aigc-apps-VideoX-Fun/scripts/qwenimage/README_TRAIN.md

Qwen-Image Full Parameter Training Guide

This document provides a complete workflow for full parameter training of Qwen-Image Diffusion Transformer, including environment configuration, data preparation, distributed training, and inference testing.

---

Table of Contents

• 1. Environment Configuration
• 2. Data Preparation
  • 2.1 Quick Test Dataset
  • 2.2 Dataset Structure
  • 2.3 metadata.json Format
  • 2.4 Relative vs Absolute Path Usage
• 3. Full Parameter Training
  • 3.1 Download Pretrained Model
  • 3.2 Quick Start (DeepSpeed-Zero-2)
  • 3.3 Common Training Parameters
  • 3.4 Training Validation
  • 3.5 Training with FSDP
  • 3.6 Other Backends
  • 3.7 Multi-Machine Distributed Training
• 4. Inference Testing
  • 4.1 Inference Parameter Parsing
  • 4.2 Single GPU Inference
  • 4.3 Multi-GPU Parallel Inference
• 5. Additional Resources

---

1. Environment Configuration

Method 1: Using requirements.txt

pip install -r requirements.txt

Method 2: Manual Dependency Installation

pip install Pillow einops safetensors timm tomesd librosa "torch>=2.1.2" torchdiffeq torchsde decord datasets numpy scikit-image
pip install omegaconf SentencePiece imageio[ffmpeg] imageio[pyav] tensorboard beautifulsoup4 ftfy func_timeout onnxruntime
pip install "peft>=0.17.0" "accelerate>=0.25.0" "gradio>=3.41.2" "diffusers>=0.30.1" "transformers>=4.46.2"
pip install yunchang xfuser modelscope openpyxl
pip uninstall opencv-python opencv-contrib-python opencv-python-headless -y
pip install opencv-python-headless
pip install deepspeed==0.17.0 numpy==1.26.4

Method 3: Using Docker

When using Docker, please ensure that the GPU driver and CUDA environment are correctly installed on your machine, then execute the following commands:

# pull image
docker pull mybigpai-public-registry.cn-beijing.cr.aliyuncs.com/easycv/torch_cuda:cogvideox_fun

# enter image
docker run -it -p 7860:7860 --network host --gpus all --security-opt seccomp:unconfined --shm-size 200g mybigpai-public-registry.cn-beijing.cr.aliyuncs.com/easycv/torch_cuda:cogvideox_fun

---

2. Data Preparation

2.1 Quick Test Dataset

We provide a test dataset containing several training samples.

# Download official demo dataset
modelscope download --dataset PAI/X-Fun-Images-Demo --local_dir ./datasets/X-Fun-Images-Demo

2.2 Dataset Structure

📦 datasets/
├── 📂 my_dataset/
│   ├── 📂 train/
│   │   ├── 📄 image001.jpg
│   │   ├── 📄 image002.png
│   │   └── 📄 ...
│   └── 📄 metadata.json

2.3 metadata.json Format

Relative Path Format (example):

[
  {
    "file_path": "train/image001.jpg",
    "text": "A beautiful sunset over the ocean, golden hour lighting",
    "width": 1024,
    "height": 1024
  },
  {
    "file_path": "train/image002.png",
    "text": "Portrait of a young woman, studio lighting, high quality",
    "width": 1328,
    "height": 1328
  }
]

Absolute Path Format:

[
  {
    "file_path": "/mnt/data/images/sunset.jpg",
    "text": "A beautiful sunset over the ocean",
    "width": 1024,
    "height": 1024
  }
]

Key Fields Description:

• file_path: Image path (relative or absolute)
• text: Image description (English prompt)
• width / height: Image dimensions (recommended to provide for bucket training; if not provided, they will be automatically read during training, which may slow down training when data is stored on slow systems like OSS)
  • You can use scripts/process_json_add_width_and_height.py to add width and height fields to JSON files without these fields, supporting both images and videos
  • Usage: python scripts/process_json_add_width_and_height.py --input_file datasets/X-Fun-Images-Demo/metadata.json --output_file datasets/X-Fun-Images-Demo/metadata_add_width_height.json

2.4 Relative vs Absolute Path Usage

Relative Paths:

If your data uses relative paths, configure the training script as follows:

export DATASET_NAME="datasets/X-Fun-Images-Demo/"
export DATASET_META_NAME="datasets/X-Fun-Images-Demo/metadata_add_width_height.json"

Absolute Paths:

If your data uses absolute paths, configure the training script as follows:

export DATASET_NAME=""
export DATASET_META_NAME="/mnt/data/metadata_add_width_height.json"

💡 Recommendation: If the dataset is small and stored locally, use relative paths. If the dataset is stored on external storage (e.g., NAS, OSS) or shared across multiple machines, use absolute paths.

