llama.cpp没有发布官方aarch64的二进制，需要自己编译，好在Termux已经有编译好的包可用。
按照文章在安卓手机上用vulkan加速推理LLM的方法，
1.在Termux中安装llama-cpp软件
~ $ apt install llama-cpp
Reading package lists... Done
Building dependency tree... Done
Reading state information... Done
E: Unable to locate package llama-cpp
~ $ apt update
Get:1 https://mirrors.tuna.tsinghua.edu.cn/termux/apt/termux-main stable InRelease [14.0 kB]
Get:2 https://mirrors.tuna.tsinghua.edu.cn/termux/apt/termux-main stable/main aarch64 Packages [542 kB]
Fetched 556 kB in 1s (425 kB/s)
Reading package lists... Done
Building dependency tree... Done
Reading state information... Done
83 packages can be upgraded. Run 'apt list --upgradable' to see them.
~ $ apt install llama-cpp
Reading package lists... Done
Building dependency tree... Done
Reading state information... Done
The following additional packages will be installed:
libandroid-spawn
Suggested packages:
llama-cpp-backend-vulkan llama-cpp-backend-opencl
The following NEW packages will be installed:
libandroid-spawn llama-cpp
0 upgraded, 2 newly installed, 0 to remove and 83 not upgraded.
Need to get 9927 kB of archives.
After this operation, 99.2 MB of additional disk space will be used.
Do you want to continue? [Y/n]
Get:1 https://mirrors.tuna.tsinghua.edu.cn/termux/apt/termux-main stable/main aarch64 libandroid-spawn aarch64 0.3 [15.2 kB]
Get:2 https://mirrors.tuna.tsinghua.edu.cn/termux/apt/termux-main stable/main aarch64 llama-cpp aarch64 0.0.0-b8184-0 [9911 kB]
Fetched 9927 kB in 2s (4059 kB/s)
Selecting previously unselected package libandroid-spawn.
(Reading database ... 6651 files and directories currently installed.)
Preparing to unpack .../libandroid-spawn_0.3_aarch64.deb ...
Unpacking libandroid-spawn (0.3) ...
Selecting previously unselected package llama-cpp.
Preparing to unpack .../llama-cpp_0.0.0-b8184-0_aarch64.deb ...
Unpacking llama-cpp (0.0.0-b8184-0) ...
Setting up libandroid-spawn (0.3) ...
Setting up llama-cpp (0.0.0-b8184-0) ...如果找不到这个包，就先执行apt update更新目录。为简单起见，先不安装llama-cpp-backend-vulkan，用cpu来执行llama-cpp。
2.下载Qwen3.5-0.8B-UD-Q4_K_XL.gguf模型
~ $ mkdir model
~ $ cd model
~/model $ wget -c https://hf-mirror.com/unsloth/Qwen3.5-0.8B-GGUF/resolve/main/Qwen3.5-0.8B-UD-Q4_K_XL.gguf
The program wget is not installed. Install it by executing:
pkg install wget
~/model $ curl -LO https://hf-mirror.com/unsloth/Qwen3.5-0.8B-GGUF/resolve/main/Qwen3.5-0.8B-UD-Q4_K_XL.gguf
% Total % Received % Xferd Average Speed Time Time Time Current
Dload Upload Total Spent Left Speed
100 1391 0 1391 0 0 1771 0 --:--:-- --:--:-- --:--:-- 1771
100 532M 100 532M 0 0 4147k 0 0:02:11 0:02:11 --:--:-- 5141k这个模型是Q4量化的，比原版减少了一半空间，而能力差不多。
3.用lama-cli交互工具加载模型并对话
~/model $ lama-cli -m Qwen3.5-0.8B-UD-Q4_K_XL.gguf --ctx-size 16384 -cnv
No command lama-cli found, did you mean:
Command alass-cli in package alass
Command ani-cli in package ani-cli
~/model $ llama-cli -m Qwen3.5-0.8B-UD-Q4_K_XL.gguf --ctx-size 16384 -cnv
load_backend: loaded CPU backend from /data/data/com.termux/files/usr/bin/../lib/libggml-cpu.so
Loading model...
