Is your feature request related to a problem? Please describe.
I want to deploy ASR related models on triton. The problem is the current implementation only features the v1 engine and thus text models.
This excludes a whole family of models from usage on triton inference server as you can see
here. There is also a lot of other use cases like translation, multi modal usage, etc.
Describe the solution you'd like
To adapt to the current engine the model.py backend has to be rewritten as well as the request.py for the input format.
Describe alternatives you've considered
There is no alternative in terms of the vllm backend, since there are no tensorrt llm implementations for all these models.
Is your feature request related to a problem? Please describe.
I want to deploy ASR related models on triton. The problem is the current implementation only features the v1 engine and thus text models.
This excludes a whole family of models from usage on triton inference server as you can see
here. There is also a lot of other use cases like translation, multi modal usage, etc.
Describe the solution you'd like
To adapt to the current engine the model.py backend has to be rewritten as well as the request.py for the input format.
Describe alternatives you've considered
There is no alternative in terms of the vllm backend, since there are no tensorrt llm implementations for all these models.