
||||
|
||||
@staticmethod
|
||||
def state_dict_converter():
|
||||
return LongCatVideoTransformer3DModelDictConverter()
|
||||
File diff suppressed because it is too large
Load Diff
@@ -7,6 +7,7 @@ from .pipeline_flux2 import Flux2Pipeline
|
||||
from .pipeline_flux2_control import Flux2ControlPipeline
|
||||
from .pipeline_hunyuanvideo import HunyuanVideoPipeline
|
||||
from .pipeline_hunyuanvideo_i2v import HunyuanVideoI2VPipeline
|
||||
from .pipeline_longcatvideo import LongCatVideoPipeline
|
||||
from .pipeline_qwenimage import QwenImagePipeline
|
||||
from .pipeline_qwenimage_edit import QwenImageEditPipeline
|
||||
from .pipeline_qwenimage_edit_plus import QwenImageEditPlusPipeline
|
||||
|
||||
@@ -0,0 +1,715 @@
|
||||
import html
|
||||
import inspect
|
||||
import math
|
||||
from dataclasses import dataclass
|
||||
from typing import Any, Callable, Dict, List, Optional, Tuple, Union
|
||||
|
||||
import ftfy
|
||||
import numpy as np
|
||||
import regex as re
|
||||
import torch
|
||||
import torch.nn.functional as F
|
||||
import torchvision.transforms.functional as TF
|
||||
from diffusers import FlowMatchEulerDiscreteScheduler
|
||||
from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
|
||||
from diffusers.image_processor import VaeImageProcessor
|
||||
from diffusers.models.embeddings import get_1d_rotary_pos_embed
|
||||
from diffusers.pipelines.pipeline_utils import DiffusionPipeline
|
||||
from diffusers.utils import BaseOutput, logging, replace_example_docstring
|
||||
from diffusers.utils.torch_utils import randn_tensor
|
||||
from diffusers.video_processor import VideoProcessor
|
||||
from einops import rearrange
|
||||
from PIL import Image
|
||||
|
||||
from ..models import (AutoencoderKLWan, AutoTokenizer,
|
||||
LongCatVideoTransformer3DModel, UMT5EncoderModel)
|
||||
from ..utils.fm_solvers import (FlowDPMSolverMultistepScheduler,
|
||||
get_sampling_sigmas)
|
||||
from ..utils.fm_solvers_unipc import FlowUniPCMultistepScheduler
|
||||
|
||||
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
|
||||
|
||||
|
||||
EXAMPLE_DOC_STRING = """
|
||||
Examples:
|
||||
```python
|
||||
pass
|
||||
```
|
||||
"""
|
||||
|
||||
|
||||
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
|
||||
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,
|
||||
):
|
||||
"""
|
||||
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 basic_clean(text):
|
||||
text = ftfy.fix_text(text)
|
||||
text = html.unescape(html.unescape(text))
|
||||
return text.strip()
|
||||
|
||||
def whitespace_clean(text):
|
||||
text = re.sub(r"\s+", " ", text)
|
||||
text = text.strip()
|
||||
return text
|
||||
|
||||
def prompt_clean(text):
|
||||
text = whitespace_clean(basic_clean(text))
|
||||
return text
|
||||
|
||||
@dataclass
|
||||
class LongCatVideoPipelineOutput(BaseOutput):
|
||||
r"""
|
||||
Output class for CogVideo pipelines.
|
||||
|
||||
Args:
|
||||
video (`torch.Tensor`, `np.ndarray`, or List[List[PIL.Image.Image]]):
|
||||
List of video outputs - It can be a nested list of length `batch_size,` with each sub-list containing
|
||||
denoised PIL image sequences of length `num_frames.` It can also be a NumPy array or Torch tensor of shape
|
||||
`(batch_size, num_frames, channels, height, width)`.
|
||||
"""
|
||||
|
||||
videos: torch.Tensor
|
||||
|
||||
|
||||
class LongCatVideoPipeline(DiffusionPipeline):
|
||||
r"""
|
||||
Pipeline for text-to-video generation using Wan.
