Misc. bug: llama-server speculative decoding not as performant as llama-speculative-simple

[hjc4869]

Name and Version

$ llama-cli --version
version: 5142 (80f19b4)
built with AMD clang version 17.0.6 (CLANG: AOCC_5.0.0-Build#1377 2024_09_24) for x86_64-unknown-linux-gnu

Operating systems

Linux

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

llama-server

Command line

llama-speculative-simple -c 512 -cd 512 -m ~/models/qwen2.5-72b-iq4xs.gguf -md ~/models/qwen2.5-0.5b-iq4xs.gguf -ngld 999 -ngl 999 --draft-max 3 --draft-min 3 -p "Repeat this sentence for 50 times: This is a test\n" --color -s 123 -n 260 -t 1 -fa -ctv q8_0 -ctk q8_0

llama-server -c 512 -cd 512 -m ~/models/qwen2.5-72b-iq4xs.gguf -md ~/models/qwen2.5-0.5b-iq4xs.gguf -ngld 999 -ngl 999 --draft-max 3 --draft-min 3 -t 1 -fa -ctv q8_0 -ctk q8_0

curl localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"max_tokens": 300, "messages": [{"role": "user", "content": "Repeat this sentence for 50 times: This is a test"}]}'

Problem description & steps to reproduce

When testing the best case scenario of speculative decoding (repeating the same content), llama-server generates less drafted tokens and performs worse than llama-speculative-simple. Using the same args to launch both programs and generate 261 tokens, we only see 154/156 drafted tokens in llama-server compared to 195 in llama-speculative-simple with similar acceptance rate (almost 100%). As a result, the performance of llama-server is around 10%-15% less than llama-speculative-simple.

ROCR_VISIBLE_DEVICES=1 ./build/bin/llama-server -c 512 -cd 512 -m ~/models/qwen2.5-72b-iq4xs.gguf -md ~/models/qwen2.5-0.5b-iq4xs.gguf -ngld 999 -ngl 999 --draft-max 3 --draft-min 3 -t 1 -fa -ctv q8_0 -ctk q8_0

curl localhost:8080/v1/chat/completions \ -H "Content-Type: application/json" \ -d '{"max_tokens": 300, "messages": [{"role": "user", "content": "Repeat this sentence for 50 times: This is a test"}]}'

Responses (2 examples):
{"choices":[{"finish_reason":"stop","index":0,"message":{"role":"assistant","content":"This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test."}}],"created":1744767822,"model":"gpt-3.5-turbo","system_fingerprint":"b5142-80f19b41","object":"chat.completion","usage":{"completion_tokens":261,"prompt_tokens":21,"total_tokens":282},"id":"chatcmpl-tojEY2jOFzBWbF2V0cHcPWMuryt4DOEI","timings":{"prompt_n":21,"prompt_ms":221.093,"prompt_per_token_ms":10.528238095238095,"prompt_per_second":94.9826543581208,"predicted_n":261,"predicted_ms":8471.726,"predicted_per_token_ms":32.45872030651341,"predicted_per_second":30.808361837953683,"draft_n":156,"draft_n_accepted":156}}

{"choices":[{"finish_reason":"stop","index":0,"message":{"role":"assistant","content":"This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test."}}],"created":1744767842,"model":"gpt-3.5-turbo","system_fingerprint":"b5142-80f19b41","object":"chat.completion","usage":{"completion_tokens":261,"prompt_tokens":21,"total_tokens":282},"id":"chatcmpl-uVHH0019M7QOf9byfNr1STS2C1JceH2w","timings":{"prompt_n":1,"prompt_ms":70.998,"prompt_per_token_ms":70.998,"prompt_per_second":14.084903800107044,"predicted_n":261,"predicted_ms":8011.701,"predicted_per_token_ms":30.696172413793104,"predicted_per_second":32.57735155118744,"draft_n":159,"draft_n_accepted":154}}

ROCR_VISIBLE_DEVICES=1 ./build/bin/llama-speculative-simple -c 512 -cd 512 -m ~/models/qwen2.5-72b-iq4xs.gguf -md ~/models/qwen2.5-0.5b-iq4xs.gguf -ngld 999 -ngl 999 --draft-max 3 --draft-min 3 -p "Repeat this sentence for 50 times: This is a test\n" --color -s 123 -n 260 -t 1 -fa -ctv q8_0 -ctk q8_0

n_draft = 3
n_predict = 261
n_drafted = 195
n_accept = 195
accept = 100.000%

First Bad Commit

No response

Relevant log output

Detailed logs of llama-server

ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 ROCm devices:
  Device 0: AMD Radeon PRO W7900 Dual Slot , gfx1100 (0x1100), VMM: no, Wave Size: 32
build: 5142 (80f19b41) with AMD clang version 17.0.6 (CLANG: AOCC_5.0.0-Build#1377 2024_09_24) for x86_64-unknown-linux-gnu
system info: n_threads = 1, n_threads_batch = 1, total_threads = 64

