mirror of
https://github.com/ikawrakow/ik_llama.cpp.git
synced 2026-05-01 11:51:53 +00:00
15 lines
589 B
Markdown
15 lines
589 B
Markdown
### 🔀 [#64](https://github.com/ikawrakow/ik_llama.cpp/pull/64) - Better sub-3-bit quantization mixes with a qkv tensor
|
|
|
|
| **Author** | `ikawrakow` |
|
|
| :--- | :--- |
|
|
| **State** | ❌ **Closed** |
|
|
| **Created** | 2024-09-28 |
|
|
| **Updated** | 2024-09-28 |
|
|
|
|
---
|
|
|
|
#### Description
|
|
|
|
Phi3.5-mini uses a combined `QKV` tensor. As a result, the quantization mix strategies used for sub-3-bit quants fail. This PR fixes it, and here is what we get as quantization error using wiki text perplexity
|
|
|
|
 |