---

3. Full Parameter Training

3.1 Download Pretrained Model

# Create model directory
mkdir -p models/Diffusion_Transformer

# Download Qwen-Image official weights
modelscope download --model Qwen/Qwen-Image --local_dir models/Diffusion_Transformer/Qwen-Image

3.2 Quick Start (DeepSpeed-Zero-2)

If you have downloaded the data as per 2.1 Quick Test Dataset and the weights as per 3.1 Download Pretrained Model, you can directly copy and run the quick start command.

DeepSpeed-Zero-2 and FSDP are recommended for training. Here we use DeepSpeed-Zero-2 as an example.

The difference between DeepSpeed-Zero-2 and FSDP lies in whether the model weights are sharded. If VRAM is insufficient when using multiple GPUs with DeepSpeed-Zero-2, you can switch to FSDP.

export MODEL_NAME="models/Diffusion_Transformer/Qwen-Image"
export DATASET_NAME="datasets/X-Fun-Images-Demo/"
export DATASET_META_NAME="datasets/X-Fun-Images-Demo/metadata_add_width_height.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/qwenimage/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_qwenimage" \
  --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 "."

3.3 Common Training Parameters

Key Parameter Descriptions:

Parameter	Description	Example Value
--pretrained_model_name_or_path	Path to pretrained model	models/Diffusion_Transformer/Qwen-Image
--train_data_dir	Training data directory	datasets/X-Fun-Images-Demo/
--train_data_meta	Training data metadata file	datasets/X-Fun-Images-Demo/metadata_add_width_height.json
--train_batch_size	Samples per batch	1
--image_sample_size	Maximum training resolution, auto bucketing	1328
--gradient_accumulation_steps	Gradient accumulation steps (equivalent to larger batch)	1
--dataloader_num_workers	DataLoader subprocesses	8
--num_train_epochs	Number of training epochs	100
--checkpointing_steps	Save checkpoint every N steps	50
--learning_rate	Initial learning rate	2e-05
--lr_scheduler	Learning rate scheduler	constant_with_warmup
--lr_warmup_steps	Learning rate warmup steps	100
--seed	Random seed	42
--output_dir	Output directory	output_dir
--gradient_checkpointing	Enable activation checkpointing	-
--mixed_precision	Mixed precision: fp16/bf16	bf16
--adam_weight_decay	AdamW weight decay	3e-2
--adam_epsilon	AdamW epsilon value	1e-10
--vae_mini_batch	Mini-batch size for VAE encoding	1
--max_grad_norm	Gradient clipping threshold	0.05
--enable_bucket	Enable bucket training: trains entire images grouped by resolution without center cropping	-
--random_hw_adapt	Auto-scale images to random size in range [512, image_sample_size]	-
--resume_from_checkpoint	Resume training from checkpoint path, use "latest" to auto-select latest	None
--uniform_sampling	Uniform timestep sampling	-
--trainable_modules	Trainable modules ("." means all modules)	"."
--validation_steps	Execute validation every N steps	100
--validation_epochs	Execute validation every N epochs	100
--validation_prompts	Prompts used during validation	"1girl, black_hair, ..."

3.4 Training Validation

You can configure validation parameters to periodically generate test images during training, allowing you to monitor training progress and model quality.

Validation Parameters:

accelerate launch --use_deepspeed --deepspeed_config_file config/zero_stage2_config.json --deepspeed_multinode_launcher standard scripts/qwenimage/train.py \
  # ... (other training parameters)
  --validation_steps=100 \
  --validation_epochs=100 \
  --validation_prompts="1girl, black_hair, brown_eyes, earrings, freckles, grey_background, jewelry, lips, long_hair, looking_at_viewer, nose, piercing, realistic, red_lips, solo, upper_body"

Parameter Descriptions:

Parameter	Description	Recommended Value
--validation_steps	Execute validation every N steps. If your dataset is large and you want to save validation time, you can set a larger value (e.g., 100 or 500)	100
--validation_epochs	Execute validation every N epochs	100
--validation_prompts	Prompt for validation image generation. Use multiple space-separated prompt strings	Space-separated prompt strings

