mirror of
https://github.com/ikawrakow/ik_llama.cpp.git
synced 2026-03-02 10:00:07 +00:00
15 lines
509 B
Markdown
15 lines
509 B
Markdown
### 🔀 [#48](https://github.com/ikawrakow/ik_llama.cpp/pull/48) - AVX2 Flash Attention
|
|
|
|
| **Author** | `ikawrakow` |
|
|
| :--- | :--- |
|
|
| **State** | ❌ **Closed** |
|
|
| **Created** | 2024-09-10 |
|
|
| **Updated** | 2024-09-10 |
|
|
|
|
---
|
|
|
|
#### Description
|
|
|
|
We don't gain as much as on a Zen4 system as there aren't as many vector registers, so we need to load/store data much more often. Still, we do get a small gain in performance.
|
|
|
|
For now it supports only `fp16` kv-cache. Quantized kv-cache will be added later. |