### 🔀 [#251](https://github.com/ikawrakow/ik_llama.cpp/pull/251) - Try using fp32 for FlashMLA | **Author** | `ikawrakow` | | :--- | :--- | | **State** | ❌ **Closed** | | **Created** | 2025-03-10 | | **Updated** | 2025-03-12 | --- #### 💬 Conversation 👤 **ikawrakow** commented the **2025-03-12** at **07:51:20**:
Closing this as the numerical issues were caused by `fp16` experts matrix multiplications.