From 200e1f3224a2d71f8affcbe311a587a56a4ea7f4 Mon Sep 17 00:00:00 2001 From: Bubbliiiing <47347516+bubbliiiing@users.noreply.github.com> Date: Fri, 28 Nov 2025 10:31:06 +0800 Subject: [PATCH] Update flux2 (#383) --- README.md | 6 +- README_ja-JP.md | 6 +- README_zh-CN.md | 6 +- examples/flux/predict_t2i.py | 8 +- examples/flux2/predict_t2i.py | 211 +++ examples/hunyuanvideo/predict_i2v.py | 2 +- examples/hunyuanvideo/predict_t2v.py | 2 +- scripts/flux/train.py | 77 +- scripts/flux/train_lora.py | 56 +- scripts/flux2/README_TRAIN.md | 133 ++ scripts/flux2/README_TRAIN_LORA.md | 136 ++ scripts/flux2/train.py | 1778 ++++++++++++++++++++ scripts/flux2/train.sh | 32 + scripts/flux2/train_lora.py | 1722 +++++++++++++++++++ scripts/flux2/train_lora.sh | 33 + scripts/hunyuanvideo/train_lora.py | 2 +- videox_fun/dist/__init__.py | 1 + videox_fun/dist/flux2_xfuser.py | 194 +++ videox_fun/dist/flux_xfuser.py | 21 +- videox_fun/models/__init__.py | 14 +- videox_fun/models/flux2_image_processor.py | 139 ++ videox_fun/models/flux2_transformer2d.py | 1278 ++++++++++++++ videox_fun/models/flux2_vae.py | 543 ++++++ videox_fun/models/flux_transformer2d.py | 130 -- videox_fun/pipeline/__init__.py | 1 + videox_fun/pipeline/pipeline_flux2.py | 900 ++++++++++ videox_fun/utils/lora_utils.py | 2 +- 27 files changed, 7168 insertions(+), 265 deletions(-) create mode 100644 examples/flux2/predict_t2i.py create mode 100644 scripts/flux2/README_TRAIN.md create mode 100644 scripts/flux2/README_TRAIN_LORA.md create mode 100644 scripts/flux2/train.py create mode 100644 scripts/flux2/train.sh create mode 100644 scripts/flux2/train_lora.py create mode 100644 scripts/flux2/train_lora.sh create mode 100644 videox_fun/dist/flux2_xfuser.py create mode 100644 videox_fun/models/flux2_image_processor.py create mode 100644 videox_fun/models/flux2_transformer2d.py create mode 100644 videox_fun/models/flux2_vae.py create mode 100644 videox_fun/pipeline/pipeline_flux2.py diff --git a/README.md b/README.md index 0be406e..f8532d4 100755 --- a/README.md +++ b/README.md @@ -615,7 +615,8 @@ V1.0: | Name | Storage | Hugging Face | Model Scope | Description | |--|--|--|--|--| -| FLUX.1-dev | [🤗Link](https://huggingface.co/black-forest-labs/FLUX.1-dev) | [😄Link](https://www.modelscope.cn/models/ai-modelscope/flux.1-dev) | Official FLUX.1-dev weights | +| FLUX.1-dev | [🤗Link](https://huggingface.co/black-forest-labs/FLUX.1-dev) | [😄Link](https://www.modelscope.cn/models/black-forest-labs/FLUX.1-dev) | Official FLUX.1-dev weights | +| FLUX.2-dev | [🤗Link](https://huggingface.co/black-forest-labs/FLUX.2-dev) | [😄Link](https://www.modelscope.cn/models/black-forest-labs/FLUX.2-dev) | Official FLUX.2-dev weights | ## 8. HunyuanVideo @@ -662,6 +663,9 @@ V1.1: - Diffusers: https://github.com/huggingface/diffusers - Qwen-Image: https://github.com/QwenLM/Qwen-Image - Self-Forcing: https://github.com/guandeh17/Self-Forcing +- Flux: https://github.com/black-forest-labs/flux +- Flux2: https://github.com/black-forest-labs/flux2 +- HunyuanVideo: https://github.com/Tencent-Hunyuan/HunyuanVideo - ComfyUI-KJNodes: https://github.com/kijai/ComfyUI-KJNodes - ComfyUI-EasyAnimateWrapper: https://github.com/kijai/ComfyUI-EasyAnimateWrapper - ComfyUI-CameraCtrl-Wrapper: https://github.com/chaojie/ComfyUI-CameraCtrl-Wrapper diff --git a/README_ja-JP.md b/README_ja-JP.md index dea5901..04a15e7 100755 --- a/README_ja-JP.md +++ b/README_ja-JP.md @@ -615,7 +615,8 @@ V1.0: | 名称 | ストレージ | Hugging Face | Model Scope | 説明 | |--|--|--|--|--| -| FLUX.1-dev | [🤗Link](https://huggingface.co/black-forest-labs/FLUX.1-dev) | [😄Link](https://www.modelscope.cn/models/ai-modelscope/flux.1-dev) | FLUX.1-dev 公式重み | +| FLUX.1-dev | [🤗Link](https://huggingface.co/black-forest-labs/FLUX.1-dev) | [😄Link](https://www.modelscope.cn/models/black-forest-labs/FLUX.1-dev)| FLUX.1-dev 公式重み | +| FLUX.2-dev | [🤗Link](https://huggingface.co/black-forest-labs/FLUX.2-dev) | [😄Link](https://www.modelscope.cn/models/black-forest-labs/FLUX.2-dev) | FLUX.2-dev 公式重み | ## 8. HunyuanVideo @@ -662,6 +663,9 @@ V1.1: - Diffusers: https://github.com/huggingface/diffusers - Qwen-Image: https://github.com/QwenLM/Qwen-Image - Self-Forcing: https://github.com/guandeh17/Self-Forcing +- Flux: https://github.com/black-forest-labs/flux +- Flux2: https://github.com/black-forest-labs/flux2 +- HunyuanVideo: https://github.com/Tencent-Hunyuan/HunyuanVideo - ComfyUI-KJNodes: https://github.com/kijai/ComfyUI-KJNodes - ComfyUI-EasyAnimateWrapper: https://github.com/kijai/ComfyUI-EasyAnimateWrapper - ComfyUI-CameraCtrl-Wrapper: https://github.com/chaojie/ComfyUI-CameraCtrl-Wrapper diff --git a/README_zh-CN.md b/README_zh-CN.md index bad5efc..4b5e219 100755 --- a/README_zh-CN.md +++ b/README_zh-CN.md @@ -604,7 +604,8 @@ V1.0: | 名称 | 存储空间 | Hugging Face | Model Scope | 描述 | |--|--|--|--|--| -| FLUX.1-dev | [🤗Link](https://huggingface.co/black-forest-labs/FLUX.1-dev) | [😄Link](https://www.modelscope.cn/models/ai-modelscope/flux.1-dev) | FLUX.1-dev官方权重 | +| FLUX.1-dev | [🤗Link](https://huggingface.co/black-forest-labs/FLUX.1-dev) | [😄Link](https://www.modelscope.cn/models/black-forest-labs/FLUX.1-dev) | FLUX.1-dev官方权重 | +| FLUX.2-dev | [🤗Link](https://huggingface.co/black-forest-labs/FLUX.2-dev) | [😄Link](https://www.modelscope.cn/models/black-forest-labs/FLUX.2-dev) | FLUX.2-dev官方权重 | ## 8. HunyuanVideo @@ -652,6 +653,9 @@ V1.1: - Diffusers: https://github.com/huggingface/diffusers - Qwen-Image: https://github.com/QwenLM/Qwen-Image - Self-Forcing: https://github.com/guandeh17/Self-Forcing +- Flux: https://github.com/black-forest-labs/flux +- Flux2: https://github.com/black-forest-labs/flux2 +- HunyuanVideo: https://github.com/Tencent-Hunyuan/HunyuanVideo - ComfyUI-KJNodes: https://github.com/kijai/ComfyUI-KJNodes - ComfyUI-EasyAnimateWrapper: https://github.com/kijai/ComfyUI-EasyAnimateWrapper - ComfyUI-CameraCtrl-Wrapper: https://github.com/chaojie/ComfyUI-CameraCtrl-Wrapper diff --git a/examples/flux/predict_t2i.py b/examples/flux/predict_t2i.py index c8518bf..5fb55e6 100644 --- a/examples/flux/predict_t2i.py +++ b/examples/flux/predict_t2i.py @@ -69,7 +69,7 @@ negative_prompt = "The video is not of a high quality, it has a low resoluti guidance_scale = 1.0 seed = 43 num_inference_steps = 50 -lora_weight = 0.70 +lora_weight = 0.55 save_path = "samples/flux-t2i" device = set_multi_gpus_devices(ulysses_degree, ring_degree) @@ -151,13 +151,11 @@ if ulysses_degree > 1 or ring_degree > 1: from functools import partial transformer.enable_multi_gpus_inference() if fsdp_dit: - shard_fn = partial(shard_model, device_id=device, param_dtype=weight_dtype) + shard_fn = partial(shard_model, device_id=device, param_dtype=weight_dtype, module_to_wrapper=list(transformer.transformer_blocks) + list(transformer.single_transformer_blocks)) pipeline.transformer = shard_fn(pipeline.transformer) print("Add FSDP DIT") if fsdp_text_encoder: - from functools import partial - from videox_fun.dist import set_multi_gpus_devices, shard_model - shard_fn = partial(shard_model, device_id=device, param_dtype=weight_dtype, module_to_wrapper=text_encoder.language_model.layers) + shard_fn = partial(shard_model, device_id=device, param_dtype=weight_dtype, module_to_wrapper=text_encoder.text_model.encoder.layers) text_encoder = shard_fn(text_encoder) print("Add FSDP TEXT ENCODER") diff --git a/examples/flux2/predict_t2i.py b/examples/flux2/predict_t2i.py new file mode 100644 index 0000000..dce3c3e --- /dev/null +++ b/examples/flux2/predict_t2i.py @@ -0,0 +1,211 @@ +import os +import sys + +import torch + +from diffusers import (FlowMatchEulerDiscreteScheduler) + +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 (AutoencoderKLFlux2, + Mistral3ForConditionalGeneration, + PixtralProcessor, Flux2Transformer2DModel) +from videox_fun.models.cache_utils import get_teacache_coefficients +from videox_fun.pipeline import Flux2Pipeline +from videox_fun.utils.fm_solvers import FlowDPMSolverMultistepScheduler +from videox_fun.utils.fm_solvers_unipc import FlowUniPCMultistepScheduler +from videox_fun.utils.fp8_optimization import (convert_model_weight_to_float8, + convert_weight_dtype_wrapper) +from videox_fun.utils.lora_utils import merge_lora, unmerge_lora + +# 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" +# Multi GPUs config +# Please ensure that the product of ulysses_degree and ring_degree equals the number of GPUs used. +# For example, if you are using 8 GPUs, you can set ulysses_degree = 2 and ring_degree = 4. +# If you are using 1 GPU, you can set ulysses_degree = 1 and ring_degree = 1. +ulysses_degree = 1 +ring_degree = 1 +# Use FSDP to save more GPU memory in multi gpus. +fsdp_dit = False +fsdp_text_encoder = False +# 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/FLUX.2-dev" + +# 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 = [1344, 768] + +# 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 +# 使用更长的neg prompt如"模糊，突变，变形，失真，画面暗，文本字幕，画面固定，连环画，漫画，线稿，没有主体。"，可以增加稳定性 +# 在neg prompt中添加"安静，固定"等词语可以增加动态性。 +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 = " " +guidance_scale = 4.00 +seed = 43 +num_inference_steps = 50 +lora_weight = 0.55 +save_path = "samples/flux2-t2i" + +device = set_multi_gpus_devices(ulysses_degree, ring_degree) + +transformer = Flux2Transformer2DModel.from_pretrained( + model_name, + subfolder="transformer", + low_cpu_mem_usage=True, + torch_dtype=weight_dtype, +).to(weight_dtype) + +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 = AutoencoderKLFlux2.from_pretrained( + model_name, + subfolder="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 and text_encoder +tokenizer = PixtralProcessor.from_pretrained( + model_name, subfolder="tokenizer" +) +text_encoder = Mistral3ForConditionalGeneration.from_pretrained( + model_name, subfolder="text_encoder", torch_dtype=weight_dtype, + low_cpu_mem_usage=True, +) + +# 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" +) + +pipeline = Flux2Pipeline( + vae=vae, + tokenizer=tokenizer, + text_encoder=text_encoder, + transformer=transformer, + scheduler=scheduler, +) + +if ulysses_degree > 1 or ring_degree > 1: + from functools import partial + transformer.enable_multi_gpus_inference() + if fsdp_dit: + shard_fn = partial(shard_model, device_id=device, param_dtype=weight_dtype, module_to_wrapper=list(transformer.transformer_blocks) + list(transformer.single_transformer_blocks)) + pipeline.transformer = shard_fn(pipeline.transformer) + print("Add FSDP DIT") + if fsdp_text_encoder: + shard_fn = partial(shard_model, device_id=device, param_dtype=weight_dtype, module_to_wrapper=text_encoder.language_model.layers) + text_encoder = shard_fn(text_encoder) + print("Add FSDP TEXT ENCODER") + +if compile_dit: + for i in range(len(pipeline.transformer.transformer_blocks)): + pipeline.transformer.transformer_blocks[i] = torch.compile(pipeline.transformer.transformer_blocks[i]) + print("Add Compile") + +if GPU_memory_mode == "sequential_cpu_offload": + 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=["img_in", "txt_in", "timestep"], 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=["img_in", "txt_in", "timestep"], 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(): + sample = pipeline( + prompt = prompt, + height = sample_size[0], + width = sample_size[1], + generator = generator, + guidance_scale = guidance_scale, + num_inference_steps = num_inference_steps, + ).images + +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) + video_path = os.path.join(save_path, prefix + ".png") + image = sample[0] + image.save(video_path) + +if ulysses_degree * ring_degree > 1: + import torch.distributed as dist + if dist.get_rank() == 0: + save_results() +else: + save_results() \ No newline at end of file diff --git a/examples/hunyuanvideo/predict_i2v.py b/examples/hunyuanvideo/predict_i2v.py index aa00e1b..22e8cf8 100644 --- a/examples/hunyuanvideo/predict_i2v.py +++ b/examples/hunyuanvideo/predict_i2v.py @@ -181,7 +181,7 @@ if ulysses_degree > 1 or ring_degree > 1: from functools import partial transformer.enable_multi_gpus_inference() if fsdp_dit: - shard_fn = partial(shard_model, device_id=device, param_dtype=weight_dtype, module_to_wrapper=transformer.transformer_blocks) + shard_fn = partial(shard_model, device_id=device, param_dtype=weight_dtype, module_to_wrapper=list(transformer.transformer_blocks) + list(transformer.single_transformer_blocks)) pipeline.transformer = shard_fn(pipeline.transformer) print("Add FSDP DIT") if fsdp_text_encoder: diff --git a/examples/hunyuanvideo/predict_t2v.py b/examples/hunyuanvideo/predict_t2v.py index 5b490f6..966186b 100644 --- a/examples/hunyuanvideo/predict_t2v.py +++ b/examples/hunyuanvideo/predict_t2v.py @@ -170,7 +170,7 @@ if ulysses_degree > 1 or ring_degree > 1: from functools import partial transformer.enable_multi_gpus_inference() if fsdp_dit: - shard_fn = partial(shard_model, device_id=device, param_dtype=weight_dtype, module_to_wrapper=transformer.transformer_blocks) + shard_fn = partial(shard_model, device_id=device, param_dtype=weight_dtype, module_to_wrapper=list(transformer.transformer_blocks) + list(transformer.single_transformer_blocks)) pipeline.transformer = shard_fn(pipeline.transformer) print("Add FSDP DIT") if fsdp_text_encoder: diff --git a/scripts/flux/train.py b/scripts/flux/train.py index 1610c7f..1bb94d3 100644 --- a/scripts/flux/train.py +++ b/scripts/flux/train.py @@ -24,6 +24,7 @@ import pickle import random import shutil import sys +from typing import List, NamedTuple, Optional, Union import accelerate import diffusers @@ -33,9 +34,6 @@ import torch.nn.functional as F import torch.utils.checkpoint import torchvision.transforms.functional as TF import transformers -from typing import NamedTuple, List, Optional, Union - - from accelerate import Accelerator from accelerate.logging import get_logger from accelerate.state import AcceleratorState @@ -60,7 +58,6 @@ from transformers.utils import ContextManagers import datasets - 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: @@ -74,14 +71,14 @@ from videox_fun.data.bucket_sampler import (ASPECT_RATIO_512, from videox_fun.data.dataset_image_video import (ImageVideoDataset, ImageVideoSampler, get_random_mask) -from videox_fun.models import (AutoencoderKL, AutoencoderKLWan, - Qwen2_5_VLForConditionalGeneration, - Qwen2Tokenizer, QwenImageTransformer2DModel) from videox_fun.dist import set_multi_gpus_devices, shard_model -from videox_fun.models import (CLIPImageProcessor, CLIPTextModel, +from videox_fun.models import (AutoencoderKL, AutoencoderKLWan, + CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection, - FluxTransformer2DModel, T5EncoderModel, - T5TokenizerFast) + FluxTransformer2DModel, + Qwen2_5_VLForConditionalGeneration, + Qwen2Tokenizer, QwenImageTransformer2DModel, + T5EncoderModel, T5TokenizerFast) from videox_fun.pipeline import FluxPipeline from videox_fun.utils.discrete_sampler import DiscreteSampling from videox_fun.utils.utils import get_image_to_video_latent, save_videos_grid @@ -109,12 +106,6 @@ def generate_timestep_with_lognorm(low, high, shape, device="cpu", generator=Non u = torch.normal(mean=0.0, std=1.0, size=shape, device=device, generator=generator) t = 1 / (1 + torch.exp(-u)) * (high - low) + low return torch.clip(t.to(torch.int32), low, high - 1) - -def _pack_latents(latents, batch_size, num_channels_latents, height, width): - latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) - latents = latents.permute(0, 2, 4, 1, 3, 5) - latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4) - return latents def calculate_shift( image_seq_len, @@ -128,6 +119,12 @@ def calculate_shift( mu = image_seq_len * m + b return mu +def _pack_latents(latents, batch_size, num_channels_latents, height, width): + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 2, 4, 1, 3, 5) + latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4) + return latents + def _extract_masked_hidden(hidden_states: torch.Tensor, mask: torch.Tensor): bool_mask = mask.bool() valid_lengths = bool_mask.sum(dim=1) @@ -149,20 +146,6 @@ def _prepare_latent_image_ids(batch_size, height, width, device, dtype): return latent_image_ids.to(device=device, dtype=dtype) -def _pack_latents(latents, batch_size, num_channels_latents, height, width): - latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) - latents = latents.permute(0, 2, 4, 1, 3, 5) - latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4) - return latents - -def _extract_masked_hidden(hidden_states: torch.Tensor, mask: torch.Tensor): - bool_mask = mask.bool() - valid_lengths = bool_mask.sum(dim=1) - selected = hidden_states[bool_mask] - split_result = torch.split(selected, valid_lengths.tolist(), dim=0) - - return split_result - def _get_t5_prompt_embeds( prompt = None, max_sequence_length = 512, @@ -645,12 +628,6 @@ def parse_args(): default=[], help='Enter a list of trainable modules with lower learning rate' ) - parser.add_argument( - '--tokenizer_max_length', - type=int, - default=1024, - help='Max length of tokenizer' - ) parser.add_argument( "--use_deepspeed", action="store_true", help="Whether or not to use deepspeed." ) @@ -660,31 +637,6 @@ def parse_args(): parser.add_argument( "--low_vram", action="store_true", help="Whether enable low_vram mode." ) - parser.add_argument( - "--prompt_template_encode", - type=str, - default="<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n", - help=( - 'The prompt template for text encoder.' - ), - ) - parser.add_argument( - "--prompt_template_encode_start_idx", - type=int, - default=34, - help=( - 'The start idx for prompt template.' - ), - ) - parser.add_argument( - "--train_mode", - type=str, - default="normal", - help=( - 'The format of training data. Support `"normal"`' - ' (default), `"i2v"`.' - ), - ) parser.add_argument( "--abnormal_norm_clip_start", type=int, @@ -1343,6 +1295,7 @@ def main(): if fsdp_stage != 0: from functools import partial + from videox_fun.dist import set_multi_gpus_devices, shard_model shard_fn = partial(shard_model, device_id=accelerator.device, param_dtype=weight_dtype, module_to_wrapper=text_encoder.text_model.encoder.layers) text_encoder = shard_fn(text_encoder) @@ -1435,7 +1388,7 @@ def main(): disable=not accelerator.is_local_main_process, ) - if args.multi_stream and args.train_mode != "normal": + if args.multi_stream: # create extra cuda streams to speedup inpaint vae computation vae_stream_1 = torch.cuda.Stream() vae_stream_2 = torch.cuda.Stream() diff --git a/scripts/flux/train_lora.py b/scripts/flux/train_lora.py index 1bb399d..ad71f90 100644 --- a/scripts/flux/train_lora.py +++ b/scripts/flux/train_lora.py @@ -16,6 +16,7 @@ # See the License for the specific language governing permissions and import argparse +import copy import gc import logging import math @@ -24,7 +25,7 @@ import pickle import random import shutil import sys -import copy +from typing import List, NamedTuple, Optional, Union import accelerate import diffusers @@ -34,9 +35,6 @@ import torch.nn.functional as F import torch.utils.checkpoint import torchvision.transforms.functional as TF import transformers -from typing import NamedTuple, List, Optional, Union - - from accelerate import Accelerator from accelerate.logging import get_logger from accelerate.state import AcceleratorState @@ -73,14 +71,14 @@ from videox_fun.data.bucket_sampler import (ASPECT_RATIO_512, from videox_fun.data.dataset_image_video import (ImageVideoDataset, ImageVideoSampler, get_random_mask) -from videox_fun.models import (AutoencoderKL, AutoencoderKLWan, - Qwen2_5_VLForConditionalGeneration, - Qwen2Tokenizer, QwenImageTransformer2DModel) from videox_fun.dist import set_multi_gpus_devices, shard_model -from videox_fun.models import (CLIPImageProcessor, CLIPTextModel, +from videox_fun.models import (AutoencoderKL, AutoencoderKLWan, + CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection, - FluxTransformer2DModel, T5EncoderModel, - T5TokenizerFast) + FluxTransformer2DModel, + Qwen2_5_VLForConditionalGeneration, + Qwen2Tokenizer, QwenImageTransformer2DModel, + T5EncoderModel, T5TokenizerFast) from videox_fun.pipeline import FluxPipeline from videox_fun.utils.discrete_sampler import DiscreteSampling from videox_fun.utils.lora_utils import (create_network, merge_lora, @@ -642,12 +640,6 @@ def parse_args(): ) parser.add_argument("--save_state", action="store_true", help="Whether or not to save state.") - parser.add_argument( - '--tokenizer_max_length', - type=int, - default=1024, - help='Max length of tokenizer' - ) parser.add_argument( "--use_deepspeed", action="store_true", help="Whether or not to use deepspeed." ) @@ -657,31 +649,6 @@ def parse_args(): parser.add_argument( "--low_vram", action="store_true", help="Whether enable low_vram mode." ) - parser.add_argument( - "--prompt_template_encode", - type=str, - default="<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n", - help=( - 'The prompt template for text encoder.' - ), - ) - parser.add_argument( - "--prompt_template_encode_start_idx", - type=int, - default=34, - help=( - 'The start idx for prompt template.' - ), - ) - parser.add_argument( - "--train_mode", - type=str, - default="normal", - help=( - 'The format of training data. Support `"normal"`' - ' (default), `"inpaint"`.' - ), - ) parser.add_argument( "--weighting_scheme", type=str, @@ -904,7 +871,8 @@ def main(): # Lora will work with this... if args.use_peft_lora: - from peft import LoraConfig, inject_adapter_in_model, get_peft_model_state_dict + from peft import (LoraConfig, get_peft_model_state_dict, + inject_adapter_in_model) lora_config = LoraConfig(r=args.rank, lora_alpha=args.network_alpha, target_modules=args.target_name.split(",")) transformer3d = inject_adapter_in_model(lora_config, transformer3d) @@ -1310,7 +1278,7 @@ def main(): from functools import partial from videox_fun.dist import set_multi_gpus_devices, shard_model - shard_fn = partial(shard_model, device_id=accelerator.device, param_dtype=weight_dtype, module_to_wrapper=transformer3d.transformer_blocks) + shard_fn = partial(shard_model, device_id=accelerator.device, param_dtype=weight_dtype, module_to_wrapper=list(transformer3d.transformer_blocks) + list(transformer3d.single_transformer_blocks)) transformer3d = shard_fn(transformer3d) if fsdp_stage != 0 or zero_stage != 0: @@ -1474,7 +1442,7 @@ def main(): disable=not accelerator.is_local_main_process, ) - if args.multi_stream and args.train_mode != "normal": + if args.multi_stream: # create extra cuda streams to speedup inpaint vae computation vae_stream_1 = torch.cuda.Stream() vae_stream_2 = torch.cuda.Stream() diff --git a/scripts/flux2/README_TRAIN.md b/scripts/flux2/README_TRAIN.md new file mode 100644 index 0000000..d1a4d49 --- /dev/null +++ b/scripts/flux2/README_TRAIN.md @@ -0,0 +1,133 @@ +## Training Code + +We can choose whether to use deepspeed or fsdp in flux2, 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 at the center, but instead, it trains the entire images after grouping them into buckets based on resolution. +- `random_hw_adapt` is used to enable automatic height and width scaling for images. When `random_hw_adapt` is enabled, the training images will have their height and width set to `image_sample_size` as the maximum and `512` as the minimum. + - For example, when `random_hw_adapt` is enabled, `image_sample_size=1024`, the resolution of image inputs for training is `512x512` to `1024x1024` +- `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=$RANK scripts/xxx/xxx.py +``` + +Without deepspeed: + +Training flux2 without DeepSpeed may result in insufficient GPU memory. +```sh +export MODEL_NAME="models/Diffusion_Transformer/FLUX.1-dev" +export DATASET_NAME="datasets/internal_datasets/" +export DATASET_META_NAME="datasets/internal_datasets/metadata.json" +# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA. +# export NCCL_IB_DISABLE=1 +# export NCCL_P2P_DISABLE=1 +NCCL_DEBUG=INFO + +accelerate launch --mixed_precision="bf16" scripts/flux2/train.py \ + --pretrained_model_name_or_path=$MODEL_NAME \ + --train_data_dir=$DATASET_NAME \ + --train_data_meta=$DATASET_META_NAME \ + --train_batch_size=1 \ + --image_sample_size=1024 \ + --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" \ + --gradient_checkpointing \ + --mixed_precision="bf16" \ + --adam_weight_decay=3e-2 \ + --adam_epsilon=1e-10 \ + --vae_mini_batch=1 \ + --max_grad_norm=0.05 \ + --enable_bucket \ + --uniform_sampling \ + --trainable_modules "." +``` + +With Deepspeed Zero-2: + +```sh +export MODEL_NAME="models/Diffusion_Transformer/FLUX.1-dev" +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/flux2/train.py \ + --pretrained_model_name_or_path=$MODEL_NAME \ + --train_data_dir=$DATASET_NAME \ + --train_data_meta=$DATASET_META_NAME \ + --train_batch_size=1 \ + --image_sample_size=1024 \ + --gradient_accumulation_steps=1 \ + --dataloader_num_workers=8 \ + --num_train_epochs=100 \ + --checkpointing_steps=50 \ + --learning_rate=2e-05 \ + --lr_scheduler="constant_with_warmup" \ + --lr_warmup_steps=100 \ + --seed=42 \ + --output_dir="output_dir" \ + --gradient_checkpointing \ + --mixed_precision="bf16" \ + --adam_weight_decay=3e-2 \ + --adam_epsilon=1e-10 \ + --vae_mini_batch=1 \ + --max_grad_norm=0.05 \ + --enable_bucket \ + --uniform_sampling \ + --trainable_modules "." +``` + +With FSDP: + +```sh +export MODEL_NAME="models/Diffusion_Transformer/FLUX.1-dev" +export DATASET_NAME="datasets/internal_datasets/" +export DATASET_META_NAME="datasets/internal_datasets/metadata.json" +# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA. +# export NCCL_IB_DISABLE=1 +# export NCCL_P2P_DISABLE=1 +NCCL_DEBUG=INFO + +accelerate launch --mixed_precision="bf16" --use_fsdp --fsdp_auto_wrap_policy TRANSFORMER_BASED_WRAP --fsdp_transformer_layer_cls_to_wrap Flux2SingleTransformerBlock,Flux2TransformerBlock --fsdp_sharding_strategy "FULL_SHARD" --fsdp_state_dict_type=SHARDED_STATE_DICT --fsdp_backward_prefetch "BACKWARD_PRE" --fsdp_cpu_ram_efficient_loading False scripts/flux2/train.py \ + --pretrained_model_name_or_path=$MODEL_NAME \ + --train_data_dir=$DATASET_NAME \ + --train_data_meta=$DATASET_META_NAME \ + --train_batch_size=1 \ + --image_sample_size=1024 \ + --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" \ + --gradient_checkpointing \ + --mixed_precision="bf16" \ + --adam_weight_decay=3e-2 \ + --adam_epsilon=1e-10 \ + --vae_mini_batch=1 \ + --max_grad_norm=0.05 \ + --enable_bucket \ + --uniform_sampling \ + --trainable_modules "." +``` \ No newline at end of file diff --git a/scripts/flux2/README_TRAIN_LORA.md b/scripts/flux2/README_TRAIN_LORA.md new file mode 100644 index 0000000..e148015 --- /dev/null +++ b/scripts/flux2/README_TRAIN_LORA.md @@ -0,0 +1,136 @@ +## Lora Training Code + +We can choose whether to use deepspeed or fsdp in flux, 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 at the center, but instead, it trains the entire images after grouping them into buckets based on resolution. +- `random_hw_adapt` is used to enable automatic height and width scaling for images. When `random_hw_adapt` is enabled, the training images will have their height and width set to `image_sample_size` as the maximum and `512` as the minimum. + - For example, when `random_hw_adapt` is enabled, `image_sample_size=1024`, the resolution of image inputs for training is `512x512` to `1024x1024` +- `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=$RANK scripts/xxx/xxx.py +``` + +Without deepspeed: + +Training flux without DeepSpeed may result in insufficient GPU memory. +```sh +export MODEL_NAME="models/Diffusion_Transformer/FLUX.1-dev" +export DATASET_NAME="datasets/internal_datasets/" +export DATASET_META_NAME="datasets/internal_datasets/metadata.json" +# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA. +# export NCCL_IB_DISABLE=1 +# export NCCL_P2P_DISABLE=1 +NCCL_DEBUG=INFO + +accelerate launch --mixed_precision="bf16" scripts/flux/train_lora.py \ + --pretrained_model_name_or_path=$MODEL_NAME \ + --train_data_dir=$DATASET_NAME \ + --train_data_meta=$DATASET_META_NAME \ + --train_batch_size=1 \ + --image_sample_size=1024 \ + --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" \ + --gradient_checkpointing \ + --mixed_precision="bf16" \ + --adam_weight_decay=3e-2 \ + --adam_epsilon=1e-10 \ + --vae_mini_batch=1 \ + --max_grad_norm=0.05 \ + --enable_bucket \ + --rank=64 \ + --network_alpha=32 \ + --target_name="to_q,to_k,to_v,ff.0,ff.2,ff_context.0,ff_context.2" \ + --use_peft_lora \ + --uniform_sampling +``` + +With Deepspeed Zero-2: + +```sh +export MODEL_NAME="models/Diffusion_Transformer/FLUX.1-dev" +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/flux/train_lora.py \ + --pretrained_model_name_or_path=$MODEL_NAME \ + --train_data_dir=$DATASET_NAME \ + --train_data_meta=$DATASET_META_NAME \ + --train_batch_size=1 \ + --image_sample_size=1024 \ + --gradient_accumulation_steps=1 \ + --dataloader_num_workers=8 \ + --num_train_epochs=100 \ + --checkpointing_steps=50 \ + --learning_rate=1e-04 \ + --seed=42 \ + --output_dir="output_dir" \ + --gradient_checkpointing \ + --mixed_precision="bf16" \ + --adam_weight_decay=3e-2 \ + --adam_epsilon=1e-10 \ + --vae_mini_batch=1 \ + --max_grad_norm=0.05 \ + --enable_bucket \ + --rank=64 \ + --network_alpha=32 \ + --target_name="to_q,to_k,to_v,ff.0,ff.2,ff_context.0,ff_context.2" \ + --use_peft_lora \ + --uniform_sampling +``` + +With FSDP: + +```sh +export MODEL_NAME="models/Diffusion_Transformer/Wan2.2-Fun-A14B-InP" +export DATASET_NAME="datasets/internal_datasets/" +export DATASET_META_NAME="datasets/internal_datasets/metadata.json" +# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA. +# export NCCL_IB_DISABLE=1 +# export NCCL_P2P_DISABLE=1 +NCCL_DEBUG=INFO + +accelerate launch --mixed_precision="bf16" --use_fsdp --fsdp_auto_wrap_policy TRANSFORMER_BASED_WRAP --fsdp_transformer_layer_cls_to_wrap Flux2SingleTransformerBlock,Flux2TransformerBlock --fsdp_sharding_strategy "FULL_SHARD" --fsdp_state_dict_type=SHARDED_STATE_DICT --fsdp_backward_prefetch "BACKWARD_PRE" --fsdp_cpu_ram_efficient_loading False scripts/flux/train_lora.py \ + --pretrained_model_name_or_path=$MODEL_NAME \ + --train_data_dir=$DATASET_NAME \ + --train_data_meta=$DATASET_META_NAME \ + --train_batch_size=1 \ + --image_sample_size=1024 \ + --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" \ + --gradient_checkpointing \ + --mixed_precision="bf16" \ + --adam_weight_decay=3e-2 \ + --adam_epsilon=1e-10 \ + --vae_mini_batch=1 \ + --max_grad_norm=0.05 \ + --enable_bucket \ + --uniform_sampling +``` \ No newline at end of file diff --git a/scripts/flux2/train.py b/scripts/flux2/train.py new file mode 100644 index 0000000..23455e4 --- /dev/null +++ b/scripts/flux2/train.py @@ -0,0 +1,1778 @@ +"""Modified from https://github.com/huggingface/diffusers/blob/main/examples/text_to_image/train_text_to_image.py +""" +#!/usr/bin/env python +# coding=utf-8 +# Copyright 2024 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and + +import argparse +import gc +import logging +import math +import os +import pickle +import random +import shutil +import sys +from typing import List, NamedTuple, Optional, Union +from typing import Any, Callable, Dict, List, Optional, Tuple, Union + +import accelerate +import diffusers +import numpy as np +import torch +import torch.nn.functional as F +import torch.utils.checkpoint +import torchvision.transforms.functional as TF +import transformers +from accelerate import Accelerator +from accelerate.logging import get_logger +from accelerate.state import AcceleratorState +from accelerate.utils import ProjectConfiguration, set_seed +from diffusers import DDIMScheduler, FlowMatchEulerDiscreteScheduler +from diffusers.optimization import get_scheduler +from diffusers.training_utils import (EMAModel, + compute_density_for_timestep_sampling, + compute_loss_weighting_for_sd3) +from diffusers.utils import check_min_version, deprecate, is_wandb_available +from diffusers.utils.torch_utils import is_compiled_module +from einops import rearrange +from omegaconf import OmegaConf +from packaging import version +from PIL import Image +from torch.utils.data import RandomSampler +from torch.utils.tensorboard import SummaryWriter +from torchvision import transforms +from tqdm.auto import tqdm +from transformers import AutoTokenizer +from transformers.utils import ContextManagers + +import datasets + +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.data.bucket_sampler import (ASPECT_RATIO_512, + ASPECT_RATIO_RANDOM_CROP_512, + ASPECT_RATIO_RANDOM_CROP_PROB, + AspectRatioBatchImageVideoSampler, + RandomSampler, get_closest_ratio) +from videox_fun.data.dataset_image_video import (ImageVideoDataset, + ImageVideoSampler, + get_random_mask) +from videox_fun.dist import set_multi_gpus_devices, shard_model +from videox_fun.models import (AutoencoderKLFlux2, + CLIPImageProcessor, Mistral3ForConditionalGeneration, + PixtralProcessor, CLIPVisionModelWithProjection, + Flux2Transformer2DModel, + Mistral3ForConditionalGeneration, + Qwen2_5_VLForConditionalGeneration, + Qwen2Tokenizer, QwenImageTransformer2DModel) +from videox_fun.pipeline import Flux2Pipeline +from videox_fun.utils.discrete_sampler import DiscreteSampling +from videox_fun.utils.utils import get_image_to_video_latent, save_videos_grid + +if is_wandb_available(): + import wandb + + +def filter_kwargs(cls, kwargs): + import inspect + sig = inspect.signature(cls.