ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-05-03 04:41:47 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add GitHub data: filename sanitization (#640)

Browse Source
This commit is contained in:
Thomas
2025-07-23 13:31:53 +02:00
committed by GitHub
parent 3600d82e98
commit eaa2510a28
626 changed files with 0 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
github-data/pull_requests/24 - softcap_ minor improvement.md Normal file
View File

@@ -0,0 +1,15 @@
### 🔀 [#24](https://github.com/ikawrakow/ik_llama.cpp/pull/24) - softcap: minor improvement
| **Author** | `ikawrakow` |
| :--- | :--- |
| **State** | ❌ **Closed** |
| **Created** | 2024-08-21 |
| **Updated** | 2024-08-21 |
---
#### Description
With this change we get 104 t/s for Gemma-2-9b with a context of 8192 tokens on a Ryzen-7950X.
For this model and context size, about 10% of the time is spent in `softcap` (5.8%) and `soft_max` (4.2%) when running on the Ryzen-7950X CPU. I wonder if it wouldn't be better to merge `softcap` and `soft_max` into a single op (for Gemma-2, `softcap` in the attention layer is immediately followed by `soft_max`)