system_info: n_threads = 1 (n_threads_batch = 1) / 64 | ROCm : NO_VMM = 1 | PEER_MAX_BATCH_SIZE = 128 | FA_ALL_QUANTS = 1 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 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: 63
main: loading model
srv    load_model: loading model '/home/david/models/qwen2.5-72b-iq4xs.gguf'
llama_model_load_from_file_impl: using device ROCm0 (AMD Radeon PRO W7900 Dual Slot ) - 49086 MiB free
llama_model_loader: loaded meta data with 35 key-value pairs and 963 tensors from /home/david/models/qwen2.5-72b-iq4xs.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.type str              = model
llama_model_loader: - kv   2:                               general.name str              = Qwen2.5 72B Instruct
llama_model_loader: - kv   3:                           general.finetune str              = Instruct
llama_model_loader: - kv   4:                           general.basename str              = Qwen2.5
llama_model_loader: - kv   5:                         general.size_label str              = 72B
llama_model_loader: - kv   6:                            general.license str              = other
llama_model_loader: - kv   7:                       general.license.name str              = qwen
llama_model_loader: - kv   8:                       general.license.link str              = https://huggingface.co/Qwen/Qwen2.5-7...
llama_model_loader: - kv   9:                   general.base_model.count u32              = 1
llama_model_loader: - kv  10:                  general.base_model.0.name str              = Qwen2.5 72B
llama_model_loader: - kv  11:          general.base_model.0.organization str              = Qwen
llama_model_loader: - kv  12:              general.base_model.0.repo_url str              = https://huggingface.co/Qwen/Qwen2.5-72B
llama_model_loader: - kv  13:                               general.tags arr[str,2]       = ["chat", "text-generation"]
llama_model_loader: - kv  14:                          general.languages arr[str,1]       = ["en"]
llama_model_loader: - kv  15:                          qwen2.block_count u32              = 80
llama_model_loader: - kv  16:                       qwen2.context_length u32              = 32768
llama_model_loader: - kv  17:                     qwen2.embedding_length u32              = 8192
llama_model_loader: - kv  18:                  qwen2.feed_forward_length u32              = 29568
llama_model_loader: - kv  19:                 qwen2.attention.head_count u32              = 64
llama_model_loader: - kv  20:              qwen2.attention.head_count_kv u32              = 8
llama_model_loader: - kv  21:                       qwen2.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  22:     qwen2.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  23:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  24:                         tokenizer.ggml.pre str              = qwen2
llama_model_loader: - kv  25:                      tokenizer.ggml.tokens arr[str,152064]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  26:                  tokenizer.ggml.token_type arr[i32,152064]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  27:                      tokenizer.ggml.merges arr[str,151387]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  28:                tokenizer.ggml.eos_token_id u32              = 151645
llama_model_loader: - kv  29:            tokenizer.ggml.padding_token_id u32              = 151643
llama_model_loader: - kv  30:                tokenizer.ggml.bos_token_id u32              = 151643
llama_model_loader: - kv  31:               tokenizer.ggml.add_bos_token bool             = false
llama_model_loader: - kv  32:                    tokenizer.chat_template str              = {%- if tools %}\n    {{- '<|im_start|>...
llama_model_loader: - kv  33:               general.quantization_version u32              = 2
llama_model_loader: - kv  34:                          general.file_type u32              = 30
llama_model_loader: - type  f32:  401 tensors
llama_model_loader: - type q5_1:   10 tensors
llama_model_loader: - type q5_K:   80 tensors
llama_model_loader: - type q6_K:    1 tensors
llama_model_loader: - type iq4_nl:   70 tensors
llama_model_loader: - type iq4_xs:  401 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = IQ4_XS - 4.25 bpw
print_info: file size   = 37.40 GiB (4.42 BPW) 
load: special tokens cache size = 22
load: token to piece cache size = 0.9310 MB
print_info: arch             = qwen2
print_info: vocab_only       = 0
print_info: n_ctx_train      = 32768
print_info: n_embd           = 8192
print_info: n_layer          = 80
print_info: n_head           = 64
print_info: n_head_kv        = 8
print_info: n_rot            = 128
print_info: n_swa            = 0
print_info: n_swa_pattern    = 1
print_info: n_embd_head_k    = 128
print_info: n_embd_head_v    = 128
print_info: n_gqa            = 8
print_info: n_embd_k_gqa     = 1024
print_info: n_embd_v_gqa     = 1024
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             = 29568
print_info: n_expert         = 0
print_info: n_expert_used    = 0
print_info: causal attn      = 1
