Static Graph Features

Sunset Features

The following capabilities are currently implemented only in static graphs (GRAPH_MODE) and are not supported in dynamic graphs (PyNative). They are marked as sunset features. New features are preferentially evolved in dynamic graphs. For details about the following features, see section "Static Graph Implementation."

Dynamic graphs (r2.0.0) currently focus on the entire training process, and the model structure is explicitly configured through the YAML model section. The following capabilities are not yet implemented in dynamic graphs and require static graphs.

Feature	Description	Link
Inference	Inference process of a trained model.	Static Graph > Inference
Service deployment	Service deployment based on vLLM and others.	Static Graph > Service Deployment
Quantification	Quantization inference integrated with MindSpore Golden Stick.	Static Graph > Quantization
Checkpoint weight	Converts or shards weights in CKPT format, for example, Checkpoint 1.0.	Static Graph > Ckpt Weights
Loading the HF model configuration	Automatically merges HF config.json through pretrained_model_dir (the dynamic graph structure must be explicitly written in the model section).	Static Graph > Loading Hugging Face Model Configuration

After the preceding capabilities are implemented in dynamic graphs, the corresponding items will be moved from this page and supplemented in the dynamic graph documentation.