compute imatrix takes a lot of time

[DavidZyy]

I use 8 A100 to compute imatrix on Meta-Llama-3-8B-Instruct-F16.gguf with file wiki.train.raw, and it takes nearly 6 hours. Is there anything wrong with it? Is this normal? Here is the output when I run it:

user @cics:~/llama.cpp$ ./llama-imatrix -m Meta-Llama-3-8B-Instruct-F16.gguf -f wiki.train.raw -ngl 99
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 8 CUDA devices:
  Device 0: NVIDIA A100-SXM4-40GB, compute capability 8.0, VMM: yes
  Device 1: NVIDIA A100-SXM4-40GB, compute capability 8.0, VMM: yes
  Device 2: NVIDIA A100-SXM4-40GB, compute capability 8.0, VMM: yes
  Device 3: NVIDIA A100-SXM4-40GB, compute capability 8.0, VMM: yes
  Device 4: NVIDIA A100-SXM4-40GB, compute capability 8.0, VMM: yes
  Device 5: NVIDIA A100-SXM4-40GB, compute capability 8.0, VMM: yes
  Device 6: NVIDIA A100-SXM4-40GB, compute capability 8.0, VMM: yes
  Device 7: NVIDIA A100-SXM4-40GB, compute capability 8.0, VMM: yes
build: 4314 (67834d25) with cc (Ubuntu 9.4.0-1ubuntu1~20.04.2) 9.4.0 for x86_64-linux-gnu
llama_load_model_from_file: using device CUDA0 (NVIDIA A100-SXM4-40GB) - 39902 MiB free
llama_load_model_from_file: using device CUDA1 (NVIDIA A100-SXM4-40GB) - 39903 MiB free
llama_load_model_from_file: using device CUDA2 (NVIDIA A100-SXM4-40GB) - 39903 MiB free
llama_load_model_from_file: using device CUDA3 (NVIDIA A100-SXM4-40GB) - 39902 MiB free
llama_load_model_from_file: using device CUDA4 (NVIDIA A100-SXM4-40GB) - 39903 MiB free
llama_load_model_from_file: using device CUDA5 (NVIDIA A100-SXM4-40GB) - 39903 MiB free
llama_load_model_from_file: using device CUDA6 (NVIDIA A100-SXM4-40GB) - 39903 MiB free
llama_load_model_from_file: using device CUDA7 (NVIDIA A100-SXM4-40GB) - 39903 MiB free
llama_model_loader: loaded meta data with 31 key-value pairs and 291 tensors from Meta-Llama-3-8B-Instruct/Meta-Llama-3-8B-Instruct-F16.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv   0:                       general.architecture str              = llama
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = Meta Llama 3 8B Instruct
llama_model_loader: - kv   3:                           general.finetune str              = Instruct
llama_model_loader: - kv   4:                           general.basename str              = Meta-Llama-3
llama_model_loader: - kv   5:                         general.size_label str              = 8B
llama_model_loader: - kv   6:                            general.license str              = other
llama_model_loader: - kv   7:                       general.license.name str              = llama3
llama_model_loader: - kv   8:                       general.license.link str              = LICENSE
llama_model_loader: - kv   9:                               general.tags arr[str,6]       = ["facebook", "meta", "pytorch", "llam...
llama_model_loader: - kv  10:                          general.languages arr[str,1]       = ["en"]
llama_model_loader: - kv  11:                          llama.block_count u32              = 32
llama_model_loader: - kv  12:                       llama.context_length u32              = 8192
llama_model_loader: - kv  13:                     llama.embedding_length u32              = 4096
llama_model_loader: - kv  14:                  llama.feed_forward_length u32              = 14336
llama_model_loader: - kv  15:                 llama.attention.head_count u32              = 32
llama_model_loader: - kv  16:              llama.attention.head_count_kv u32              = 8
llama_model_loader: - kv  17:                       llama.rope.freq_base f32              = 500000.000000
llama_model_loader: - kv  18:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  19:                          general.file_type u32              = 1
llama_model_loader: - kv  20:                           llama.vocab_size u32              = 128256
llama_model_loader: - kv  21:                 llama.rope.dimension_count u32              = 128
llama_model_loader: - kv  22:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  23:                         tokenizer.ggml.pre str              = llama-bpe
llama_model_loader: - kv  24:                      tokenizer.ggml.tokens arr[str,128256]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  25:                  tokenizer.ggml.token_type arr[i32,128256]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  26:                      tokenizer.ggml.merges arr[str,280147]  = ["Ġ Ġ", "Ġ ĠĠĠ", "ĠĠ ĠĠ", "...
llama_model_loader: - kv  27:                tokenizer.ggml.bos_token_id u32              = 128000
llama_model_loader: - kv  28:                tokenizer.ggml.eos_token_id u32              = 128009
llama_model_loader: - kv  29:                    tokenizer.chat_template str              = {% set loop_messages = messages %}{% ...
llama_model_loader: - kv  30:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:   65 tensors
llama_model_loader: - type  f16:  226 tensors
llm_load_vocab: special tokens cache size = 256
llm_load_vocab: token to piece cache size = 0.8000 MB
llm_load_print_meta: format           = GGUF V3 (latest)
