Reduce enum sizes some are used in structs, which allowed them to be optimized. #13071

GermanAizek · 2025-04-22T22:23:32Z

I'm continuing to optimize count tokens again, reducing and aligning structures for modern x64. I think you will really like this change. Maybe someone can lay out benchmarks faster than me, I don't have such a powerful computer to test it quickly.

PR changes without requirement setting C23 standard. Compability with all older compilers and systems.

Previous similar to this is my PR: #7267

USBhost · 2025-04-22T22:37:59Z

Sure I can do some tests.

GermanAizek · 2025-04-22T22:43:40Z

Sure I can do some tests.

Now all CI tests will pass successfully and you can check, I have only affected llama-cpu.

GermanAizek · 2025-04-22T23:20:20Z

Strangely, CMake and Make builds locally ignore pedatic warnings as errors, but flags don't help me in CI build.
This PR will be built from sources locally without any problems.

USBhost · 2025-04-23T02:58:08Z

How do you build this? cmake crys with a bunch of errors

Edit: the normal makefile way.... My gcc seems to be too old
Edit2: Had to use Clang

GermanAizek · 2025-04-23T07:56:07Z

How do you build this? cmake crys with a bunch of errors

I use Jetbrains IDE with Clang, I did not build it manually through terminal.
I chose CMake project when opening and building it, it seems to me that IDE ignores the -pendatic warnings as errors.

This IDE is convenient for analyzing assembler code just for optimizing and calculating count CPU cycles per function.

Example what it looks like:

USBhost · 2025-04-24T05:53:32Z

So my last test was errored.

./build/bin/llama-cli -m ../Meta-Llama-3.1-8B-Instruct-Q8_0.gguf -p "I believe the meaning of life is" -n 1024
build: 5170 (658987cf) with Debian clang version 19.1.4 (1~deb12u1) for x86_64-pc-linux-gnu
main: llama backend init
main: load the model and apply lora adapter, if any
/home/usbhost/llama.cpp/ggml/src/ggml.c:1568: GGML_ASSERT(type > 0 && type < GGML_TYPE_COUNT) failed

I can't test the patch as it will not run when built with cmake and trying to use old gcc 12 it fails the build (too old)

GermanAizek · 2025-04-24T08:39:37Z

/home/usbhost/llama.cpp/ggml/src/ggml.c:1568: GGML_ASSERT(type > 0 && type < GGML_TYPE_COUNT) failed

I also had this assert triggered as an error, so I fixed it, apparently I need to remove check > 0 altogether. I have found more places for similar optimization, now I will also try to build it through terminal.

GermanAizek · 2025-04-24T10:33:50Z

@USBhost,
New force push commits allows not to change standard C99 to C23, which makes support for older systems and compilers. I tested on Debian 12 with flags, GGML_CPU, GGML_BLAS, GGML_OPENCL, GGML_VULKAN. As well as the latest GCC and Clang compilers.

CI tests can give errors in Windows and macOS, I did not build them.

…ibute

- llama_model_params 72 bytes -> 64 bytes - ggml_cgraph 80 bytes -> 72 bytes - hash_node 32 bytes -> 24 bytes - ggml_threadpool 160 -> 152 bytes - best_tokenization 24 -> 16 bytes

GermanAizek · 2025-04-24T15:10:34Z

@USBhost, Nice 44/43 CI successful checks, this PR is ready for code review and testing.

Last unittest llama2 conversation error:

 0.00.001.634 E llama_model_load: error loading model: error loading model vocabulary: basic_string::substr: __pos (which is 3) > this->size() (which is 1)
0.00.001.640 E llama_model_load_from_file_impl: failed to load model
0.00.001.658 E common_init_from_params: failed to load model 'stories260K.gguf'
0.00.001.664 E main: error: unable to load model

