(章节 3.1) 本地运行 AI 有多慢 ? 大模型推理测速 (llama.cpp, Intel GPU A770)

由于本文太长, 分开发布, 方便阅读.


3.1 CPU (i5-6200U, 2C/4T/2.8GHz) x86_64 AVX2

在 4 号 PC (物理机) 上运行. 版本:

sh 复制代码
> ./llama-b3617-bin-ubuntu-x64/build/bin/llama-cli --version
version: 3617 (a07c32ea)
built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu

运行模型 llama2-7B.q4, 生成长度 100:

sh 复制代码
> ./llama-b3617-bin-ubuntu-x64/build/bin/llama-cli -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -n 100
Log start
main: build = 3617 (a07c32ea)
main: built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
main: seed  = 1724500181
llama_model_loader: loaded meta data with 19 key-value pairs and 291 tensors from llama-2-7b.Q4_K_M.gguf (version GGUF V2)
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.context_length u32              = 4096
llama_model_loader: - kv   3:                     llama.embedding_length u32              = 4096
llama_model_loader: - kv   4:                          llama.block_count u32              = 32
llama_model_loader: - kv   5:                  llama.feed_forward_length u32              = 11008
llama_model_loader: - kv   6:                 llama.rope.dimension_count u32              = 128
llama_model_loader: - kv   7:                 llama.attention.head_count u32              = 32
llama_model_loader: - kv   8:              llama.attention.head_count_kv u32              = 32
llama_model_loader: - kv   9:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  10:                          general.file_type u32              = 15
llama_model_loader: - kv  11:                       tokenizer.ggml.model str              = llama
llama_model_loader: - kv  12:                      tokenizer.ggml.tokens arr[str,32000]   = ["<unk>", "<s>", "</s>", "<0x00>", "<...
llama_model_loader: - kv  13:                      tokenizer.ggml.scores arr[f32,32000]   = [0.000000, 0.000000, 0.000000, 0.0000...
llama_model_loader: - kv  14:                  tokenizer.ggml.token_type arr[i32,32000]   = [2, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 6, ...
llama_model_loader: - kv  15:                tokenizer.ggml.bos_token_id u32              = 1
llama_model_loader: - kv  16:                tokenizer.ggml.eos_token_id u32              = 2
llama_model_loader: - kv  17:            tokenizer.ggml.unknown_token_id u32              = 0
llama_model_loader: - kv  18:               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 cache size = 3
llm_load_vocab: token to piece cache size = 0.1684 MB
llm_load_print_meta: format           = GGUF V2
llm_load_print_meta: arch             = llama
llm_load_print_meta: vocab type       = SPM
llm_load_print_meta: n_vocab          = 32000
llm_load_print_meta: n_merges         = 0
llm_load_print_meta: vocab_only       = 0
llm_load_print_meta: n_ctx_train      = 4096
llm_load_print_meta: n_embd           = 4096
llm_load_print_meta: n_layer          = 32
llm_load_print_meta: n_head           = 32
llm_load_print_meta: n_head_kv        = 32
llm_load_print_meta: n_rot            = 128
llm_load_print_meta: n_swa            = 0
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_ctx_orig_yarn  = 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: ssm_dt_b_c_rms   = 0
llm_load_print_meta: model type       = 7B
llm_load_print_meta: model ftype      = Q4_K - Medium
llm_load_print_meta: model params     = 6.74 B
llm_load_print_meta: model size       = 3.80 GiB (4.84 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>'
llm_load_print_meta: max token length = 48
llm_load_tensors: ggml ctx size =    0.14 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 =  3891.24 MiB
..................................................................................................
llama_new_context_with_model: n_ctx      = 4096
llama_new_context_with_model: n_batch    = 2048
llama_new_context_with_model: n_ubatch   = 512
llama_new_context_with_model: flash_attn = 0
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 =  2048.00 MiB
llama_new_context_with_model: KV self size  = 2048.00 MiB, K (f16): 1024.00 MiB, V (f16): 1024.00 MiB
llama_new_context_with_model:        CPU  output buffer size =     0.12 MiB
llama_new_context_with_model:        CPU compute buffer size =   296.01 MiB
llama_new_context_with_model: graph nodes  = 1030
llama_new_context_with_model: graph splits = 1

system_info: n_threads = 2 / 4 | AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | FMA = 1 | NEON = 0 | SVE = 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 | 
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 = 4096, n_batch = 2048, n_predict = 100, n_keep = 1


 hello, this is a very very long story. nobody wants to read this much. just tell me what happened.

(此处省略一部分)

llama_print_timings:        load time =    2666.87 ms
llama_print_timings:      sample time =       5.38 ms /   100 runs   (    0.05 ms per token, 18580.45 tokens per second)
llama_print_timings: prompt eval time =    1898.40 ms /    10 tokens (  189.84 ms per token,     5.27 tokens per second)
llama_print_timings:        eval time =   28113.06 ms /    99 runs   (  283.97 ms per token,     3.52 tokens per second)
llama_print_timings:       total time =   30034.85 ms /   109 tokens
Log end

运行模型 llama2-7B.q4, 生成长度 200:

sh 复制代码
> ./llama-b3617-bin-ubuntu-x64/build/bin/llama-cli -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -n 200

(此处省略一部分)

llama_print_timings:        load time =    2703.62 ms
llama_print_timings:      sample time =      12.85 ms /   200 runs   (    0.06 ms per token, 15560.57 tokens per second)
llama_print_timings: prompt eval time =    1873.80 ms /    10 tokens (  187.38 ms per token,     5.34 tokens per second)
llama_print_timings:        eval time =   59352.84 ms /   199 runs   (  298.26 ms per token,     3.35 tokens per second)
llama_print_timings:       total time =   61281.14 ms /   209 tokens

运行模型 llama2-7B.q4, 生成长度 500:

sh 复制代码
> ./llama-b3617-bin-ubuntu-x64/build/bin/llama-cli -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -n 500

(此处省略一部分)

llama_print_timings:        load time =    2706.04 ms
llama_print_timings:      sample time =      33.77 ms /   500 runs   (    0.07 ms per token, 14808.23 tokens per second)
llama_print_timings: prompt eval time =    1866.60 ms /    10 tokens (  186.66 ms per token,     5.36 tokens per second)
llama_print_timings:        eval time =  154145.54 ms /   499 runs   (  308.91 ms per token,     3.24 tokens per second)
llama_print_timings:       total time =  156146.19 ms /   509 tokens

运行模型 llama2-7B.q4, 生成长度 1000:

sh 复制代码
> ./llama-b3617-bin-ubuntu-x64/build/bin/llama-cli -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -n 1000

(此处省略一部分)

llama_print_timings:        load time =    2912.39 ms
llama_print_timings:      sample time =      60.76 ms /  1000 runs   (    0.06 ms per token, 16457.65 tokens per second)
llama_print_timings: prompt eval time =    1870.87 ms /    10 tokens (  187.09 ms per token,     5.35 tokens per second)
llama_print_timings:        eval time =  335019.17 ms /   999 runs   (  335.35 ms per token,     2.98 tokens per second)
llama_print_timings:       total time =  337155.40 ms /  1009 tokens

运行模型 qwen2-7B.q8, 生成长度 100:

sh 复制代码
> ./llama-b3617-bin-ubuntu-x64/build/bin/llama-cli -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -n 100
Log start
main: build = 3617 (a07c32ea)
main: built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
main: seed  = 1724501237
llama_model_loader: loaded meta data with 26 key-value pairs and 339 tensors from qwen2-7b-instruct-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              = qwen2
llama_model_loader: - kv   1:                               general.name str              = qwen2-7b-instruct
llama_model_loader: - kv   2:                          qwen2.block_count u32              = 28
llama_model_loader: - kv   3:                       qwen2.context_length u32              = 32768
llama_model_loader: - kv   4:                     qwen2.embedding_length u32              = 3584
llama_model_loader: - kv   5:                  qwen2.feed_forward_length u32              = 18944
llama_model_loader: - kv   6:                 qwen2.attention.head_count u32              = 28
llama_model_loader: - kv   7:              qwen2.attention.head_count_kv u32              = 4
llama_model_loader: - kv   8:                       qwen2.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv   9:     qwen2.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  10:                          general.file_type u32              = 7
llama_model_loader: - kv  11:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  12:                         tokenizer.ggml.pre str              = qwen2
llama_model_loader: - kv  13:                      tokenizer.ggml.tokens arr[str,152064]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  14:                  tokenizer.ggml.token_type arr[i32,152064]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  15:                      tokenizer.ggml.merges arr[str,151387]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  16:                tokenizer.ggml.eos_token_id u32              = 151645
llama_model_loader: - kv  17:            tokenizer.ggml.padding_token_id u32              = 151643
llama_model_loader: - kv  18:                tokenizer.ggml.bos_token_id u32              = 151643
llama_model_loader: - kv  19:                    tokenizer.chat_template str              = {% for message in messages %}{% if lo...
llama_model_loader: - kv  20:               tokenizer.ggml.add_bos_token bool             = false
llama_model_loader: - kv  21:               general.quantization_version u32              = 2
llama_model_loader: - kv  22:                      quantize.imatrix.file str              = ../Qwen2/gguf/qwen2-7b-imatrix/imatri...
llama_model_loader: - kv  23:                   quantize.imatrix.dataset str              = ../sft_2406.txt
llama_model_loader: - kv  24:             quantize.imatrix.entries_count i32              = 196
llama_model_loader: - kv  25:              quantize.imatrix.chunks_count i32              = 1937
llama_model_loader: - type  f32:  141 tensors
llama_model_loader: - type q8_0:  198 tensors
llm_load_vocab: special tokens cache size = 421
llm_load_vocab: token to piece cache size = 0.9352 MB
llm_load_print_meta: format           = GGUF V3 (latest)
llm_load_print_meta: arch             = qwen2
llm_load_print_meta: vocab type       = BPE
llm_load_print_meta: n_vocab          = 152064
llm_load_print_meta: n_merges         = 151387
llm_load_print_meta: vocab_only       = 0
llm_load_print_meta: n_ctx_train      = 32768
llm_load_print_meta: n_embd           = 3584
llm_load_print_meta: n_layer          = 28
llm_load_print_meta: n_head           = 28
llm_load_print_meta: n_head_kv        = 4
llm_load_print_meta: n_rot            = 128
llm_load_print_meta: n_swa            = 0
llm_load_print_meta: n_embd_head_k    = 128
llm_load_print_meta: n_embd_head_v    = 128
llm_load_print_meta: n_gqa            = 7
llm_load_print_meta: n_embd_k_gqa     = 512
llm_load_print_meta: n_embd_v_gqa     = 512
llm_load_print_meta: f_norm_eps       = 0.0e+00
llm_load_print_meta: f_norm_rms_eps   = 1.0e-06
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             = 18944
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        = 2
llm_load_print_meta: rope scaling     = linear
llm_load_print_meta: freq_base_train  = 1000000.0
llm_load_print_meta: freq_scale_train = 1
llm_load_print_meta: n_ctx_orig_yarn  = 32768
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: ssm_dt_b_c_rms   = 0
llm_load_print_meta: model type       = ?B
llm_load_print_meta: model ftype      = Q8_0
llm_load_print_meta: model params     = 7.62 B
llm_load_print_meta: model size       = 7.54 GiB (8.50 BPW) 
llm_load_print_meta: general.name     = qwen2-7b-instruct
llm_load_print_meta: BOS token        = 151643 '<|endoftext|>'
llm_load_print_meta: EOS token        = 151645 '<|im_end|>'
llm_load_print_meta: PAD token        = 151643 '<|endoftext|>'
llm_load_print_meta: LF token         = 148848 'ÄĬ'
llm_load_print_meta: EOT token        = 151645 '<|im_end|>'
llm_load_print_meta: max token length = 256
llm_load_tensors: ggml ctx size =    0.15 MiB
llm_load_tensors: offloading 0 repeating layers to GPU
llm_load_tensors: offloaded 0/29 layers to GPU
llm_load_tensors:        CPU buffer size =  7717.68 MiB
........................................................................................
llama_new_context_with_model: n_ctx      = 32768
llama_new_context_with_model: n_batch    = 2048
llama_new_context_with_model: n_ubatch   = 512
llama_new_context_with_model: flash_attn = 0
llama_new_context_with_model: freq_base  = 1000000.0
llama_new_context_with_model: freq_scale = 1
llama_kv_cache_init:        CPU KV buffer size =  1792.00 MiB
llama_new_context_with_model: KV self size  = 1792.00 MiB, K (f16):  896.00 MiB, V (f16):  896.00 MiB
llama_new_context_with_model:        CPU  output buffer size =     0.58 MiB
llama_new_context_with_model:        CPU compute buffer size =  1884.01 MiB
llama_new_context_with_model: graph nodes  = 986
llama_new_context_with_model: graph splits = 1

system_info: n_threads = 2 / 4 | AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | FMA = 1 | NEON = 0 | SVE = 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 | 
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 = 32768, n_batch = 2048, n_predict = 100, n_keep = 0


hello, this is a very very long story and it is very complicated.

