llama-server : Prompt eval very slow / Kimi-K2 and gpt-oss 120B

[YannFollet]

the prompt eval is the same speed of eval
during the prompt eval the GPU is 30% usage and only 1 cpu is used

tell me if there are better parameters or if there is some more test to do

thanks again for the great work on llama.cpp of everyone

./build/bin/llama-server -m ./models/Kimi-K2-Instruct-Q2_K-00001-of-00008.gguf -c 131072 -ngl 99 --host 0.0.0.0 --port 8080 --prio 2 --no-webui --temp 0.6 --top-k 20 --top-p 0.95 --min-p 0 -ot "blk\.([4-9]|[1-9][0-9])\.ffn_.*_exps\.weight=CPU" -fa --jinja

ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 CUDA devices:
  Device 0: NVIDIA GeForce RTX 5090, compute capability 12.0, VMM: yes
build: 6107 (36d3f00e) with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
system info: n_threads = 56, n_threads_batch = 56, total_threads = 112

system_info: n_threads = 56 (n_threads_batch = 56) / 112 | CUDA : ARCHS = 500,610,700,750,800,860 | USE_GRAPHS = 1
| PEER_MAX_BATCH_SIZE = 128 | CPU : SSE3 = 1 | SSSE3 = 1 
| AVX = 1 | AVX_VNNI = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1
| AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 
| AMX_INT8 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 |

