Misc. bug: LLAMA-SERVER is 40% slower than LLAMA-CLI when using identical parameters including -ot option for tensor offloading

[joesixpaq]

Name and Version

ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 2 CUDA devices:
Device 0: NVIDIA GeForce RTX 3090, compute capability 8.6, VMM: yes
Device 1: Quadro M2000, compute capability 5.2, VMM: yes
version: 5614 (8f47e25)
built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu

Operating systems

Linux

Which llama.cpp modules do you know to be affected?

llama-server

Command line

CUDA_VISIBLE_DEVICES="0," \
numactl --physcpubind="8,10,12,14, 24,26,28,30, 9,11,13,15, 25,27,29,31" --membind=1 /home/ai/LLAMA_CPP/8f47e25f56e9792093b7497c68e9f80bab82ed19/llama.cpp/build/bin/llama-server \
--model /mnt/AI/LLM/DeepSeek-R1-UD-Q2_K_XL/DeepSeek-R1-UD-Q2_K_XL-00001-of-00006.gguf \
--threads 16 \
--n-gpu-layers 99 \
--override-tensor ".ffn_.*_exps.=CPU"

--cpunodebind=1 can be used instead of --physcpubind="8,10,12,14, 24,26,28,30, 9,11,13,15, 25,27,29,31" to the same effect. Essentially, it is about exclusively using numa's NODE1 and make sure that the model is loaded accordingly close to CPU0 

Problem description & steps to reproduce

I get 4t/s when I run LLAMA-CLI with the same parameters while I get only 2.4t/s when running LLAMA-SERVER. What I also observe, is that CPU usage (for those carefully picked 16 cores) is at 100% while it is only 75-80% evenly distributed over 16 cores in case of LLAMA-SERVER. GPU/VRAM usage is equal in both cases.

Tested with https://huggingface.co/unsloth/DeepSeek-R1-0528-GGUF/tree/main/Q2_K_L
and https://huggingface.co/unsloth/DeepSeek-R1-GGUF-UD/tree/main/UD-Q2_K_XL

First Bad Commit

I used this commit: 8f47e25
Unfornunally, the next commit (f470bc3) has another issues which I might discuss in another bug report.

Relevant log output

user@pop-os:~$ CUDA_VISIBLE_DEVICES="0," \
numactl --physcpubind="8,10,12,14, 24,26,28,30, 9,11,13,15, 25,27,29,31" --membind=1 /home/ai/LLAMA_CPP/8f47e25f56e9792093b7497c68e9f80bab82ed19/llama.cpp/build/bin/llama-server \
--model /mnt/AI/LLM/DeepSeek-R1-UD-Q2_K_XL/DeepSeek-R1-UD-Q2_K_XL-00001-of-00006.gguf \
--threads 16 \
--n-gpu-layers 99 \
--temp 0.6 --top_p 0.95 --min_p 0.01 \
--ctx-size 32768 \
--flash-attn \
--override-tensor ".ffn_.*_exps.=CPU" --seed 1234567890
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 3090, compute capability 8.6, VMM: yes
build: 5614 (8f47e25f) with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
system info: n_threads = 16, n_threads_batch = 16, total_threads = 32

system_info: n_threads = 16 (n_threads_batch = 16) / 32 | CUDA : ARCHS = 500,610,700,750,800,860,890 | USE_GRAPHS = 1 | PEER_MAX_BATCH_SIZE = 128 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | AARCH64_REPACK = 1 | 

main: binding port with default address family
main: HTTP server is listening, hostname: 127.0.0.1, port: 8080, http threads: 31
main: loading model
srv    load_model: loading model '/mnt/AI/LLM/DeepSeek-R1-UD-Q2_K_XL/DeepSeek-R1-UD-Q2_K_XL-00001-of-00006.gguf'
llama_model_load_from_file_impl: using device CUDA0 (NVIDIA GeForce RTX 3090) - 23858 MiB free
llama_model_loader: additional 5 GGUFs metadata loaded.
llama_model_loader: loaded meta data with 62 key-value pairs and 1086 tensors from /mnt/AI/LLM/DeepSeek-R1-UD-Q2_K_XL/DeepSeek-R1-UD-Q2_K_XL-00001-of-00006.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              = Deepseek-R1
llama_model_loader: - kv   3:                           general.basename str              = Deepseek-R1
llama_model_loader: - kv   4:                       general.quantized_by str              = Unsloth
llama_model_loader: - kv   5:                         general.size_label str              = 256x20B
llama_model_loader: - kv   6:                            general.license str              = mit
llama_model_loader: - kv   7:                           general.repo_url str              = https://huggingface.co/unsloth
llama_model_loader: - kv   8:                   general.base_model.count u32              = 1
llama_model_loader: - kv   9:                  general.base_model.0.name str              = DeepSeek R1
llama_model_loader: - kv  10:          general.base_model.0.organization str              = Deepseek Ai
llama_model_loader: - kv  11:              general.base_model.0.repo_url str              = https://huggingface.co/deepseek-ai/De...
llama_model_loader: - kv  12:                               general.tags arr[str,3]       = ["deepseek", "unsloth", "transformers"]
llama_model_loader: - kv  13:                          general.languages arr[str,1]       = ["en"]
llama_model_loader: - kv  14:                      deepseek2.block_count u32              = 61
llama_model_loader: - kv  15:                   deepseek2.context_length u32              = 163840
llama_model_loader: - kv  16:                 deepseek2.embedding_length u32              = 7168
llama_model_loader: - kv  17:              deepseek2.feed_forward_length u32              = 18432
llama_model_loader: - kv  18:             deepseek2.attention.head_count u32              = 128
llama_model_loader: - kv  19:          deepseek2.attention.head_count_kv u32              = 1
llama_model_loader: - kv  20:                   deepseek2.rope.freq_base f32              = 10000.000000
llama_model_loader: - kv  21: deepseek2.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  22:                deepseek2.expert_used_count u32              = 8
llama_model_loader: - kv  23:        deepseek2.leading_dense_block_count u32              = 3
llama_model_loader: - kv  24:                       deepseek2.vocab_size u32              = 129280
llama_model_loader: - kv  25:            deepseek2.attention.q_lora_rank u32              = 1536
llama_model_loader: - kv  26:           deepseek2.attention.kv_lora_rank u32              = 512
llama_model_loader: - kv  27:             deepseek2.attention.key_length u32              = 576
llama_model_loader: - kv  28:           deepseek2.attention.value_length u32              = 512
llama_model_loader: - kv  29:         deepseek2.attention.key_length_mla u32              = 192
llama_model_loader: - kv  30:       deepseek2.attention.value_length_mla u32              = 128
llama_model_loader: - kv  31:       deepseek2.expert_feed_forward_length u32              = 2048
llama_model_loader: - kv  32:                     deepseek2.expert_count u32              = 256
llama_model_loader: - kv  33:              deepseek2.expert_shared_count u32              = 1
llama_model_loader: - kv  34:             deepseek2.expert_weights_scale f32              = 2.500000
llama_model_loader: - kv  35:              deepseek2.expert_weights_norm bool             = true
llama_model_loader: - kv  36:               deepseek2.expert_gating_func u32              = 2
llama_model_loader: - kv  37:             deepseek2.rope.dimension_count u32              = 64
llama_model_loader: - kv  38:                deepseek2.rope.scaling.type str              = yarn
llama_model_loader: - kv  39:              deepseek2.rope.scaling.factor f32              = 40.000000
llama_model_loader: - kv  40: deepseek2.rope.scaling.original_context_length u32              = 4096
llama_model_loader: - kv  41: deepseek2.rope.scaling.yarn_log_multiplier f32              = 0.100000
llama_model_loader: - kv  42:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  43:                         tokenizer.ggml.pre str              = deepseek-v3
llama_model_loader: - kv  44:                      tokenizer.ggml.tokens arr[str,129280]  = ["<｜begin▁of▁sentence｜>", "<�...
llama_model_loader: - kv  45:                  tokenizer.ggml.token_type arr[i32,129280]  = [3, 3, 3, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  46:                      tokenizer.ggml.merges arr[str,127741]  = ["Ġ t", "Ġ a", "i n", "Ġ Ġ", "h e...
llama_model_loader: - kv  47:                tokenizer.ggml.bos_token_id u32              = 0
llama_model_loader: - kv  48:                tokenizer.ggml.eos_token_id u32              = 1
llama_model_loader: - kv  49:            tokenizer.ggml.padding_token_id u32              = 2
llama_model_loader: - kv  50:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  51:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  52:                    tokenizer.chat_template str              = {% if not add_generation_prompt is de...
llama_model_loader: - kv  53:               general.quantization_version u32              = 2
llama_model_loader: - kv  54:                          general.file_type u32              = 10
llama_model_loader: - kv  55:                      quantize.imatrix.file str              = DeepSeek-R1-GGUF/imatrix_unsloth.dat
llama_model_loader: - kv  56:                   quantize.imatrix.dataset str              = unsloth_calibration_DeepSeek-R1.txt
llama_model_loader: - kv  57:             quantize.imatrix.entries_count i32              = 720
llama_model_loader: - kv  58:              quantize.imatrix.chunks_count i32              = 35
llama_model_loader: - kv  59:                                   split.no u16              = 0
