Chinese-LLaMA-Alpaca-2模型量化部署测试

简介

Chinese-LLaMA-Alpaca-2基于Meta发布的可商用大模型Llama-2开发, 是中文LLaMA&Alpaca大模型的第二期项目.

量化

模型的下载还是应用脚本

bash hfd.sh hfl/chinese-alpaca-2-13b --tool aria2c -x 8

应用llama.cpp进行量化, 主要参考该教程.
其中比较折腾的是与BLAS一起编译.

OpenBLAS

这个真是一言难尽, 非常折腾也没起作用(issue1 & issue2). 而且提升很小, 后续再尝试能不能成功.

cuBLAS

这个提升较为明显, 在有Nvidia GPU的情况下, 需要折腾应该就只有非root用户手动安装一下CUDA toolkit, 然后在CMakeLists.txt中指定一下路径即可.
手动安装CUDA toolkit和cuDnn后, 在CMakeLists.txt中加入:

# ${cuda path}示例: /home/orange/software/cuda-118
set(CUDA_TOOLKIT_ROOT_DIR ${cuda path})

进行编译即可

mkdir build
cd build
cmake .. -DLLAMA_CUBLAS=ON
cmake --build . --config Release

量化

编译完成llama.cpp后, 进行量化

python convert.py zh-models/chinese-alpaca-2-7b/
./build/bin/quantize ./zh-models/chinese-alpaca-2-7b/ggml-model-f16.gguf ./zh-models/chinese-alpaca-2-7b/ggml-model-q8_0.gguf q8_0

部署测试

直接使用./build/bin/main -m ./zh-models/chinese-alpaca-2-7b/ggml-model-q8_0.gguf不能进行对话, 加入参数-i表示交互模式, 也可以使用教程中的脚本形式.
按照tutorial, 新建chat.sh文件并填入以下内容

#!/bin/bash# temporary script to chat with Chinese Alpaca-2 model
# usage: ./chat.sh alpaca2-ggml-model-path your-first-instructionSYSTEM='You are a helpful assistant. 你是一个乐于助人的助手。'
FIRST_INSTRUCTION=$2./build/bin/main -m $1 \
--color -i -c 4096 -t 8 --temp 0.5 --top_k 40 --top_p 0.9 --repeat_penalty 1.1 \
--in-prefix-bos --in-prefix ' [INST] ' --in-suffix ' [/INST]' -p \
"[INST] <<SYS>>
$SYSTEM
<</SYS>>$FIRST_INSTRUCTION [/INST]"

运行

bash chat.sh ./zh-models/chinese-alpaca-2-7b/ggml-model-q8_0.gguf '请列举5条文明乘车的建议'

成功实现对话, 部署测试成功.

测试

下载并解压测试数据

chinese-alpaca-2-1.3b

测试命令:

./build/bin/perplexity -m ./zh-models/chinese-alpaca-2-1.3b/ggml-model-q8_0.gguf -f ./wikitext-2-raw/wiki.test.raw -ngl 20

由于使用cmake编译, 可执行文件位于build/bin下, 注意执行文件和模型, 数据的路径替换即可.
测试数据如下:

main: build = 2509 (50ccaf5e)
main: built with cc (Ubuntu 9.4.0-1ubuntu1~20.04.2) 9.4.0 for x86_64-linux-gnu
main: seed  = 1711210157
llama_model_loader: loaded meta data with 23 key-value pairs and 39 tensors from ./zh-models/chinese-alpaca-2-1.3b/ggml-model-q8_0.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              = 55296
llama_model_loader: - kv   3:                       llama.context_length u32              = 4096
llama_model_loader: - kv   4:                     llama.embedding_length u32              = 4096
llama_model_loader: - kv   5:                          llama.block_count u32              = 4
llama_model_loader: - kv   6:                  llama.feed_forward_length u32              = 11008
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              = 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              = 7
llama_model_loader: - kv  13:                       tokenizer.ggml.model str              = llama
llama_model_loader: - kv  14:                      tokenizer.ggml.tokens arr[str,55296]   = ["<unk>", "<s>", "</s>", "<0x00>", "<...
llama_model_loader: - kv  15:                      tokenizer.ggml.scores arr[f32,55296]   = [0.000000, 0.000000, 0.000000, 0.0000...
llama_model_loader: - kv  16:                  tokenizer.ggml.token_type arr[i32,55296]   = [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              = 2
llama_model_loader: - kv  19:            tokenizer.ggml.padding_token_id u32              = 0
llama_model_loader: - kv  20:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  21:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  22:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:    9 tensors
llama_model_loader: - type q8_0:   30 tensors
llm_load_vocab: mismatch in special tokens definition ( 889/55296 vs 259/55296 ).
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          = 55296
llm_load_print_meta: n_merges         = 0
llm_load_print_meta: n_ctx_train      = 4096
llm_load_print_meta: n_embd           = 4096
llm_load_print_meta: n_head           = 32
llm_load_print_meta: n_head_kv        = 32
llm_load_print_meta: n_layer          = 4
llm_load_print_meta: n_rot            = 128
llm_load_print_meta: n_embd_head_k    = 128
llm_load_print_meta: n_embd_head_v    = 128
llm_load_print_meta: n_gqa            = 1
llm_load_print_meta: n_embd_k_gqa     = 4096
llm_load_print_meta: n_embd_v_gqa     = 4096
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             = 11008
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       = ?B
llm_load_print_meta: model ftype      = Q8_0
llm_load_print_meta: model params     = 1.26 B
llm_load_print_meta: model size       = 1.25 GiB (8.50 BPW)
llm_load_print_meta: general.name     = LLaMA v2
llm_load_print_meta: BOS token        = 1 '<s>'
llm_load_print_meta: EOS token        = 2 '</s>'
llm_load_print_meta: UNK token        = 0 '<unk>'
llm_load_print_meta: PAD token        = 0 '<unk>'
llm_load_print_meta: LF token         = 13 '<0x0A>'
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:   no
ggml_cuda_init: CUDA_USE_TENSOR_CORES: yes
ggml_cuda_init: found 2 CUDA devices:Device 0: NVIDIA A100-PCIE-40GB, compute capability 8.0, VMM: yesDevice 1: NVIDIA A100-PCIE-40GB, compute capability 8.0, VMM: yes
llm_load_tensors: ggml ctx size =    0.05 MiB
llm_load_tensors: offloading 4 repeating layers to GPU
llm_load_tensors: offloading non-repeating layers to GPU
llm_load_tensors: offloaded 5/5 layers to GPU
llm_load_tensors:        CPU buffer size =   229.50 MiB
llm_load_tensors:      CUDA0 buffer size =   615.28 MiB
llm_load_tensors:      CUDA1 buffer size =   434.61 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
llama_new_context_with_model: freq_base  = 10000.0
llama_new_context_with_model: freq_scale = 1
llama_kv_cache_init:      CUDA0 KV buffer size =    96.00 MiB
llama_kv_cache_init:      CUDA1 KV buffer size =    32.00 MiB
llama_new_context_with_model: KV self size  =  128.00 MiB, K (f16):   64.00 MiB, V (f16):   64.00 MiB
llama_new_context_with_model:  CUDA_Host  output buffer size =   432.00 MiB
llama_new_context_with_model: pipeline parallelism enabled (n_copies=4)
llama_new_context_with_model:      CUDA0 compute buffer size =   208.01 MiB
llama_new_context_with_model:      CUDA1 compute buffer size =   200.01 MiB
llama_new_context_with_model:  CUDA_Host compute buffer size =    24.02 MiB
llama_new_context_with_model: graph nodes  = 136
llama_new_context_with_model: graph splits = 3system_info: n_threads = 76 / 152 | AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | FMA = 1 | NEON = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 |
perplexity: tokenizing the input ..
perplexity: tokenization took 1187.98 ms
perplexity: calculating perplexity over 655 chunks, n_ctx=512, batch_size=2048, n_seq=4
perplexity: 0.06 seconds per pass - ETA 0.17 minutes
[1]35.2055,[2]3151.7331,[3]9745.8526,[4]3056.9236
......
[653]1226.9638,[654]1219.7704,[655]1213.9217,
Final estimate: PPL = 1213.9217 +/- 16.09822llama_print_timings:        load time =    2998.83 ms
llama_print_timings:      sample time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
llama_print_timings: prompt eval time =    8371.13 ms / 335360 tokens (    0.02 ms per token, 40061.49 tokens per second)
llama_print_timings:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
llama_print_timings:       total time =   17937.28 ms / 335361 tokens

chinese-alpaca-2-13b

测试命令:

./build/bin/perplexity -m ./zh-models/chinese-alpaca-2-13b/ggml-model-q8_0.gguf -f ./wikitext-2-raw/wiki.test.raw -ngl 10

测试数据如下:

main: build = 2509 (50ccaf5e)
main: built with cc (Ubuntu 9.4.0-1ubuntu1~20.04.2) 9.4.0 for x86_64-linux-gnu
main: seed  = 1711210012
llama_model_loader: loaded meta data with 21 key-value pairs and 363 tensors from ./zh-models/chinese-alpaca-2-13b/ggml-model-q8_0.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              = 55296
llama_model_loader: - kv   3:                       llama.context_length u32              = 4096
llama_model_loader: - kv   4:                     llama.embedding_length u32              = 5120
llama_model_loader: - kv   5:                          llama.block_count u32              = 40
llama_model_loader: - kv   6:                  llama.feed_forward_length u32              = 13824
llama_model_loader: - kv   7:                 llama.rope.dimension_count u32              = 128
llama_model_loader: - kv   8:                 llama.attention.head_count u32              = 40
llama_model_loader: - kv   9:              llama.attention.head_count_kv u32              = 40
llama_model_loader: - kv  10:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  11:                          general.file_type u32              = 7
llama_model_loader: - kv  12:                       tokenizer.ggml.model str              = llama
llama_model_loader: - kv  13:                      tokenizer.ggml.tokens arr[str,55296]   = ["<unk>", "<s>", "</s>", "<0x00>", "<...
llama_model_loader: - kv  14:                      tokenizer.ggml.scores arr[f32,55296]   = [0.000000, 0.000000, 0.000000, 0.0000...
llama_model_loader: - kv  15:                  tokenizer.ggml.token_type arr[i32,55296]   = [2, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 6, ...
llama_model_loader: - kv  16:                tokenizer.ggml.bos_token_id u32              = 1
llama_model_loader: - kv  17:                tokenizer.ggml.eos_token_id u32              = 2
llama_model_loader: - kv  18:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  19:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  20:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:   81 tensors
llama_model_loader: - type q8_0:  282 tensors
llm_load_vocab: mismatch in special tokens definition ( 889/55296 vs 259/55296 ).
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          = 55296
llm_load_print_meta: n_merges         = 0
llm_load_print_meta: n_ctx_train      = 4096
llm_load_print_meta: n_embd           = 5120
llm_load_print_meta: n_head           = 40
llm_load_print_meta: n_head_kv        = 40
llm_load_print_meta: n_layer          = 40
llm_load_print_meta: n_rot            = 128
llm_load_print_meta: n_embd_head_k    = 128
llm_load_print_meta: n_embd_head_v    = 128
llm_load_print_meta: n_gqa            = 1
llm_load_print_meta: n_embd_k_gqa     = 5120
llm_load_print_meta: n_embd_v_gqa     = 5120
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             = 13824
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       = 13B
llm_load_print_meta: model ftype      = Q8_0
llm_load_print_meta: model params     = 13.25 B
llm_load_print_meta: model size       = 13.12 GiB (8.50 BPW)
llm_load_print_meta: general.name     = LLaMA v2
llm_load_print_meta: BOS token        = 1 '<s>'
llm_load_print_meta: EOS token        = 2 '</s>'
llm_load_print_meta: UNK token        = 0 '<unk>'
llm_load_print_meta: LF token         = 13 '<0x0A>'
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:   no
ggml_cuda_init: CUDA_USE_TENSOR_CORES: yes
ggml_cuda_init: found 2 CUDA devices:Device 0: NVIDIA A100-PCIE-40GB, compute capability 8.0, VMM: yesDevice 1: NVIDIA A100-PCIE-40GB, compute capability 8.0, VMM: yes
llm_load_tensors: ggml ctx size =    0.42 MiB
llm_load_tensors: offloading 10 repeating layers to GPU
llm_load_tensors: offloaded 10/41 layers to GPU
llm_load_tensors:        CPU buffer size = 13431.58 MiB
llm_load_tensors:      CUDA0 buffer size =  1607.23 MiB
llm_load_tensors:      CUDA1 buffer size =  1607.23 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
llama_new_context_with_model: freq_base  = 10000.0
llama_new_context_with_model: freq_scale = 1
llama_kv_cache_init:  CUDA_Host KV buffer size =  1200.00 MiB
llama_kv_cache_init:      CUDA0 KV buffer size =   200.00 MiB
llama_kv_cache_init:      CUDA1 KV buffer size =   200.00 MiB
llama_new_context_with_model: KV self size  = 1600.00 MiB, K (f16):  800.00 MiB, V (f16):  800.00 MiB
llama_new_context_with_model:  CUDA_Host  output buffer size =   432.00 MiB
llama_new_context_with_model:      CUDA0 compute buffer size =   404.88 MiB
llama_new_context_with_model:      CUDA1 compute buffer size =   204.00 MiB
llama_new_context_with_model:  CUDA_Host compute buffer size =    24.00 MiB
llama_new_context_with_model: graph nodes  = 1324
llama_new_context_with_model: graph splits = 335system_info: n_threads = 76 / 152 | AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | FMA = 1 | NEON = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 |
perplexity: tokenizing the input ..
perplexity: tokenization took 728.604 ms
perplexity: calculating perplexity over 655 chunks, n_ctx=512, batch_size=2048, n_seq=4
perplexity: 7.36 seconds per pass - ETA 20.08 minutes
[1]4.8998,[2]5.3381,[3]6.0623,
......
[654]6.3736,[655]6.3713,
Final estimate: PPL = 6.3713 +/- 0.03705llama_print_timings:        load time =   17705.89 ms
llama_print_timings:      sample time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
llama_print_timings: prompt eval time = 1130068.93 ms / 335360 tokens (    3.37 ms per token,   296.76 tokens per second)
llama_print_timings:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
llama_print_timings:       total time = 1137969.38 ms / 335361 tokens