### 🔀 [#138](https://github.com/ikawrakow/ik_llama.cpp/pull/138) - IQ4_K_R4

| **Author** | `ikawrakow` |
| :--- | :--- |
| **State** | ❌ **Closed** |
| **Created** | 2024-12-12 |
| **Updated** | 2024-12-12 |

---

#### Description

On to R4 implementation of the new iqk quants.

First `IQ4_K`

We get very signifiant performance gains on `ARM_NEON` and more modest gains on `AVX2/Zen4`. I suspect my `AVX2/Zen4` implementation is not optimum, but I did not see a better way for now.

Here is `PP-512` for LLaMA-3.1-8B on `Zen4` (Ryzen-7950X), `ARM_NEON` (M2-Max) and `AVX2` (Ryzen-5975WX)

| Platform |  Threads | IQ4_K | IQ4_K_R4 | Speedup |
| ---: | ---: | ---: | ---: | ---: |
| ARM_NEON |  8 |  58.20 ± 1.03  | 108.02 ± 1.10 | 1.856 |
| Zen4            | 16 | 182.20 ± 0.38 | 232.63 ± 0.39  | 1.277 |
| AVX2           | 32 | 206.43 ± 0.49 |  227.60 ± 0.46  | 1.103 |

We get decent performance gains for TG as well.
Here results for TG-128 on LLaMA-3.1-8B with different numbers of threads:

| Platform |  Threads | Q2_K_S | Q2_K_R4 | Speedup |
| ---: | ---: | ---: | ---: | ---: |
| ARM_NEON | 2 | 8.44 ± 0.02 | 10.56 ± 0.01 | 1.251 |
|                      | 4 | 15.90 ± 0.05 | 19.32 ± 0.14 | 1.215 |
|                      | 8 | 24.54 ± 0.15 | 25.16 ± 0.03  | 1.025 |
| Zen4            | 1 |  5.26 ± 0.00  | 6.73 ± 0.00  |  1.279 |
|                      | 2 |  9.71 ± 0.01 | 12.43 ± 0.00  |  1.269 |
|                      | 4 |  13.48 ± 0.06  | 14.00 ± 0.03  |  1.039 |
| AVX2           | 2 | 4.02 ± 0.00   | 6.91 ± 0.00 | 1.719 |
|                     | 4 | 8.03 ± 0.00    |  11.13 ± 0.00 | 1.386 |
|                     | 8 |  11.81 ± 0.00  | 12.75 ± 0.00  | 1.079 |

- [x] I have read the [contributing guidelines](https://github.com/ggerganov/llama.cpp/blob/master/CONTRIBUTING.md)
- Self-reported review complexity:
  - [ ] Low
  - [ ] Medium
  - [ ] High