llama.cpp实现大模型的格式转换、量化、推理、部署
概述
llama.cpp的主要目标是能够在各种硬件上实现LLM推理，只需最少的设置，并提供最先进的性能。提供1.5位、2位、3位、4位、5位、6位和8位整数量化，以加快推理速度并减少内存使用。
GitHub：https://github.com/ggerganov/llama.cpp
克隆和编译
克隆最新版llama.cpp仓库代码
git clone https://github.com/ggerganov/llama.cpp
对llama.cpp项目进行编译，在目录下会生成一系列可执行文件
main：使用模型进行推理
quantize：量化模型
server：提供模型API服务
1.编译构建CPU执行环境，安装简单，适用于没有GPU的操作系统
cd llama.cpp
mkdir
2.编译构建GPU执行环境，确保安装CUDA工具包，适用于有GPU的操作系统
如果CUDA设置正确，那么执行
nvidia-smi、nvcc --version没有错误提示，则表示一切设置正确。
make clean && make LLAMA_CUDA=1
3.如果编译失败或者需要重新编译，可尝试清理并重新编译，直至编译成功
make clean
环境准备
1.下载受支持的模型
要使用llamma.cpp，首先需要准备它支持的模型。在官方文档中给出了说明，这里仅仅截取其中一部分
2.安装依赖
llama.cpp项目下带有requirements.txt 文件，直接安装依赖即可。
pip install -r requirements.txt
模型格式转换
根据模型架构，可以使用
convert.py或convert-hf-to-gguf.py文件。
转换脚本读取模型配置、分词器、张量名称+数据，并将它们转换为GGUF元数据和张量。
GGUF格式
Llama-3相比其前两代显著扩充了词表大小，由32K扩充至128K，并且改为BPE词表。因此需要使用
--vocab-type参数指定分词算法，默认值是spm，如果是bpe，需要显示指定
注意：
官方文档说convert.py不支持LLaMA 3，喊使用convert-hf-to-gguf.py，但它不支持
--vocab-type，且出现异常：error: unrecognized arguments: --vocab-type bpe，因此使用convert.py且没出问题
使用llama.cpp项目中的convert.py脚本转换模型为GGUF格式
root@master:~/work/llama.cpp# python3 ./convert.py /root/work/models/Llama3-Chinese-8B-Instruct/ --outtype f16 --vocab-type bpe --outfile ./models/Llama3-FP16.gguf
INFO:convert:Loading model file /root/work/models/Llama3-Chinese-8B-Instruct/model-00001-of-00004.safetensors
INFO:convert:Loading model file /root/work/models/Llama3-Chinese-8B-Instruct/model-00001-of-00004.safetensors
INFO:convert:Loading model file /root/work/models/Llama3-Chinese-8B-Instruct/model-00002-of-00004.safetensors
INFO:convert:Loading model file /root/work/models/Llama3-Chinese-8B-Instruct/model-00003-of-00004.safetensors
INFO:convert:Loading model file /root/work/models/Llama3-Chinese-8B-Instruct/model-00004-of-00004.safetensors
INFO:convert:model parameters count : 8030261248 (8B)
INFO:convert:params = Params(n_vocab=128256, n_embd=4096, n_layer=32, n_ctx=8192, n_ff=14336, n_head=32, n_head_kv=8, n_experts=None, n_experts_used=None, f_norm_eps=1e-05, rope_scaling_type=None, f_rope_freq_base=500000.0, f_rope_scale=None, n_orig_ctx=None, rope_finetuned=None, ftype=<GGMLFileType.MostlyF16: 1>, path_model=PosixPath('/root/work/models/Llama3-Chinese-8B-Instruct'))
INFO:convert:Loaded vocab file PosixPath('/root/work/models/Llama3-Chinese-8B-Instruct/tokenizer.json'), type 'bpe'
INFO:convert:Vocab info: <BpeVocab with 128000 base tokens and 256 added tokens>
INFO:convert:Special vocab info: <SpecialVocab with 280147 merges, special tokens {'bos': 128000, 'eos': 128001}, add special tokens unset>
INFO:convert:Writing models/Llama3-FP16.gguf, format 1
WARNING:convert:Ignoring added_tokens.json since model matches vocab size without it.
INFO:gguf.gguf_writer:gguf: This GGUF file is for Little Endian only
INFO:gguf.vocab:Adding 280147 merge(s).
INFO:gguf.vocab:Setting special token type bos to 128000
INFO:gguf.vocab:Setting special token type eos to 128001
INFO:gguf.vocab:Setting chat_template to {% set loop_messages = messages %}{% for message in loop_messages %}{% set content = '<|start_header_id|>' + message['role'] + '<|end_header_id|>
'+ message['content'] | trim + '<|eot_id|>' %}{% if loop.index0 == 0 %}{% set content = bos_token + content %}{% endif %}{{ content }}{% endfor %}{{ '<|start_header_id|>assistant<|end_header_id|>
' }}
INFO:convert:[ 1/291] Writing tensor token_embd.weight | size 128256 x 4096 | type F16 | T+ 1
INFO:convert:[ 2/291] Writing tensor blk.0.attn_norm.weight | size 4096 | type F32 | T+ 2
INFO:convert:[ 3/291] Writing tensor blk.0.ffn_down.weight | size 4096 x 14336 | type F16 | T+ 2
INFO:convert:[ 4/291] Writing tensor blk.0.ffn_gate.weight | size 14336 x 4096 | type F16 | T+ 2
INFO:convert:[ 5/291] Writing tensor blk.0.ffn_up.weight | size 14336 x 4096 | type F16 | T+ 2
INFO:convert:[ 6/291] Writing tensor blk.0.ffn_norm.weight | size 4096 | type F32 | T+ 2
INFO:convert:[ 7/291] Writing tensor blk.0.attn_k.weight | size 1024 x 4096 | type F16 | T+ 2
INFO:convert:[ 8/291] Writing tensor blk.0.attn_output.weight | size 4096 x 4096 | type F16 | T+ 2
INFO:convert:[ 9/291] Writing tensor blk.0.attn_q.weight | size 4096 x 4096 | type F16 | T+ 3
INFO:convert:[ 10/291] Writing tensor blk.0.attn_v.weight | size 1024 x 4096 | type F16 | T+ 3
INFO:convert:[ 11/291] Writing tensor blk.1.attn_norm.weight | size 4096 | type F32 | T+ 3
转换为FP16的GGUF格式，模型体积大概15G。
root@master:~/work/llama.cpp# ll models -h
-rw-r--r-- 1 root root 15G May 17 07:47 Llama3-FP16.gguf
bin格式
root@master:~/work/llama.cpp# python3 ./convert.py /root/work/models/Llama3-Chinese-8B-Instruct/ --outtype f16 --vocab-type bpe --outfile ./models/Llama3-FP16.bin
INFO:convert:Loading model file /root/work/models/Llama3-Chinese-8B-Instruct/model-00001-of-00004.safetensors
INFO:convert:Loading model file /root/work/models/Llama3-Chinese-8B-Instruct/model-00001-of-00004.safetensors
INFO:convert:Loading model file /root/work/models/Llama3-Chinese-8B-Instruct/model-00002-of-00004.safetensors
INFO:convert:Loading model file /root/work/models/Llama3-Chinese-8B-Instruct/model-00003-of-00004.safetensors
INFO:convert:Loading model file /root/work/models/Llama3-Chinese-8B-Instruct/model-00004-of-00004.safetensors
INFO:convert:model parameters count : 8030261248 (8B)
INFO:convert:params = Params(n_vocab=128256, n_embd=4096, n_layer=32, n_ctx=8192, n_ff=14336, n_head=32, n_head_kv=8, n_experts=None, n_experts_used=None, f_norm_eps=1e-05, rope_scaling_type=None, f_rope_freq_base=500000.0, f_rope_scale=None, n_orig_ctx=None, rope_finetuned=None, ftype=<GGMLFileType.MostlyF16: 1>, path_model=PosixPath('/root/work/models/Llama3-Chinese-8B-Instruct'))
INFO:convert:Loaded vocab file PosixPath('/root/work/models/Llama3-Chinese-8B-Instruct/tokenizer.json'), type 'bpe'
INFO:convert:Vocab info: <BpeVocab with 128000 base tokens and 256 added tokens>
INFO:convert:Special vocab info: <SpecialVocab with 280147 merges, special tokens {'bos': 128000, 'eos': 128001}, add special tokens unset>
INFO:convert:Writing models/Llama3-FP16.bin, format 1
WARNING:convert:Ignoring added_tokens.json since model matches vocab size without it.
INFO:gguf.gguf_writer:gguf: This GGUF file is for Little Endian only
INFO:gguf.vocab:Adding 280147 merge(s).
INFO:gguf.vocab:Setting special token type bos to 128000
INFO:gguf.vocab:Setting special token type eos to 128001
INFO:gguf.vocab:Setting chat_template to {% set loop_messages = messages %}{% for message in loop_messages %}{% set content = '<|start_header_id|>' + message['role'] + '<|end_header_id|>
'+ message['content'] | trim + '<|eot_id|>' %}{% if loop.index0 == 0 %}{% set content = bos_token + content %}{% endif %}{{ content }}{% endfor %}{{ '<|start_header_id|>assistant<|end_header_id|>
' }}
INFO:convert:[ 1/291] Writing tensor token_embd.weight | size 128256 x 4096 | type F16 | T+ 4
INFO:convert:[ 2/291] Writing tensor blk.0.attn_norm.weight | size 4096 | type F32 | T+ 4
INFO:convert:[ 3/291] Writing tensor blk.0.ffn_down.weight | size 4096 x 14336 | type F16 | T+ 4
INFO:convert:[ 4/291] Writing tensor blk.0.ffn_gate.weight | size 14336 x 4096 | type F16 | T+ 5
INFO:convert:[ 5/291] Writing tensor blk.0.ffn_up.weight | size 14336 x 4096 | type F16 | T+ 5
INFO:convert:[ 6/291] Writing tensor blk.0.ffn_norm.weight | size 4096 | type F32 | T+ 5
INFO:convert:[ 7/291] Writing tensor blk.0.attn_k.weight | size 1024 x 4096 | type F16 | T+ 5
INFO:convert:[ 8/291] Writing tensor blk.0.attn_output.weight | size 4096 x 4096 | type F16 | T+ 5
INFO:convert:[ 9/291] Writing tensor blk.0.attn_q.weight | size 4096 x 4096 | type F16 | T+ 5
INFO:convert:[ 10/291] Writing tensor blk.0.attn_v.weight | size 1024 x 4096 | type F16 | T+ 5
INFO:convert:[ 11/291] Writing tensor blk.1.attn_norm.weight | size 4096 | type F32 | T+ 5
INFO:convert:[ 12/291] Writing tensor blk.1.ffn_down.weight | size 4096 x 14336 | type F16 | T+ 5
INFO:convert:[ 13/291] Writing tensor blk.1.ffn_gate.weight | size 14336 x 4096 | type F16 | T+ 5
root@master:~/work/llama.cpp# ll models -h
-rw-r--r-- 1 root root 15G May 17 07:47 Llama3-FP16.gguf
-rw-r--r-- 1 root root 15G May 17 08:02 Llama3-FP16.bin
模型量化
模型量化使用quantize命令，其具体可用参数与允许量化的类型如下：
root@master:~/work/llama.cpp# ./quantize
usage: ./quantize [--help] [--allow-requantize] [--leave-output-tensor] [--pure] [--imatrix] [--include-weights] [--exclude-weights] [--output-tensor-type] [--token-embedding-type] [--override-kv] model-f32.gguf [model-quant.gguf] type [nthreads]
--allow-requantize: Allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit
--leave-output-tensor: Will leave output.weight un(re)quantized. Increases model size but may also increase quality, especially when requantizing
--pure: Disable k-quant mixtures and quantize all tensors to the same type
--imatrix file_name: use data in file_name as importance matrix for quant optimizations
--include-weights tensor_name: use importance matrix for this/these tensor(s)
--exclude-weights tensor_name: use importance matrix for this/these tensor(s)
--output-tensor-type ggml_type: use this ggml_type for the output.weight tensor
--token-embedding-type ggml_type: use this ggml_type for the token embeddings tensor
--keep-split: will generate quatized model in the same shards as input --override-kv KEY=TYPE:VALUE
Advanced option to override model metadata by key in the quantized model. May be specified multiple times.
