# 增加软件包可执行权限，{version}表示软件版本号，{arch}表示CPU架构，{soc}表示昇腾AI处理器的版本。
chmod +x ./Ascend-cann-toolkit_{version}_linux-{arch}.run
chmod +x ./Ascend-cann-kernels-{soc}_{version}_linux.run
# 校验软件包安装文件的一致性和完整性
./Ascend-cann-toolkit_{version}_linux-{arch}.run --check
./Ascend-cann-kernels-{soc}_{version}_linux.run --check
# 安装
./Ascend-cann-toolkit_{version}_linux-{arch}.run --install
./Ascend-cann-kernels-{soc}_{version}_linux.run --install

# 设置环境变量
source /usr/local/Ascend/ascend-toolkit/set_env.sh

# 增加软件包可执行权限，{version}表示软件版本号，{arch}表示CPU架构。
chmod +x ./Ascend-mindie_${version}_linux-${arch}.run
./Ascend-mindie_${version}_linux-${arch}.run --check

# 方式一：默认路径安装
./Ascend-mindie_${version}_linux-${arch}.run --install
# 设置环境变量
cd /usr/local/Ascend/mindie && source set_env.sh

# 方式二：指定路径安装
./Ascend-mindie_${version}_linux-${arch}.run --install-path=${AieInstallPath}
# 设置环境变量
cd ${AieInstallPath}/mindie && source set_env.sh

tar -xzvf pytorch_v{pytorchversion}_py{pythonversion}.tar.gz
# 解压后，会有whl包
pip install torch_npu-{pytorchversion}.xxxx.{arch}.whl

https://huggingface.co/Qwen/Qwen-Image-Edit-2509

#  1. 下载代码
git clone https://modelers.cn/MindIE/Qwen-Image-Edit-2509.git && cd Qwen-Image-Edit-2509

# 2. python相关依赖安装
pip install diffusers==0.35.1
pip install transformers==4.52.4
pip install yunchang==0.6.0

# 等价优化
export ROPE_FUSE=1
export ADALN_FUSE=1

python run_edit_2509.py  \
  --model_path ./Qwen-Image-Edit-2509  \
  --device_id 0  \
  --img_paths ./yarn-art-pikachu.png \
  --prompt_file "./edit_prompts.txt" \
  --width 1024 \
  --height 1024 \
  --vae_tiling \
  --vae_slicing \

# 等价优化
export ROPE_FUSE=1
export ADALN_FUSE=1
# cache算法优化
export COND_CACHE=1
export UNCOND_CACHE=1

python run_edit_2509.py  \
  --model_path ./Qwen-Image-Edit-2509  \
  --device_id 0  \
  --img_paths ./yarn-art-pikachu.png
  --prompt_file "./edit_prompts.txt" \
  --width 1024 \
  --height 1024 \
  --vae_tiling \
  --vae_slicing \

export LCCL_DETERMINISTIC=true  
export HCCL_DETERMINISTIC=true  
export ATB_MATMUL_SHUFFLE_K_ENABLE=0  
export ATB_LLM_LCOC_ENABLE=true      
export CLOSE_MATMUL_K_SHIFT=true     

model_path="/mnt/weights/Qwen-Image-Edit-2509"

# 等价优化  需要时开启
# export ROPE_FUSE=1
# export ADALN_FUSE=1

# cache算法优化 需要时开启
# export COND_CACHE=1
# export UNCOND_CACHE=1

# 等价优化+cache算法优化  需要时开启
# export ROPE_FUSE=1
# export ADALN_FUSE=1
# export COND_CACHE=1
# export UNCOND_CACHE=1

# 8卡并行，cfg_size * ulysses_size = 8
# (cfg_size=2, ulysses_size=4) 优于 (cfg_size=1, ulysses_size=8)
export ASCEND_RT_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
# export ASCEND_RT_VISIBLE_DEVICES=8,9,10,11,12,13,14,15
torchrun --nproc_per_node=8 --master-port 29508 run_edit_2509_cfg_usp.py \
    --model_path ${model_path} \
    --num_inference_steps 50 \
    --seed 42 \
    --output_dir "baseline_cfg2_ulysses4_optimize_1+2_50steps" \
    --ulysses_size 4 \
    --cfg_size 2 \
    --img_paths ./yarn-art-pikachu.png \
    --prompt_file "./edit_prompts.txt" \
    --width 1024 \
    --height 1024 \
    --vae_tiling \
    --vae_slicing \

export LCCL_DETERMINISTIC=true  
export HCCL_DETERMINISTIC=true  
export ATB_MATMUL_SHUFFLE_K_ENABLE=0  
export ATB_LLM_LCOC_ENABLE=true      
export CLOSE_MATMUL_K_SHIFT=true     

model_path="/mnt/weights/Qwen-Image-Edit-2509"

