Skip to content

Dig into config issues when setting vllm options like --max-num-seqs #94

@joerunde

Description

@joerunde

With the continuous batching support, we noticed that the VllmConfig objects that are created are different in the main process and the worker processes: #66 (comment)

This caused us to introduce new environment variables instead of using the existing vllm args, because we needed to ensure that the batch size and context length were accurate everywhere. We should look into this and understand why there are differences in the config objects, and fix any bugs upstream so that we can accept the usual vllm args.

Metadata

Metadata

Labels

No labels
No labels

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions