ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-23 14:44:09 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add GitHub data (#637)

Browse Source
This commit is contained in:
Thomas
2025-07-22 18:18:40 +02:00
committed by GitHub
parent 9513222ba5
commit 94aa54df76
626 changed files with 175142 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
github-data/pull_requests/582-Vulkan_ adding GGML_OP_MULTI_ADD implementation.md Normal file
View File

@@ -0,0 +1,15 @@
### 🔀 [#582](https://github.com/ikawrakow/ik_llama.cpp/pull/582) - Vulkan: adding GGML_OP_MULTI_ADD implementation
| **Author** | `ikawrakow` |
| :--- | :--- |
| **State** | ❌ **Closed** |
| **Created** | 2025-07-04 |
| **Updated** | 2025-07-04 |
---
#### Description
This is relevant for MoE models. The performance improvement is surprisingly small. Somewhere it was mentioned that Vulkan kernel launch overhead is significantly larger than CUDA, so I would have expected a more significant performance benefit. For DeepSeek-Lite, the number of graph nodes in `ik_llama.cpp` with this PR is 1420 vs 1871 in mainline `llama.cpp`.
But, if nothing else, this removes the last Vulkan special-casing when building the compute graph.