__init__) + valid_params = set(sig.parameters.keys()) - {'self', 'cls'} + filtered_kwargs = {k: v for k, v in kwargs.items() if k in valid_params} + return filtered_kwargs + +def linear_decay(initial_value, final_value, total_steps, current_step): + if current_step >= total_steps: + return final_value + current_step = max(0, current_step) + step_size = (final_value - initial_value) / total_steps + current_value = initial_value + step_size * current_step + return current_value + +def generate_timestep_with_lognorm(low, high, shape, device="cpu", generator=None): + u = torch.normal(mean=0.0, std=1.0, size=shape, device=device, generator=generator) + t = 1 / (1 + torch.exp(-u)) * (high - low) + low + return torch.clip(t.to(torch.int32), low, high - 1) + +def compute_empirical_mu(image_seq_len: int, num_steps: int) -> float: + a1, b1 = 8.73809524e-05, 1.89833333 + a2, b2 = 0.00016927, 0.45666666 + + if image_seq_len > 4300: + mu = a2 * image_seq_len + b2 + return float(mu) + + m_200 = a2 * image_seq_len + b2 + m_10 = a1 * image_seq_len + b1 + + a = (m_200 - m_10) / 190.0 + b = m_200 - 200.0 * a + mu = a * num_steps + b + + return float(mu) + +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + +def _prepare_latent_ids( + latents: torch.Tensor, # (B, C, H, W) +): + r""" + Generates 4D position coordinates (T, H, W, L) for latent tensors. + + Args: + latents (torch.Tensor): + Latent tensor of shape (B, C, H, W) + + Returns: + torch.Tensor: + Position IDs tensor of shape (B, H*W, 4) All batches share the same coordinate structure: T=0, + H=[0..H-1], W=[0..W-1], L=0 + """ + + batch_size, _, height, width = latents.shape + + t = torch.arange(1) # [0] - time dimension + h = torch.arange(height) + w = torch.arange(width) + l = torch.arange(1) # [0] - layer dimension + + # Create position IDs: (H*W, 4) + latent_ids = torch.cartesian_prod(t, h, w, l) + + # Expand to batch: (B, H*W, 4) + latent_ids = latent_ids.unsqueeze(0).expand(batch_size, -1, -1) + + return latent_ids + +def _patchify_latents(latents): + batch_size, num_channels_latents, height, width = latents.shape + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 1, 3, 5, 2, 4) + latents = latents.reshape(batch_size, num_channels_latents * 4, height // 2, width // 2) + return latents + +def _pack_latents(latents): + """ + pack latents: (batch_size, num_channels, height, width) -> (batch_size, height * width, num_channels) + """ + + batch_size, num_channels, height, width = latents.shape + latents = latents.reshape(batch_size, num_channels, height * width).permute(0, 2, 1) + + return latents + +def format_text_input(prompts: List[str], system_message: str = None): + # Remove [IMG] tokens from prompts to avoid Pixtral validation issues + # when truncation is enabled. The processor counts [IMG] tokens and fails + # if the count changes after truncation. + cleaned_txt = [prompt.replace("[IMG]", "") for prompt in prompts] + + return [ + [ + { + "role": "system", + "content": [{"type": "text", "text": system_message}], + }, + {"role": "user", "content": [{"type": "text", "text": prompt}]}, + ] + for prompt in cleaned_txt + ] + +def _get_mistral_3_small_prompt_embeds( + text_encoder: Mistral3ForConditionalGeneration, + tokenizer: PixtralProcessor, + prompt: Union[str, List[str]], + dtype: Optional[torch.dtype] = None, + device: Optional[torch.device] = None, + max_sequence_length: int = 512, + # fmt: off + system_message: str = "You are an AI that reasons about image descriptions. You give structured responses focusing on object relationships, object attribution and actions without speculation.", + # fmt: on + hidden_states_layers: List[int] = (10, 20, 30), +): + dtype = text_encoder.dtype if dtype is None else dtype + device = text_encoder.device if device is None else device + + prompt = [prompt] if isinstance(prompt, str) else prompt + + # Format input messages + messages_batch = format_text_input(prompts=prompt, system_message=system_message) + + # Process all messages at once + inputs = tokenizer.apply_chat_template( + messages_batch, + add_generation_prompt=False, + tokenize=True, + return_dict=True, + return_tensors="pt", + padding="max_length", + truncation=True, + max_length=max_sequence_length, + ) + + # Move to device + input_ids = inputs["input_ids"].to(device) + attention_mask = inputs["attention_mask"].to(device) + + # Forward pass through the model + output = text_encoder( + input_ids=input_ids, + attention_mask=attention_mask, + output_hidden_states=True, + use_cache=False, + ) + + # Only use outputs from intermediate layers and stack them + out = torch.stack([output.hidden_states[k] for k in hidden_states_layers], dim=1) + out = out.to(dtype=dtype, device=device) + + batch_size, num_channels, seq_len, hidden_dim = out.shape + prompt_embeds = out.permute(0, 2, 1, 3).reshape(batch_size, seq_len, num_channels * hidden_dim) + + return prompt_embeds + +def _prepare_text_ids( + x: torch.Tensor, # (B, L, D) or (L, D) + t_coord: Optional[torch.Tensor] = None, +): + B, L, _ = x.shape + out_ids = [] + + for i in range(B): + t = torch.arange(1) if t_coord is None else t_coord[i] + h = torch.arange(1) + w = torch.arange(1) + l = torch.arange(L) + + coords = torch.cartesian_prod(t, h, w, l) + out_ids.append(coords) + + return torch.stack(out_ids) + +def encode_prompt( + prompt: Union[str, List[str]], + device: Optional[torch.device] = None, + text_encoder=None, + tokenizer=None, + num_images_per_prompt: int = 1, + prompt_embeds: Optional[torch.Tensor] = None, + max_sequence_length: int = 512, + text_encoder_out_layers: Tuple[int] = (10, 20, 30), + system_message = "You are an AI that reasons about image descriptions. You give structured responses focusing on object relationships, object attribution and actions without speculation." +): + if prompt is None: + prompt = "" + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt_embeds is None: + prompt_embeds = _get_mistral_3_small_prompt_embeds( + text_encoder=text_encoder, + tokenizer=tokenizer, + prompt=prompt, + device=device, + max_sequence_length=max_sequence_length, + system_message=system_message, + hidden_states_layers=text_encoder_out_layers, + ) + + batch_size, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + text_ids = _prepare_text_ids(prompt_embeds) + text_ids = text_ids.to(device) + return prompt_embeds, text_ids + +# Will error if the minimal version of diffusers is not installed. Remove at your own risks. +check_min_version("0.18.0.dev0") + +logger = get_logger(__name__, log_level="INFO") + +def log_validation(vae, text_encoder, tokenizer, transformer3d, network, args, accelerator, weight_dtype, global_step): + try: + logger.info("Running validation... ") + + transformer3d_val = Flux2Transformer2DModel.from_pretrained( + args.pretrained_model_name_or_path, subfolder="transformer", torch_dtype=weight_dtype, + low_cpu_mem_usage=True, + ).to(weight_dtype) + transformer3d_val.load_state_dict(accelerator.unwrap_model(transformer3d).state_dict()) + scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained( + args.pretrained_model_name_or_path, + subfolder="scheduler" + ) + transformer3d = transformer3d.to("cpu") + pipeline = Flux2Pipeline( + vae=accelerator.unwrap_model(vae).to(weight_dtype), + text_encoder=accelerator.unwrap_model(text_encoder), + tokenizer=tokenizer, + transformer=transformer3d_val, + scheduler=scheduler, + ) + pipeline = pipeline.to(accelerator.device) + + if args.seed is None: + generator = None + else: + generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) + + for i in range(len(args.validation_prompts)): + with torch.no_grad(): + sample = pipeline( + args.validation_prompts[i], + negative_prompt = "bad detailed", + height = args.image_sample_size, + width = args.image_sample_size, + generator = generator + ).images + os.makedirs(os.path.join(args.output_dir, "sample"), exist_ok=True) + image = sample[0].save(os.path.join(args.output_dir, f"sample/sample-{global_step}-image-{i}.gif")) + + del pipeline + del transformer3d_val + gc.collect() + torch.cuda.empty_cache() + torch.cuda.ipc_collect() + transformer3d = transformer3d.to(accelerator.device) + except Exception as e: + gc.collect() + torch.cuda.empty_cache() + torch.cuda.ipc_collect() + print(f"Eval error with info {e}") + transformer3d = transformer3d.to(accelerator.device) + +def parse_args(): + parser = argparse.ArgumentParser(description="Simple example of a training script.") + parser.add_argument( + "--input_perturbation", type=float, default=0, help="The scale of input perturbation. Recommended 0.1." + ) + parser.add_argument( + "--pretrained_model_name_or_path", + type=str, + default=None, + required=True, + help="Path to pretrained model or model identifier from huggingface.co/models.", + ) + parser.add_argument( + "--revision", + type=str, + default=None, + required=False, + help="Revision of pretrained model identifier from huggingface.co/models.", + ) + parser.add_argument( + "--variant", + type=str, + default=None, + help="Variant of the model files of the pretrained model identifier from huggingface.co/models, 'e.g.' fp16", + ) + parser.add_argument( + "--train_data_dir", + type=str, + default=None, + help=( + "A folder containing the training data. " + ), + ) + parser.add_argument( + "--train_data_meta", + type=str, + default=None, + help=( + "A csv containing the training data. " + ), + ) + parser.add_argument( + "--max_train_samples", + type=int, + default=None, + help=( + "For debugging purposes or quicker training, truncate the number of training examples to this " + "value if set." + ), + ) + parser.add_argument( + "--validation_prompts", + type=str, + default=None, + nargs="+", + help=("A set of prompts evaluated every `--validation_epochs` and logged to `--report_to`."), + ) + parser.add_argument( + "--output_dir", + type=str, + default="sd-model-finetuned", + help="The output directory where the model predictions and checkpoints will be written.", + ) + parser.add_argument( + "--cache_dir", + type=str, + default=None, + help="The directory where the downloaded models and datasets will be stored.", + ) + parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.") + parser.add_argument( + "--random_flip", + action="store_true", + help="whether to randomly flip images horizontally", + ) + parser.add_argument( + "--use_came", + action="store_true", + help="whether to use came", + ) + parser.add_argument( + "--multi_stream", + action="store_true", + help="whether to use cuda multi-stream", + ) + parser.add_argument( + "--train_batch_size", type=int, default=16, help="Batch size (per device) for the training dataloader." + ) + parser.add_argument( + "--vae_mini_batch", type=int, default=32, help="mini batch size for vae." + ) + parser.add_argument("--num_train_epochs", type=int, default=100) + parser.add_argument( + "--max_train_steps", + type=int, + default=None, + help="Total number of training steps to perform. If provided, overrides num_train_epochs.", + ) + parser.add_argument( + "--gradient_accumulation_steps", + type=int, + default=1, + help="Number of updates steps to accumulate before performing a backward/update pass.", + ) + parser.add_argument( + "--gradient_checkpointing", + action="store_true", + help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.", + ) + parser.add_argument( + "--learning_rate", + type=float, + default=1e-4, + help="Initial learning rate (after the potential warmup period) to use.", + ) + parser.add_argument( + "--scale_lr", + action="store_true", + default=False, + help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.", + ) + parser.add_argument( + "--lr_scheduler", + type=str, + default="constant", + help=( + 'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",' + ' "constant", "constant_with_warmup"]' + ), + ) + parser.add_argument( + "--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler." + ) + parser.add_argument( + "--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes." + ) + parser.add_argument( + "--allow_tf32", + action="store_true", + help=( + "Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see" + " https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices" + ), + ) + parser.add_argument("--use_ema", action="store_true", help="Whether to use EMA model.") + parser.add_argument( + "--non_ema_revision", + type=str, + default=None, + required=False, + help=( + "Revision of pretrained non-ema model identifier. Must be a branch, tag or git identifier of the local or" + " remote repository specified with --pretrained_model_name_or_path." + ), + ) + parser.add_argument( + "--dataloader_num_workers", + type=int, + default=0, + help=( + "Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process." + ), + ) + parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.") + parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.") + parser.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.") + parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer") + parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.") + parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.") + parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.") + parser.add_argument( + "--prediction_type", + type=str, + default=None, + help="The prediction_type that shall be used for training. Choose between 'epsilon' or 'v_prediction' or leave `None`. If left to `None` the default prediction type of the scheduler: `noise_scheduler.config.prediciton_type` is chosen.", + ) + parser.add_argument( + "--hub_model_id", + type=str, + default=None, + help="The name of the repository to keep in sync with the local `output_dir`.", + ) + parser.add_argument( + "--logging_dir", + type=str, + default="logs", + help=( + "[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to" + " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***." + ), + ) + parser.add_argument( + "--report_model_info", action="store_true", help="Whether or not to report more info about model (such as norm, grad)." + ) + parser.add_argument( + "--mixed_precision", + type=str, + default=None, + choices=["no", "fp16", "bf16"], + help=( + "Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >=" + " 1.10.and an Nvidia Ampere GPU. Default to the value of accelerate config of the current system or the" + " flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config." + ), + ) + parser.add_argument( + "--report_to", + type=str, + default="tensorboard", + help=( + 'The integration to report the results and logs to. Supported platforms are `"tensorboard"`' + ' (default), `"wandb"` and `"comet_ml"`. Use `"all"` to report to all integrations.' + ), + ) + parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank") + parser.add_argument( + "--checkpointing_steps", + type=int, + default=500, + help=( + "Save a checkpoint of the training state every X updates. These checkpoints are only suitable for resuming" + " training using `--resume_from_checkpoint`." + ), + ) + parser.add_argument( + "--checkpoints_total_limit", + type=int, + default=None, + help=("Max number of checkpoints to store."), + ) + parser.add_argument( + "--resume_from_checkpoint", + type=str, + default=None, + help=( + "Whether training should be resumed from a previous checkpoint. Use a path saved by" + ' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.' + ), + ) + parser.add_argument("--noise_offset", type=float, default=0, help="The scale of noise offset.") + parser.add_argument( + "--validation_epochs", + type=int, + default=5, + help="Run validation every X epochs.", + ) + parser.add_argument( + "--validation_steps", + type=int, + default=2000, + help="Run validation every X steps.", + ) + parser.add_argument( + "--tracker_project_name", + type=str, + default="text2image-fine-tune", + help=( + "The `project_name` argument passed to Accelerator.init_trackers for" + " more information see https://huggingface.co/docs/accelerate/v0.17.0/en/package_reference/accelerator#accelerate.Accelerator" + ), + ) + + parser.add_argument( + "--snr_loss", action="store_true", help="Whether or not to use snr_loss." + ) + parser.add_argument( + "--uniform_sampling", action="store_true", help="Whether or not to use uniform_sampling." + ) + parser.add_argument( + "--enable_text_encoder_in_dataloader", action="store_true", help="Whether or not to use text encoder in dataloader." + ) + parser.add_argument( + "--enable_bucket", action="store_true", help="Whether enable bucket sample in datasets." + ) + parser.add_argument( + "--random_ratio_crop", action="store_true", help="Whether enable random ratio crop sample in datasets." + ) + parser.add_argument( + "--random_hw_adapt", action="store_true", help="Whether enable random adapt height and width in datasets." + ) + parser.add_argument( + "--train_sampling_steps", + type=int, + default=1000, + help="Run train_sampling_steps.", + ) + parser.add_argument( + "--image_sample_size", + type=int, + default=512, + help="Sample size of the image.", + ) + parser.add_argument( + "--fix_sample_size", + nargs=2, type=int, default=None, + help="Fix Sample size [height, width] when using bucket and collate_fn." + ) + parser.add_argument( + "--config_path", + type=str, + default=None, + help=( + "The config of the model in training." + ), + ) + parser.add_argument( + "--transformer_path", + type=str, + default=None, + help=("If you want to load the weight from other transformers, input its path."), + ) + parser.add_argument( + "--vae_path", + type=str, + default=None, + help=("If you want to load the weight from other vaes, input its path."), + ) + + parser.add_argument( + '--trainable_modules', + nargs='+', + help='Enter a list of trainable modules' + ) + parser.add_argument( + '--trainable_modules_low_learning_rate', + nargs='+', + default=[], + help='Enter a list of trainable modules with lower learning rate' + ) + parser.add_argument( + '--tokenizer_max_length', + type=int, + default=512, + help='Max length of tokenizer' + ) + parser.add_argument( + "--use_deepspeed", action="store_true", help="Whether or not to use deepspeed." + ) + parser.add_argument( + "--use_fsdp", action="store_true", help="Whether or not to use fsdp." + ) + parser.add_argument( + "--low_vram", action="store_true", help="Whether enable low_vram mode." + ) + parser.add_argument( + "--prompt_template_encode", + type=str, + default="<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n", + help=( + 'The prompt template for text encoder.' + ), + ) + parser.add_argument( + "--prompt_template_encode_start_idx", + type=int, + default=34, + help=( + 'The start idx for prompt template.' + ), + ) + parser.add_argument( + "--train_mode", + type=str, + default="normal", + help=( + 'The format of training data. Support `"normal"`' + ' (default), `"i2v"`.' + ), + ) + parser.add_argument( + "--abnormal_norm_clip_start", + type=int, + default=1000, + help=( + 'When do we start doing additional processing on abnormal gradients. ' + ), + ) + parser.add_argument( + "--initial_grad_norm_ratio", + type=int, + default=5, + help=( + 'The initial gradient is relative to the multiple of the max_grad_norm. ' + ), + ) + parser.add_argument( + "--weighting_scheme", + type=str, + default="none", + choices=["sigma_sqrt", "logit_normal", "mode", "cosmap", "none"], + help=('We default to the "none" weighting scheme for uniform sampling and uniform loss'), + ) + parser.add_argument( + "--logit_mean", type=float, default=0.0, help="mean to use when using the `'logit_normal'` weighting scheme." + ) + parser.add_argument( + "--logit_std", type=float, default=1.0, help="std to use when using the `'logit_normal'` weighting scheme." + ) + parser.add_argument( + "--mode_scale", + type=float, + default=1.29, + help="Scale of mode weighting scheme. Only effective when using the `'mode'` as the `weighting_scheme`.", + ) + parser.add_argument( + "--guidance_scale", + type=float, + default=3.5, + help="the FLUX.1 dev variant is a guidance distilled model", + ) + + args = parser.parse_args() + env_local_rank = int(os.environ.get("LOCAL_RANK", -1)) + if env_local_rank != -1 and env_local_rank != args.local_rank: + args.local_rank = env_local_rank + + # default to using the same revision for the non-ema model if not specified + if args.non_ema_revision is None: + args.non_ema_revision = args.revision + + return args + + +def main(): + args = parse_args() + + if args.report_to == "wandb" and args.hub_token is not None: + raise ValueError( + "You cannot use both --report_to=wandb and --hub_token due to a security risk of exposing your token." + " Please use `huggingface-cli login` to authenticate with the Hub." + ) + + if args.non_ema_revision is not None: + deprecate( + "non_ema_revision!=None", + "0.15.0", + message=( + "Downloading 'non_ema' weights from revision branches of the Hub is deprecated. Please make sure to" + " use `--variant=non_ema` instead." + ), + ) + logging_dir = os.path.join(args.output_dir, args.logging_dir) + + accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir) + + accelerator = Accelerator( + gradient_accumulation_steps=args.gradient_accumulation_steps, + mixed_precision=args.mixed_precision, + log_with=args.report_to, + project_config=accelerator_project_config, + ) + + deepspeed_plugin = accelerator.state.deepspeed_plugin if hasattr(accelerator.state, "deepspeed_plugin") else None + fsdp_plugin = accelerator.state.fsdp_plugin if hasattr(accelerator.state, "fsdp_plugin") else None + if deepspeed_plugin is not None: + zero_stage = int(deepspeed_plugin.zero_stage) + fsdp_stage = 0 + print(f"Using DeepSpeed Zero stage: {zero_stage}") + + args.use_deepspeed = True + if zero_stage == 3: + print(f"Auto set save_state to True because zero_stage == 3") + args.save_state = True + elif fsdp_plugin is not None: + from torch.distributed.fsdp import ShardingStrategy + zero_stage = 0 + if fsdp_plugin.sharding_strategy is ShardingStrategy.FULL_SHARD: + fsdp_stage = 3 + elif fsdp_plugin.sharding_strategy is None: # The fsdp_plugin.sharding_strategy is None in FSDP 2. + fsdp_stage = 3 + elif fsdp_plugin.sharding_strategy is ShardingStrategy.SHARD_GRAD_OP: + fsdp_stage = 2 + else: + fsdp_stage = 0 + print(f"Using FSDP stage: {fsdp_stage}") + + args.use_fsdp = True + if fsdp_stage == 3: + print(f"Auto set save_state to True because fsdp_stage == 3") + args.save_state = True + else: + zero_stage = 0 + fsdp_stage = 0 + print("DeepSpeed is not enabled.") + + if accelerator.is_main_process: + writer = SummaryWriter(log_dir=logging_dir) + + # Make one log on every process with the configuration for debugging. + logging.basicConfig( + format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", + datefmt="%m/%d/%Y %H:%M:%S", + level=logging.INFO, + ) + logger.info(accelerator.state, main_process_only=False) + if accelerator.is_local_main_process: + datasets.utils.logging.set_verbosity_warning() + transformers.utils.logging.set_verbosity_warning() + diffusers.utils.logging.set_verbosity_info() + else: + datasets.utils.logging.set_verbosity_error() + transformers.utils.logging.set_verbosity_error() + diffusers.utils.logging.set_verbosity_error() + + # If passed along, set the training seed now. + if args.seed is not None: + set_seed(args.seed) + rng = np.random.default_rng(np.random.PCG64(args.seed + accelerator.process_index)) + torch_rng = torch.Generator(accelerator.device).manual_seed(args.seed + accelerator.process_index) + else: + rng = None + torch_rng = None + index_rng = np.random.default_rng(np.random.PCG64(43)) + print(f"Init rng with seed {args.seed + accelerator.process_index}. Process_index is {accelerator.process_index}") + + # Handle the repository creation + if accelerator.is_main_process: + if args.output_dir is not None: + os.makedirs(args.output_dir, exist_ok=True) + + # For mixed precision training we cast all non-trainable weigths (vae, non-lora text_encoder and non-lora transformer3d) to half-precision + # as these weights are only used for inference, keeping weights in full precision is not required. + weight_dtype = torch.float32 + if accelerator.mixed_precision == "fp16": + weight_dtype = torch.float16 + args.mixed_precision = accelerator.mixed_precision + elif accelerator.mixed_precision == "bf16": + weight_dtype = torch.bfloat16 + args.mixed_precision = accelerator.mixed_precision + + # Load scheduler, tokenizer and models. + noise_scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained( + args.pretrained_model_name_or_path, + subfolder="scheduler" + ) + + # Get Tokenizer + tokenizer = PixtralProcessor.from_pretrained( + args.pretrained_model_name_or_path, subfolder="tokenizer" + ) + + def deepspeed_zero_init_disabled_context_manager(): + """ + returns either a context list that includes one that will disable zero.Init or an empty context list + """ + deepspeed_plugin = AcceleratorState().deepspeed_plugin if accelerate.state.is_initialized() else None + if deepspeed_plugin is None: + return [] + + return [deepspeed_plugin.zero3_init_context_manager(enable=False)] + + # Currently Accelerate doesn't know how to handle multiple models under Deepspeed ZeRO stage 3. + # For this to work properly all models must be run through `accelerate.prepare`. But accelerate + # will try to assign the same optimizer with the same weights to all models during + # `deepspeed.initialize`, which of course doesn't work. + # + # For now the following workaround will partially support Deepspeed ZeRO-3, by excluding the 2 + # frozen models from being partitioned during `zero.Init` which gets called during + # `from_pretrained` So Mistral3ForConditionalGeneration and AutoencoderKLFlux2 will not enjoy the parameter sharding + # across multiple gpus and only UNet2DConditionModel will get ZeRO sharded. + with ContextManagers(deepspeed_zero_init_disabled_context_manager()): + # Get Text encoder + text_encoder = Mistral3ForConditionalGeneration.from_pretrained( + args.pretrained_model_name_or_path, subfolder="text_encoder", torch_dtype=weight_dtype + ) + text_encoder = text_encoder.eval() + # Get Vae + vae = AutoencoderKLFlux2.from_pretrained( + args.pretrained_model_name_or_path, + subfolder="vae" + ).to(weight_dtype) + vae.eval() + latents_bn_mean = vae.bn.running_mean.view(1, -1, 1, 1).to(accelerator.device, weight_dtype) + latents_bn_std = torch.sqrt(vae.bn.running_var.view(1, -1, 1, 1) + vae.config.batch_norm_eps).to(accelerator.device, weight_dtype) + + # Get Transformer + transformer3d = Flux2Transformer2DModel.from_pretrained( + args.pretrained_model_name_or_path, + subfolder="transformer", + torch_dtype=weight_dtype, + ).to(weight_dtype) + + # Freeze vae and text_encoder and set transformer3d to trainable + vae.requires_grad_(False) + text_encoder.requires_grad_(False) + transformer3d.requires_grad_(False) + + if args.transformer_path is not None: + print(f"From checkpoint: {args.transformer_path}") + if args.transformer_path.endswith("safetensors"): + from safetensors.torch import load_file, safe_open + state_dict = load_file(args.transformer_path) + else: + state_dict = torch.load(args.transformer_path, map_location="cpu") + state_dict = state_dict["state_dict"] if "state_dict" in state_dict else state_dict + + m, u = transformer3d.load_state_dict(state_dict, strict=False) + print(f"missing keys: {len(m)}, unexpected keys: {len(u)}") + assert len(u) == 0 + + if args.vae_path is not None: + print(f"From checkpoint: {args.vae_path}") + if args.vae_path.endswith("safetensors"): + from safetensors.torch import load_file, safe_open + state_dict = load_file(args.vae_path) + else: + state_dict = torch.load(args.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)}") + assert len(u) == 0 + + # A good trainable modules is showed below now. + # For 3D Patch: trainable_modules = ['ff.net', 'pos_embed', 'attn2', 'proj_out', 'timepositionalencoding', 'h_position', 'w_position'] + # For 2D Patch: trainable_modules = ['ff.net', 'attn2', 'timepositionalencoding', 'h_position', 'w_position'] + transformer3d.train() + if accelerator.is_main_process: + accelerator.print( + f"Trainable modules '{args.trainable_modules}'." + ) + for name, param in transformer3d.named_parameters(): + for trainable_module_name in args.trainable_modules + args.trainable_modules_low_learning_rate: + if trainable_module_name in name: + param.requires_grad = True + break + + # Create EMA for the transformer3d. + if args.use_ema: + if zero_stage == 3: + raise NotImplementedError("FSDP does not support EMA.") + + ema_transformer3d = Flux2Transformer2DModel.from_pretrained( + args.pretrained_model_name_or_path, + subfolder="transformer", + torch_dtype=weight_dtype, + ).to(weight_dtype) + + ema_transformer3d = EMAModel(ema_transformer3d.parameters(), model_cls=Flux2Transformer2DModel, model_config=ema_transformer3d.config) + + # `accelerate` 0.16.0 will have better support for customized saving + if version.parse(accelerate.