Note: --include-weights and --exclude-weights cannot be used together
Allowed quantization types:
2 or Q4_0 : 3.56G, +0.2166 ppl @ LLaMA-v1-7B
3 or Q4_1 : 3.90G, +0.1585 ppl @ LLaMA-v1-7B
8 or Q5_0 : 4.33G, +0.0683 ppl @ LLaMA-v1-7B
9 or Q5_1 : 4.70G, +0.0349 ppl @ LLaMA-v1-7B
19 or IQ2_XXS : 2.06 bpw quantization
20 or IQ2_XS : 2.31 bpw quantization
28 or IQ2_S : 2.5 bpw quantization
29 or IQ2_M : 2.7 bpw quantization
24 or IQ1_S : 1.56 bpw quantization
31 or IQ1_M : 1.75 bpw quantization
10 or Q2_K : 2.63G, +0.6717 ppl @ LLaMA-v1-7B
21 or Q2_K_S : 2.16G, +9.0634 ppl @ LLaMA-v1-7B
23 or IQ3_XXS : 3.06 bpw quantization
26 or IQ3_S : 3.44 bpw quantization
27 or IQ3_M : 3.66 bpw quantization mix
12 or Q3_K : alias for Q3_K_M
22 or IQ3_XS : 3.3 bpw quantization
11 or Q3_K_S : 2.75G, +0.5551 ppl @ LLaMA-v1-7B
12 or Q3_K_M : 3.07G, +0.2496 ppl @ LLaMA-v1-7B
13 or Q3_K_L : 3.35G, +0.1764 ppl @ LLaMA-v1-7B
25 or IQ4_NL : 4.50 bpw non-linear quantization
30 or IQ4_XS : 4.25 bpw non-linear quantization
15 or Q4_K : alias for Q4_K_M
14 or Q4_K_S : 3.59G, +0.0992 ppl @ LLaMA-v1-7B
15 or Q4_K_M : 3.80G, +0.0532 ppl @ LLaMA-v1-7B
17 or Q5_K : alias for Q5_K_M
16 or Q5_K_S : 4.33G, +0.0400 ppl @ LLaMA-v1-7B
17 or Q5_K_M : 4.45G, +0.0122 ppl @ LLaMA-v1-7B
18 or Q6_K : 5.15G, +0.0008 ppl @ LLaMA-v1-7B
7 or Q8_0 : 6.70G, +0.0004 ppl @ LLaMA-v1-7B
1 or F16 : 14.00G, -0.0020 ppl @ Mistral-7B
32 or BF16 : 14.00G, -0.0050 ppl @ Mistral-7B
0 or F32 : 26.00G @ 7B
COPY : only copy tensors, no quantizing
使用quantize量化模型，它提供各种量化位数的模型：Q2、Q3、Q4、Q5、Q6、Q8、F16。
量化模型的命名方法遵循: Q + 量化比特位 + 变种。量化位数越少，对硬件资源的要求越低，但是模型的精度也越低。
模型经过量化之后，可以发现模型的大小从15G降低到8G，但模型精度从16位浮点数降低到8位整数。
root@master:~/work/llama.cpp# ./quantize ./models/Llama3-FP16.gguf ./models/Llama3-q8.gguf q8_0
main: build = 2908 (359cbe3f)
main: built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
main: quantizing '/root/work/models/Llama3-FP16.gguf' to '/root/work/models/Llama3-q8.gguf' as Q8_0
llama_model_loader: loaded meta data with 21 key-value pairs and 291 tensors from /root/work/models/Llama3-FP16.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 = Llama3-Chinese-8B-Instruct
llama_model_loader: - kv 2: llama.vocab_size u32 = 128256
llama_model_loader: - kv 3: llama.context_length u32 = 8192
llama_model_loader: - kv 4: llama.embedding_length u32 = 4096
llama_model_loader: - kv 5: llama.block_count u32 = 32
llama_model_loader: - kv 6: llama.feed_forward_length u32 = 14336
llama_model_loader: - kv 7: llama.rope.dimension_count u32 = 128
llama_model_loader: - kv 8: llama.attention.head_count u32 = 32
llama_model_loader: - kv 9: llama.attention.head_count_kv u32 = 8
llama_model_loader: - kv 10: llama.attention.layer_norm_rms_epsilon f32 = 0.000010
llama_model_loader: - kv 11: llama.rope.freq_base f32 = 500000.000000
llama_model_loader: - kv 12: general.file_type u32 = 1
llama_model_loader: - kv 13: tokenizer.ggml.model str = gpt2
llama_model_loader: - kv 14: tokenizer.ggml.tokens arr[str,128256] = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv 15: tokenizer.ggml.scores arr[f32,128256] = [0.000000, 0.000000, 0.000000, 0.0000...
