Enhancement: Add a preflight fit check for model weights and context memory requirements

Hi everyone,

would it make sense to add a small preflight fit check to FastFlowLM?

The goal would be to check before loading whether a model fits on the target system for a given context length. The check should estimate required resources for model weights, KV/context cache, runtime overhead, and relevant NPU/platform limits.

This would help Lemonade avoid late load failures and OOM situations when memory is tight, and provide clearer feedback to users (e.g. model does not fit, reduce context length, or choose another backend).

Benefits:

• fail early before starting/loading the model
• clearer error messages
• better Lemonade integration
• possible max-context suggestion
• cleaner distinction between “FLM setup is valid” and “this specific model config fits”

I’d be happy to work on this from the Lemonade side and contribute the backend feature (useful for any frontend or caller) back to FastFlowLM.

No files, tests, or entry points are named. First clarify the FastFlowLM API and Lemonade integration boundary, then define how to estimate model weights, KV/context cache, runtime overhead, and NPU/platform limits; done means early fit failures provide actionable feedback before loading.