mirror of
https://github.com/ikawrakow/ik_llama.cpp.git
synced 2026-05-01 03:41:53 +00:00
15 lines
573 B
Markdown
15 lines
573 B
Markdown
### 🔀 [#488](https://github.com/ikawrakow/ik_llama.cpp/pull/488) - Faster CPU prompt processing for Trellis quants and MoE models
|
|
|
|
| **Author** | `ikawrakow` |
|
|
| :--- | :--- |
|
|
| **State** | ❌ **Closed** |
|
|
| **Created** | 2025-06-03 |
|
|
| **Updated** | 2025-06-05 |
|
|
|
|
---
|
|
|
|
#### Description
|
|
|
|
This PR is a follow up to #482, and applies the same dequantizing GEMM for MoE matrix multiplications.
|
|
|
|
For a DeepSeek-Lite model where only the `ffn_up` and `ffn_gate` tensors are quantized with `IQ2_KT` I observe a ~35% improvement in PP performance compared to te main branch. |