print_info: pooling type     = 0
print_info: rope type        = 2
print_info: rope scaling     = linear
print_info: freq_base_train  = 1000000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn  = 32768
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       = 70B
print_info: model params     = 72.71 B
print_info: general.name     = Qwen2.5 72B Instruct
print_info: vocab type       = BPE
print_info: n_vocab          = 152064
print_info: n_merges         = 151387
print_info: BOS token        = 151643 '<|endoftext|>'
print_info: EOS token        = 151645 '<|im_end|>'
print_info: EOT token        = 151645 '<|im_end|>'
print_info: PAD token        = 151643 '<|endoftext|>'
print_info: LF token         = 198 'Ċ'
print_info: FIM PRE token    = 151659 '<|fim_prefix|>'
print_info: FIM SUF token    = 151661 '<|fim_suffix|>'
print_info: FIM MID token    = 151660 '<|fim_middle|>'
print_info: FIM PAD token    = 151662 '<|fim_pad|>'
print_info: FIM REP token    = 151663 '<|repo_name|>'
print_info: FIM SEP token    = 151664 '<|file_sep|>'
print_info: EOG token        = 151643 '<|endoftext|>'
print_info: EOG token        = 151645 '<|im_end|>'
print_info: EOG token        = 151662 '<|fim_pad|>'
print_info: EOG token        = 151663 '<|repo_name|>'
print_info: EOG token        = 151664 '<|file_sep|>'
print_info: max token length = 256
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: offloading 80 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 81/81 layers to GPU
load_tensors:        ROCm0 model buffer size = 37665.82 MiB
load_tensors:   CPU_Mapped model buffer size =   631.12 MiB
..................................................................................................
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 512
llama_context: n_ctx_per_seq = 512
llama_context: n_batch       = 512
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = 1
llama_context: freq_base     = 1000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_per_seq (512) < n_ctx_train (32768) -- the full capacity of the model will not be utilized
llama_context:  ROCm_Host  output buffer size =     0.58 MiB
init: kv_size = 512, offload = 1, type_k = 'q8_0', type_v = 'q8_0', n_layer = 80, can_shift = 1
init:      ROCm0 KV buffer size =    85.00 MiB
llama_context: KV self size  =   85.00 MiB, K (q8_0):   42.50 MiB, V (q8_0):   42.50 MiB
llama_context:      ROCm0 compute buffer size =   313.00 MiB
llama_context:  ROCm_Host compute buffer size =    17.01 MiB
llama_context: graph nodes  = 2647
llama_context: graph splits = 2
common_init_from_params: setting dry_penalty_last_n to ctx_size = 512
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
srv    load_model: loading draft model '/home/david/models/qwen2.5-0.5b-iq4xs.gguf'
llama_model_load_from_file_impl: using device ROCm0 (AMD Radeon PRO W7900 Dual Slot ) - 10848 MiB free
llama_model_loader: loaded meta data with 34 key-value pairs and 290 tensors from /home/david/models/qwen2.5-0.5b-iq4xs.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.type str              = model
llama_model_loader: - kv   2:                               general.name str              = Qwen2.5 0.5B Instruct
llama_model_loader: - kv   3:                           general.finetune str              = Instruct
llama_model_loader: - kv   4:                           general.basename str              = Qwen2.5
llama_model_loader: - kv   5:                         general.size_label str              = 0.5B
llama_model_loader: - kv   6:                            general.license str              = apache-2.0
llama_model_loader: - kv   7:                       general.license.link str              = https://huggingface.co/Qwen/Qwen2.5-0...
llama_model_loader: - kv   8:                   general.base_model.count u32              = 1
llama_model_loader: - kv   9:                  general.base_model.0.name str              = Qwen2.5 0.5B
llama_model_loader: - kv  10:          general.base_model.0.organization str              = Qwen
llama_model_loader: - kv  11:              general.base_model.0.repo_url str              = https://huggingface.co/Qwen/Qwen2.5-0.5B
llama_model_loader: - kv  12:                               general.tags arr[str,2]       = ["chat", "text-generation"]
llama_model_loader: - kv  13:                          general.languages arr[str,1]       = ["en"]
llama_model_loader: - kv  14:                          qwen2.block_count u32              = 24
llama_model_loader: - kv  15:                       qwen2.context_length u32              = 32768
llama_model_loader: - kv  16:                     qwen2.embedding_length u32              = 896
llama_model_loader: - kv  17:                  qwen2.feed_forward_length u32              = 4864
llama_model_loader: - kv  18:                 qwen2.attention.head_count u32              = 14
llama_model_loader: - kv  19:              qwen2.attention.head_count_kv u32              = 2
llama_model_loader: - kv  20:                       qwen2.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  21:     qwen2.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  22:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  23:                         tokenizer.ggml.pre str              = qwen2
