How to achieve faster prompt eval time with a single GPU: NVIDIA A100-PCIE-40GB?

[Dominik7131]

Right now I'm getting these results with Mixtral-8x7b-instruct-v0.1.Q5_K_M.gguf:

• prompt eval time = ~1.3 seconds / 600 tokens
• prompt eval time = ~6.0 seconds / 3900 tokens

In general, my goal is to make an interactive application where the user uploads a document and then asks questions. This document usually contains at most 5000 tokens. With RAG I'm automatically finding the most relevant sentences in this document based on what the user wants therefore I cannot cache it beforehand.

Is there a way to achieve faster prompt eval time (faster time to first token) so the user doesn't have to wait a few seconds before he starts seeing some output? I tried Exllamav2 with Mixtral-8x7B-instruct-exl2-6.0bpw and achieved this:

• prompt eval time = ~0.7 seconds / 600 tokens
• prompt eval time = ~2.0 seconds / 3900 tokens

However, the quality of the output is worse than with Mixtral-8x7b-instruct-v0.1.Q5_K_M.gguf (~5.52 bpw) and I don't have VRAM for a better version of Mixtral so I want to stick with Llama-cpp (or is there some better option for my case?).

These are my settings:

~/llama.cpp$ ./main -m /data/LLMs/models--TheBloke--Mixtral-8x7B-Instruct-v0.1-GGUF/mixtral-8x7b-instruct-v0.1.Q5_K_M.gguf -i -b 8192 -ub 4096 -ngl 100 -c 8192 --temp 0 --repeat_penalty 1 -p "$(cat prompts/prompt.txt)"
Log start
main: build = 2603 (7a2c9263)
main: built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
main: seed  = 1712230636
llama_model_loader: loaded meta data with 26 key-value pairs and 995 tensors from /data/LLMs/models--TheBloke--Mixtral-8x7B-Instruct-v0.1-GGUF/mixtral-8x7b-instruct-v0.1.Q5_K_M.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv   0:                       general.architecture str              = llama
llama_model_loader: - kv   1:                               general.name str              = mistralai_mixtral-8x7b-instruct-v0.1
llama_model_loader: - kv   2:                       llama.context_length u32              = 32768
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              = 14336
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              = 8
llama_model_loader: - kv   9:                         llama.expert_count u32              = 8
llama_model_loader: - kv  10:                    llama.expert_used_count u32              = 2
llama_model_loader: - kv  11:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  12:                       llama.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  13:                          general.file_type u32              = 17
llama_model_loader: - kv  14:                       tokenizer.ggml.model str              = llama
llama_model_loader: - kv  15:                      tokenizer.ggml.tokens arr[str,32000]   = ["<unk>", "<s>", "</s>", "<0x00>", "<...
llama_model_loader: - kv  16:                      tokenizer.ggml.scores arr[f32,32000]   = [0.000000, 0.000000, 0.000000, 0.0000...
llama_model_loader: - kv  17:                  tokenizer.ggml.token_type arr[i32,32000]   = [2, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 6, ...
llama_model_loader: - kv  18:                tokenizer.ggml.bos_token_id u32              = 1
llama_model_loader: - kv  19:                tokenizer.ggml.eos_token_id u32              = 2
llama_model_loader: - kv  20:            tokenizer.ggml.unknown_token_id u32              = 0
llama_model_loader: - kv  21:            tokenizer.ggml.padding_token_id u32              = 0
llama_model_loader: - kv  22:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  23:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  24:                    tokenizer.chat_template str              = {{ bos_token }}{% for message in mess...
llama_model_loader: - kv  25:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:   65 tensors
llama_model_loader: - type  f16:   32 tensors
llama_model_loader: - type q8_0:   64 tensors
llama_model_loader: - type q5_K:  833 tensors
llama_model_loader: - type q6_K:    1 tensors
llm_load_vocab: special tokens definition check successful ( 259/32000 ).
llm_load_print_meta: format           = GGUF V3 (latest)
llm_load_print_meta: arch             = llama
llm_load_print_meta: vocab type       = SPM
llm_load_print_meta: n_vocab          = 32000
llm_load_print_meta: n_merges         = 0
llm_load_print_meta: n_ctx_train      = 32768
llm_load_print_meta: n_embd           = 4096
llm_load_print_meta: n_head           = 32
llm_load_print_meta: n_head_kv        = 8
llm_load_print_meta: n_layer          = 32
llm_load_print_meta: n_rot            = 128
llm_load_print_meta: n_embd_head_k    = 128
llm_load_print_meta: n_embd_head_v    = 128
llm_load_print_meta: n_gqa            = 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         = 8
llm_load_print_meta: n_expert_used    = 2
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  = 1000000.0
llm_load_print_meta: freq_scale_train = 1
llm_load_print_meta: n_yarn_orig_ctx  = 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: model type       = 7B
llm_load_print_meta: model ftype      = Q5_K - Medium
llm_load_print_meta: model params     = 46.70 B
llm_load_print_meta: model size       = 30.02 GiB (5.52 BPW) 
llm_load_print_meta: general.name     = mistralai_mixtral-8x7b-instruct-v0.1
llm_load_print_meta: BOS token        = 1 '<s>'
llm_load_print_meta: EOS token        = 2 '</s>'
llm_load_print_meta: UNK token        = 0 '<unk>'
llm_load_print_meta: PAD token        = 0 '<unk>'
llm_load_print_meta: LF token         = 13 '<0x0A>'
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:   no
ggml_cuda_init: CUDA_USE_TENSOR_CORES: yes
ggml_cuda_init: found 1 CUDA devices:
  Device 0: NVIDIA A100-PCIE-40GB, compute capability 8.0, VMM: yes
llm_load_tensors: ggml ctx size =    0.96 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:  CUDA_Host buffer size =    85.94 MiB
llm_load_tensors:      CUDA0 buffer size = 30649.55 MiB
....................................................................................................
llama_new_context_with_model: n_ctx      = 8192
llama_new_context_with_model: n_batch    = 8192
llama_new_context_with_model: n_ubatch   = 4096
llama_new_context_with_model: freq_base  = 1000000.0
llama_new_context_with_model: freq_scale = 1
llama_kv_cache_init:      CUDA0 KV buffer size =  1024.00 MiB
llama_new_context_with_model: KV self size  = 1024.00 MiB, K (f16):  512.00 MiB, V (f16):  512.00 MiB
llama_new_context_with_model:  CUDA_Host  output buffer size =     0.12 MiB
llama_new_context_with_model:      CUDA0 compute buffer size =  4640.28 MiB
llama_new_context_with_model:  CUDA_Host compute buffer size =   192.05 MiB
llama_new_context_with_model: graph nodes  = 1638
llama_new_context_with_model: graph splits = 2

system_info: n_threads = 4 / 4 | AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 1 | AVX512_VBMI = 0 | AVX512_VNNI = 1 | FMA = 1 | NEON = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | 
main: interactive mode on.
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 = 0.000
        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 = 8192, n_batch = 8192, n_predict = -1, n_keep = 1

[ggerganov]

#6505 seems related

[Dominik7131]

Thanks! After downloading commit from #6505 these are my new measurements with Mixtral-8x7b-instruct-v0.1.Q5_K_M.gguf:

• prompt eval time = ~ 0.4 seconds / 600 tokens (3.25x speedup)
• prompt eval time = ~1.5 seconds / 3900 tokens (4.00x speedup)

Sounds too good to be true. 👍