llama_model_loader: - kv 16: tokenizer.ggml.token_type arr[i32,128256] = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv 17: tokenizer.ggml.merges arr[str,280147] = ["Ġ Ġ", "Ġ ĠĠĠ", "ĠĠ ĠĠ", "...
llama_model_loader: - kv 18: tokenizer.ggml.bos_token_id u32 = 128000
llama_model_loader: - kv 19: tokenizer.ggml.eos_token_id u32 = 128001
llama_model_loader: - kv 20: tokenizer.chat_template str = {% set loop_messages = messages %}{% ...
llama_model_loader: - type f32: 65 tensors
llama_model_loader: - type f16: 226 tensors
[ 1/ 291] token_embd.weight - [ 4096, 128256, 1, 1], type = f16, converting to q8_0 .. size = 1002.00 MiB -> 532.31 MiB
[ 2/ 291] blk.0.attn_norm.weight - [ 4096, 1, 1, 1], type = f32, size = 0.016 MB
[ 3/ 291] blk.0.ffn_down.weight - [14336, 4096, 1, 1], type = f16, converting to q8_0 .. size = 112.00 MiB -> 59.50 MiB
[ 4/ 291] blk.0.ffn_gate.weight - [ 4096, 14336, 1, 1], type = f16, converting to q8_0 .. size = 112.00 MiB -> 59.50 MiB
[ 5/ 291] blk.0.ffn_up.weight - [ 4096, 14336, 1, 1], type = f16, converting to q8_0 .. size = 112.00 MiB -> 59.50 MiB
[ 6/ 291] blk.0.ffn_norm.weight - [ 4096, 1, 1, 1], type = f32, size = 0.016 MB
[ 7/ 291] blk.0.attn_k.weight - [ 4096, 1024, 1, 1], type = f16, converting to q8_0 .. size = 8.00 MiB -> 4.25 MiB
[ 8/ 291] blk.0.attn_output.weight - [ 4096, 4096, 1, 1], type = f16, converting to q8_0 .. size = 32.00 MiB -> 17.00 MiB
[ 9/ 291] blk.0.attn_q.weight - [ 4096, 4096, 1, 1], type = f16, converting to q8_0 .. size = 32.00 MiB -> 17.00 MiB
[ 10/ 291] blk.0.attn_v.weight - [ 4096, 1024, 1, 1], type = f16, converting to q8_0 .. size = 8.00 MiB -> 4.25 MiB
[ 11/ 291] blk.1.attn_norm.weight - [ 4096, 1, 1, 1], type = f32, size = 0.016 MB
[ 12/ 291] blk.1.ffn_down.weight - [14336, 4096, 1, 1], type = f16, converting to q8_0 .. size = 112.00 MiB -> 59.50 MiB
[ 13/ 291] blk.1.ffn_gate.weight - [ 4096, 14336, 1, 1], type = f16, converting to q8_0 .. size = 112.00 MiB -> 59.50 MiB
[ 14/ 291] blk.1.ffn_up.weight - [ 4096, 14336, 1, 1], type = f16, converting to q8_0 .. size = 112.00 MiB -> 59.50 MiB
[ 15/ 291] blk.1.ffn_norm.weight - [ 4096, 1, 1, 1], type = f32, size = 0.016 MB
[ 16/ 291] blk.1.attn_k.weight - [ 4096, 1024, 1, 1], type = f16, converting to q8_0 .. size = 8.00 MiB -> 4.25 MiB
[ 17/ 291] blk.1.attn_output.weight - [ 4096, 4096, 1, 1], type = f16, converting to q8_0 .. size = 32.00 MiB -> 17.00 MiB
[ 18/ 291] blk.1.attn_q.weight - [ 4096, 4096, 1, 1], type = f16, converting to q8_0 .. size = 32.00 MiB -> 17.00 MiB
[ 19/ 291] blk.1.attn_v.weight - [ 4096, 1024, 1, 1], type = f16, converting to q8_0 .. size = 8.00 MiB -> 4.25 MiB
root@master:~/work/llama.cpp# ll -h models/
-rw-r--r-- 1 root root 8.0G May 17 07:54 Llama3-q8.gguf
模型加载与推理