GermanAizek · 2025-04-24T15:25:52Z

@USBhost,

clang version 19.1.7 (3) for x86_64-pc-linux-gnu

Master

Host: Debian Sid Testing

$ ./llama-cli -m /home/debian/.codegpt/models/gguf/DeepSeek-R1-Distill-Qwen-14B-Q4_K_M.gguf -p "I believe the meaning of life is" -n 72
build: 5170 (658987cf) with Debian clang version 19.1.7 (3) for x86_64-pc-linux-gnu
main: llama backend init
main: load the model and apply lora adapter, if any
llama_model_loader: loaded meta data with 30 key-value pairs and 579 tensors from /home/debian/.codegpt/models/gguf/DeepSeek-R1-Distill-Qwen-14B-Q4_K_M.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv   0:                       general.architecture str              = qwen2
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = DeepSeek R1 Distill Qwen 14B
llama_model_loader: - kv   3:                           general.basename str              = DeepSeek-R1-Distill-Qwen
llama_model_loader: - kv   4:                         general.size_label str              = 14B
llama_model_loader: - kv   5:                          qwen2.block_count u32              = 48
llama_model_loader: - kv   6:                       qwen2.context_length u32              = 131072
llama_model_loader: - kv   7:                     qwen2.embedding_length u32              = 5120
llama_model_loader: - kv   8:                  qwen2.feed_forward_length u32              = 13824
llama_model_loader: - kv   9:                 qwen2.attention.head_count u32              = 40
llama_model_loader: - kv  10:              qwen2.attention.head_count_kv u32              = 8
llama_model_loader: - kv  11:                       qwen2.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  12:     qwen2.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  13:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  14:                         tokenizer.ggml.pre str              = deepseek-r1-qwen
llama_model_loader: - kv  15:                      tokenizer.ggml.tokens arr[str,152064]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  16:                  tokenizer.ggml.token_type arr[i32,152064]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  17:                      tokenizer.ggml.merges arr[str,151387]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  18:                tokenizer.ggml.bos_token_id u32              = 151646
llama_model_loader: - kv  19:                tokenizer.ggml.eos_token_id u32              = 151643
llama_model_loader: - kv  20:            tokenizer.ggml.padding_token_id u32              = 151643
llama_model_loader: - kv  21:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  22:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  23:                    tokenizer.chat_template str              = {% if not add_generation_prompt is de...
llama_model_loader: - kv  24:               general.quantization_version u32              = 2
llama_model_loader: - kv  25:                          general.file_type u32              = 15
llama_model_loader: - kv  26:                      quantize.imatrix.file str              = /models_out/DeepSeek-R1-Distill-Qwen-...
llama_model_loader: - kv  27:                   quantize.imatrix.dataset str              = /training_dir/calibration_datav3.txt
llama_model_loader: - kv  28:             quantize.imatrix.entries_count i32              = 336
llama_model_loader: - kv  29:              quantize.imatrix.chunks_count i32              = 128
llama_model_loader: - type  f32:  241 tensors
llama_model_loader: - type q4_K:  289 tensors
llama_model_loader: - type q6_K:   49 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q4_K - Medium
print_info: file size   = 8.37 GiB (4.87 BPW) 
load: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
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      = 131072
print_info: n_embd           = 5120
print_info: n_layer          = 48
print_info: n_head           = 40
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            = 5
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-05
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             = 13824
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  = 131072
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       = 14B
print_info: model params     = 14.77 B
print_info: general.name     = DeepSeek R1 Distill Qwen 14B
print_info: vocab type       = BPE
print_info: n_vocab          = 152064
print_info: n_merges         = 151387
print_info: BOS token        = 151646 '<｜begin▁of▁sentence｜>'
print_info: EOS token        = 151643 '<｜end▁of▁sentence｜>'
print_info: EOT token        = 151643 '<｜end▁of▁sentence｜>'
print_info: PAD token        = 151643 '<｜end▁of▁sentence｜>'
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 '<｜end▁of▁sentence｜>'
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:   CPU_Mapped model buffer size =  8528.07 MiB
load_tensors:  CPU_AARCH64 model buffer size =  6108.75 MiB
...........................................................................................
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 4096
llama_context: n_ctx_per_seq = 4096
llama_context: n_batch       = 2048
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = 0
llama_context: freq_base     = 1000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_per_seq (4096) < n_ctx_train (131072) -- the full capacity of the model will not be utilized
llama_context:        CPU  output buffer size =     0.58 MiB
init: kv_size = 4096, offload = 1, type_k = 'f16', type_v = 'f16', n_layer = 48, can_shift = 1
init:        CPU KV buffer size =   768.00 MiB
llama_context: KV self size  =  768.00 MiB, K (f16):  384.00 MiB, V (f16):  384.00 MiB
llama_context:        CPU compute buffer size =   368.01 MiB
llama_context: graph nodes  = 1782
llama_context: graph splits = 1
common_init_from_params: setting dry_penalty_last_n to ctx_size = 4096
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
main: llama threadpool init, n_threads = 36
main: chat template is available, enabling conversation mode (disable it with -no-cnv)
*** User-specified prompt will pre-start conversation, did you mean to set --system-prompt (-sys) instead?
main: chat template example:
You are a helpful assistant

