diffusers-ascend-weight-prep

Diffusers Model Weight Preparation

This Skill provides two core capabilities:

Download model weights from HuggingFace or ModelScope
Generate fake weights (random initialization) based on the Diffusers Pipeline config files, for business validation

Quick Start

1. Download Model Weights

Download from HuggingFace:

pip install -U huggingface_hub

# 基础用法
hf download Wan-AI/Wan2.2-I2V-A14B-Diffusers

Using a proxy (for access within China)

export HF_ENDPOINT=https://hf-mirror.com hf download Wan-AI/Wan2.2-I2V-A14B-Diffusers

Download via script (supports dry-run testing)

python scripts/download_weights.py hf Wan-AI/Wan2.2-I2V-A14B-Diffusers --proxy https://hf-mirror.com python scripts/download_weights.py hf Wan-AI/Wan2.2-I2V-A14B-Diffusers --dry-run


**从 ModelScope 下载：**

```bash
pip install modelscope

# 基础用法
modelscope download --model Wan-AI/Wan2.2-T2V-A14B

# 通过脚本下载
python scripts/download_weights.py modelscope Wan-AI/Wan2.2-T2V-A14B -o ./models
python scripts/download_weights.py modelscope Wan-AI/Wan2.2-T2V-A14B --dry-run

2. Generate Fake Weights

Automatically download metadata from HuggingFace (config.json, model_index.json, tokenizer, etc.,without downloading weight files), then randomly initialize each sub-model based on the config and save it, generating fake weightswith a file structure completely identicalto the original model.

Generate directly from HuggingFace:

# 生成 Z-Image-Turbo 假权重
python scripts/generate_fake_weights.py from-hub Tongyi-MAI/Z-Image-Turbo -o ./fake_z_image --dtype bfloat16

# 生成 Qwen-Image-2512 假权重（国内使用镜像）
python scripts/generate_fake_weights.py from-hub Qwen/Qwen-Image-2512 -o ./fake_qwen --proxy https://hf-mirror.com

Generate from an existing local metadata directory:

python scripts/generate_fake_weights.py from-local ./model_metadata_dir -o ./fake_weights --dtype bfloat16

Verify that the fake weights can be loaded normally:

import torch
import diffusers

pipe = diffusers.DiffusionPipeline.from_pretrained(
    "./fake_z_image",
    torch_dtype=torch.bfloat16,
)
print(f"Pipeline: {type(pipe).__name__}")
for name, comp in pipe.components.items():
    if hasattr(comp, "parameters"):
        pc = sum(p.numel() for p in comp.parameters())
        print(f"  {name}: {type(comp).__name__} ({pc:,} params)")

Environment Requirements

Component	Installation Command	Description
huggingface_hub	`pip install -U huggingface_hub`	HuggingFace download & fake weight metadata download
modelscope	`pip install modelscope`	ModelScope download
torch	`pip install torch`	Fake weight generation
diffusers	`pip install diffusers`	Pipeline model instantiation
transformers	`pip install transformers`	text_encoder instantiation

Fake Weight Generation Principle

Generation process:

Download metadata files from HuggingFace via snapshot_download (excluding *.safetensors、*.bin）
Read model_index.json, parse each component of the Pipeline (scheduler, text_encoder, tokenizer, transformer, vae)
For each model component, use from_config() to instantiate the model from config.json (random weights)
Save via save_pretrained() , automatically generating correctly formatted safetensors files and index.json
tokenizer and scheduler directly copy the config files

The generated directory structure is completely identical to the original model, for example Z-Image-Turbo:

fake_z_image_turbo/
├── model_index.json
├── scheduler/
│   └── scheduler_config.json
├── text_encoder/
│   ├── config.json
│   ├── generation_config.json
│   ├── model.safetensors
│   └── model.safetensors.index.json
├── tokenizer/
│   ├── merges.txt
│   ├── tokenizer_config.json
│   ├── tokenizer.json
│   └── vocab.json
├── transformer/
│   ├── config.json
│   ├── diffusion_pytorch_model-00001-of-00002.safetensors
│   ├── diffusion_pytorch_model-00002-of-00002.safetensors
│   └── diffusion_pytorch_model.safetensors.index.json
└── vae/
    ├── config.json
    └── diffusion_pytorch_model.safetensors

Script Reference

download_weights.py

Parameter	Required	Description
`source`	Yes	Download source:`hf` or `modelscope`
`model_id`	Yes	Model ID (e.g., `Wan-AI/Wan2.2-I2V-A14B-Diffusers`）
`-o, --output-dir`	No	Output directory
`--proxy`	No	HuggingFace proxy URL (only for `hf` source)
`--dry-run`	No	Test mode, does not actually download

generate_fake_weights.py

Subcommand from-hub: Download metadata from HuggingFace + generate fake weights

Parameter	Required	Description
`model_id`	Yes	HuggingFace Model ID
`-o, --output-dir`	Yes	Output directory
`--proxy`	No	HuggingFace proxy URL
`--dtype`	No	Data type:`float32`、`float16`、`bfloat16`(default:`bfloat16`）

Subcommand from-local: Generate fake weights from a local metadata directory

Parameter	Required	Description
`metadata_dir`	Yes	Local metadata directory (must contain model_index.json)
`-o, --output-dir`	Yes	Output directory
`--dtype`	No	Data type (default:`bfloat16`）

Verified models

Model	Pipeline type	Component	Verification status
Qwen/Qwen-Image-2512	QwenImagePipeline	vae(127M) + text_encoder(8.3B) + transformer(20.4B)	✅
Tongyi-MAI/Z-Image-Turbo	ZImagePipeline	vae(84M) + text_encoder(4.0B) + transformer(6.2B)	✅

Notes

Fake weights are randomly initialized; the generated images are meaningless and are only used to verify the code workflow
Generating fake weights for large models (>10B parameters) may require a lot of memory and time
In mainland China, it is recommended to use --proxy https://hf-mirror.com or set the HF_ENDPOINT environment variable
If you encounter a socks proxy conflict, first unset all_proxy ALL_PROXY and then run

About this skill