mirror of
https://github.com/ikawrakow/ik_llama.cpp.git
synced 2026-05-01 11:51:53 +00:00
17 lines
561 B
Markdown
17 lines
561 B
Markdown
### 🔀 [#43](https://github.com/ikawrakow/ik_llama.cpp/pull/43) - iq2_tn: slightly faster PP on Zen4
|
|
|
|
| **Author** | `ikawrakow` |
|
|
| :--- | :--- |
|
|
| **State** | ❌ **Closed** |
|
|
| **Created** | 2024-09-08 |
|
|
| **Updated** | 2024-09-08 |
|
|
|
|
---
|
|
|
|
#### Description
|
|
|
|
With this change we get `PP512 = 494 t/s` (using flash attention), up from `468 t/s` (~5% improvement) running on a Ryzen-7950X CPU.
|
|
|
|
Compared to the initial `IQ2_TN` PR #13 the cumulative improvement is 15%.
|
|
|
|
Compared to `TQ2_0` in `llama.cpp`, which has now been merged, we are now 80% faster. |