Gemma 3 270M を Raspberry Pi 400 で動かす

Tue, 19 Aug 2025 18:00:00 +0900

Gemma 3 270MモデルがRaspberry Pi 400上で動作させることが確認されました。

ビルドと実行方法

ビルドコマンド

以下のコマンドを使用して、llama.cppをビルドします。

cmake -B build -DGGML_NATIVE=ON -DLLAMA_NEON=ON -DLLAMA_CURL=OFF
cmake --build build --config Release -j4

ベンチマーク実行ビルド後、ベンチマークを実行します。

./llama.cpp/build/bin/llama-bench -m models/gemma-3-270m-it-Q2_K.gguf

ベンチマーク結果は以下の通りです。

model	size	params	backend	threads	test	t/s
gemma3 270M Q2_K - Medium	219.87 MiB	268.10 M	CPU	4	pp512	47.36 ± 0.04
gemma3 270M Q2_K - Medium	219.87 MiB	268.10 M	CPU	4	tg128	2.48 ± 0.00

対話モードの実行とテスト

以下のコマンドで対話モードを実行します。

./llama.cpp/build/bin/llama-cli -m models/gemma-3-270m-it-Q2_K.gguf

build: 6201 (9d262f4b) main: llama backend init main: load the model llama_model_loader: llama_model_loader: llama_model_loader: - kv 0: llama_model_loader: - kv 1: llama_model_loader: - kv 2: llama_model_loader: - kv 3: llama_model_loader: - kv 4: llama_model_loader: - kv 5: llama_model_loader: - kv 6: llama_model_loader: - kv 7: llama_model_loader: - kv 8: llama_model_loader: - kv 9: llama_model_loader: - kv 10: llama_model_loader: - kv 11: llama_model_loader: - kv 12: llama_model_loader: - kv 13: llama_model_loader: - kv 14: llama_model_loader: - kv 15: llama_model_loader: - kv 16: llama_model_loader: - kv 17: llama_model_loader: - kv 18: llama_model_loader: - kv 19: llama_model_loader: - kv 20: llama_model_loader: - kv 21: llama_model_loader: - kv 22: llama_model_loader: - kv 23: llama_model_loader: - kv 24: llama_model_loader: - kv 25: llama_model_loader: - kv 26: llama_model_loader: - kv 27: llama_model_loader: - kv 28: llama_model_loader: - kv 29: llama_model_loader: - kv 30: llama_model_loader: - kv 31: llama_model_loader: - kv 32: llama_model_loader: - kv 33: llama_model_loader: - kv 34: llama_model_loader: - kv 35: llama_model_loader: - kv 36: llama_model_loader: - kv 37: llama_model_loader: - kv 38: llama_model_loader: - kv 39: llama_model_loader: - kv 40: llama_model_loader: - kv 41: llama_model_loader: - kv 42: llama_model_loader: - kv 43: llama_model_loader: - kv 44: llama_model_loader: - type llama_model_loader: - type q5_0: llama_model_loader: - type q8_0: llama_model_loader: - type q3_K: llama_model_loader: - type iq4_nl: print_info: file format print_info: file type = Q2_K - Medium print_info: file size load: printing all EOG tokens: load: - 106 (’<end load: special tokens cache size = 6414 load: token to piece print_info: arch = gemma3 print_info: vocab_only = 0 print_info: n_ctx_train = 32768 print_info: n_embd = 640 print_info: n_layer = 18 print_info: n_head = 4 print_info: n_head_kv = 1 print_info: n_rot = 256 print_info: n_swa = 512 print_info: is_swa_any = 1 print_info: n_embd_head_k = 256 print_info: n_embd_head_v = 256 print_info: n_gqa = 4 print_info: n_embd_k_gqa = 256 print_info: n_embd_v_gqa = 256 print_info: f_norm_eps = 0.0e+00 print_info: f_norm_rms_eps = 1.0e-06 print_info: f_clamp_kqv = 0.0e+00 print_info: f_max_alibi_bias = 0.0e+00 print_info: f_logit_scale = 0.0e+00 print_info: f_attn_scale = 6.2e-02 print_info: n_ff = 2048 print_info: n_expert = 0 print_info: n_expert_used = 0 print_info: causal attn = 1 print_info: pooling type = 0 print_info: rope type = 2 print_info: rope scaling = linear print_info: freq_base_train print_info: freq_scale_train = 1 print_info: n_ctx_orig_yarn = 32768 print_info: rope_finetuned = unknown print_info: model type = ?B print_info: model params = 268.10 M print_info: general.name print_info: vocab type = SPM print_info: n_vocab = 262144 print_info: n_merges = 0 print_info: BOS token print_info: EOS token print_info: EOT token print_info: UNK token print_info: PAD token print_info: LF token print_info: EOG token print_info: max token length = 48 load_tensors: loading load_tensors: CPU_Mapped ………&hell llama_context: constructing llama_context: n_seq_max = 1 llama_context: n_ctx = 4096 llama_context: n_ctx_per_seq = 4096 llama_context: n_batch = 2048 llama_context: n_ubatch = 512 llama_context: causal_attn = 1 llama_context: flash_attn = 0 llama_context: kv_unified = false