模型加载与推理使用main命令，其支持如下可用参数：
root@master:~/work/llama.cpp# ./main -h
usage: ./main [options]
options:
-h, --help show this help message and exit
--version show version and build info
-i, --interactive run in interactive mode
--interactive-specials allow special tokens in user text, in interactive mode
--interactive-first run in interactive mode and wait for input right away
-cnv, --conversation run in conversation mode (does not print special tokens and suffix/prefix)
-ins, --instruct run in instruction mode (use with Alpaca models)
-cml, --chatml run in chatml mode (use with ChatML-compatible models)
--multiline-input allows you to write or paste multiple lines without ending each in '\'
-r PROMPT, --reverse-prompt PROMPT
halt generation at PROMPT, return control in interactive mode
(can be specified more than once for multiple prompts).
--color colorise output to distinguish prompt and user input from generations
-s SEED, --seed SEED RNG seed (default: -1, use random seed for < 0)
-t N, --threads N number of threads to use during generation (default: 30)
-tb N, --threads-batch N
number of threads to use during batch and prompt processing (default: same as --threads)
-td N, --threads-draft N number of threads to use during generation (default: same as --threads)
-tbd N, --threads-batch-draft N
number of threads to use during batch and prompt processing (default: same as --threads-draft)
-p PROMPT, --prompt PROMPT
prompt to start generation with (default: empty)
可以加载预训练模型或者经过量化之后的模型，这里选择加载量化后的模型进行推理。
在llama.cpp项目的根目录，执行如下命令，加载模型进行推理。
root@master:~/work/llama.cpp# ./main -m models/Llama3-q8.gguf --color -f prompts/alpaca.txt -ins -c 2048 --temp 0.2 -n 256 --repeat_penalty 1.1
Log start
main: build = 2908 (359cbe3f)
main: built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
main: seed = 1715935175
llama_model_loader: loaded meta data with 22 key-value pairs and 291 tensors from models/Llama3-q8.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 = Llama3-Chinese-8B-Instruct
llama_model_loader: - kv 2: llama.vocab_size u32 = 128256
llama_model_loader: - kv 3: llama.context_length u32 = 8192
llama_model_loader: - kv 4: llama.embedding_length u32 = 4096
llama_model_loader: - kv 5: llama.block_count u32 = 32
llama_model_loader: - kv 6: llama.feed_forward_length u32 = 14336
llama_model_loader: - kv 7: llama.rope.dimension_count u32 = 128
== Running in interactive mode. ==
- Press Ctrl+C to interject at any time.
- Press Return to return control to LLaMa.
- To return control without starting a new line, end your input with '/'.
- If you want to submit another line, end your input with '\'.
<|begin_of_text|>Below is an instruction that describes a task. Write a response that appropriately completes the request.