Web UI is disabled
main: binding port with default address family
main: HTTP server is listening, hostname: 0.0.0.0, port: 8080, http threads: 111
main: loading model
srv    load_model: loading model '/home/yann/dev/models/Kimi-K2-Instruct-Q2_K-00001-of-00008.gguf'
llama_model_load_from_file_impl: using device CUDA0 (NVIDIA GeForce RTX 5090) - 31587 MiB free
llama_model_loader: additional 7 GGUFs metadata loaded.
llama_model_loader: loaded meta data with 61 key-value pairs and 1096 tensors from /home/yann/dev/models/Kimi-K2-Instruct-Q2_K-00001-of-00008.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              = deepseek2
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = Kimi-K2-Instruct
llama_model_loader: - kv   3:                           general.finetune str              = Instruct
llama_model_loader: - kv   4:                           general.basename str              = Kimi-K2-Instruct
llama_model_loader: - kv   5:                       general.quantized_by str              = Unsloth
llama_model_loader: - kv   6:                         general.size_label str              = 384x14B
llama_model_loader: - kv   7:                            general.license str              = other
llama_model_loader: - kv   8:                       general.license.name str              = modified-mit
llama_model_loader: - kv   9:                           general.repo_url str              = https://huggingface.co/unsloth
llama_model_loader: - kv  10:                   general.base_model.count u32              = 1
llama_model_loader: - kv  11:                  general.base_model.0.name str              = Kimi K2 Instruct
llama_model_loader: - kv  12:          general.base_model.0.organization str              = Moonshotai
llama_model_loader: - kv  13:              general.base_model.0.repo_url str              = https://huggingface.co/moonshotai/Kim...
llama_model_loader: - kv  14:                               general.tags arr[str,1]       = ["unsloth"]
llama_model_loader: - kv  15:                      deepseek2.block_count u32              = 61
llama_model_loader: - kv  16:                   deepseek2.context_length u32              = 131072
llama_model_loader: - kv  17:                 deepseek2.embedding_length u32              = 7168
llama_model_loader: - kv  18:              deepseek2.feed_forward_length u32              = 18432
llama_model_loader: - kv  19:             deepseek2.attention.head_count u32              = 64
llama_model_loader: - kv  20:          deepseek2.attention.head_count_kv u32              = 1
llama_model_loader: - kv  21:                   deepseek2.rope.freq_base f32              = 50000.000000
llama_model_loader: - kv  22: deepseek2.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  23:                deepseek2.expert_used_count u32              = 8
llama_model_loader: - kv  24:        deepseek2.leading_dense_block_count u32              = 1
llama_model_loader: - kv  25:                       deepseek2.vocab_size u32              = 163840
llama_model_loader: - kv  26:            deepseek2.attention.q_lora_rank u32              = 1536
llama_model_loader: - kv  27:           deepseek2.attention.kv_lora_rank u32              = 512
llama_model_loader: - kv  28:             deepseek2.attention.key_length u32              = 576
llama_model_loader: - kv  29:           deepseek2.attention.value_length u32              = 512
llama_model_loader: - kv  30:         deepseek2.attention.key_length_mla u32              = 192
llama_model_loader: - kv  31:       deepseek2.attention.value_length_mla u32              = 128
llama_model_loader: - kv  32:       deepseek2.expert_feed_forward_length u32              = 2048
llama_model_loader: - kv  33:                     deepseek2.expert_count u32              = 384
llama_model_loader: - kv  34:              deepseek2.expert_shared_count u32              = 1
llama_model_loader: - kv  35:             deepseek2.expert_weights_scale f32              = 2.827000
llama_model_loader: - kv  36:              deepseek2.expert_weights_norm bool             = true
llama_model_loader: - kv  37:               deepseek2.expert_gating_func u32              = 2
llama_model_loader: - kv  38:             deepseek2.rope.dimension_count u32              = 64
llama_model_loader: - kv  39:                deepseek2.rope.scaling.type str              = yarn
llama_model_loader: - kv  40:              deepseek2.rope.scaling.factor f32              = 32.000000
llama_model_loader: - kv  41: deepseek2.rope.scaling.original_context_length u32              = 4096
llama_model_loader: - kv  42: deepseek2.rope.scaling.yarn_log_multiplier f32              = 0.100000
llama_model_loader: - kv  43:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  44:                         tokenizer.ggml.pre str              = kimi-k2
llama_model_loader: - kv  45:                      tokenizer.ggml.tokens arr[str,163840]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  46:                  tokenizer.ggml.token_type arr[i32,163840]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  47:                      tokenizer.ggml.merges arr[str,163328]  = ["Ġ Ġ", "ĠĠ ĠĠ", "Ġ t", "i n",...
llama_model_loader: - kv  48:                tokenizer.ggml.bos_token_id u32              = 163584
llama_model_loader: - kv  49:                tokenizer.ggml.eos_token_id u32              = 163586
llama_model_loader: - kv  50:            tokenizer.ggml.padding_token_id u32              = 163839
