jinaai/xlm-roberta-flash-implementation

Jina AI org Jul 23

•

In this PR, we try to infer the task type from the input directly. If the task type is specified then use that else try to infer from the input. If the prompt can be inferred from the input then the respective lora adapter is set else, no lora adapter is used. Embedding returned in that case will be from the backend model.
Changes:

parameter name changed to task_type based on the usage on the API side
tokenizer moved from encode() to init in XLMRobertaModel class
Remove __no_lora_update__ and set default value as None
Use only one input to infer the prompt and set the adapter.
Adapt both encode and forward function of XLMRobertaLoRA class

feat: set adapter based on prompt71b163e3

jupyterjazz

Jina AI org Jul 24

LGTM

fix: read prompts from config51411ff2

makram93 changed pull request status to open Jul 30

fix: check for exact task namesb3c540ce

jupyterjazz

Jina AI org Jul 30

Might be a good idea to set prompts to the default values in the configuration_xlm_roberta.py if no prompts are provided, otherwise we'll have to copy the dict everytime

makram93

Jina AI org Jul 31

By not setting it to default we enforce that the LoRA config should always have prompts param, same as the lora_adaptations param. I feel this will be needed when we move out the instructions adding part from the training script directly here

jupyterjazz changed pull request status to merged Aug 2

jinaai
/

xlm-roberta-flash-implementation

feat-routing