> hi
Hello! How can I help you today?<|eot_id|>
>
在提示符>之后输入prompt，使用ctrl+c中断输出，多行信息以\作为行尾。执行./main -h命令查看帮助和参数说明，以下是一些常用的参数： `
| 命令 | 描述 |
|---|---|
| -m | 指定 LLaMA 模型文件的路径 |
| -mu | 指定远程 http url 来下载文件 |
| -i | 以交互模式运行程序，允许直接提供输入并接收实时响应。 |
| -ins | 以指令模式运行程序，这在处理羊驼模型时特别有用。 |
| -f | 指定prompt模板，alpaca模型请加载prompts/alpaca.txt |
| -n | 控制回复生成的最大长度（默认：128） |
| -c | 设置提示上下文的大小，值越大越能参考更长的对话历史（默认：512） |
| -b | 控制batch size（默认：8），可适当增加 |
| -t | 控制线程数量（默认：4），可适当增加 |
--repeat_penalty | 控制生成回复中对重复文本的惩罚力度 |
--temp | 温度系数，值越低回复的随机性越小，反之越大 |
--top_p, top_k | 控制解码采样的相关参数 |
--color | 区分用户输入和生成的文本 |
模型API服务
llama.cpp提供了完全与OpenAI API兼容的API接口，使用经过编译生成的server可执行文件启动API服务。
root@master:~/work/llama.cpp# ./server -m models/Llama3-q8.gguf --host 0.0.0.0 --port 8000
{"tid":"140018656950080","timestamp":1715936504,"level":"INFO","function":"main","line":2942,"msg":"build info","build":2908,"commit":"359cbe3f"}
{"tid":"140018656950080","timestamp":1715936504,"level":"INFO","function":"main","line":2947,"msg":"system info","n_threads":30,"n_threads_batch":-1,"total_threads":30,"system_info":"AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 1 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | FMA = 1 | NEON = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 0 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 1 | "}
llama_model_loader: loaded meta data with 22 key-value pairs and 291 tensors from models/Llama3-q8.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 = Llama3-Chinese-8B-Instruct
llama_model_loader: - kv 2: llama.vocab_size u32 = 128256
llama_model_loader: - kv 3: llama.context_length u32 = 8192
llama_model_loader: - kv 4: llama.embedding_length u32 = 4096
llama_model_loader: - kv 5: llama.block_count u32 = 32
llama_model_loader: - kv 6: llama.feed_forward_length u32 = 14336
启动API服务后，可以使用curl命令进行测试
curl --request POST \
--url http://localhost:8000/completion \
--header "Content-Type: application/json" \
--data '{"prompt": "Hi"}'
模型API服务(第三方)
在llamm.cpp项目中有提到各种语言编写的第三方工具包，可以使用这些工具包提供API服务，这里以Python为例，使用llama-cpp-python提供API服务。
安装依赖
pip install llama-cpp-python
pip install llama-cpp-python -i https://mirrors.aliyun.com/pypi/simple/
注意：可能还需要安装以下缺失依赖，可根据启动时的异常提示分别安装。
pip install sse_starlette starlette_context pydantic_settings
启动API服务，默认运行在http://localhost:8000
python -m llama_cpp.server --model models/Llama3-q8.gguf
安装openai依赖
pip install openai
使用openai调用API服务
import os
from openai import OpenAI # 导入OpenAI库
# 设置OpenAI的BASE_URL
os.environ["OPENAI_BASE_URL"] = "http://localhost:8000/v1"
client = OpenAI() # 创建OpenAI客户端对象
# 调用模型
completion = client.chat.completions.create(
model="llama3", # 任意填
messages=[
{"role": "system", "content": "你是一个乐于助人的助手。"},
{"role": "user", "content": "你好!"}
]
)
# 输出模型回复
print(completion.choices[0].message)
GPU推理
如果编译构建了GPU执行环境，可以使用
-ngl N或--n-gpu-layers N参数，指定offload层数，让模型在GPU上运行推理
例如：
-ngl 40表示offload 40层模型参数到GPU
未使用-ngl N或 --n-gpu-layers N参数，程序默认在CPU上运行
root@master:~/work/llama.cpp# ./server -m models/Llama3-FP16.gguf --host 0.0.0.0 --port 8000
可从以下关键启动日志看出，模型并没有在GPU上执行
ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no
ggml_cuda_init: CUDA_USE_TENSOR_CORES: yes
ggml_cuda_init: found 1 CUDA devices:
Device 0: Tesla V100S-PCIE-32GB, compute capability 7.0, VMM: yes
llm_load_tensors: ggml ctx size = 0.15 MiB
llm_load_tensors: offloading 0 repeating layers to GPU
llm_load_tensors: offloaded 0/33 layers to GPU
llm_load_tensors: CPU buffer size = 8137.64 MiB
.........................................................................................