llama_model_loader: - kv  60:                        split.tensors.count i32              = 1086
llama_model_loader: - kv  61:                                split.count u16              = 6
llama_model_loader: - type  f32:  361 tensors
llama_model_loader: - type q8_0:  122 tensors
llama_model_loader: - type q2_K:  122 tensors
llama_model_loader: - type q3_K:   54 tensors
llama_model_loader: - type q4_K:  389 tensors
llama_model_loader: - type q5_K:   23 tensors
llama_model_loader: - type q6_K:   15 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q2_K - Medium
print_info: file size   = 233.18 GiB (2.98 BPW) 
load: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
load: special tokens cache size = 818
load: token to piece cache size = 0.8223 MB
print_info: arch             = deepseek2
print_info: vocab_only       = 0
print_info: n_ctx_train      = 163840
print_info: n_embd           = 7168
print_info: n_layer          = 61
print_info: n_head           = 128
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            = 128
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         = 256
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  = 10000.0
print_info: freq_scale_train = 0.025
print_info: n_ctx_orig_yarn  = 4096
print_info: rope_finetuned   = unknown
print_info: ssm_d_conv       = 0
print_info: ssm_d_inner      = 0
print_info: ssm_d_state      = 0
print_info: ssm_dt_rank      = 0
print_info: ssm_dt_b_c_rms   = 0
print_info: model type       = 671B
print_info: model params     = 671.03 B
print_info: general.name     = Deepseek-R1
print_info: n_layer_dense_lead   = 3
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 = 2.5
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          = 129280
print_info: n_merges         = 127741
print_info: BOS token        = 0 '<｜begin▁of▁sentence｜>'
print_info: EOS token        = 1 '<｜end▁of▁sentence｜>'
print_info: EOT token        = 1 '<｜end▁of▁sentence｜>'
print_info: PAD token        = 2 '<｜▁pad▁｜>'
print_info: LF token         = 201 'Ċ'
print_info: FIM PRE token    = 128801 '<｜fim▁begin｜>'
print_info: FIM SUF token    = 128800 '<｜fim▁hole｜>'
print_info: FIM MID token    = 128802 '<｜fim▁end｜>'
print_info: EOG token        = 1 '<｜end▁of▁sentence｜>'
print_info: max token length = 256
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:   CPU_Mapped model buffer size = 45875.33 MiB
load_tensors:   CPU_Mapped model buffer size = 47175.69 MiB
load_tensors:   CPU_Mapped model buffer size = 46713.92 MiB
load_tensors:   CPU_Mapped model buffer size = 47079.97 MiB
load_tensors:   CPU_Mapped model buffer size = 46513.37 MiB
load_tensors:   CPU_Mapped model buffer size =  4394.39 MiB
load_tensors:        CUDA0 model buffer size =  9686.98 MiB
....................................................................................................
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 32768
llama_context: n_ctx_per_seq = 32768
llama_context: n_batch       = 2048
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = 1
llama_context: freq_base     = 10000.0
llama_context: freq_scale    = 0.025
llama_context: n_ctx_per_seq (32768) < n_ctx_train (163840) -- the full capacity of the model will not be utilized
llama_context:  CUDA_Host  output buffer size =     0.49 MiB
llama_kv_cache_unified:      CUDA0 KV buffer size =  4148.00 MiB
llama_kv_cache_unified: size = 4148.00 MiB ( 32768 cells,  61 layers,  1 seqs), K (f16): 2196.00 MiB, V (f16): 1952.00 MiB
llama_context:      CUDA0 compute buffer size =  3253.50 MiB
llama_context:  CUDA_Host compute buffer size =    78.01 MiB
llama_context: graph nodes  = 4904
llama_context: graph splits = 176 (with bs=512), 118 (with bs=1)
common_init_from_params: setting dry_penalty_last_n to ctx_size = 32768
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 = 32768
main: model loaded
main: chat template, chat_template: {% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% set ns = namespace(is_first=false, is_tool=false, is_output_first=true, system_prompt='', is_first_sp=true) %}{%- for message in messages %}{%- if message['role'] == 'system' %}{%- if ns.is_first_sp %}{% set ns.system_prompt = ns.system_prompt + message['content'] %}{% set ns.is_first_sp = false %}{%- else %}{% set ns.system_prompt = ns.system_prompt + '\n\n' + message['content'] %}{%- endif %}{%- endif %}{%- endfor %}{{ bos_token }}{{ ns.system_prompt }}{%- for message in messages %}{%- if message['role'] == 'user' %}{%- set ns.is_tool = false -%}{{'<｜User｜>' + message['content']}}{%- endif %}{%- if message['role'] == 'assistant' and 'tool_calls' in message %}{%- set ns.is_tool = false -%}{%- for tool in message['tool_calls'] %}{%- if not ns.is_first %}{%- if message['content'] is none %}{{'<｜Assistant｜><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\n' + '' + '\n' + tool['function']['arguments'] + '\n' + '' + '<｜tool▁call▁end｜>'}}{%- else %}{{'<｜Assistant｜>' + message['content'] + '<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\n' + '' + '\n' + tool['function']['arguments'] + '\n' + '' + '<｜tool▁call▁end｜>'}}{%- endif %}{%- set ns.is_first = true -%}{%- else %}{{'\n' + '<｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\n' + '' + '\n' + tool['function']['arguments'] + '\n' + '' + '<｜tool▁call▁end｜>'}}{%- endif %}{%- endfor %}{{'<｜tool▁calls▁end｜><｜end▁of▁sentence｜>'}}{%- endif %}{%- if message['role'] == 'assistant' and 'tool_calls' not in message %}{%- if ns.is_tool %}{{'<｜tool▁outputs▁end｜>' + message['content'] + '<｜end▁of▁sentence｜>'}}{%- set ns.is_tool = false -%}{%- else %}{% set content = message['content'] %}{% if '</think>' in content %}{% set content = content.split('</think>')[-1] %}{% endif %}{{'<｜Assistant｜>' + content + '<｜end▁of▁sentence｜>'}}{%- endif %}{%- endif %}{%- if message['role'] == 'tool' %}{%- set ns.is_tool = true -%}{%- if ns.is_output_first %}{{'<｜tool▁outputs▁begin｜><｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}}{%- set ns.is_output_first = false %}{%- else %}{{'<｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}}{%- endif %}{%- endif %}{%- endfor -%}{% if ns.is_tool %}{{'<｜tool▁outputs▁end｜>'}}{% endif %}{% if add_generation_prompt and not ns.is_tool %}{{'<｜Assistant｜><think>\n'}}{% endif %}, example_format: 'You are a helpful assistant

