Skip to content

[Real-time Inference] Real-time inference endpoints do NOT support trained LM Models #1338

@zhjwy9343

Description

@zhjwy9343

In GraphStorm, there are two main methods to use language models (LMs) during graph construction: 1/ tokenizing text attributes into integers of tokens by using with HuggingFace tokenizer;2/ embedding text attributes into numerical features by using HuggingFace Bert models.

After graph construction, GraphStorm can directly use embeddings as one of node features without retraining the Bert model used in construction. And this method supports real-time inference.

Converting tokens into embeddings for the downstream training w/wt LM retraining, however, do not support real-time inference. In the transform_fn GraphStorm will raise KeyError: the key 'token_ids' is not found in node feature.

Metadata

Metadata

Assignees

Labels

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions