Skip to content

[CB] requesting only last token per default #254

@yannicks1

Description

@yannicks1

[CB] requesting only last token per default

To minimize data transfer, we only want to request the last tokens logits instead of the the entire prompts for CB.
The flag only_last_token in the fms forward pass should enable this, however, for right padding we get different results. This branch is to debug this issue in fms and is eventually to be merged when support in fms is established.

-> needs changes in fms and Spyre compiler, vllm-spyre code is ready.

related PR: #253

Metadata

Metadata

Assignees

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