__version__) >= version.parse("0.16.0"): + # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format + if fsdp_stage != 0: + def save_model_hook(models, weights, output_dir): + accelerate_state_dict = accelerator.get_state_dict(models[-1], unwrap=True) + if accelerator.is_main_process: + from safetensors.torch import save_file + + safetensor_save_path = os.path.join(output_dir, f"diffusion_pytorch_model.safetensors") + accelerate_state_dict = {k: v.to(dtype=weight_dtype) for k, v in accelerate_state_dict.items()} + save_file(accelerate_state_dict, safetensor_save_path, metadata={"format": "pt"}) + + with open(os.path.join(output_dir, "sampler_pos_start.pkl"), 'wb') as file: + pickle.dump([batch_sampler.sampler._pos_start, first_epoch], file) + + def load_model_hook(models, input_dir): + pkl_path = os.path.join(input_dir, "sampler_pos_start.pkl") + if os.path.exists(pkl_path): + with open(pkl_path, 'rb') as file: + loaded_number, _ = pickle.load(file) + batch_sampler.sampler._pos_start = max(loaded_number - args.dataloader_num_workers * accelerator.num_processes * 2, 0) + print(f"Load pkl from {pkl_path}. Get loaded_number = {loaded_number}.") + + elif zero_stage == 3: + # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format + def save_model_hook(models, weights, output_dir): + accelerate_state_dict = accelerator.get_state_dict(models[-1], unwrap=True) + if accelerator.is_main_process: + from safetensors.torch import save_file + safetensor_save_path = os.path.join(output_dir, f"diffusion_pytorch_model.safetensors") + save_file(accelerate_state_dict, safetensor_save_path, metadata={"format": "pt"}) + + with open(os.path.join(output_dir, "sampler_pos_start.pkl"), 'wb') as file: + pickle.dump([batch_sampler.sampler._pos_start, first_epoch], file) + + def load_model_hook(models, input_dir): + pkl_path = os.path.join(input_dir, "sampler_pos_start.pkl") + if os.path.exists(pkl_path): + with open(pkl_path, 'rb') as file: + loaded_number, _ = pickle.load(file) + batch_sampler.sampler._pos_start = max(loaded_number - args.dataloader_num_workers * accelerator.num_processes * 2, 0) + print(f"Load pkl from {pkl_path}. Get loaded_number = {loaded_number}.") + else: + # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format + def save_model_hook(models, weights, output_dir): + if accelerator.is_main_process: + if args.use_ema: + ema_transformer3d.save_pretrained(os.path.join(output_dir, "transformer_ema")) + + models[0].save_pretrained(os.path.join(output_dir, "transformer")) + if not args.use_deepspeed: + weights.pop() + + with open(os.path.join(output_dir, "sampler_pos_start.pkl"), 'wb') as file: + pickle.dump([batch_sampler.sampler._pos_start, first_epoch], file) + + def load_model_hook(models, input_dir): + if args.use_ema: + ema_path = os.path.join(input_dir, "transformer_ema") + _, ema_kwargs = Flux2Transformer2DModel.load_config(ema_path, return_unused_kwargs=True) + load_model = Flux2Transformer2DModel.from_pretrained( + input_dir, subfolder="transformer_ema", + ) + load_model = EMAModel(load_model.parameters(), model_cls=Flux2Transformer2DModel, model_config=load_model.config) + load_model.load_state_dict(ema_kwargs) + + ema_transformer3d.load_state_dict(load_model.state_dict()) + ema_transformer3d.to(accelerator.device) + del load_model + + for i in range(len(models)): + # pop models so that they are not loaded again + model = models.pop() + + # load diffusers style into model + load_model = Flux2Transformer2DModel.from_pretrained( + input_dir, subfolder="transformer" + ) + model.register_to_config(**load_model.config) + + model.load_state_dict(load_model.state_dict()) + del load_model + + pkl_path = os.path.join(input_dir, "sampler_pos_start.pkl") + if os.path.exists(pkl_path): + with open(pkl_path, 'rb') as file: + loaded_number, _ = pickle.load(file) + batch_sampler.sampler._pos_start = max(loaded_number - args.dataloader_num_workers * accelerator.num_processes * 2, 0) + print(f"Load pkl from {pkl_path}. Get loaded_number = {loaded_number}.") + + accelerator.register_save_state_pre_hook(save_model_hook) + accelerator.register_load_state_pre_hook(load_model_hook) + + if args.gradient_checkpointing: + transformer3d.enable_gradient_checkpointing() + + # Enable TF32 for faster training on Ampere GPUs, + # cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices + if args.allow_tf32: + torch.backends.cuda.matmul.allow_tf32 = True + + if args.scale_lr: + args.learning_rate = ( + args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes + ) + + # Initialize the optimizer + if args.use_8bit_adam: + try: + import bitsandbytes as bnb + except ImportError: + raise ImportError( + "Please install bitsandbytes to use 8-bit Adam. You can do so by running `pip install bitsandbytes`" + ) + + optimizer_cls = bnb.optim.AdamW8bit + elif args.use_came: + try: + from came_pytorch import CAME + except: + raise ImportError( + "Please install came_pytorch to use CAME. You can do so by running `pip install came_pytorch`" + ) + + optimizer_cls = CAME + else: + optimizer_cls = torch.optim.AdamW + + trainable_params = list(filter(lambda p: p.requires_grad, transformer3d.parameters())) + trainable_params_optim = [ + {'params': [], 'lr': args.learning_rate}, + {'params': [], 'lr': args.learning_rate / 2}, + ] + in_already = [] + for name, param in transformer3d.named_parameters(): + high_lr_flag = False + if name in in_already: + continue + for trainable_module_name in args.trainable_modules: + if trainable_module_name in name: + in_already.append(name) + high_lr_flag = True + trainable_params_optim[0]['params'].append(param) + if accelerator.is_main_process: + print(f"Set {name} to lr : {args.learning_rate}") + break + if high_lr_flag: + continue + for trainable_module_name in args.trainable_modules_low_learning_rate: + if trainable_module_name in name: + in_already.append(name) + trainable_params_optim[1]['params'].append(param) + if accelerator.is_main_process: + print(f"Set {name} to lr : {args.learning_rate / 2}") + break + + if args.use_came: + optimizer = optimizer_cls( + trainable_params_optim, + lr=args.learning_rate, + # weight_decay=args.adam_weight_decay, + betas=(0.9, 0.999, 0.9999), + eps=(1e-30, 1e-16) + ) + else: + optimizer = optimizer_cls( + trainable_params_optim, + lr=args.learning_rate, + betas=(args.adam_beta1, args.adam_beta2), + weight_decay=args.adam_weight_decay, + eps=args.adam_epsilon, + ) + + # Get the training dataset + if args.fix_sample_size is not None and args.enable_bucket: + args.image_sample_size = max(max(args.fix_sample_size), args.image_sample_size) + args.random_hw_adapt = False + + # Get the dataset + train_dataset = ImageVideoDataset( + args.train_data_meta, args.train_data_dir, + image_sample_size=args.image_sample_size, + enable_bucket=args.enable_bucket, + ) + + def worker_init_fn(_seed): + _seed = _seed * 256 + def _worker_init_fn(worker_id): + print(f"worker_init_fn with {_seed + worker_id}") + np.random.seed(_seed + worker_id) + random.seed(_seed + worker_id) + return _worker_init_fn + + if args.enable_bucket: + aspect_ratio_sample_size = {key : [x / 512 * args.image_sample_size for x in ASPECT_RATIO_512[key]] for key in ASPECT_RATIO_512.keys()} + batch_sampler_generator = torch.Generator().manual_seed(args.seed) + batch_sampler = AspectRatioBatchImageVideoSampler( + sampler=RandomSampler(train_dataset, generator=batch_sampler_generator), dataset=train_dataset.dataset, + batch_size=args.train_batch_size, train_folder = args.train_data_dir, drop_last=True, + aspect_ratios=aspect_ratio_sample_size, + ) + + def collate_fn(examples): + def get_random_downsample_ratio(sample_size, image_ratio=[], + all_choices=False, rng=None): + def _create_special_list(length): + if length == 1: + return [1.0] + if length >= 2: + first_element = 0.90 + remaining_sum = 1.0 - first_element + other_elements_value = remaining_sum / (length - 1) + special_list = [first_element] + [other_elements_value] * (length - 1) + return special_list + + if sample_size >= 1536: + number_list = [1, 1.25, 1.5, 2, 2.5, 3] + image_ratio + elif sample_size >= 1024: + number_list = [1, 1.25, 1.5, 2] + image_ratio + elif sample_size >= 768: + number_list = [1, 1.25, 1.5] + image_ratio + elif sample_size >= 512: + number_list = [1] + image_ratio + else: + number_list = [1] + + if all_choices: + return number_list + + number_list_prob = np.array(_create_special_list(len(number_list))) + if rng is None: + return np.random.choice(number_list, p = number_list_prob) + else: + return rng.choice(number_list, p = number_list_prob) + + # Create new output + new_examples = {} + new_examples["pixel_values"] = [] + new_examples["text"] = [] + + # Get downsample ratio in image + pixel_value = examples[0]["pixel_values"] + data_type = examples[0]["data_type"] + f, h, w, c = np.shape(pixel_value) + + random_downsample_ratio = 1 if not args.random_hw_adapt else get_random_downsample_ratio(args.image_sample_size) + + aspect_ratio_sample_size = {key : [x / 512 * args.image_sample_size / random_downsample_ratio for x in ASPECT_RATIO_512[key]] for key in ASPECT_RATIO_512.keys()} + aspect_ratio_random_crop_sample_size = {key : [x / 512 * args.image_sample_size / random_downsample_ratio for x in ASPECT_RATIO_RANDOM_CROP_512[key]] for key in ASPECT_RATIO_RANDOM_CROP_512.keys()} + + if args.fix_sample_size is not None: + fix_sample_size = [int(x / 16) * 16 for x in args.fix_sample_size] + elif args.random_ratio_crop: + if rng is None: + random_sample_size = aspect_ratio_random_crop_sample_size[ + np.random.choice(list(aspect_ratio_random_crop_sample_size.keys()), p = ASPECT_RATIO_RANDOM_CROP_PROB) + ] + else: + random_sample_size = aspect_ratio_random_crop_sample_size[ + rng.choice(list(aspect_ratio_random_crop_sample_size.keys()), p = ASPECT_RATIO_RANDOM_CROP_PROB) + ] + random_sample_size = [int(x / 16) * 16 for x in random_sample_size] + else: + closest_size, closest_ratio = get_closest_ratio(h, w, ratios=aspect_ratio_sample_size) + closest_size = [int(x / 16) * 16 for x in closest_size] + + for example in examples: + if args.fix_sample_size is not None: + # To 0~1 + pixel_values = torch.from_numpy(example["pixel_values"]).permute(0, 3, 1, 2).contiguous() + pixel_values = pixel_values / 255. + + # Get adapt hw for resize + fix_sample_size = list(map(lambda x: int(x), fix_sample_size)) + transform = transforms.Compose([ + transforms.Resize(fix_sample_size, interpolation=transforms.InterpolationMode.BILINEAR), # Image.BICUBIC + transforms.CenterCrop(fix_sample_size), + transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], inplace=True), + ]) + elif args.random_ratio_crop: + # To 0~1 + pixel_values = torch.from_numpy(example["pixel_values"]).permute(0, 3, 1, 2).contiguous() + pixel_values = pixel_values / 255. + + # Get adapt hw for resize + b, c, h, w = pixel_values.size() + th, tw = random_sample_size + if th / tw > h / w: + nh = int(th) + nw = int(w / h * nh) + else: + nw = int(tw) + nh = int(h / w * nw) + + transform = transforms.Compose([ + transforms.Resize([nh, nw]), + transforms.CenterCrop([int(x) for x in random_sample_size]), + transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], inplace=True), + ]) + else: + # To 0~1 + pixel_values = torch.from_numpy(example["pixel_values"]).permute(0, 3, 1, 2).contiguous() + pixel_values = pixel_values / 255. + + # Get adapt hw for resize + closest_size = list(map(lambda x: int(x), closest_size)) + if closest_size[0] / h > closest_size[1] / w: + resize_size = closest_size[0], int(w * closest_size[0] / h) + else: + resize_size = int(h * closest_size[1] / w), closest_size[1] + + transform = transforms.Compose([ + transforms.Resize(resize_size, interpolation=transforms.InterpolationMode.BILINEAR), # Image.BICUBIC + transforms.CenterCrop(closest_size), + transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], inplace=True), + ]) + new_examples["pixel_values"].append(transform(pixel_values)) + new_examples["text"].append(example["text"]) + + # Limit the number of frames to the same + new_examples["pixel_values"] = torch.stack([example for example in new_examples["pixel_values"]]) + + # Encode prompts when enable_text_encoder_in_dataloader=True + if args.enable_text_encoder_in_dataloader: + prompt_embeds, text_ids = encode_prompt( + batch['text'], device="cpu", + text_encoder=text_encoder, + tokenizer=tokenizer, + ) + + new_examples['prompt_embeds'] = prompt_embeds + new_examples['text_ids'] = text_ids + + return new_examples + + # DataLoaders creation: + train_dataloader = torch.utils.data.DataLoader( + train_dataset, + batch_sampler=batch_sampler, + collate_fn=collate_fn, + persistent_workers=True if args.dataloader_num_workers != 0 else False, + num_workers=args.dataloader_num_workers, + worker_init_fn=worker_init_fn(args.seed + accelerator.process_index) + ) + else: + # DataLoaders creation: + batch_sampler_generator = torch.Generator().manual_seed(args.seed) + batch_sampler = ImageVideoSampler(RandomSampler(train_dataset, generator=batch_sampler_generator), train_dataset, args.train_batch_size) + train_dataloader = torch.utils.data.DataLoader( + train_dataset, + batch_sampler=batch_sampler, + persistent_workers=True if args.dataloader_num_workers != 0 else False, + num_workers=args.dataloader_num_workers, + worker_init_fn=worker_init_fn(args.seed + accelerator.process_index) + ) + + # Scheduler and math around the number of training steps. + overrode_max_train_steps = False + num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps) + if args.max_train_steps is None: + args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch + overrode_max_train_steps = True + + lr_scheduler = get_scheduler( + args.lr_scheduler, + optimizer=optimizer, + num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes, + num_training_steps=args.max_train_steps * accelerator.num_processes, + ) + + # Prepare everything with our `accelerator`. + transformer3d, optimizer, train_dataloader, lr_scheduler = accelerator.prepare( + transformer3d, optimizer, train_dataloader, lr_scheduler + ) + + if fsdp_stage != 0 or zero_stage != 0: + from functools import partial + + from videox_fun.dist import set_multi_gpus_devices, shard_model + shard_fn = partial(shard_model, device_id=accelerator.device, param_dtype=weight_dtype, module_to_wrapper=text_encoder.language_model.layers) + text_encoder = shard_fn(text_encoder) + + if args.use_ema: + ema_transformer3d.to(accelerator.device) + + # Move text_encode and vae to gpu and cast to weight_dtype + vae.to(accelerator.device if not args.low_vram else "cpu", dtype=weight_dtype) + if not args.enable_text_encoder_in_dataloader: + text_encoder.to(accelerator.device if not args.low_vram else "cpu", dtype=weight_dtype) + + # We need to recalculate our total training steps as the size of the training dataloader may have changed. + num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps) + if overrode_max_train_steps: + args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch + # Afterwards we recalculate our number of training epochs + args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch) + + # We need to initialize the trackers we use, and also store our configuration. + # The trackers initializes automatically on the main process. + if accelerator.is_main_process: + tracker_config = dict(vars(args)) + keys_to_pop = [k for k, v in tracker_config.items() if isinstance(v, list)] + for k in keys_to_pop: + tracker_config.pop(k) + print(f"Removed tracker_config['{k}']") + accelerator.init_trackers(args.tracker_project_name, tracker_config) + + # Function for unwrapping if model was compiled with `torch.compile`. + def unwrap_model(model): + model = accelerator.unwrap_model(model) + model = model._orig_mod if is_compiled_module(model) else model + return model + + # Train! + total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps + + logger.info("***** Running training *****") + logger.info(f" Num examples = {len(train_dataset)}") + logger.info(f" Num Epochs = {args.num_train_epochs}") + logger.info(f" Instantaneous batch size per device = {args.train_batch_size}") + logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}") + logger.info(f" Gradient Accumulation steps = {args.gradient_accumulation_steps}") + logger.info(f" Total optimization steps = {args.max_train_steps}") + global_step = 0 + first_epoch = 0 + + # Potentially load in the weights and states from a previous save + if args.resume_from_checkpoint: + if args.resume_from_checkpoint != "latest": + path = os.path.basename(args.resume_from_checkpoint) + else: + # Get the most recent checkpoint + dirs = os.listdir(args.output_dir) + dirs = [d for d in dirs if d.startswith("checkpoint")] + dirs = sorted(dirs, key=lambda x: int(x.split("-")[1])) + path = dirs[-1] if len(dirs) > 0 else None + + if path is None: + accelerator.print( + f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run." + ) + args.resume_from_checkpoint = None + initial_global_step = 0 + else: + global_step = int(path.split("-")[1]) + + initial_global_step = global_step + + pkl_path = os.path.join(os.path.join(args.output_dir, path), "sampler_pos_start.pkl") + if os.path.exists(pkl_path): + with open(pkl_path, 'rb') as file: + _, first_epoch = pickle.load(file) + else: + first_epoch = global_step // num_update_steps_per_epoch + print(f"Load pkl from {pkl_path}. Get first_epoch = {first_epoch}.") + + accelerator.print(f"Resuming from checkpoint {path}") + accelerator.load_state(os.path.join(args.output_dir, path)) + else: + initial_global_step = 0 + + progress_bar = tqdm( + range(0, args.max_train_steps), + initial=initial_global_step, + desc="Steps", + # Only show the progress bar once on each machine. + disable=not accelerator.is_local_main_process, + ) + + if args.multi_stream and args.train_mode != "normal": + # create extra cuda streams to speedup inpaint vae computation + vae_stream_1 = torch.cuda.Stream() + vae_stream_2 = torch.cuda.Stream() + else: + vae_stream_1 = None + vae_stream_2 = None + + # Calculate the index we need】 + idx_sampling = DiscreteSampling(args.train_sampling_steps, uniform_sampling=args.uniform_sampling) + + for epoch in range(first_epoch, args.num_train_epochs): + train_loss = 0.0 + batch_sampler.sampler.generator = torch.Generator().manual_seed(args.seed + epoch) + for step, batch in enumerate(train_dataloader): + # Data batch sanity check + if epoch == first_epoch and step == 0: + pixel_values, texts = batch['pixel_values'].cpu(), batch['text'] + pixel_values = rearrange(pixel_values, "b f c h w -> b c f h w") + os.makedirs(os.path.join(args.output_dir, "sanity_check"), exist_ok=True) + for idx, (pixel_value, text) in enumerate(zip(pixel_values, texts)): + pixel_value = pixel_value[None, ...] + gif_name = '-'.join(text.replace('/', '').split()[:10]) if not text == '' else f'{global_step}-{idx}' + save_videos_grid(pixel_value, f"{args.output_dir}/sanity_check/{gif_name[:10]}.gif", rescale=True) + + with accelerator.accumulate(transformer3d): + # Convert images to latent space + pixel_values = batch["pixel_values"].to(weight_dtype) + + if args.low_vram: + torch.cuda.empty_cache() + vae.to(accelerator.device) + if not args.enable_text_encoder_in_dataloader: + text_encoder.to("cpu") + + with torch.no_grad(): + # This way is quicker when batch grows up + def _batch_encode_vae(pixel_values): + pixel_values = pixel_values.squeeze(1) + bs = args.vae_mini_batch + new_pixel_values = [] + for i in range(0, pixel_values.shape[0], bs): + pixel_values_bs = pixel_values[i : i + bs] + pixel_values_bs = vae.encode(pixel_values_bs)[0] + pixel_values_bs = pixel_values_bs.sample() + new_pixel_values.append(pixel_values_bs) + return torch.cat(new_pixel_values, dim = 0) + if vae_stream_1 is not None: + vae_stream_1.wait_stream(torch.cuda.current_stream()) + with torch.cuda.stream(vae_stream_1): + latents = _batch_encode_vae(pixel_values) + else: + latents = _batch_encode_vae(pixel_values) + + # wait for latents = vae.encode(pixel_values) to complete + if vae_stream_1 is not None: + torch.cuda.current_stream().wait_stream(vae_stream_1) + + if args.low_vram: + vae.to('cpu') + torch.cuda.empty_cache() + if not args.enable_text_encoder_in_dataloader: + text_encoder.to(accelerator.device) + + if args.enable_text_encoder_in_dataloader: + prompt_embeds = batch['prompt_embeds'].to(dtype=latents.dtype, device=accelerator.device) + text_ids = batch['text_ids'] + else: + with torch.no_grad(): + prompt_embeds, text_ids = encode_prompt( + batch['text'], device=accelerator.device, + text_encoder=text_encoder, + tokenizer=tokenizer, + ) + + if args.low_vram and not args.enable_text_encoder_in_dataloader: + text_encoder.to('cpu') + torch.cuda.empty_cache() + + bsz, channel, height, width = latents.size() + latents = _patchify_latents(latents) + latent_image_ids = _prepare_latent_ids(latents) + latents = ((latents - latents_bn_mean) / latents_bn_std).to(dtype=weight_dtype) + latents = _pack_latents(latents) + + noise = torch.randn(latents.size(), device=latents.device, generator=torch_rng, dtype=weight_dtype) + # handle guidance + guidance = torch.tensor([args.guidance_scale], device=accelerator.device) + guidance = guidance.expand(latents.shape[0]) + + if not args.uniform_sampling: + u = compute_density_for_timestep_sampling( + weighting_scheme=args.weighting_scheme, + batch_size=bsz, + logit_mean=args.logit_mean, + logit_std=args.logit_std, + mode_scale=args.mode_scale, + ) + indices = (u * noise_scheduler.config.num_train_timesteps).long() + else: + # Sample a random timestep for each image + # timesteps = generate_timestep_with_lognorm(0, args.train_sampling_steps, (bsz,), device=latents.device, generator=torch_rng) + # timesteps = torch.randint(0, args.train_sampling_steps, (bsz,), device=latents.device, generator=torch_rng) + indices = idx_sampling(bsz, generator=torch_rng, device=latents.device) + indices = indices.long().cpu() + + sigmas = np.linspace(1.0, 1 / args.train_sampling_steps, args.train_sampling_steps) + image_seq_len = latents.shape[1] + mu = calculate_shift( + image_seq_len, + noise_scheduler.config.get("base_image_seq_len", 256), + noise_scheduler.config.get("max_image_seq_len", 4096), + noise_scheduler.config.get("base_shift", 0.5), + noise_scheduler.config.get("max_shift", 1.15), + ) + noise_scheduler.set_timesteps(sigmas=sigmas, device=latents.device, mu=mu) + timesteps = noise_scheduler.timesteps[indices].to(device=latents.device) + + def get_sigmas(timesteps, n_dim=4, dtype=torch.float32): + sigmas = noise_scheduler.sigmas.to(device=accelerator.device, dtype=dtype) + schedule_timesteps = noise_scheduler.timesteps.to(accelerator.device) + timesteps = timesteps.to(accelerator.device) + step_indices = [(schedule_timesteps == t).nonzero().item() for t in timesteps] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < n_dim: + sigma = sigma.unsqueeze(-1) + return sigma + + # Add noise according to flow matching. + # zt = (1 - texp) * x + texp * z1 + sigmas = get_sigmas(timesteps, n_dim=latents.ndim, dtype=latents.dtype) + noisy_latents = (1.0 - sigmas) * latents + sigmas * noise + + # Add noise + target = noise - latents + + # Predict the noise residual + with torch.cuda.amp.autocast(dtype=weight_dtype), torch.cuda.device(device=accelerator.device): + noise_pred = transformer3d( + hidden_states=noisy_latents, + timestep=timesteps / 1000, + guidance=guidance, + encoder_hidden_states=prompt_embeds, + txt_ids=text_ids, + img_ids=latent_image_ids, + return_dict=False, + )[0] + + def custom_mse_loss(noise_pred, target, weighting=None, threshold=50): + noise_pred = noise_pred.float() + target = target.float() + diff = noise_pred - target + mse_loss = F.mse_loss(noise_pred, target, reduction='none') + mask = (diff.abs() <= threshold).float() + masked_loss = mse_loss * mask + if weighting is not None: + masked_loss = masked_loss * weighting + final_loss = masked_loss.mean() + return final_loss + + weighting = compute_loss_weighting_for_sd3(weighting_scheme=args.weighting_scheme, sigmas=sigmas) + loss = custom_mse_loss(noise_pred.float(), target.float(), weighting.float()) + loss = loss.mean() + + # Gather the losses across all processes for logging (if we use distributed training). + avg_loss = accelerator.gather(loss.repeat(args.train_batch_size)).mean() + train_loss += avg_loss.item() / args.gradient_accumulation_steps + + # Backpropagate + accelerator.backward(loss) + if accelerator.sync_gradients: + if not args.use_deepspeed and not args.use_fsdp: + trainable_params_grads = [p.grad for p in trainable_params if p.grad is not None] + trainable_params_total_norm = torch.norm(torch.stack([torch.norm(g.detach(), 2) for g in trainable_params_grads]), 2) + max_grad_norm = linear_decay(args.max_grad_norm * args.initial_grad_norm_ratio, args.max_grad_norm, args.abnormal_norm_clip_start, global_step) + if trainable_params_total_norm / max_grad_norm > 5 and global_step > args.abnormal_norm_clip_start: + actual_max_grad_norm = max_grad_norm / min((trainable_params_total_norm / max_grad_norm), 10) + else: + actual_max_grad_norm = max_grad_norm + else: + actual_max_grad_norm = args.max_grad_norm + + if not args.use_deepspeed and not args.use_fsdp and args.report_model_info and accelerator.is_main_process: + if trainable_params_total_norm > 1 and global_step > args.abnormal_norm_clip_start: + for name, param in transformer3d.named_parameters(): + if param.requires_grad: + writer.add_scalar(f'gradients/before_clip_norm/{name}', param.grad.norm(), global_step=global_step) + + norm_sum = accelerator.clip_grad_norm_(trainable_params, actual_max_grad_norm) + if not args.use_deepspeed and not args.use_fsdp and args.report_model_info and accelerator.is_main_process: + writer.add_scalar(f'gradients/norm_sum', norm_sum, global_step=global_step) + writer.add_scalar(f'gradients/actual_max_grad_norm', actual_max_grad_norm, global_step=global_step) + optimizer.step() + lr_scheduler.step() + optimizer.zero_grad() + + # Checks if the accelerator has performed an optimization step behind the scenes + if accelerator.sync_gradients: + + if args.use_ema: + ema_transformer3d.step(transformer3d.parameters()) + progress_bar.update(1) + global_step += 1 + accelerator.log({"train_loss": train_loss}, step=global_step) + train_loss = 0.0 + + if global_step % args.checkpointing_steps == 0: + if args.use_deepspeed or args.use_fsdp or accelerator.is_main_process: + # _before_ saving state, check if this save would set us over the `checkpoints_total_limit` + if args.checkpoints_total_limit is not None: + checkpoints = os.listdir(args.output_dir) + checkpoints = [d for d in checkpoints if d.startswith("checkpoint")] + checkpoints = sorted(checkpoints, key=lambda x: int(x.split("-")[1])) + + # before we save the new checkpoint, we need to have at _most_ `checkpoints_total_limit - 1` checkpoints + if len(checkpoints) >= args.checkpoints_total_limit: + num_to_remove = len(checkpoints) - args.checkpoints_total_limit + 1 + removing_checkpoints = checkpoints[0:num_to_remove] + + logger.info( + f"{len(checkpoints)} checkpoints already exist, removing {len(removing_checkpoints)} checkpoints" + ) + logger.info(f"removing checkpoints: {', '.join(removing_checkpoints)}") + + for removing_checkpoint in removing_checkpoints: + removing_checkpoint = os.path.join(args.output_dir, removing_checkpoint) + shutil.rmtree(removing_checkpoint) + + gc.collect() + torch.cuda.empty_cache() + torch.cuda.ipc_collect() + save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}") + accelerator.save_state(save_path) + logger.info(f"Saved state to {save_path}") + + if accelerator.is_main_process: + if args.validation_prompts is not None and global_step % args.validation_steps == 0: + if args.use_ema: + # Store the UNet parameters temporarily and load the EMA parameters to perform inference. + ema_transformer3d.store(transformer3d.parameters()) + ema_transformer3d.copy_to(transformer3d.parameters()) + log_validation( + vae, + text_encoder, + tokenizer, + transformer3d, + network, + args, + accelerator, + weight_dtype, + global_step, + ) + if args.use_ema: + # Switch back to the original transformer3d parameters. + ema_transformer3d.restore(transformer3d.parameters()) + + logs = {"step_loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]} + progress_bar.set_postfix(**logs) + + if global_step >= args.max_train_steps: + break + + if accelerator.is_main_process: + if args.validation_prompts is not None and epoch % args.validation_epochs == 0: + if args.use_ema: + # Store the UNet parameters temporarily and load the EMA parameters to perform inference. + ema_transformer3d.store(transformer3d.parameters()) + ema_transformer3d.copy_to(transformer3d.parameters()) + log_validation( + vae, + text_encoder, + tokenizer, + tokenizer_2, + transformer3d, + network, + args, + accelerator, + weight_dtype, + global_step, + ) + if args.use_ema: + # Switch back to the original transformer3d parameters. + ema_transformer3d.restore(transformer3d.parameters()) + + # Create the pipeline using the trained modules and save it. + accelerator.wait_for_everyone() + if args.use_deepspeed or args.use_fsdp or accelerator.is_main_process: + gc.collect() + torch.cuda.empty_cache() + torch.cuda.ipc_collect() + save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}") + accelerator.save_state(save_path) + logger.info(f"Saved state to {save_path}") + + accelerator.end_training() + + +if __name__ == "__main__": + main() diff --git a/scripts/flux2/train.sh b/scripts/flux2/train.sh new file mode 100644 index 0000000..84271f6 --- /dev/null +++ b/scripts/flux2/train.sh @@ -0,0 +1,32 @@ +export MODEL_NAME="models/Diffusion_Transformer/FLUX.2-dev" +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/flux2/train.py \ + --pretrained_model_name_or_path=$MODEL_NAME \ + --train_data_dir=$DATASET_NAME \ + --train_data_meta=$DATASET_META_NAME \ + --train_batch_size=1 \ + --image_sample_size=1328 \ + --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" \ + --gradient_checkpointing \ + --mixed_precision="bf16" \ + --adam_weight_decay=3e-2 \ + --adam_epsilon=1e-10 \ + --vae_mini_batch=1 \ + --max_grad_norm=0.05 \ + --enable_bucket \ + --uniform_sampling \ + --trainable_modules "." \ No newline at end of file diff --git a/scripts/flux2/train_lora.py b/scripts/flux2/train_lora.py new file mode 100644 index 0000000..9727ab1 --- /dev/null +++ b/scripts/flux2/train_lora.py @@ -0,0 +1,1722 @@ +"""Modified from https://github.com/huggingface/diffusers/blob/main/examples/text_to_image/train_text_to_image.py +""" +#!/usr/bin/env python +# coding=utf-8 +# Copyright 2024 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and + +import argparse +import gc +import logging +import math +import os +import pickle +import random +import shutil +import sys +from typing import List, NamedTuple, Optional, Union +from typing import Any, Callable, Dict, List, Optional, Tuple, Union + +import accelerate +import diffusers +import numpy as np +import torch +import torch.nn.functional as F +import torch.utils.checkpoint +import torchvision.transforms.functional as TF +import transformers +from accelerate import Accelerator +from accelerate.logging import get_logger +from accelerate.state import AcceleratorState +from accelerate.utils import ProjectConfiguration, set_seed +from diffusers import DDIMScheduler, FlowMatchEulerDiscreteScheduler +from diffusers.optimization import get_scheduler +from diffusers.training_utils import (EMAModel, + compute_density_for_timestep_sampling, + compute_loss_weighting_for_sd3) +from diffusers.utils import check_min_version, deprecate, is_wandb_available +from diffusers.utils.torch_utils import is_compiled_module +from einops import rearrange +from omegaconf import OmegaConf +from packaging import version +from PIL import Image +from torch.utils.data import RandomSampler +from torch.utils.tensorboard import SummaryWriter +from torchvision import transforms +from tqdm.auto import tqdm +from transformers import AutoTokenizer +from transformers.utils import ContextManagers + +import datasets + +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.data.bucket_sampler import (ASPECT_RATIO_512, + ASPECT_RATIO_RANDOM_CROP_512, + ASPECT_RATIO_RANDOM_CROP_PROB, + AspectRatioBatchImageVideoSampler, + RandomSampler, get_closest_ratio) +from videox_fun.data.dataset_image_video import (ImageVideoDataset, + ImageVideoSampler, + get_random_mask) +from videox_fun.dist import set_multi_gpus_devices, shard_model +from videox_fun.models import (AutoencoderKLFlux2, + CLIPImageProcessor, Mistral3ForConditionalGeneration, + PixtralProcessor, CLIPVisionModelWithProjection, + Flux2Transformer2DModel, + Mistral3ForConditionalGeneration, + Qwen2_5_VLForConditionalGeneration, + Qwen2Tokenizer, QwenImageTransformer2DModel) +from videox_fun.pipeline import Flux2Pipeline +from videox_fun.utils.discrete_sampler import DiscreteSampling +from videox_fun.utils.utils import get_image_to_video_latent, save_videos_grid + +if is_wandb_available(): + import wandb + + +def filter_kwargs(cls, kwargs): + import inspect + sig = inspect.signature(cls.