No way to predict VRAM/GPU contention when running more than one local model — calls silently time out and fall back

Summary

An M1 developer finds that one small local model runs fine, but with two or three concurrent model calls every request times out and silently falls back to non-LLM behaviour. This is the concurrency blind spot of local inference: nothing tells a user how much VRAM/unified memory a model plus its KV cache will consume at a given context, how much headroom remains, or how many concurrent instances are actually safe. Tools like Ollama give no pre-flight memory estimate or admission control, so failures surface as unexplained timeouts and degraded quality rather than a clear 'you can't fit a second instance at this context length.' For indie developers and small teams shipping local-AI features, this is a debugging nightmare — the symptom looks like a model bug, not a capacity bug — and it directly blocks multi-agent or multi-session local apps. No mainstream product maps workload (models × quantization × context × concurrency) to a memory budget, so the problem remains unsolved and would be worth paying for before shipping.