llama_model_loader: - kv  24:                      tokenizer.ggml.tokens arr[str,151936]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  25:                  tokenizer.ggml.token_type arr[i32,151936]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  26:                      tokenizer.ggml.merges arr[str,151387]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  27:                tokenizer.ggml.eos_token_id u32              = 151645
llama_model_loader: - kv  28:            tokenizer.ggml.padding_token_id u32              = 151643
llama_model_loader: - kv  29:                tokenizer.ggml.bos_token_id u32              = 151643
llama_model_loader: - kv  30:               tokenizer.ggml.add_bos_token bool             = false
llama_model_loader: - kv  31:                    tokenizer.chat_template str              = {%- if tools %}\n    {{- '<|im_start|>...
llama_model_loader: - kv  32:               general.quantization_version u32              = 2
llama_model_loader: - kv  33:                          general.file_type u32              = 30
llama_model_loader: - type  f32:  121 tensors
llama_model_loader: - type q5_1:   24 tensors
llama_model_loader: - type q8_0:    1 tensors
llama_model_loader: - type q5_K:    3 tensors
llama_model_loader: - type iq4_nl:  120 tensors
llama_model_loader: - type iq4_xs:   21 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = IQ4_XS - 4.25 bpw
print_info: file size   = 329.49 MiB (5.59 BPW) 
load: special tokens cache size = 22
load: token to piece cache size = 0.9310 MB
print_info: arch             = qwen2
print_info: vocab_only       = 0
print_info: n_ctx_train      = 32768
print_info: n_embd           = 896
print_info: n_layer          = 24
print_info: n_head           = 14
print_info: n_head_kv        = 2
print_info: n_rot            = 64
print_info: n_swa            = 0
print_info: n_swa_pattern    = 1
print_info: n_embd_head_k    = 64
print_info: n_embd_head_v    = 64
print_info: n_gqa            = 7
print_info: n_embd_k_gqa     = 128
print_info: n_embd_v_gqa     = 128
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             = 4864
print_info: n_expert         = 0
print_info: n_expert_used    = 0
print_info: causal attn      = 1
print_info: pooling type     = 0
print_info: rope type        = 2
print_info: rope scaling     = linear
print_info: freq_base_train  = 1000000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn  = 32768
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       = 1B
print_info: model params     = 494.03 M
print_info: general.name     = Qwen2.5 0.5B Instruct
print_info: vocab type       = BPE
print_info: n_vocab          = 151936
print_info: n_merges         = 151387
print_info: BOS token        = 151643 '<|endoftext|>'
print_info: EOS token        = 151645 '<|im_end|>'
print_info: EOT token        = 151645 '<|im_end|>'
print_info: PAD token        = 151643 '<|endoftext|>'
print_info: LF token         = 198 'Ċ'
print_info: FIM PRE token    = 151659 '<|fim_prefix|>'
print_info: FIM SUF token    = 151661 '<|fim_suffix|>'
print_info: FIM MID token    = 151660 '<|fim_middle|>'
print_info: FIM PAD token    = 151662 '<|fim_pad|>'
print_info: FIM REP token    = 151663 '<|repo_name|>'
print_info: FIM SEP token    = 151664 '<|file_sep|>'
print_info: EOG token        = 151643 '<|endoftext|>'
print_info: EOG token        = 151645 '<|im_end|>'
print_info: EOG token        = 151662 '<|fim_pad|>'
print_info: EOG token        = 151663 '<|repo_name|>'
print_info: EOG token        = 151664 '<|file_sep|>'
print_info: max token length = 256
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: offloading 24 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 25/25 layers to GPU
load_tensors:        ROCm0 model buffer size =   329.52 MiB
load_tensors:   CPU_Mapped model buffer size =   137.94 MiB
...........................................
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 512
llama_context: n_ctx_per_seq = 512
llama_context: n_batch       = 512
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = 1
llama_context: freq_base     = 1000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_per_seq (512) < n_ctx_train (32768) -- the full capacity of the model will not be utilized
llama_context:  ROCm_Host  output buffer size =     0.58 MiB
init: kv_size = 512, offload = 1, type_k = 'f16', type_v = 'f16', n_layer = 24, can_shift = 1
init:      ROCm0 KV buffer size =     6.00 MiB
llama_context: KV self size  =    6.00 MiB, K (f16):    3.00 MiB, V (f16):    3.00 MiB
llama_context:      ROCm0 compute buffer size =   298.50 MiB
llama_context:  ROCm_Host compute buffer size =     2.76 MiB
llama_context: graph nodes  = 799
llama_context: graph splits = 2
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
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 512
llama_context: n_ctx_per_seq = 512
llama_context: n_batch       = 512
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = 1
llama_context: freq_base     = 1000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_per_seq (512) < n_ctx_train (32768) -- the full capacity of the model will not be utilized