<｜User｜>Hello<｜Assistant｜>Hi there<｜end▁of▁sentence｜><｜User｜>How are you?<｜Assistant｜>'
main: server is listening on http://127.0.0.1:8080 - starting the main loop
srv  update_slots: all slots are idle
srv  log_server_r: request: GET / 127.0.0.1 200
srv  log_server_r: request: GET /favicon.ico 127.0.0.1 404
srv  log_server_r: request: GET /props 127.0.0.1 200
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 = 32768, n_keep = 0, n_prompt_tokens = 10846
slot update_slots: id  0 | task 0 | kv cache rm [0, end)
slot update_slots: id  0 | task 0 | prompt processing progress, n_past = 2048, n_tokens = 2048, progress = 0.188825
slot update_slots: id  0 | task 0 | kv cache rm [2048, end)
slot update_slots: id  0 | task 0 | prompt processing progress, n_past = 4096, n_tokens = 2048, progress = 0.377651
slot update_slots: id  0 | task 0 | kv cache rm [4096, end)
slot update_slots: id  0 | task 0 | prompt processing progress, n_past = 6144, n_tokens = 2048, progress = 0.566476
slot update_slots: id  0 | task 0 | kv cache rm [6144, end)
slot update_slots: id  0 | task 0 | prompt processing progress, n_past = 8192, n_tokens = 2048, progress = 0.755301
slot update_slots: id  0 | task 0 | kv cache rm [8192, end)
slot update_slots: id  0 | task 0 | prompt processing progress, n_past = 10240, n_tokens = 2048, progress = 0.944127
slot update_slots: id  0 | task 0 | kv cache rm [10240, end)
slot update_slots: id  0 | task 0 | prompt processing progress, n_past = 10846, n_tokens = 606, progress = 1.000000
slot update_slots: id  0 | task 0 | prompt done, n_past = 10846, n_tokens = 606
slot      release: id  0 | task 0 | stop processing: n_past = 11656, truncated = 0
slot print_timing: id  0 | task 0 | 
prompt eval time =  961813.26 ms / 10846 tokens (   88.68 ms per token,    11.28 tokens per second)
       eval time =  347536.86 ms /   811 tokens (  428.53 ms per token,     2.33 tokens per second)
      total time = 1309350.13 ms / 11657 tokens
srv  update_slots: all slots are idle
srv  log_server_r: request: POST /v1/chat/completions 127.0.0.1 200

