Handle models with smaller context length

[deboer-tim]

Some models (e.g. granite3.1-dense:8b) have a short context length that runs out after a few prompts. In Ollama I don't notice, I assume it is either managing the length or continually clearing it. In Ramalama I see the following message and the process exits:

context size exceeded
llama_decode: failed to decode, ret = 1
failed to decode
failed to generate response

[ericcurtin]

Maybe we are supposed to read the context length from the model itself and use that:

https://ollama.com/library/granite3.1-dense/blobs/0a922eb99317

[deboer-tim]

Yes, I think that's the expectation.