(此处省略一部分)

llama_print_timings:        load time =    5355.79 ms
llama_print_timings:      sample time =      16.50 ms /   100 runs   (    0.17 ms per token,  6059.14 tokens per second)
llama_print_timings: prompt eval time =    1727.39 ms /     9 tokens (  191.93 ms per token,     5.21 tokens per second)
llama_print_timings:        eval time =   41066.65 ms /    99 runs   (  414.81 ms per token,     2.41 tokens per second)
llama_print_timings:       total time =   42914.72 ms /   108 tokens
Log end

运行模型 qwen2-7B.q8, 生成长度 200:

sh 复制代码
> ./llama-b3617-bin-ubuntu-x64/build/bin/llama-cli -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -n 200

(此处省略一部分)

llama_print_timings:        load time =    4641.45 ms
llama_print_timings:      sample time =      34.69 ms /   200 runs   (    0.17 ms per token,  5765.85 tokens per second)
llama_print_timings: prompt eval time =    1735.51 ms /     9 tokens (  192.83 ms per token,     5.19 tokens per second)
llama_print_timings:        eval time =   84374.46 ms /   199 runs   (  423.99 ms per token,     2.36 tokens per second)
llama_print_timings:       total time =   86360.14 ms /   208 tokens

运行模型 qwen2-7B.q8, 生成长度 500:

sh 复制代码
> ./llama-b3617-bin-ubuntu-x64/build/bin/llama-cli -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -n 500

(此处省略一部分)

llama_print_timings:        load time =    5026.41 ms
llama_print_timings:      sample time =      91.64 ms /   500 runs   (    0.18 ms per token,  5456.37 tokens per second)
llama_print_timings: prompt eval time =    1713.90 ms /     9 tokens (  190.43 ms per token,     5.25 tokens per second)
llama_print_timings:        eval time =  214729.88 ms /   499 runs   (  430.32 ms per token,     2.32 tokens per second)
llama_print_timings:       total time =  217097.31 ms /   508 tokens

运行模型 qwen2-7B.q8, 生成长度 1000:

sh 复制代码
> ./llama-b3617-bin-ubuntu-x64/build/bin/llama-cli -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -n 1000

(此处省略一部分)

llama_print_timings:        load time =    4939.31 ms
llama_print_timings:      sample time =     194.02 ms /  1000 runs   (    0.19 ms per token,  5154.00 tokens per second)
llama_print_timings: prompt eval time =    1879.29 ms /     9 tokens (  208.81 ms per token,     4.79 tokens per second)
llama_print_timings:        eval time =  440575.12 ms /   999 runs   (  441.02 ms per token,     2.27 tokens per second)
llama_print_timings:       total time =  443841.74 ms /  1008 tokens

3.2 CPU (E5-2650v3, 10C/10T/3.0GHz) x86_64 AVX2

在 5 号 (物理机) 上运行. 版本:

sh 复制代码
fc-test@MiWiFi-RA74-srv:~/llama-cpp$ ./llama-b3617-bin-ubuntu-x64/build/bin/llama-cli --version
./llama-b3617-bin-ubuntu-x64/build/bin/llama-cli: /lib64/libcurl.so.4: no version information available (required by ./llama-b3617-bin-ubuntu-x64/build/bin/llama-cli)
version: 3617 (a07c32ea)
built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu

运行模型 llama2-7B.q4, 生成长度 100:

sh 复制代码
fc-test@MiWiFi-RA74-srv:~/llama-cpp$ ./llama-b3617-bin-ubuntu-x64/build/bin/llama-cli -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -n 100
./llama-b3617-bin-ubuntu-x64/build/bin/llama-cli: /lib64/libcurl.so.4: no version information available (required by ./llama-b3617-bin-ubuntu-x64/build/bin/llama-cli)
Log start
main: build = 3617 (a07c32ea)
main: built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
main: seed  = 1724498199

(此处省略一部分)

llm_load_print_meta: max token length = 48
llm_load_tensors: ggml ctx size =    0.14 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 =  3891.24 MiB
..................................................................................................
llama_new_context_with_model: n_ctx      = 4096
llama_new_context_with_model: n_batch    = 2048
llama_new_context_with_model: n_ubatch   = 512
llama_new_context_with_model: flash_attn = 0
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 =  2048.00 MiB
llama_new_context_with_model: KV self size  = 2048.00 MiB, K (f16): 1024.00 MiB, V (f16): 1024.00 MiB
llama_new_context_with_model:        CPU  output buffer size =     0.12 MiB
llama_new_context_with_model:        CPU compute buffer size =   296.01 MiB
llama_new_context_with_model: graph nodes  = 1030
llama_new_context_with_model: graph splits = 1

system_info: n_threads = 10 / 10 | AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | FMA = 1 | NEON = 0 | SVE = 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 | 
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 = 4096, n_batch = 2048, n_predict = 100, n_keep = 1


 hello, this is a very very long story, but this is the only way I could explain what I did to solve this problem. everyone here said it cannot be done, but I did it. I don't know why I can solve it, but I did.

(此处省略一部分)

llama_print_timings:        load time =    1542.10 ms
llama_print_timings:      sample time =       4.82 ms /   100 runs   (    0.05 ms per token, 20768.43 tokens per second)
llama_print_timings: prompt eval time =     493.57 ms /    10 tokens (   49.36 ms per token,    20.26 tokens per second)
llama_print_timings:        eval time =   10175.47 ms /    99 runs   (  102.78 ms per token,     9.73 tokens per second)
llama_print_timings:       total time =   10693.97 ms /   109 tokens
Log end

运行模型 llama2-7B.q4, 生成长度 200:

sh 复制代码
$ ./llama-b3617-bin-ubuntu-x64/build/bin/llama-cli -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -n 200

(此处省略一部分)

llama_print_timings:        load time =    1607.02 ms
llama_print_timings:      sample time =       9.29 ms /   200 runs   (    0.05 ms per token, 21528.53 tokens per second)
llama_print_timings: prompt eval time =     494.35 ms /    10 tokens (   49.44 ms per token,    20.23 tokens per second)
llama_print_timings:        eval time =   20434.74 ms /   199 runs   (  102.69 ms per token,     9.74 tokens per second)
llama_print_timings:       total time =   20978.91 ms /   209 tokens

运行模型 llama2-7B.q4, 生成长度 500:

sh 复制代码
$ ./llama-b3617-bin-ubuntu-x64/build/bin/llama-cli -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -n 500

(此处省略一部分)

llama_print_timings:        load time =    1583.59 ms
llama_print_timings:      sample time =      23.55 ms /   500 runs   (    0.05 ms per token, 21226.92 tokens per second)
llama_print_timings: prompt eval time =     499.12 ms /    10 tokens (   49.91 ms per token,    20.04 tokens per second)
llama_print_timings:        eval time =   52358.53 ms /   499 runs   (  104.93 ms per token,     9.53 tokens per second)
llama_print_timings:       total time =   52987.01 ms /   509 tokens

运行模型 llama2-7B.q4, 生成长度 1000:

sh 复制代码
$ ./llama-b3617-bin-ubuntu-x64/build/bin/llama-cli -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -n 1000

(此处省略一部分)

llama_print_timings:        load time =    3247.78 ms
llama_print_timings:      sample time =      47.13 ms /  1000 runs   (    0.05 ms per token, 21218.81 tokens per second)
llama_print_timings: prompt eval time =    2596.30 ms /    10 tokens (  259.63 ms per token,     3.85 tokens per second)
llama_print_timings:        eval time =  118042.47 ms /   999 runs   (  118.16 ms per token,     8.46 tokens per second)
llama_print_timings:       total time =  120896.74 ms /  1009 tokens

运行模型 qwen2-7B.q8, 生成长度 100:

sh 复制代码
fc-test@MiWiFi-RA74-srv:~/llama-cpp$ ./llama-b3617-bin-ubuntu-x64/build/bin/llama-cli -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -n 100
./llama-b3617-bin-ubuntu-x64/build/bin/llama-cli: /lib64/libcurl.so.4: no version information available (required by ./llama-b3617-bin-ubuntu-x64/build/bin/llama-cli)
Log start
main: build = 3617 (a07c32ea)
main: built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
main: seed  = 1724498632
llama_model_loader: loaded meta data with 26 key-value pairs and 339 tensors from qwen2-7b-instruct-q8_0.gguf (version GGUF V3 (latest))

(此处省略一部分)

llm_load_print_meta: max token length = 256
llm_load_tensors: ggml ctx size =    0.15 MiB
llm_load_tensors: offloading 0 repeating layers to GPU
llm_load_tensors: offloaded 0/29 layers to GPU
llm_load_tensors:        CPU buffer size =  7717.68 MiB
........................................................................................
llama_new_context_with_model: n_ctx      = 32768
llama_new_context_with_model: n_batch    = 2048
llama_new_context_with_model: n_ubatch   = 512
llama_new_context_with_model: flash_attn = 0
llama_new_context_with_model: freq_base  = 1000000.0
llama_new_context_with_model: freq_scale = 1
llama_kv_cache_init:        CPU KV buffer size =  1792.00 MiB
llama_new_context_with_model: KV self size  = 1792.00 MiB, K (f16):  896.00 MiB, V (f16):  896.00 MiB
llama_new_context_with_model:        CPU  output buffer size =     0.58 MiB
llama_new_context_with_model:        CPU compute buffer size =  1884.01 MiB
llama_new_context_with_model: graph nodes  = 986
llama_new_context_with_model: graph splits = 1

system_info: n_threads = 10 / 10 | AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | FMA = 1 | NEON = 0 | SVE = 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 | 
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 = 32768, n_batch = 2048, n_predict = 100, n_keep = 0


hello, this is a very very long story, so i will split it into parts.

(此处省略一部分)

llama_print_timings:        load time =    1626.44 ms
llama_print_timings:      sample time =      14.31 ms /   100 runs   (    0.14 ms per token,  6987.63 tokens per second)
llama_print_timings: prompt eval time =     507.61 ms /     9 tokens (   56.40 ms per token,    17.73 tokens per second)
llama_print_timings:        eval time =   14615.79 ms /    99 runs   (  147.63 ms per token,     6.77 tokens per second)
llama_print_timings:       total time =   15238.41 ms /   108 tokens
Log end

运行模型 qwen2-7B.q8, 生成长度 200:

sh 复制代码
$ ./llama-b3617-bin-ubuntu-x64/build/bin/llama-cli -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -n 200

(此处省略一部分)

llama_print_timings:        load time =    1577.00 ms
llama_print_timings:      sample time =      28.41 ms /   200 runs   (    0.14 ms per token,  7039.03 tokens per second)
llama_print_timings: prompt eval time =     503.02 ms /     9 tokens (   55.89 ms per token,    17.89 tokens per second)
llama_print_timings:        eval time =   28940.41 ms /   199 runs   (  145.43 ms per token,     6.88 tokens per second)
llama_print_timings:       total time =   29668.90 ms /   208 tokens

运行模型 qwen2-7B.q8, 生成长度 500:

sh 复制代码
$ ./llama-b3617-bin-ubuntu-x64/build/bin/llama-cli -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -n 500

(此处省略一部分)

llama_print_timings:        load time =    1598.72 ms
llama_print_timings:      sample time =      72.10 ms /   500 runs   (    0.14 ms per token,  6935.01 tokens per second)
llama_print_timings: prompt eval time =     502.73 ms /     9 tokens (   55.86 ms per token,    17.90 tokens per second)
llama_print_timings:        eval time =   72983.23 ms /   499 runs   (  146.26 ms per token,     6.84 tokens per second)
llama_print_timings:       total time =   74061.66 ms /   508 tokens

