Thông báo bảo mật: Bài viết này là ghi chú phát triển suy luận AI cục bộ, chỉ ghi lại các bước kiểm tra môi trường CUDA trên Jetson Orin Nano, biên dịch mã nguồn chính thức llama.cpp, thử nghiệm các mô hình GGUF mã nguồn mở và quan sát hiệu năng. Trang này không cung cấp gói cài đặt có thể thực thi và không yêu cầu bạn nhập tài khoản, mật khẩu, thông tin thanh toán hoặc bất kỳ thông tin nào khác.
Chuẩn bị môi trường
Kiểm tra xem thư mục /usr/local/cuda/bin/nvcc có tồn tại hay không; nếu có, hãy cấu hình môi trường.
# 临时在当前终端生效
export PATH=/usr/local/cuda/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH
# 验证 nvcc 是否能找到
nvcc --version如果 CUDA toolkit 未配置，可按 Jetson 官方软件源准备对应工具包。
sudo -E apt install cuda-toolkit获取 llama.cpp 源码（官方仓库）
git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp准备 ModelScope 命令行工具。
python3 -m pip install modelscope编译
mkdir build
cmake .. -DGGML_CUDA=ON
cmake --build . --config Release -j $(nproc)跑模型
低功耗模式
mkdir model_zoo
modelscope download --model Qwen/Qwen2.5-1.5B-Instruct-GGUF qwen2.5-1.5b-instruct-q4_0.gguf --local_dir ./model_zoo换个笑话./build/bin/llama-cli -m ../model_zoo/qwen2.5-1.5b-instruct-q4_0.gguf --perf --show-timings -lv 3
换个笑话
|slot get_availabl: id 0 | task -1 | selected slot by LCP similarity, sim_best = 0.952 (> 0.100 thold), f_keep = 1.000
slot launch_slot_: id 0 | task -1 | sampler chain: logits -> ?penalties -> ?dry -> ?top-n-sigma -> top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist
slot launch_slot_: id 0 | task 184 | processing task, is_child = 0
slot update_slots: id 0 | task 184 | new prompt, n_ctx_slot = 4096, n_keep = 0, task.n_tokens = 248
slot update_slots: id 0 | task 184 | n_tokens = 236, memory_seq_rm [236, end)
slot update_slots: id 0 | task 184 | prompt processing progress, n_tokens = 248, batch.n_tokens = 12, progress = 1.000000
slot update_slots: id 0 | task 184 | prompt done, n_tokens = 248, batch.n_tokens = 12
slot init_sampler: id 0 | task 184 | init sampler, took 0.11 ms, tokens: text = 248, total = 248
好的，以下是一个不同的笑话：
为什么电脑没有呼吸？
因为它只有一个输入口，一个输出口。slot print_timing: id 0 | task 184 |
prompt eval time = 214.26 ms / 12 tokens ( 17.85 ms per token, 56.01 tokens per second)
eval time = 1691.48 ms / 22 tokens ( 76.89 ms per token, 13.01 tokens per second)
total time = 1905.74 ms / 34 tokens
slot release: id 0 | task 184 | stop processing: n_tokens = 269, truncated = 0
srv update_slots: all slots are idle
[ Prompt: 56.0 t/s | Generation: 13.0 t/s ]看起来速度不是很快。 再跑一个3B的试试看。
./build/bin/llama-cli -m ../model_zoo/qwen2.5-3b-instruct-q4_0.gguf --perf --show-timings -lv 3
prompt eval time = 1412.32 ms / 15 tokens ( 94.15 ms per token, 10.62 tokens per second)
eval time = 48064.10 ms / 244 tokens ( 196.98 ms per token, 5.08 tokens per second)
total time = 49476.41 ms / 259 tokens
