mirror of
https://github.com/ikawrakow/ik_llama.cpp.git
synced 2026-01-31 03:29:52 +00:00
29 lines
1.1 KiB
Markdown
29 lines
1.1 KiB
Markdown
### 🐛 [#604](https://github.com/ikawrakow/ik_llama.cpp/pull/604) - Fix attn_v conditionality when quantizing.
|
|
|
|
| **Author** | `Nexesenex` |
|
|
| :--- | :--- |
|
|
| **State** | ❌ **Closed** |
|
|
| **Created** | 2025-07-12 |
|
|
| **Updated** | 2025-07-13 |
|
|
|
|
---
|
|
|
|
#### Description
|
|
|
|
To retain compatibility with : https://github.com/ikawrakow/ik_llama.cpp/pull/91 We need "else if" and not "if", otherwise the MOE and 70b condition takes precedence over the specified quant in the CLI.
|
|
|
|
I can also expand this legacy custom quant to the IQ1 and IQ2 types quant strategies tree, and add the shexp tensor to it, if that's all right.
|
|
|
|
- [x] I have read the [contributing guidelines](https://github.com/ggerganov/llama.cpp/blob/master/CONTRIBUTING.md)
|
|
- Self-reported review complexity:
|
|
- [x] Low
|
|
- [ ] Medium
|
|
- [ ] High
|
|
|
|
---
|
|
|
|
#### 💬 Conversation
|
|
|
|
👤 **ikawrakow** submitted a review the **2025-07-13** at **09:24:27**: ✅ `APPROVED`<br>
|
|
|
|
This is OK, but I think you should really start using `--custom-q`. That way you can make the mixes any way you like without relying on the logic in this function. |