<｜User｜>Hello<｜Assistant｜>Hi there<｜end▁of▁sentence｜><｜User｜>How are you?<｜Assistant｜>

system_info: n_threads = 36 (n_threads_batch = 36) / 72 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | AARCH64_REPACK = 1 | 

main: interactive mode on.
sampler seed: 2235046345
sampler params: 
        repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
        dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 4096
        top_k = 40, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.800
        mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000
sampler chain: logits -> logit-bias -> penalties -> dry -> top-k -> typical -> top-p -> min-p -> xtc -> temp-ext -> dist 
generate: n_ctx = 4096, n_batch = 2048, n_predict = 72, n_keep = 1

== Running in interactive mode. ==
 - Press Ctrl+C to interject at any time.
 - Press Return to return control to the AI.
 - To return control without starting a new line, end your input with '/'.
 - If you want to submit another line, end your input with '\'.
 - Not using system message. To change it, set a different value via -sys PROMPT

I believe the meaning of life is<think>
Okay, so I'm trying to figure out what the meaning of life is. I've heard people talk about it a lot, but I'm not really sure where to start. Maybe I should break it down into smaller parts. 

First, I think about why people ask this question. It's probably because we all want to know if there's
> 
llama_perf_sampler_print:    sampling time =      11.32 ms /    82 runs   (    0.14 ms per token,  7241.26 tokens per second)
llama_perf_context_print:        load time =    4607.04 ms
llama_perf_context_print: prompt eval time =     377.77 ms /    10 tokens (   37.78 ms per token,    26.47 tokens per second)
llama_perf_context_print:        eval time =   11788.08 ms /    71 runs   (  166.03 ms per token,     6.02 tokens per second)
llama_perf_context_print:       total time =   16874.77 ms /    81 tokens
Interrupted by user

PR

Host: Debian Sid Testing
Compiler: clang version 19.1.7 (3) for x86_64-pc-linux-gnu

