Files
ik_llama.cpp/github-data/pull_requests/155 - IQ2_XS_R4.md
2025-07-23 13:31:53 +02:00

41 lines
1.9 KiB
Markdown

### 🔀 [#155](https://github.com/ikawrakow/ik_llama.cpp/pull/155) - IQ2_XS_R4
| **Author** | `ikawrakow` |
| :--- | :--- |
| **State** | ❌ **Closed** |
| **Created** | 2024-12-21 |
| **Updated** | 2024-12-21 |
---
#### Description
Sub-4 bpw i-quants have a terrible CPU performance, so I was curious to see if we can improve by interleaving rows.
This PR adds `IQ2_XS_R4`, a 4-row interleaved version of `IQ2_XS`.
We get very modest performance gains. I guess, the combination of loading data from a large table, blocks of 16 quants, and perhaps me not having found the optimum bit packing kills the performance.
Anyway, here is `PP-512` for LLaMA-3.1-8B on `Zen4` (Ryzen-7950X), `ARM_NEON` (M2-Max) and `AVX2` (Ryzen-5975WX)
| Platform | Threads | IQ2_XS | IQ2_XS_R4 | Speedup |
| ---: | ---: | ---: | ---: | ---: |
| ARM_NEON | 8 | 45.55 ± 0.28 | 54.13 ± 0.19 | 1.188 |
| Zen4 | 16 | 135.43 ± 0.65 | 156.55 ± 0.51 | 1.156 |
| AVX2 | 32 | 157.34 ± 0.27 | 192.60 ± 0.37 | 1.224 |
We get some performance gains for TG as well, especially on `AVX2`.
Here results for TG-128 on LLaMA-3.1-8B with different numbers of threads:
| Platform | Threads | IQ2_XS | IQ2_XS_R4 | Speedup |
| ---: | ---: | ---: | ---: | ---: |
| ARM_NEON | 2 | 5.10 ± 0.02 | 5.91 ± 0.01 | 1.159 |
| | 4 | 9.71 ± 0.09 | 10.90 ± 0.03 | 1.123 |
| | 8 | 17.21 ± 0.77 | 19.30 ± 0.56 | 1.121 |
| Zen4 | 2 | 6.54 ± 0.01 | 6.90 ± 0.00 | 1.055 |
| | 4 | 12.23 ± 0.02 | 12.79 ± 0.00 | 1.046 |
| | 8 | 21.19 ± 0.01 | 22.12 ± 0.01 | 1.044 |
| AVX2 | 2 | 3.16 ± 0.00 | 4.54 ± 0.00 | 1.437 |
| | 4 | 6.13 ± 0.00 | 8.75 ± 0.00 | 1.427 |
| | 8 | 11.31 ± 0.05 | 15.67 ± 0.05 | 1.385 |
| | 16 | 19.41 ± 0.01 | 22.28 ± 0.00 | 1.148 |