源码编译llama.cpp for android

源码编译llama.cpp for android

我这有已经编译好的版本，直接下载使用：

https://github.com/turingevo/llama.cpp-build/releases/tag/b4331

准备 android-ndk

已下载：

/media/wmx/ws1/software/qtAndroid/Sdk/ndk/23.1.7779620

版本 ： llama.cpp-b4331
下载源码
切换到 llama.cpp/目录

编译脚本 llama.cpp/build-android.sh

#!/bin/bashANDROID_NDK_PATH=/media/wmx/ws1/software/qtAndroid/Sdk/ndk/23.1.7779620
build_dir=build-android
src_dir=.
install_dir=bin/androidcmake \-DCMAKE_TOOLCHAIN_FILE=${ANDROID_NDK_PATH}/build/cmake/android.toolchain.cmake \-DANDROID_ABI=arm64-v8a \-DANDROID_PLATFORM=android-28 \-DCMAKE_C_FLAGS="-march=armv8.7a" \-DCMAKE_CXX_FLAGS="-march=armv8.7a" \-DGGML_OPENMP=OFF \-DGGML_LLAMAFILE=OFF \-B ${build_dir} \-S ${src_dir}cmake --build ${build_dir} --config Release -j48cmake --install ${build_dir} --prefix ${install_dir} --config Release

