I've been looking at the solver setup and hybrid quantization stuff. Right now capacity seems fixed.
What do you think about making it adaptive? For example, lowering load when the device gets hot, low on memory, or when there's high churn based on simple telemetry.
Could help low-end devices participate better without breaking things. Just an observation from running it locally.
Happy to test if useful.
I've been looking at the solver setup and hybrid quantization stuff. Right now capacity seems fixed.
What do you think about making it adaptive? For example, lowering load when the device gets hot, low on memory, or when there's high churn based on simple telemetry.
Could help low-end devices participate better without breaking things. Just an observation from running it locally.
Happy to test if useful.