Follow up to autocomplete pr #317

[oandreeva-nv]

This is a follow up PR to : #317 . Unfortunately, I missed essential part.

After we retrieved updated model configuration from core during the call:

python_backend/src/python_be.cc

Line 1864 in 0f12211

RETURN_IF_ERROR((*state)->SetModelConfig());

decoupled_ is not updated on the python backend site, since it is a part of ModelState.

I've added PropagateAutoCompletedConfig() that does exactly that, i.e. looks for model_transaction_policy in the model config (already filled by core), and if decoupled is found in model_transaction_policy , we update decoupled_ property.

Actual functionality is tested with vllm backend on this PR: triton-inference-server/vllm_backend#20

[the-david-oy]

Great work!