运行模型 qwen2-7B.q8, 生成长度 1000:

sh 复制代码
$ ./llama-b3617-bin-ubuntu-x64/build/bin/llama-cli -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -n 1000

(此处省略一部分)

llama_print_timings:        load time =    1602.06 ms
llama_print_timings:      sample time =     144.15 ms /  1000 runs   (    0.14 ms per token,  6937.31 tokens per second)
llama_print_timings: prompt eval time =     509.66 ms /     9 tokens (   56.63 ms per token,    17.66 tokens per second)
llama_print_timings:        eval time =  149336.77 ms /   999 runs   (  149.49 ms per token,     6.69 tokens per second)
llama_print_timings:       total time =  150983.01 ms /  1008 tokens

3.3 CPU (r5-5600g, 6C/12T/4.4GHz) x86_64 AVX2

在 6 号 PC (物理机) 上运行. 版本:

sh 复制代码
> ./llama-b3617-bin-ubuntu-x64/build/bin/llama-cli --version
version: 3617 (a07c32ea)
built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu

运行模型 llama2-7B.q4, 生成长度 100:

sh 复制代码
> ./llama-b3617-bin-ubuntu-x64/build/bin/llama-cli -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -n 100
Log start
main: build = 3617 (a07c32ea)
main: built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
main: seed  = 1724488187

(此处省略一部分)

llm_load_print_meta: max token length = 48
llm_load_tensors: ggml ctx size =    0.14 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 =  3891.24 MiB
..................................................................................................
llama_new_context_with_model: n_ctx      = 4096
llama_new_context_with_model: n_batch    = 2048
llama_new_context_with_model: n_ubatch   = 512
llama_new_context_with_model: flash_attn = 0
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 =  2048.00 MiB
llama_new_context_with_model: KV self size  = 2048.00 MiB, K (f16): 1024.00 MiB, V (f16): 1024.00 MiB
llama_new_context_with_model:        CPU  output buffer size =     0.12 MiB
llama_new_context_with_model:        CPU compute buffer size =   296.01 MiB
llama_new_context_with_model: graph nodes  = 1030
llama_new_context_with_model: graph splits = 1

system_info: n_threads = 6 / 12 | AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | FMA = 1 | NEON = 0 | SVE = 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 | 
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 = 4096, n_batch = 2048, n_predict = 100, n_keep = 1


 hello, this is a very very long story, but i think it's important to read.

(此处省略一部分)

llama_print_timings:        load time =     649.76 ms
llama_print_timings:      sample time =       2.40 ms /   100 runs   (    0.02 ms per token, 41701.42 tokens per second)
llama_print_timings: prompt eval time =     311.37 ms /    10 tokens (   31.14 ms per token,    32.12 tokens per second)
llama_print_timings:        eval time =    9771.88 ms /    99 runs   (   98.71 ms per token,    10.13 tokens per second)
llama_print_timings:       total time =   10092.46 ms /   109 tokens
Log end

运行模型 llama2-7B.q4, 生成长度 200:

sh 复制代码
> ./llama-b3617-bin-ubuntu-x64/build/bin/llama-cli -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -n 200

(此处省略一部分)

llama_print_timings:        load time =     650.76 ms
llama_print_timings:      sample time =       5.08 ms /   200 runs   (    0.03 ms per token, 39331.37 tokens per second)
llama_print_timings: prompt eval time =     308.01 ms /    10 tokens (   30.80 ms per token,    32.47 tokens per second)
llama_print_timings:        eval time =   19887.24 ms /   199 runs   (   99.94 ms per token,    10.01 tokens per second)
llama_print_timings:       total time =   20214.70 ms /   209 tokens

运行模型 llama2-7B.q4, 生成长度 500:

sh 复制代码
> ./llama-b3617-bin-ubuntu-x64/build/bin/llama-cli -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -n 500

(此处省略一部分)

llama_print_timings:        load time =     648.51 ms
llama_print_timings:      sample time =      12.16 ms /   500 runs   (    0.02 ms per token, 41128.57 tokens per second)
llama_print_timings: prompt eval time =     308.95 ms /    10 tokens (   30.89 ms per token,    32.37 tokens per second)
llama_print_timings:        eval time =   51687.76 ms /   499 runs   (  103.58 ms per token,     9.65 tokens per second)
llama_print_timings:       total time =   52043.21 ms /   509 tokens

运行模型 llama2-7B.q4, 生成长度 1000:

sh 复制代码
> ./llama-b3617-bin-ubuntu-x64/build/bin/llama-cli -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -n 1000

(此处省略一部分)

llama_print_timings:        load time =     648.60 ms
llama_print_timings:      sample time =      24.13 ms /  1000 runs   (    0.02 ms per token, 41438.75 tokens per second)
llama_print_timings: prompt eval time =     311.58 ms /    10 tokens (   31.16 ms per token,    32.09 tokens per second)
llama_print_timings:        eval time =  107409.32 ms /   999 runs   (  107.52 ms per token,     9.30 tokens per second)
llama_print_timings:       total time =  107815.70 ms /  1009 tokens

运行模型 qwen2-7B.q8, 生成长度 100:

sh 复制代码
> ./llama-b3617-bin-ubuntu-x64/build/bin/llama-cli -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -n 100
Log start
main: build = 3617 (a07c32ea)
main: built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
main: seed  = 1724489633
llama_model_loader: loaded meta data with 26 key-value pairs and 339 tensors from qwen2-7b-instruct-q8_0.gguf (version GGUF V3 (latest))

(此处省略一部分)

llm_load_print_meta: max token length = 256
llm_load_tensors: ggml ctx size =    0.15 MiB
llm_load_tensors: offloading 0 repeating layers to GPU
llm_load_tensors: offloaded 0/29 layers to GPU
llm_load_tensors:        CPU buffer size =  7717.68 MiB
........................................................................................
llama_new_context_with_model: n_ctx      = 32768
llama_new_context_with_model: n_batch    = 2048
llama_new_context_with_model: n_ubatch   = 512
llama_new_context_with_model: flash_attn = 0
llama_new_context_with_model: freq_base  = 1000000.0
llama_new_context_with_model: freq_scale = 1
llama_kv_cache_init:        CPU KV buffer size =  1792.00 MiB
llama_new_context_with_model: KV self size  = 1792.00 MiB, K (f16):  896.00 MiB, V (f16):  896.00 MiB
llama_new_context_with_model:        CPU  output buffer size =     0.58 MiB
llama_new_context_with_model:        CPU compute buffer size =  1884.01 MiB
llama_new_context_with_model: graph nodes  = 986
llama_new_context_with_model: graph splits = 1

system_info: n_threads = 6 / 12 | AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | FMA = 1 | NEON = 0 | SVE = 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 | 
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 = 32768, n_batch = 2048, n_predict = 100, n_keep = 0


hello, this is a very very long story about my friend and her husband, so please bear with me.

(此处省略一部分)

llama_print_timings:        load time =    1158.78 ms
llama_print_timings:      sample time =       8.32 ms /   100 runs   (    0.08 ms per token, 12025.01 tokens per second)
llama_print_timings: prompt eval time =     457.69 ms /     9 tokens (   50.85 ms per token,    19.66 tokens per second)
llama_print_timings:        eval time =   17878.08 ms /    99 runs   (  180.59 ms per token,     5.54 tokens per second)
llama_print_timings:       total time =   18402.49 ms /   108 tokens
Log end

运行模型 qwen2-7B.q8, 生成长度 200:

sh 复制代码
> ./llama-b3617-bin-ubuntu-x64/build/bin/llama-cli -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -n 200

(此处省略一部分)

llama_print_timings:        load time =    1109.41 ms
llama_print_timings:      sample time =      13.17 ms /   200 runs   (    0.07 ms per token, 15181.42 tokens per second)
llama_print_timings: prompt eval time =     496.57 ms /     9 tokens (   55.17 ms per token,    18.12 tokens per second)
llama_print_timings:        eval time =   35791.00 ms /   199 runs   (  179.85 ms per token,     5.56 tokens per second)
llama_print_timings:       total time =   36411.02 ms /   208 tokens

运行模型 qwen2-7B.q8, 生成长度 500:

sh 复制代码
> ./llama-b3617-bin-ubuntu-x64/build/bin/llama-cli -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -n 500

(此处省略一部分)

llama_print_timings:        load time =    1061.77 ms
llama_print_timings:      sample time =      40.61 ms /   500 runs   (    0.08 ms per token, 12311.03 tokens per second)
llama_print_timings: prompt eval time =     409.44 ms /     9 tokens (   45.49 ms per token,    21.98 tokens per second)
llama_print_timings:        eval time =   90250.99 ms /   499 runs   (  180.86 ms per token,     5.53 tokens per second)
llama_print_timings:       total time =   90991.53 ms /   508 tokens

运行模型 qwen2-7B.q8, 生成长度 1000:

sh 复制代码
> ./llama-b3617-bin-ubuntu-x64/build/bin/llama-cli -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -n 1000

(此处省略一部分)

llama_print_timings:        load time =     977.25 ms
llama_print_timings:      sample time =      60.87 ms /  1000 runs   (    0.06 ms per token, 16428.99 tokens per second)
llama_print_timings: prompt eval time =     479.25 ms /     9 tokens (   53.25 ms per token,    18.78 tokens per second)
llama_print_timings:        eval time =  182514.10 ms /   999 runs   (  182.70 ms per token,     5.47 tokens per second)
llama_print_timings:       total time =  183593.03 ms /  1008 tokens

3.4 iGPU (Intel HD520, i5-6200U) vulkan

在 4 号 PC (物理机) 上运行. 版本:

sh 复制代码
> ./llama-cli-vulkan-b3617 --version
version: 1 (a07c32e)
built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu

运行模型 llama2-7B.q4, 生成长度 100:

sh 复制代码
> ./llama-cli-vulkan-b3617 -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -ngl 33 -n 100
Log start
main: build = 1 (a07c32e)
main: built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
main: seed  = 1724502840

(此处省略一部分)

llm_load_print_meta: max token length = 48
ggml_vulkan: Found 1 Vulkan devices:
Vulkan0: Intel(R) HD Graphics 520 (SKL GT2) (Intel open-source Mesa driver) | uma: 1 | fp16: 1 | warp size: 32
llm_load_tensors: ggml ctx size =    0.27 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 =    70.31 MiB
llm_load_tensors: Intel(R) HD Graphics 520 (SKL GT2) buffer size =  3820.93 MiB
..................................................................................................
llama_new_context_with_model: n_ctx      = 4096
llama_new_context_with_model: n_batch    = 2048
llama_new_context_with_model: n_ubatch   = 512
llama_new_context_with_model: flash_attn = 0
llama_new_context_with_model: freq_base  = 10000.0
llama_new_context_with_model: freq_scale = 1
llama_kv_cache_init: Intel(R) HD Graphics 520 (SKL GT2) KV buffer size =  2048.00 MiB
llama_new_context_with_model: KV self size  = 2048.00 MiB, K (f16): 1024.00 MiB, V (f16): 1024.00 MiB
llama_new_context_with_model: Vulkan_Host  output buffer size =     0.12 MiB
llama_new_context_with_model: Intel(R) HD Graphics 520 (SKL GT2) compute buffer size =   296.00 MiB
llama_new_context_with_model: Vulkan_Host compute buffer size =    16.01 MiB
llama_new_context_with_model: graph nodes  = 1030
llama_new_context_with_model: graph splits = 2

system_info: n_threads = 2 / 4 | AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | FMA = 1 | NEON = 0 | SVE = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 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 = 4096, n_batch = 2048, n_predict = 100, n_keep = 1


 hello, this is a very very long story but i will try and make it short

(此处省略一部分)

llama_print_timings:        load time =   27305.92 ms
llama_print_timings:      sample time =      20.64 ms /   100 runs   (    0.21 ms per token,  4844.49 tokens per second)
llama_print_timings: prompt eval time =   10725.27 ms /    10 tokens ( 1072.53 ms per token,     0.93 tokens per second)
llama_print_timings:        eval time =  104246.69 ms /    99 runs   ( 1053.00 ms per token,     0.95 tokens per second)
llama_print_timings:       total time =  115065.04 ms /   109 tokens
Log end

运行模型 llama2-7B.q4, 生成长度 200:

sh 复制代码
> ./llama-cli-vulkan-b3617 -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -ngl 33 -n 200

(此处省略一部分)

llama_print_timings:        load time =   26358.11 ms
llama_print_timings:      sample time =      43.34 ms /   200 runs   (    0.22 ms per token,  4615.21 tokens per second)
llama_print_timings: prompt eval time =   10579.07 ms /    10 tokens ( 1057.91 ms per token,     0.95 tokens per second)
llama_print_timings:        eval time =  209900.70 ms /   199 runs   ( 1054.78 ms per token,     0.95 tokens per second)
llama_print_timings:       total time =  220666.27 ms /   209 tokens

运行模型 llama2-7B.q4, 生成长度 500:

sh 复制代码
> ./llama-cli-vulkan-b3617 -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -ngl 33 -n 500

(此处省略一部分)

llama_print_timings:        load time =   27769.47 ms
llama_print_timings:      sample time =     100.38 ms /   500 runs   (    0.20 ms per token,  4981.17 tokens per second)
llama_print_timings: prompt eval time =   10573.54 ms /    10 tokens ( 1057.35 ms per token,     0.95 tokens per second)
llama_print_timings:        eval time =  532338.80 ms /   499 runs   ( 1066.81 ms per token,     0.94 tokens per second)
llama_print_timings:       total time =  543350.42 ms /   509 tokens

运行模型 llama2-7B.q4, 生成长度 1000:

sh 复制代码
> ./llama-cli-vulkan-b3617 -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -ngl 33 -n 1000

(此处省略一部分)

llama_print_timings:        load time =   29646.65 ms
llama_print_timings:      sample time =     179.74 ms /  1000 runs   (    0.18 ms per token,  5563.62 tokens per second)
llama_print_timings: prompt eval time =   10538.36 ms /    10 tokens ( 1053.84 ms per token,     0.95 tokens per second)
llama_print_timings:        eval time = 1089916.74 ms /   999 runs   ( 1091.01 ms per token,     0.92 tokens per second)
llama_print_timings:       total time = 1101057.43 ms /  1009 tokens

运行模型 qwen2-7B.q8. 错误, 无法运行, 提示内存不足:

sh 复制代码
> ./llama-cli-vulkan-b3617 -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -ngl 33 -n 100
Log start
main: build = 1 (a07c32e)
main: built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
main: seed  = 1724508115
llama_model_loader: loaded meta data with 26 key-value pairs and 339 tensors from qwen2-7b-instruct-q8_0.gguf (version GGUF V3 (latest))

(此处省略一部分)

llm_load_print_meta: max token length = 256
ggml_vulkan: Found 1 Vulkan devices:
Vulkan0: Intel(R) HD Graphics 520 (SKL GT2) (Intel open-source Mesa driver) | uma: 1 | fp16: 1 | warp size: 32
llm_load_tensors: ggml ctx size =    0.30 MiB
llm_load_tensors: offloading 28 repeating layers to GPU
llm_load_tensors: offloading non-repeating layers to GPU
llm_load_tensors: offloaded 29/29 layers to GPU
llm_load_tensors:        CPU buffer size =   552.23 MiB
llm_load_tensors: Intel(R) HD Graphics 520 (SKL GT2) buffer size =  7165.44 MiB
........................................................................................
llama_new_context_with_model: n_ctx      = 32768
llama_new_context_with_model: n_batch    = 2048
llama_new_context_with_model: n_ubatch   = 512
llama_new_context_with_model: flash_attn = 0
llama_new_context_with_model: freq_base  = 1000000.0
llama_new_context_with_model: freq_scale = 1
ggml_vulkan: Device memory allocation of size 1879048192 failed.
ggml_vulkan: vk::Device::allocateMemory: ErrorOutOfDeviceMemory
llama_kv_cache_init: failed to allocate buffer for kv cache
llama_new_context_with_model: llama_kv_cache_init() failed for self-attention cache
llama_init_from_gpt_params: error: failed to create context with model 'qwen2-7b-instruct-q8_0.gguf'
main: error: unable to load model

3.5 iGPU (AMD Radeon Vega 7, r5-5600g) vulkan

在 6 号 PC (物理机) 上运行. 版本:

sh 复制代码
> ./llama-cli-vulkan-b3617 --version
version: 1 (a07c32e)
built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu

运行模型 llama2-7B.q4, 生成长度 100:

sh 复制代码
> ./llama-cli-vulkan-b3617 -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -n 100 -ngl 33
Log start
main: build = 1 (a07c32e)
main: built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
main: seed  = 1724488777

(此处省略一部分)

llm_load_print_meta: max token length = 48
ggml_vulkan: Found 1 Vulkan devices:
Vulkan0: AMD Radeon Graphics (RADV RENOIR) (radv) | uma: 1 | fp16: 1 | warp size: 64
llm_load_tensors: ggml ctx size =    0.27 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 =    70.31 MiB
llm_load_tensors: AMD Radeon Graphics (RADV RENOIR) buffer size =  3820.93 MiB
..................................................................................................
llama_new_context_with_model: n_ctx      = 4096
llama_new_context_with_model: n_batch    = 2048
llama_new_context_with_model: n_ubatch   = 512
llama_new_context_with_model: flash_attn = 0
llama_new_context_with_model: freq_base  = 10000.0
llama_new_context_with_model: freq_scale = 1
llama_kv_cache_init: AMD Radeon Graphics (RADV RENOIR) KV buffer size =  2048.00 MiB
llama_new_context_with_model: KV self size  = 2048.00 MiB, K (f16): 1024.00 MiB, V (f16): 1024.00 MiB
llama_new_context_with_model: Vulkan_Host  output buffer size =     0.12 MiB
llama_new_context_with_model: AMD Radeon Graphics (RADV RENOIR) compute buffer size =   296.00 MiB
llama_new_context_with_model: Vulkan_Host compute buffer size =    16.01 MiB
llama_new_context_with_model: graph nodes  = 1030
llama_new_context_with_model: graph splits = 2

system_info: n_threads = 6 / 12 | AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | FMA = 1 | NEON = 0 | SVE = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 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 = 4096, n_batch = 2048, n_predict = 100, n_keep = 1


 hello, this is a very very long story and it's only the first episode.

(此处省略一部分)

llama_print_timings:        load time =    3300.29 ms
llama_print_timings:      sample time =       4.41 ms /   100 runs   (    0.04 ms per token, 22686.03 tokens per second)
llama_print_timings: prompt eval time =    1028.22 ms /    10 tokens (  102.82 ms per token,     9.73 tokens per second)
llama_print_timings:        eval time =   23080.64 ms /    99 runs   (  233.14 ms per token,     4.29 tokens per second)
llama_print_timings:       total time =   24122.46 ms /   109 tokens
Log end

运行模型 llama2-7B.q4, 生成长度 200:

sh 复制代码
> ./llama-cli-vulkan-b3617 -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -n 200 -ngl 33

(此处省略一部分)

llama_print_timings:        load time =    3410.94 ms
llama_print_timings:      sample time =       8.64 ms /   200 runs   (    0.04 ms per token, 23153.51 tokens per second)
llama_print_timings: prompt eval time =    1027.37 ms /    10 tokens (  102.74 ms per token,     9.73 tokens per second)
llama_print_timings:        eval time =   46620.34 ms /   199 runs   (  234.27 ms per token,     4.27 tokens per second)
llama_print_timings:       total time =   47674.32 ms /   209 tokens

运行模型 llama2-7B.q4, 生成长度 500:

sh 复制代码
> ./llama-cli-vulkan-b3617 -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -n 500 -ngl 33

(此处省略一部分)

llama_print_timings:        load time =    3389.70 ms
llama_print_timings:      sample time =      21.42 ms /   500 runs   (    0.04 ms per token, 23339.40 tokens per second)
llama_print_timings: prompt eval time =    1026.09 ms /    10 tokens (  102.61 ms per token,     9.75 tokens per second)
llama_print_timings:        eval time =  118409.44 ms /   499 runs   (  237.29 ms per token,     4.21 tokens per second)
llama_print_timings:       total time =  119502.95 ms /   509 tokens

运行模型 llama2-7B.q4, 生成长度 1000:

sh 复制代码
> ./llama-cli-vulkan-b3617 -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -n 1000 -ngl 33

(此处省略一部分)

llama_print_timings:        load time =    3362.42 ms
llama_print_timings:      sample time =      43.25 ms /  1000 runs   (    0.04 ms per token, 23120.85 tokens per second)
llama_print_timings: prompt eval time =    1027.78 ms /    10 tokens (  102.78 ms per token,     9.73 tokens per second)
llama_print_timings:        eval time =  242531.02 ms /   999 runs   (  242.77 ms per token,     4.12 tokens per second)
llama_print_timings:       total time =  243694.80 ms /  1009 tokens

运行模型 qwen2-7B.q8, 生成长度 100:

sh 复制代码
> ./llama-cli-vulkan-b3617 -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -n 100 -ngl 33
Log start
main: build = 1 (a07c32e)
main: built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
main: seed  = 1724490279
llama_model_loader: loaded meta data with 26 key-value pairs and 339 tensors from qwen2-7b-instruct-q8_0.gguf (version GGUF V3 (latest))

(此处省略一部分)

llm_load_print_meta: max token length = 256
ggml_vulkan: Found 1 Vulkan devices:
Vulkan0: AMD Radeon Graphics (RADV RENOIR) (radv) | uma: 1 | fp16: 1 | warp size: 64
llm_load_tensors: ggml ctx size =    0.30 MiB
llm_load_tensors: offloading 28 repeating layers to GPU
llm_load_tensors: offloading non-repeating layers to GPU
llm_load_tensors: offloaded 29/29 layers to GPU
llm_load_tensors:        CPU buffer size =   552.23 MiB
llm_load_tensors: AMD Radeon Graphics (RADV RENOIR) buffer size =  7165.44 MiB
........................................................................................
llama_new_context_with_model: n_ctx      = 32768
llama_new_context_with_model: n_batch    = 2048
llama_new_context_with_model: n_ubatch   = 512
llama_new_context_with_model: flash_attn = 0
llama_new_context_with_model: freq_base  = 1000000.0
llama_new_context_with_model: freq_scale = 1
llama_kv_cache_init: AMD Radeon Graphics (RADV RENOIR) KV buffer size =  1792.00 MiB
llama_new_context_with_model: KV self size  = 1792.00 MiB, K (f16):  896.00 MiB, V (f16):  896.00 MiB
llama_new_context_with_model: Vulkan_Host  output buffer size =     0.58 MiB
llama_new_context_with_model: AMD Radeon Graphics (RADV RENOIR) compute buffer size =  1884.00 MiB
llama_new_context_with_model: Vulkan_Host compute buffer size =    71.01 MiB
llama_new_context_with_model: graph nodes  = 986
llama_new_context_with_model: graph splits = 2

system_info: n_threads = 6 / 12 | AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | FMA = 1 | NEON = 0 | SVE = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 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 = 32768, n_batch = 2048, n_predict = 100, n_keep = 0


hello, this is a very very long story, but I'm going to do my best to explain in a concise manner:

(此处省略一部分)

llama_print_timings:        load time =    8781.85 ms
llama_print_timings:      sample time =       9.19 ms /   100 runs   (    0.09 ms per token, 10880.21 tokens per second)
llama_print_timings: prompt eval time =     913.76 ms /     9 tokens (  101.53 ms per token,     9.85 tokens per second)
llama_print_timings:        eval time =   34897.82 ms /    99 runs   (  352.50 ms per token,     2.84 tokens per second)
llama_print_timings:       total time =   35889.02 ms /   108 tokens
Log end

运行模型 qwen2-7B.q8, 生成长度 200:

sh 复制代码
> ./llama-cli-vulkan-b3617 -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -n 200 -ngl 33

(此处省略一部分)

llama_print_timings:        load time =    8249.67 ms
llama_print_timings:      sample time =      17.88 ms /   200 runs   (    0.09 ms per token, 11185.68 tokens per second)
llama_print_timings: prompt eval time =     909.22 ms /     9 tokens (  101.02 ms per token,     9.90 tokens per second)
llama_print_timings:        eval time =   70426.63 ms /   199 runs   (  353.90 ms per token,     2.83 tokens per second)
llama_print_timings:       total time =   71489.45 ms /   208 tokens

运行模型 qwen2-7B.q8, 生成长度 500:

sh 复制代码
> ./llama-cli-vulkan-b3617 -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -n 500 -ngl 33