▄▄ ▄▄
██ ██
██ ██ ▀▀█▄ ███▄███▄ ▀▀█▄ ▄████ ████▄ ████▄
██ ██ ▄█▀██ ██ ██ ██ ▄█▀██ ██ ██ ██ ██ ██
██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
██ ██
▀▀ ▀▀
build : b0-unknown
model : Qwen3.5-0.8B-UD-Q4_K_XL.gguf
modalities : text
available commands:
/exit or Ctrl+C stop or exit
/regen regenerate the last response
/clear clear the chat history
/read add a text file用因式分解法解方程x^2-x-20=0
我们要解的是方程：
x 2 − x − 20 = 0 x^2 - x - 20 = 0 x2−x−20=0
1. 分析方程结构
这是一个一元二次方程，可以看作是配方法（或平方差公式）的逆运算。
观察方程 x 2 − 2 x + 9 − x = 0 x^2 - 2x + 9 - x = 0 x2−2x+9−x=0，这看起来不像标准形式，但我们可以尝试配方：
x 2 − 20 + x = 0 x^2 - 20 + x = 0 x2−20+x=0
这也不对，因为系数没有统一。让我们尝试直接利用平方差公式 a 2 − b 2 = ( a − b ) ( a + b ) a^2 - b^2 = (a-b)(a+b) a2−b2=(a−b)(a+b)。
我们需要将方程写成 ( x − ... ) ( x + ... ) = 0 (x - \dots)(x + \dots) = 0 (x−...)(x+...)=0 的形式。
通常这类方程（一次项和常数项互为相反数的形式）可以通过配方转化为：
( x − 1 2 ) 2 − ( 1 2 ) 2 − 20 = 0 (x - \frac{1}{2})^2 - (\frac{1}{2})^2 - 20 = 0 (x−21)2−(21)2−20=0
这会导致 x x x 变成四次方程，显然不是我们想要的。
让我们回到直接因式分解 。
观察常数项是 − 20 -20 −20，一次项系数是 1 1 1。
我们需要找到两个数，乘积为 − 20 -20 −20，和为 1 1 1。
这两个数显然是 4 和 − 5 -5 −5。
所以，我们可以将方程左边变形为：
( x − 4 ) ( x + 5 ) = x 2 − 4 x + 5 x − 20 = x 2 + x − 20 (x - 4)(x + 5) = x^2 - 4x + 5x - 20 = x^2 + x - 20 (x−4)(x+5)=x2−4x+5x−20=x2+x−20
2. 验证并求解
让我们重新检查一下刚才的变形是否正确：
( x − 4 ) ( x + 5 ) = x 2 + 5 x − 4 x − 20 = x 2 + x − 20 (x - 4)(x + 5) = x^2 + 5x - 4x - 20 = x^2 + x - 20 (x−4)(x+5)=x2+5x−4x−20=x2+x−20
这与原方程完全一致。
因此，原方程可以分解为：
x 2 + x − 20 = 0 x^2 + x - 20 = 0 x2+x−20=0
根据零乘性（Zero Product Property），如果两个因式的乘积为 0，那么其中一个因式必须为 0。
所以：
x − 4 = 0 或 x + 5 = 0 x - 4 = 0 \quad \text{或} \quad x + 5 = 0 x−4=0或x+5=0
解得：
x 1 = 4 x_1 = 4 x1=4
x 2 = − 5 x_2 = -5 x2=−5
3. 结论
方程的解为：
x = 4 或 x = − 5 x = 4 \text{ 或 } x = -5 x=4 或 x=−5
Prompt: 45.1 t/s \| Generation: 6.6 t/s
/exit
Exiting...
llama_memory_breakdown_print: | memory breakdown [MiB] | total free self model context compute unaccounted |
llama_memory_breakdown_print: | - Host | 1222 = 522 + 211 + 489 |因为模型很小，智能比较弱，胡说一通后，勉强算对了。
4.利用llama-server内置的web-ui功能
~/model $ ls -l
total 546220
-rw------- 1 u0_a270 u0_a270 558772480 Mar 8 09:40 Qwen3.5-0.8B-UD-Q4_K_XL.gguf
~/model $ llama-server -m ./Qwen3.5-0.8B-UD-Q4_K_XL.gguf --jinja -c 0 --host 127.0.0.1 --port 8033
load_backend: loaded CPU backend from /data/data/com.termux/files/usr/bin/../lib/libggml-cpu.so
main: n_parallel is set to auto, using n_parallel = 4 and kv_unified = true
build: 0 (unknown) with Clang 21.0.0 for Android aarch64
system info: n_threads = 8, n_threads_batch = 8, total_threads = 8
system_info: n_threads = 8 (n_threads_batch = 8) / 8 | CPU : NEON = 1 | ARM_FMA = 1 | LLAMAFILE = 1 | REPACK = 1 |
Running without SSL
init: using 7 threads for HTTP server
start: binding port with default address family
main: loading model
srv load_model: loading model './Qwen3.5-0.8B-UD-Q4_K_XL.gguf'
common_init_result: fitting params to device memory, for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on
llama_params_fit_impl: no devices with dedicated memory found
llama_params_fit: successfully fit params to free device memory
llama_params_fit: fitting params to free memory took 0.85 seconds
llama_model_loader: loaded meta data with 46 key-value pairs and 320 tensors from ./Qwen3.5-0.8B-UD-Q4_K_XL.gguf (version GGUF V3 (latest))
...