Notes:

• Validation images will be saved to the output_dir directory
• Setting --validation_steps=1 means validation is performed every step, which may slow down training. Adjust according to your needs
• For multi-prompt validation, use: --validation_prompts "prompt1" "prompt2" "prompt3"

3.5 Training with FSDP

If VRAM is insufficient when using multiple GPUs with DeepSpeed-Zero-2, you can switch to FSDP.

export MODEL_NAME="models/Diffusion_Transformer/Qwen-Image"
export DATASET_NAME="datasets/X-Fun-Images-Demo/"
export DATASET_META_NAME="datasets/X-Fun-Images-Demo/metadata_add_width_height.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=QwenImageTransformerBlock --fsdp_sharding_strategy "FULL_SHARD" --fsdp_state_dict_type=SHARDED_STATE_DICT --fsdp_backward_prefetch "BACKWARD_PRE" --fsdp_cpu_ram_efficient_loading False scripts/qwenimage/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_qwenimage" \
  --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 "."

3.6 Other Backends

3.6.1 Training with DeepSpeed-Zero-3

DeepSpeed Zero-3 is not highly recommended at the moment. In this repository, using FSDP has fewer errors and is more stable.

DeepSpeed Zero-3:

After training, you can use the following command to get the final model:

python scripts/zero_to_bf16.py output_dir/checkpoint-{our-num-steps} output_dir/checkpoint-{your-num-steps}-outputs --max_shard_size 80GB --safe_serialization

Training shell command:

export MODEL_NAME="models/Diffusion_Transformer/Qwen-Image"
export DATASET_NAME="datasets/X-Fun-Images-Demo/"
export DATASET_META_NAME="datasets/X-Fun-Images-Demo/metadata_add_width_height.json"
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA. 
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO

accelerate launch --zero_stage 3 --zero3_save_16bit_model true --zero3_init_flag true --use_deepspeed --deepspeed_config_file config/zero_stage3_config.json --deepspeed_multinode_launcher standard scripts/qwenimage/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 "."

3.6.2 Training Without DeepSpeed or FSDP

This approach is not recommended as it lacks VRAM-saving backends and may easily cause out-of-memory errors. This is provided for reference only.

export MODEL_NAME="models/Diffusion_Transformer/Qwen-Image"
export DATASET_NAME="datasets/X-Fun-Images-Demo/"
export DATASET_META_NAME="datasets/X-Fun-Images-Demo/metadata_add_width_height.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/qwenimage/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_qwenimage" \
  --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 "."

3.7 Multi-Machine Distributed Training

Suitable for: Ultra-large-scale datasets, faster training speed

3.7.1 Environment Configuration

Assuming 2 machines with 8 GPUs each:

Machine 0 (Master):

export MODEL_NAME="models/Diffusion_Transformer/Qwen-Image"
export DATASET_NAME="datasets/X-Fun-Images-Demo/"
export DATASET_META_NAME="datasets/X-Fun-Images-Demo/metadata_add_width_height.json"
export MASTER_ADDR="192.168.1.100"  # Master machine IP
export MASTER_PORT=10086
export WORLD_SIZE=2                  # Total number of machines
export NUM_PROCESS=16                # Total processes = machines × 8
export RANK=0                        # Current machine rank (0 or 1)
# 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" --main_process_ip=$MASTER_ADDR --main_process_port=$MASTER_PORT --num_machines=$WORLD_SIZE --num_processes=$NUM_PROCESS --machine_rank=$RANK --use_deepspeed --deepspeed_config_file config/zero_stage2_config.json --deepspeed_multinode_launcher standard scripts/qwenimage/train.py \
  --pretrained_model_name_or_path=$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_qwenimage" \
  --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 "."

Machine 1 (Worker):

export MODEL_NAME="models/Diffusion_Transformer/Qwen-Image"
export DATASET_NAME="datasets/X-Fun-Images-Demo/"
export DATASET_META_NAME="datasets/X-Fun-Images-Demo/metadata_add_width_height.json"
export MASTER_ADDR="192.168.1.100"  # Same as Master
export MASTER_PORT=10086
export WORLD_SIZE=2
export NUM_PROCESS=16
export RANK=1  # Note this is 1
# 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

# Use the same accelerate launch command as Machine 0

3.7.2 Multi-Machine Training Notes

• Network Requirements:

  • RDMA/InfiniBand recommended (high performance)
  • Without RDMA, add environment variables:

    export NCCL_IB_DISABLE=1
    export NCCL_P2P_DISABLE=1
    

• Data Synchronization: All machines must be able to access the same data paths (NFS/shared storage)

4. Inference Testing

4.1 Inference Parameter Parsing

Key Parameter Descriptions:

Parameter	Description	Example Value
GPU_memory_mode	VRAM management mode, see table below for options	model_group_offload
ulysses_degree	Head dimension parallelism degree, set to 1 for single GPU	1
ring_degree	Sequence dimension parallelism degree, set to 1 for single GPU	1
fsdp_dit	Use FSDP for Transformer during multi-GPU inference to save VRAM	False
fsdp_text_encoder	Use FSDP for text encoder during multi-GPU inference	False
compile_dit	Compile Transformer for faster inference (effective at fixed resolution)	False
enable_teacache	Enable TeaCache for faster inference	True
teacache_threshold	TeaCache threshold, recommended 0.05~0.30, higher is faster but may reduce quality	0.25
num_skip_start_steps	Number of steps to skip at inference start, reduces impact on generation quality	5
teacache_offload	Offload TeaCache tensors to CPU to save VRAM	False
cfg_skip_ratio	Skip some CFG steps for faster inference, recommended 0.00~0.25	0
model_name	Model path	models/Diffusion_Transformer/Qwen-Image
sampler_name	Sampler type: Flow, Flow_Unipc, Flow_DPM++	Flow
transformer_path	Path to load trained Transformer weights	None
vae_path	Path to load trained VAE weights	None
lora_path	LoRA weights path	None
sample_size	Generated image resolution [height, width]	[1344, 768]
weight_dtype	Model weight precision, use torch.float16 for GPUs without bf16 support	torch.bfloat16
prompt	Positive prompt describing the generation content	"1girl, black_hair..."
negative_prompt	Negative prompt for content to avoid	" "
guidance_scale	Guidance strength	4.0
seed	Random seed for reproducible results	43
num_inference_steps	Number of inference steps	50
lora_weight	LoRA weight strength	0.55
save_path	Path to save generated images	samples/qwenimage-t2i

VRAM Management Mode Description:

Mode	Description	VRAM Usage
model_full_load	Load entire model to GPU	Highest
model_full_load_and_qfloat8	Full load + FP8 quantization	High
model_cpu_offload	Offload model to CPU after use	Medium
model_cpu_offload_and_qfloat8	CPU offload + FP8 quantization	Medium-Low
model_group_offload	Layer groups switch between CPU/CUDA	Low
sequential_cpu_offload	Sequential layer offload (slowest)	Lowest

4.2 Single GPU Inference

Quick Start

Run the following command for single GPU inference:

python examples/qwenimage/predict_t2i.py

Edit examples/qwenimage/predict_t2i.py according to your needs. For first-time inference, focus on these parameters. For other parameters, refer to the inference parameter parsing above.

# Choose based on GPU VRAM
GPU_memory_mode = "model_group_offload"
# Based on actual model path
model_name = "models/Diffusion_Transformer/Qwen-Image"  
# Path to trained weights, e.g., "output_dir_qwenimage/checkpoint-xxx/diffusion_pytorch_model.safetensors"
transformer_path = None  
# Write based on generation content
prompt = "a young girl with flowing long hair, wearing a white halter dress"  
# ...

4.3 Multi-GPU Parallel Inference

Suitable for: High-resolution generation, accelerated inference

Install Parallel Inference Dependencies

pip install xfuser==0.4.2 yunchang==0.6.2

Configure Parallel Strategy

Edit examples/qwenimage/predict_t2i.py:

# Ensure ulysses_degree × ring_degree = number of GPUs
# For example, using 2 GPUs:
ulysses_degree = 2  # Head dimension parallelization
ring_degree = 1     # Sequence dimension parallelization

Configuration Principles:

• ulysses_degree must evenly divide the model's number of heads
• ring_degree splits on sequence dimension, affecting communication overhead; avoid using it when heads can be divided

Example Configurations:

GPU Count	ulysses_degree	ring_degree	Description
1	1	1	Single GPU
4	4	1	Head parallelization
8	8	1	Head parallelization
8	4	2	Hybrid parallelization

Run Multi-GPU Inference

torchrun --nproc-per-node=2 examples/qwenimage/predict_t2i.py

5. Additional Resources

• Official GitHub: https://github.com/aigc-apps/VideoX-Fun