llama_model_loader: - kv  51:                    tokenizer.chat_template str              = {%- if tools -%}\n  <|im_system|>tool_...
llama_model_loader: - kv  52:               general.quantization_version u32              = 2
llama_model_loader: - kv  53:                          general.file_type u32              = 10
llama_model_loader: - kv  54:                      quantize.imatrix.file str              = Kimi-K2-Instruct-GGUF/imatrix_unsloth...
llama_model_loader: - kv  55:                   quantize.imatrix.dataset str              = unsloth_calibration_Kimi-K2-Instruct-...
llama_model_loader: - kv  56:             quantize.imatrix.entries_count u32              = 667
llama_model_loader: - kv  57:              quantize.imatrix.chunks_count u32              = 704
llama_model_loader: - kv  58:                                   split.no u16              = 0
llama_model_loader: - kv  59:                        split.tensors.count i32              = 1096
llama_model_loader: - kv  60:                                split.count u16              = 8
llama_model_loader: - type  f32:  365 tensors
llama_model_loader: - type q8_0:  122 tensors
llama_model_loader: - type q2_K:  425 tensors
llama_model_loader: - type q3_K:  182 tensors
llama_model_loader: - type q4_K:    1 tensors
llama_model_loader: - type q6_K:    1 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q2_K - Medium
print_info: file size   = 347.80 GiB (2.91 BPW)
load: printing all EOG tokens:
load:   - 163586 ('<|im_end|>')
load: special tokens cache size = 256
load: token to piece cache size = 1.0607 MB
print_info: arch             = deepseek2
print_info: vocab_only       = 0
print_info: n_ctx_train      = 131072
print_info: n_embd           = 7168
print_info: n_layer          = 61
print_info: n_head           = 64
print_info: n_head_kv        = 1
print_info: n_rot            = 64
print_info: n_swa            = 0
print_info: is_swa_any       = 0
print_info: n_embd_head_k    = 576
print_info: n_embd_head_v    = 512
print_info: n_gqa            = 64
print_info: n_embd_k_gqa     = 576
print_info: n_embd_v_gqa     = 512
print_info: f_norm_eps       = 0.0e+00
print_info: f_norm_rms_eps   = 1.0e-06
print_info: f_clamp_kqv      = 0.0e+00
print_info: f_max_alibi_bias = 0.0e+00
print_info: f_logit_scale    = 0.0e+00
print_info: f_attn_scale     = 0.0e+00
print_info: n_ff             = 18432
print_info: n_expert         = 384
print_info: n_expert_used    = 8
print_info: causal attn      = 1
print_info: pooling type     = 0
print_info: rope type        = 0
print_info: rope scaling     = yarn
print_info: freq_base_train  = 50000.0
print_info: freq_scale_train = 0.03125
print_info: n_ctx_orig_yarn  = 4096
print_info: rope_finetuned   = unknown
print_info: model type       = 671B
print_info: model params     = 1.03 T
print_info: general.name     = Kimi-K2-Instruct
print_info: n_layer_dense_lead   = 1
print_info: n_lora_q             = 1536
print_info: n_lora_kv            = 512
print_info: n_embd_head_k_mla    = 192
print_info: n_embd_head_v_mla    = 128
print_info: n_ff_exp             = 2048
print_info: n_expert_shared      = 1
print_info: expert_weights_scale = 2.8
print_info: expert_weights_norm  = 1
print_info: expert_gating_func   = sigmoid
print_info: rope_yarn_log_mul    = 0.1000
print_info: vocab type       = BPE
print_info: n_vocab          = 163840
print_info: n_merges         = 163328
print_info: BOS token        = 163584 '[BOS]'
print_info: EOS token        = 163586 '<|im_end|>'
print_info: EOT token        = 163586 '<|im_end|>'
print_info: PAD token        = 163839 '[PAD]'
print_info: LF token         = 198 'Ċ'
print_info: EOG token        = 163586 '<|im_end|>'
print_info: max token length = 512
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: offloading 61 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 62/62 layers to GPU
load_tensors:        CUDA0 model buffer size =  5237.29 MiB
load_tensors:   CPU_Mapped model buffer size = 46282.17 MiB
load_tensors:   CPU_Mapped model buffer size = 47240.66 MiB
load_tensors:   CPU_Mapped model buffer size = 47240.66 MiB
load_tensors:   CPU_Mapped model buffer size = 47240.66 MiB
load_tensors:   CPU_Mapped model buffer size = 47240.66 MiB
load_tensors:   CPU_Mapped model buffer size = 47240.66 MiB
load_tensors:   CPU_Mapped model buffer size = 47240.66 MiB
load_tensors:   CPU_Mapped model buffer size = 25391.89 MiB
....................................................................................................
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 131072
llama_context: n_ctx_per_seq = 131072
llama_context: n_batch       = 2048
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = 1
llama_context: kv_unified    = false
llama_context: freq_base     = 50000.0
llama_context: freq_scale    = 0.03125
llama_context:  CUDA_Host  output buffer size =     0.62 MiB
llama_kv_cache_unified:      CUDA0 KV buffer size = 16592.00 MiB
llama_kv_cache_unified: size = 16592.00 MiB (131072 cells,  61 layers,  1/1 seqs), K (f16): 8784.00 MiB, V (f16): 7808.00 MiB
llama_context:      CUDA0 compute buffer size =  2893.01 MiB
llama_context:  CUDA_Host compute buffer size =   270.01 MiB
llama_context: graph nodes  = 4853
llama_context: graph splits = 182 (with bs=512), 122 (with bs=1)
common_init_from_params: added <|im_end|> logit bias = -inf
common_init_from_params: setting dry_penalty_last_n to ctx_size = 131072