[Rotatingxenomorph]

Off-topic but I don't get the 0.01 min-p used with top-p.

My thinking with deepseek r1 (the real one) is that for the thinking part, the ~0.6 temp and top-p 0.95 is so it doesn't get stuck in repetitive loops with high prob tokens. Once it's /thunk, change it to whatever kind of sampling seems appropriate for the situation.

[artraze]

I was meaning to report this as well. I actually have an even larger performance dropoff using -ot ... -ngl 99 vs pure CPU inference: my llama-cli is much faster (3.8x) than the reporter's while my llama-server is not (1.6x). Indeed I have just stopped using mixed inference on llama-server anymore.

I tested as far back as daa4228 and collected these results from 6c9ee3b. All built with CUDA, using CUDA_VISIBLE_DEVICES to hide the GPU in tests. The CPU is an Epyc 9654 (in 48c VM) with 12ch memory, the GPU is a 4090.

llama.cpp/build/bin/llama-cli --n-gpu-layers 99 -ot .ffn_.*_exps.=CPU --ctx-size 60000 --cache-type-k f16 --cache-type-v f16 --temp 0.6 --conversation --system-prompt-file prompt-default.txt -m DeepSeek-V3-0324-Q4_K_M.gguf --no-context-shift --flash-attn

CUDA_VISIBLE_DEVICES=-1
llama_perf_context_print: prompt eval time =    2692.58 ms /    73 tokens (   36.88 ms per token,    27.11 tokens per second)
llama_perf_context_print:        eval time =   51397.24 ms /   458 runs   (  112.22 ms per token,     8.91 tokens per second)
llama_perf_context_print:       total time =  136939.32 ms /   531 tokens

CUDA_VISIBLE_DEVICES=0
llama_perf_context_print: prompt eval time =   18146.54 ms /    73 tokens (  248.58 ms per token,     4.02 tokens per second)
llama_perf_context_print:        eval time =   33537.09 ms /   504 runs   (   66.54 ms per token,    15.03 tokens per second)
llama_perf_context_print:       total time =   59074.46 ms /   577 tokens


CUDA_VISIBLE_DEVICES=-1 llama.cpp/build/bin/llama-bench -p 512 -n 512 -r 3 -d 0 -m DeepSeek-V3-0324-Q4_K_M.gguf 
ggml_cuda_init: failed to initialize CUDA: no CUDA-capable device is detected
| model                          |       size |     params | backend   | ngl |    test |           t/s |
| ------------------------------ | ---------: | ---------: | --------- | --: | ------: | ------------: |
| deepseek2 671B Q4_K - Medium   | 378.02 GiB |   671.03 B | CUDA      |  99 |   pp512 |  31.30 ± 0.06 |
| deepseek2 671B Q4_K - Medium   | 378.02 GiB |   671.03 B | CUDA      |  99 |   tg512 |   8.45 ± 0.03 |

CUDA_VISIBLE_DEVICES=0 llama.cpp/build/bin/llama-bench -p 512 -n 512 -r 3 -d 0 -ngl 99 -ot .ffn_.*_exps.=CPU -m DeepSeek-V3-0324-Q4_K_M.gguf
| model                          |       size |     params | backend   | ngl | ot                  |    test |           t/s |
| ------------------------------ | ---------: | ---------: | --------- | --: | ------------------- | ------: | ------------: |
| deepseek2 671B Q4_K - Medium   | 378.02 GiB |   671.03 B | CUDA      |  99 | .ffn_.*_exps.=CPU   |   pp512 |  27.57 ± 0.01 |
| deepseek2 671B Q4_K - Medium   | 378.02 GiB |   671.03 B | CUDA      |  99 | .ffn_.*_exps.=CPU   |   tg512 |  14.20 ± 0.01 |


llama.cpp/build/bin/llama-server --n-gpu-layers 99 -ot .ffn_.*_exps.=CPU --ctx-size 60000 --cache-type-k f16 --cache-type-v f16 --temp 0.6 --no-cont-batching --log-prefix --log-timestamps --log-colors --threads-http 1 -np 1 -m DeepSeek-V3-0324-Q4_K_M.gguf --no-context-shift --flash-attn

