mirror of
https://github.com/ikawrakow/ik_llama.cpp.git
synced 2026-05-01 11:51:53 +00:00
15 lines
683 B
Markdown
15 lines
683 B
Markdown
### 🔀 [#603](https://github.com/ikawrakow/ik_llama.cpp/pull/603) - Check if MMQ should be used before using it
|
|
|
|
| **Author** | `ikawrakow` |
|
|
| :--- | :--- |
|
|
| **State** | ❌ **Closed** |
|
|
| **Created** | 2025-07-12 |
|
|
| **Updated** | 2025-07-13 |
|
|
|
|
---
|
|
|
|
#### Description
|
|
|
|
In #589 I added an optimization of the fused ffn_up/gate op to not repeat the quantization of the activations when `ffn_up` and `ffn_gate` are quantized with the same type. But the check to use the direct route did not consider the possibility that some quantization types do not have MMQ implementation (e.g., `IQ1_M`), which then results in an assert.
|
|
|
|
This PR adds the missing check, which should fix #596 |