__init__) + valid_params = set(sig.parameters.keys()) - {'self', 'cls'} + filtered_kwargs = {k: v for k, v in kwargs.items() if k in valid_params} + return filtered_kwargs + +def linear_decay(initial_value, final_value, total_steps, current_step): + if current_step >= total_steps: + return final_value + current_step = max(0, current_step) + step_size = (final_value - initial_value) / total_steps + current_value = initial_value + step_size * current_step + return current_value + +def generate_timestep_with_lognorm(low, high, shape, device="cpu", generator=None): + u = torch.normal(mean=0.0, std=1.0, size=shape, device=device, generator=generator) + t = 1 / (1 + torch.exp(-u)) * (high - low) + low + return torch.clip(t.to(torch.int32), low, high - 1) + +def compute_empirical_mu(image_seq_len: int, num_steps: int) -> float: + a1, b1 = 8.73809524e-05, 1.89833333 + a2, b2 = 0.00016927, 0.45666666 + + if image_seq_len > 4300: + mu = a2 * image_seq_len + b2 + return float(mu) + + m_200 = a2 * image_seq_len + b2 + m_10 = a1 * image_seq_len + b1 + + a = (m_200 - m_10) / 190.0 + b = m_200 - 200.0 * a + mu = a * num_steps + b + + return float(mu) + +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + +def _prepare_latent_ids( + latents: torch.Tensor, # (B, C, H, W) +): + r""" + Generates 4D position coordinates (T, H, W, L) for latent tensors. + + Args: + latents (torch.Tensor): + Latent tensor of shape (B, C, H, W) + + Returns: + torch.Tensor: + Position IDs tensor of shape (B, H*W, 4) All batches share the same coordinate structure: T=0, + H=[0..H-1], W=[0..W-1], L=0 + """ + + batch_size, _, height, width = latents.shape + + t = torch.arange(1) # [0] - time dimension + h = torch.arange(height) + w = torch.arange(width) + l = torch.arange(1) # [0] - layer dimension + + # Create position IDs: (H*W, 4) + latent_ids = torch.cartesian_prod(t, h, w, l) + + # Expand to batch: (B, H*W, 4) + latent_ids = latent_ids.unsqueeze(0).expand(batch_size, -1, -1) + + return latent_ids + +def _patchify_latents(latents): + batch_size, num_channels_latents, height, width = latents.shape + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 1, 3, 5, 2, 4) + latents = latents.reshape(batch_size, num_channels_latents * 4, height // 2, width // 2) + return latents + +def _pack_latents(latents): + """ + pack latents: (batch_size, num_channels, height, width) -> (batch_size, height * width, num_channels) + """ + + batch_size, num_channels, height, width = latents.shape + latents = latents.reshape(batch_size, num_channels, height * width).permute(0, 2, 1) + + return latents + +def format_text_input(prompts: List[str], system_message: str = None): + # Remove [IMG] tokens from prompts to avoid Pixtral validation issues + # when truncation is enabled. The processor counts [IMG] tokens and fails + # if the count changes after truncation. + cleaned_txt = [prompt.replace("[IMG]", "") for prompt in prompts] + + return [ + [ + { + "role": "system", + "content": [{"type": "text", "text": system_message}], + }, + {"role": "user", "content": [{"type": "text", "text": prompt}]}, + ] + for prompt in cleaned_txt + ] + +def _get_mistral_3_small_prompt_embeds( + text_encoder: Mistral3ForConditionalGeneration, + tokenizer: PixtralProcessor, + prompt: Union[str, List[str]], + dtype: Optional[torch.dtype] = None, + device: Optional[torch.device] = None, + max_sequence_length: int = 512, + # fmt: off + system_message: str = "You are an AI that reasons about image descriptions. You give structured responses focusing on object relationships, object attribution and actions without speculation.", + # fmt: on + hidden_states_layers: List[int] = (10, 20, 30), +): + dtype = text_encoder.dtype if dtype is None else dtype + device = text_encoder.device if device is None else device + + prompt = [prompt] if isinstance(prompt, str) else prompt + + # Format input messages + messages_batch = format_text_input(prompts=prompt, system_message=system_message) + + # Process all messages at once + inputs = tokenizer.apply_chat_template( + messages_batch, + add_generation_prompt=False, + tokenize=True, + return_dict=True, + return_tensors="pt", + padding="max_length", + truncation=True, + max_length=max_sequence_length, + ) + + # Move to device + input_ids = inputs["input_ids"].to(device) + attention_mask = inputs["attention_mask"].to(device) + + # Forward pass through the model + output = text_encoder( + input_ids=input_ids, + attention_mask=attention_mask, + output_hidden_states=True, + use_cache=False, + ) + + # Only use outputs from intermediate layers and stack them + out = torch.stack([output.hidden_states[k] for k in hidden_states_layers], dim=1) + out = out.to(dtype=dtype, device=device) + + batch_size, num_channels, seq_len, hidden_dim = out.shape + prompt_embeds = out.permute(0, 2, 1, 3).reshape(batch_size, seq_len, num_channels * hidden_dim) + + return prompt_embeds + +def _prepare_text_ids( + x: torch.Tensor, # (B, L, D) or (L, D) + t_coord: Optional[torch.Tensor] = None, +): + B, L, _ = x.shape + out_ids = [] + + for i in range(B): + t = torch.arange(1) if t_coord is None else t_coord[i] + h = torch.arange(1) + w = torch.arange(1) + l = torch.arange(L) + + coords = torch.cartesian_prod(t, h, w, l) + out_ids.append(coords) + + return torch.stack(out_ids) + +def encode_prompt( + prompt: Union[str, List[str]], + device: Optional[torch.device] = None, + text_encoder=None, + tokenizer=None, + num_images_per_prompt: int = 1, + prompt_embeds: Optional[torch.Tensor] = None, + max_sequence_length: int = 512, + text_encoder_out_layers: Tuple[int] = (10, 20, 30), + system_message = "You are an AI that reasons about image descriptions. You give structured responses focusing on object relationships, object attribution and actions without speculation." +): + if prompt is None: + prompt = "" + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt_embeds is None: + prompt_embeds = _get_mistral_3_small_prompt_embeds( + text_encoder=text_encoder, + tokenizer=tokenizer, + prompt=prompt, + device=device, + max_sequence_length=max_sequence_length, + system_message=system_message, + hidden_states_layers=text_encoder_out_layers, + ) + + batch_size, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + text_ids = _prepare_text_ids(prompt_embeds) + text_ids = text_ids.to(device) + return prompt_embeds, text_ids + +# Will error if the minimal version of diffusers is not installed. Remove at your own risks. +check_min_version("0.18.0.dev0") + +logger = get_logger(__name__, log_level="INFO") + +def log_validation(vae, text_encoder, tokenizer, transformer3d, network, args, accelerator, weight_dtype, global_step): + try: + logger.info("Running validation... ") + + transformer3d_val = Flux2Transformer2DModel.from_pretrained( + args.pretrained_model_name_or_path, subfolder="transformer", torch_dtype=weight_dtype, + low_cpu_mem_usage=True, + ).to(weight_dtype) + transformer3d_val.load_state_dict(accelerator.unwrap_model(transformer3d).state_dict()) + scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained( + args.pretrained_model_name_or_path, + subfolder="scheduler" + ) + transformer3d = transformer3d.to("cpu") + pipeline = Flux2Pipeline( + vae=accelerator.unwrap_model(vae).to(weight_dtype), + text_encoder=accelerator.unwrap_model(text_encoder), + tokenizer=tokenizer, + transformer=transformer3d_val, + scheduler=scheduler, + ) + pipeline = pipeline.to(accelerator.device) + pipeline = merge_lora( + pipeline, None, 1, accelerator.device, state_dict=accelerator.unwrap_model(network).state_dict(), transformer_only=True + ) + + if args.seed is None: + generator = None + else: + generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) + + for i in range(len(args.validation_prompts)): + with torch.no_grad(): + sample = pipeline( + args.validation_prompts[i], + negative_prompt = "bad detailed", + height = args.image_sample_size, + width = args.image_sample_size, + generator = generator + ).images + os.makedirs(os.path.join(args.output_dir, "sample"), exist_ok=True) + image = sample[0].save(os.path.join(args.output_dir, f"sample/sample-{global_step}-image-{i}.gif")) + + del pipeline + del transformer3d_val + gc.collect() + torch.cuda.empty_cache() + torch.cuda.ipc_collect() + transformer3d = transformer3d.to(accelerator.device) + except Exception as e: + gc.collect() + torch.cuda.empty_cache() + torch.cuda.ipc_collect() + print(f"Eval error with info {e}") + transformer3d = transformer3d.to(accelerator.device) + +def parse_args(): + parser = argparse.ArgumentParser(description="Simple example of a training script.") + parser.add_argument( + "--input_perturbation", type=float, default=0, help="The scale of input perturbation. Recommended 0.1." + ) + parser.add_argument( + "--pretrained_model_name_or_path", + type=str, + default=None, + required=True, + help="Path to pretrained model or model identifier from huggingface.co/models.", + ) + parser.add_argument( + "--revision", + type=str, + default=None, + required=False, + help="Revision of pretrained model identifier from huggingface.co/models.", + ) + parser.add_argument( + "--variant", + type=str, + default=None, + help="Variant of the model files of the pretrained model identifier from huggingface.co/models, 'e.g.' fp16", + ) + parser.add_argument( + "--train_data_dir", + type=str, + default=None, + help=( + "A folder containing the training data. " + ), + ) + parser.add_argument( + "--train_data_meta", + type=str, + default=None, + help=( + "A csv containing the training data. " + ), + ) + parser.add_argument( + "--max_train_samples", + type=int, + default=None, + help=( + "For debugging purposes or quicker training, truncate the number of training examples to this " + "value if set." + ), + ) + parser.add_argument( + "--validation_prompts", + type=str, + default=None, + nargs="+", + help=("A set of prompts evaluated every `--validation_epochs` and logged to `--report_to`."), + ) + parser.add_argument( + "--output_dir", + type=str, + default="sd-model-finetuned", + help="The output directory where the model predictions and checkpoints will be written.", + ) + parser.add_argument( + "--cache_dir", + type=str, + default=None, + help="The directory where the downloaded models and datasets will be stored.", + ) + parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.") + parser.add_argument( + "--random_flip", + action="store_true", + help="whether to randomly flip images horizontally", + ) + parser.add_argument( + "--use_came", + action="store_true", + help="whether to use came", + ) + parser.add_argument( + "--multi_stream", + action="store_true", + help="whether to use cuda multi-stream", + ) + parser.add_argument( + "--train_batch_size", type=int, default=16, help="Batch size (per device) for the training dataloader." + ) + parser.add_argument( + "--vae_mini_batch", type=int, default=32, help="mini batch size for vae." + ) + parser.add_argument("--num_train_epochs", type=int, default=100) + parser.add_argument( + "--max_train_steps", + type=int, + default=None, + help="Total number of training steps to perform. If provided, overrides num_train_epochs.", + ) + parser.add_argument( + "--gradient_accumulation_steps", + type=int, + default=1, + help="Number of updates steps to accumulate before performing a backward/update pass.", + ) + parser.add_argument( + "--gradient_checkpointing", + action="store_true", + help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.", + ) + parser.add_argument( + "--learning_rate", + type=float, + default=1e-4, + help="Initial learning rate (after the potential warmup period) to use.", + ) + parser.add_argument( + "--scale_lr", + action="store_true", + default=False, + help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.", + ) + parser.add_argument( + "--lr_scheduler", + type=str, + default="constant", + help=( + 'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",' + ' "constant", "constant_with_warmup"]' + ), + ) + parser.add_argument( + "--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler." + ) + parser.add_argument( + "--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes." + ) + parser.add_argument( + "--allow_tf32", + action="store_true", + help=( + "Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see" + " https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices" + ), + ) + parser.add_argument("--use_ema", action="store_true", help="Whether to use EMA model.") + parser.add_argument( + "--non_ema_revision", + type=str, + default=None, + required=False, + help=( + "Revision of pretrained non-ema model identifier. Must be a branch, tag or git identifier of the local or" + " remote repository specified with --pretrained_model_name_or_path." + ), + ) + parser.add_argument( + "--dataloader_num_workers", + type=int, + default=0, + help=( + "Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process." + ), + ) + parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.") + parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.") + parser.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.") + parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer") + parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.") + parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.") + parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.") + parser.add_argument( + "--prediction_type", + type=str, + default=None, + help="The prediction_type that shall be used for training. Choose between 'epsilon' or 'v_prediction' or leave `None`. If left to `None` the default prediction type of the scheduler: `noise_scheduler.config.prediciton_type` is chosen.", + ) + parser.add_argument( + "--hub_model_id", + type=str, + default=None, + help="The name of the repository to keep in sync with the local `output_dir`.", + ) + parser.add_argument( + "--logging_dir", + type=str, + default="logs", + help=( + "[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to" + " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***." + ), + ) + parser.add_argument( + "--mixed_precision", + type=str, + default=None, + choices=["no", "fp16", "bf16"], + help=( + "Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >=" + " 1.10.and an Nvidia Ampere GPU. Default to the value of accelerate config of the current system or the" + " flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config." + ), + ) + parser.add_argument( + "--report_to", + type=str, + default="tensorboard", + help=( + 'The integration to report the results and logs to. Supported platforms are `"tensorboard"`' + ' (default), `"wandb"` and `"comet_ml"`. Use `"all"` to report to all integrations.' + ), + ) + parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank") + parser.add_argument( + "--checkpointing_steps", + type=int, + default=500, + help=( + "Save a checkpoint of the training state every X updates. These checkpoints are only suitable for resuming" + " training using `--resume_from_checkpoint`." + ), + ) + parser.add_argument( + "--checkpoints_total_limit", + type=int, + default=None, + help=("Max number of checkpoints to store."), + ) + parser.add_argument( + "--resume_from_checkpoint", + type=str, + default=None, + help=( + "Whether training should be resumed from a previous checkpoint. Use a path saved by" + ' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.' + ), + ) + parser.add_argument("--noise_offset", type=float, default=0, help="The scale of noise offset.") + parser.add_argument( + "--validation_epochs", + type=int, + default=5, + help="Run validation every X epochs.", + ) + parser.add_argument( + "--validation_steps", + type=int, + default=2000, + help="Run validation every X steps.", + ) + parser.add_argument( + "--tracker_project_name", + type=str, + default="text2image-fine-tune", + help=( + "The `project_name` argument passed to Accelerator.init_trackers for" + " more information see https://huggingface.co/docs/accelerate/v0.17.0/en/package_reference/accelerator#accelerate.Accelerator" + ), + ) + + parser.add_argument( + "--rank", + type=int, + default=128, + help=("The dimension of the LoRA update matrices."), + ) + parser.add_argument( + "--network_alpha", + type=int, + default=64, + help=("The dimension of the LoRA update matrices."), + ) + parser.add_argument( + "--use_peft_lora", action="store_true", help="Whether or not to use peft lora." + ) + parser.add_argument( + "--train_text_encoder", + action="store_true", + help="Whether to train the text encoder. If set, the text encoder should be float32 precision.", + ) + parser.add_argument( + "--snr_loss", action="store_true", help="Whether or not to use snr_loss." + ) + parser.add_argument( + "--uniform_sampling", action="store_true", help="Whether or not to use uniform_sampling." + ) + parser.add_argument( + "--enable_text_encoder_in_dataloader", action="store_true", help="Whether or not to use text encoder in dataloader." + ) + parser.add_argument( + "--enable_bucket", action="store_true", help="Whether enable bucket sample in datasets." + ) + parser.add_argument( + "--random_ratio_crop", action="store_true", help="Whether enable random ratio crop sample in datasets." + ) + parser.add_argument( + "--random_hw_adapt", action="store_true", help="Whether enable random adapt height and width in datasets." + ) + parser.add_argument( + "--train_sampling_steps", + type=int, + default=1000, + help="Run train_sampling_steps.", + ) + parser.add_argument( + "--image_sample_size", + type=int, + default=512, + help="Sample size of the image.", + ) + parser.add_argument( + "--fix_sample_size", + nargs=2, type=int, default=None, + help="Fix Sample size [height, width] when using bucket and collate_fn." + ) + parser.add_argument( + "--config_path", + type=str, + default=None, + help=( + "The config of the model in training." + ), + ) + parser.add_argument( + "--transformer_path", + type=str, + default=None, + help=("If you want to load the weight from other transformers, input its path."), + ) + parser.add_argument( + "--vae_path", + type=str, + default=None, + help=("If you want to load the weight from other vaes, input its path."), + ) + parser.add_argument("--save_state", action="store_true", help="Whether or not to save state.") + + parser.add_argument( + "--use_deepspeed", action="store_true", help="Whether or not to use deepspeed." + ) + parser.add_argument( + "--use_fsdp", action="store_true", help="Whether or not to use fsdp." + ) + parser.add_argument( + "--low_vram", action="store_true", help="Whether enable low_vram mode." + ) + parser.add_argument( + "--weighting_scheme", + type=str, + default="none", + choices=["sigma_sqrt", "logit_normal", "mode", "cosmap", "none"], + help=('We default to the "none" weighting scheme for uniform sampling and uniform loss'), + ) + parser.add_argument( + "--logit_mean", type=float, default=0.0, help="mean to use when using the `'logit_normal'` weighting scheme." + ) + parser.add_argument( + "--logit_std", type=float, default=1.0, help="std to use when using the `'logit_normal'` weighting scheme." + ) + parser.add_argument( + "--mode_scale", + type=float, + default=1.29, + help="Scale of mode weighting scheme. Only effective when using the `'mode'` as the `weighting_scheme`.", + ) + parser.add_argument( + "--lora_skip_name", + type=str, + default=None, + help=("The module is not trained in loras. "), + ) + parser.add_argument( + "--guidance_scale", + type=float, + default=3.5, + help="the FLUX.1 dev variant is a guidance distilled model", + ) + parser.add_argument( + "--target_name", + type=str, + default=None, + help=("The module is trained in loras. "), + ) + + args = parser.parse_args() + env_local_rank = int(os.environ.get("LOCAL_RANK", -1)) + if env_local_rank != -1 and env_local_rank != args.local_rank: + args.local_rank = env_local_rank + + # default to using the same revision for the non-ema model if not specified + if args.non_ema_revision is None: + args.non_ema_revision = args.revision + + return args + + +def main(): + args = parse_args() + + if args.report_to == "wandb" and args.hub_token is not None: + raise ValueError( + "You cannot use both --report_to=wandb and --hub_token due to a security risk of exposing your token." + " Please use `huggingface-cli login` to authenticate with the Hub." + ) + + if args.non_ema_revision is not None: + deprecate( + "non_ema_revision!=None", + "0.15.0", + message=( + "Downloading 'non_ema' weights from revision branches of the Hub is deprecated. Please make sure to" + " use `--variant=non_ema` instead." + ), + ) + logging_dir = os.path.join(args.output_dir, args.logging_dir) + + accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir) + + accelerator = Accelerator( + gradient_accumulation_steps=args.gradient_accumulation_steps, + mixed_precision=args.mixed_precision, + log_with=args.report_to, + project_config=accelerator_project_config, + ) + + deepspeed_plugin = accelerator.state.deepspeed_plugin if hasattr(accelerator.state, "deepspeed_plugin") else None + fsdp_plugin = accelerator.state.fsdp_plugin if hasattr(accelerator.state, "fsdp_plugin") else None + if deepspeed_plugin is not None: + zero_stage = int(deepspeed_plugin.zero_stage) + fsdp_stage = 0 + print(f"Using DeepSpeed Zero stage: {zero_stage}") + + args.use_deepspeed = True + if zero_stage == 3: + print(f"Auto set save_state to True because zero_stage == 3") + args.save_state = True + elif fsdp_plugin is not None: + from torch.distributed.fsdp import ShardingStrategy + zero_stage = 0 + if fsdp_plugin.sharding_strategy is ShardingStrategy.FULL_SHARD: + fsdp_stage = 3 + elif fsdp_plugin.sharding_strategy is None: # The fsdp_plugin.sharding_strategy is None in FSDP 2. + fsdp_stage = 3 + elif fsdp_plugin.sharding_strategy is ShardingStrategy.SHARD_GRAD_OP: + fsdp_stage = 2 + else: + fsdp_stage = 0 + print(f"Using FSDP stage: {fsdp_stage}") + + args.use_fsdp = True + if fsdp_stage == 3: + print(f"Auto set save_state to True because fsdp_stage == 3") + args.save_state = True + else: + zero_stage = 0 + fsdp_stage = 0 + print("DeepSpeed is not enabled.") + + if accelerator.is_main_process: + writer = SummaryWriter(log_dir=logging_dir) + + # Make one log on every process with the configuration for debugging. + logging.basicConfig( + format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", + datefmt="%m/%d/%Y %H:%M:%S", + level=logging.INFO, + ) + logger.info(accelerator.state, main_process_only=False) + if accelerator.is_local_main_process: + datasets.utils.logging.set_verbosity_warning() + transformers.utils.logging.set_verbosity_warning() + diffusers.utils.logging.set_verbosity_info() + else: + datasets.utils.logging.set_verbosity_error() + transformers.utils.logging.set_verbosity_error() + diffusers.utils.logging.set_verbosity_error() + + # If passed along, set the training seed now. + if args.seed is not None: + set_seed(args.seed) + rng = np.random.default_rng(np.random.PCG64(args.seed + accelerator.process_index)) + torch_rng = torch.Generator(accelerator.device).manual_seed(args.seed + accelerator.process_index) + else: + rng = None + torch_rng = None + index_rng = np.random.default_rng(np.random.PCG64(43)) + print(f"Init rng with seed {args.seed + accelerator.process_index}. Process_index is {accelerator.process_index}") + + # Handle the repository creation + if accelerator.is_main_process: + if args.output_dir is not None: + os.makedirs(args.output_dir, exist_ok=True) + + # For mixed precision training we cast all non-trainable weigths (vae, non-lora text_encoder and non-lora transformer3d) to half-precision + # as these weights are only used for inference, keeping weights in full precision is not required. + weight_dtype = torch.float32 + if accelerator.mixed_precision == "fp16": + weight_dtype = torch.float16 + args.mixed_precision = accelerator.mixed_precision + elif accelerator.mixed_precision == "bf16": + weight_dtype = torch.bfloat16 + args.mixed_precision = accelerator.mixed_precision + + # Load scheduler, tokenizer and models. + noise_scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained( + args.pretrained_model_name_or_path, + subfolder="scheduler" + ) + + # Get Tokenizer + tokenizer = PixtralProcessor.from_pretrained( + args.pretrained_model_name_or_path, subfolder="tokenizer" + ) + + def deepspeed_zero_init_disabled_context_manager(): + """ + returns either a context list that includes one that will disable zero.Init or an empty context list + """ + deepspeed_plugin = AcceleratorState().deepspeed_plugin if accelerate.state.is_initialized() else None + if deepspeed_plugin is None: + return [] + + return [deepspeed_plugin.zero3_init_context_manager(enable=False)] + + # Currently Accelerate doesn't know how to handle multiple models under Deepspeed ZeRO stage 3. + # For this to work properly all models must be run through `accelerate.prepare`. But accelerate + # will try to assign the same optimizer with the same weights to all models during + # `deepspeed.initialize`, which of course doesn't work. + # + # For now the following workaround will partially support Deepspeed ZeRO-3, by excluding the 2 + # frozen models from being partitioned during `zero.Init` which gets called during + # `from_pretrained` So Mistral3ForConditionalGeneration and AutoencoderKLFlux2 will not enjoy the parameter sharding + # across multiple gpus and only UNet2DConditionModel will get ZeRO sharded. + with ContextManagers(deepspeed_zero_init_disabled_context_manager()): + # Get Text encoder + text_encoder = Mistral3ForConditionalGeneration.from_pretrained( + args.pretrained_model_name_or_path, subfolder="text_encoder", torch_dtype=weight_dtype + ) + text_encoder = text_encoder.eval() + # Get Vae + vae = AutoencoderKLFlux2.from_pretrained( + args.pretrained_model_name_or_path, + subfolder="vae" + ).to(weight_dtype) + vae.eval() + latents_bn_mean = vae.bn.running_mean.view(1, -1, 1, 1).to(accelerator.device, weight_dtype) + latents_bn_std = torch.sqrt(vae.bn.running_var.view(1, -1, 1, 1) + vae.config.batch_norm_eps).to(accelerator.device, weight_dtype) + + # Get Transformer + transformer3d = Flux2Transformer2DModel.from_pretrained( + args.pretrained_model_name_or_path, + subfolder="transformer", + torch_dtype=weight_dtype, + ).to(weight_dtype) + + # Freeze vae and text_encoder and set transformer3d to trainable + vae.requires_grad_(False) + text_encoder.requires_grad_(False) + transformer3d.requires_grad_(False) + + # Lora will work with this... + if args.use_peft_lora: + from peft import (LoraConfig, get_peft_model_state_dict, + inject_adapter_in_model) + lora_config = LoraConfig(r=args.rank, lora_alpha=args.network_alpha, target_modules=args.target_name.split(",")) + transformer3d = inject_adapter_in_model(lora_config, transformer3d) + + network = None + else: + network = create_network( + 1.0, + args.rank, + args.network_alpha, + text_encoder, + transformer3d, + neuron_dropout=None, + target_name=args.target_name, + skip_name=args.lora_skip_name, + ) + network = network.to(weight_dtype) + network.apply_to(text_encoder, transformer3d, args.train_text_encoder and not args.training_with_video_token_length, True) + + if args.transformer_path is not None: + print(f"From checkpoint: {args.transformer_path}") + if args.transformer_path.endswith("safetensors"): + from safetensors.torch import load_file, safe_open + state_dict = load_file(args.transformer_path) + else: + state_dict = torch.load(args.transformer_path, map_location="cpu") + state_dict = state_dict["state_dict"] if "state_dict" in state_dict else state_dict + + m, u = transformer3d.load_state_dict(state_dict, strict=False) + print(f"missing keys: {len(m)}, unexpected keys: {len(u)}") + assert len(u) == 0 + + if args.vae_path is not None: + print(f"From checkpoint: {args.vae_path}") + if args.vae_path.endswith("safetensors"): + from safetensors.torch import load_file, safe_open + state_dict = load_file(args.vae_path) + else: + state_dict = torch.load(args.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)}") + assert len(u) == 0 + + # `accelerate` 0.16.0 will have better support for customized saving + if version.parse(accelerate.__version__) >= version.parse("0.16.0"): + # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format + if fsdp_stage != 0: + def save_model_hook(models, weights, output_dir): + accelerate_state_dict = accelerator.get_state_dict(models[-1], unwrap=True) + if accelerator.is_main_process: + from safetensors.torch import save_file + safetensor_save_path = os.path.join(output_dir, f"lora_diffusion_pytorch_model.safetensors") + if args.use_peft_lora: + network_state_dict = get_peft_model_state_dict(accelerator.unwrap_model(models[-1]), accelerate_state_dict) + else: + network_state_dict = {} + for key in accelerate_state_dict: + if "network" in key: + network_state_dict[key.replace("network.", "")] = accelerate_state_dict[key].to(weight_dtype) + save_file(network_state_dict, safetensor_save_path, metadata={"format": "pt"}) + + with open(os.path.join(output_dir, "sampler_pos_start.pkl"), 'wb') as file: + pickle.dump([batch_sampler.sampler._pos_start, first_epoch], file) + + def load_model_hook(models, input_dir): + pkl_path = os.path.join(input_dir, "sampler_pos_start.pkl") + if os.path.exists(pkl_path): + with open(pkl_path, 'rb') as file: + loaded_number, _ = pickle.load(file) + batch_sampler.sampler._pos_start = max(loaded_number - args.dataloader_num_workers * accelerator.num_processes * 2, 0) + print(f"Load pkl from {pkl_path}. Get loaded_number = {loaded_number}.") + + elif zero_stage == 3: + # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format + def save_model_hook(models, weights, output_dir): + accelerate_state_dict = accelerator.get_state_dict(models[-1], unwrap=True) + if accelerator.is_main_process: + from safetensors.torch import save_file + safetensor_save_path = os.path.join(output_dir, f"lora_diffusion_pytorch_model.safetensors") + if args.use_peft_lora: + network_state_dict = get_peft_model_state_dict(accelerator.unwrap_model(models[-1]), accelerate_state_dict) + else: + network_state_dict = accelerate_state_dict + save_file(network_state_dict, safetensor_save_path, metadata={"format": "pt"}) + + with open(os.path.join(output_dir, "sampler_pos_start.pkl"), 'wb') as file: + pickle.dump([batch_sampler.sampler._pos_start, first_epoch], file) + + def load_model_hook(models, input_dir): + pkl_path = os.path.join(input_dir, "sampler_pos_start.pkl") + if os.path.exists(pkl_path): + with open(pkl_path, 'rb') as file: + loaded_number, _ = pickle.load(file) + batch_sampler.sampler._pos_start = max(loaded_number - args.dataloader_num_workers * accelerator.num_processes * 2, 0) + print(f"Load pkl from {pkl_path}. Get loaded_number = {loaded_number}.") + else: + # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format + def save_model_hook(models, weights, output_dir): + if accelerator.is_main_process: + safetensor_save_path = os.path.join(output_dir, f"lora_diffusion_pytorch_model.safetensors") + if args.use_peft_lora: + save_model(safetensor_save_path, get_peft_model_state_dict(accelerator.unwrap_model(models[-1]))) + else: + save_model(safetensor_save_path, accelerator.unwrap_model(models[-1])) + + if not args.use_deepspeed: + for _ in range(len(weights)): + weights.pop() + + with open(os.path.join(output_dir, "sampler_pos_start.pkl"), 'wb') as file: + pickle.dump([batch_sampler.sampler._pos_start, first_epoch], file) + + def load_model_hook(models, input_dir): + pkl_path = os.path.join(input_dir, "sampler_pos_start.pkl") + if os.path.exists(pkl_path): + with open(pkl_path, 'rb') as file: + loaded_number, _ = pickle.load(file) + batch_sampler.sampler._pos_start = max(loaded_number - args.dataloader_num_workers * accelerator.num_processes * 2, 0) + print(f"Load pkl from {pkl_path}. Get loaded_number = {loaded_number}.") + + accelerator.register_save_state_pre_hook(save_model_hook) + accelerator.register_load_state_pre_hook(load_model_hook) + + if args.gradient_checkpointing: + transformer3d.enable_gradient_checkpointing() + + # Enable TF32 for faster training on Ampere GPUs, + # cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices + if args.allow_tf32: + torch.backends.cuda.matmul.allow_tf32 = True + + if args.scale_lr: + args.learning_rate = ( + args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes + ) + + # Initialize the optimizer + if args.use_8bit_adam: + try: + import bitsandbytes as bnb + except ImportError: + raise ImportError( + "Please install bitsandbytes to use 8-bit Adam. You can do so by running `pip install bitsandbytes`" + ) + + optimizer_cls = bnb.optim.AdamW8bit + elif args.use_came: + try: + from came_pytorch import CAME + except: + raise ImportError( + "Please install came_pytorch to use CAME. You can do so by running `pip install came_pytorch`" + ) + + optimizer_cls = CAME + else: + optimizer_cls = torch.optim.AdamW + + if args.use_peft_lora: + logging.info("Add peft parameters") + trainable_params = list(filter(lambda p: p.requires_grad, transformer3d.parameters())) + trainable_params_optim = list(filter(lambda p: p.requires_grad, transformer3d.parameters())) + else: + logging.info("Add network parameters") + trainable_params = list(filter(lambda p: p.requires_grad, network.parameters())) + trainable_params_optim = network.prepare_optimizer_params(args.learning_rate / 2, args.learning_rate, args.learning_rate) + + if args.use_came: + optimizer = optimizer_cls( + trainable_params_optim, + lr=args.learning_rate, + # weight_decay=args.adam_weight_decay, + betas=(0.9, 0.999, 0.9999), + eps=(1e-30, 1e-16) + ) + else: + optimizer = optimizer_cls( + trainable_params_optim, + lr=args.learning_rate, + betas=(args.adam_beta1, args.adam_beta2), + weight_decay=args.adam_weight_decay, + eps=args.adam_epsilon, + ) + + # Get the training dataset + if args.fix_sample_size is not None and args.enable_bucket: + args.image_sample_size = max(max(args.fix_sample_size), args.image_sample_size) + args.random_hw_adapt = False + + # Get the dataset + train_dataset = ImageVideoDataset( + args.train_data_meta, args.train_data_dir, + image_sample_size=args.image_sample_size, + enable_bucket=args.enable_bucket, + ) + + def worker_init_fn(_seed): + _seed = _seed * 256 + def _worker_init_fn(worker_id): + print(f"worker_init_fn with {_seed + worker_id}") + np.random.seed(_seed + worker_id) + random.seed(_seed + worker_id) + return _worker_init_fn + + if args.enable_bucket: + aspect_ratio_sample_size = {key : [x / 512 * args.image_sample_size for x in ASPECT_RATIO_512[key]] for key in ASPECT_RATIO_512.keys()} + batch_sampler_generator = torch.Generator().manual_seed(args.seed) + batch_sampler = AspectRatioBatchImageVideoSampler( + sampler=RandomSampler(train_dataset, generator=batch_sampler_generator), dataset=train_dataset.dataset, + batch_size=args.train_batch_size, train_folder = args.train_data_dir, drop_last=True, + aspect_ratios=aspect_ratio_sample_size, + ) + + def collate_fn(examples): + def get_random_downsample_ratio(sample_size, image_ratio=[], + all_choices=False, rng=None): + def _create_special_list(length): + if length == 1: + return [1.0] + if length >= 2: + first_element = 0.90 + remaining_sum = 1.0 - first_element + other_elements_value = remaining_sum / (length - 1) + special_list = [first_element] + [other_elements_value] * (length - 1) + return special_list + + if sample_size >= 1536: + number_list = [1, 1.25, 1.5, 2, 2.5, 3] + image_ratio + elif sample_size >= 1024: + number_list = [1, 1.25, 1.5, 2] + image_ratio + elif sample_size >= 768: + number_list = [1, 1.25, 1.5] + image_ratio + elif sample_size >= 512: + number_list = [1] + image_ratio + else: + number_list = [1] + + if all_choices: + return number_list + + number_list_prob = np.array(_create_special_list(len(number_list))) + if rng is None: + return np.random.choice(number_list, p = number_list_prob) + else: + return rng.choice(number_list, p = number_list_prob) + + # Create new output + new_examples = {} + new_examples["pixel_values"] = [] + new_examples["text"] = [] + + # Get downsample ratio in image + pixel_value = examples[0]["pixel_values"] + data_type = examples[0]["data_type"] + f, h, w, c = np.shape(pixel_value) + + random_downsample_ratio = 1 if not args.random_hw_adapt else get_random_downsample_ratio(args.image_sample_size) + + aspect_ratio_sample_size = {key : [x / 512 * args.image_sample_size / random_downsample_ratio for x in ASPECT_RATIO_512[key]] for key in ASPECT_RATIO_512.keys()} + aspect_ratio_random_crop_sample_size = {key : [x / 512 * args.image_sample_size / random_downsample_ratio for x in ASPECT_RATIO_RANDOM_CROP_512[key]] for key in ASPECT_RATIO_RANDOM_CROP_512.keys()} + + if args.fix_sample_size is not None: + fix_sample_size = [int(x / 16) * 16 for x in args.fix_sample_size] + elif args.random_ratio_crop: + if rng is None: + random_sample_size = aspect_ratio_random_crop_sample_size[ + np.random.choice(list(aspect_ratio_random_crop_sample_size.keys()), p = ASPECT_RATIO_RANDOM_CROP_PROB) + ] + else: + random_sample_size = aspect_ratio_random_crop_sample_size[ + rng.choice(list(aspect_ratio_random_crop_sample_size.keys()), p = ASPECT_RATIO_RANDOM_CROP_PROB) + ] + random_sample_size = [int(x / 16) * 16 for x in random_sample_size] + else: + closest_size, closest_ratio = get_closest_ratio(h, w, ratios=aspect_ratio_sample_size) + closest_size = [int(x / 16) * 16 for x in closest_size] + + for example in examples: + if args.fix_sample_size is not None: + # To 0~1 + pixel_values = torch.from_numpy(example["pixel_values"]).permute(0, 3, 1, 2).contiguous() + pixel_values = pixel_values / 255. + + # Get adapt hw for resize + fix_sample_size = list(map(lambda x: int(x), fix_sample_size)) + transform = transforms.Compose([ + transforms.Resize(fix_sample_size, interpolation=transforms.InterpolationMode.BILINEAR), # Image.BICUBIC + transforms.CenterCrop(fix_sample_size), + transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], inplace=True), + ]) + elif args.random_ratio_crop: + # To 0~1 + pixel_values = torch.from_numpy(example["pixel_values"]).permute(0, 3, 1, 2).contiguous() + pixel_values = pixel_values / 255. + + # Get adapt hw for resize + b, c, h, w = pixel_values.size() + th, tw = random_sample_size + if th / tw > h / w: + nh = int(th) + nw = int(w / h * nh) + else: + nw = int(tw) + nh = int(h / w * nw) + + transform = transforms.Compose([ + transforms.Resize([nh, nw]), + transforms.CenterCrop([int(x) for x in random_sample_size]), + transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], inplace=True), + ]) + else: + # To 0~1 + pixel_values = torch.from_numpy(example["pixel_values"]).permute(0, 3, 1, 2).contiguous() + pixel_values = pixel_values / 255. + + # Get adapt hw for resize + closest_size = list(map(lambda x: int(x), closest_size)) + if closest_size[0] / h > closest_size[1] / w: + resize_size = closest_size[0], int(w * closest_size[0] / h) + else: + resize_size = int(h * closest_size[1] / w), closest_size[1] + + transform = transforms.Compose([ + transforms.Resize(resize_size, interpolation=transforms.InterpolationMode.BILINEAR), # Image.BICUBIC + transforms.CenterCrop(closest_size), + transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], inplace=True), + ]) + new_examples["pixel_values"].append(transform(pixel_values)) + new_examples["text"].append(example["text"]) + + # Limit the number of frames to the same + new_examples["pixel_values"] = torch.stack([example for example in new_examples["pixel_values"]]) + + # Encode prompts when enable_text_encoder_in_dataloader=True + if args.enable_text_encoder_in_dataloader: + prompt_embeds, text_ids = encode_prompt( + batch['text'], device="cpu", + text_encoder=text_encoder, + tokenizer=tokenizer, + ) + + new_examples['prompt_embeds'] = prompt_embeds + new_examples['text_ids'] = text_ids + + return new_examples + + # DataLoaders creation: + train_dataloader = torch.utils.data.DataLoader( + train_dataset, + batch_sampler=batch_sampler, + collate_fn=collate_fn, + persistent_workers=True if args.dataloader_num_workers != 0 else False, + num_workers=args.dataloader_num_workers, + worker_init_fn=worker_init_fn(args.seed + accelerator.process_index) + ) + else: + # DataLoaders creation: + batch_sampler_generator = torch.Generator().manual_seed(args.seed) + batch_sampler = ImageVideoSampler(RandomSampler(train_dataset, generator=batch_sampler_generator), train_dataset, args.train_batch_size) + train_dataloader = torch.utils.data.DataLoader( + train_dataset, + batch_sampler=batch_sampler, + persistent_workers=True if args.dataloader_num_workers != 0 else False, + num_workers=args.dataloader_num_workers, + worker_init_fn=worker_init_fn(args.seed + accelerator.process_index) + ) + + # Scheduler and math around the number of training steps. + overrode_max_train_steps = False + num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps) + if args.max_train_steps is None: + args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch + overrode_max_train_steps = True + + lr_scheduler = get_scheduler( + args.lr_scheduler, + optimizer=optimizer, + num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes, + num_training_steps=args.max_train_steps * accelerator.num_processes, + ) + # Prepare everything with our `accelerator`. + if args.use_peft_lora: + transformer3d, optimizer, train_dataloader, lr_scheduler = accelerator.prepare( + transformer3d, optimizer, train_dataloader, lr_scheduler + ) + elif fsdp_stage != 0: + transformer3d.network = network + transformer3d = transformer3d.to(dtype=weight_dtype) + transformer3d, optimizer, train_dataloader, lr_scheduler = accelerator.prepare( + transformer3d, optimizer, train_dataloader, lr_scheduler + ) + else: + network, optimizer, train_dataloader, lr_scheduler = accelerator.prepare( + network, optimizer, train_dataloader, lr_scheduler + ) + + if zero_stage != 0 and not args.use_peft_lora: + from functools import partial + + from videox_fun.dist import set_multi_gpus_devices, shard_model + shard_fn = partial(shard_model, device_id=accelerator.device, param_dtype=weight_dtype, module_to_wrapper=list(transformer3d.transformer_blocks) + list(transformer3d.single_transformer_blocks)) + transformer3d = shard_fn(transformer3d) + + if fsdp_stage != 0 or zero_stage != 0: + from functools import partial + + from videox_fun.dist import set_multi_gpus_devices, shard_model + shard_fn = partial(shard_model, device_id=accelerator.device, param_dtype=weight_dtype, module_to_wrapper=text_encoder.language_model.layers) + text_encoder = shard_fn(text_encoder) + + # Move text_encode and vae to gpu and cast to weight_dtype + vae.to(accelerator.device if not args.low_vram else "cpu", dtype=weight_dtype) + transformer3d.to(accelerator.device, dtype=weight_dtype) + if not args.enable_text_encoder_in_dataloader: + text_encoder.to(accelerator.device if not args.low_vram else "cpu", dtype=weight_dtype) + + # We need to recalculate our total training steps as the size of the training dataloader may have changed. + num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps) + if overrode_max_train_steps: + args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch + # Afterwards we recalculate our number of training epochs + args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch) + + # We need to initialize the trackers we use, and also store our configuration. + # The trackers initializes automatically on the main process. + if accelerator.is_main_process: + tracker_config = dict(vars(args)) + keys_to_pop = [k for k, v in tracker_config.items() if isinstance(v, list)] + for k in keys_to_pop: + tracker_config.pop(k) + print(f"Removed tracker_config['{k}']") + accelerator.init_trackers(args.tracker_project_name, tracker_config) + + # Function for unwrapping if model was compiled with `torch.compile`. + def unwrap_model(model): + model = accelerator.unwrap_model(model) + model = model._orig_mod if is_compiled_module(model) else model + return model + + # Train! + total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps + + logger.info("***** Running training *****") + logger.info(f" Num examples = {len(train_dataset)}") + logger.info(f" Num Epochs = {args.num_train_epochs}") + logger.info(f" Instantaneous batch size per device = {args.train_batch_size}") + logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}") + logger.info(f" Gradient Accumulation steps = {args.gradient_accumulation_steps}") + logger.info(f" Total optimization steps = {args.max_train_steps}") + global_step = 0 + first_epoch = 0 + + # Potentially load in the weights and states from a previous save + if args.resume_from_checkpoint: + if args.resume_from_checkpoint != "latest": + path = os.path.basename(args.resume_from_checkpoint) + else: + # Get the most recent checkpoint + dirs = os.listdir(args.output_dir) + dirs = [d for d in dirs if d.startswith("checkpoint")] + dirs = sorted(dirs, key=lambda x: int(x.split("-")[1])) + path = dirs[-1] if len(dirs) > 0 else None + + if path is None: + accelerator.print( + f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run." + ) + args.resume_from_checkpoint = None + initial_global_step = 0 + else: + global_step = int(path.split("-")[1]) + + initial_global_step = global_step + + pkl_path = os.path.join(os.path.join(args.output_dir, path), "sampler_pos_start.pkl") + if os.path.exists(pkl_path): + with open(pkl_path, 'rb') as file: + _, first_epoch = pickle.load(file) + else: + first_epoch = global_step // num_update_steps_per_epoch + print(f"Load pkl from {pkl_path}. Get first_epoch = {first_epoch}.") + + accelerator.print(f"Resuming from checkpoint {path}") + accelerator.load_state(os.path.join(args.output_dir, path)) + else: + initial_global_step = 0 + + # function for saving/removing + def save_model(ckpt_file, unwrapped_nw): + os.makedirs(args.output_dir, exist_ok=True) + accelerator.print(f"\nsaving checkpoint: {ckpt_file}") + if isinstance(unwrapped_nw, dict): + from safetensors.torch import save_file + save_file(unwrapped_nw, ckpt_file, metadata={"format": "pt"}) + return ckpt_file + unwrapped_nw.save_weights(ckpt_file, weight_dtype, None) + + progress_bar = tqdm( + range(0, args.max_train_steps), + initial=initial_global_step, + desc="Steps", + # Only show the progress bar once on each machine. + disable=not accelerator.is_local_main_process, + ) + + if args.multi_stream: + # create extra cuda streams to speedup inpaint vae computation + vae_stream_1 = torch.cuda.Stream() + vae_stream_2 = torch.cuda.Stream() + else: + vae_stream_1 = None + vae_stream_2 = None + + # Calculate the index we need】 + idx_sampling = DiscreteSampling(args.train_sampling_steps, uniform_sampling=args.uniform_sampling) + + for epoch in range(first_epoch, args.num_train_epochs): + train_loss = 0.0 + batch_sampler.sampler.generator = torch.Generator().manual_seed(args.seed + epoch) + for step, batch in enumerate(train_dataloader): + # Data batch sanity check + if epoch == first_epoch and step == 0: + pixel_values, texts = batch['pixel_values'].cpu(), batch['text'] + pixel_values = rearrange(pixel_values, "b f c h w -> b c f h w") + os.makedirs(os.path.join(args.output_dir, "sanity_check"), exist_ok=True) + for idx, (pixel_value, text) in enumerate(zip(pixel_values, texts)): + pixel_value = pixel_value[None, ...] + gif_name = '-'.join(text.replace('/', '').split()[:10]) if not text == '' else f'{global_step}-{idx}' + save_videos_grid(pixel_value, f"{args.output_dir}/sanity_check/{gif_name[:10]}.gif", rescale=True) + + with accelerator.accumulate(transformer3d): + # Convert images to latent space + pixel_values = batch["pixel_values"].to(weight_dtype) + + if args.low_vram: + torch.cuda.empty_cache() + vae.to(accelerator.device) + if not args.enable_text_encoder_in_dataloader: + text_encoder.to("cpu") + + with torch.no_grad(): + # This way is quicker when batch grows up + def _batch_encode_vae(pixel_values): + pixel_values = pixel_values.squeeze(1) + bs = args.vae_mini_batch + new_pixel_values = [] + for i in range(0, pixel_values.shape[0], bs): + pixel_values_bs = pixel_values[i : i + bs] + pixel_values_bs = vae.encode(pixel_values_bs)[0] + pixel_values_bs = pixel_values_bs.sample() + new_pixel_values.append(pixel_values_bs) + return torch.cat(new_pixel_values, dim = 0) + if vae_stream_1 is not None: + vae_stream_1.wait_stream(torch.cuda.current_stream()) + with torch.cuda.stream(vae_stream_1): + latents = _batch_encode_vae(pixel_values) + else: + latents = _batch_encode_vae(pixel_values) + + # wait for latents = vae.encode(pixel_values) to complete + if vae_stream_1 is not None: + torch.cuda.current_stream().wait_stream(vae_stream_1) + + if args.low_vram: + vae.to('cpu') + torch.cuda.empty_cache() + if not args.enable_text_encoder_in_dataloader: + text_encoder.to(accelerator.device) + + if args.enable_text_encoder_in_dataloader: + prompt_embeds = batch['prompt_embeds'].to(dtype=latents.dtype, device=accelerator.device) + text_ids = batch['text_ids'] + else: + with torch.no_grad(): + prompt_embeds, text_ids = encode_prompt( + batch['text'], device=accelerator.device, + text_encoder=text_encoder, + tokenizer=tokenizer, + ) + + if args.low_vram and not args.enable_text_encoder_in_dataloader: + text_encoder.to('cpu') + torch.cuda.empty_cache() + + bsz, channel, height, width = latents.size() + latents = _patchify_latents(latents) + latent_image_ids = _prepare_latent_ids(latents) + latents = ((latents - latents_bn_mean) / latents_bn_std).to(dtype=weight_dtype) + latents = _pack_latents(latents) + + noise = torch.randn(latents.size(), device=latents.device, generator=torch_rng, dtype=weight_dtype) + # handle guidance + guidance = torch.tensor([args.guidance_scale], device=accelerator.device) + guidance = guidance.expand(latents.shape[0]) + + if not args.uniform_sampling: + u = compute_density_for_timestep_sampling( + weighting_scheme=args.weighting_scheme, + batch_size=bsz, + logit_mean=args.logit_mean, + logit_std=args.logit_std, + mode_scale=args.mode_scale, + ) + indices = (u * noise_scheduler.config.num_train_timesteps).long() + else: + # Sample a random timestep for each image + # timesteps = generate_timestep_with_lognorm(0, args.train_sampling_steps, (bsz,), device=latents.device, generator=torch_rng) + # timesteps = torch.randint(0, args.train_sampling_steps, (bsz,), device=latents.device, generator=torch_rng) + indices = idx_sampling(bsz, generator=torch_rng, device=latents.device) + indices = indices.long().cpu() + + sigmas = np.linspace(1.0, 1 / args.train_sampling_steps, args.train_sampling_steps) + image_seq_len = latents.shape[1] + mu = calculate_shift( + image_seq_len, + noise_scheduler.config.get("base_image_seq_len", 256), + noise_scheduler.config.get("max_image_seq_len", 4096), + noise_scheduler.config.get("base_shift", 0.5), + noise_scheduler.config.get("max_shift", 1.15), + ) + noise_scheduler.set_timesteps(sigmas=sigmas, device=latents.device, mu=mu) + timesteps = noise_scheduler.timesteps[indices].to(device=latents.device) + + def get_sigmas(timesteps, n_dim=4, dtype=torch.float32): + sigmas = noise_scheduler.sigmas.to(device=accelerator.device, dtype=dtype) + schedule_timesteps = noise_scheduler.timesteps.to(accelerator.device) + timesteps = timesteps.to(accelerator.device) + step_indices = [(schedule_timesteps == t).nonzero().item() for t in timesteps] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < n_dim: + sigma = sigma.unsqueeze(-1) + return sigma + + # Add noise according to flow matching. + # zt = (1 - texp) * x + texp * z1 + sigmas = get_sigmas(timesteps, n_dim=latents.ndim, dtype=latents.dtype) + noisy_latents = (1.0 - sigmas) * latents + sigmas * noise + + # Add noise + target = noise - latents + + # Predict the noise residual + with torch.cuda.amp.autocast(dtype=weight_dtype), torch.cuda.device(device=accelerator.device): + noise_pred = transformer3d( + hidden_states=noisy_latents, + timestep=timesteps / 1000, + guidance=guidance, + encoder_hidden_states=prompt_embeds, + txt_ids=text_ids, + img_ids=latent_image_ids, + return_dict=False, + )[0] + + def custom_mse_loss(noise_pred, target, weighting=None, threshold=50): + noise_pred = noise_pred.float() + target = target.float() + diff = noise_pred - target + mse_loss = F.mse_loss(noise_pred, target, reduction='none') + mask = (diff.abs() <= threshold).float() + masked_loss = mse_loss * mask + if weighting is not None: + masked_loss = masked_loss * weighting + final_loss = masked_loss.mean() + return final_loss + + weighting = compute_loss_weighting_for_sd3(weighting_scheme=args.weighting_scheme, sigmas=sigmas) + loss = custom_mse_loss(noise_pred.float(), target.float(), weighting.float()) + loss = loss.mean() + + # Gather the losses across all processes for logging (if we use distributed training). + avg_loss = accelerator.gather(loss.repeat(args.train_batch_size)).mean() + train_loss += avg_loss.item() / args.gradient_accumulation_steps + + # Backpropagate + accelerator.backward(loss) + if accelerator.sync_gradients: + accelerator.clip_grad_norm_(trainable_params, args.max_grad_norm) + optimizer.step() + lr_scheduler.step() + optimizer.zero_grad() + + # Checks if the accelerator has performed an optimization step behind the scenes + if accelerator.sync_gradients: + progress_bar.update(1) + global_step += 1 + accelerator.log({"train_loss": train_loss}, step=global_step) + train_loss = 0.0 + + if global_step % args.checkpointing_steps == 0: + if args.use_deepspeed or args.use_fsdp or accelerator.is_main_process: + # _before_ saving state, check if this save would set us over the `checkpoints_total_limit` + if args.checkpoints_total_limit is not None: + checkpoints = os.listdir(args.output_dir) + checkpoints = [d for d in checkpoints if d.startswith("checkpoint")] + checkpoints = sorted(checkpoints, key=lambda x: int(x.split("-")[1])) + + # before we save the new checkpoint, we need to have at _most_ `checkpoints_total_limit - 1` checkpoints + if len(checkpoints) >= args.checkpoints_total_limit: + num_to_remove = len(checkpoints) - args.checkpoints_total_limit + 1 + removing_checkpoints = checkpoints[0:num_to_remove] + + logger.info( + f"{len(checkpoints)} checkpoints already exist, removing {len(removing_checkpoints)} checkpoints" + ) + logger.info(f"removing checkpoints: {', '.join(removing_checkpoints)}") + + for removing_checkpoint in removing_checkpoints: + removing_checkpoint = os.path.join(args.output_dir, removing_checkpoint) + shutil.rmtree(removing_checkpoint) + gc.collect() + torch.cuda.empty_cache() + torch.cuda.ipc_collect() + if not args.save_state: + if args.use_peft_lora: + safetensor_save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}.safetensors") + save_model(safetensor_save_path, get_peft_model_state_dict(accelerator.unwrap_model(transformer3d))) + logger.info(f"Saved safetensor to {safetensor_save_path}") + else: + safetensor_save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}.safetensors") + save_model(safetensor_save_path, accelerator.unwrap_model(network)) + logger.info(f"Saved safetensor to {safetensor_save_path}") + else: + accelerator_save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}") + accelerator.save_state(accelerator_save_path) + logger.info(f"Saved state to {accelerator_save_path}") + + if accelerator.is_main_process: + if args.validation_prompts is not None and global_step % args.validation_steps == 0: + log_validation( + vae, + text_encoder, + tokenizer, + transformer3d, + network, + args, + accelerator, + weight_dtype, + global_step, + ) + + logs = {"step_loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]} + progress_bar.set_postfix(**logs) + + if global_step >= args.max_train_steps: + break + + if accelerator.is_main_process: + if args.validation_prompts is not None and epoch % args.validation_epochs == 0: + log_validation( + vae, + text_encoder, + tokenizer, + tokenizer_2, + transformer3d, + network, + args, + accelerator, + weight_dtype, + global_step, + ) + + # Create the pipeline using the trained modules and save it. + accelerator.wait_for_everyone() + if args.use_deepspeed or args.use_fsdp or accelerator.is_main_process: + gc.collect() + torch.cuda.empty_cache() + torch.cuda.ipc_collect() + if not args.save_state: + if args.use_peft_lora: + safetensor_save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}.safetensors") + save_model(safetensor_save_path, get_peft_model_state_dict(accelerator.unwrap_model(transformer3d))) + logger.info(f"Saved safetensor to {safetensor_save_path}") + else: + safetensor_save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}.safetensors") + save_model(safetensor_save_path, accelerator.unwrap_model(network)) + logger.info(f"Saved safetensor to {safetensor_save_path}") + else: + accelerator_save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}") + accelerator.save_state(accelerator_save_path) + logger.info(f"Saved state to {accelerator_save_path}") + + accelerator.end_training() + + +if __name__ == "__main__": + main() diff --git a/scripts/flux2/train_lora.sh b/scripts/flux2/train_lora.sh new file mode 100644 index 0000000..5b1078d --- /dev/null +++ b/scripts/flux2/train_lora.sh @@ -0,0 +1,33 @@ +export MODEL_NAME="models/Diffusion_Transformer/FLUX.2-dev" +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/flux2/train_lora.py \ + --pretrained_model_name_or_path=$MODEL_NAME \ + --train_data_dir=$DATASET_NAME \ + --train_data_meta=$DATASET_META_NAME \ + --train_batch_size=1 \ + --image_sample_size=1328 \ + --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_lora" \ + --gradient_checkpointing \ + --mixed_precision="bf16" \ + --adam_weight_decay=3e-2 \ + --adam_epsilon=1e-10 \ + --vae_mini_batch=1 \ + --max_grad_norm=0.05 \ + --enable_bucket \ + --rank=64 \ + --network_alpha=32 \ + --target_name="to_q,to_k,to_v,ff.0,ff.2,ff_context.0,ff_context.2" \ + --use_peft_lora \ + --uniform_sampling \ No newline at end of file diff --git a/scripts/hunyuanvideo/train_lora.py b/scripts/hunyuanvideo/train_lora.py index 0ad4c2c..d1fa84c 100644 --- a/scripts/hunyuanvideo/train_lora.py +++ b/scripts/hunyuanvideo/train_lora.py @@ -1423,7 +1423,7 @@ def main(): from functools import partial from videox_fun.dist import set_multi_gpus_devices, shard_model - shard_fn = partial(shard_model, device_id=accelerator.device, param_dtype=weight_dtype, module_to_wrapper=transformer3d.transformer_blocks) + shard_fn = partial(shard_model, device_id=accelerator.device, param_dtype=weight_dtype, module_to_wrapper=list(transformer3d.transformer_blocks) + list(transformer3d.single_transformer_blocks)) transformer3d = shard_fn(transformer3d) if fsdp_stage != 0 or zero_stage != 0: diff --git a/videox_fun/dist/__init__.py b/videox_fun/dist/__init__.py index 313f0ef..d019f5f 100755 --- a/videox_fun/dist/__init__.py +++ b/videox_fun/dist/__init__.py @@ -1,6 +1,7 @@ import importlib.util from .cogvideox_xfuser import CogVideoXMultiGPUsAttnProcessor2_0 +from .flux2_xfuser import Flux2MultiGPUsAttnProcessor2_0 from .flux_xfuser import FluxMultiGPUsAttnProcessor2_0 from .fsdp import shard_model from .fuser import (get_sequence_parallel_rank, diff --git a/videox_fun/dist/flux2_xfuser.py b/videox_fun/dist/flux2_xfuser.py new file mode 100644 index 0000000..bfce2b2 --- /dev/null +++ b/videox_fun/dist/flux2_xfuser.py @@ -0,0 +1,194 @@ +from typing import Optional, Tuple, Union + +import torch +import torch.nn.functional as F +from diffusers.models.attention_processor import Attention + +from .fuser import xFuserLongContextAttention + + +def _get_projections(attn: "FluxAttention", hidden_states, encoder_hidden_states=None): + query = attn.to_q(hidden_states) + key = attn.to_k(hidden_states) + value = attn.to_v(hidden_states) + + encoder_query = encoder_key = encoder_value = None + if encoder_hidden_states is not None and attn.added_kv_proj_dim is not None: + encoder_query = attn.add_q_proj(encoder_hidden_states) + encoder_key = attn.add_k_proj(encoder_hidden_states) + encoder_value = attn.add_v_proj(encoder_hidden_states) + + return query, key, value, encoder_query, encoder_key, encoder_value + + +def _get_qkv_projections(attn: "FluxAttention", hidden_states, encoder_hidden_states=None): + return _get_projections(attn, hidden_states, encoder_hidden_states) + + +def apply_rotary_emb( + x: torch.Tensor, + freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]], + use_real: bool = True, + use_real_unbind_dim: int = -1, + sequence_dim: int = 2, +) -> Tuple[torch.Tensor, torch.Tensor]: + """ + Apply rotary embeddings to input tensors using the given frequency tensor. This function applies rotary embeddings + to the given query or key 'x' tensors using the provided frequency tensor 'freqs_cis'. The input tensors are + reshaped as complex numbers, and the frequency tensor is reshaped for broadcasting compatibility. The resulting + tensors contain rotary embeddings and are returned as real tensors. + + Args: + x (`torch.Tensor`): + Query or key tensor to apply rotary embeddings. [B, H, S, D] xk (torch.Tensor): Key tensor to apply + freqs_cis (`Tuple[torch.Tensor]`): Precomputed frequency tensor for complex exponentials. ([S, D], [S, D],) + + Returns: + Tuple[torch.Tensor, torch.Tensor]: Tuple of modified query tensor and key tensor with rotary embeddings. + """ + if use_real: + cos, sin = freqs_cis # [S, D] + if sequence_dim == 2: + cos = cos[None, None, :, :] + sin = sin[None, None, :, :] + elif sequence_dim == 1: + cos = cos[None, :, None, :] + sin = sin[None, :, None, :] + else: + raise ValueError(f"`sequence_dim={sequence_dim}` but should be 1 or 2.") + + cos, sin = cos.to(x.device), sin.to(x.device) + + if use_real_unbind_dim == -1: + # Used for flux, cogvideox, hunyuan-dit + x_real, x_imag = x.reshape(*x.shape[:-1], -1, 2).unbind(-1) # [B, H, S, D//2] + x_rotated = torch.stack([-x_imag, x_real], dim=-1).flatten(3) + elif use_real_unbind_dim == -2: + # Used for Stable Audio, OmniGen, CogView4 and Cosmos + x_real, x_imag = x.reshape(*x.shape[:-1], 2, -1).unbind(-2) # [B, H, S, D//2] + x_rotated = torch.cat([-x_imag, x_real], dim=-1) + else: + raise ValueError(f"`use_real_unbind_dim={use_real_unbind_dim}` but should be -1 or -2.") + + out = (x.float() * cos + x_rotated.float() * sin).to(x.dtype) + + return out + else: + # used for lumina + x_rotated = torch.view_as_complex(x.float().reshape(*x.shape[:-1], -1, 2)) + freqs_cis = freqs_cis.unsqueeze(2) + x_out = torch.view_as_real(x_rotated * freqs_cis).flatten(3) + + return x_out.type_as(x) + + +class Flux2MultiGPUsAttnProcessor2_0: + r""" + Processor for implementing scaled dot-product attention for the CogVideoX model. It applies a rotary embedding on + query and key vectors, but does not include spatial normalization. + """ + + def __init__(self): + if not hasattr(F, "scaled_dot_product_attention"): + raise ImportError("Flux2MultiGPUsAttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.") + + def __call__( + self, + attn: "FluxAttention", + hidden_states: torch.Tensor, + encoder_hidden_states: Optional[torch.Tensor] = None, + attention_mask: Optional[torch.Tensor] = None, + image_rotary_emb: Optional[torch.Tensor] = None, + text_seq_len: int = None, + ) -> torch.FloatTensor: + # Determine which type of attention we're processing + is_parallel_self_attn = hasattr(attn, 'to_qkv_mlp_proj') and attn.to_qkv_mlp_proj is not None + + if is_parallel_self_attn: + # Parallel in (QKV + MLP in) projection + hidden_states = attn.to_qkv_mlp_proj(hidden_states) + qkv, mlp_hidden_states = torch.split( + hidden_states, [3 * attn.inner_dim, attn.mlp_hidden_dim * attn.mlp_mult_factor], dim=-1 + ) + + # Handle the attention logic + query, key, value = qkv.chunk(3, dim=-1) + + else: + query, key, value, encoder_query, encoder_key, encoder_value = _get_qkv_projections( + attn, hidden_states, encoder_hidden_states + ) + + # Common processing for query, key, value + query = query.unflatten(-1, (attn.heads, -1)) + key = key.unflatten(-1, (attn.heads, -1)) + value = value.unflatten(-1, (attn.heads, -1)) + + query = attn.norm_q(query) + key = attn.norm_k(key) + + # Handle encoder projections (only for standard attention) + if not is_parallel_self_attn and attn.added_kv_proj_dim is not None: + encoder_query = encoder_query.unflatten(-1, (attn.heads, -1)) + encoder_key = encoder_key.unflatten(-1, (attn.heads, -1)) + encoder_value = encoder_value.unflatten(-1, (attn.heads, -1)) + + encoder_query = attn.norm_added_q(encoder_query) + encoder_key = attn.norm_added_k(encoder_key) + + query = torch.cat([encoder_query, query], dim=1) + key = torch.cat([encoder_key, key], dim=1) + value = torch.cat([encoder_value, value], dim=1) + + # Apply rotary embeddings + if image_rotary_emb is not None: + query = apply_rotary_emb(query, image_rotary_emb, sequence_dim=1) + key = apply_rotary_emb(key, image_rotary_emb, sequence_dim=1) + + if not is_parallel_self_attn and attn.added_kv_proj_dim is not None and text_seq_len is None: + text_seq_len = encoder_query.shape[1] + + txt_query, txt_key, txt_value = query[:, :text_seq_len], key[:, :text_seq_len], value[:, :text_seq_len] + img_query, img_key, img_value = query[:, text_seq_len:], key[:, text_seq_len:], value[:, text_seq_len:] + + half_dtypes = (torch.float16, torch.bfloat16) + def half(x): + return x if x.dtype in half_dtypes else x.to(torch.bfloat16) + + hidden_states = xFuserLongContextAttention()( + None, + half(img_query), half(img_key), half(img_value), dropout_p=0.0, causal=False, + joint_tensor_query=half(txt_query) if txt_query is not None else None, + joint_tensor_key=half(txt_key) if txt_key is not None else None, + joint_tensor_value=half(txt_value) if txt_value is not None else None, + joint_strategy='front', + ) + hidden_states = hidden_states.flatten(2, 3) + hidden_states = hidden_states.to(query.dtype) + + if is_parallel_self_attn: + # Handle the feedforward (FF) logic + mlp_hidden_states = attn.mlp_act_fn(mlp_hidden_states) + + # Concatenate and parallel output projection + hidden_states = torch.cat([hidden_states, mlp_hidden_states], dim=-1) + hidden_states = attn.to_out(hidden_states) + + return hidden_states + + else: + # Split encoder and latent hidden states if encoder was used + if encoder_hidden_states is not None: + encoder_hidden_states, hidden_states = hidden_states.split_with_sizes( + [encoder_hidden_states.shape[1], hidden_states.shape[1] - encoder_hidden_states.shape[1]], dim=1 + ) + encoder_hidden_states = attn.to_add_out(encoder_hidden_states) + + # Project output + hidden_states = attn.to_out[0](hidden_states) + hidden_states = attn.to_out[1](hidden_states) + + if encoder_hidden_states is not None: + return hidden_states, encoder_hidden_states + else: + return hidden_states \ No newline at end of file diff --git a/videox_fun/dist/flux_xfuser.py b/videox_fun/dist/flux_xfuser.py index f08ba93..6fb91f4 100644 --- a/videox_fun/dist/flux_xfuser.py +++ b/videox_fun/dist/flux_xfuser.py @@ -124,23 +124,20 @@ class FluxMultiGPUsAttnProcessor2_0: key = torch.cat([encoder_key, key], dim=1) value = torch.cat([encoder_value, value], dim=1) - if image_rotary_emb is not None: - query = apply_rotary_emb(query, image_rotary_emb, sequence_dim=1) - key = apply_rotary_emb(key, image_rotary_emb, sequence_dim=1) + # Apply rotary embeddings + if image_rotary_emb is not None: + query = apply_rotary_emb(query, image_rotary_emb, sequence_dim=1) + key = apply_rotary_emb(key, image_rotary_emb, sequence_dim=1) + if attn.added_kv_proj_dim is not None and text_seq_len is None: text_seq_len = encoder_query.shape[1] - txt_query, txt_key, txt_value = query[:, :text_seq_len], key[:, :text_seq_len], value[:, :text_seq_len] - img_query, img_key, img_value = query[:, text_seq_len:], key[:, text_seq_len:], value[:, text_seq_len:] - else: - if image_rotary_emb is not None: - query = apply_rotary_emb(query, image_rotary_emb, sequence_dim=1) - key = apply_rotary_emb(key, image_rotary_emb, sequence_dim=1) - txt_query, txt_key, txt_value = query[:, :text_seq_len], key[:, :text_seq_len], value[:, :text_seq_len] - img_query, img_key, img_value = query[:, text_seq_len:], key[:, text_seq_len:], value[:, text_seq_len:] + + txt_query, txt_key, txt_value = query[:, :text_seq_len], key[:, :text_seq_len], value[:, :text_seq_len] + img_query, img_key, img_value = query[:, text_seq_len:], key[:, text_seq_len:], value[:, text_seq_len:] half_dtypes = (torch.float16, torch.bfloat16) def half(x): - return x if x.dtype in half_dtypes else x.to(dtype) + return x if x.dtype in half_dtypes else x.to(torch.bfloat16) hidden_states = xFuserLongContextAttention()( None, diff --git a/videox_fun/models/__init__.py b/videox_fun/models/__init__.py index 4f5f262..3fb64cc 100755 --- a/videox_fun/models/__init__.py +++ b/videox_fun/models/__init__.py @@ -1,11 +1,12 @@ import importlib.util from diffusers import AutoencoderKL -from transformers import (AutoTokenizer, CLIPImageProcessor, CLIPTextModel, - CLIPTokenizer, CLIPVisionModelWithProjection, - LlamaModel, LlamaTokenizerFast, - LlavaForConditionalGeneration, T5EncoderModel, - T5Tokenizer, T5TokenizerFast) +from transformers import (AutoProcessor, AutoTokenizer, CLIPImageProcessor, + CLIPTextModel, CLIPTokenizer, + CLIPVisionModelWithProjection, LlamaModel, + LlamaTokenizerFast, LlavaForConditionalGeneration, + Mistral3ForConditionalGeneration, PixtralProcessor, + T5EncoderModel, T5Tokenizer, T5TokenizerFast) try: from transformers import (Qwen2_5_VLConfig, @@ -20,6 +21,9 @@ from .cogvideox_transformer3d import CogVideoXTransformer3DModel from .cogvideox_vae import AutoencoderKLCogVideoX from .fantasytalking_audio_encoder import FantasyTalkingAudioEncoder from .fantasytalking_transformer3d import FantasyTalkingTransformer3DModel +from .flux2_image_processor import Flux2ImageProcessor +from .flux2_transformer2d import Flux2Transformer2DModel +from .flux2_vae import AutoencoderKLFlux2 from .flux_transformer2d import FluxTransformer2DModel from .hunyuanvideo_transformer3d import HunyuanVideoTransformer3DModel from .hunyuanvideo_vae import AutoencoderKLHunyuanVideo diff --git a/videox_fun/models/flux2_image_processor.py b/videox_fun/models/flux2_image_processor.py new file mode 100644 index 0000000..d2edc9e --- /dev/null +++ b/videox_fun/models/flux2_image_processor.py @@ -0,0 +1,139 @@ +# Modified from https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/flux2/image_processor.py +# Copyright 2025 The Black Forest Labs Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import math +from typing import Tuple + +import PIL.Image + +from diffusers.configuration_utils import register_to_config +from diffusers.image_processor import VaeImageProcessor + + +class Flux2ImageProcessor(VaeImageProcessor): + r""" + Image processor to preprocess the reference (character) image for the Flux2 model. + + Args: + do_resize (`bool`, *optional*, defaults to `True`): + Whether to downscale the image's (height, width) dimensions to multiples of `vae_scale_factor`. Can accept + `height` and `width` arguments from [`image_processor.VaeImageProcessor.preprocess`] method. + vae_scale_factor (`int`, *optional*, defaults to `16`): + VAE (spatial) scale factor. If `do_resize` is `True`, the image is automatically resized to multiples of + this factor. + vae_latent_channels (`int`, *optional*, defaults to `32`): + VAE latent channels. + do_normalize (`bool`, *optional*, defaults to `True`): + Whether to normalize the image to [-1,1]. + do_convert_rgb (`bool`, *optional*, defaults to be `True`): + Whether to convert the images to RGB format. + """ + + @register_to_config + def __init__( + self, + do_resize: bool = True, + vae_scale_factor: int = 16, + vae_latent_channels: int = 32, + do_normalize: bool = True, + do_convert_rgb: bool = True, + ): + super().