|
||||
|
||||
This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
|
||||
library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
|
||||
"""
|
||||
|
||||
_optional_components = []
|
||||
model_cpu_offload_seq = "text_encoder->transformer->vae"
|
||||
|
||||
_callback_tensor_inputs = [
|
||||
"latents",
|
||||
"prompt_embeds",
|
||||
"negative_prompt_embeds",
|
||||
]
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
tokenizer: AutoTokenizer,
|
||||
text_encoder: UMT5EncoderModel,
|
||||
vae: AutoencoderKLWan,
|
||||
transformer: LongCatVideoTransformer3DModel,
|
||||
scheduler: FlowMatchEulerDiscreteScheduler,
|
||||
):
|
||||
super().__init__()
|
||||
|
||||
self.register_modules(
|
||||
tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler
|
||||
)
|
||||
self.video_processor = VideoProcessor(vae_scale_factor=self.vae.scale_factor_spatial)
|
||||
self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae.scale_factor_spatial)
|
||||
self.mask_processor = VaeImageProcessor(
|
||||
vae_scale_factor=self.vae.scale_factor_spatial, do_normalize=False, do_binarize=True, do_convert_grayscale=True
|
||||
)
|
||||
|
||||
def _get_t5_prompt_embeds(
|
||||
self,
|
||||
prompt: Union[str, List[str]] = None,
|
||||
num_videos_per_prompt: int = 1,
|
||||
max_sequence_length: int = 512,
|
||||
device: Optional[torch.device] = None,
|
||||
dtype: Optional[torch.dtype] = None,
|
||||
):
|
||||
dtype = dtype or self.text_encoder.dtype
|
||||
|
||||
prompt = [prompt] if isinstance(prompt, str) else prompt
|
||||
prompt = [prompt_clean(u) for u in prompt]
|
||||
batch_size = len(prompt)
|
||||
|
||||
text_inputs = self.tokenizer(
|
||||
prompt,
|
||||
padding="max_length",
|
||||
max_length=max_sequence_length,
|
||||
truncation=True,
|
||||
add_special_tokens=True,
|
||||
return_attention_mask=True,
|
||||
return_tensors="pt",
|
||||
)
|
||||
text_input_ids, mask = text_inputs.input_ids, text_inputs.attention_mask
|
||||
|
||||
prompt_embeds = self.text_encoder(text_input_ids.to(device), mask.to(device)).last_hidden_state
|
||||
prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
|
||||
mask = mask.to(device=device)
|
||||
|
||||
# duplicate text embeddings for each generation per prompt, using mps friendly method
|
||||
_, seq_len, _ = prompt_embeds.shape
|
||||
prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
|
||||
prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, 1, seq_len, -1)
|
||||
|
||||
return prompt_embeds, mask
|
||||
|
||||
def encode_prompt(
|
||||
self,
|
||||
prompt: Union[str, List[str]],
|
||||
negative_prompt: Optional[Union[str, List[str]]] = None,
|
||||
do_classifier_free_guidance: bool = True,
|
||||
num_videos_per_prompt: int = 1,
|
||||
max_sequence_length: int = 512,
|
||||
device: Optional[torch.device] = None,
|
||||
dtype: Optional[torch.dtype] = None,
|
||||
):
|
||||
r"""
|
||||
Encodes the prompt into text encoder hidden states.
|
||||
|
||||
Args:
|
||||
prompt (`str` or `List[str]`, *optional*):
|
||||
prompt to be encoded
|
||||
num_videos_per_prompt (`int`, *optional*, defaults to 1):
|
||||
Number of videos that should be generated per prompt. torch device to place the resulting embeddings on
|
||||
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.
|
||||
negative_prompt_embeds (`torch.Tensor`, *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.
|
||||
device: (`torch.device`, *optional*):
|
||||
torch device
|
||||
dtype: (`torch.dtype`, *optional*):
|
||||
torch dtype
|
||||
"""
|
||||
|
||||
prompt = [prompt] if isinstance(prompt, str) else prompt
|
||||
batch_size = len(prompt)
|
||||
|
||||
prompt_embeds, prompt_attention_mask = self._get_t5_prompt_embeds(
|
||||
prompt=prompt,
|
||||
num_videos_per_prompt=num_videos_per_prompt,
|
||||
max_sequence_length=max_sequence_length,
|
||||
device=device,
|
||||
dtype=dtype,
|
||||
)
|
||||
|
||||
if do_classifier_free_guidance:
|
||||
negative_prompt = negative_prompt or ""
|
||||
negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
|
||||
|
||||
if 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)}."