(此处省略一部分)

llama_print_timings:        load time =    6014.76 ms
llama_print_timings:      sample time =      46.23 ms /   500 runs   (    0.09 ms per token, 10815.96 tokens per second)
llama_print_timings: prompt eval time =     916.14 ms /     9 tokens (  101.79 ms per token,     9.82 tokens per second)
llama_print_timings:        eval time =  177508.81 ms /   499 runs   (  355.73 ms per token,     2.81 tokens per second)
llama_print_timings:       total time =  178809.12 ms /   508 tokens

运行模型 qwen2-7B.q8, 生成长度 1000:

sh 复制代码
> ./llama-cli-vulkan-b3617 -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -n 1000 -ngl 33

(此处省略一部分)

llama_print_timings:        load time =    6662.38 ms
llama_print_timings:      sample time =      89.55 ms /  1000 runs   (    0.09 ms per token, 11167.57 tokens per second)
llama_print_timings: prompt eval time =     916.79 ms /     9 tokens (  101.87 ms per token,     9.82 tokens per second)
llama_print_timings:        eval time =  358831.15 ms /   999 runs   (  359.19 ms per token,     2.78 tokens per second)
llama_print_timings:       total time =  360504.90 ms /  1008 tokens

3.6 dGPU (A770) vulkan

在 6 号 (虚拟机) 上运行. 版本:

sh 复制代码
a2@a2s:~$ ./llama-cli-vulkan-b3617 --version
version: 1 (a07c32e)
built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu

运行模型 llama2-7B.q4, 生成长度 100:

sh 复制代码
a2@a2s:~$ ./llama-cli-vulkan-b3617 -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -n 100 -ngl 33
Log start
main: build = 1 (a07c32e)
main: built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
main: seed  = 1724492722

(此处省略一部分)

llm_load_print_meta: max token length = 48
ggml_vulkan: Found 1 Vulkan devices:
Vulkan0: Intel(R) Arc(tm) A770 Graphics (DG2) (Intel open-source Mesa driver) | uma: 0 | fp16: 1 | warp size: 32
llm_load_tensors: ggml ctx size =    0.27 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 =    70.31 MiB
llm_load_tensors: Intel(R) Arc(tm) A770 Graphics (DG2) buffer size =  3820.93 MiB
..................................................................................................
llama_new_context_with_model: n_ctx      = 4096
llama_new_context_with_model: n_batch    = 2048
llama_new_context_with_model: n_ubatch   = 512
llama_new_context_with_model: flash_attn = 0
llama_new_context_with_model: freq_base  = 10000.0
llama_new_context_with_model: freq_scale = 1
llama_kv_cache_init: Intel(R) Arc(tm) A770 Graphics (DG2) KV buffer size =  2048.00 MiB
llama_new_context_with_model: KV self size  = 2048.00 MiB, K (f16): 1024.00 MiB, V (f16): 1024.00 MiB
llama_new_context_with_model: Vulkan_Host  output buffer size =     0.12 MiB
llama_new_context_with_model: Intel(R) Arc(tm) A770 Graphics (DG2) compute buffer size =   296.00 MiB
llama_new_context_with_model: Vulkan_Host compute buffer size =    16.01 MiB
llama_new_context_with_model: graph nodes  = 1030
llama_new_context_with_model: graph splits = 2

system_info: n_threads = 4 / 4 | AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | FMA = 1 | NEON = 0 | SVE = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 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 = 4096, n_batch = 2048, n_predict = 100, n_keep = 1


 hello, this is a very very long story, and you can ignore most of it, i'm just putting it here for reference in case anyone has a similar issue.

(此处省略一部分)

llama_print_timings:        load time =    2274.09 ms
llama_print_timings:      sample time =       4.14 ms /   100 runs   (    0.04 ms per token, 24148.76 tokens per second)
llama_print_timings: prompt eval time =     440.70 ms /    10 tokens (   44.07 ms per token,    22.69 tokens per second)
llama_print_timings:        eval time =    3809.51 ms /    99 runs   (   38.48 ms per token,    25.99 tokens per second)
llama_print_timings:       total time =    4262.46 ms /   109 tokens
Log end

运行模型 llama2-7B.q4, 生成长度 200:

sh 复制代码
a2@a2s:~$ ./llama-cli-vulkan-b3617 -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -n 200 -ngl 33

(此处省略一部分)

llama_print_timings:        load time =    2308.60 ms
llama_print_timings:      sample time =       8.50 ms /   200 runs   (    0.04 ms per token, 23518.34 tokens per second)
llama_print_timings: prompt eval time =     441.26 ms /    10 tokens (   44.13 ms per token,    22.66 tokens per second)
llama_print_timings:        eval time =    7704.86 ms /   199 runs   (   38.72 ms per token,    25.83 tokens per second)
llama_print_timings:       total time =    8171.87 ms /   209 tokens

运行模型 llama2-7B.q4, 生成长度 500:

sh 复制代码
a2@a2s:~$ ./llama-cli-vulkan-b3617 -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -n 500 -ngl 33

(此处省略一部分)

llama_print_timings:        load time =    2296.68 ms
llama_print_timings:      sample time =      21.31 ms /   500 runs   (    0.04 ms per token, 23460.96 tokens per second)
llama_print_timings: prompt eval time =     440.77 ms /    10 tokens (   44.08 ms per token,    22.69 tokens per second)
llama_print_timings:        eval time =   19597.74 ms /   499 runs   (   39.27 ms per token,    25.46 tokens per second)
llama_print_timings:       total time =   20102.66 ms /   509 tokens

运行模型 llama2-7B.q4, 生成长度 1000:

sh 复制代码
a2@a2s:~$ ./llama-cli-vulkan-b3617 -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -n 1000 -ngl 33

(此处省略一部分)

llama_print_timings:        load time =    2273.46 ms
llama_print_timings:      sample time =      42.10 ms /  1000 runs   (    0.04 ms per token, 23751.84 tokens per second)
llama_print_timings: prompt eval time =     441.47 ms /    10 tokens (   44.15 ms per token,    22.65 tokens per second)
llama_print_timings:        eval time =   40262.07 ms /   999 runs   (   40.30 ms per token,    24.81 tokens per second)
llama_print_timings:       total time =   40827.46 ms /  1009 tokens

运行模型 qwen2-7B.q8, 生成长度 100:

sh 复制代码
a2@a2s:~$ ./llama-cli-vulkan-b3617 -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -n 100 -ngl 33
Log start
main: build = 1 (a07c32e)
main: built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
main: seed  = 1724493121
llama_model_loader: loaded meta data with 26 key-value pairs and 339 tensors from qwen2-7b-instruct-q8_0.gguf (version GGUF V3 (latest))

(此处省略一部分)

llm_load_print_meta: max token length = 256
ggml_vulkan: Found 1 Vulkan devices:
Vulkan0: Intel(R) Arc(tm) A770 Graphics (DG2) (Intel open-source Mesa driver) | uma: 0 | fp16: 1 | warp size: 32
llm_load_tensors: ggml ctx size =    0.30 MiB
llm_load_tensors: offloading 28 repeating layers to GPU
llm_load_tensors: offloading non-repeating layers to GPU
llm_load_tensors: offloaded 29/29 layers to GPU
llm_load_tensors:        CPU buffer size =   552.23 MiB
llm_load_tensors: Intel(R) Arc(tm) A770 Graphics (DG2) buffer size =  7165.44 MiB
........................................................................................
llama_new_context_with_model: n_ctx      = 32768
llama_new_context_with_model: n_batch    = 2048
llama_new_context_with_model: n_ubatch   = 512
llama_new_context_with_model: flash_attn = 0
llama_new_context_with_model: freq_base  = 1000000.0
llama_new_context_with_model: freq_scale = 1
llama_kv_cache_init: Intel(R) Arc(tm) A770 Graphics (DG2) KV buffer size =  1792.00 MiB
llama_new_context_with_model: KV self size  = 1792.00 MiB, K (f16):  896.00 MiB, V (f16):  896.00 MiB
llama_new_context_with_model: Vulkan_Host  output buffer size =     0.58 MiB
llama_new_context_with_model: Intel(R) Arc(tm) A770 Graphics (DG2) compute buffer size =  1884.00 MiB
llama_new_context_with_model: Vulkan_Host compute buffer size =    71.01 MiB
llama_new_context_with_model: graph nodes  = 986
llama_new_context_with_model: graph splits = 2

system_info: n_threads = 4 / 4 | AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | FMA = 1 | NEON = 0 | SVE = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 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 = 32768, n_batch = 2048, n_predict = 100, n_keep = 0


hello, this is a very very long story with multiple characters, but i will try to write it in a way that makes it easy to follow. 

(此处省略一部分)

llama_print_timings:        load time =    8202.05 ms
llama_print_timings:      sample time =      10.16 ms /   100 runs   (    0.10 ms per token,  9839.61 tokens per second)
llama_print_timings: prompt eval time =     587.73 ms /     9 tokens (   65.30 ms per token,    15.31 tokens per second)
llama_print_timings:        eval time =    4755.44 ms /    99 runs   (   48.03 ms per token,    20.82 tokens per second)
llama_print_timings:       total time =    5460.46 ms /   108 tokens
Log end

运行模型 qwen2-7B.q8, 生成长度 200:

sh 复制代码
a2@a2s:~$ ./llama-cli-vulkan-b3617 -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -n 200 -ngl 33

(此处省略一部分)

llama_print_timings:        load time =    6642.05 ms
llama_print_timings:      sample time =      19.91 ms /   200 runs   (    0.10 ms per token, 10043.19 tokens per second)
llama_print_timings: prompt eval time =     587.07 ms /     9 tokens (   65.23 ms per token,    15.33 tokens per second)
llama_print_timings:        eval time =    9581.81 ms /   199 runs   (   48.15 ms per token,    20.77 tokens per second)
llama_print_timings:       total time =   10348.91 ms /   208 tokens

运行模型 qwen2-7B.q8, 生成长度 500:

sh 复制代码
a2@a2s:~$ ./llama-cli-vulkan-b3617 -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -n 500 -ngl 33

(此处省略一部分)

llama_print_timings:        load time =    6756.91 ms
llama_print_timings:      sample time =      51.43 ms /   500 runs   (    0.10 ms per token,  9722.33 tokens per second)
llama_print_timings: prompt eval time =     588.10 ms /     9 tokens (   65.34 ms per token,    15.30 tokens per second)
llama_print_timings:        eval time =   24196.44 ms /   499 runs   (   48.49 ms per token,    20.62 tokens per second)
llama_print_timings:       total time =   25212.38 ms /   508 tokens

运行模型 qwen2-7B.q8, 生成长度 1000:

sh 复制代码
a2@a2s:~$ ./llama-cli-vulkan-b3617 -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -n 1000 -ngl 33

(此处省略一部分)

llama_print_timings:        load time =    6664.69 ms
llama_print_timings:      sample time =      92.37 ms /  1000 runs   (    0.09 ms per token, 10825.91 tokens per second)
llama_print_timings: prompt eval time =     586.92 ms /     9 tokens (   65.21 ms per token,    15.33 tokens per second)
llama_print_timings:        eval time =   48610.18 ms /   999 runs   (   48.66 ms per token,    20.55 tokens per second)
llama_print_timings:       total time =   49939.72 ms /  1008 tokens

3.7 dGPU (A770) SYCL

在 6 号 (虚拟机) 上运行. 准备工作:

sh 复制代码
a2@a2s:~$ source /opt/intel/oneapi/setvars.sh
 
:: initializing oneAPI environment ...
   -bash: BASH_VERSION = 5.1.16(1)-release
   args: Using "$@" for setvars.sh arguments: 
:: ccl -- latest
:: compiler -- latest
:: debugger -- latest
:: dev-utilities -- latest
:: mkl -- latest
:: mpi -- latest
:: tbb -- latest
:: oneAPI environment initialized ::
 
a2@a2s:~$ export ZES_ENABLE_SYSMAN=1
a2@a2s:~$ export USE_XETLA=OFF
a2@a2s:~$ export SYCL_CACHE_PERSISTENT=1
a2@a2s:~$ export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
a2@a2s:~$ sycl-ls
[opencl:cpu][opencl:0] Intel(R) OpenCL, AMD Ryzen 5 5600G with Radeon Graphics          OpenCL 3.0 (Build 0) [2024.18.7.0.11_160000]
[opencl:gpu][opencl:1] Intel(R) OpenCL Graphics, Intel(R) Arc(TM) A770 Graphics OpenCL 3.0 NEO  [24.22.29735.27]
[level_zero:gpu][level_zero:0] Intel(R) Level-Zero, Intel(R) Arc(TM) A770 Graphics 1.3 [1.3.29735]