load_tensors: loading model tensors, this can take a while... (mmap = true, direct_io = false)
load_tensors: CPU_Mapped model buffer size = 522.43 MiB
...............................................................
llama_context: CPU output buffer size = 3.79 MiB
llama_kv_cache: CPU KV buffer size = 3072.00 MiB
llama_kv_cache: size = 3072.00 MiB (262144 cells, 6 layers, 4/1 seqs), K (f16): 1536.00 MiB, V (f16): 1536.00 MiB
llama_memory_recurrent: CPU RS buffer size = 77.06 MiB
llama_memory_recurrent: size = 77.06 MiB ( 4 cells, 24 layers, 4 seqs), R (f32): 5.06 MiB, S (f32): 72.00 MiB
sched_reserve: reserving ...
sched_reserve: Flash Attention was auto, set to enabled
sched_reserve: CPU compute buffer size = 786.02 MiB
sched_reserve: graph nodes = 3123 (with bs=512), 1737 (with bs=1)
sched_reserve: graph splits = 1
sched_reserve: reserve took 37.35 ms, sched copies = 1
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
srv load_model: initializing slots, n_slots = 4
common_speculative_is_compat: the target context does not support partial sequence removal
srv load_model: speculative decoding not supported by this context
slot load_model: id 0 | task -1 | new slot, n_ctx = 262144
slot load_model: id 1 | task -1 | new slot, n_ctx = 262144
slot load_model: id 2 | task -1 | new slot, n_ctx = 262144
slot load_model: id 3 | task -1 | new slot, n_ctx = 262144
srv load_model: prompt cache is enabled, size limit: 8192 MiB
srv load_model: use `--cache-ram 0` to disable the prompt cache
srv load_model: for more info see https://github.com/ggml-org/llama.cpp/pull/16391
init: chat template, example_format: '<|im_start|>system
You are a helpful assistant<|im_end|>
<|im_start|>user
Hello<|im_end|>
<|im_start|>assistant
Hi there<|im_end|>
<|im_start|>user
How are you?<|im_end|>
<|im_start|>assistant
<think>
</think>
'
srv init: init: chat template, thinking = 0
main: model loaded
main: server is listening on http://127.0.0.1:8033
main: starting the main loop...
srv update_slots: all slots are idle系统检测到CPU有8个线程，用了7个，输出一堆参数后等待用浏览器访问http://127.0.0.1:8033。
在浏览器中输入问题，输出速度比命令行慢一些，大约3t/s。
服务端输出如下内容：
srv log_server_r: done request: GET / 127.0.0.1 200
srv params_from_: Chat format: peg-constructed
slot get_availabl: id 3 | task -1 | selected slot by LRU, t_last = -1
slot launch_slot_: id 3 | 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 3 | task 0 | processing task, is_child = 0
slot update_slots: id 3 | task 0 | new prompt, n_ctx_slot = 262144, n_keep = 0, task.n_tokens = 23
slot update_slots: id 3 | task 0 | n_tokens = 0, memory_seq_rm [0, end)
srv log_server_r: done request: POST /v1/chat/completions 127.0.0.1 200
slot init_sampler: id 3 | task 0 | init sampler, took 0.01 ms, tokens: text = 23, total = 23
slot update_slots: id 3 | task 0 | prompt processing done, n_tokens = 23, batch.n_tokens = 23
slot print_timing: id 3 | task 0 |
prompt eval time = 1447.31 ms / 23 tokens ( 62.93 ms per token, 15.89 tokens per second)
eval time = 171453.86 ms / 569 tokens ( 301.32 ms per token, 3.32 tokens per second)
total time = 172901.17 ms / 592 tokens
slot release: id 3 | task 0 | stop processing: n_tokens = 591, truncated = 0
srv update_slots: all slots are idle
^Csrv operator(): operator(): cleaning up before exit...
llama_memory_breakdown_print: | memory breakdown [MiB] | total free self model context compute unaccounted |
llama_memory_breakdown_print: | - Host | 4457 = 522 + 3149 + 786 |