mboy2k's picture
Upload README.md with huggingface_hub
69cc85f verified
|
Raw History Blame Contribute Delete
951 Bytes
---
license: mit
tags:
- llama.cpp
- cuda
- sm75
- turing
- qwen
---
# llama.cpp CUDA binary (sm_75 / Tesla T4)
Binary `llama-server` build sẵn cho **NVIDIA Tesla T4 (sm_75, Turing)**, dùng để
cache trong kernel Kaggle — tránh build lại từ source (~25 phút) mỗi lần khởi động.
- Commit llama.cpp: `4e416ee` (2026-09-23)
- Kiến trúc hỗ trợ: `qwen35moe` (Qwen3.8-35B-A3B MoE)
- Build flags: `-DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=75`
## Cách dùng trong kernel Kaggle
```bash
mkdir -p /kaggle/working/llama-bin && cd /kaggle/working/llama-bin
for f in llama-server libggml-base.so libggml-cpu.so libggml-cuda.so libggml.so \
libllama-common.so libllama-server-impl.so libllama.so libmtmd.so; do
curl -sL -o "$f" "https://huggingface.co/mboy2k/llama-cpp-sm75-cuda-binary/resolve/main/$f"
done
chmod +x llama-server
```
Sau đó chạy `llama-server` với `LD_LIBRARY_PATH` trỏ vào thư mục này.