llm_load_print_meta: arch             = llama
llm_load_print_meta: vocab type       = BPE
llm_load_print_meta: n_vocab          = 128256
llm_load_print_meta: n_merges         = 280147
llm_load_print_meta: vocab_only       = 0
llm_load_print_meta: n_ctx_train      = 8192
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        = 8
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            = 4
llm_load_print_meta: n_embd_k_gqa     = 1024
llm_load_print_meta: n_embd_v_gqa     = 1024
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             = 14336
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  = 500000.0
llm_load_print_meta: freq_scale_train = 1
llm_load_print_meta: n_ctx_orig_yarn  = 8192
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       = 8B
llm_load_print_meta: model ftype      = F16
llm_load_print_meta: model params     = 8.03 B
llm_load_print_meta: model size       = 14.96 GiB (16.00 BPW) 
llm_load_print_meta: general.name     = Meta Llama 3 8B Instruct
llm_load_print_meta: BOS token        = 128000 '<|begin_of_text|>'
llm_load_print_meta: EOS token        = 128009 '<|eot_id|>'
llm_load_print_meta: EOT token        = 128009 '<|eot_id|>'
llm_load_print_meta: LF token         = 128 'Ä'
llm_load_print_meta: EOG token        = 128009 '<|eot_id|>'
llm_load_print_meta: max token length = 256
llm_load_tensors: offloading 32 repeating layers to GPU
llm_load_tensors: offloading output layer to GPU
llm_load_tensors: offloaded 33/33 layers to GPU
llm_load_tensors:   CPU_Mapped model buffer size =  1002.00 MiB
llm_load_tensors:        CUDA0 model buffer size =  2080.16 MiB
llm_load_tensors:        CUDA1 model buffer size =  1664.12 MiB
llm_load_tensors:        CUDA2 model buffer size =  1664.12 MiB
llm_load_tensors:        CUDA3 model buffer size =  1664.12 MiB
llm_load_tensors:        CUDA4 model buffer size =  1664.12 MiB
llm_load_tensors:        CUDA5 model buffer size =  1664.12 MiB
llm_load_tensors:        CUDA6 model buffer size =  1664.12 MiB
llm_load_tensors:        CUDA7 model buffer size =  2250.11 MiB
..........................................................................................
llama_new_context_with_model: n_seq_max     = 1
llama_new_context_with_model: n_ctx         = 512
llama_new_context_with_model: n_ctx_per_seq = 512
llama_new_context_with_model: n_batch       = 512
llama_new_context_with_model: n_ubatch      = 512
llama_new_context_with_model: flash_attn    = 0
llama_new_context_with_model: freq_base     = 500000.0
llama_new_context_with_model: freq_scale    = 1
llama_new_context_with_model: n_ctx_per_seq (512) < n_ctx_train (8192) -- the full capacity of the model will not be utilized
llama_kv_cache_init:      CUDA0 KV buffer size =    10.00 MiB
llama_kv_cache_init:      CUDA1 KV buffer size =     8.00 MiB
llama_kv_cache_init:      CUDA2 KV buffer size =     8.00 MiB
llama_kv_cache_init:      CUDA3 KV buffer size =     8.00 MiB
llama_kv_cache_init:      CUDA4 KV buffer size =     8.00 MiB
llama_kv_cache_init:      CUDA5 KV buffer size =     8.00 MiB
llama_kv_cache_init:      CUDA6 KV buffer size =     8.00 MiB
llama_kv_cache_init:      CUDA7 KV buffer size =     6.00 MiB
llama_new_context_with_model: KV self size  =   64.00 MiB, K (f16):   32.00 MiB, V (f16):   32.00 MiB
llama_new_context_with_model:  CUDA_Host  output buffer size =     0.49 MiB
llama_new_context_with_model: pipeline parallelism enabled (n_copies=4)
llama_new_context_with_model:      CUDA0 compute buffer size =   108.01 MiB
llama_new_context_with_model:      CUDA1 compute buffer size =   108.01 MiB
llama_new_context_with_model:      CUDA2 compute buffer size =   108.01 MiB
llama_new_context_with_model:      CUDA3 compute buffer size =   108.01 MiB
llama_new_context_with_model:      CUDA4 compute buffer size =   108.01 MiB
llama_new_context_with_model:      CUDA5 compute buffer size =   108.01 MiB
llama_new_context_with_model:      CUDA6 compute buffer size =   108.01 MiB
llama_new_context_with_model:      CUDA7 compute buffer size =   294.52 MiB
llama_new_context_with_model:  CUDA_Host compute buffer size =    12.02 MiB
llama_new_context_with_model: graph nodes  = 1030
llama_new_context_with_model: graph splits = 9

system_info: n_threads = 128 (n_threads_batch = 128) / 256 | CUDA : ARCHS = 800 | USE_GRAPHS = 1 | PEER_MAX_BATCH_SIZE = 128 | CPU : LLAMAFILE = 1 | OPENMP = 1 | AARCH64_REPACK = 1 | 
compute_imatrix: tokenizing the input ..
compute_imatrix: tokenization took 2471.01 ms
compute_imatrix: computing over 4758 chunks with batch_size 512
compute_imatrix: 4.48 seconds per pass - ETA 5 hours 55.38 minutes

[DavidZyy]

I use the following command to compile it:

make LLAMA_MAKEFILE=1 GGML_CUDA=1 -j 8 llama-imatrix

And I noticed params.n_ctx and parms.n_batch are set to 512 in imatrix.cpp, and each GPU only uses a small amount of vram.

[ExtReMLapin]

NGL is already on, maybe disable tensor gpu split using CUDA_VISIBLE_DEVICES=0 or --split-mode none

Also it seems that this dataset might be huge, I personally use the one from there , not sure if it's any good

[DavidZyy]

Thanks for your advice! I will try it.