|
||||
)
|
||||
|
||||
negative_prompt_embeds, negative_prompt_attention_mask = self._get_t5_prompt_embeds(
|
||||
prompt=negative_prompt,
|
||||
num_videos_per_prompt=num_videos_per_prompt,
|
||||
max_sequence_length=max_sequence_length,
|
||||
device=device,
|
||||
dtype=dtype,
|
||||
)
|
||||
else:
|
||||
negative_prompt_embeds = None
|
||||
negative_prompt_attention_mask = None
|
||||
|
||||
return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask
|
||||
|
||||
def prepare_latents(
|
||||
self, batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents=None
|
||||
):
|
||||
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."
|
||||
)
|
||||
|
||||
shape = (
|
||||
batch_size,
|
||||
num_channels_latents,
|
||||
(num_frames - 1) // self.vae.scale_factor_temporal + 1,
|
||||
height // self.vae.scale_factor_spatial,
|
||||
width // self.vae.scale_factor_spatial,
|
||||
)
|
||||
|
||||
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
|
||||
if hasattr(self.scheduler, "init_noise_sigma"):
|
||||
latents = latents * self.scheduler.init_noise_sigma
|
||||
return latents
|
||||
|
||||
def prepare_mask_latents(
|
||||
self, mask, masked_image, batch_size, height, width, dtype, device, generator, do_classifier_free_guidance, noise_aug_strength
|
||||
):
|
||||
# resize the mask to latents shape as we concatenate the mask to the latents
|
||||
# we do that before converting to dtype to avoid breaking in case we're using cpu_offload
|
||||
# and half precision
|
||||
|
||||
if mask is not None:
|
||||
mask = mask.to(device=device, dtype=self.vae.dtype)
|
||||
bs = 1
|
||||
new_mask = []
|
||||
for i in range(0, mask.shape[0], bs):
|
||||
mask_bs = mask[i : i + bs]
|
||||
mask_bs = self.vae.encode(mask_bs)[0]
|
||||
mask_bs = mask_bs.mode()
|
||||
new_mask.append(mask_bs)
|
||||
mask = torch.cat(new_mask, dim = 0)
|
||||
mask = self.normalize_latents(mask)
|
||||
# mask = mask * self.vae.config.scaling_factor
|
||||
|
||||
if masked_image is not None:
|
||||
masked_image = masked_image.to(device=device, dtype=self.vae.dtype)
|
||||
bs = 1
|
||||
new_mask_pixel_values = []
|
||||
for i in range(0, masked_image.shape[0], bs):
|
||||
mask_pixel_values_bs = masked_image[i : i + bs]
|
||||
mask_pixel_values_bs = self.vae.encode(mask_pixel_values_bs)[0]
|
||||
mask_pixel_values_bs = mask_pixel_values_bs.mode()
|
||||
new_mask_pixel_values.append(mask_pixel_values_bs)
|
||||
masked_image_latents = torch.cat(new_mask_pixel_values, dim = 0)
|
||||
masked_image_latents = self.normalize_latents(masked_image_latents)
|
||||
# masked_image_latents = masked_image_latents * self.vae.config.scaling_factor
|
||||
else:
|
||||
masked_image_latents = None
|
||||
|
||||
return mask, masked_image_latents
|
||||
|
||||
def optimized_scale(self, positive_flat, negative_flat):
|
||||
""" from CFG-zero paper
|
||||
"""
|
||||
# Calculate dot production
|
||||
dot_product = torch.sum(positive_flat * negative_flat, dim=1, keepdim=True)
|
||||
# Squared norm of uncondition
|
||||
squared_norm = torch.sum(negative_flat ** 2, dim=1, keepdim=True) + 1e-8
|
||||
# st_star = v_condˆT * v_uncond / ||v_uncond||ˆ2
|
||||
st_star = dot_product / squared_norm
|
||||
return st_star
|
||||
|
||||
def normalize_latents(self, latents):
|
||||
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
|
||||
)
|
||||
return (latents - latents_mean) * latents_std
|
||||
|
||||
def denormalize_latents(self, latents):
|
||||
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
|
||||
)
|
||||
return latents / latents_std + latents_mean
|
||||
|
||||
def decode_latents(self, latents: torch.Tensor) -> torch.Tensor:
|
||||
frames = self.vae.decode(latents.to(self.vae.dtype)).sample
|
||||
frames = (frames / 2 + 0.5).clamp(0, 1)
|
||||
# we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
|
||||
frames = frames.cpu().float().numpy()
|
||||
return frames
|
||||
|
||||
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
|
||||
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
|
||||
|
||||
# Copied from diffusers.pipelines.latte.pipeline_latte.LattePipeline.check_inputs
|
||||
def check_inputs(
|
||||
self,
|
||||
prompt,
|
||||
height,
|
||||
width,
|
||||
negative_prompt,
|
||||
callback_on_step_end_tensor_inputs,
|
||||
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_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 prompt is not None and negative_prompt_embeds is not None:
|
||||
raise ValueError(
|
||||
f"Cannot forward both `prompt`: {prompt} and `negative_prompt_embeds`:"
|
||||
f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
|
||||
)
|
||||
|
||||
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}."