CUDA_VISIBLE_DEVICES=-1
prompt eval time =    2461.55 ms /    66 tokens (   37.30 ms per token,    26.81 tokens per second)
       eval time =   55492.33 ms /   469 tokens (  118.32 ms per token,     8.45 tokens per second)
      total time =   57953.88 ms /   535 tokens

CUDA_VISIBLE_DEVICES=0
prompt eval time =   17565.03 ms /    66 tokens (  266.14 ms per token,     3.76 tokens per second)
       eval time =  121675.67 ms /   476 tokens (  255.62 ms per token,     3.91 tokens per second)
      total time =  139240.71 ms /   542 tokens

So the numbers agree with CPU-only inference. With the partial offload, llama-cli and llama-bench agree, while llama-server falls off terribly, dropping from 14t/s to 4t/s. (You might also note that PP for short prompts with offload is terrible but that's not too interesting as it's within 10% for longer prompts as can be seen in llama-bench and that seems reasonable.)

I did more experiments, but realistically they were all pointless because llama-server only uses ~1600% CPU when run with GPU assist and 4600% when CPU only, so degradation is expected. To compare, llama-bench uses the full 4800% regardless of GPU.

[Rotatingxenomorph]

When I tested this a couple of weeks ago I couldn't get any speedup using cli instead of server. This was on an ancient v3 xeon and quad channel ddr4 though.

[JeroenAdam]

Same issue on 2x16 GB VRAM + 64 GB DDR5 6000Mhz, 6.5 t/s with llama-cli and 2~3 t/s less with llama-server
llama-server -m .\Qwen3-235B-A22B-Instruct-2507-Q2_K-00001-of-00002.gguf -ngl 99 -fa -c 30720 -ctk q8_0 -ctv q8_0 --main-gpu 0 -ot ".ffn_(up|down)_exps.=CPU" -t 10 -ts 0.95,1

[Rotatingxenomorph]

I finally set up ubuntu (again) and tested... sure enough, CUDA llama-server is almost 70% faster on linux than llama-server on Windows with the same settings. 256GB quad channel DDR 4, a E5-2697v3 and dual 3060 12GB.

The Windows tax is real.

[joesixpaq]

70% faster on linux than llama-server on Windows

The point made was that server is slower than cli

[Rotatingxenomorph]

70% faster on linux than llama-server on Windows

The point made was that server is slower than cli

Yes. Although that was with an older version of llama.cpp. I will test tomorrow to see if cli is faster than server on linux though.

[artraze]

70% faster on linux than llama-server on Windows

The point made was that server is slower than cli

Yes. Although that was with an older version of llama.cpp. I will test tomorrow to see if cli is faster than server on linux though.

I just tested with 74f52f7 (current head) and using Qwen3-Coder-480B-Q4_K_M under Linux (in VM FWIW, Epyc CPU + Pro6000 GPU) I get:

• 6.3t/s with llama-server (CPU+GPU)
• 15.2t/s with lama-server (CPU only)
• 20t/s with llama-cli (CPU+GPU)

So I can confirm it's still happening and isn't a Windows or older problem

[Rotatingxenomorph]

I just tested with 74f52f7 (current head) and using Qwen3-Coder-480B-Q4_K_M under Linux (in VM FWIW, Epyc CPU + Pro6000 GPU) I get:

* 6.3t/s with llama-server (CPU+GPU)

* 15.2t/s with lama-server (CPU only)

* 20t/s with llama-cli (CPU+GPU)

So I can confirm it's still happening and isn't a Windows or older problem

I tested and I get the same performance on cli as with server on this single socket quad channel ddr 4 machine with cpu+gpu on CUDA. Not sure if it's a useful data point but it is what it is.

[github-actions]

This issue was closed because it has been inactive for 14 days since being marked as stale.