Implementation:Zai org CogVideo Args Parse Args

Implementation Metadata
Name	Args_Parse_Args
Type	API Doc
Category	Configuration_Management
Domains	Fine_Tuning, Diffusion_Models
Knowledge Sources	CogVideo Repository
Last Updated	2026-02-10 00:00 GMT

Overview

Args_Parse_Args is a concrete tool for parsing and validating CogVideoX fine-tuning configuration, provided by the CogVideo finetune package.

Description

This implementation provides the Args Pydantic model class and its parse_args classmethod for constructing a fully validated training configuration from command-line arguments. The Args class defines all fields needed for CogVideoX fine-tuning, including model paths, LoRA settings, training hyperparameters, and validation options. Pydantic validators enforce constraints such as frame count divisibility, resolution compatibility, and precision requirements.

Usage

Use at the entry point of any CogVideoX fine-tuning script. Call Args.parse_args() to parse command-line arguments and receive a validated configuration object. This is typically the first operation in the training pipeline.

Code Reference

Source Location

• finetune/schemas/args.py:L10-177 -- Args Pydantic model definition
• finetune/schemas/args.py:L178-254 -- parse_args classmethod

Signature

class Args(BaseModel):
    model_path: Path
    model_name: str
    model_type: Literal["i2v", "t2v"]
    training_type: Literal["lora", "sft"] = "lora"
    output_dir: Path
    data_root: Path
    caption_column: Path
    video_column: Path
    train_epochs: int
    batch_size: int
    train_resolution: Tuple[int, int, int]  # (frames, height, width)
    mixed_precision: Literal["no", "fp16", "bf16"]
    rank: int = 128
    lora_alpha: int = 64
    target_modules: List[str] = ["to_q", "to_k", "to_v", "to_out.0"]
    learning_rate: float = 1e-4
    gradient_accumulation_steps: int = 1
    max_grad_norm: float = 1.0
    checkpointing_steps: int = 200
    checkpointing_limit: int = 10
    do_validation: bool = True
    validation_steps: int = 200
    seed: int = 42
    nccl_timeout: int = 1800

    @classmethod
    def parse_args(cls) -> "Args":
        """Parse command-line arguments and return a validated Args instance."""
        ...

Import

from finetune.schemas import Args

Key Parameters

Parameter	Type	Default	Description
model_path	Path	required	Path to pretrained CogVideoX model checkpoint.
model_name	str	required	Model variant name (e.g., "cogvideox-5b").
model_type	Literal["i2v", "t2v"]	required	Whether the model is text-to-video or image-to-video.
training_type	Literal["lora", "sft"]	"lora"	Training mode: LoRA adapter or full supervised fine-tuning.
train_resolution	Tuple[int, int, int]	required	Target resolution as (frames, height, width).
rank	int	128	LoRA rank (dimension of low-rank matrices).
lora_alpha	int	64	LoRA scaling factor.
target_modules	List[str]	["to_q", "to_k", "to_v", "to_out.0"]	Transformer modules to apply LoRA adapters to.
learning_rate	float	1e-4	Optimizer learning rate.
mixed_precision	Literal["no", "fp16", "bf16"]	required	Mixed precision training mode.

External Dependencies

• pydantic -- Schema validation and type enforcement
• argparse -- Command-line argument parsing

I/O Contract

Inputs

Input	Format	Description
Command-line arguments	Shell arguments or script variables	Training parameters passed via CLI (e.g., --model_path, --train_resolution 49 480 720).

Outputs

Output	Format	Description
Configuration object	Args (Pydantic model instance)	Fully validated training configuration with all fields populated and constraints checked.

Usage Examples

Basic Argument Parsing

from finetune.schemas import Args

# Parse and validate from command line
args = Args.parse_args()

# Access validated fields
print(f"Model: {args.model_name}")
print(f"Resolution: {args.train_resolution}")
print(f"LoRA rank: {args.rank}")
print(f"Training type: {args.training_type}")

Shell Script Invocation

python train.py \
    --model_path /models/cogvideox-5b \
    --model_name cogvideox-5b \
    --model_type t2v \
    --training_type lora \
    --output_dir /output/lora_run \
    --data_root /data/my_videos \
    --caption_column prompts.txt \
    --video_column videos.txt \
    --train_epochs 100 \
    --batch_size 1 \
    --train_resolution 49 480 720 \
    --mixed_precision bf16 \
    --rank 128 \
    --lora_alpha 64

Related Pages

• Principle:Zai_org_CogVideo_Training_Configuration
• Environment:Zai_org_CogVideo_Diffusers_Finetuning_Environment
• Heuristic:Zai_org_CogVideo_BF16_FP16_Precision_Selection
• Heuristic:Zai_org_CogVideo_Frame_Count_and_Resolution_Constraints
• Heuristic:Zai_org_CogVideo_LoRA_Configuration_Tips
• Heuristic:Zai_org_CogVideo_Training_Hyperparameter_Defaults
• Implementation:Zai_org_CogVideo_T2V_I2V_Dataset_Loader
• Implementation:Zai_org_CogVideo_CogVideoX_LoRA_Trainer_Load_Components