__init__( + do_resize=do_resize, + vae_scale_factor=vae_scale_factor, + vae_latent_channels=vae_latent_channels, + do_normalize=do_normalize, + do_convert_rgb=do_convert_rgb, + ) + + @staticmethod + def check_image_input( + image: PIL.Image.Image, max_aspect_ratio: int = 8, min_side_length: int = 64, max_area: int = 1024 * 1024 + ) -> PIL.Image.Image: + """ + Check if image meets minimum size and aspect ratio requirements. + + Args: + image: PIL Image to validate + max_aspect_ratio: Maximum allowed aspect ratio (width/height or height/width) + min_side_length: Minimum pixels required for width and height + max_area: Maximum allowed area in pixels² + + Returns: + The input image if valid + + Raises: + ValueError: If image is too small or aspect ratio is too extreme + """ + if not isinstance(image, PIL.Image.Image): + raise ValueError(f"Image must be a PIL.Image.Image, got {type(image)}") + + width, height = image.size + + # Check minimum dimensions + if width < min_side_length or height < min_side_length: + raise ValueError( + f"Image too small: {width}×{height}. Both dimensions must be at least {min_side_length}px" + ) + + # Check aspect ratio + aspect_ratio = max(width / height, height / width) + if aspect_ratio > max_aspect_ratio: + raise ValueError( + f"Aspect ratio too extreme: {width}×{height} (ratio: {aspect_ratio:.1f}:1). " + f"Maximum allowed ratio is {max_aspect_ratio}:1" + ) + + return image + + @staticmethod + def _resize_to_target_area(image: PIL.Image.Image, target_area: int = 1024 * 1024) -> Tuple[int, int]: + image_width, image_height = image.size + + scale = math.sqrt(target_area / (image_width * image_height)) + width = int(image_width * scale) + height = int(image_height * scale) + + return image.resize((width, height), PIL.Image.Resampling.LANCZOS) + + def _resize_and_crop( + self, + image: PIL.Image.Image, + width: int, + height: int, + ) -> PIL.Image.Image: + r""" + center crop the image to the specified width and height. + + Args: + image (`PIL.Image.Image`): + The image to resize and crop. + width (`int`): + The width to resize the image to. + height (`int`): + The height to resize the image to. + + Returns: + `PIL.Image.Image`: + The resized and cropped image. + """ + image_width, image_height = image.size + + left = (image_width - width) // 2 + top = (image_height - height) // 2 + right = left + width + bottom = top + height + + return image.crop((left, top, right, bottom)) \ No newline at end of file diff --git a/videox_fun/models/flux2_transformer2d.py b/videox_fun/models/flux2_transformer2d.py new file mode 100644 index 0000000..0fdecb6 --- /dev/null +++ b/videox_fun/models/flux2_transformer2d.py @@ -0,0 +1,1278 @@ +# Modified from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/transformers/transformer_flux2.py +# Copyright 2025 Black Forest Labs, The HuggingFace Team and The InstantX Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import glob +import inspect +import json +import os +from typing import Any, Dict, List, Optional, Tuple, Union + +import torch +import torch.nn as nn +import torch.nn.functional as F +from diffusers.configuration_utils import ConfigMixin, register_to_config +from diffusers.loaders import FromOriginalModelMixin +from diffusers.models.attention_processor import Attention, AttentionProcessor +from diffusers.models.embeddings import (TimestepEmbedding, Timesteps, + apply_rotary_emb, + get_1d_rotary_pos_embed) +from diffusers.models.modeling_outputs import Transformer2DModelOutput +from diffusers.models.modeling_utils import ModelMixin +from diffusers.models.normalization import AdaLayerNormContinuous +from diffusers.utils import (USE_PEFT_BACKEND, is_torch_npu_available, + is_torch_version, logging, scale_lora_layers, + unscale_lora_layers) + +from ..dist import (Flux2MultiGPUsAttnProcessor2_0, get_sequence_parallel_rank, + get_sequence_parallel_world_size, get_sp_group) +from .attention_utils import attention + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +def _get_projections(attn: "Flux2Attention", hidden_states, encoder_hidden_states=None): + query = attn.to_q(hidden_states) + key = attn.to_k(hidden_states) + value = attn.to_v(hidden_states) + + encoder_query = encoder_key = encoder_value = None + if encoder_hidden_states is not None and attn.added_kv_proj_dim is not None: + encoder_query = attn.add_q_proj(encoder_hidden_states) + encoder_key = attn.add_k_proj(encoder_hidden_states) + encoder_value = attn.add_v_proj(encoder_hidden_states) + + return query, key, value, encoder_query, encoder_key, encoder_value + + +def _get_qkv_projections(attn: "Flux2Attention", hidden_states, encoder_hidden_states=None): + return _get_projections(attn, hidden_states, encoder_hidden_states) + + +def apply_rotary_emb( + x: torch.Tensor, + freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]], + use_real: bool = True, + use_real_unbind_dim: int = -1, + sequence_dim: int = 2, +) -> Tuple[torch.Tensor, torch.Tensor]: + """ + Apply rotary embeddings to input tensors using the given frequency tensor. This function applies rotary embeddings + to the given query or key 'x' tensors using the provided frequency tensor 'freqs_cis'. The input tensors are + reshaped as complex numbers, and the frequency tensor is reshaped for broadcasting compatibility. The resulting + tensors contain rotary embeddings and are returned as real tensors. + + Args: + x (`torch.Tensor`): + Query or key tensor to apply rotary embeddings. [B, H, S, D] xk (torch.Tensor): Key tensor to apply + freqs_cis (`Tuple[torch.Tensor]`): Precomputed frequency tensor for complex exponentials. ([S, D], [S, D],) + + Returns: + Tuple[torch.Tensor, torch.Tensor]: Tuple of modified query tensor and key tensor with rotary embeddings. + """ + if use_real: + cos, sin = freqs_cis # [S, D] + if sequence_dim == 2: + cos = cos[None, None, :, :] + sin = sin[None, None, :, :] + elif sequence_dim == 1: + cos = cos[None, :, None, :] + sin = sin[None, :, None, :] + else: + raise ValueError(f"`sequence_dim={sequence_dim}` but should be 1 or 2.") + + cos, sin = cos.to(x.device), sin.to(x.device) + + if use_real_unbind_dim == -1: + # Used for flux, cogvideox, hunyuan-dit + x_real, x_imag = x.reshape(*x.shape[:-1], -1, 2).unbind(-1) # [B, H, S, D//2] + x_rotated = torch.stack([-x_imag, x_real], dim=-1).flatten(3) + elif use_real_unbind_dim == -2: + # Used for Stable Audio, OmniGen, CogView4 and Cosmos + x_real, x_imag = x.reshape(*x.shape[:-1], 2, -1).unbind(-2) # [B, H, S, D//2] + x_rotated = torch.cat([-x_imag, x_real], dim=-1) + else: + raise ValueError(f"`use_real_unbind_dim={use_real_unbind_dim}` but should be -1 or -2.") + + out = (x.float() * cos + x_rotated.float() * sin).to(x.dtype) + + return out + else: + # used for lumina + x_rotated = torch.view_as_complex(x.float().reshape(*x.shape[:-1], -1, 2)) + freqs_cis = freqs_cis.unsqueeze(2) + x_out = torch.view_as_real(x_rotated * freqs_cis).flatten(3) + + return x_out.type_as(x) + + +class Flux2SwiGLU(nn.Module): + """ + Flux 2 uses a SwiGLU-style activation in the transformer feedforward sub-blocks, but with the linear projection + layer fused into the first linear layer of the FF sub-block. Thus, this module has no trainable parameters. + """ + + def __init__(self): + super().__init__() + self.gate_fn = nn.SiLU() + + def forward(self, x: torch.Tensor) -> torch.Tensor: + x1, x2 = x.chunk(2, dim=-1) + x = self.gate_fn(x1) * x2 + return x + + +class Flux2FeedForward(nn.Module): + def __init__( + self, + dim: int, + dim_out: Optional[int] = None, + mult: float = 3.0, + inner_dim: Optional[int] = None, + bias: bool = False, + ): + super().__init__() + if inner_dim is None: + inner_dim = int(dim * mult) + dim_out = dim_out or dim + + # Flux2SwiGLU will reduce the dimension by half + self.linear_in = nn.Linear(dim, inner_dim * 2, bias=bias) + self.act_fn = Flux2SwiGLU() + self.linear_out = nn.Linear(inner_dim, dim_out, bias=bias) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + x = self.linear_in(x) + x = self.act_fn(x) + x = self.linear_out(x) + return x + + +class Flux2AttnProcessor: + _attention_backend = None + _parallel_config = None + + def __init__(self): + if not hasattr(F, "scaled_dot_product_attention"): + raise ImportError(f"{self.__class__.__name__} requires PyTorch 2.0. Please upgrade your pytorch version.") + + def __call__( + self, + attn: Union["Flux2Attention", "Flux2ParallelSelfAttention"], + hidden_states: torch.Tensor, + encoder_hidden_states: Optional[torch.Tensor] = None, + attention_mask: Optional[torch.Tensor] = None, + image_rotary_emb: Optional[torch.Tensor] = None, + text_seq_len: int = None, + ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]: + """ + Unified processor for both Flux2Attention and Flux2ParallelSelfAttention. + + Args: + attn: Attention module (either Flux2Attention or Flux2ParallelSelfAttention) + hidden_states: Input hidden states + encoder_hidden_states: Optional encoder hidden states (only for Flux2Attention) + attention_mask: Optional attention mask + image_rotary_emb: Optional rotary embeddings + + Returns: + For Flux2Attention with encoder_hidden_states: (hidden_states, encoder_hidden_states) + For Flux2Attention without encoder_hidden_states: hidden_states + For Flux2ParallelSelfAttention: hidden_states + """ + # Determine which type of attention we're processing + is_parallel_self_attn = hasattr(attn, 'to_qkv_mlp_proj') and attn.to_qkv_mlp_proj is not None + + if is_parallel_self_attn: + # ============================================ + # Parallel Self-Attention Path (with MLP) + # ============================================ + # Parallel in (QKV + MLP in) projection + hidden_states = attn.to_qkv_mlp_proj(hidden_states) + qkv, mlp_hidden_states = torch.split( + hidden_states, [3 * attn.inner_dim, attn.mlp_hidden_dim * attn.mlp_mult_factor], dim=-1 + ) + + # Handle the attention logic + query, key, value = qkv.chunk(3, dim=-1) + + else: + # ============================================ + # Standard Attention Path (possibly with encoder) + # ============================================ + query, key, value, encoder_query, encoder_key, encoder_value = _get_qkv_projections( + attn, hidden_states, encoder_hidden_states + ) + + # Common processing for query, key, value + query = query.unflatten(-1, (attn.heads, -1)) + key = key.unflatten(-1, (attn.heads, -1)) + value = value.unflatten(-1, (attn.heads, -1)) + + query = attn.norm_q(query) + key = attn.norm_k(key) + + # Handle encoder projections (only for standard attention) + if not is_parallel_self_attn and attn.added_kv_proj_dim is not None: + encoder_query = encoder_query.unflatten(-1, (attn.heads, -1)) + encoder_key = encoder_key.unflatten(-1, (attn.heads, -1)) + encoder_value = encoder_value.unflatten(-1, (attn.heads, -1)) + + encoder_query = attn.norm_added_q(encoder_query) + encoder_key = attn.norm_added_k(encoder_key) + + query = torch.cat([encoder_query, query], dim=1) + key = torch.cat([encoder_key, key], dim=1) + value = torch.cat([encoder_value, value], dim=1) + + # Apply rotary embeddings + if image_rotary_emb is not None: + query = apply_rotary_emb(query, image_rotary_emb, sequence_dim=1) + key = apply_rotary_emb(key, image_rotary_emb, sequence_dim=1) + + # Perform attention + hidden_states = attention( + query, key, value, attn_mask=attention_mask, + ) + hidden_states = hidden_states.flatten(2, 3) + hidden_states = hidden_states.to(query.dtype) + + if is_parallel_self_attn: + # ============================================ + # Parallel Self-Attention Output Path + # ============================================ + # Handle the feedforward (FF) logic + mlp_hidden_states = attn.mlp_act_fn(mlp_hidden_states) + + # Concatenate and parallel output projection + hidden_states = torch.cat([hidden_states, mlp_hidden_states], dim=-1) + hidden_states = attn.to_out(hidden_states) + + return hidden_states + + else: + # ============================================ + # Standard Attention Output Path + # ============================================ + # Split encoder and latent hidden states if encoder was used + if encoder_hidden_states is not None: + encoder_hidden_states, hidden_states = hidden_states.split_with_sizes( + [encoder_hidden_states.shape[1], hidden_states.shape[1] - encoder_hidden_states.shape[1]], dim=1 + ) + encoder_hidden_states = attn.to_add_out(encoder_hidden_states) + + # Project output + hidden_states = attn.to_out[0](hidden_states) + hidden_states = attn.to_out[1](hidden_states) + + if encoder_hidden_states is not None: + return hidden_states, encoder_hidden_states + else: + return hidden_states + + +class Flux2Attention(torch.nn.Module): + _default_processor_cls = Flux2AttnProcessor + _available_processors = [Flux2AttnProcessor] + + def __init__( + self, + query_dim: int, + heads: int = 8, + dim_head: int = 64, + dropout: float = 0.0, + bias: bool = False, + added_kv_proj_dim: Optional[int] = None, + added_proj_bias: Optional[bool] = True, + out_bias: bool = True, + eps: float = 1e-5, + out_dim: int = None, + elementwise_affine: bool = True, + processor=None, + ): + super().__init__() + + self.head_dim = dim_head + self.inner_dim = out_dim if out_dim is not None else dim_head * heads + self.query_dim = query_dim + self.out_dim = out_dim if out_dim is not None else query_dim + self.heads = out_dim // dim_head if out_dim is not None else heads + + self.use_bias = bias + self.dropout = dropout + + self.added_kv_proj_dim = added_kv_proj_dim + self.added_proj_bias = added_proj_bias + + self.to_q = torch.nn.Linear(query_dim, self.inner_dim, bias=bias) + self.to_k = torch.nn.Linear(query_dim, self.inner_dim, bias=bias) + self.to_v = torch.nn.Linear(query_dim, self.inner_dim, bias=bias) + + # QK Norm + self.norm_q = torch.nn.RMSNorm(dim_head, eps=eps, elementwise_affine=elementwise_affine) + self.norm_k = torch.nn.RMSNorm(dim_head, eps=eps, elementwise_affine=elementwise_affine) + + self.to_out = torch.nn.ModuleList([]) + self.to_out.append(torch.nn.Linear(self.inner_dim, self.out_dim, bias=out_bias)) + self.to_out.append(torch.nn.Dropout(dropout)) + + if added_kv_proj_dim is not None: + self.norm_added_q = torch.nn.RMSNorm(dim_head, eps=eps) + self.norm_added_k = torch.nn.RMSNorm(dim_head, eps=eps) + self.add_q_proj = torch.nn.Linear(added_kv_proj_dim, self.inner_dim, bias=added_proj_bias) + self.add_k_proj = torch.nn.Linear(added_kv_proj_dim, self.inner_dim, bias=added_proj_bias) + self.add_v_proj = torch.nn.Linear(added_kv_proj_dim, self.inner_dim, bias=added_proj_bias) + self.to_add_out = torch.nn.Linear(self.inner_dim, query_dim, bias=out_bias) + + if processor is None: + processor = self._default_processor_cls() + self.set_processor(processor) + + def set_processor(self, processor: AttentionProcessor) -> None: + """ + Set the attention processor to use. + + Args: + processor (`AttnProcessor`): + The attention processor to use. + """ + # if current processor is in `self._modules` and if passed `processor` is not, we need to + # pop `processor` from `self._modules` + if ( + hasattr(self, "processor") + and isinstance(self.processor, torch.nn.Module) + and not isinstance(processor, torch.nn.Module) + ): + logger.info(f"You are removing possibly trained weights of {self.processor} with {processor}") + self._modules.pop("processor") + + self.processor = processor + + def get_processor(self, return_deprecated_lora: bool = False) -> "AttentionProcessor": + """ + Get the attention processor in use. + + Args: + return_deprecated_lora (`bool`, *optional*, defaults to `False`): + Set to `True` to return the deprecated LoRA attention processor. + + Returns: + "AttentionProcessor": The attention processor in use. + """ + if not return_deprecated_lora: + return self.processor + + def forward( + self, + hidden_states: torch.Tensor, + encoder_hidden_states: Optional[torch.Tensor] = None, + attention_mask: Optional[torch.Tensor] = None, + image_rotary_emb: Optional[torch.Tensor] = None, + **kwargs, + ) -> torch.Tensor: + attn_parameters = set(inspect.signature(self.processor.__call__).parameters.keys()) + unused_kwargs = [k for k, _ in kwargs.items() if k not in attn_parameters] + if len(unused_kwargs) > 0: + logger.warning( + f"joint_attention_kwargs {unused_kwargs} are not expected by {self.processor.__class__.__name__} and will be ignored." + ) + kwargs = {k: w for k, w in kwargs.items() if k in attn_parameters} + return self.processor(self, hidden_states, encoder_hidden_states, attention_mask, image_rotary_emb, **kwargs) + + +class Flux2ParallelSelfAttention(torch.nn.Module): + """ + Flux 2 parallel self-attention for the Flux 2 single-stream transformer blocks. + + This implements a parallel transformer block, where the attention QKV projections are fused to the feedforward (FF) + input projections, and the attention output projections are fused to the FF output projections. See the [ViT-22B + paper](https://arxiv.org/abs/2302.05442) for a visual depiction of this type of transformer block. + """ + + _default_processor_cls = Flux2AttnProcessor + _available_processors = [Flux2AttnProcessor] + # Does not support QKV fusion as the QKV projections are always fused + _supports_qkv_fusion = False + + def __init__( + self, + query_dim: int, + heads: int = 8, + dim_head: int = 64, + dropout: float = 0.0, + bias: bool = False, + out_bias: bool = True, + eps: float = 1e-5, + out_dim: int = None, + elementwise_affine: bool = True, + mlp_ratio: float = 4.0, + mlp_mult_factor: int = 2, + processor=None, + ): + super().__init__() + + self.head_dim = dim_head + self.inner_dim = out_dim if out_dim is not None else dim_head * heads + self.query_dim = query_dim + self.out_dim = out_dim if out_dim is not None else query_dim + self.heads = out_dim // dim_head if out_dim is not None else heads + + self.use_bias = bias + self.dropout = dropout + + self.mlp_ratio = mlp_ratio + self.mlp_hidden_dim = int(query_dim * self.mlp_ratio) + self.mlp_mult_factor = mlp_mult_factor + + # Fused QKV projections + MLP input projection + self.to_qkv_mlp_proj = torch.nn.Linear( + self.query_dim, self.inner_dim * 3 + self.mlp_hidden_dim * self.mlp_mult_factor, bias=bias + ) + self.mlp_act_fn = Flux2SwiGLU() + + # QK Norm + self.norm_q = torch.nn.RMSNorm(dim_head, eps=eps, elementwise_affine=elementwise_affine) + self.norm_k = torch.nn.RMSNorm(dim_head, eps=eps, elementwise_affine=elementwise_affine) + + # Fused attention output projection + MLP output projection + self.to_out = torch.nn.Linear(self.inner_dim + self.mlp_hidden_dim, self.out_dim, bias=out_bias) + + if processor is None: + processor = self._default_processor_cls() + self.set_processor(processor) + + def set_processor(self, processor: AttentionProcessor) -> None: + """ + Set the attention processor to use. + + Args: + processor (`AttnProcessor`): + The attention processor to use. + """ + # if current processor is in `self._modules` and if passed `processor` is not, we need to + # pop `processor` from `self._modules` + if ( + hasattr(self, "processor") + and isinstance(self.processor, torch.nn.Module) + and not isinstance(processor, torch.nn.Module) + ): + logger.info(f"You are removing possibly trained weights of {self.processor} with {processor}") + self._modules.pop("processor") + + self.processor = processor + + def get_processor(self, return_deprecated_lora: bool = False) -> "AttentionProcessor": + """ + Get the attention processor in use. + + Args: + return_deprecated_lora (`bool`, *optional*, defaults to `False`): + Set to `True` to return the deprecated LoRA attention processor. + + Returns: + "AttentionProcessor": The attention processor in use. + """ + if not return_deprecated_lora: + return self.processor + + def forward( + self, + hidden_states: torch.Tensor, + encoder_hidden_states: Optional[torch.Tensor] = None, + attention_mask: Optional[torch.Tensor] = None, + image_rotary_emb: Optional[torch.Tensor] = None, + **kwargs, + ) -> torch.Tensor: + attn_parameters = set(inspect.signature(self.processor.__call__).parameters.keys()) + unused_kwargs = [k for k, _ in kwargs.items() if k not in attn_parameters] + if len(unused_kwargs) > 0: + logger.warning( + f"joint_attention_kwargs {unused_kwargs} are not expected by {self.processor.__class__.__name__} and will be ignored." + ) + kwargs = {k: w for k, w in kwargs.items() if k in attn_parameters} + return self.processor(self, hidden_states, encoder_hidden_states, attention_mask, image_rotary_emb, **kwargs) + + +class Flux2SingleTransformerBlock(nn.Module): + def __init__( + self, + dim: int, + num_attention_heads: int, + attention_head_dim: int, + mlp_ratio: float = 3.0, + eps: float = 1e-6, + bias: bool = False, + ): + super().__init__() + + self.norm = nn.LayerNorm(dim, elementwise_affine=False, eps=eps) + + # Note that the MLP in/out linear layers are fused with the attention QKV/out projections, respectively; this + # is often called a "parallel" transformer block. See the [ViT-22B paper](https://arxiv.org/abs/2302.05442) + # for a visual depiction of this type of transformer block. + self.attn = Flux2ParallelSelfAttention( + query_dim=dim, + dim_head=attention_head_dim, + heads=num_attention_heads, + out_dim=dim, + bias=bias, + out_bias=bias, + eps=eps, + mlp_ratio=mlp_ratio, + mlp_mult_factor=2, + processor=Flux2AttnProcessor(), + ) + + def forward( + self, + hidden_states: torch.Tensor, + encoder_hidden_states: Optional[torch.Tensor], + temb_mod_params: Tuple[torch.Tensor, torch.Tensor, torch.Tensor], + image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, + joint_attention_kwargs: Optional[Dict[str, Any]] = None, + ) -> Tuple[torch.Tensor, torch.Tensor]: + # If encoder_hidden_states is None, hidden_states is assumed to have encoder_hidden_states already + # concatenated + if encoder_hidden_states is not None: + text_seq_len = encoder_hidden_states.shape[1] + hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1) + + mod_shift, mod_scale, mod_gate = temb_mod_params + + norm_hidden_states = self.norm(hidden_states) + norm_hidden_states = (1 + mod_scale) * norm_hidden_states + mod_shift + + joint_attention_kwargs = joint_attention_kwargs or {} + attn_output = self.attn( + hidden_states=norm_hidden_states, + image_rotary_emb=image_rotary_emb, + text_seq_len=text_seq_len, + **joint_attention_kwargs, + ) + + hidden_states = hidden_states + mod_gate * attn_output + if hidden_states.dtype == torch.float16: + hidden_states = hidden_states.clip(-65504, 65504) + + encoder_hidden_states, hidden_states = hidden_states[:, :text_seq_len], hidden_states[:, text_seq_len:] + return encoder_hidden_states, hidden_states + + +class Flux2TransformerBlock(nn.Module): + def __init__( + self, + dim: int, + num_attention_heads: int, + attention_head_dim: int, + mlp_ratio: float = 3.0, + eps: float = 1e-6, + bias: bool = False, + ): + super().__init__() + self.mlp_hidden_dim = int(dim * mlp_ratio) + + self.norm1 = nn.LayerNorm(dim, elementwise_affine=False, eps=eps) + self.norm1_context = nn.LayerNorm(dim, elementwise_affine=False, eps=eps) + + self.attn = Flux2Attention( + query_dim=dim, + added_kv_proj_dim=dim, + dim_head=attention_head_dim, + heads=num_attention_heads, + out_dim=dim, + bias=bias, + added_proj_bias=bias, + out_bias=bias, + eps=eps, + processor=Flux2AttnProcessor(), + ) + + self.norm2 = nn.LayerNorm(dim, elementwise_affine=False, eps=eps) + self.ff = Flux2FeedForward(dim=dim, dim_out=dim, mult=mlp_ratio, bias=bias) + + self.norm2_context = nn.LayerNorm(dim, elementwise_affine=False, eps=eps) + self.ff_context = Flux2FeedForward(dim=dim, dim_out=dim, mult=mlp_ratio, bias=bias) + + def forward( + self, + hidden_states: torch.Tensor, + encoder_hidden_states: torch.Tensor, + temb_mod_params_img: Tuple[Tuple[torch.Tensor, torch.Tensor, torch.Tensor], ...], + temb_mod_params_txt: Tuple[Tuple[torch.Tensor, torch.Tensor, torch.Tensor], ...], + image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, + joint_attention_kwargs: Optional[Dict[str, Any]] = None, + ) -> Tuple[torch.Tensor, torch.Tensor]: + joint_attention_kwargs = joint_attention_kwargs or {} + + # Modulation parameters shape: [1, 1, self.dim] + (shift_msa, scale_msa, gate_msa), (shift_mlp, scale_mlp, gate_mlp) = temb_mod_params_img + (c_shift_msa, c_scale_msa, c_gate_msa), (c_shift_mlp, c_scale_mlp, c_gate_mlp) = temb_mod_params_txt + + # Img stream + norm_hidden_states = self.norm1(hidden_states) + norm_hidden_states = (1 + scale_msa) * norm_hidden_states + shift_msa + + # Conditioning txt stream + norm_encoder_hidden_states = self.norm1_context(encoder_hidden_states) + norm_encoder_hidden_states = (1 + c_scale_msa) * norm_encoder_hidden_states + c_shift_msa + + # Attention on concatenated img + txt stream + attention_outputs = self.attn( + hidden_states=norm_hidden_states, + encoder_hidden_states=norm_encoder_hidden_states, + image_rotary_emb=image_rotary_emb, + **joint_attention_kwargs, + ) + + attn_output, context_attn_output = attention_outputs + + # Process attention outputs for the image stream (`hidden_states`). + attn_output = gate_msa * attn_output + hidden_states = hidden_states + attn_output + + norm_hidden_states = self.norm2(hidden_states) + norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp + + ff_output = self.ff(norm_hidden_states) + hidden_states = hidden_states + gate_mlp * ff_output + + # Process attention outputs for the text stream (`encoder_hidden_states`). + context_attn_output = c_gate_msa * context_attn_output + encoder_hidden_states = encoder_hidden_states + context_attn_output + + norm_encoder_hidden_states = self.norm2_context(encoder_hidden_states) + norm_encoder_hidden_states = norm_encoder_hidden_states * (1 + c_scale_mlp) + c_shift_mlp + + context_ff_output = self.ff_context(norm_encoder_hidden_states) + encoder_hidden_states = encoder_hidden_states + c_gate_mlp * context_ff_output + if encoder_hidden_states.dtype == torch.float16: + encoder_hidden_states = encoder_hidden_states.clip(-65504, 65504) + + return encoder_hidden_states, hidden_states + + +class Flux2PosEmbed(nn.Module): + # modified from https://github.com/black-forest-labs/flux/blob/c00d7c60b085fce8058b9df845e036090873f2ce/src/flux/modules/layers.py#L11 + def __init__(self, theta: int, axes_dim: List[int]): + super().__init__() + self.theta = theta + self.axes_dim = axes_dim + + def forward(self, ids: torch.Tensor) -> torch.Tensor: + # Expected ids shape: [S, len(self.axes_dim)] + cos_out = [] + sin_out = [] + pos = ids.float() + is_mps = ids.device.type == "mps" + is_npu = ids.device.type == "npu" + freqs_dtype = torch.float32 if (is_mps or is_npu) else torch.float64 + # Unlike Flux 1, loop over len(self.axes_dim) rather than ids.shape[-1] + for i in range(len(self.axes_dim)): + cos, sin = get_1d_rotary_pos_embed( + self.axes_dim[i], + pos[..., i], + theta=self.theta, + repeat_interleave_real=True, + use_real=True, + freqs_dtype=freqs_dtype, + ) + cos_out.append(cos) + sin_out.append(sin) + freqs_cos = torch.cat(cos_out, dim=-1).to(ids.device) + freqs_sin = torch.cat(sin_out, dim=-1).to(ids.device) + return freqs_cos, freqs_sin + + +class Flux2TimestepGuidanceEmbeddings(nn.Module): + def __init__(self, in_channels: int = 256, embedding_dim: int = 6144, bias: bool = False): + super().__init__() + + self.time_proj = Timesteps(num_channels=in_channels, flip_sin_to_cos=True, downscale_freq_shift=0) + self.timestep_embedder = TimestepEmbedding( + in_channels=in_channels, time_embed_dim=embedding_dim, sample_proj_bias=bias + ) + + self.guidance_embedder = TimestepEmbedding( + in_channels=in_channels, time_embed_dim=embedding_dim, sample_proj_bias=bias + ) + + def forward(self, timestep: torch.Tensor, guidance: torch.Tensor) -> torch.Tensor: + timesteps_proj = self.time_proj(timestep) + timesteps_emb = self.timestep_embedder(timesteps_proj.to(timestep.dtype)) # (N, D) + + guidance_proj = self.time_proj(guidance) + guidance_emb = self.guidance_embedder(guidance_proj.to(guidance.dtype)) # (N, D) + + time_guidance_emb = timesteps_emb + guidance_emb + + return time_guidance_emb + + +class Flux2Modulation(nn.Module): + def __init__(self, dim: int, mod_param_sets: int = 2, bias: bool = False): + super().