llama_context:  ROCm_Host  output buffer size =     0.58 MiB
init: kv_size = 512, offload = 1, type_k = 'f16', type_v = 'f16', n_layer = 24, can_shift = 1
init:      ROCm0 KV buffer size =     6.00 MiB
llama_context: KV self size  =    6.00 MiB, K (f16):    3.00 MiB, V (f16):    3.00 MiB
llama_context:      ROCm0 compute buffer size =   298.50 MiB
llama_context:  ROCm_Host compute buffer size =     2.76 MiB
llama_context: graph nodes  = 799
llama_context: graph splits = 2
slot         init: id  0 | task -1 | new slot n_ctx_slot = 512
main: model loaded
main: chat template, chat_template: {%- if tools %}
    {{- '<|im_start|>system\n' }}
    {%- if messages[0]['role'] == 'system' %}
        {{- messages[0]['content'] }}
    {%- else %}
        {{- 'You are Qwen, created by Alibaba Cloud. You are a helpful assistant.' }}
    {%- endif %}
    {{- "\n\n# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
    {%- for tool in tools %}
        {{- "\n" }}
        {{- tool | tojson }}
    {%- endfor %}
    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
{%- else %}
    {%- if messages[0]['role'] == 'system' %}
        {{- '<|im_start|>system\n' + messages[0]['content'] + '<|im_end|>\n' }}
    {%- else %}
        {{- '<|im_start|>system\nYou are Qwen, created by Alibaba Cloud. You are a helpful assistant.<|im_end|>\n' }}
    {%- endif %}
{%- endif %}
{%- for message in messages %}
    {%- if (message.role == "user") or (message.role == "system" and not loop.first) or (message.role == "assistant" and not message.tool_calls) %}
        {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>' + '\n' }}
    {%- elif message.role == "assistant" %}
        {{- '<|im_start|>' + message.role }}
        {%- if message.content %}
            {{- '\n' + message.content }}
        {%- endif %}
        {%- for tool_call in message.tool_calls %}
            {%- if tool_call.function is defined %}
                {%- set tool_call = tool_call.function %}
            {%- endif %}
            {{- '\n<tool_call>\n{"name": "' }}
            {{- tool_call.name }}
            {{- '", "arguments": ' }}
            {{- tool_call.arguments | tojson }}
            {{- '}\n</tool_call>' }}
        {%- endfor %}
        {{- '<|im_end|>\n' }}
    {%- elif message.role == "tool" %}
        {%- if (loop.index0 == 0) or (messages[loop.index0 - 1].role != "tool") %}
            {{- '<|im_start|>user' }}
        {%- endif %}
        {{- '\n<tool_response>\n' }}
        {{- message.content }}
        {{- '\n</tool_response>' }}
        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
            {{- '<|im_end|>\n' }}
        {%- endif %}
    {%- endif %}
{%- endfor %}
{%- if add_generation_prompt %}
    {{- '<|im_start|>assistant\n' }}
{%- endif %}
, example_format: '<|im_start|>system
You are a helpful assistant<|im_end|>
<|im_start|>user
Hello<|im_end|>
<|im_start|>assistant
Hi there<|im_end|>
<|im_start|>user
How are you?<|im_end|>
<|im_start|>assistant
'
main: server is listening on http://127.0.0.1:8080 - 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 = 512, n_keep = 0, n_prompt_tokens = 21
slot update_slots: id  0 | task 0 | kv cache rm [0, end)
slot update_slots: id  0 | task 0 | prompt processing progress, n_past = 21, n_tokens = 21, progress = 1.000000
slot update_slots: id  0 | task 0 | prompt done, n_past = 21, n_tokens = 21
slot      release: id  0 | task 0 | stop processing: n_past = 281, truncated = 0
slot print_timing: id  0 | task 0 | 
prompt eval time =     221.09 ms /    21 tokens (   10.53 ms per token,    94.98 tokens per second)
       eval time =    8471.73 ms /   261 tokens (   32.46 ms per token,    30.81 tokens per second)
      total time =    8692.82 ms /   282 tokens
slot print_timing: id  0 | task 0 | 
draft acceptance rate = 1.00000 (  156 accepted /   156 generated)
srv  update_slots: all slots are idle
srv  log_server_r: request: POST /v1/chat/completions 127.0.0.1 200
srv  params_from_: Chat format: Content-only
slot launch_slot_: id  0 | task 54 | processing task
slot update_slots: id  0 | task 54 | new prompt, n_ctx_slot = 512, n_keep = 0, n_prompt_tokens = 21
slot update_slots: id  0 | task 54 | need to evaluate at least 1 token to generate logits, n_past = 21, n_prompt_tokens = 21
slot update_slots: id  0 | task 54 | kv cache rm [20, end)
slot update_slots: id  0 | task 54 | prompt processing progress, n_past = 21, n_tokens = 1, progress = 0.047619
slot update_slots: id  0 | task 54 | prompt done, n_past = 21, n_tokens = 1
slot      release: id  0 | task 54 | stop processing: n_past = 281, truncated = 0
slot print_timing: id  0 | task 54 | 
prompt eval time =      71.00 ms /     1 tokens (   71.00 ms per token,    14.08 tokens per second)
       eval time =    8011.70 ms /   261 tokens (   30.70 ms per token,    32.58 tokens per second)
      total time =    8082.70 ms /   262 tokens
slot print_timing: id  0 | task 54 | 
draft acceptance rate = 0.96855 (  154 accepted /   159 generated)
srv  update_slots: all slots are idle
srv  log_server_r: request: POST /v1/chat/completions 127.0.0.1 200