# 等价优化  需要时开启
# export ROPE_FUSE=1
# export ADALN_FUSE=1

# cache算法优化 需要时开启
# export COND_CACHE=1
# export UNCOND_CACHE=1

# 等价优化+cache算法优化  需要时开启
# export ROPE_FUSE=1
# export ADALN_FUSE=1
# export COND_CACHE=1
# export UNCOND_CACHE=1

# 8卡并行，cfg_size * ulysses_size = 16
export ASCEND_RT_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
# export ASCEND_RT_VISIBLE_DEVICES=8,9,10,11,12,13,14,15
torchrun --nproc_per_node=8 --master-port 29508 run_edit_2509_cfg_usp.py \
    --model_path ${model_path} \
    --num_inference_steps 50 \
    --seed 42 \
    --output_dir "baseline_cfg2_ulysses4_optimize_1+2_50steps" \
    --ulysses_size 8 \
    --cfg_size 2 \
    --img_paths ./yarn-art-pikachu.png \
    --prompt_file "./edit_prompts.txt" \
    --width 1024 \
    --height 1024 \
    --vae_tiling \
    --vae_slicing \

export PYTHONPATH=".:$PYTHONPATH"  # 添加当前目录到Python搜索路径
model_path="/mnt/weights/Qwen-Image-Edit-2509"
# w8a8量化
python quantization/quant.py \
  --model_path ${model_path} \
  --device_id 0 \
  --quant_mode w8a8 \
  --w_sym \
  --is_dynamic \
  --act_method 3 \
  --quant_save_dir ./quant_w8a8_dynamic_withoutData_use_disable_quant_layers

export PYTHONPATH=".:$PYTHONPATH"  # 添加当前目录到Python搜索路径
model_path="/mnt/weights/Qwen-Image-Edit-2509"
# w8a16量化
python quantization/quant.py \
  --model_path ${model_path} \
  --device_id 0 \
  --quant_mode w8a16 \
  --w_sym \
  --act_method 3 \
  --quant_save_dir ./quant_w8a16_withoutData_use_disable_quant_layers

# 增加软件包可执行权限，{version}表示软件版本号，{arch}表示CPU架构。
chmod +x Ascend-cann-nnal_<version>_linux-<arch>.run
# 默认路径安装:
./Ascend-cann-nnal_<version>_linux-<arch>.run --install --torch_atb
# 配置环境变量:
source ${HOME}/Ascend/nnal/atb/set_env.sh

model_path="/mnt/weights/Qwen-Image-Edit-2509"
quant_desc_path="/home/Qwen-Image-Edit-2509/quant_w8a8_dynamic_withoutData_use_disable_quant_layers/quant_model_description_w8a8_dynamic.json"

# 算子优化  需要时开启
# export ROPE_FUSE=1
# export ADALN_FUSE=1

# cache 算法优化  需要时开启
# export COND_CACHE=1
# export UNCOND_CACHE=1

python run_edit_2509.py  \
    --model_path ${model_path}  \
    --device_id 0 \
    --quant_desc_path ${quant_desc_path} \
    --img_paths ./yarn-art-pikachu.png
    --prompt_file "./edit_prompts.txt" \
    --width 1024 \
    --height 1024 \

export LCCL_DETERMINISTIC=true  
export HCCL_DETERMINISTIC=true  
export ATB_MATMUL_SHUFFLE_K_ENABLE=0  
export ATB_LLM_LCOC_ENABLE=true      
export CLOSE_MATMUL_K_SHIFT=true  

model_path="/mnt/weights/Qwen-Image-Edit-2509"
quant_desc_path="/home/Qwen-Image-Edit-2509/quant_w8a8_dynamic_withoutData_use_disable_quant_layers/quant_model_description_w8a8_dynamic.json"

