Metal: ~6% token-generation regression on gemma-4-26B-A4B since #28164 (fusion packing chains patterns, reducing concurrency)

在 Apple Silicon(Metal 后端)上用 llama.cpp 跑 Gemma 26B 系列模型做 token 生成时,若发现相对 b10908 及之前版本速度下降约 6–9%,通常是自 #28164(b10909 起)引入的 fusion packing 变更把多个 pattern 串
![[Bug]: Knowledge base creators cannot access datasets created via upload or RAG Pipeline](https://www.chat-gpts.plus/wp-content/uploads/2026/09/42836-411fb8f3-768x403.jpg)
![[Bug]: Multi card issue and multi token prediction (mtp) issue with Intel/Qwen3.6-35B-A3B-int4-mixed-AutoRound](https://www.chat-gpts.plus/wp-content/uploads/2026/09/53119-9a400882-768x403.jpg)




![[Bug] Ollama hardcodes Gemma 4 image token budget (max_soft_tokens=280), breaking OCR on high-resolution images](https://www.chat-gpts.plus/wp-content/uploads/2026/09/17152-d9c601c0-768x403.jpg)