Detailed logs of llama-speculative-simple

ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 ROCm devices:
  Device 0: AMD Radeon PRO W7900 Dual Slot , gfx1100 (0x1100), VMM: no, Wave Size: 32
build: 5142 (80f19b41) with AMD clang version 17.0.6 (CLANG: AOCC_5.0.0-Build#1377 2024_09_24) for x86_64-unknown-linux-gnu
llama_model_load_from_file_impl: using device ROCm0 (AMD Radeon PRO W7900 Dual Slot ) - 49086 MiB free
llama_model_loader: loaded meta data with 35 key-value pairs and 963 tensors from /home/david/models/qwen2.5-72b-iq4xs.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.type str              = model
llama_model_loader: - kv   2:                               general.name str              = Qwen2.5 72B Instruct
llama_model_loader: - kv   3:                           general.finetune str              = Instruct
llama_model_loader: - kv   4:                           general.basename str              = Qwen2.5
llama_model_loader: - kv   5:                         general.size_label str              = 72B
llama_model_loader: - kv   6:                            general.license str              = other
llama_model_loader: - kv   7:                       general.license.name str              = qwen
llama_model_loader: - kv   8:                       general.license.link str              = https://huggingface.co/Qwen/Qwen2.5-7...
llama_model_loader: - kv   9:                   general.base_model.count u32              = 1
llama_model_loader: - kv  10:                  general.base_model.0.name str              = Qwen2.5 72B
llama_model_loader: - kv  11:          general.base_model.0.organization str              = Qwen
llama_model_loader: - kv  12:              general.base_model.0.repo_url str              = https://huggingface.co/Qwen/Qwen2.5-72B
llama_model_loader: - kv  13:                               general.tags arr[str,2]       = ["chat", "text-generation"]
llama_model_loader: - kv  14:                          general.languages arr[str,1]       = ["en"]
llama_model_loader: - kv  15:                          qwen2.block_count u32              = 80
llama_model_loader: - kv  16:                       qwen2.context_length u32              = 32768
llama_model_loader: - kv  17:                     qwen2.embedding_length u32              = 8192
llama_model_loader: - kv  18:                  qwen2.feed_forward_length u32              = 29568
llama_model_loader: - kv  19:                 qwen2.attention.head_count u32              = 64
llama_model_loader: - kv  20:              qwen2.attention.head_count_kv u32              = 8
llama_model_loader: - kv  21:                       qwen2.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  22:     qwen2.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  23:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  24:                         tokenizer.ggml.pre str              = qwen2
llama_model_loader: - kv  25:                      tokenizer.ggml.tokens arr[str,152064]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  26:                  tokenizer.ggml.token_type arr[i32,152064]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  27:                      tokenizer.ggml.merges arr[str,151387]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  28:                tokenizer.ggml.eos_token_id u32              = 151645
llama_model_loader: - kv  29:            tokenizer.ggml.padding_token_id u32              = 151643
llama_model_loader: - kv  30:                tokenizer.ggml.bos_token_id u32              = 151643
llama_model_loader: - kv  31:               tokenizer.ggml.add_bos_token bool             = false
llama_model_loader: - kv  32:                    tokenizer.chat_template str              = {%- if tools %}\n    {{- '<|im_start|>...
llama_model_loader: - kv  33:               general.quantization_version u32              = 2
llama_model_loader: - kv  34:                          general.file_type u32              = 30
llama_model_loader: - type  f32:  401 tensors
llama_model_loader: - type q5_1:   10 tensors
llama_model_loader: - type q5_K:   80 tensors
llama_model_loader: - type q6_K:    1 tensors
llama_model_loader: - type iq4_nl:   70 tensors
llama_model_loader: - type iq4_xs:  401 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = IQ4_XS - 4.25 bpw
print_info: file size   = 37.40 GiB (4.42 BPW) 
load: special tokens cache size = 22
load: token to piece cache size = 0.9310 MB
print_info: arch             = qwen2
print_info: vocab_only       = 0
print_info: n_ctx_train      = 32768
print_info: n_embd           = 8192
print_info: n_layer          = 80
print_info: n_head           = 64
print_info: n_head_kv        = 8
print_info: n_rot            = 128
print_info: n_swa            = 0
print_info: n_swa_pattern    = 1
print_info: n_embd_head_k    = 128
print_info: n_embd_head_v    = 128
print_info: n_gqa            = 8
print_info: n_embd_k_gqa     = 1024
print_info: n_embd_v_gqa     = 1024
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             = 29568
print_info: n_expert         = 0
print_info: n_expert_used    = 0
print_info: causal attn      = 1
print_info: pooling type     = 0
print_info: rope type        = 2
print_info: rope scaling     = linear
print_info: freq_base_train  = 1000000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn  = 32768
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       = 70B
print_info: model params     = 72.71 B
print_info: general.name     = Qwen2.5 72B Instruct
print_info: vocab type       = BPE
print_info: n_vocab          = 152064
print_info: n_merges         = 151387
print_info: BOS token        = 151643 '<|endoftext|>'
print_info: EOS token        = 151645 '<|im_end|>'
print_info: EOT token        = 151645 '<|im_end|>'
print_info: PAD token        = 151643 '<|endoftext|>'
print_info: LF token         = 198 'Ċ'
print_info: FIM PRE token    = 151659 '<|fim_prefix|>'
print_info: FIM SUF token    = 151661 '<|fim_suffix|>'
print_info: FIM MID token    = 151660 '<|fim_middle|>'
print_info: FIM PAD token    = 151662 '<|fim_pad|>'
print_info: FIM REP token    = 151663 '<|repo_name|>'
print_info: FIM SEP token    = 151664 '<|file_sep|>'
print_info: EOG token        = 151643 '<|endoftext|>'
print_info: EOG token        = 151645 '<|im_end|>'
print_info: EOG token        = 151662 '<|fim_pad|>'
print_info: EOG token        = 151663 '<|repo_name|>'
print_info: EOG token        = 151664 '<|file_sep|>'
print_info: max token length = 256
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: offloading 80 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 81/81 layers to GPU
load_tensors:        ROCm0 model buffer size = 37665.82 MiB
load_tensors:   CPU_Mapped model buffer size =   631.12 MiB
..................................................................................................
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 512
llama_context: n_ctx_per_seq = 512
llama_context: n_batch       = 512
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = 1
llama_context: freq_base     = 1000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_per_seq (512) < n_ctx_train (32768) -- the full capacity of the model will not be utilized
llama_context:  ROCm_Host  output buffer size =     0.58 MiB
init: kv_size = 512, offload = 1, type_k = 'q8_0', type_v = 'q8_0', n_layer = 80, can_shift = 1
init:      ROCm0 KV buffer size =    85.00 MiB
llama_context: KV self size  =   85.00 MiB, K (q8_0):   42.50 MiB, V (q8_0):   42.50 MiB
llama_context:      ROCm0 compute buffer size =   313.00 MiB
