COG-based ones, have a specific interface -- i.e each model, even each revision, can have it's own interface. Also the input/output json follows a particular schema (so there is a model specific bit, but also common bits outside). Encode this in CogBaseModel, and the common bits in CogRequest, CogResponse, so people can even query their own custom COGs by specifying the model-specific bits.
LLM custom models have the same interface as other text-generation models.
Also implement SDXL as a CogBaseModel. I think the cog models should live in their own directory (not in text-to-image as it's now), but that is a minor correction.
I'm a bit new to typescript, so I might have messed something up.
I'll fix the TextToImage type in another patch, the current one assumes all SDXL params are available to all TextToImage models, but that is not the case.
DeepInfra supports 2 types of custom models:
COG-based ones, have a specific interface -- i.e each model, even each revision, can have it's own interface. Also the input/output json follows a particular schema (so there is a model specific bit, but also common bits outside). Encode this in
CogBaseModel
, and the common bits inCogRequest
,CogResponse
, so people can even query their own custom COGs by specifying the model-specific bits.LLM custom models have the same interface as other text-generation models.
Also implement
SDXL
as aCogBaseModel
. I think the cog models should live in their own directory (not intext-to-image
as it's now), but that is a minor correction.I'm a bit new to typescript, so I might have messed something up.
I'll fix the TextToImage type in another patch, the current one assumes all SDXL params are available to all TextToImage models, but that is not the case.