# 算子优化  需要时开启
# export ROPE_FUSE=1
# export ADALN_FUSE=1

# 算法优化 需要时开启
# export COND_CACHE=1
# export UNCOND_CACHE=1

# 8卡并行
export ASCEND_RT_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
torchrun --nproc_per_node=8 --master-port 29508 run_edit_2509_cfg_usp.py \
    --model_path ${model_path} \
    --num_inference_steps 50 \
    --seed 42 \
    --output_dir "baseline_cfg2_ulysses4_optimize_1+2_50steps" \
    --ulysses_size 4 \
    --cfg_size 2 \
    --quant_desc_path ${quant_desc_path} \
    --img_paths ./yarn-art-pikachu.png \
    --prompt_file "./edit_prompts.txt" \
    --width 1024 \
    --height 1024 \

model_path="/home/weight/Qwen-Image-Edit-2509/" 
quant_desc_path="/home/Qwen-Image-Edit-2509/quant_w8a8_withoutData_use_disable_quant_layers/quant_model_description_w8a8_dynamic.json"

# 算子优化  需要时开启
# export ROPE_FUSE=1
# export ADALN_FUSE=1

# cache 算法优化  需要时开启
# export COND_CACHE=1
# export UNCOND_CACHE=1

# GEdit-Bench数据集的使用示例
dataset_path="/mnt/datasets/GEdit-Bench"
python run_edit_2509.py \
    --model_path ${model_path} \
    --device_id 0 \
    --dataset_name "GEdit-Bench" \
    --dataset_path ${dataset_path} \
    --num_inference_steps 50 \
    --seed 42 \
    --output_dir "GEdit-Bench_baseline" \
    --vae_tiling \
    --vae_slicing \
    # 有量化权重时候，使用量化时才开启
    # --quant_desc_path ${quant_desc_path}

# 确定性运算
export LCCL_DETERMINISTIC=true  
export HCCL_DETERMINISTIC=true  
export ATB_MATMUL_SHUFFLE_K_ENABLE=0  
export ATB_LLM_LCOC_ENABLE=true       
export CLOSE_MATMUL_K_SHIFT=true    

model_path="/home/weight/Qwen-Image-Edit-2509/"
quant_desc_path="/home/Qwen-Image-Edit-2509/quant_w8a8_withoutData_use_disable_quant_layers/quant_model_description_w8a8_dynamic.json"

# 算子优化  需要时开启
# export ROPE_FUSE=1
# export ADALN_FUSE=1

# cache 算法优化  需要时开启
# export COND_CACHE=1
# export UNCOND_CACHE=1

# GEdit-Bench数据集的使用示例, 8卡并行 cfg_size=2 ulysses_size=4
dataset_path="/mnt/datasets/GEdit-Bench"
export ASCEND_RT_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
torchrun --nproc_per_node=8 --master-port 29503 run_edit_2509_cfg_usp.py \
    --model_path ${model_path} \
    --dataset_name "GEdit-Bench" \
    --dataset_path ${dataset_path} \
    --num_inference_steps 50 \
    --seed 42 \
    --output_dir "GEdit-Bench_cfg2_ulysses4_optimize_1+2_50steps" \
    --ulysses_size 4 \
    --cfg_size 2 \
    --vae_tiling \
    --vae_slicing \
    # 有量化权重时候，使用量化时才开启
    # --quant_desc_path ${quant_desc_path}

# 4张卡
export ASCEND_RT_VISIBLE_DEVICES=0,1,2,3
export QWEN_MODEL_PATH="/mnt/share/weight/Qwen2.5-VL-72B-Instruct"
python run_gedit_score.py \
    --model_name Qwen-Image-Edit-2509 \
    --dataset_dir /mnt/datasets/GEdit-Bench \
    --edited_images_dir /mnt/Qwen-Image-Edit-2509/GEdit-Bench_cfg2_ulysses4_optimize_1+2_40steps \
    --backbone qwen25vl \
    --save_dir ./score_dir \

python calculate_statistics.py \
    --model_name Qwen-Image-Edit-2509 \
    --save_path ./score_dir \
    --backbone qwen25vl \
    --language all