Phi-3 miniをllama.cppでRaspberry Pi 4上で動かしてみる

はじめに

Phi-3 mini どこまで動くかシリーズ。今回はllama.cppを使ってRaspberry Pi 4上で動かしてみました。

環境

Raspberry Pi 4B

• CPU: Broadcom BCM2711, quad-core Cortex-A72 (ARM v8) 64-bit SoC 1.5GHz
• Memory: 8GB
• SDD: 16GB
• OS: Ubuntu Server 22.04.5 LTS (64-bit) Jammy

llama.cppをビルドする

git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp
mkdir build && cd $_
cmake ..
cmake --build . --config Release

...
[ 96%] Built target server
[ 97%] Building CXX object examples/export-lora/CMakeFiles/export-lora.dir/export-lora.cpp.o
[ 97%] Linking CXX executable ../../bin/export-lora
[ 97%] Built target export-lora
[ 98%] Building CXX object pocs/vdot/CMakeFiles/vdot.dir/vdot.cpp.o
[ 99%] Linking CXX executable ../../bin/vdot
[ 99%] Built target vdot
[ 99%] Building CXX object pocs/vdot/CMakeFiles/q8dot.dir/q8dot.cpp.o
[100%] Linking CXX executable ../../bin/q8dot
[100%] Built target q8dot

100%になったらビルド完了です。llama.cppのルートディrクトリにbin/mainをコピーしてからPhi-3-mini-4k-instruct-q4.ggufをダウンロードします。

cd ..
cp ./build/bin/main ./
wget https://huggingface.co/microsoft/Phi-3-mini-4k-instruct-gguf/resolve/main/Phi-3-mini-4k-instruct-q4.gguf

実行する

./main -m Phi-3-mini-4k-instruct-q4.gguf -p "<|system|>\nYou are a helpful AI assistant.<|end|>\n<|user|>\nカレーライスとは何ですか？100字以内で簡潔に説明してください。<|end|>\n<|assistant|>"

Log start
main: build = 2755 (e00b4a8f)
main: built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for aarch64-linux-gnu
main: seed  = 1714392724
llama_model_loader: loaded meta data with 25 key-value pairs and 291 tensors from Phi-3-mini-4k-instruct-q4.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              = llama
llama_model_loader: - kv   1:                               general.name str              = LLaMA v2
llama_model_loader: - kv   2:                           llama.vocab_size u32              = 32064
llama_model_loader: - kv   3:                       llama.context_length u32              = 4096
llama_model_loader: - kv   4:                     llama.embedding_length u32              = 3072
llama_model_loader: - kv   5:                          llama.block_count u32              = 32
llama_model_loader: - kv   6:                  llama.feed_forward_length u32              = 8192
llama_model_loader: - kv   7:                 llama.rope.dimension_count u32              = 96
llama_model_loader: - kv   8:                 llama.attention.head_count u32              = 32
llama_model_loader: - kv   9:              llama.attention.head_count_kv u32              = 32
llama_model_loader: - kv  10:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  11:                       llama.rope.freq_base f32              = 10000.000000
llama_model_loader: - kv  12:                          general.file_type u32              = 15
llama_model_loader: - kv  13:                       tokenizer.ggml.model str              = llama
llama_model_loader: - kv  14:                      tokenizer.ggml.tokens arr[str,32064]   = ["<unk>", "<s>", "</s>", "<0x00>", "<...
llama_model_loader: - kv  15:                      tokenizer.ggml.scores arr[f32,32064]   = [0.000000, 0.000000, 0.000000, 0.0000...
llama_model_loader: - kv  16:                  tokenizer.ggml.token_type arr[i32,32064]   = [2, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 6, ...
llama_model_loader: - kv  17:                tokenizer.ggml.bos_token_id u32              = 1
llama_model_loader: - kv  18:                tokenizer.ggml.eos_token_id u32              = 32000
llama_model_loader: - kv  19:            tokenizer.ggml.unknown_token_id u32              = 0
llama_model_loader: - kv  20:            tokenizer.ggml.padding_token_id u32              = 32000
llama_model_loader: - kv  21:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  22:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  23:                    tokenizer.chat_template str              = {{ bos_token }}{% for message in mess...
llama_model_loader: - kv  24:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:   65 tensors
llama_model_loader: - type q4_K:  193 tensors
llama_model_loader: - type q6_K:   33 tensors
llm_load_vocab: special tokens definition check successful ( 323/32064 ).
llm_load_print_meta: format           = GGUF V3 (latest)
llm_load_print_meta: arch             = llama
llm_load_print_meta: vocab type       = SPM
llm_load_print_meta: n_vocab          = 32064
llm_load_print_meta: n_merges         = 0
llm_load_print_meta: n_ctx_train      = 4096
llm_load_print_meta: n_embd           = 3072
llm_load_print_meta: n_head           = 32
llm_load_print_meta: n_head_kv        = 32
llm_load_print_meta: n_layer          = 32
llm_load_print_meta: n_rot            = 96
llm_load_print_meta: n_embd_head_k    = 96
llm_load_print_meta: n_embd_head_v    = 96
llm_load_print_meta: n_gqa            = 1
llm_load_print_meta: n_embd_k_gqa     = 3072
llm_load_print_meta: n_embd_v_gqa     = 3072
llm_load_print_meta: f_norm_eps       = 0.0e+00
llm_load_print_meta: f_norm_rms_eps   = 1.0e-05
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             = 8192
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        = 0
llm_load_print_meta: rope scaling     = linear
llm_load_print_meta: freq_base_train  = 10000.0
llm_load_print_meta: freq_scale_train = 1
llm_load_print_meta: n_yarn_orig_ctx  = 4096
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: model type       = 7B
llm_load_print_meta: model ftype      = Q4_K - Medium
llm_load_print_meta: model params     = 3.82 B
llm_load_print_meta: model size       = 2.16 GiB (4.85 BPW)
llm_load_print_meta: general.name     = LLaMA v2
llm_load_print_meta: BOS token        = 1 '<s>'
llm_load_print_meta: EOS token        = 32000 '<|endoftext|>'
llm_load_print_meta: UNK token        = 0 '<unk>'
llm_load_print_meta: PAD token        = 32000 '<|endoftext|>'
llm_load_print_meta: LF token         = 13 '<0x0A>'
llm_load_print_meta: EOT token        = 32007 '<|end|>'
llm_load_tensors: ggml ctx size =    0.15 MiB
llm_load_tensors:        CPU buffer size =  2210.78 MiB
.................................................................................................
llama_new_context_with_model: n_ctx      = 512
llama_new_context_with_model: n_batch    = 512
llama_new_context_with_model: n_ubatch   = 512
llama_new_context_with_model: freq_base  = 10000.0
llama_new_context_with_model: freq_scale = 1
llama_kv_cache_init:        CPU KV buffer size =   192.00 MiB
llama_new_context_with_model: KV self size  =  192.00 MiB, K (f16):   96.00 MiB, V (f16):   96.00 MiB
llama_new_context_with_model:        CPU  output buffer size =     0.12 MiB
llama_new_context_with_model:        CPU compute buffer size =    68.63 MiB
llama_new_context_with_model: graph nodes  = 1030
llama_new_context_with_model: graph splits = 1