push 到android设备测试

下面是 华为mate40pro 上的测试结果

build llama.cpp/bin/android

adb shell "mkdir /data/local/tmp/llama.cpp"
adb push bin/android /data/local/tmp/llama.cpp/
adb push qwen2.5-0.5b-instruct-q4_k_m.gguf /data/local/tmp/llama.cpp/adb shell
cd /data/local/tmp/llama.cpp/androidtouch test.sh
chmod a+x test.sh
cat " LD_LIBRARY_PATH=lib ./bin/llama-simple -m qwen2.5-0.5b-instruct-q4_k_m.gguf -p \"你是谁?\"  "  > test.sh./test.shHWNOH:/data/local/tmp/llama.cpp/android $ ./test.sh                                                                                           
llama_model_loader: loaded meta data with 26 key-value pairs and 291 tensors from /sdcard/a-wmx/models/qwen2.5-0.5b-instruct-q4_k_m.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv   0:                       general.architecture str              = qwen2
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = qwen2.5-0.5b-instruct
llama_model_loader: - kv   3:                            general.version str              = v0.1
llama_model_loader: - kv   4:                           general.finetune str              = qwen2.5-0.5b-instruct
llama_model_loader: - kv   5:                         general.size_label str              = 630M
llama_model_loader: - kv   6:                          qwen2.block_count u32              = 24
llama_model_loader: - kv   7:                       qwen2.context_length u32              = 32768
llama_model_loader: - kv   8:                     qwen2.embedding_length u32              = 896
llama_model_loader: - kv   9:                  qwen2.feed_forward_length u32              = 4864
llama_model_loader: - kv  10:                 qwen2.attention.head_count u32              = 14
llama_model_loader: - kv  11:              qwen2.attention.head_count_kv u32              = 2
llama_model_loader: - kv  12:                       qwen2.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  13:     qwen2.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  14:                          general.file_type u32              = 15
llama_model_loader: - kv  15:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  16:                         tokenizer.ggml.pre str              = qwen2
llama_model_loader: - kv  17:                      tokenizer.ggml.tokens arr[str,151936]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  18:                  tokenizer.ggml.token_type arr[i32,151936]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  19:                      tokenizer.ggml.merges arr[str,151387]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  20:                tokenizer.ggml.eos_token_id u32              = 151645
llama_model_loader: - kv  21:            tokenizer.ggml.padding_token_id u32              = 151643
llama_model_loader: - kv  22:                tokenizer.ggml.bos_token_id u32              = 151643
llama_model_loader: - kv  23:               tokenizer.ggml.add_bos_token bool             = false
llama_model_loader: - kv  24:                    tokenizer.chat_template str              = {%- if tools %}\n    {{- '<|im_start|>...
llama_model_loader: - kv  25:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:  121 tensors
llama_model_loader: - type q5_0:  133 tensors
llama_model_loader: - type q8_0:   13 tensors
llama_model_loader: - type q4_K:   12 tensors
llama_model_loader: - type q6_K:   12 tensors
llm_load_vocab: control token: 151659 '<|fim_prefix|>' is not marked as EOG
llm_load_vocab: control token: 151656 '<|video_pad|>' is not marked as EOG
llm_load_vocab: control token: 151655 '<|image_pad|>' is not marked as EOG
llm_load_vocab: control token: 151653 '<|vision_end|>' is not marked as EOG
llm_load_vocab: control token: 151652 '<|vision_start|>' is not marked as EOG
llm_load_vocab: control token: 151651 '<|quad_end|>' is not marked as EOG
llm_load_vocab: control token: 151649 '<|box_end|>' is not marked as EOG
llm_load_vocab: control token: 151648 '<|box_start|>' is not marked as EOG
llm_load_vocab: control token: 151646 '<|object_ref_start|>' is not marked as EOG
llm_load_vocab: control token: 151644 '<|im_start|>' is not marked as EOG
llm_load_vocab: control token: 151661 '<|fim_suffix|>' is not marked as EOG
llm_load_vocab: control token: 151647 '<|object_ref_end|>' is not marked as EOG
llm_load_vocab: control token: 151660 '<|fim_middle|>' is not marked as EOG
llm_load_vocab: control token: 151654 '<|vision_pad|>' is not marked as EOG
llm_load_vocab: control token: 151650 '<|quad_start|>' is not marked as EOG
llm_load_vocab: special tokens cache size = 22
llm_load_vocab: token to piece cache size = 0.9310 MB
llm_load_print_meta: format           = GGUF V3 (latest)
llm_load_print_meta: arch             = qwen2
llm_load_print_meta: vocab type       = BPE