__init__() + self.mod_param_sets = mod_param_sets + + self.linear = nn.Linear(dim, dim * 3 * self.mod_param_sets, bias=bias) + self.act_fn = nn.SiLU() + + def forward(self, temb: torch.Tensor) -> Tuple[Tuple[torch.Tensor, torch.Tensor, torch.Tensor], ...]: + mod = self.act_fn(temb) + mod = self.linear(mod) + + if mod.ndim == 2: + mod = mod.unsqueeze(1) + mod_params = torch.chunk(mod, 3 * self.mod_param_sets, dim=-1) + # Return tuple of 3-tuples of modulation params shift/scale/gate + return tuple(mod_params[3 * i : 3 * (i + 1)] for i in range(self.mod_param_sets)) + + +class Flux2Transformer2DModel( + ModelMixin, + ConfigMixin, + FromOriginalModelMixin, +): + """ + The Transformer model introduced in Flux 2. + + Reference: https://blackforestlabs.ai/announcing-black-forest-labs/ + + Args: + patch_size (`int`, defaults to `1`): + Patch size to turn the input data into small patches. + in_channels (`int`, defaults to `128`): + The number of channels in the input. + out_channels (`int`, *optional*, defaults to `None`): + The number of channels in the output. If not specified, it defaults to `in_channels`. + num_layers (`int`, defaults to `8`): + The number of layers of dual stream DiT blocks to use. + num_single_layers (`int`, defaults to `48`): + The number of layers of single stream DiT blocks to use. + attention_head_dim (`int`, defaults to `128`): + The number of dimensions to use for each attention head. + num_attention_heads (`int`, defaults to `48`): + The number of attention heads to use. + joint_attention_dim (`int`, defaults to `15360`): + The number of dimensions to use for the joint attention (embedding/channel dimension of + `encoder_hidden_states`). + pooled_projection_dim (`int`, defaults to `768`): + The number of dimensions to use for the pooled projection. + guidance_embeds (`bool`, defaults to `True`): + Whether to use guidance embeddings for guidance-distilled variant of the model. + axes_dims_rope (`Tuple[int]`, defaults to `(32, 32, 32, 32)`): + The dimensions to use for the rotary positional embeddings. + """ + + _supports_gradient_checkpointing = True + # _no_split_modules = ["Flux2TransformerBlock", "Flux2SingleTransformerBlock"] + # _skip_layerwise_casting_patterns = ["pos_embed", "norm"] + # _repeated_blocks = ["Flux2TransformerBlock", "Flux2SingleTransformerBlock"] + + @register_to_config + def __init__( + self, + patch_size: int = 1, + in_channels: int = 128, + out_channels: Optional[int] = None, + num_layers: int = 8, + num_single_layers: int = 48, + attention_head_dim: int = 128, + num_attention_heads: int = 48, + joint_attention_dim: int = 15360, + timestep_guidance_channels: int = 256, + mlp_ratio: float = 3.0, + axes_dims_rope: Tuple[int, ...] = (32, 32, 32, 32), + rope_theta: int = 2000, + eps: float = 1e-6, + ): + super().__init__() + self.out_channels = out_channels or in_channels + self.inner_dim = num_attention_heads * attention_head_dim + + # 1. Sinusoidal positional embedding for RoPE on image and text tokens + self.pos_embed = Flux2PosEmbed(theta=rope_theta, axes_dim=axes_dims_rope) + + # 2. Combined timestep + guidance embedding + self.time_guidance_embed = Flux2TimestepGuidanceEmbeddings( + in_channels=timestep_guidance_channels, embedding_dim=self.inner_dim, bias=False + ) + + # 3. Modulation (double stream and single stream blocks share modulation parameters, resp.) + # Two sets of shift/scale/gate modulation parameters for the double stream attn and FF sub-blocks + self.double_stream_modulation_img = Flux2Modulation(self.inner_dim, mod_param_sets=2, bias=False) + self.double_stream_modulation_txt = Flux2Modulation(self.inner_dim, mod_param_sets=2, bias=False) + # Only one set of modulation parameters as the attn and FF sub-blocks are run in parallel for single stream + self.single_stream_modulation = Flux2Modulation(self.inner_dim, mod_param_sets=1, bias=False) + + # 4. Input projections + self.x_embedder = nn.Linear(in_channels, self.inner_dim, bias=False) + self.context_embedder = nn.Linear(joint_attention_dim, self.inner_dim, bias=False) + + # 5. Double Stream Transformer Blocks + self.transformer_blocks = nn.ModuleList( + [ + Flux2TransformerBlock( + dim=self.inner_dim, + num_attention_heads=num_attention_heads, + attention_head_dim=attention_head_dim, + mlp_ratio=mlp_ratio, + eps=eps, + bias=False, + ) + for _ in range(num_layers) + ] + ) + + # 6. Single Stream Transformer Blocks + self.single_transformer_blocks = nn.ModuleList( + [ + Flux2SingleTransformerBlock( + dim=self.inner_dim, + num_attention_heads=num_attention_heads, + attention_head_dim=attention_head_dim, + mlp_ratio=mlp_ratio, + eps=eps, + bias=False, + ) + for _ in range(num_single_layers) + ] + ) + + # 7. Output layers + self.norm_out = AdaLayerNormContinuous( + self.inner_dim, self.inner_dim, elementwise_affine=False, eps=eps, bias=False + ) + self.proj_out = nn.Linear(self.inner_dim, patch_size * patch_size * self.out_channels, bias=False) + + self.gradient_checkpointing = False + + self.sp_world_size = 1 + self.sp_world_rank = 0 + + def _set_gradient_checkpointing(self, *args, **kwargs): + if "value" in kwargs: + self.gradient_checkpointing = kwargs["value"] + elif "enable" in kwargs: + self.gradient_checkpointing = kwargs["enable"] + else: + raise ValueError("Invalid set gradient checkpointing") + + def enable_multi_gpus_inference(self,): + self.sp_world_size = get_sequence_parallel_world_size() + self.sp_world_rank = get_sequence_parallel_rank() + self.all_gather = get_sp_group().all_gather + self.set_attn_processor(Flux2MultiGPUsAttnProcessor2_0()) + + @property + # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors + def attn_processors(self) -> Dict[str, AttentionProcessor]: + r""" + Returns: + `dict` of attention processors: A dictionary containing all attention processors used in the model with + indexed by its weight name. + """ + # set recursively + processors = {} + + def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]): + if hasattr(module, "get_processor"): + processors[f"{name}.processor"] = module.get_processor() + + for sub_name, child in module.named_children(): + fn_recursive_add_processors(f"{name}.{sub_name}", child, processors) + + return processors + + for name, module in self.named_children(): + fn_recursive_add_processors(name, module, processors) + + return processors + + # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor + def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]): + r""" + Sets the attention processor to use to compute attention. + + Parameters: + processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`): + The instantiated processor class or a dictionary of processor classes that will be set as the processor + for **all** `Attention` layers. + + If `processor` is a dict, the key needs to define the path to the corresponding cross attention + processor. This is strongly recommended when setting trainable attention processors. + + """ + count = len(self.attn_processors.keys()) + + if isinstance(processor, dict) and len(processor) != count: + raise ValueError( + f"A dict of processors was passed, but the number of processors {len(processor)} does not match the" + f" number of attention layers: {count}. Please make sure to pass {count} processor classes." + ) + + def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor): + if hasattr(module, "set_processor"): + if not isinstance(processor, dict): + module.set_processor(processor) + else: + module.set_processor(processor.pop(f"{name}.processor")) + + for sub_name, child in module.named_children(): + fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor) + + for name, module in self.named_children(): + fn_recursive_attn_processor(name, module, processor) + + def forward( + self, + hidden_states: torch.Tensor, + encoder_hidden_states: torch.Tensor = None, + timestep: torch.LongTensor = None, + img_ids: torch.Tensor = None, + txt_ids: torch.Tensor = None, + guidance: torch.Tensor = None, + joint_attention_kwargs: Optional[Dict[str, Any]] = None, + return_dict: bool = True, + ) -> Union[torch.Tensor, Transformer2DModelOutput]: + """ + The [`FluxTransformer2DModel`] forward method. + + Args: + hidden_states (`torch.Tensor` of shape `(batch_size, image_sequence_length, in_channels)`): + Input `hidden_states`. + encoder_hidden_states (`torch.Tensor` of shape `(batch_size, text_sequence_length, joint_attention_dim)`): + Conditional embeddings (embeddings computed from the input conditions such as prompts) to use. + timestep ( `torch.LongTensor`): + Used to indicate denoising step. + block_controlnet_hidden_states: (`list` of `torch.Tensor`): + A list of tensors that if specified are added to the residuals of transformer blocks. + joint_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~models.transformer_2d.Transformer2DModelOutput`] instead of a plain + tuple. + + Returns: + If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a + `tuple` where the first element is the sample tensor. + """ + # 0. Handle input arguments + if joint_attention_kwargs is not None: + joint_attention_kwargs = joint_attention_kwargs.copy() + lora_scale = joint_attention_kwargs.pop("scale", 1.0) + else: + lora_scale = 1.0 + + num_txt_tokens = encoder_hidden_states.shape[1] + + # 1. Calculate timestep embedding and modulation parameters + timestep = timestep.to(hidden_states.dtype) * 1000 + guidance = guidance.to(hidden_states.dtype) * 1000 + + temb = self.time_guidance_embed(timestep, guidance) + + double_stream_mod_img = self.double_stream_modulation_img(temb) + double_stream_mod_txt = self.double_stream_modulation_txt(temb) + single_stream_mod = self.single_stream_modulation(temb)[0] + + # 2. Input projection for image (hidden_states) and conditioning text (encoder_hidden_states) + hidden_states = self.x_embedder(hidden_states) + encoder_hidden_states = self.context_embedder(encoder_hidden_states) + + # 3. Calculate RoPE embeddings from image and text tokens + # NOTE: the below logic means that we can't support batched inference with images of different resolutions or + # text prompts of differents lengths. Is this a use case we want to support? + if img_ids.ndim == 3: + img_ids = img_ids[0] + if txt_ids.ndim == 3: + txt_ids = txt_ids[0] + + if is_torch_npu_available(): + freqs_cos_image, freqs_sin_image = self.pos_embed(img_ids.cpu()) + image_rotary_emb = (freqs_cos_image.npu(), freqs_sin_image.npu()) + freqs_cos_text, freqs_sin_text = self.pos_embed(txt_ids.cpu()) + text_rotary_emb = (freqs_cos_text.npu(), freqs_sin_text.npu()) + else: + image_rotary_emb = self.pos_embed(img_ids) + text_rotary_emb = self.pos_embed(txt_ids) + concat_rotary_emb = ( + torch.cat([text_rotary_emb[0], image_rotary_emb[0]], dim=0), + torch.cat([text_rotary_emb[1], image_rotary_emb[1]], dim=0), + ) + + # Context Parallel + if self.sp_world_size > 1: + hidden_states = torch.chunk(hidden_states, self.sp_world_size, dim=1)[self.sp_world_rank] + if concat_rotary_emb is not None: + txt_rotary_emb = ( + concat_rotary_emb[0][:encoder_hidden_states.shape[1]], + concat_rotary_emb[1][:encoder_hidden_states.shape[1]] + ) + concat_rotary_emb = ( + torch.chunk(concat_rotary_emb[0][encoder_hidden_states.shape[1]:], self.sp_world_size, dim=0)[self.sp_world_rank], + torch.chunk(concat_rotary_emb[1][encoder_hidden_states.shape[1]:], self.sp_world_size, dim=0)[self.sp_world_rank], + ) + concat_rotary_emb = [torch.cat([_txt_rotary_emb, _image_rotary_emb], dim=0) \ + for _txt_rotary_emb, _image_rotary_emb in zip(txt_rotary_emb, concat_rotary_emb)] + + # 4. Double Stream Transformer Blocks + for index_block, block in enumerate(self.transformer_blocks): + if torch.is_grad_enabled() and self.gradient_checkpointing: + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + encoder_hidden_states, hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(block), + hidden_states, + encoder_hidden_states, + double_stream_mod_img, + double_stream_mod_txt, + concat_rotary_emb, + joint_attention_kwargs, + **ckpt_kwargs, + ) + else: + encoder_hidden_states, hidden_states = block( + hidden_states=hidden_states, + encoder_hidden_states=encoder_hidden_states, + temb_mod_params_img=double_stream_mod_img, + temb_mod_params_txt=double_stream_mod_txt, + image_rotary_emb=concat_rotary_emb, + joint_attention_kwargs=joint_attention_kwargs, + ) + + # 5. Single Stream Transformer Blocks + for index_block, block in enumerate(self.single_transformer_blocks): + if torch.is_grad_enabled() and self.gradient_checkpointing: + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + encoder_hidden_states, hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(block), + hidden_states, + encoder_hidden_states, + single_stream_mod, + concat_rotary_emb, + joint_attention_kwargs, + **ckpt_kwargs, + ) + else: + encoder_hidden_states, hidden_states = block( + hidden_states=hidden_states, + encoder_hidden_states=encoder_hidden_states, + temb_mod_params=single_stream_mod, + image_rotary_emb=concat_rotary_emb, + joint_attention_kwargs=joint_attention_kwargs, + ) + + # 6. Output layers + hidden_states = self.norm_out(hidden_states, temb) + output = self.proj_out(hidden_states) + + if self.sp_world_size > 1: + output = self.all_gather(output, dim=1) + + if not return_dict: + return (output,) + + return Transformer2DModelOutput(sample=output) + + @classmethod + def from_pretrained( + cls, pretrained_model_path, subfolder=None, transformer_additional_kwargs={}, + low_cpu_mem_usage=False, torch_dtype=torch.bfloat16 + ): + if subfolder is not None: + pretrained_model_path = os.path.join(pretrained_model_path, subfolder) + print(f"loaded 3D transformer's pretrained weights from {pretrained_model_path} ...") + + config_file = os.path.join(pretrained_model_path, 'config.json') + if not os.path.isfile(config_file): + raise RuntimeError(f"{config_file} does not exist") + with open(config_file, "r") as f: + config = json.load(f) + + from diffusers.utils import WEIGHTS_NAME + model_file = os.path.join(pretrained_model_path, WEIGHTS_NAME) + model_file_safetensors = model_file.replace(".bin", ".safetensors") + + if "dict_mapping" in transformer_additional_kwargs.keys(): + for key in transformer_additional_kwargs["dict_mapping"]: + transformer_additional_kwargs[transformer_additional_kwargs["dict_mapping"][key]] = config[key] + + if low_cpu_mem_usage: + try: + import re + + from diffusers import __version__ as diffusers_version + if diffusers_version >= "0.33.0": + from diffusers.models.model_loading_utils import \ + load_model_dict_into_meta + else: + from diffusers.models.modeling_utils import \ + load_model_dict_into_meta + from diffusers.utils import is_accelerate_available + if is_accelerate_available(): + import accelerate + + # Instantiate model with empty weights + with accelerate.init_empty_weights(): + model = cls.from_config(config, **transformer_additional_kwargs) + + param_device = "cpu" + if os.path.exists(model_file): + state_dict = torch.load(model_file, map_location="cpu") + elif os.path.exists(model_file_safetensors): + from safetensors.torch import load_file, safe_open + state_dict = load_file(model_file_safetensors) + else: + from safetensors.torch import load_file, safe_open + model_files_safetensors = glob.glob(os.path.join(pretrained_model_path, "*.safetensors")) + state_dict = {} + print(model_files_safetensors) + for _model_file_safetensors in model_files_safetensors: + _state_dict = load_file(_model_file_safetensors) + for key in _state_dict: + state_dict[key] = _state_dict[key] + + filtered_state_dict = {} + for key in state_dict: + if key in model.state_dict() and model.state_dict()[key].size() == state_dict[key].size(): + filtered_state_dict[key] = state_dict[key] + else: + print(f"Skipping key '{key}' due to size mismatch or absence in model.") + + model_keys = set(model.state_dict().keys()) + loaded_keys = set(filtered_state_dict.keys()) + missing_keys = model_keys - loaded_keys + + def initialize_missing_parameters(missing_keys, model_state_dict, torch_dtype=None): + initialized_dict = {} + + with torch.no_grad(): + for key in missing_keys: + param_shape = model_state_dict[key].shape + param_dtype = torch_dtype if torch_dtype is not None else model_state_dict[key].dtype + if 'weight' in key: + if any(norm_type in key for norm_type in ['norm', 'ln_', 'layer_norm', 'group_norm', 'batch_norm']): + initialized_dict[key] = torch.ones(param_shape, dtype=param_dtype) + elif 'embedding' in key or 'embed' in key: + initialized_dict[key] = torch.randn(param_shape, dtype=param_dtype) * 0.02 + elif 'head' in key or 'output' in key or 'proj_out' in key: + initialized_dict[key] = torch.zeros(param_shape, dtype=param_dtype) + elif len(param_shape) >= 2: + initialized_dict[key] = torch.empty(param_shape, dtype=param_dtype) + nn.init.xavier_uniform_(initialized_dict[key]) + else: + initialized_dict[key] = torch.randn(param_shape, dtype=param_dtype) * 0.02 + elif 'bias' in key: + initialized_dict[key] = torch.zeros(param_shape, dtype=param_dtype) + elif 'running_mean' in key: + initialized_dict[key] = torch.zeros(param_shape, dtype=param_dtype) + elif 'running_var' in key: + initialized_dict[key] = torch.ones(param_shape, dtype=param_dtype) + elif 'num_batches_tracked' in key: + initialized_dict[key] = torch.zeros(param_shape, dtype=torch.long) + else: + initialized_dict[key] = torch.zeros(param_shape, dtype=param_dtype) + + return initialized_dict + + if missing_keys: + print(f"Missing keys will be initialized: {sorted(missing_keys)}") + initialized_params = initialize_missing_parameters( + missing_keys, + model.state_dict(), + torch_dtype + ) + filtered_state_dict.update(initialized_params) + + if diffusers_version >= "0.33.0": + # Diffusers has refactored `load_model_dict_into_meta` since version 0.33.0 in this commit: + # https://github.com/huggingface/diffusers/commit/f5929e03060d56063ff34b25a8308833bec7c785. + load_model_dict_into_meta( + model, + filtered_state_dict, + dtype=torch_dtype, + model_name_or_path=pretrained_model_path, + ) + else: + model._convert_deprecated_attention_blocks(filtered_state_dict) + unexpected_keys = load_model_dict_into_meta( + model, + filtered_state_dict, + device=param_device, + dtype=torch_dtype, + model_name_or_path=pretrained_model_path, + ) + + if cls._keys_to_ignore_on_load_unexpected is not None: + for pat in cls._keys_to_ignore_on_load_unexpected: + unexpected_keys = [k for k in unexpected_keys if re.search(pat, k) is None] + + if len(unexpected_keys) > 0: + print( + f"Some weights of the model checkpoint were not used when initializing {cls.__name__}: \n {[', '.join(unexpected_keys)]}" + ) + + return model + except Exception as e: + print( + f"The low_cpu_mem_usage mode is not work because {e}. Use low_cpu_mem_usage=False instead." + ) + + model = cls.from_config(config, **transformer_additional_kwargs) + if os.path.exists(model_file): + state_dict = torch.load(model_file, map_location="cpu") + elif os.path.exists(model_file_safetensors): + from safetensors.torch import load_file, safe_open + state_dict = load_file(model_file_safetensors) + else: + from safetensors.torch import load_file, safe_open + model_files_safetensors = glob.glob(os.path.join(pretrained_model_path, "*.safetensors")) + state_dict = {} + for _model_file_safetensors in model_files_safetensors: + _state_dict = load_file(_model_file_safetensors) + for key in _state_dict: + state_dict[key] = _state_dict[key] + + tmp_state_dict = {} + for key in state_dict: + if key in model.state_dict().keys() and model.state_dict()[key].size() == state_dict[key].size(): + tmp_state_dict[key] = state_dict[key] + else: + print(key, "Size don't match, skip") + + state_dict = tmp_state_dict + + m, u = model.load_state_dict(state_dict, strict=False) + print(f"### missing keys: {len(m)}; \n### unexpected keys: {len(u)};") + print(m) + + params = [p.numel() if "." in n else 0 for n, p in model.named_parameters()] + print(f"### All Parameters: {sum(params) / 1e6} M") + + params = [p.numel() if "attn1." in n else 0 for n, p in model.named_parameters()] + print(f"### attn1 Parameters: {sum(params) / 1e6} M") + + model = model.to(torch_dtype) + return model \ No newline at end of file diff --git a/videox_fun/models/flux2_vae.py b/videox_fun/models/flux2_vae.py new file mode 100644 index 0000000..76326c5 --- /dev/null +++ b/videox_fun/models/flux2_vae.py @@ -0,0 +1,543 @@ +# Modified from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/autoencoders/autoencoder_kl_flux2.py +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +import math +from typing import Dict, Optional, Tuple, Union + +import torch +import torch.nn as nn +from diffusers.configuration_utils import ConfigMixin, register_to_config +from diffusers.loaders.single_file_model import FromOriginalModelMixin +from diffusers.models.attention_processor import ( + ADDED_KV_ATTENTION_PROCESSORS, CROSS_ATTENTION_PROCESSORS, Attention, + AttentionProcessor, AttnAddedKVProcessor, AttnProcessor, + FusedAttnProcessor2_0) +from diffusers.models.autoencoders.vae import (Decoder, + DecoderOutput, + DiagonalGaussianDistribution, + Encoder) +from diffusers.models.modeling_outputs import AutoencoderKLOutput +from diffusers.models.modeling_utils import ModelMixin +from diffusers.utils import deprecate +from diffusers.utils.accelerate_utils import apply_forward_hook + + +class AutoencoderKLFlux2(ModelMixin, ConfigMixin, FromOriginalModelMixin): + r""" + A VAE model with KL loss for encoding images into latents and decoding latent representations into images. + + This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented + for all models (such as downloading or saving). + + Parameters: + in_channels (int, *optional*, defaults to 3): Number of channels in the input image. + out_channels (int, *optional*, defaults to 3): Number of channels in the output. + down_block_types (`Tuple[str]`, *optional*, defaults to `("DownEncoderBlock2D",)`): + Tuple of downsample block types. + up_block_types (`Tuple[str]`, *optional*, defaults to `("UpDecoderBlock2D",)`): + Tuple of upsample block types. + block_out_channels (`Tuple[int]`, *optional*, defaults to `(64,)`): + Tuple of block output channels. + act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use. + latent_channels (`int`, *optional*, defaults to 4): Number of channels in the latent space. + sample_size (`int`, *optional*, defaults to `32`): Sample input size. + force_upcast (`bool`, *optional*, default to `True`): + If enabled it will force the VAE to run in float32 for high image resolution pipelines, such as SD-XL. VAE + can be fine-tuned / trained to a lower range without losing too much precision in which case `force_upcast` + can be set to `False` - see: https://huggingface.co/madebyollin/sdxl-vae-fp16-fix + mid_block_add_attention (`bool`, *optional*, default to `True`): + If enabled, the mid_block of the Encoder and Decoder will have attention blocks. If set to false, the + mid_block will only have resnet blocks + """ + + _supports_gradient_checkpointing = True + _no_split_modules = ["BasicTransformerBlock", "ResnetBlock2D"] + + @register_to_config + def __init__( + self, + in_channels: int = 3, + out_channels: int = 3, + down_block_types: Tuple[str, ...] = ( + "DownEncoderBlock2D", + "DownEncoderBlock2D", + "DownEncoderBlock2D", + "DownEncoderBlock2D", + ), + up_block_types: Tuple[str, ...] = ( + "UpDecoderBlock2D", + "UpDecoderBlock2D", + "UpDecoderBlock2D", + "UpDecoderBlock2D", + ), + block_out_channels: Tuple[int, ...] = ( + 128, + 256, + 512, + 512, + ), + layers_per_block: int = 2, + act_fn: str = "silu", + latent_channels: int = 32, + norm_num_groups: int = 32, + sample_size: int = 1024, # YiYi notes: not sure + force_upcast: bool = True, + use_quant_conv: bool = True, + use_post_quant_conv: bool = True, + mid_block_add_attention: bool = True, + batch_norm_eps: float = 1e-4, + batch_norm_momentum: float = 0.1, + patch_size: Tuple[int, int] = (2, 2), + ): + super().__init__() + + # pass init params to Encoder + self.encoder = Encoder( + in_channels=in_channels, + out_channels=latent_channels, + down_block_types=down_block_types, + block_out_channels=block_out_channels, + layers_per_block=layers_per_block, + act_fn=act_fn, + norm_num_groups=norm_num_groups, + double_z=True, + mid_block_add_attention=mid_block_add_attention, + ) + + # pass init params to Decoder + self.decoder = Decoder( + in_channels=latent_channels, + out_channels=out_channels, + up_block_types=up_block_types, + block_out_channels=block_out_channels, + layers_per_block=layers_per_block, + norm_num_groups=norm_num_groups, + act_fn=act_fn, + mid_block_add_attention=mid_block_add_attention, + ) + + self.quant_conv = nn.Conv2d(2 * latent_channels, 2 * latent_channels, 1) if use_quant_conv else None + self.post_quant_conv = nn.Conv2d(latent_channels, latent_channels, 1) if use_post_quant_conv else None + + self.bn = nn.BatchNorm2d( + math.prod(patch_size) * latent_channels, + eps=batch_norm_eps, + momentum=batch_norm_momentum, + affine=False, + track_running_stats=True, + ) + + self.use_slicing = False + self.use_tiling = False + + # only relevant if vae tiling is enabled + self.tile_sample_min_size = self.config.sample_size + sample_size = ( + self.config.sample_size[0] + if isinstance(self.config.sample_size, (list, tuple)) + else self.config.sample_size + ) + self.tile_latent_min_size = int(sample_size / (2 ** (len(self.config.block_out_channels) - 1))) + self.tile_overlap_factor = 0.25 + + @property + # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors + def attn_processors(self) -> Dict[str, AttentionProcessor]: + r""" + Returns: + `dict` of attention processors: A dictionary containing all attention processors used in the model with + indexed by its weight name. + """ + # set recursively + processors = {} + + def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]): + if hasattr(module, "get_processor"): + processors[f"{name}.processor"] = module.get_processor() + + for sub_name, child in module.named_children(): + fn_recursive_add_processors(f"{name}.{sub_name}", child, processors) + + return processors + + for name, module in self.named_children(): + fn_recursive_add_processors(name, module, processors) + + return processors + + # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor + def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]): + r""" + Sets the attention processor to use to compute attention. + + Parameters: + processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`): + The instantiated processor class or a dictionary of processor classes that will be set as the processor + for **all** `Attention` layers. + + If `processor` is a dict, the key needs to define the path to the corresponding cross attention + processor. This is strongly recommended when setting trainable attention processors. + + """ + count = len(self.attn_processors.keys()) + + if isinstance(processor, dict) and len(processor) != count: + raise ValueError( + f"A dict of processors was passed, but the number of processors {len(processor)} does not match the" + f" number of attention layers: {count}. Please make sure to pass {count} processor classes." + ) + + def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor): + if hasattr(module, "set_processor"): + if not isinstance(processor, dict): + module.set_processor(processor) + else: + module.set_processor(processor.pop(f"{name}.processor")) + + for sub_name, child in module.named_children(): + fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor) + + for name, module in self.named_children(): + fn_recursive_attn_processor(name, module, processor) + + # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor + def set_default_attn_processor(self): + """ + Disables custom attention processors and sets the default attention implementation. + """ + if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()): + processor = AttnAddedKVProcessor() + elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()): + processor = AttnProcessor() + else: + raise ValueError( + f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}" + ) + + self.set_attn_processor(processor) + + def _encode(self, x: torch.Tensor) -> torch.Tensor: + batch_size, num_channels, height, width = x.shape + + if self.use_tiling and (width > self.tile_sample_min_size or height > self.tile_sample_min_size): + return self._tiled_encode(x) + + enc = self.encoder(x) + if self.quant_conv is not None: + enc = self.quant_conv(enc) + + return enc + + @apply_forward_hook + def encode( + self, x: torch.Tensor, return_dict: bool = True + ) -> Union[AutoencoderKLOutput, Tuple[DiagonalGaussianDistribution]]: + """ + Encode a batch of images into latents. + + Args: + x (`torch.Tensor`): Input batch of images. + return_dict (`bool`, *optional*, defaults to `True`): + Whether to return a [`~models.autoencoder_kl.AutoencoderKLOutput`] instead of a plain tuple. + + Returns: + The latent representations of the encoded images. If `return_dict` is True, a + [`~models.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain `tuple` is returned. + """ + if self.use_slicing and x.shape[0] > 1: + encoded_slices = [self._encode(x_slice) for x_slice in x.split(1)] + h = torch.cat(encoded_slices) + else: + h = self._encode(x) + + posterior = DiagonalGaussianDistribution(h) + + if not return_dict: + return (posterior,) + + return AutoencoderKLOutput(latent_dist=posterior) + + def _decode(self, z: torch.Tensor, return_dict: bool = True) -> Union[DecoderOutput, torch.Tensor]: + if self.use_tiling and (z.shape[-1] > self.tile_latent_min_size or z.shape[-2] > self.tile_latent_min_size): + return self.tiled_decode(z, return_dict=return_dict) + + if self.post_quant_conv is not None: + z = self.post_quant_conv(z) + + dec = self.decoder(z) + + if not return_dict: + return (dec,) + + return DecoderOutput(sample=dec) + + @apply_forward_hook + def decode( + self, z: torch.FloatTensor, return_dict: bool = True, generator=None + ) -> Union[DecoderOutput, torch.FloatTensor]: + """ + Decode a batch of images. + + Args: + z (`torch.Tensor`): Input batch of latent vectors. + return_dict (`bool`, *optional*, defaults to `True`): + Whether to return a [`~models.vae.DecoderOutput`] instead of a plain tuple. + + Returns: + [`~models.vae.DecoderOutput`] or `tuple`: + If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is + returned. + + """ + if self.use_slicing and z.shape[0] > 1: + decoded_slices = [self._decode(z_slice).sample for z_slice in z.split(1)] + decoded = torch.cat(decoded_slices) + else: + decoded = self._decode(z).sample + + if not return_dict: + return (decoded,) + + return DecoderOutput(sample=decoded) + + def blend_v(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor: + blend_extent = min(a.shape[2], b.shape[2], blend_extent) + for y in range(blend_extent): + b[:, :, y, :] = a[:, :, -blend_extent + y, :] * (1 - y / blend_extent) + b[:, :, y, :] * (y / blend_extent) + return b + + def blend_h(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor: + blend_extent = min(a.shape[3], b.shape[3], blend_extent) + for x in range(blend_extent): + b[:, :, :, x] = a[:, :, :, -blend_extent + x] * (1 - x / blend_extent) + b[:, :, :, x] * (x / blend_extent) + return b + + def _tiled_encode(self, x: torch.Tensor) -> torch.Tensor: + r"""Encode a batch of images using a tiled encoder. + + When this option is enabled, the VAE will split the input tensor into tiles to compute encoding in several + steps. This is useful to keep memory use constant regardless of image size. The end result of tiled encoding is + different from non-tiled encoding because each tile uses a different encoder. To avoid tiling artifacts, the + tiles overlap and are blended together to form a smooth output. You may still see tile-sized changes in the + output, but they should be much less noticeable. + + Args: + x (`torch.Tensor`): Input batch of images. + + Returns: + `torch.Tensor`: + The latent representation of the encoded videos. + """ + + overlap_size = int(self.tile_sample_min_size * (1 - self.tile_overlap_factor)) + blend_extent = int(self.tile_latent_min_size * self.tile_overlap_factor) + row_limit = self.tile_latent_min_size - blend_extent + + # Split the image into 512x512 tiles and encode them separately. + rows = [] + for i in range(0, x.shape[2], overlap_size): + row = [] + for j in range(0, x.shape[3], overlap_size): + tile = x[:, :, i : i + self.tile_sample_min_size, j : j + self.tile_sample_min_size] + tile = self.encoder(tile) + if self.config.use_quant_conv: + tile = self.quant_conv(tile) + row.append(tile) + rows.append(row) + result_rows = [] + for i, row in enumerate(rows): + result_row = [] + for j, tile in enumerate(row): + # blend the above tile and the left tile + # to the current tile and add the current tile to the result row + if i > 0: + tile = self.blend_v(rows[i - 1][j], tile, blend_extent) + if j > 0: + tile = self.blend_h(row[j - 1], tile, blend_extent) + result_row.append(tile[:, :, :row_limit, :row_limit]) + result_rows.append(torch.cat(result_row, dim=3)) + + enc = torch.cat(result_rows, dim=2) + return enc + + def tiled_encode(self, x: torch.Tensor, return_dict: bool = True) -> AutoencoderKLOutput: + r"""Encode a batch of images using a tiled encoder. + + When this option is enabled, the VAE will split the input tensor into tiles to compute encoding in several + steps. This is useful to keep memory use constant regardless of image size. The end result of tiled encoding is + different from non-tiled encoding because each tile uses a different encoder. To avoid tiling artifacts, the + tiles overlap and are blended together to form a smooth output. You may still see tile-sized changes in the + output, but they should be much less noticeable. + + Args: + x (`torch.Tensor`): Input batch of images. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~models.autoencoder_kl.AutoencoderKLOutput`] instead of a plain tuple. + + Returns: + [`~models.autoencoder_kl.AutoencoderKLOutput`] or `tuple`: + If return_dict is True, a [`~models.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain + `tuple` is returned. + """ + deprecation_message = ( + "The tiled_encode implementation supporting the `return_dict` parameter is deprecated. In the future, the " + "implementation of this method will be replaced with that of `_tiled_encode` and you will no longer be able " + "to pass `return_dict`. You will also have to create a `DiagonalGaussianDistribution()` from the returned value." + ) + deprecate("tiled_encode", "1.0.0", deprecation_message, standard_warn=False) + + overlap_size = int(self.tile_sample_min_size * (1 - self.tile_overlap_factor)) + blend_extent = int(self.tile_latent_min_size * self.tile_overlap_factor) + row_limit = self.tile_latent_min_size - blend_extent + + # Split the image into 512x512 tiles and encode them separately. + rows = [] + for i in range(0, x.shape[2], overlap_size): + row = [] + for j in range(0, x.shape[3], overlap_size): + tile = x[:, :, i : i + self.tile_sample_min_size, j : j + self.tile_sample_min_size] + tile = self.encoder(tile) + if self.config.use_quant_conv: + tile = self.quant_conv(tile) + row.append(tile) + rows.append(row) + result_rows = [] + for i, row in enumerate(rows): + result_row = [] + for j, tile in enumerate(row): + # blend the above tile and the left tile + # to the current tile and add the current tile to the result row + if i > 0: + tile = self.blend_v(rows[i - 1][j], tile, blend_extent) + if j > 0: + tile = self.blend_h(row[j - 1], tile, blend_extent) + result_row.append(tile[:, :, :row_limit, :row_limit]) + result_rows.append(torch.cat(result_row, dim=3)) + + moments = torch.cat(result_rows, dim=2) + posterior = DiagonalGaussianDistribution(moments) + + if not return_dict: + return (posterior,) + + return AutoencoderKLOutput(latent_dist=posterior) + + def tiled_decode(self, z: torch.Tensor, return_dict: bool = True) -> Union[DecoderOutput, torch.Tensor]: + r""" + Decode a batch of images using a tiled decoder. + + Args: + z (`torch.Tensor`): Input batch of latent vectors. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~models.vae.DecoderOutput`] instead of a plain tuple. + + Returns: + [`~models.vae.DecoderOutput`] or `tuple`: + If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is + returned. + """ + overlap_size = int(self.tile_latent_min_size * (1 - self.tile_overlap_factor)) + blend_extent = int(self.tile_sample_min_size * self.tile_overlap_factor) + row_limit = self.tile_sample_min_size - blend_extent + + # Split z into overlapping 64x64 tiles and decode them separately. + # The tiles have an overlap to avoid seams between tiles. + rows = [] + for i in range(0, z.shape[2], overlap_size): + row = [] + for j in range(0, z.shape[3], overlap_size): + tile = z[:, :, i : i + self.tile_latent_min_size, j : j + self.tile_latent_min_size] + if self.config.use_post_quant_conv: + tile = self.post_quant_conv(tile) + decoded = self.decoder(tile) + row.append(decoded) + rows.append(row) + result_rows = [] + for i, row in enumerate(rows): + result_row = [] + for j, tile in enumerate(row): + # blend the above tile and the left tile + # to the current tile and add the current tile to the result row + if i > 0: + tile = self.blend_v(rows[i - 1][j], tile, blend_extent) + if j > 0: + tile = self.blend_h(row[j - 1], tile, blend_extent) + result_row.append(tile[:, :, :row_limit, :row_limit]) + result_rows.append(torch.cat(result_row, dim=3)) + + dec = torch.cat(result_rows, dim=2) + if not return_dict: + return (dec,) + + return DecoderOutput(sample=dec) + + def forward( + self, + sample: torch.Tensor, + sample_posterior: bool = False, + return_dict: bool = True, + generator: Optional[torch.Generator] = None, + ) -> Union[DecoderOutput, torch.Tensor]: + r""" + Args: + sample (`torch.Tensor`): Input sample. + sample_posterior (`bool`, *optional*, defaults to `False`): + Whether to sample from the posterior. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`DecoderOutput`] instead of a plain tuple. + """ + x = sample + posterior = self.encode(x).latent_dist + if sample_posterior: + z = posterior.sample(generator=generator) + else: + z = posterior.mode() + dec = self.decode(z).sample + + if not return_dict: + return (dec,) + + return DecoderOutput(sample=dec) + + # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections + def fuse_qkv_projections(self): + """ + Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value) + are fused. For cross-attention modules, key and value projection matrices are fused. + + > [!WARNING] > This API is 🧪 experimental. + """ + self.original_attn_processors = None + + for _, attn_processor in self.attn_processors.items(): + if "Added" in str(attn_processor.__class__.__name__): + raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.") + + self.original_attn_processors = self.attn_processors + + for module in self.modules(): + if isinstance(module, Attention): + module.fuse_projections(fuse=True) + + self.set_attn_processor(FusedAttnProcessor2_0()) + + # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections + def unfuse_qkv_projections(self): + """Disables the fused QKV projection if enabled. + + > [!WARNING] > This API is 🧪 experimental. + + """ + if self.original_attn_processors is not None: + self.set_attn_processor(self.original_attn_processors) \ No newline at end of file diff --git a/videox_fun/models/flux_transformer2d.py b/videox_fun/models/flux_transformer2d.py index b8020b6..b77175b 100644 --- a/videox_fun/models/flux_transformer2d.py +++ b/videox_fun/models/flux_transformer2d.py @@ -178,140 +178,10 @@ class FluxAttnProcessor: return hidden_states -class FluxIPAdapterAttnProcessor(torch.nn.Module): - """Flux Attention processor for IP-Adapter.""" - - _attention_backend = None - - def __init__( - self, hidden_size: int, cross_attention_dim: int, num_tokens=(4,), scale=1.0, device=None, dtype=None - ): - super().__init__() - - if not hasattr(F, "scaled_dot_product_attention"): - raise ImportError( - f"{self.__class__.__name__} requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0." - ) - - self.hidden_size = hidden_size - self.cross_attention_dim = cross_attention_dim - - if not isinstance(num_tokens, (tuple, list)): - num_tokens = [num_tokens] - - if not isinstance(scale, list): - scale = [scale] * len(num_tokens) - if len(scale) != len(num_tokens): - raise ValueError("`scale` should be a list of integers with the same length as `num_tokens`.") - self.scale = scale - - self.to_k_ip = nn.ModuleList( - [ - nn.Linear(cross_attention_dim, hidden_size, bias=True, device=device, dtype=dtype) - for _ in range(len(num_tokens)) - ] - ) - self.to_v_ip = nn.ModuleList( - [ - nn.Linear(cross_attention_dim, hidden_size, bias=True, device=device, dtype=dtype) - for _ in range(len(num_tokens)) - ] - ) - - def __call__( - self, - attn: "FluxAttention", - hidden_states: torch.Tensor, - encoder_hidden_states: torch.Tensor = None, - attention_mask: Optional[torch.Tensor] = None, - image_rotary_emb: Optional[torch.Tensor] = None, - ip_hidden_states: Optional[List[torch.Tensor]] = None, - ip_adapter_masks: Optional[torch.Tensor] = None, - ) -> torch.Tensor: - batch_size = hidden_states.shape[0] - - query, key, value, encoder_query, encoder_key, encoder_value = _get_qkv_projections( - attn, hidden_states, encoder_hidden_states - ) - - query = query.unflatten(-1, (attn.heads, -1)) - key = key.unflatten(-1, (attn.heads, -1)) - value = value.unflatten(-1, (attn.heads, -1)) - - query = attn.norm_q(query) - key = attn.norm_k(key) - ip_query = query - - if encoder_hidden_states is not None: - encoder_query = encoder_query.unflatten(-1, (attn.heads, -1)) - encoder_key = encoder_key.unflatten(-1, (attn.heads, -1)) - encoder_value = encoder_value.unflatten(-1, (attn.heads, -1)) - - encoder_query = attn.norm_added_q(encoder_query) - encoder_key = attn.norm_added_k(encoder_key) - - query = torch.cat([encoder_query, query], dim=1) - key = torch.cat([encoder_key, key], dim=1) - value = torch.cat([encoder_value, value], dim=1) - - if image_rotary_emb is not None: - query = apply_rotary_emb(query, image_rotary_emb, sequence_dim=1) - key = apply_rotary_emb(key, image_rotary_emb, sequence_dim=1) - - hidden_states = attention( - query, - key, - value, - attn_mask=attention_mask, - dropout_p=0.0, - is_causal=False, - ) - hidden_states = hidden_states.flatten(2, 3) - hidden_states = hidden_states.to(query.dtype) - - if encoder_hidden_states is not None: - encoder_hidden_states, hidden_states = hidden_states.split_with_sizes( - [encoder_hidden_states.shape[1], hidden_states.shape[1] - encoder_hidden_states.shape[1]], dim=1 - ) - hidden_states = attn.to_out[0](hidden_states) - hidden_states = attn.to_out[1](hidden_states) - encoder_hidden_states = attn.to_add_out(encoder_hidden_states) - - # IP-adapter - ip_attn_output = torch.zeros_like(hidden_states) - - for current_ip_hidden_states, scale, to_k_ip, to_v_ip in zip( - ip_hidden_states, self.scale, self.to_k_ip, self.to_v_ip - ): - ip_key = to_k_ip(current_ip_hidden_states) - ip_value = to_v_ip(current_ip_hidden_states) - - ip_key = ip_key.view(batch_size, -1, attn.heads, attn.head_dim) - ip_value = ip_value.view(batch_size, -1, attn.heads, attn.head_dim) - - current_ip_hidden_states = dispatch_attention_fn( - ip_query, - ip_key, - ip_value, - attn_mask=None, - dropout_p=0.0, - is_causal=False, - backend=self._attention_backend, - ) - current_ip_hidden_states = current_ip_hidden_states.reshape(batch_size, -1, attn.heads * attn.head_dim) - current_ip_hidden_states = current_ip_hidden_states.to(ip_query.dtype) - ip_attn_output += scale * current_ip_hidden_states - - return hidden_states, encoder_hidden_states, ip_attn_output - else: - return hidden_states - - class FluxAttention(torch.nn.Module): _default_processor_cls = FluxAttnProcessor _available_processors = [ FluxAttnProcessor, - FluxIPAdapterAttnProcessor, ] def __init__( diff --git a/videox_fun/pipeline/__init__.py b/videox_fun/pipeline/__init__.py index 274ba40..703e2ac 100755 --- a/videox_fun/pipeline/__init__.py +++ b/videox_fun/pipeline/__init__.py @@ -3,6 +3,7 @@ from .pipeline_cogvideox_fun_control import CogVideoXFunControlPipeline from .pipeline_cogvideox_fun_inpaint import CogVideoXFunInpaintPipeline from .pipeline_fantasy_talking import FantasyTalkingPipeline from .pipeline_flux import FluxPipeline +from .pipeline_flux2 import Flux2Pipeline from .pipeline_hunyuanvideo import HunyuanVideoPipeline from .pipeline_hunyuanvideo_i2v import HunyuanVideoI2VPipeline from .pipeline_qwenimage import QwenImagePipeline diff --git a/videox_fun/pipeline/pipeline_flux2.py b/videox_fun/pipeline/pipeline_flux2.py new file mode 100644 index 0000000..26a8741 --- /dev/null +++ b/videox_fun/pipeline/pipeline_flux2.py @@ -0,0 +1,900 @@ +# https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/flux2/pipeline_flux2.py +# Copyright 2025 Black Forest Labs and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from diffusers.utils import (BaseOutput, is_torch_xla_available, logging, + replace_example_docstring) +from dataclasses import dataclass +from typing import Any, Callable, Dict, List, Optional, Tuple, Union + +import numpy as np +import PIL +import torch +from diffusers.pipelines.pipeline_utils import DiffusionPipeline +from diffusers.schedulers import FlowMatchEulerDiscreteScheduler +from diffusers.utils import (is_torch_xla_available, logging, + replace_example_docstring) +from diffusers.utils.torch_utils import randn_tensor + +from ..models import (AutoencoderKLFlux2, Flux2ImageProcessor, + Flux2Transformer2DModel, Mistral3ForConditionalGeneration, AutoProcessor) + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import Flux2Pipeline + + >>> pipe = Flux2Pipeline.from_pretrained("black-forest-labs/FLUX.2-dev", torch_dtype=torch.bfloat16) + >>> pipe.to("cuda") + >>> prompt = "A cat holding a sign that says hello world" + >>> # Depending on the variant being used, the pipeline call will slightly vary. + >>> # Refer to the pipeline documentation for more details. + >>> image = pipe(prompt, num_inference_steps=50, guidance_scale=2.5).images[0] + >>> image.save("flux.png") + ``` +""" + + +def format_text_input(prompts: List[str], system_message: str = None): + # Remove [IMG] tokens from prompts to avoid Pixtral validation issues + # when truncation is enabled. The processor counts [IMG] tokens and fails + # if the count changes after truncation. + cleaned_txt = [prompt.replace("[IMG]", "") for prompt in prompts] + + return [ + [ + { + "role": "system", + "content": [{"type": "text", "text": system_message}], + }, + {"role": "user", "content": [{"type": "text", "text": prompt}]}, + ] + for prompt in cleaned_txt + ] + + +def compute_empirical_mu(image_seq_len: int, num_steps: int) -> float: + a1, b1 = 8.73809524e-05, 1.89833333 + a2, b2 = 0.00016927, 0.45666666 + + if image_seq_len > 4300: + mu = a2 * image_seq_len + b2 + return float(mu) + + m_200 = a2 * image_seq_len + b2 + m_10 = a1 * image_seq_len + b1 + + a = (m_200 - m_10) / 190.0 + b = m_200 - 200.0 * a + mu = a * num_steps + b + + return float(mu) + + +# 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, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`List[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`List[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +@dataclass +class Flux2PipelineOutput(BaseOutput): + """ + Output class for Flux2 image generation pipelines. + + Args: + images (`List[PIL.Image.Image]` or `torch.Tensor` or `np.ndarray`) + List of denoised PIL images of length `batch_size` or numpy array or torch tensor of shape `(batch_size, + height, width, num_channels)`. PIL images or numpy array present the denoised images of the diffusion + pipeline. Torch tensors can represent either the denoised images or the intermediate latents ready to be + passed to the decoder. + """ + + images: Union[List[PIL.Image.Image], np.ndarray] + + +class Flux2Pipeline(DiffusionPipeline): + r""" + The Flux2 pipeline for text-to-image generation. + + Reference: [https://bfl.ai/blog/flux-2](https://bfl.ai/blog/flux-2) + + Args: + transformer ([`Flux2Transformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKLFlux2`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`Mistral3ForConditionalGeneration`]): + [Mistral3ForConditionalGeneration](https://huggingface.co/docs/transformers/en/model_doc/mistral3#transformers.Mistral3ForConditionalGeneration) + tokenizer (`AutoProcessor`): + Tokenizer of class + [PixtralProcessor](https://huggingface.co/docs/transformers/en/model_doc/pixtral#transformers.PixtralProcessor). + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKLFlux2, + text_encoder: Mistral3ForConditionalGeneration, + tokenizer: AutoProcessor, + transformer: Flux2Transformer2DModel, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + scheduler=scheduler, + transformer=transformer, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + # Flux latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible + # by the patch size. So the vae scale factor is multiplied by the patch size to account for this + self.image_processor = Flux2ImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + self.tokenizer_max_length = 512 + self.default_sample_size = 128 + + # fmt: off + self.system_message = "You are an AI that reasons about image descriptions. You give structured responses focusing on object relationships, object attribution and actions without speculation." + # fmt: on + + @staticmethod + def _get_mistral_3_small_prompt_embeds( + text_encoder: Mistral3ForConditionalGeneration, + tokenizer: AutoProcessor, + prompt: Union[str, List[str]], + dtype: Optional[torch.dtype] = None, + device: Optional[torch.device] = None, + max_sequence_length: int = 512, + # fmt: off + system_message: str = "You are an AI that reasons about image descriptions. You give structured responses focusing on object relationships, object attribution and actions without speculation.", + # fmt: on + hidden_states_layers: List[int] = (10, 20, 30), + ): + dtype = text_encoder.dtype if dtype is None else dtype + device = text_encoder.device if device is None else device + + prompt = [prompt] if isinstance(prompt, str) else prompt + + # Format input messages + messages_batch = format_text_input(prompts=prompt, system_message=system_message) + + # Process all messages at once + inputs = tokenizer.apply_chat_template( + messages_batch, + add_generation_prompt=False, + tokenize=True, + return_dict=True, + return_tensors="pt", + padding="max_length", + truncation=True, + max_length=max_sequence_length, + ) + + # Move to device + input_ids = inputs["input_ids"].to(device) + attention_mask = inputs["attention_mask"].to(device) + + # Forward pass through the model + output = text_encoder( + input_ids=input_ids, + attention_mask=attention_mask, + output_hidden_states=True, + use_cache=False, + ) + + # Only use outputs from intermediate layers and stack them + out = torch.stack([output.hidden_states[k] for k in hidden_states_layers], dim=1) + out = out.to(dtype=dtype, device=device) + + batch_size, num_channels, seq_len, hidden_dim = out.shape + prompt_embeds = out.permute(0, 2, 1, 3).reshape(batch_size, seq_len, num_channels * hidden_dim) + + return prompt_embeds + + @staticmethod + def _prepare_text_ids( + x: torch.Tensor, # (B, L, D) or (L, D) + t_coord: Optional[torch.Tensor] = None, + ): + B, L, _ = x.shape + out_ids = [] + + for i in range(B): + t = torch.arange(1) if t_coord is None else t_coord[i] + h = torch.arange(1) + w = torch.arange(1) + l = torch.arange(L) + + coords = torch.cartesian_prod(t, h, w, l) + out_ids.append(coords) + + return torch.stack(out_ids) + + def encode_prompt( + self, + prompt: Union[str, List[str]], + device: Optional[torch.device] = None, + num_images_per_prompt: int = 1, + prompt_embeds: Optional[torch.Tensor] = None, + max_sequence_length: int = 512, + text_encoder_out_layers: Tuple[int] = (10, 20, 30), + ): + device = device or self._execution_device + + if prompt is None: + prompt = "" + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt_embeds is None: + prompt_embeds = self._get_mistral_3_small_prompt_embeds( + text_encoder=self.text_encoder, + tokenizer=self.tokenizer, + prompt=prompt, + device=device, + max_sequence_length=max_sequence_length, + system_message=self.system_message, + hidden_states_layers=text_encoder_out_layers, + ) + + batch_size, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + text_ids = self._prepare_text_ids(prompt_embeds) + text_ids = text_ids.to(device) + return prompt_embeds, text_ids + + @staticmethod + def _prepare_latent_ids( + latents: torch.Tensor, # (B, C, H, W) + ): + r""" + Generates 4D position coordinates (T, H, W, L) for latent tensors. + + Args: + latents (torch.Tensor): + Latent tensor of shape (B, C, H, W) + + Returns: + torch.Tensor: + Position IDs tensor of shape (B, H*W, 4) All batches share the same coordinate structure: T=0, + H=[0..H-1], W=[0..W-1], L=0 + """ + + batch_size, _, height, width = latents.shape + + t = torch.arange(1) # [0] - time dimension + h = torch.arange(height) + w = torch.arange(width) + l = torch.arange(1) # [0] - layer dimension + + # Create position IDs: (H*W, 4) + latent_ids = torch.cartesian_prod(t, h, w, l) + + # Expand to batch: (B, H*W, 4) + latent_ids = latent_ids.unsqueeze(0).expand(batch_size, -1, -1) + + return latent_ids + + @staticmethod + def _prepare_image_ids( + image_latents: List[torch.Tensor], # [(1, C, H, W), (1, C, H, W), ...] + scale: int = 10, + ): + r""" + Generates 4D time-space coordinates (T, H, W, L) for a sequence of image latents. + + This function creates a unique coordinate for every pixel/patch across all input latent with different + dimensions. + + Args: + image_latents (List[torch.Tensor]): + A list of image latent feature tensors, typically of shape (C, H, W). + scale (int, optional): + A factor used to define the time separation (T-coordinate) between latents. T-coordinate for the i-th + latent is: 'scale + scale * i'. Defaults to 10. + + Returns: + torch.Tensor: + The combined coordinate tensor. Shape: (1, N_total, 4) Where N_total is the sum of (H * W) for all + input latents. + + Coordinate Components (Dimension 4): + - T (Time): The unique index indicating which latent image the coordinate belongs to. + - H (Height): The row index within that latent image. + - W (Width): The column index within that latent image. + - L (Seq. Length): A sequence length dimension, which is always fixed at 0 (size 1) + """ + + if not isinstance(image_latents, list): + raise ValueError(f"Expected `image_latents` to be a list, got {type(image_latents)}.") + + # create time offset for each reference image + t_coords = [scale + scale * t for t in torch.arange(0, len(image_latents))] + t_coords = [t.view(-1) for t in t_coords] + + image_latent_ids = [] + for x, t in zip(image_latents, t_coords): + x = x.squeeze(0) + _, height, width = x.shape + + x_ids = torch.cartesian_prod(t, torch.arange(height), torch.arange(width), torch.arange(1)) + image_latent_ids.append(x_ids) + + image_latent_ids = torch.cat(image_latent_ids, dim=0) + image_latent_ids = image_latent_ids.unsqueeze(0) + + return image_latent_ids + + @staticmethod + def _patchify_latents(latents): + batch_size, num_channels_latents, height, width = latents.shape + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 1, 3, 5, 2, 4) + latents = latents.reshape(batch_size, num_channels_latents * 4, height // 2, width // 2) + return latents + + @staticmethod + def _unpatchify_latents(latents): + batch_size, num_channels_latents, height, width = latents.shape + latents = latents.reshape(batch_size, num_channels_latents // (2 * 2), 2, 2, height, width) + latents = latents.permute(0, 1, 4, 2, 5, 3) + latents = latents.reshape(batch_size, num_channels_latents // (2 * 2), height * 2, width * 2) + return latents + + @staticmethod + def _pack_latents(latents): + """ + pack latents: (batch_size, num_channels, height, width) -> (batch_size, height * width, num_channels) + """ + + batch_size, num_channels, height, width = latents.shape + latents = latents.reshape(batch_size, num_channels, height * width).permute(0, 2, 1) + + return latents + + @staticmethod + def _unpack_latents_with_ids(x: torch.Tensor, x_ids: torch.Tensor) -> list[torch.Tensor]: + """ + using position ids to scatter tokens into place + """ + x_list = [] + for data, pos in zip(x, x_ids): + _, ch = data.shape # noqa: F841 + h_ids = pos[:, 1].to(torch.int64) + w_ids = pos[:, 2].to(torch.int64) + + h = torch.max(h_ids) + 1 + w = torch.max(w_ids) + 1 + + flat_ids = h_ids * w + w_ids + + out = torch.zeros((h * w, ch), device=data.device, dtype=data.dtype) + out.scatter_(0, flat_ids.unsqueeze(1).expand(-1, ch), data) + + # reshape from (H * W, C) to (H, W, C) and permute to (C, H, W) + + out = out.view(h, w, ch).permute(2, 0, 1) + x_list.append(out) + + return torch.stack(x_list, dim=0) + + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + if image.ndim != 4: + raise ValueError(f"Expected image dims 4, got {image.ndim}.") + + image_latents = retrieve_latents(self.vae.encode(image), generator=generator, sample_mode="argmax") + image_latents = self._patchify_latents(image_latents) + + latents_bn_mean = self.vae.bn.running_mean.view(1, -1, 1, 1).to(image_latents.device, image_latents.dtype) + latents_bn_std = torch.sqrt(self.vae.bn.running_var.view(1, -1, 1, 1) + self.vae.config.batch_norm_eps) + image_latents = (image_latents - latents_bn_mean) / latents_bn_std + + return image_latents + + def prepare_latents( + self, + batch_size, + num_latents_channels, + height, + width, + dtype, + device, + generator: torch.Generator, + latents: Optional[torch.Tensor] = None, + ): + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + + shape = (batch_size, num_latents_channels * 4, height // 2, width // 2) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device=device, dtype=dtype) + + latent_ids = self._prepare_latent_ids(latents) + latent_ids = latent_ids.to(device) + + latents = self._pack_latents(latents) # [B, C, H, W] -> [B, H*W, C] + return latents, latent_ids + + def prepare_image_latents( + self, + images: List[torch.Tensor], + batch_size, + generator: torch.Generator, + device, + dtype, + ): + image_latents = [] + for image in images: + image = image.to(device=device, dtype=dtype) + imagge_latent = self._encode_vae_image(image=image, generator=generator) + image_latents.append(imagge_latent) # (1, 128, 32, 32) + + image_latent_ids = self._prepare_image_ids(image_latents) + + # Pack each latent and concatenate + packed_latents = [] + for latent in image_latents: + # latent: (1, 128, 32, 32) + packed = self._pack_latents(latent) # (1, 1024, 128) + packed = packed.squeeze(0) # (1024, 128) - remove batch dim + packed_latents.append(packed) + + # Concatenate all reference tokens along sequence dimension + image_latents = torch.cat(packed_latents, dim=0) # (N*1024, 128) + image_latents = image_latents.unsqueeze(0) # (1, N*1024, 128) + + image_latents = image_latents.repeat(batch_size, 1, 1) + image_latent_ids = image_latent_ids.repeat(batch_size, 1, 1) + image_latent_ids = image_latent_ids.to(device) + + return image_latents, image_latent_ids + + def check_inputs( + self, + prompt, + height, + width, + prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if ( + height is not None + and height % (self.vae_scale_factor * 2) != 0 + or width is not None + and width % (self.vae_scale_factor * 2) != 0 + ): + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + image: Optional[Union[List[PIL.Image.Image], PIL.Image.Image]] = None, + prompt: Union[str, List[str]] = None, + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 50, + sigmas: Optional[List[float]] = None, + guidance_scale: Optional[float] = 4.0, + num_images_per_prompt: int = 1, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + attention_kwargs: Optional[Dict[str, Any]] = None, + callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + max_sequence_length: int = 512, + text_encoder_out_layers: Tuple[int] = (10, 20, 30), + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both + numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list + or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a + list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image + latents as `image`, but if passing latents directly it is not encoded again. + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + guidance_scale (`float`, *optional*, defaults to 1.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`List[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + 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. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.qwenimage.QwenImagePipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`. + text_encoder_out_layers (`Tuple[int]`): + Layer indices to use in the `text_encoder` to derive the final prompt embeddings. + + Examples: + + Returns: + [`~pipelines.flux2.Flux2PipelineOutput`] or `tuple`: [`~pipelines.flux2.Flux2PipelineOutput`] if + `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the + generated images. + """ + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt=prompt, + height=height, + width=width, + prompt_embeds=prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + ) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. prepare text embeddings + prompt_embeds, text_ids = self.encode_prompt( + prompt=prompt, + prompt_embeds=prompt_embeds, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + text_encoder_out_layers=text_encoder_out_layers, + ) + + # 4. process images + if image is not None and not isinstance(image, list): + image = [image] + + condition_images = None + if image is not None: + for img in image: + self.image_processor.check_image_input(img) + + condition_images = [] + for img in image: + image_width, image_height = img.size + if image_width * image_height > 1024 * 1024: + img = self.image_processor._resize_to_target_area(img, 1024 * 1024) + image_width, image_height = img.size + + multiple_of = self.vae_scale_factor * 2 + image_width = (image_width // multiple_of) * multiple_of + image_height = (image_height // multiple_of) * multiple_of + img = self.image_processor.preprocess(img, height=image_height, width=image_width, resize_mode="crop") + condition_images.append(img) + height = height or image_height + width = width or image_width + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + # 5. prepare latent variables + num_channels_latents = self.transformer.config.in_channels // 4 + latents, latent_ids = self.prepare_latents( + batch_size=batch_size * num_images_per_prompt, + num_latents_channels=num_channels_latents, + height=height, + width=width, + dtype=prompt_embeds.dtype, + device=device, + generator=generator, + latents=latents, + ) + + image_latents = None + image_latent_ids = None + if condition_images is not None: + image_latents, image_latent_ids = self.prepare_image_latents( + images=condition_images, + batch_size=batch_size * num_images_per_prompt, + generator=generator, + device=device, + dtype=self.vae.dtype, + ) + + # 6. Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + if hasattr(self.scheduler.config, "use_flow_sigmas") and self.scheduler.config.use_flow_sigmas: + sigmas = None + image_seq_len = latents.shape[1] + mu = compute_empirical_mu(image_seq_len=image_seq_len, num_steps=num_inference_steps) + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + sigmas=sigmas, + mu=mu, + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # handle guidance + guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32) + guidance = guidance.expand(latents.shape[0]) + + # 7. Denoising loop + # We set the index here to remove DtoH sync, helpful especially during compilation. + # Check out more details here: https://github.com/huggingface/diffusers/pull/11696 + self.scheduler.set_begin_index(0) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]).to(latents.dtype) + + latent_model_input = latents.to(self.transformer.dtype) + latent_image_ids = latent_ids + + if image_latents is not None: + latent_model_input = torch.cat([latents, image_latents], dim=1).to(self.transformer.dtype) + latent_image_ids = torch.cat([latent_ids, image_latent_ids], dim=1) + + noise_pred = self.transformer( + hidden_states=latent_model_input, # (B, image_seq_len, C) + timestep=timestep / 1000, + guidance=guidance, + encoder_hidden_states=prompt_embeds, + txt_ids=text_ids, # B, text_seq_len, 4 + img_ids=latent_image_ids, # B, image_seq_len, 4 + joint_attention_kwargs=self._attention_kwargs, + return_dict=False, + )[0] + + noise_pred = noise_pred[:, : latents.size(1) :] + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + + if output_type == "latent": + image = latents + else: + torch.save({"pred": latents}, "pred_d.pt") + latents = self._unpack_latents_with_ids(latents, latent_ids) + + latents_bn_mean = self.vae.bn.running_mean.view(1, -1, 1, 1).to(latents.device, latents.dtype) + latents_bn_std = torch.sqrt(self.vae.bn.running_var.view(1, -1, 1, 1) + self.vae.config.batch_norm_eps).to( + latents.device, latents.dtype + ) + latents = latents * latents_bn_std + latents_bn_mean + latents = self._unpatchify_latents(latents) + + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return Flux2PipelineOutput(images=image) \ No newline at end of file diff --git a/videox_fun/utils/lora_utils.py b/videox_fun/utils/lora_utils.py index 4853a01..e4001da 100755 --- a/videox_fun/utils/lora_utils.py +++ b/videox_fun/utils/lora_utils.py @@ -156,7 +156,7 @@ class LoRANetwork(torch.nn.Module): "CogVideoXTransformer3DModel", "WanTransformer3DModel", \ "Wan2_2Transformer3DModel", "FluxTransformer2DModel", "QwenImageTransformer2DModel", \ "Wan2_2Transformer3DModel_Animate", "Wan2_2Transformer3DModel_S2V", "FantasyTalkingTransformer3DModel", \ - "HunyuanVideoTransformer3DModel" + "HunyuanVideoTransformer3DModel", "Flux2Transformer2DModel" ] TEXT_ENCODER_TARGET_REPLACE_MODULE = ["T5LayerSelfAttention", "T5LayerFF", "BertEncoder", "T5SelfAttention", "T5CrossAttention"] LORA_PREFIX_TRANSFORMER = "lora_unet"