版本:

sh 复制代码
a2@a2s:~$ ./llama-cli-sycl-b3617-f32 --version
version: 1 (a07c32e)
built with Intel(R) oneAPI DPC++/C++ Compiler 2024.2.1 (2024.2.1.20240711) for x86_64-unknown-linux-gnu
a2@a2s:~$ ./llama-cli-sycl-b3617-f16 --version
version: 1 (a07c32e)
built with Intel(R) oneAPI DPC++/C++ Compiler 2024.2.1 (2024.2.1.20240711) for x86_64-unknown-linux-gnu

运行模型 llama2-7B.q4, 生成长度 100 (f32):

sh 复制代码
a2@a2s:~$ ./llama-cli-sycl-b3617-f32 -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -ngl 33 -sm none -n 100
Log start
main: build = 1 (a07c32e)
main: built with Intel(R) oneAPI DPC++/C++ Compiler 2024.2.1 (2024.2.1.20240711) for x86_64-unknown-linux-gnu
main: seed  = 1724493845

(此处省略一部分)

llm_load_print_meta: max token length = 48
ggml_sycl_init: GGML_SYCL_FORCE_MMQ:   no
ggml_sycl_init: SYCL_USE_XMX: yes
ggml_sycl_init: found 1 SYCL devices:
llm_load_tensors: ggml ctx size =    0.27 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:      SYCL0 buffer size =  3820.94 MiB
llm_load_tensors:        CPU buffer size =    70.31 MiB
..................................................................................................
llama_new_context_with_model: n_ctx      = 4096
llama_new_context_with_model: n_batch    = 2048
llama_new_context_with_model: n_ubatch   = 512
llama_new_context_with_model: flash_attn = 0
llama_new_context_with_model: freq_base  = 10000.0
llama_new_context_with_model: freq_scale = 1
[SYCL] call ggml_check_sycl
ggml_check_sycl: GGML_SYCL_DEBUG: 0
ggml_check_sycl: GGML_SYCL_F16: no
found 1 SYCL devices:
|  |                   |                                       |       |Max    |        |Max  |Global |                     |
|  |                   |                                       |       |compute|Max work|sub  |mem    |                     |
|ID|        Device Type|                                   Name|Version|units  |group   |group|size   |       Driver version|
|--|-------------------|---------------------------------------|-------|-------|--------|-----|-------|---------------------|
| 0| [level_zero:gpu:0]|                Intel Arc A770 Graphics|    1.3|    512|    1024|   32| 16225M|            1.3.29735|
llama_kv_cache_init:      SYCL0 KV buffer size =  2048.00 MiB
llama_new_context_with_model: KV self size  = 2048.00 MiB, K (f16): 1024.00 MiB, V (f16): 1024.00 MiB
llama_new_context_with_model:  SYCL_Host  output buffer size =     0.12 MiB
llama_new_context_with_model:      SYCL0 compute buffer size =   296.00 MiB
llama_new_context_with_model:  SYCL_Host compute buffer size =    16.01 MiB
llama_new_context_with_model: graph nodes  = 1030
llama_new_context_with_model: graph splits = 2

system_info: n_threads = 4 / 4 | AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | FMA = 1 | NEON = 0 | SVE = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 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 = 4096, n_batch = 2048, n_predict = 100, n_keep = 1


 hello, this is a very very long story, but i promise it's worth it.

(此处省略一部分)

llama_print_timings:        load time =    2066.71 ms
llama_print_timings:      sample time =       2.90 ms /   100 runs   (    0.03 ms per token, 34542.31 tokens per second)
llama_print_timings: prompt eval time =     180.84 ms /    10 tokens (   18.08 ms per token,    55.30 tokens per second)
llama_print_timings:        eval time =    2852.87 ms /    99 runs   (   28.82 ms per token,    34.70 tokens per second)
llama_print_timings:       total time =    3044.63 ms /   109 tokens
Log end

运行模型 llama2-7B.q4, 生成长度 200 (f32):

sh 复制代码
a2@a2s:~$ ./llama-cli-sycl-b3617-f32 -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -ngl 33 -sm none -n 200

(此处省略一部分)

llama_print_timings:        load time =    2040.98 ms
llama_print_timings:      sample time =       5.98 ms /   200 runs   (    0.03 ms per token, 33450.41 tokens per second)
llama_print_timings: prompt eval time =     179.29 ms /    10 tokens (   17.93 ms per token,    55.78 tokens per second)
llama_print_timings:        eval time =    5765.54 ms /   199 runs   (   28.97 ms per token,    34.52 tokens per second)
llama_print_timings:       total time =    5968.10 ms /   209 tokens

运行模型 llama2-7B.q4, 生成长度 500 (f32):

sh 复制代码
a2@a2s:~$ ./llama-cli-sycl-b3617-f32 -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -ngl 33 -sm none -n 500

(此处省略一部分)

llama_print_timings:        load time =    1994.74 ms
llama_print_timings:      sample time =      15.04 ms /   500 runs   (    0.03 ms per token, 33246.89 tokens per second)
llama_print_timings: prompt eval time =     177.09 ms /    10 tokens (   17.71 ms per token,    56.47 tokens per second)
llama_print_timings:        eval time =   14675.46 ms /   499 runs   (   29.41 ms per token,    34.00 tokens per second)
llama_print_timings:       total time =   14911.41 ms /   509 tokens

运行模型 llama2-7B.q4, 生成长度 1000 (f32):

sh 复制代码
a2@a2s:~$ ./llama-cli-sycl-b3617-f32 -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -ngl 33 -sm none -n 1000

(此处省略一部分)

llama_print_timings:        load time =    2071.28 ms
llama_print_timings:      sample time =      28.05 ms /  1000 runs   (    0.03 ms per token, 35646.81 tokens per second)
llama_print_timings: prompt eval time =     178.45 ms /    10 tokens (   17.85 ms per token,    56.04 tokens per second)
llama_print_timings:        eval time =   30044.60 ms /   999 runs   (   30.07 ms per token,    33.25 tokens per second)
llama_print_timings:       total time =   30329.49 ms /  1009 tokens

运行模型 qwen2-7B.q8, 生成长度 100 (f32):

sh 复制代码
a2@a2s:~$ ./llama-cli-sycl-b3617-f32 -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -ngl 33 -sm none -n 100
Log start
main: build = 1 (a07c32e)
main: built with Intel(R) oneAPI DPC++/C++ Compiler 2024.2.1 (2024.2.1.20240711) for x86_64-unknown-linux-gnu
main: seed  = 1724494148
llama_model_loader: loaded meta data with 26 key-value pairs and 339 tensors from qwen2-7b-instruct-q8_0.gguf (version GGUF V3 (latest))

(此处省略一部分)

llm_load_print_meta: max token length = 256
ggml_sycl_init: GGML_SYCL_FORCE_MMQ:   no
ggml_sycl_init: SYCL_USE_XMX: yes
ggml_sycl_init: found 1 SYCL devices:
llm_load_tensors: ggml ctx size =    0.30 MiB
llm_load_tensors: offloading 28 repeating layers to GPU
llm_load_tensors: offloading non-repeating layers to GPU
llm_load_tensors: offloaded 29/29 layers to GPU
llm_load_tensors:      SYCL0 buffer size =  7165.44 MiB
llm_load_tensors:        CPU buffer size =   552.23 MiB
.......................................................................................
llama_new_context_with_model: n_ctx      = 32768
llama_new_context_with_model: n_batch    = 2048
llama_new_context_with_model: n_ubatch   = 512
llama_new_context_with_model: flash_attn = 0
llama_new_context_with_model: freq_base  = 1000000.0
llama_new_context_with_model: freq_scale = 1
[SYCL] call ggml_check_sycl
ggml_check_sycl: GGML_SYCL_DEBUG: 0
ggml_check_sycl: GGML_SYCL_F16: no
found 1 SYCL devices:
|  |                   |                                       |       |Max    |        |Max  |Global |                     |
|  |                   |                                       |       |compute|Max work|sub  |mem    |                     |
|ID|        Device Type|                                   Name|Version|units  |group   |group|size   |       Driver version|
|--|-------------------|---------------------------------------|-------|-------|--------|-----|-------|---------------------|
| 0| [level_zero:gpu:0]|                Intel Arc A770 Graphics|    1.3|    512|    1024|   32| 16225M|            1.3.29735|
llama_kv_cache_init:      SYCL0 KV buffer size =  1792.00 MiB
llama_new_context_with_model: KV self size  = 1792.00 MiB, K (f16):  896.00 MiB, V (f16):  896.00 MiB
llama_new_context_with_model:  SYCL_Host  output buffer size =     0.58 MiB
llama_new_context_with_model:      SYCL0 compute buffer size =  1884.00 MiB
llama_new_context_with_model:  SYCL_Host compute buffer size =    71.01 MiB
llama_new_context_with_model: graph nodes  = 986
llama_new_context_with_model: graph splits = 2

system_info: n_threads = 4 / 4 | AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | FMA = 1 | NEON = 0 | SVE = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 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 = 32768, n_batch = 2048, n_predict = 100, n_keep = 0


hello, this is a very very long story and I would like to ask for help.

(此处省略一部分)

llama_print_timings:        load time =    9055.51 ms
llama_print_timings:      sample time =       8.22 ms /   100 runs   (    0.08 ms per token, 12158.05 tokens per second)
llama_print_timings: prompt eval time =     395.27 ms /     9 tokens (   43.92 ms per token,    22.77 tokens per second)
llama_print_timings:        eval time =    5195.18 ms /    99 runs   (   52.48 ms per token,    19.06 tokens per second)
llama_print_timings:       total time =    5679.84 ms /   108 tokens
Log end

运行模型 qwen2-7B.q8, 生成长度 200 (f32):

sh 复制代码
a2@a2s:~$ ./llama-cli-sycl-b3617-f32 -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -ngl 33 -sm none -n 200

(此处省略一部分)

llama_print_timings:        load time =    8413.38 ms
llama_print_timings:      sample time =      16.47 ms /   200 runs   (    0.08 ms per token, 12141.08 tokens per second)
llama_print_timings: prompt eval time =     405.85 ms /     9 tokens (   45.09 ms per token,    22.18 tokens per second)
llama_print_timings:        eval time =   10455.78 ms /   199 runs   (   52.54 ms per token,    19.03 tokens per second)
llama_print_timings:       total time =   11017.44 ms /   208 tokens

运行模型 qwen2-7B.q8, 生成长度 500 (f32):

sh 复制代码
a2@a2s:~$ ./llama-cli-sycl-b3617-f32 -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -ngl 33 -sm none -n 500

(此处省略一部分)

llama_print_timings:        load time =    9179.45 ms
llama_print_timings:      sample time =      47.42 ms /   500 runs   (    0.09 ms per token, 10544.74 tokens per second)
llama_print_timings: prompt eval time =     402.42 ms /     9 tokens (   44.71 ms per token,    22.36 tokens per second)
llama_print_timings:        eval time =   26367.77 ms /   499 runs   (   52.84 ms per token,    18.92 tokens per second)
llama_print_timings:       total time =   27130.93 ms /   508 tokens

运行模型 qwen2-7B.q8, 生成长度 1000 (f32):

sh 复制代码
a2@a2s:~$ ./llama-cli-sycl-b3617-f32 -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -ngl 33 -sm none -n 1000

(此处省略一部分)

llama_print_timings:        load time =    9531.60 ms
llama_print_timings:      sample time =      96.63 ms /  1000 runs   (    0.10 ms per token, 10348.86 tokens per second)
llama_print_timings: prompt eval time =     401.50 ms /     9 tokens (   44.61 ms per token,    22.42 tokens per second)
llama_print_timings:        eval time =   53212.71 ms /   999 runs   (   53.27 ms per token,    18.77 tokens per second)
llama_print_timings:       total time =   54321.34 ms /  1008 tokens