|
||||
)
|
||||
|
||||
@property
|
||||
def guidance_scale(self):
|
||||
return self._guidance_scale
|
||||
|
||||
@property
|
||||
def num_timesteps(self):
|
||||
return self._num_timesteps
|
||||
|
||||
@property
|
||||
def attention_kwargs(self):
|
||||
return self._attention_kwargs
|
||||
|
||||
@property
|
||||
def interrupt(self):
|
||||
return self._interrupt
|
||||
|
||||
@torch.no_grad()
|
||||
@replace_example_docstring(EXAMPLE_DOC_STRING)
|
||||
def __call__(
|
||||
self,
|
||||
prompt: Optional[Union[str, List[str]]] = None,
|
||||
negative_prompt: Optional[Union[str, List[str]]] = None,
|
||||
height: int = 480,
|
||||
width: int = 720,
|
||||
video: Union[torch.FloatTensor] = None,
|
||||
mask_video: Union[torch.FloatTensor] = None,
|
||||
num_frames: int = 49,
|
||||
num_inference_steps: int = 50,
|
||||
timesteps: Optional[List[int]] = None,
|
||||
guidance_scale: float = 6,
|
||||
num_videos_per_prompt: 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: str = "numpy",
|
||||
return_dict: bool = False,
|
||||
callback_on_step_end: Optional[
|
||||
Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks]
|
||||
] = None,
|
||||
attention_kwargs: Optional[Dict[str, Any]] = None,
|
||||
callback_on_step_end_tensor_inputs: List[str] = ["latents"],
|
||||
max_sequence_length: int = 512,
|
||||
comfyui_progressbar: bool = False,
|
||||
shift: int = 5,
|
||||
) -> Union[LongCatVideoPipelineOutput, Tuple]:
|
||||
"""
|
||||
Function invoked when calling the pipeline for generation.
|
||||
Args:
|
||||
|
||||
Examples:
|
||||
|
||||
Returns:
|
||||
|
||||
"""
|
||||
|
||||
if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
|
||||
callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
|
||||
num_videos_per_prompt = 1
|
||||
|
||||
# 1. Check inputs. Raise error if not correct
|
||||
self.check_inputs(
|
||||
prompt,
|
||||
height,
|
||||
width,
|
||||
negative_prompt,
|
||||
callback_on_step_end_tensor_inputs,
|
||||
prompt_embeds,
|
||||
negative_prompt_embeds,
|
||||
)
|
||||
self._guidance_scale = guidance_scale
|
||||
self._attention_kwargs = attention_kwargs
|
||||
self._interrupt = False
|
||||
|
||||
# 2. Default 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
|
||||
weight_dtype = self.text_encoder.dtype
|
||||
|
||||
# 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
|
||||
(
|
||||
prompt_embeds,
|
||||
prompt_attention_mask,
|
||||
negative_prompt_embeds,
|
||||
negative_prompt_attention_mask,
|
||||
) = self.encode_prompt(
|
||||
prompt=prompt,
|
||||
negative_prompt=negative_prompt,
|
||||
do_classifier_free_guidance=do_classifier_free_guidance,
|
||||
num_videos_per_prompt=num_videos_per_prompt,
|
||||
max_sequence_length=max_sequence_length,
|
||||
device=device,
|
||||
)
|
||||
if do_classifier_free_guidance:
|
||||
prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
|
||||
prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0)
|
||||
|
||||
# 4. Prepare timesteps
|
||||
if isinstance(self.scheduler, FlowMatchEulerDiscreteScheduler):
|
||||
timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps, mu=1)
|
||||
elif isinstance(self.scheduler, FlowUniPCMultistepScheduler):
|
||||
self.scheduler.set_timesteps(num_inference_steps, device=device, shift=shift)
