Add README
Browse files
README.md
ADDED
|
@@ -0,0 +1,63 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
tags:
|
| 3 |
+
- llama-cpp
|
| 4 |
+
- turboquant
|
| 5 |
+
- triattention
|
| 6 |
+
- kv-cache
|
| 7 |
+
- windows
|
| 8 |
+
- cuda
|
| 9 |
+
license: mit
|
| 10 |
+
---
|
| 11 |
+
|
| 12 |
+
# llama.cpp TurboQuant + TriAttention — Windows CUDA 13 Binaries
|
| 13 |
+
|
| 14 |
+
Pre-built Windows x64 Release binaries for the [atomicmilkshake/llama-cpp-turboquant](https://github.com/atomicmilkshake/llama-cpp-turboquant) fork.
|
| 15 |
+
|
| 16 |
+
This builds adds **TurboQuant** (custom quantization) and **TriAttention** (GPU-accelerated KV cache pruning based on [arXiv 2604.04921](https://arxiv.org/abs/2604.04921)) on top of llama.cpp.
|
| 17 |
+
|
| 18 |
+
## Download
|
| 19 |
+
|
| 20 |
+
**[llama-turboquant-triattention-win-cu13-x64.zip](llama-turboquant-triattention-win-cu13-x64.zip)** (~179 MB)
|
| 21 |
+
|
| 22 |
+
## Requirements
|
| 23 |
+
|
| 24 |
+
- Windows 10/11 x64
|
| 25 |
+
- NVIDIA GPU (Turing+, GTX 1600 / RTX 2000 series or newer)
|
| 26 |
+
- CUDA 13.x runtime — install from [developer.nvidia.com/cuda-downloads](https://developer.nvidia.com/cuda-downloads) (the cublasLt64_13.dll is NOT included in the zip due to its 432 MB size)
|
| 27 |
+
|
| 28 |
+
## Usage
|
| 29 |
+
|
| 30 |
+
`
|
| 31 |
+
llama-server.exe -m YourModel.gguf -c 32768 -ngl 99 --port 8080 ^
|
| 32 |
+
--triattention-stats model.triattention ^
|
| 33 |
+
--triattention-budget 4096 ^
|
| 34 |
+
--triattention-window 256 ^
|
| 35 |
+
--triattention-log
|
| 36 |
+
`
|
| 37 |
+
|
| 38 |
+
## TriAttention Performance
|
| 39 |
+
|
| 40 |
+
Tested on Qwen3-8B Q4_K_M, RTX 3080, -c 512, udget=256:
|
| 41 |
+
|
| 42 |
+
| Mode | Prune time | Generation |
|
| 43 |
+
|------|-----------|------------|
|
| 44 |
+
| No pruning | — | 17.5 tok/s |
|
| 45 |
+
| CPU scoring | ~5900 ms/event | 17.5 tok/s |
|
| 46 |
+
| **GPU scoring** | **~4-9 ms/event** | **75.0 tok/s** |
|
| 47 |
+
|
| 48 |
+
~1000x speedup on pruning events; 4.3x overall throughput improvement.
|
| 49 |
+
|
| 50 |
+
## TriAttention Flags
|
| 51 |
+
|
| 52 |
+
| Flag | Description | Default |
|
| 53 |
+
|------|-------------|---------|
|
| 54 |
+
| --triattention-stats <file> | Calibration file (**required** to enable) | — |
|
| 55 |
+
| --triattention-budget <n> | Max KV tokens to retain | 512 |
|
| 56 |
+
| --triattention-window <n> | Recent-token protection window | 64 |
|
| 57 |
+
| --triattention-trigger | slack\|interval\| ill | slack |
|
| 58 |
+
| --triattention-log | Log each prune event | off |
|
| 59 |
+
| --triattention-no-protect-prefill | Allow evicting prompt tokens | off |
|
| 60 |
+
|
| 61 |
+
## Source
|
| 62 |
+
|
| 63 |
+
[github.com/atomicmilkshake/llama-cpp-turboquant](https://github.com/atomicmilkshake/llama-cpp-turboquant) — branch eature/triattention
|