运行模型 llama2-7B.q4, 生成长度 100 (f16):

sh 复制代码
a2@a2s:~$ ./llama-cli-sycl-b3617-f16 -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -ngl 33 -sm none -n 100
Log start
main: build = 1 (a07c32e)
main: built with Intel(R) oneAPI DPC++/C++ Compiler 2024.2.1 (2024.2.1.20240711) for x86_64-unknown-linux-gnu
main: seed  = 1724494475

(此处省略一部分)

llm_load_print_meta: max token length = 48
ggml_sycl_init: GGML_SYCL_FORCE_MMQ:   no
ggml_sycl_init: SYCL_USE_XMX: yes
ggml_sycl_init: found 1 SYCL devices:
llm_load_tensors: ggml ctx size =    0.27 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:      SYCL0 buffer size =  3820.94 MiB
llm_load_tensors:        CPU buffer size =    70.31 MiB
..................................................................................................
llama_new_context_with_model: n_ctx      = 4096
llama_new_context_with_model: n_batch    = 2048
llama_new_context_with_model: n_ubatch   = 512
llama_new_context_with_model: flash_attn = 0
llama_new_context_with_model: freq_base  = 10000.0
llama_new_context_with_model: freq_scale = 1
[SYCL] call ggml_check_sycl
ggml_check_sycl: GGML_SYCL_DEBUG: 0
ggml_check_sycl: GGML_SYCL_F16: yes
found 1 SYCL devices:
|  |                   |                                       |       |Max    |        |Max  |Global |                     |
|  |                   |                                       |       |compute|Max work|sub  |mem    |                     |
|ID|        Device Type|                                   Name|Version|units  |group   |group|size   |       Driver version|
|--|-------------------|---------------------------------------|-------|-------|--------|-----|-------|---------------------|
| 0| [level_zero:gpu:0]|                Intel Arc A770 Graphics|    1.3|    512|    1024|   32| 16225M|            1.3.29735|
llama_kv_cache_init:      SYCL0 KV buffer size =  2048.00 MiB
llama_new_context_with_model: KV self size  = 2048.00 MiB, K (f16): 1024.00 MiB, V (f16): 1024.00 MiB
llama_new_context_with_model:  SYCL_Host  output buffer size =     0.12 MiB
llama_new_context_with_model:      SYCL0 compute buffer size =   296.00 MiB
llama_new_context_with_model:  SYCL_Host compute buffer size =    16.01 MiB
llama_new_context_with_model: graph nodes  = 1030
llama_new_context_with_model: graph splits = 2

system_info: n_threads = 4 / 4 | AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | FMA = 1 | NEON = 0 | SVE = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 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 = 4096, n_batch = 2048, n_predict = 100, n_keep = 1


 hello, this is a very very long story, and I hope you will read it, because I need to tell you something.

(此处省略一部分)

llama_print_timings:        load time =    1866.40 ms
llama_print_timings:      sample time =       3.23 ms /   100 runs   (    0.03 ms per token, 30998.14 tokens per second)
llama_print_timings: prompt eval time =     187.70 ms /    10 tokens (   18.77 ms per token,    53.28 tokens per second)
llama_print_timings:        eval time =    2873.84 ms /    99 runs   (   29.03 ms per token,    34.45 tokens per second)
llama_print_timings:       total time =    3074.08 ms /   109 tokens
Log end

运行模型 llama2-7B.q4, 生成长度 200 (f16):

sh 复制代码
a2@a2s:~$ ./llama-cli-sycl-b3617-f16 -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -ngl 33 -sm none -n 200

(此处省略一部分)

llama_print_timings:        load time =    1867.46 ms
llama_print_timings:      sample time =       5.99 ms /   200 runs   (    0.03 ms per token, 33411.29 tokens per second)
llama_print_timings: prompt eval time =     194.39 ms /    10 tokens (   19.44 ms per token,    51.44 tokens per second)
llama_print_timings:        eval time =    5783.95 ms /   199 runs   (   29.07 ms per token,    34.41 tokens per second)
llama_print_timings:       total time =    6003.07 ms /   209 tokens

运行模型 llama2-7B.q4, 生成长度 500 (f16):

sh 复制代码
a2@a2s:~$ ./llama-cli-sycl-b3617-f16 -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -ngl 33 -sm none -n 500

(此处省略一部分)

llama_print_timings:        load time =    1909.92 ms
llama_print_timings:      sample time =      15.56 ms /   500 runs   (    0.03 ms per token, 32123.35 tokens per second)
llama_print_timings: prompt eval time =     186.10 ms /    10 tokens (   18.61 ms per token,    53.73 tokens per second)
llama_print_timings:        eval time =   14680.81 ms /   499 runs   (   29.42 ms per token,    33.99 tokens per second)
llama_print_timings:       total time =   14925.64 ms /   509 tokens

运行模型 llama2-7B.q4, 生成长度 1000 (f16):

sh 复制代码
a2@a2s:~$ ./llama-cli-sycl-b3617-f16 -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -ngl 33 -sm none -n 1000

(此处省略一部分)

llama_print_timings:        load time =    2017.43 ms
llama_print_timings:      sample time =      13.53 ms /   461 runs   (    0.03 ms per token, 34067.40 tokens per second)
llama_print_timings: prompt eval time =     189.74 ms /    10 tokens (   18.97 ms per token,    52.70 tokens per second)
llama_print_timings:        eval time =   13480.19 ms /   460 runs   (   29.30 ms per token,    34.12 tokens per second)
llama_print_timings:       total time =   13722.36 ms /   470 tokens

运行模型 qwen2-7B.q8, 生成长度 100 (f16):

sh 复制代码
a2@a2s:~$ ./llama-cli-sycl-b3617-f16 -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -ngl 33 -sm none -n 100
Log start
main: build = 1 (a07c32e)
main: built with Intel(R) oneAPI DPC++/C++ Compiler 2024.2.1 (2024.2.1.20240711) for x86_64-unknown-linux-gnu
main: seed  = 1724494717
llama_model_loader: loaded meta data with 26 key-value pairs and 339 tensors from qwen2-7b-instruct-q8_0.gguf (version GGUF V3 (latest))

(此处省略一部分)

llm_load_print_meta: max token length = 256
ggml_sycl_init: GGML_SYCL_FORCE_MMQ:   no
ggml_sycl_init: SYCL_USE_XMX: yes
ggml_sycl_init: found 1 SYCL devices:
llm_load_tensors: ggml ctx size =    0.30 MiB
llm_load_tensors: offloading 28 repeating layers to GPU
llm_load_tensors: offloading non-repeating layers to GPU
llm_load_tensors: offloaded 29/29 layers to GPU
llm_load_tensors:      SYCL0 buffer size =  7165.44 MiB
llm_load_tensors:        CPU buffer size =   552.23 MiB
.......................................................................................
llama_new_context_with_model: n_ctx      = 32768
llama_new_context_with_model: n_batch    = 2048
llama_new_context_with_model: n_ubatch   = 512
llama_new_context_with_model: flash_attn = 0
llama_new_context_with_model: freq_base  = 1000000.0
llama_new_context_with_model: freq_scale = 1
[SYCL] call ggml_check_sycl
ggml_check_sycl: GGML_SYCL_DEBUG: 0
ggml_check_sycl: GGML_SYCL_F16: yes
found 1 SYCL devices:
|  |                   |                                       |       |Max    |        |Max  |Global |                     |
|  |                   |                                       |       |compute|Max work|sub  |mem    |                     |
|ID|        Device Type|                                   Name|Version|units  |group   |group|size   |       Driver version|
|--|-------------------|---------------------------------------|-------|-------|--------|-----|-------|---------------------|
| 0| [level_zero:gpu:0]|                Intel Arc A770 Graphics|    1.3|    512|    1024|   32| 16225M|            1.3.29735|
llama_kv_cache_init:      SYCL0 KV buffer size =  1792.00 MiB
llama_new_context_with_model: KV self size  = 1792.00 MiB, K (f16):  896.00 MiB, V (f16):  896.00 MiB
llama_new_context_with_model:  SYCL_Host  output buffer size =     0.58 MiB
llama_new_context_with_model:      SYCL0 compute buffer size =  1884.00 MiB
llama_new_context_with_model:  SYCL_Host compute buffer size =    71.01 MiB
llama_new_context_with_model: graph nodes  = 986
llama_new_context_with_model: graph splits = 2

system_info: n_threads = 4 / 4 | AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | FMA = 1 | NEON = 0 | SVE = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 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 = 32768, n_batch = 2048, n_predict = 100, n_keep = 0


hello, this is a very very long story but I will try to make it as short as possible.

(此处省略一部分)

llama_print_timings:        load time =    8893.71 ms
llama_print_timings:      sample time =       9.80 ms /   100 runs   (    0.10 ms per token, 10204.08 tokens per second)
llama_print_timings: prompt eval time =     295.81 ms /     9 tokens (   32.87 ms per token,    30.42 tokens per second)
llama_print_timings:        eval time =    5931.59 ms /    99 runs   (   59.91 ms per token,    16.69 tokens per second)
llama_print_timings:       total time =    6305.29 ms /   108 tokens
Log end

运行模型 qwen2-7B.q8, 生成长度 200 (f16):

sh 复制代码
a2@a2s:~$ ./llama-cli-sycl-b3617-f16 -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -ngl 33 -sm none -n 200

(此处省略一部分)

llama_print_timings:        load time =    8474.98 ms
llama_print_timings:      sample time =      18.22 ms /   200 runs   (    0.09 ms per token, 10978.15 tokens per second)
llama_print_timings: prompt eval time =     298.85 ms /     9 tokens (   33.21 ms per token,    30.12 tokens per second)
llama_print_timings:        eval time =   11935.47 ms /   199 runs   (   59.98 ms per token,    16.67 tokens per second)
llama_print_timings:       total time =   12379.13 ms /   208 tokens

运行模型 qwen2-7B.q8, 生成长度 500 (f16):

sh 复制代码
a2@a2s:~$ ./llama-cli-sycl-b3617-f16 -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -ngl 33 -sm none -n 500

(此处省略一部分)

llama_print_timings:        load time =    8836.66 ms
llama_print_timings:      sample time =      41.76 ms /   500 runs   (    0.08 ms per token, 11972.32 tokens per second)
llama_print_timings: prompt eval time =     304.28 ms /     9 tokens (   33.81 ms per token,    29.58 tokens per second)
llama_print_timings:        eval time =   30052.85 ms /   499 runs   (   60.23 ms per token,    16.60 tokens per second)
llama_print_timings:       total time =   30722.98 ms /   508 tokens

运行模型 qwen2-7B.q8, 生成长度 1000 (f16):

sh 复制代码
a2@a2s:~$ ./llama-cli-sycl-b3617-f16 -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -ngl 33 -sm none -n 1000

(此处省略一部分)

llama_print_timings:        load time =    8206.19 ms
llama_print_timings:      sample time =      96.05 ms /  1000 runs   (    0.10 ms per token, 10411.24 tokens per second)
llama_print_timings: prompt eval time =     312.47 ms /     9 tokens (   34.72 ms per token,    28.80 tokens per second)
llama_print_timings:        eval time =   60716.89 ms /   999 runs   (   60.78 ms per token,    16.45 tokens per second)
llama_print_timings:       total time =   61768.29 ms /  1008 tokens

3.8 Windows (CPU) r5-5600g AVX2

在 6 号 PC (物理机) 上运行. 版本:

sh 复制代码
>.\llama-b3617-bin-win-avx2-x64\llama-cli.exe --version
version: 3617 (a07c32ea)
built with MSVC 19.29.30154.0 for x64

运行模型 llama2-7B.q4, 生成长度 100:

sh 复制代码
p>.\llama-b3617-bin-win-avx2-x64\llama-cli.exe -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -n 100
Log start
main: build = 3617 (a07c32ea)
main: built with MSVC 19.29.30154.0 for x64
main: seed  = 1724480697

llama_print_timings:        load time =    1005.41 ms
llama_print_timings:      sample time =       4.11 ms /   100 runs   (    0.04 ms per token, 24354.60 tokens per second)
llama_print_timings: prompt eval time =     399.08 ms /    10 tokens (   39.91 ms per token,    25.06 tokens per second)
llama_print_timings:        eval time =    9688.39 ms /    99 runs   (   97.86 ms per token,    10.22 tokens per second)
llama_print_timings:       total time =   10110.42 ms /   109 tokens