llama_context:  ROCm_Host compute buffer size =    17.01 MiB
llama_context: graph nodes  = 2647
llama_context: graph splits = 2
common_init_from_params: setting dry_penalty_last_n to ctx_size = 512
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
llama_model_load_from_file_impl: using device ROCm0 (AMD Radeon PRO W7900 Dual Slot ) - 10848 MiB free
llama_model_loader: loaded meta data with 34 key-value pairs and 290 tensors from /home/david/models/qwen2.5-0.5b-iq4xs.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.type str              = model
llama_model_loader: - kv   2:                               general.name str              = Qwen2.5 0.5B Instruct
llama_model_loader: - kv   3:                           general.finetune str              = Instruct
llama_model_loader: - kv   4:                           general.basename str              = Qwen2.5
llama_model_loader: - kv   5:                         general.size_label str              = 0.5B
llama_model_loader: - kv   6:                            general.license str              = apache-2.0
llama_model_loader: - kv   7:                       general.license.link str              = https://huggingface.co/Qwen/Qwen2.5-0...
llama_model_loader: - kv   8:                   general.base_model.count u32              = 1
llama_model_loader: - kv   9:                  general.base_model.0.name str              = Qwen2.5 0.5B
llama_model_loader: - kv  10:          general.base_model.0.organization str              = Qwen
llama_model_loader: - kv  11:              general.base_model.0.repo_url str              = https://huggingface.co/Qwen/Qwen2.5-0.5B
llama_model_loader: - kv  12:                               general.tags arr[str,2]       = ["chat", "text-generation"]
llama_model_loader: - kv  13:                          general.languages arr[str,1]       = ["en"]
llama_model_loader: - kv  14:                          qwen2.block_count u32              = 24
llama_model_loader: - kv  15:                       qwen2.context_length u32              = 32768
llama_model_loader: - kv  16:                     qwen2.embedding_length u32              = 896
llama_model_loader: - kv  17:                  qwen2.feed_forward_length u32              = 4864
llama_model_loader: - kv  18:                 qwen2.attention.head_count u32              = 14
llama_model_loader: - kv  19:              qwen2.attention.head_count_kv u32              = 2
llama_model_loader: - kv  20:                       qwen2.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  21:     qwen2.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  22:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  23:                         tokenizer.ggml.pre str              = qwen2
llama_model_loader: - kv  24:                      tokenizer.ggml.tokens arr[str,151936]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  25:                  tokenizer.ggml.token_type arr[i32,151936]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  26:                      tokenizer.ggml.merges arr[str,151387]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  27:                tokenizer.ggml.eos_token_id u32              = 151645
llama_model_loader: - kv  28:            tokenizer.ggml.padding_token_id u32              = 151643
llama_model_loader: - kv  29:                tokenizer.ggml.bos_token_id u32              = 151643
llama_model_loader: - kv  30:               tokenizer.ggml.add_bos_token bool             = false
llama_model_loader: - kv  31:                    tokenizer.chat_template str              = {%- if tools %}\n    {{- '<|im_start|>...
llama_model_loader: - kv  32:               general.quantization_version u32              = 2
llama_model_loader: - kv  33:                          general.file_type u32              = 30
llama_model_loader: - type  f32:  121 tensors
llama_model_loader: - type q5_1:   24 tensors
llama_model_loader: - type q8_0:    1 tensors
llama_model_loader: - type q5_K:    3 tensors
llama_model_loader: - type iq4_nl:  120 tensors
llama_model_loader: - type iq4_xs:   21 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = IQ4_XS - 4.25 bpw
print_info: file size   = 329.49 MiB (5.59 BPW) 
load: special tokens cache size = 22
load: token to piece cache size = 0.9310 MB
print_info: arch             = qwen2
print_info: vocab_only       = 0
print_info: n_ctx_train      = 32768
print_info: n_embd           = 896
print_info: n_layer          = 24
print_info: n_head           = 14
print_info: n_head_kv        = 2
print_info: n_rot            = 64
print_info: n_swa            = 0
print_info: n_swa_pattern    = 1
print_info: n_embd_head_k    = 64
print_info: n_embd_head_v    = 64
print_info: n_gqa            = 7
print_info: n_embd_k_gqa     = 128
print_info: n_embd_v_gqa     = 128
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             = 4864
print_info: n_expert         = 0
print_info: n_expert_used    = 0
print_info: causal attn      = 1
print_info: pooling type     = 0
print_info: rope type        = 2
print_info: rope scaling     = linear
print_info: freq_base_train  = 1000000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn  = 32768
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       = 1B
print_info: model params     = 494.03 M
print_info: general.name     = Qwen2.5 0.5B Instruct
print_info: vocab type       = BPE
print_info: n_vocab          = 151936
print_info: n_merges         = 151387
print_info: BOS token        = 151643 '<|endoftext|>'
print_info: EOS token        = 151645 '<|im_end|>'
print_info: EOT token        = 151645 '<|im_end|>'
print_info: PAD token        = 151643 '<|endoftext|>'
print_info: LF token         = 198 'Ċ'
print_info: FIM PRE token    = 151659 '<|fim_prefix|>'
print_info: FIM SUF token    = 151661 '<|fim_suffix|>'
print_info: FIM MID token    = 151660 '<|fim_middle|>'
print_info: FIM PAD token    = 151662 '<|fim_pad|>'
print_info: FIM REP token    = 151663 '<|repo_name|>'
print_info: FIM SEP token    = 151664 '<|file_sep|>'
print_info: EOG token        = 151643 '<|endoftext|>'
print_info: EOG token        = 151645 '<|im_end|>'
print_info: EOG token        = 151662 '<|fim_pad|>'
print_info: EOG token        = 151663 '<|repo_name|>'
print_info: EOG token        = 151664 '<|file_sep|>'
print_info: max token length = 256
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: offloading 24 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 25/25 layers to GPU
load_tensors:        ROCm0 model buffer size =   329.52 MiB
load_tensors:   CPU_Mapped model buffer size =   137.94 MiB
...........................................
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 512
llama_context: n_ctx_per_seq = 512
llama_context: n_batch       = 512
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = 1
llama_context: freq_base     = 1000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_per_seq (512) < n_ctx_train (32768) -- the full capacity of the model will not be utilized
llama_context:  ROCm_Host  output buffer size =     0.58 MiB
init: kv_size = 512, offload = 1, type_k = 'q8_0', type_v = 'q8_0', n_layer = 24, can_shift = 1
init:      ROCm0 KV buffer size =     3.19 MiB
llama_context: KV self size  =    3.19 MiB, K (q8_0):    1.59 MiB, V (q8_0):    1.59 MiB
llama_context:      ROCm0 compute buffer size =   300.25 MiB
llama_context:  ROCm_Host compute buffer size =     4.51 MiB
llama_context: graph nodes  = 799
llama_context: graph splits = 50
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)