llm_load_print_meta: n_vocab          = 151936
llm_load_print_meta: n_merges         = 151387
llm_load_print_meta: vocab_only       = 0
llm_load_print_meta: n_ctx_train      = 32768
llm_load_print_meta: n_embd           = 896
llm_load_print_meta: n_layer          = 24
llm_load_print_meta: n_head           = 14
llm_load_print_meta: n_head_kv        = 2
llm_load_print_meta: n_rot            = 64
llm_load_print_meta: n_swa            = 0
llm_load_print_meta: n_embd_head_k    = 64
llm_load_print_meta: n_embd_head_v    = 64
llm_load_print_meta: n_gqa            = 7
llm_load_print_meta: n_embd_k_gqa     = 128
llm_load_print_meta: n_embd_v_gqa     = 128
llm_load_print_meta: f_norm_eps       = 0.0e+00
llm_load_print_meta: f_norm_rms_eps   = 1.0e-06
llm_load_print_meta: f_clamp_kqv      = 0.0e+00
llm_load_print_meta: f_max_alibi_bias = 0.0e+00
llm_load_print_meta: f_logit_scale    = 0.0e+00
llm_load_print_meta: n_ff             = 4864
llm_load_print_meta: n_expert         = 0
llm_load_print_meta: n_expert_used    = 0
llm_load_print_meta: causal attn      = 1
llm_load_print_meta: pooling type     = 0
llm_load_print_meta: rope type        = 2
llm_load_print_meta: rope scaling     = linear
llm_load_print_meta: freq_base_train  = 1000000.0
llm_load_print_meta: freq_scale_train = 1
llm_load_print_meta: n_ctx_orig_yarn  = 32768
llm_load_print_meta: rope_finetuned   = unknown
llm_load_print_meta: ssm_d_conv       = 0
llm_load_print_meta: ssm_d_inner      = 0
llm_load_print_meta: ssm_d_state      = 0
llm_load_print_meta: ssm_dt_rank      = 0
llm_load_print_meta: ssm_dt_b_c_rms   = 0
llm_load_print_meta: model type       = 1B
llm_load_print_meta: model ftype      = Q4_K - Medium
llm_load_print_meta: model params     = 630.17 M
llm_load_print_meta: model size       = 462.96 MiB (6.16 BPW) 
llm_load_print_meta: general.name     = qwen2.5-0.5b-instruct
llm_load_print_meta: BOS token        = 151643 '<|endoftext|>'
llm_load_print_meta: EOS token        = 151645 '<|im_end|>'
llm_load_print_meta: EOT token        = 151645 '<|im_end|>'
llm_load_print_meta: PAD token        = 151643 '<|endoftext|>'
llm_load_print_meta: LF token         = 148848 'ÄĬ'
llm_load_print_meta: FIM PRE token    = 151659 '<|fim_prefix|>'
llm_load_print_meta: FIM SUF token    = 151661 '<|fim_suffix|>'
llm_load_print_meta: FIM MID token    = 151660 '<|fim_middle|>'
llm_load_print_meta: FIM PAD token    = 151662 '<|fim_pad|>'
llm_load_print_meta: FIM REP token    = 151663 '<|repo_name|>'
llm_load_print_meta: FIM SEP token    = 151664 '<|file_sep|>'
llm_load_print_meta: EOG token        = 151643 '<|endoftext|>'
llm_load_print_meta: EOG token        = 151645 '<|im_end|>'
llm_load_print_meta: EOG token        = 151662 '<|fim_pad|>'
llm_load_print_meta: EOG token        = 151663 '<|repo_name|>'
llm_load_print_meta: EOG token        = 151664 '<|file_sep|>'
llm_load_print_meta: max token length = 256
llm_load_tensors: tensor 'token_embd.weight' (q5_0) (and 290 others) cannot be used with preferred buffer type CPU_AARCH64, using CPU instead
llm_load_tensors:   CPU_Mapped model buffer size =   462.96 MiB
.....................................................
llama_new_context_with_model: n_batch is less than GGML_KQ_MASK_PAD - increasing to 32
llama_new_context_with_model: n_seq_max     = 1
llama_new_context_with_model: n_ctx         = 64
llama_new_context_with_model: n_ctx_per_seq = 64
llama_new_context_with_model: n_batch       = 32
llama_new_context_with_model: n_ubatch      = 32
llama_new_context_with_model: flash_attn    = 0
llama_new_context_with_model: freq_base     = 1000000.0
llama_new_context_with_model: freq_scale    = 1
llama_new_context_with_model: n_ctx_per_seq (64) < n_ctx_train (32768) -- the full capacity of the model will not be utilized
llama_kv_cache_init:        CPU KV buffer size =     0.75 MiB
llama_new_context_with_model: KV self size  =    0.75 MiB, K (f16):    0.38 MiB, V (f16):    0.38 MiB
llama_new_context_with_model:        CPU  output buffer size =     0.58 MiB
llama_new_context_with_model:        CPU compute buffer size =    18.66 MiB
llama_new_context_with_model: graph nodes  = 846
llama_new_context_with_model: graph splits = 1
-p 你是谁?我是阿里云开发的超大规模语言模型，我叫通义千问。通义是“通义天下”，千问是“千问天下
main: decoded 32 tokens in 2.21 s, speed: 14.49 t/sllama_perf_sampler_print:    sampling time =       5.69 ms /    32 runs   (    0.18 ms per token,  5622.91 tokens per second)
llama_perf_context_print:        load time =    1907.15 ms
llama_perf_context_print: prompt eval time =     165.11 ms /     5 tokens (   33.02 ms per token,    30.28 tokens per second)
llama_perf_context_print:        eval time =    2000.08 ms /    31 runs   (   64.52 ms per token,    15.50 tokens per second)
llama_perf_context_print:       total time =    3950.19 ms /    36 tokens