llama_context: freq_base llama_context: freq_scale = 1 llama_context: n_ctx_per_seq llama_context: llama_kv_cache_unified_iswa: llama_kv_cache_unified: llama_kv_cache_unified: size = llama_kv_cache_unified_iswa: creating llama_kv_cache_unified: llama_kv_cache_unified: size = llama_context: llama_context: graph nodes = 799 llama_context: graph splits = 1 common_init_from_params: common_init_from_params: common_init_from_params: common_init_from_params: main: llama threadpool main: chat template main: chat template example: <start_of_turn>user You are a helpful assistant with cc (Debian 12.2.0-14+deb12u1) 12.2.0 for aarch64-linux-gnu and apply lora adapter, if any loaded meta data with 45 key-value pairs and 236 tensors from models/gemma-3-270m-it-Q2_K.gguf (version GGUF V3 (latest)) Dumping metadata keys/values. Note: KV overrides do not apply in this output. general.architecture str = gemma3 general.type str = model general.name str = Gemma-3-270M-It general.finetune str = it general.basename str = Gemma-3-270M-It general.quantized_by str = Unsloth general.size_label str = 270M general.license str = gemma general.repo_url str = https://huggingface.co/unsloth general.base_model.count u32 = 1 general.base_model.0.name str = Gemma 3 270m It general.base_model.0.organization str = Gg Hf Gm general.base_model.0.repo_url str = https://huggingface.co/gg-hf-gm/gemma… general.tags arr[str,5] = [“gemma3”, “unsloth”, “gemma”, “googl… gemma3.context_length u32 = 32768 gemma3.embedding_length u32 = 640 gemma3.block_count u32 = 18 gemma3.feed_forward_length u32 = 2048 gemma3.attention.head_count u32 = 4 gemma3.attention.layer_norm_rms_epsilon f32 = 0.000001 gemma3.attention.key_length u32 = 256 gemma3.attention.value_length u32 = 256 gemma3.rope.freq_base f32 = 1000000.000000 gemma3.attention.sliding_window u32 = 512 gemma3.attention.head_count_kv u32 = 1 tokenizer.ggml.model str = llama tokenizer.ggml.pre str = default tokenizer.ggml.tokens arr[str,262144] = ["”, “”, “”, “”, … tokenizer.ggml.scores arr[f32,262144] = [-1000.000000, -1000.000000, -1000.00… tokenizer.ggml.token_type arr[i32,262144] = [3, 3, 3, 3, 3, 4, 3, 3, 3, 3, 3, 3, … tokenizer.ggml.bos_token_id u32 = 2 tokenizer.ggml.eos_token_id u32 = 106 tokenizer.ggml.unknown_token_id u32 = 3 tokenizer.ggml.padding_token_id u32 = 0 tokenizer.ggml.add_bos_token bool = true tokenizer.ggml.add_sep_token bool = false tokenizer.ggml.add_eos_token bool = false tokenizer.chat_template str = {# Unsloth Chat template fixes #}\n{{ … tokenizer.ggml.add_space_prefix bool = false general.quantization_version u32 = 2 general.file_type u32 = 10 quantize.imatrix.file str = gemma-3-270m-it-GGUF/imatrix_unsloth…. quantize.imatrix.dataset str = unsloth_calibration_gemma-3-270m-it.txt quantize.imatrix.entries_count u32 = 126 quantize.imatrix.chunks_count u32 = 141 f32: 109 tensors 18 tensors 1 tensors 36 tensors 72 tensors = GGUF V3 (latest) = 219.87 MiB (6.88 BPW) _of_turn>’) cache size = 1.9446 MB = 1000000.0 = Gemma-3-270M-It = 2 ‘’ = 106 ‘<end_of_turn>’ = 106 ‘<end_of_turn>’ = 3 ‘’ = 0 ‘’ = 248 ‘<0x0A>’ = 106 ‘<end_of_turn>’ model tensors, this can take a while… (mmap = true) model buffer size = 219.87 MiB ip;………… llama_context = 1000000.0 (4096) < n_ctx_train (32768) – the full capacity of the model will not be utilized CPU output buffer size = 1.00 MiB creating non-SWA KV cache, size = 4096 cells CPU KV buffer size = 12.00 MiB 12.00 MiB ( 4096 cells, 3 layers, 1/1 seqs), K (f16): 6.00 MiB, V (f16): 6.00 MiB SWA KV cache, size = 1024 cells CPU KV buffer size = 15.00 MiB 15.00 MiB ( 1024 cells, 15 layers, 1/1 seqs), K (f16): 7.50 MiB, V (f16): 7.50 MiB CPU compute buffer size = 513.25 MiB KV cache shifting is not supported for this context, disabling KV cache shifting added <end_of_turn> logit bias = -inf setting dry_penalty_last_n to ctx_size = 4096 warming up the model with an empty run - please wait … (–no-warmup to disable) init, n_threads = 4 is available, enabling conversation mode (disable it with -no-cnv)

Llama.cpp on Grayrecord Technow Blog

Gemma 3 270M を Raspberry Pi 400 で動かす

ビルドと実行方法