Implementation:Zai org CogVideo SAT Get Model Load Checkpoint

Attribute	Value
Implementation Name	SAT Get Model Load Checkpoint
Workflow	SAT Video Generation
Step	2 of 5
Type	Wrapper Doc
Source File	sat/sample_video.py:L137-144
Repository	zai-org/CogVideo
External Dependencies	sat
Last Updated	2026-02-10 00:00 GMT

Overview

Implementation of model loading for SAT-based CogVideoX inference. This wrapper combines SAT framework's get_model and load_checkpoint functions to instantiate the SATVideoDiffusionEngine with pretrained weights and prepare it for inference.

Description

The model loading implementation consists of three sequential calls:

1. get_model(args, model_cls=SATVideoDiffusionEngine) -- Constructs the model architecture from configuration
2. load_checkpoint(model, args) -- Loads pretrained weights from the checkpoint path
3. model.eval() -- Sets the model to evaluation mode

The get_model function reads model architecture parameters from args.model_config (populated from the YAML config file) and instantiates the specified model class. The load_checkpoint function reads weights from args.load and applies them to the model.

Usage

from sat.model.base_model import get_model
from sat.training.model_io import load_checkpoint
from diffusion_video import SATVideoDiffusionEngine

# args is obtained from get_args()
model = get_model(args, model_cls=SATVideoDiffusionEngine)
load_checkpoint(model, args)
model.eval()

# Model is now ready for inference
with torch.no_grad():
    samples = model.sample(cond, uc, batch_size=1, shape=shape)

Code Reference

Source Location

File	Lines	Description
sat/sample_video.py	L137-144	Model loading wrapper

Signature

from sat.model.base_model import get_model
from sat.training.model_io import load_checkpoint

model = get_model(args, model_cls=SATVideoDiffusionEngine)
load_checkpoint(model, args)
model.eval()

Import

from sat.model.base_model import get_model
from sat.training.model_io import load_checkpoint

I/O Contract

Inputs

Parameter	Type	Default	Description
args	argparse.Namespace	Required	Parsed arguments containing model config and checkpoint path
args.model_config	dict	From YAML	Model architecture configuration loaded from YAML
args.load	str	Required	Path to pretrained checkpoint directory
model_cls	type	SATVideoDiffusionEngine	Model class to instantiate

Outputs

Output	Type	Description
model	SATVideoDiffusionEngine	Loaded model in eval mode, ready for inference

Usage Examples

Example 1: Standard model loading

from arguments import get_args
from sat.model.base_model import get_model
from sat.training.model_io import load_checkpoint
from diffusion_video import SATVideoDiffusionEngine

args = get_args()
model = get_model(args, model_cls=SATVideoDiffusionEngine)
load_checkpoint(model, args)
model.eval()

Example 2: Model loading with distributed setup

import torch
from arguments import get_args
from sat.model.base_model import get_model
from sat.training.model_io import load_checkpoint
from diffusion_video import SATVideoDiffusionEngine

args = get_args()
model = get_model(args, model_cls=SATVideoDiffusionEngine)
load_checkpoint(model, args)
model.eval()

# Distributed inference uses SAT's built-in model parallelism
# configured via args (--model-parallel-size, etc.)

Related Pages

• Principle:Zai_org_CogVideo_SAT_Model_Loading_for_Inference -- Principle governing model loading for inference
• Environment:Zai_org_CogVideo_SAT_Framework_Environment
• Zai_org_CogVideo_SAT_Inference_Get_Args -- Previous step: argument parsing that provides model config
• Zai_org_CogVideo_SAT_Read_From_CLI_File -- Next step: reading prompts for generation
• Zai_org_CogVideo_SAT_Diffusion_Sample -- Sampling step using the loaded model