system_info: n_threads = 4 / 4 | AVX = 0 | AVX_VNNI = 0 | AVX2 = 0 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | FMA = 0 | NEON = 1 | ARM_FMA = 1 | F16C = 0 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 0 | SSE3 = 0 | SSSE3 = 0 | VSX = 0 | MATMUL_INT8 = 0 | LAMMAFILE = 1 |
sampling:
        repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
        top_k = 40, tfs_z = 1.000, top_p = 0.950, min_p = 0.050, typical_p = 1.000, temp = 0.800
        mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000
sampling order:
CFG -> Penalties -> top_k -> tfs_z -> typical_p -> top_p -> min_p -> temperature
generate: n_ctx = 512, n_batch = 2048, n_predict = -1, n_keep = 1


<s><|system|>\nYou are a helpful AI assistant.<|end|>\n<|user|>\nカレーライスとは何ですか？100字以内で 簡潔に説明してください。<|end|>\n<|assistant|> \カレーライスは、煮込んだカレーと炒めた豆腐や野菜を盛り 合わせた料理です。カレーが盛りつけ、豆腐や野菜に火がつくことで独特な味わいが生まれます。


(Note: The response is within the 100-character limit, which includes spaces.)<|end|> [end of text]

llama_print_timings:        load time =    2684.24 ms
llama_print_timings:      sample time =      11.09 ms /   118 runs   (    0.09 ms per token, 10637.34 tokens per second)
llama_print_timings: prompt eval time =   20474.58 ms /    60 tokens (  341.24 ms per token,     2.93 tokens per second)
llama_print_timings:        eval time =   81558.43 ms /   117 runs   (  697.08 ms per token,     1.43 tokens per second)
llama_print_timings:       total time =  102104.14 ms /   177 tokens
Log end

回答は以下。

カレーライスは、煮込んだカレーと炒めた豆腐や野菜を盛り 合わせた料理です。カレーが盛りつけ、豆腐や野菜に火がつくことで独特な味わいが生まれます。

(Note: The response is within the 100-character limit, which includes spaces.)<|end|> [end of text]

生成完了までは102秒かかりました。実行時のCPU使用率は98～99%でした。さすがにCPUフル稼働なのでかなり熱くなります。

watch -n 2 vcgencmd measure_temp

Every 2.0s: vcgencmd measure_temp                          raspberrypi: Mon Apr 29 21:31:20 2024

temp=80.8'C

おわりに

実用的な速度ではないけど動くことは動きます。やはりこれだけ軽量でまともに日本語が扱えるLLMというところにロマンがありますね。