Repeat this sentence for 50 times: This is a test
This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This is a test. This

encoded   14 tokens in    0.012 seconds, speed: 1163.370 t/s
decoded  261 tokens in    7.355 seconds, speed:   35.484 t/s

n_draft   = 3
n_predict = 261
n_drafted = 195
n_accept  = 195
accept    = 100.000%

draft:

llama_perf_context_print:        load time =     161.07 ms
llama_perf_context_print: prompt eval time =    1006.70 ms /   142 tokens (    7.09 ms per token,   141.05 tokens per second)
llama_perf_context_print:        eval time =     943.49 ms /   131 runs   (    7.20 ms per token,   138.85 tokens per second)
llama_perf_context_print:       total time =    7369.47 ms /   273 tokens

target:

llama_perf_sampler_print:    sampling time =      12.24 ms /   261 runs   (    0.05 ms per token, 21327.01 tokens per second)
llama_perf_context_print:        load time =    6076.40 ms
llama_perf_context_print: prompt eval time =    5327.03 ms /   273 tokens (   19.51 ms per token,    51.25 tokens per second)
llama_perf_context_print:        eval time =      64.45 ms /     1 runs   (   64.45 ms per token,    15.52 tokens per second)
llama_perf_context_print:       total time =    7530.74 ms /   274 tokens

[ggerganov]

If you enable verbose output with -lv 1 can you figure out what is causing the discrepancy?

[hjc4869]

Collected the verbose logs from both. It seems that llama-speculative-simple keeps invoking common_speculative_gen_draft in a loop, while llama-server always do decoding batch, n_tokens = 1 first to generate a single token without speculative decoding, then call common_speculative_gen_draft to generate another 4, then goes back to generate a single token again.

Not sure if this is the expected behavior of llama-server as its code is a lot more complex than llama-speculative-simple, but the different behavior matches what I read from server.cpp and speculative-simple.cpp

Attached log files:
llama-server.log
llama-speculative-simple.log

[ggerganov]

Right, the server implementation can be improved to do it more efficiently as the simple example. For now it was easier to do it like this.