|
||||
timesteps = self.scheduler.timesteps
|
||||
elif isinstance(self.scheduler, FlowDPMSolverMultistepScheduler):
|
||||
sampling_sigmas = get_sampling_sigmas(num_inference_steps, shift)
|
||||
timesteps, _ = retrieve_timesteps(
|
||||
self.scheduler,
|
||||
device=device,
|
||||
sigmas=sampling_sigmas)
|
||||
else:
|
||||
timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
|
||||
self._num_timesteps = len(timesteps)
|
||||
if comfyui_progressbar:
|
||||
from comfy.utils import ProgressBar
|
||||
pbar = ProgressBar(num_inference_steps + 1)
|
||||
|
||||
if video is not None:
|
||||
video_length = video.shape[2]
|
||||
init_video = self.image_processor.preprocess(rearrange(video, "b c f h w -> (b f) c h w"), height=height, width=width)
|
||||
init_video = init_video.to(dtype=torch.float32)
|
||||
init_video = rearrange(init_video, "(b f) c h w -> b c f h w", f=video_length)
|
||||
else:
|
||||
init_video = None
|
||||
|
||||
# 5. Prepare latents
|
||||
latent_channels = self.transformer.config.in_channels
|
||||
latents = self.prepare_latents(
|
||||
batch_size * num_videos_per_prompt,
|
||||
latent_channels,
|
||||
num_frames,
|
||||
height,
|
||||
width,
|
||||
weight_dtype,
|
||||
device,
|
||||
generator,
|
||||
latents,
|
||||
)
|
||||
if comfyui_progressbar:
|
||||
pbar.update(1)
|
||||
|
||||
# Prepare mask latent variables
|
||||
if init_video is not None and not (mask_video == 255).all():
|
||||
bs, _, video_length, height, width = video.size()
|
||||
mask_condition = self.mask_processor.preprocess(rearrange(mask_video, "b c f h w -> (b f) c h w"), height=height, width=width)
|
||||
mask_condition = mask_condition.to(dtype=torch.float32)
|
||||
mask_condition = rearrange(mask_condition, "(b f) c h w -> b c f h w", f=video_length)
|
||||
|
||||
masked_video = init_video * (torch.tile(mask_condition, [1, 3, 1, 1, 1]) < 0.5)
|
||||
_, masked_video_latents = self.prepare_mask_latents(
|
||||
None,
|
||||
masked_video,
|
||||
batch_size,
|
||||
height,
|
||||
width,
|
||||
weight_dtype,
|
||||
device,
|
||||
generator,
|
||||
do_classifier_free_guidance,
|
||||
noise_aug_strength=None,
|
||||
)
|
||||
|
||||
mask_condition = torch.concat(
|
||||
[
|
||||
torch.repeat_interleave(mask_condition[:, :, 0:1], repeats=4, dim=2),
|
||||
mask_condition[:, :, 1:]
|
||||
], dim=2
|
||||
)
|
||||
mask_condition = mask_condition.view(bs, mask_condition.shape[2] // 4, 4, height, width)
|
||||
mask_condition = mask_condition.transpose(1, 2)
|
||||
|
||||
mask = F.interpolate(mask_condition[:, :1], size=latents.size()[-3:], mode='trilinear', align_corners=True).to(device, weight_dtype)
|
||||
latents = (1 - mask) * masked_video_latents + mask * latents
|
||||
else:
|
||||
init_video = None
|
||||
|
||||
# 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 = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
|
||||
self.transformer.num_inference_steps = num_inference_steps
|
||||
with self.progress_bar(total=num_inference_steps) as progress_bar:
|
||||
for i, t in enumerate(timesteps):
|
||||
self.transformer.current_steps = i
|
||||
|
||||
if self.interrupt:
|
||||
continue
|
||||
|
||||
latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
|
||||
if hasattr(self.scheduler, "scale_model_input"):
|
||||
