mirror of
https://github.com/ikawrakow/ik_llama.cpp.git
synced 2026-05-01 11:51:53 +00:00
17 lines
499 B
Markdown
17 lines
499 B
Markdown
### 🔀 [#248](https://github.com/ikawrakow/ik_llama.cpp/pull/248) - Faster MoE token generation on CUDA
|
|
|
|
| **Author** | `ikawrakow` |
|
|
| :--- | :--- |
|
|
| **State** | ❌ **Closed** |
|
|
| **Created** | 2025-03-09 |
|
|
| **Updated** | 2025-03-10 |
|
|
|
|
---
|
|
|
|
#### Description
|
|
|
|
This PR adds special purpose matrix-vector multiplications for MoE models.
|
|
|
|
For DeepSeek-Lite this results in a ~25% speedup for token generation.
|
|
|
|
For now only implemented ~with the `-fmoe` option and only~ for quantized experts. |