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
srv          init: initializing slots, n_slots = 1
slot         init: id  0 | task -1 | new slot n_ctx_slot = 131072
main: model loaded
main: chat template, chat_template: {%- if tools -%}
  <|im_system|>tool_declare<|im_middle|>{{ tools | tojson }}<|im_end|>
{%- endif -%}
{%- for message in messages -%}
  {%- if loop.first and messages[0]['role'] != 'system' -%}
    <|im_system|>system<|im_middle|>You are Kimi, an AI assistant created by Moonshot AI.<|im_end|>
  {%- endif -%}
  {%- if message['role'] == 'system' -%}
    <|im_system|>system<|im_middle|>
  {%- elif message['role'] == 'user' -%}
    <|im_user|>user<|im_middle|>
  {%- elif message['role'] == 'assistant' -%}
    <|im_assistant|>assistant<|im_middle|>
  {%- elif message['role'] == 'tool' -%}
    <|im_system|>tool<|im_middle|>
  {%- endif -%}
  {%- if message['role'] == 'assistant' and message.get('tool_calls') -%}
    {%- if message['content'] -%}{{ message['content'] }}{%- endif -%}
    <|tool_calls_section_begin|>
    {%- for tool_call in message['tool_calls'] -%}
      {%- set formatted_id = tool_call['id'] -%}
      <|tool_call_begin|>{{ formatted_id }}<|tool_call_argument_begin|>{% if tool_call['function']['arguments'] is string %}{{ tool_call['function']['arguments'] }}{% else %}{{ tool_call['function']['arguments'] | tojson }}{% endif %}<|tool_call_end|>
    {%- endfor -%}
    <|tool_calls_section_end|>
  {%- elif message['role'] == 'tool' -%}
    ## Return of {{ message.tool_call_id }}
    {{ message['content'] }}
  {%- elif message['content'] is string -%}
    {{ message['content'] }}
  {%- elif message['content'] is not none -%}
    {% for content in message['content'] -%}
      {% if content['type'] == 'image' or 'image' in content or 'image_url' in content -%}
        <|media_start|>image<|media_content|><|media_pad|><|media_end|>
      {% else -%}
        {{ content['text'] }}
      {%- endif -%}
    {%- endfor -%}
  {%- endif -%}
  <|im_end|>
{%- endfor -%}
{%- if add_generation_prompt -%}
  <|im_assistant|>assistant<|im_middle|>
{%- endif -%}, example_format: '<|im_system|>system<|im_middle|>You are a helpful assistant<|im_end|><|im_user|>user<|im_middle|>Hello<|im_end|><|im_assistant|>assistant<|im_middle|>Hi there<|im_end|><|im_user|>user<|im_middle|>How are you?<|im_end|><|im_assistant|>assistant<|im_middle|>'
main: server is listening on http://0.0.0.0:8080 - starting the main loop
srv  update_slots: all slots are idle
srv  params_from_: Chat format: Content-only
slot launch_slot_: id  0 | task 0 | processing task
slot update_slots: id  0 | task 0 | new prompt, n_ctx_slot = 131072, n_keep = 0, n_prompt_tokens = 54
slot update_slots: id  0 | task 0 | kv cache rm [0, end)
slot update_slots: id  0 | task 0 | prompt processing progress, n_past = 54, n_tokens = 54, progress = 1.000000
slot update_slots: id  0 | task 0 | prompt done, n_past = 54, n_tokens = 54
slot      release: id  0 | task 0 | stop processing: n_past = 63, truncated = 0
slot print_timing: id  0 | task 0 |
prompt eval time =   70271.80 ms /    54 tokens ( 1301.33 ms per token,     0.77 tokens per second)
       eval time =     863.00 ms /    10 tokens (   86.30 ms per token,    11.59 tokens per second)
      total time =   71134.80 ms /    64 tokens
srv  update_slots: all slots are idle
srv  log_server_r: request: POST /v1/chat/completions 192.168.26.23 200
srv  params_from_: Chat format: Content-only
slot launch_slot_: id  0 | task 11 | processing task
slot update_slots: id  0 | task 11 | new prompt, n_ctx_slot = 131072, n_keep = 0, n_prompt_tokens = 88
slot update_slots: id  0 | task 11 | kv cache rm [63, end)
slot update_slots: id  0 | task 11 | prompt processing progress, n_past = 88, n_tokens = 25, progress = 0.284091
slot update_slots: id  0 | task 11 | prompt done, n_past = 88, n_tokens = 25
slot      release: id  0 | task 11 | stop processing: n_past = 90, truncated = 0
slot print_timing: id  0 | task 11 |
prompt eval time =     917.39 ms /    25 tokens (   36.70 ms per token,    27.25 tokens per second)
       eval time =     193.12 ms /     3 tokens (   64.37 ms per token,    15.53 tokens per second)
      total time =    1110.51 ms /    28 tokens
srv  update_slots: all slots are idle
srv  log_server_r: request: POST /v1/chat/completions 192.168.26.23 200
srv  params_from_: Chat format: Content-only
slot launch_slot_: id  0 | task 15 | processing task
slot update_slots: id  0 | task 15 | new prompt, n_ctx_slot = 131072, n_keep = 0, n_prompt_tokens = 1743
slot update_slots: id  0 | task 15 | kv cache rm [90, end)
slot update_slots: id  0 | task 15 | prompt processing progress, n_past = 1743, n_tokens = 1653, progress = 0.948365
slot update_slots: id  0 | task 15 | prompt done, n_past = 1743, n_tokens = 1653
slot      release: id  0 | task 15 | stop processing: n_past = 3212, truncated = 0
slot print_timing: id  0 | task 15 |
prompt eval time =  162988.04 ms /  1653 tokens (   98.60 ms per token,    10.14 tokens per second)
       eval time =  144322.12 ms /  1470 tokens (   98.18 ms per token,    10.19 tokens per second)
      total time =  307310.16 ms /  3123 tokens
srv  update_slots: all slots are idle
srv  log_server_r: request: POST /v1/chat/completions 192.168.26.23 200

[YannFollet]

when I laungh in verbose mode I got this message I don't understand :
load_tensors: tensor 'token_embd.weight' (q4_K) (and 171 others) cannot be used with preferred buffer type CUDA_Host, using CPU instead