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
|
||||
|
||||
# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
|
||||
timestep = t.expand(latent_model_input.shape[0])
|
||||
if init_video is not None:
|
||||
timestep = timestep.unsqueeze(-1).repeat(1, latent_model_input.shape[2])
|
||||
timestep[:, :1] = 0
|
||||
|
||||
# predict noise model_output
|
||||
with torch.cuda.amp.autocast(dtype=weight_dtype), torch.cuda.device(device=device):
|
||||
noise_pred = self.transformer(
|
||||
hidden_states=latent_model_input,
|
||||
timestep=timestep,
|
||||
encoder_hidden_states=prompt_embeds,
|
||||
encoder_attention_mask=prompt_attention_mask,
|
||||
num_cond_latents=1 if init_video is not None else 0
|
||||
)
|
||||
|
||||
# perform guidance
|
||||
if do_classifier_free_guidance:
|
||||
noise_pred_uncond, noise_pred_cond = noise_pred.chunk(2)
|
||||
|
||||
B = noise_pred_cond.shape[0]
|
||||
positive = noise_pred_cond.reshape(B, -1)
|
||||
negative = noise_pred_uncond.reshape(B, -1)
|
||||
# Calculate the optimized scale
|
||||
st_star = self.optimized_scale(positive, negative)
|
||||
# Reshape for broadcasting
|
||||
st_star = st_star.view(B, 1, 1, 1)
|
||||
# print(f'step i: {i} --> scale: {st_star}')
|
||||
|
||||
noise_pred = noise_pred_uncond * st_star + guidance_scale * (noise_pred_cond - noise_pred_uncond * st_star)
|
||||
|
||||
# negate for scheduler compatibility
|
||||
noise_pred = -noise_pred
|
||||
|
||||
# compute the previous noisy sample x_t -> x_t-1
|
||||
if init_video is not None:
|
||||
latents[:, :, 1:] = self.scheduler.step(noise_pred[:, :, 1:], t, latents[:, :, 1:], return_dict=False)[0]
|
||||
else:
|
||||
latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
|
||||
|
||||
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)
|
||||
|
||||
if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
|
||||
progress_bar.update()
|
||||
if comfyui_progressbar:
|
||||
pbar.update(1)
|
||||
|
||||
if output_type == "numpy":
|
||||
latents = self.denormalize_latents(latents)
|
||||
video = self.decode_latents(latents)
|
||||
elif not output_type == "latent":
|
||||
latents = self.denormalize_latents(latents)
|
||||
video = self.decode_latents(latents)
|
||||
video = self.video_processor.postprocess_video(video=video, output_type=output_type)
|
||||
else:
|
||||
video = latents
|
||||
|
||||
# Offload all models
|
||||
self.maybe_free_model_hooks()
|
||||
|
||||
if not return_dict:
|
||||
video = torch.from_numpy(video)
|
||||
|
||||
return LongCatVideoPipelineOutput(videos=video)
|
||||
@@ -157,7 +157,7 @@ class LoRANetwork(torch.nn.Module):
|
||||
"Wan2_2Transformer3DModel", "FluxTransformer2DModel", "QwenImageTransformer2DModel", \
|
||||
"Wan2_2Transformer3DModel_Animate", "Wan2_2Transformer3DModel_S2V", "FantasyTalkingTransformer3DModel", \
|
||||
"HunyuanVideoTransformer3DModel", "Flux2Transformer2DModel", "ZImageTransformer2DModel", \
|
||||
"ZImageOmniTransformer2DModel",
|
||||
"LongCatVideoTransformer3DModel", "ZImageOmniTransformer2DModel",
|
||||
]
|
||||
TEXT_ENCODER_TARGET_REPLACE_MODULE = ["T5LayerSelfAttention", "T5LayerFF", "BertEncoder", "T5SelfAttention", "T5CrossAttention"]
|
||||
LORA_PREFIX_TRANSFORMER = "lora_unet"
|
||||
|
||||
Reference in New Issue
Block a user