快速开始
备注
阅读本篇前，请确保已按照 安装指南 准备好昇腾环境及 llama.cpp ！
本教程聚焦大语言模型（Large Language Model，LLM）的推理过程，以 Qwen2.5-7B 模型为例，讲述如何使用 llama.cpp 在昇腾 NPU 上进行推理。
模型文件准备及量化
llama.cpp 的推理需要使用 gguf 格式文件，llama.cpp 提供了两种方式转换 Hugging Face 模型文件：
使用 GGUF-my-repo 将模型进行转换。
使用项目中的 convert_hf_to_gguf.py 文件将 Hugging Face 模型转换为 gguf 格式:
1python convert_hf_to_gguf.py path/to/model
详情请参考 Prepare and Quantize 。
注意：目前仅支持 FP16 精度及 Q4_0/Q8_0 量化模型。
推理
有两种设备选择模式:
单设备：使用用户指定的一个设备目标。
多设备：自动选择具有相同后端的设备。
设备选择 |
参数 |
|---|---|
单设备 |
--split-mode none --main-gpu DEVICE_ID |
多设备 |
--split-mode layer (default) |
使用单卡推理
1./build/bin/llama-cli -m path_to_model -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm none -mg 0
使用多卡推理
1./build/bin/llama-cli -m path_to_model -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm layer
以下为正常推理结果：
1Log start
2main: build = 3520 (8e707118)
3main: built with cc (Ubuntu 9.4.0-1ubuntu1~20.04.2) 9.4.0 for aarch64-linux-gnu
4main: seed = 1728907816
5llama_model_loader: loaded meta data with 22 key-value pairs and 291 tensors from /home/jiahao/models/llama3-8b-instruct-fp16.gguf (version GGUF V3 (latest))
6llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
7llama_model_loader: - kv 0: general.architecture str = llama
8llama_model_loader: - kv 1: general.name str = Meta-Llama-3-8B-Instruct
9llama_model_loader: - kv 2: llama.block_count u32 = 32
10llama_model_loader: - kv 3: llama.context_length u32 = 8192
11llama_model_loader: - kv 4: llama.embedding_length u32 = 4096
12llama_model_loader: - kv 5: llama.feed_forward_length u32 = 14336
13llama_model_loader: - kv 6: llama.attention.head_count u32 = 32
14llama_model_loader: - kv 7: llama.attention.head_count_kv u32 = 8
15llama_model_loader: - kv 8: llama.rope.freq_base f32 = 500000.000000
16llama_model_loader: - kv 9: llama.attention.layer_norm_rms_epsilon f32 = 0.000010
17llama_model_loader: - kv 10: general.file_type u32 = 1
18llama_model_loader: - kv 11: llama.vocab_size u32 = 128256
19llama_model_loader: - kv 12: llama.rope.dimension_count u32 = 128
20llama_model_loader: - kv 13: tokenizer.ggml.model str = gpt2
21llama_model_loader: - kv 14: tokenizer.ggml.pre str = llama-bpe
22llama_model_loader: - kv 15: tokenizer.ggml.tokens arr[str,128256] = ["!", "\"", "#", "$", "%", "&", "'", ...
23llama_model_loader: - kv 16: tokenizer.ggml.token_type arr[i32,128256] = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
24llama_model_loader: - kv 17: tokenizer.ggml.merges arr[str,280147] = ["Ġ Ġ", "Ġ ĠĠĠ", "ĠĠ ĠĠ", "...
25llama_model_loader: - kv 18: tokenizer.ggml.bos_token_id u32 = 128000
26llama_model_loader: - kv 19: tokenizer.ggml.eos_token_id u32 = 128009
27llama_model_loader: - kv 20: tokenizer.chat_template str = {% set loop_messages = messages %}{% ...