运行模型 llama2-7B.q4, 生成长度 200:

sh 复制代码
>.\llama-b3617-bin-win-avx2-x64\llama-cli.exe -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -n 200

llama_print_timings:        load time =    1045.93 ms
llama_print_timings:      sample time =       8.82 ms /   200 runs   (    0.04 ms per token, 22673.17 tokens per second)
llama_print_timings: prompt eval time =     436.84 ms /    10 tokens (   43.68 ms per token,    22.89 tokens per second)
llama_print_timings:        eval time =   19960.35 ms /   199 runs   (  100.30 ms per token,     9.97 tokens per second)
llama_print_timings:       total time =   20439.79 ms /   209 tokens

运行模型 llama2-7B.q4, 生成长度 500:

sh 复制代码
>.\llama-b3617-bin-win-avx2-x64\llama-cli.exe -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -n 500

llama_print_timings:        load time =    1028.02 ms
llama_print_timings:      sample time =      18.32 ms /   500 runs   (    0.04 ms per token, 27300.03 tokens per second)
llama_print_timings: prompt eval time =     382.15 ms /    10 tokens (   38.22 ms per token,    26.17 tokens per second)
llama_print_timings:        eval time =   51622.99 ms /   499 runs   (  103.45 ms per token,     9.67 tokens per second)
llama_print_timings:       total time =   52107.10 ms /   509 tokens

运行模型 llama2-7B.q4, 生成长度 1000:

sh 复制代码
>.\llama-b3617-bin-win-avx2-x64\llama-cli.exe -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -n 1000

llama_print_timings:        load time =    1241.78 ms
llama_print_timings:      sample time =      41.52 ms /  1000 runs   (    0.04 ms per token, 24084.78 tokens per second)
llama_print_timings: prompt eval time =     484.10 ms /    10 tokens (   48.41 ms per token,    20.66 tokens per second)
llama_print_timings:        eval time =  114393.05 ms /   999 runs   (  114.51 ms per token,     8.73 tokens per second)
llama_print_timings:       total time =  115084.29 ms /  1009 tokens

运行模型 qwen2-7B.q8, 生成长度 100:

sh 复制代码
>.\llama-b3617-bin-win-avx2-x64\llama-cli.exe -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -n 100

llama_print_timings:        load time =    1429.29 ms
llama_print_timings:      sample time =      15.21 ms /   100 runs   (    0.15 ms per token,  6572.89 tokens per second)
llama_print_timings: prompt eval time =     523.07 ms /     9 tokens (   58.12 ms per token,    17.21 tokens per second)
llama_print_timings:        eval time =   17786.69 ms /    99 runs   (  179.66 ms per token,     5.57 tokens per second)
llama_print_timings:       total time =   18409.82 ms /   108 tokens

运行模型 qwen2-7B.q8, 生成长度 200:

sh 复制代码
>.\llama-b3617-bin-win-avx2-x64\llama-cli.exe -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -n 200

llama_print_timings:        load time =    1424.62 ms
llama_print_timings:      sample time =      31.78 ms /   200 runs   (    0.16 ms per token,  6292.47 tokens per second)
llama_print_timings: prompt eval time =     564.79 ms /     9 tokens (   62.75 ms per token,    15.93 tokens per second)
llama_print_timings:        eval time =   36148.33 ms /   199 runs   (  181.65 ms per token,     5.51 tokens per second)
llama_print_timings:       total time =   36919.37 ms /   208 tokens

运行模型 qwen2-7B.q8, 生成长度 500:

sh 复制代码
>.\llama-b3617-bin-win-avx2-x64\llama-cli.exe -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -n 500

llama_print_timings:        load time =    1462.26 ms
llama_print_timings:      sample time =      80.31 ms /   500 runs   (    0.16 ms per token,  6225.64 tokens per second)
llama_print_timings: prompt eval time =     720.86 ms /     9 tokens (   80.10 ms per token,    12.49 tokens per second)
llama_print_timings:        eval time =   90566.92 ms /   499 runs   (  181.50 ms per token,     5.51 tokens per second)
llama_print_timings:       total time =   91801.55 ms /   508 tokens

运行模型 qwen2-7B.q8, 生成长度 1000:

sh 复制代码
>.\llama-b3617-bin-win-avx2-x64\llama-cli.exe -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -n 1000

llama_print_timings:        load time =    1439.21 ms
llama_print_timings:      sample time =     165.06 ms /  1000 runs   (    0.17 ms per token,  6058.48 tokens per second)
llama_print_timings: prompt eval time =     555.15 ms /     9 tokens (   61.68 ms per token,    16.21 tokens per second)
llama_print_timings:        eval time =  184706.64 ms /   999 runs   (  184.89 ms per token,     5.41 tokens per second)
llama_print_timings:       total time =  186313.82 ms /  1008 tokens

3.9 Windows (GPU) A770 vulkan

在 6 号 PC (物理机) 上运行. 版本:

sh 复制代码
>.\llama-b3617-bin-win-vulkan-x64\llama-cli.exe --version
version: 3617 (a07c32ea)
built with MSVC 19.29.30154.0 for x64

运行模型 llama2-7B.q4, 生成长度 100:

sh 复制代码
>.\llama-b3617-bin-win-vulkan-x64\llama-cli.exe -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -n 100 -ngl 33
Log start
main: build = 3617 (a07c32ea)
main: built with MSVC 19.29.30154.0 for x64
main: seed  = 1724482103

llama_print_timings:        load time =    3375.14 ms
llama_print_timings:      sample time =       4.04 ms /   100 runs   (    0.04 ms per token, 24764.74 tokens per second)
llama_print_timings: prompt eval time =     471.87 ms /    10 tokens (   47.19 ms per token,    21.19 tokens per second)
llama_print_timings:        eval time =    5913.11 ms /    99 runs   (   59.73 ms per token,    16.74 tokens per second)
llama_print_timings:       total time =    6408.49 ms /   109 tokens

运行模型 llama2-7B.q4, 生成长度 200:

sh 复制代码
>.\llama-b3617-bin-win-vulkan-x64\llama-cli.exe -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -n 200 -ngl 33

llama_print_timings:        load time =    2932.55 ms
llama_print_timings:      sample time =       8.03 ms /   200 runs   (    0.04 ms per token, 24915.91 tokens per second)
llama_print_timings: prompt eval time =     471.34 ms /    10 tokens (   47.13 ms per token,    21.22 tokens per second)
llama_print_timings:        eval time =   11931.98 ms /   199 runs   (   59.96 ms per token,    16.68 tokens per second)
llama_print_timings:       total time =   12452.04 ms /   209 tokens

运行模型 llama2-7B.q4, 生成长度 500:

sh 复制代码
>.\llama-b3617-bin-win-vulkan-x64\llama-cli.exe -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -n 500 -ngl 33

llama_print_timings:        load time =    2913.84 ms
llama_print_timings:      sample time =      19.84 ms /   500 runs   (    0.04 ms per token, 25204.15 tokens per second)
llama_print_timings: prompt eval time =     471.64 ms /    10 tokens (   47.16 ms per token,    21.20 tokens per second)
llama_print_timings:        eval time =   30253.41 ms /   499 runs   (   60.63 ms per token,    16.49 tokens per second)
llama_print_timings:       total time =   30844.12 ms /   509 tokens

运行模型 llama2-7B.q4, 生成长度 1000:

sh 复制代码
>.\llama-b3617-bin-win-vulkan-x64\llama-cli.exe -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -n 1000 -ngl 33

llama_print_timings:        load time =    2909.30 ms
llama_print_timings:      sample time =      40.91 ms /  1000 runs   (    0.04 ms per token, 24443.90 tokens per second)
llama_print_timings: prompt eval time =     471.58 ms /    10 tokens (   47.16 ms per token,    21.21 tokens per second)
llama_print_timings:        eval time =   61725.41 ms /   999 runs   (   61.79 ms per token,    16.18 tokens per second)
llama_print_timings:       total time =   62433.39 ms /  1009 tokens

运行模型 qwen2-7B.q8, 生成长度 100:

sh 复制代码
>.\llama-b3617-bin-win-vulkan-x64\llama-cli.exe -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -n 100 -ngl 33

llama_print_timings:        load time =    4785.92 ms
llama_print_timings:      sample time =       9.08 ms /   100 runs   (    0.09 ms per token, 11016.86 tokens per second)
llama_print_timings: prompt eval time =     609.77 ms /     9 tokens (   67.75 ms per token,    14.76 tokens per second)
llama_print_timings:        eval time =    6401.98 ms /    99 runs   (   64.67 ms per token,    15.46 tokens per second)
llama_print_timings:       total time =    7100.18 ms /   108 tokens

运行模型 qwen2-7B.q8, 生成长度 200:

sh 复制代码
>.\llama-b3617-bin-win-vulkan-x64\llama-cli.exe -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -n 200 -ngl 33

llama_print_timings:        load time =    4783.54 ms
llama_print_timings:      sample time =      18.63 ms /   200 runs   (    0.09 ms per token, 10735.37 tokens per second)
llama_print_timings: prompt eval time =     610.60 ms /     9 tokens (   67.84 ms per token,    14.74 tokens per second)
llama_print_timings:        eval time =   12910.01 ms /   199 runs   (   64.87 ms per token,    15.41 tokens per second)
llama_print_timings:       total time =   13698.94 ms /   208 tokens

运行模型 qwen2-7B.q8, 生成长度 500:

sh 复制代码
>.\llama-b3617-bin-win-vulkan-x64\llama-cli.exe -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -n 500 -ngl 33

llama_print_timings:        load time =    4798.07 ms
llama_print_timings:      sample time =      46.32 ms /   500 runs   (    0.09 ms per token, 10794.47 tokens per second)
llama_print_timings: prompt eval time =     610.28 ms /     9 tokens (   67.81 ms per token,    14.75 tokens per second)
llama_print_timings:        eval time =   32517.07 ms /   499 runs   (   65.16 ms per token,    15.35 tokens per second)
llama_print_timings:       total time =   33565.60 ms /   508 tokens

运行模型 qwen2-7B.q8, 生成长度 1000:

sh 复制代码
>.\llama-b3617-bin-win-vulkan-x64\llama-cli.exe -m qwen2-7b-instruct-q8_0.gguf -p "hello, this is a very very long story" -n 1000 -ngl 33

llama_print_timings:        load time =    4802.01 ms
llama_print_timings:      sample time =      93.21 ms /   989 runs   (    0.09 ms per token, 10610.22 tokens per second)
llama_print_timings: prompt eval time =     610.76 ms /     9 tokens (   67.86 ms per token,    14.74 tokens per second)
llama_print_timings:        eval time =   64868.89 ms /   988 runs   (   65.66 ms per token,    15.23 tokens per second)
llama_print_timings:       total time =   66351.20 ms /   997 tokens

(未完待续)

相关推荐
Jurio.几秒前
Conda 管理项目环境
人工智能·python·深度学习·conda·virtualenv·pip
曼城周杰伦12 分钟前
自然语言处理:第六十二章 KAG 超越GraphRAG的图谱框架
人工智能·pytorch·神经网络·自然语言处理·chatgpt·nlp·gpt-3
Donvink15 分钟前
多模态大语言模型——《动手学大模型》实践教程第六章
人工智能·深度学习·语言模型·自然语言处理·llama
Joyner201832 分钟前
pytorch训练的双卡,一个显卡占有20GB,另一个卡占有8GB,怎么均衡?
人工智能·pytorch·python
我爱学Python!33 分钟前
解决复杂查询难题:如何通过 Self-querying Prompting 提高 RAG 系统效率?
人工智能·程序人生·自然语言处理·大模型·llm·大语言模型·rag
AI视觉网奇35 分钟前
pytorch3d linux安装
linux·人工智能·pytorch
OBOO鸥柏44 分钟前
OBOO鸥柏28.6寸液晶广告屏:创新技术引领智能显示新时代
人工智能·科技·大屏端·广告一体机
封步宇AIGC1 小时前
量化交易系统开发-实时行情自动化交易-4.2.1.简单移动平均线实现
人工智能·python·机器学习·数据挖掘
封步宇AIGC1 小时前
量化交易系统开发-实时行情自动化交易-4.1.4.A股布林带(BOLL)实现
人工智能·python·机器学习·数据挖掘
HengCeResearch881 小时前
中国【食品检测实验室自动化】程度相对欧美等发达国家相对落后,并且技术层面存在明显的代差,未来有比较大的发展空间
人工智能·百度·自动化