Gemma 4 iOS hits CPU fallback, buffer limit

This looks less like a Gemma problem than a runtime packing problem: the model can run, but the current iOS delegate path appears to hit a Metal backend constraint before GPU acceleration ever starts.

• –The reported failure mode is specific and reproducible: Metal rejects `buffer(31)`, so the app drops to CPU fallback and tanks latency
• –Google AI Edge Gallery reportedly runs fast on the same hardware, which suggests the newer LiteRT-LM stack may already avoid this limitation
• –The post highlights the current fragmentation in iOS local-model tooling: MediaPipe, LiteRT-LM, MLX-Swift, and custom bridges all trade off maturity versus performance
• –For developers, the practical takeaway is that Gemma 4 on iOS may be gated more by backend/runtime choice than by raw device capability
• –This is a strong signal that on-device LLM work on iPhone is shifting toward lower-level, Apple-native inference paths