ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-05-11 00:20:19 +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

13
github-data/pull_requests/102 - Add support for Granite and GraniteMoE models.md Normal file
View File

@@ -0,0 +1,13 @@
### 🔀 [#102](https://github.com/ikawrakow/ik_llama.cpp/pull/102) - Add support for Granite and GraniteMoE models
| **Author** | `ikawrakow` |
| :--- | :--- |
| **State** | ❌ **Closed** |
| **Created** | 2024-10-22 |
| **Updated** | 2024-10-22 |
---
#### Description
On CUDA GraniteMoE-1b suffers from precision issues in the attention portion, so I became curious to see why. One way to avoid the NaNs is to set the precision of the `K*Q` matrix multiplication to `F32`. What also fixes it is to apply the attention scale on `Q` before the `K*Q` multiplication (the solution I went with in this PR). One can apply the scale before or after RoPE. It works in both cases, so this really narrows it down to the `K*Q` multiplication suffering from precision issues when done in `f16`. Strange how these models were trained in the first place.