Best workstation layout for local models: anyone testing this?

[tensorworks]

That matches what I was seeing. The pattern around continuous batching has been pretty similar for me. The biggest change for me was simplifying the stack instead of adding one more optimization layer.