slot release: id 0 | task 148 | stop processing: n_tokens = 469, truncated = 0
srv update_slots: all slots are idle
[ Prompt: 10.6 t/s | Generation: 5.1 t/s ]看起来不太行。 加上个参数重新编译试试，针对CUDA的能力计算数值，加这个参数，llama.cpp 的 cmake 默认可能不会启用针对特定 GPU 架构的深度优化
rm -rf build && mkdir build && cd build
cmake .. -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=87
#或者cmake .. -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=native
cmake --build . --config Release -j$(nproc)再测试执行：
./build/bin/llama-cli -m ../model_zoo/qwen2.5-3b-instruct-q4_0.gguf --perf --show-timings -lv 3 -ngl 99
prompt eval time = 201.92 ms / 13 tokens ( 15.53 ms per token, 64.38 tokens per second)
eval time = 39903.36 ms / 386 tokens ( 103.38 ms per token, 9.67 tokens per second)
total time = 40105.28 ms / 399 tokens
slot release: id 0 | task 159 | stop processing: n_tokens = 598, truncated = 0
srv update_slots: all slots are idle
[ Prompt: 64.4 t/s | Generation: 9.7 t/s ]推理速度上来了，稳定在10 token/s。
再跑跑bench
./build/bin/llama-bench -m ../model_zoo/qwen2.5-3b-instruct-q4_0.gguf -ngl 99
ggml_cuda_init: found 1 CUDA devices:
Device 0: Orin, compute capability 8.7, VMM: yes
| model | size | params | backend | ngl | test | t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | --------------: | -------------------: |
| qwen2 3B Q4_0 | 1.86 GiB | 3.40 B | CUDA | 99 | pp512 | 259.35 ± 0.46 |
| qwen2 3B Q4_0 | 1.86 GiB | 3.40 B | CUDA | 99 | tg128 | 8.98 ± 0.03 |
build: 37c35f0e1 (7787)
(base) bianbu@ubuntu:~/llama.cpp$
sudo nvpmodel -m 0 # 开启最大功率模式
sudo jetson_clocks # 锁定 CPU/GPU/内存频率到最高
(base) bianbu@ubuntu:~/llama.cpp$ ./build/bin/llama-bench -m ../model_zoo/qwen2.5-3b-instruct-q4_0.gguf -ngl 99
ggml_cuda_init: found 1 CUDA devices:
Device 0: Orin, compute capability 8.7, VMM: yes
| model | size | params | backend | ngl | test | t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | --------------: | -------------------: |
| qwen2 3B Q4_0 | 1.86 GiB | 3.40 B | CUDA | 99 | pp512 | 264.42 ± 0.27 |
| qwen2 3B Q4_0 | 1.86 GiB | 3.40 B | CUDA | 99 | tg128 | 9.17 ± 0.01 |
build: 37c35f0e1 (7787)超级性能模式
上面跑了一下看起来还是很慢，看了一下jetson官网，要跑性能，需要设置电源为超级性能模式，点击 Ubuntu 桌面顶部栏右侧的 NVIDIA 图标
Power mode
0： 15W
1： 25W
2： MAXN SUPER如果没有MAXN SUPER，执行下面的命令安装。
sudo apt remove nvidia-l4t-bootloader
sudo apt install nvidia-l4t-bootloader
sudo reboot如果还是无法配置的话，参考下：super mode 接下来，接着跑。
./build/bin/llama-cli -m ../model_zoo_gguf/model_zoo/qwen2.5-1.5b-instruct-q4_0.gguf --perf --show-timings -lv 3 -ngl 99
./build/bin/llama-bench -m ../model_zoo_gguf/model_zoo/qwen2.5-1.5b-instruct-q4_0.gguf -ngl 99
在跑个3B的
./build/bin/llama-bench -m ../model_zoo_gguf/model_zoo/qwen2.5-3b-instruct-q4_0.gguf -ngl 99
./build/bin/llama-cli -m ../model_zoo_gguf/model_zoo/qwen2.5-3b-instruct-q4_0.gguf -ngl 99 --perf -lv 3
Leave a Reply