28llama_model_loader: - kv 21: general.quantization_version u32 = 2
29llama_model_loader: - type f32: 65 tensors
30llama_model_loader: - type f16: 226 tensors
31llm_load_vocab: special tokens cache size = 256
32llm_load_vocab: token to piece cache size = 0.8000 MB
33llm_load_print_meta: format = GGUF V3 (latest)
34llm_load_print_meta: arch = llama
35llm_load_print_meta: vocab type = BPE
36llm_load_print_meta: n_vocab = 128256
37llm_load_print_meta: n_merges = 280147
38llm_load_print_meta: vocab_only = 0
39llm_load_print_meta: n_ctx_train = 8192
40llm_load_print_meta: n_embd = 4096
41llm_load_print_meta: n_layer = 32
42llm_load_print_meta: n_head = 32
43llm_load_print_meta: n_head_kv = 8
44llm_load_print_meta: n_rot = 128
45llm_load_print_meta: n_swa = 0
46llm_load_print_meta: n_embd_head_k = 128
47llm_load_print_meta: n_embd_head_v = 128
48llm_load_print_meta: n_gqa = 4
49llm_load_print_meta: n_embd_k_gqa = 1024
50llm_load_print_meta: n_embd_v_gqa = 1024
51llm_load_print_meta: f_norm_eps = 0.0e+00
52llm_load_print_meta: f_norm_rms_eps = 1.0e-05
53llm_load_print_meta: f_clamp_kqv = 0.0e+00
54llm_load_print_meta: f_max_alibi_bias = 0.0e+00
55llm_load_print_meta: f_logit_scale = 0.0e+00
56llm_load_print_meta: n_ff = 14336
57llm_load_print_meta: n_expert = 0
58llm_load_print_meta: n_expert_used = 0
59llm_load_print_meta: causal attn = 1
60llm_load_print_meta: pooling type = 0
61llm_load_print_meta: rope type = 0
62llm_load_print_meta: rope scaling = linear
63llm_load_print_meta: freq_base_train = 500000.0
64llm_load_print_meta: freq_scale_train = 1
65llm_load_print_meta: n_ctx_orig_yarn = 8192
66llm_load_print_meta: rope_finetuned = unknown
67llm_load_print_meta: ssm_d_conv = 0
68llm_load_print_meta: ssm_d_inner = 0
69llm_load_print_meta: ssm_d_state = 0
70llm_load_print_meta: ssm_dt_rank = 0
71llm_load_print_meta: model type = 8B
72llm_load_print_meta: model ftype = F16
73llm_load_print_meta: model params = 8.03 B
74llm_load_print_meta: model size = 14.96 GiB (16.00 BPW)
75llm_load_print_meta: general.name = Meta-Llama-3-8B-Instruct
76llm_load_print_meta: BOS token = 128000 '<|begin_of_text|>'
77llm_load_print_meta: EOS token = 128009 '<|eot_id|>'
78llm_load_print_meta: LF token = 128 'Ä'
79llm_load_print_meta: EOT token = 128009 '<|eot_id|>'
80llm_load_print_meta: max token length = 256
81llm_load_tensors: ggml ctx size = 0.27 MiB
82llm_load_tensors: CPU buffer size = 15317.02 MiB
83llm_load_tensors: CANN buffer size = 13313.00 MiB
84.........................................................................................
85llama_new_context_with_model: n_ctx = 8192
86llama_new_context_with_model: n_batch = 2048
87llama_new_context_with_model: n_ubatch = 512
88llama_new_context_with_model: flash_attn = 0
89llama_new_context_with_model: freq_base = 500000.0
90llama_new_context_with_model: freq_scale = 1
91llama_kv_cache_init: CANN KV buffer size = 1024.00 MiB
92llama_new_context_with_model: KV self size = 1024.00 MiB, K (f16): 512.00 MiB, V (f16): 512.00 MiB
93llama_new_context_with_model: CPU output buffer size = 0.49 MiB
94llama_new_context_with_model: CANN compute buffer size = 1260.50 MiB
95llama_new_context_with_model: CPU compute buffer size = 24.01 MiB
96llama_new_context_with_model: graph nodes = 1030
97llama_new_context_with_model: graph splits = 4
98
99system_info: n_threads = 192 / 192 | AVX = 0 | AVX_VNNI = 0 | AVX2 = 0 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | FMA = 0 | NEON = 1 | SVE = 0 | ARM_FMA = 1 | F16C = 0 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 0 | SSE3 = 0 | SSSE3 = 0 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 1 |
100sampling:
101 repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
102 top_k = 40, tfs_z = 1.000, top_p = 0.950, min_p = 0.050, typical_p = 1.000, temp = 0.800
103 mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000
104sampling order:
105CFG -> Penalties -> top_k -> tfs_z -> typical_p -> top_p -> min_p -> temperature
106generate: n_ctx = 8192, n_batch = 2048, n_predict = -1, n_keep = 1
107
108
109Building a website can be done in 10 simple steps: 1. Define your website's purpose and target audience 2. Choose a domain name and register it with a registrar 3. Select a web hosting service and set up your hosting account 4. Design your website's layout and structure 5. Create content for your website, including text, images, and other media 6. Build a responsive website design that adapts to different devices and screen sizes 7. Choose a Content Management System (CMS) and install it on your website 8. Customize your website's design and layout using a CMS
110
111llama_print_timings: load time = 9074.69 ms
112llama_print_timings: sample time = 31.97 ms / 112 runs ( 0.29 ms per token, 3503.28 tokens per second)
113llama_print_timings: prompt eval time = 238.53 ms / 13 tokens ( 18.35 ms per token, 54.50 tokens per second)
114llama_print_timings: eval time = 13152.29 ms / 111 runs ( 118.49 ms per token, 8.44 tokens per second)
115llama_print_timings: total time = 13623.53 ms / 124 tokens