llama_new_context_with_model: n_ctx = 2048
llama_new_context_with_model: n_batch = 2048
llama_new_context_with_model: n_ubatch = 512
使用-ngl N或 --n-gpu-layers N参数，程序默认在GPU上运行
root@master:~/work/llama.cpp# ./server -m models/Llama3-FP16.gguf --host 0.0.0.0 --port 8000 --n-gpu-layers 1000
可从以下关键启动日志看出，模型在GPU上执行
ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no
ggml_cuda_init: CUDA_USE_TENSOR_CORES: yes
ggml_cuda_init: found 1 CUDA devices:
Device 0: Tesla V100S-PCIE-32GB, compute capability 7.0, VMM: yes
llm_load_tensors: ggml ctx size = 0.30 MiB
llm_load_tensors: offloading 32 repeating layers to GPU
llm_load_tensors: offloading non-repeating layers to GPU
llm_load_tensors: offloaded 33/33 layers to GPU
llm_load_tensors: CPU buffer size = 1002.00 MiB
llm_load_tensors: CUDA0 buffer size = 14315.02 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: flash_attn = 0
执行nvidia-smi命令，可以���一步验证模型已在GPU上运行。 
如何学习AI大模型？
我在一线互联网企业工作十余年里，指导过不少同行后辈。帮助很多人得到了学习和成长。
我意识到有很多经验和知识值得分享给大家，也可以通过我们的能力和经验解答大家在人工智能学习中的很多困惑，所以在工作繁忙的情况下还是坚持各种整理和分享。但苦于知识传播途径有限，很多互联网行业朋友无法获得正确的资料得到学习提升，故此将并将重要的AI大模型资料包括AI大模型入门学习思维导图、精品AI大模型学习书籍手册、视频教程、实战学习等录播视频免费分享出来。
第一阶段： 从大模型系统设计入手，讲解大模型的主要方法；
第二阶段： 在通过大模型提示词工程从Prompts角度入手更好发挥模型的作用；
第三阶段： 大模型平台应用开发借助阿里云PAI平台构建电商领域虚拟试衣系统；
第四阶段： 大模型知识库应用开发以LangChain框架为例，构建物流行业咨询智能问答系统；
第五阶段： 大模型微调开发借助以大健康、新零售、新媒体领域构建适合当前领域大模型；
第六阶段： 以SD多模态大模型为主，搭建了文生图小程序案例；
第七阶段： 以大模型平台应用与开发为主，通过星火大模型，文心大模型等成熟大模型构建大模型行业应用。
👉学会后的收获：👈
• 基于大模型全栈工程实现（前端、后端、产品经理、设计、数据分析等），通过这门课可获得不同能力；
• 能够利用大模型解决相关实际项目需求： 大数据时代，越来越多的企业和机构需要处理海量数据，利用大模型技术可以更好地处理这些数据，提高数据分析和决策的准确性。因此，掌握大模型应用开发技能，可以让程序员更好地应对实际项目需求；
• 基于大模型和企业数据AI应用开发，实现大模型理论、掌握GPU算力、硬件、LangChain开发框架和项目实战技能， 学会Fine-tuning垂直训练大模型（数据准备、数据蒸馏、大模型部署）一站式掌握；
• 能够完成时下热门大模型垂直领域模型训练能力，提高程序员的编码能力： 大模型应用开发需要掌握机器学习算法、深度学习框架等技术，这些技术的掌握可以提高程序员的编码能力和分析能力，让程序员更加熟练地编写高质量的代码。
1.AI大模型学习路线图
2.100套AI大模型商业化落地方案
3.100集大模型视频教程
4.200本大模型PDF书籍
5.LLM面试题合集
6.AI产品经理资源合集
👉获取方式：
😝有需要的小伙伴，可以保存图片到wx扫描二v码免费领取【保证100%免费】🆓
扫一扫
2万+
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
