Files
ik_llama.cpp/github-data/pull_requests/544 - New integer trellis on ARM_NEON.md
2025-07-23 13:31:53 +02:00

35 lines
1.4 KiB
Markdown

### 🔀 [#544](https://github.com/ikawrakow/ik_llama.cpp/pull/544) - New integer trellis on ARM_NEON
| **Author** | `ikawrakow` |
| :--- | :--- |
| **State** | ❌ **Closed** |
| **Created** | 2025-06-20 |
| **Updated** | 2025-06-20 |
---
#### Description
This PR adapts the ARM_NEON trellis implementation to the new integer trellis.
Test done on an M2-Max CPU using LlaMA-3.1-8B-Instruct.
Very respectable PP performance:
| model | size | test | t/s |
| ---------------- | ---------: | ------------: | ---------------: |
| llama 8B IQ2_KT | 2.77 GiB | pp512 | 129.19 ± 0.22 |
| llama 8B IQ3_KT | 3.58 GiB | pp512 | 127.66 ± 0.38 |
| llama 8B IQ4_KT | 4.30 GiB | pp512 | 125.23 ± 0.44 |
Still very low TG performance:
| model | size | test | t/s |
| ---------------- | ---------: | ------------: | ---------------: |
| llama 8B IQ2_KT | 2.77 GiB | tg128 | 12.59 ± 0.15 |
| llama 8B IQ3_KT | 3.58 GiB | tg128 | 9.92 ± 0.02 |
| llama 8B IQ4_KT | 4.30 GiB | tg128 | 9.73 ± 0.05 |
Don't ask Apple Silicon to do too much work with a piece of data fetched from memory.
Nevertheless, compared to PR #471 we observe ~13% speedup for `IQ2_KT`, ~30% speedup for `IQ3_KT`, and nearly 70% speedup for `Q4_KT`.