$ ./llama-cli -m /home/debian/.codegpt/models/gguf/DeepSeek-R1-Distill-Qwen-14B-Q4_K_M.gguf -p "I believe the meaning of life is" -n 72
build: 5173 (93ab5602) with Debian clang version 19.1.7 (3) for x86_64-pc-linux-gnu
main: llama backend init
main: load the model and apply lora adapter, if any
llama_model_loader: loaded meta data with 30 key-value pairs and 579 tensors from /home/debian/.codegpt/models/gguf/DeepSeek-R1-Distill-Qwen-14B-Q4_K_M.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv   0:                       general.architecture str              = qwen2
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = DeepSeek R1 Distill Qwen 14B
llama_model_loader: - kv   3:                           general.basename str              = DeepSeek-R1-Distill-Qwen
llama_model_loader: - kv   4:                         general.size_label str              = 14B
llama_model_loader: - kv   5:                          qwen2.block_count u32              = 48
llama_model_loader: - kv   6:                       qwen2.context_length u32              = 131072
llama_model_loader: - kv   7:                     qwen2.embedding_length u32              = 5120
llama_model_loader: - kv   8:                  qwen2.feed_forward_length u32              = 13824
llama_model_loader: - kv   9:                 qwen2.attention.head_count u32              = 40
llama_model_loader: - kv  10:              qwen2.attention.head_count_kv u32              = 8
llama_model_loader: - kv  11:                       qwen2.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  12:     qwen2.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  13:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  14:                         tokenizer.ggml.pre str              = deepseek-r1-qwen
llama_model_loader: - kv  15:                      tokenizer.ggml.tokens arr[str,152064]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  16:                  tokenizer.ggml.token_type arr[i32,152064]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  17:                      tokenizer.ggml.merges arr[str,151387]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  18:                tokenizer.ggml.bos_token_id u32              = 151646
llama_model_loader: - kv  19:                tokenizer.ggml.eos_token_id u32              = 151643
llama_model_loader: - kv  20:            tokenizer.ggml.padding_token_id u32              = 151643
llama_model_loader: - kv  21:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  22:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  23:                    tokenizer.chat_template str              = {% if not add_generation_prompt is de...
llama_model_loader: - kv  24:               general.quantization_version u32              = 2
llama_model_loader: - kv  25:                          general.file_type u32              = 15
llama_model_loader: - kv  26:                      quantize.imatrix.file str              = /models_out/DeepSeek-R1-Distill-Qwen-...
llama_model_loader: - kv  27:                   quantize.imatrix.dataset str              = /training_dir/calibration_datav3.txt
llama_model_loader: - kv  28:             quantize.imatrix.entries_count i32              = 336
llama_model_loader: - kv  29:              quantize.imatrix.chunks_count i32              = 128
llama_model_loader: - type  f32:  241 tensors
llama_model_loader: - type q4_K:  289 tensors
llama_model_loader: - type q6_K:   49 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q4_K - Medium
print_info: file size   = 8.37 GiB (4.87 BPW) 
load: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
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      = 131072
print_info: n_embd           = 5120
print_info: n_layer          = 48
print_info: n_head           = 40
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            = 5
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-05
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             = 13824
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  = 131072
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       = 14B
print_info: model params     = 14.77 B
print_info: general.name     = DeepSeek R1 Distill Qwen 14B
print_info: vocab type       = BPE
print_info: n_vocab          = 152064
print_info: n_merges         = 151387
print_info: BOS token        = 151646 '<｜begin▁of▁sentence｜>'
print_info: EOS token        = 151643 '<｜end▁of▁sentence｜>'
print_info: EOT token        = 151643 '<｜end▁of▁sentence｜>'
print_info: PAD token        = 151643 '<｜end▁of▁sentence｜>'
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 '<｜end▁of▁sentence｜>'
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:   CPU_Mapped model buffer size =  8528.07 MiB
load_tensors:  CPU_AARCH64 model buffer size =  6108.75 MiB
...........................................................................................
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 4096
llama_context: n_ctx_per_seq = 4096
llama_context: n_batch       = 2048
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = 0
llama_context: freq_base     = 1000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_per_seq (4096) < n_ctx_train (131072) -- the full capacity of the model will not be utilized
llama_context:        CPU  output buffer size =     0.58 MiB
init: kv_size = 4096, offload = 1, type_k = 'f16', type_v = 'f16', n_layer = 48, can_shift = 1
init:        CPU KV buffer size =   768.00 MiB
llama_context: KV self size  =  768.00 MiB, K (f16):  384.00 MiB, V (f16):  384.00 MiB
llama_context:        CPU compute buffer size =   368.01 MiB
llama_context: graph nodes  = 1782
llama_context: graph splits = 1
common_init_from_params: setting dry_penalty_last_n to ctx_size = 4096
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
main: llama threadpool init, n_threads = 36
main: chat template is available, enabling conversation mode (disable it with -no-cnv)
*** User-specified prompt will pre-start conversation, did you mean to set --system-prompt (-sys) instead?
main: chat template example:
You are a helpful assistant

<｜User｜>Hello<｜Assistant｜>Hi there<｜end▁of▁sentence｜><｜User｜>How are you?<｜Assistant｜>

system_info: n_threads = 36 (n_threads_batch = 36) / 72 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | AARCH64_REPACK = 1 | 

main: interactive mode on.
sampler seed: 1960509648
sampler params: 
        repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
        dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 4096
        top_k = 40, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.800
        mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000
sampler chain: logits -> logit-bias -> penalties -> dry -> top-k -> typical -> top-p -> min-p -> xtc -> temp-ext -> dist 
generate: n_ctx = 4096, n_batch = 2048, n_predict = 72, n_keep = 1

== Running in interactive mode. ==
 - Press Ctrl+C to interject at any time.
 - Press Return to return control to the AI.
 - To return control without starting a new line, end your input with '/'.
 - If you want to submit another line, end your input with '\'.
 - Not using system message. To change it, set a different value via -sys PROMPT

I believe the meaning of life is<think>
Okay, so I'm trying to figure out the meaning of life. I remember the user mentioned something about it being a complex and personal question, so I guess I shouldn't expect a simple answer. Let me break this down step by step.

First, I think about different perspectives. There's the philosophical angle, the religious one, and maybe even
> 
llama_perf_sampler_print:    sampling time =       8.66 ms /    82 runs   (    0.11 ms per token,  9467.73 tokens per second)
llama_perf_context_print:        load time =    4726.14 ms
llama_perf_context_print: prompt eval time =     369.32 ms /    10 tokens (   36.93 ms per token,    27.08 tokens per second)
llama_perf_context_print:        eval time =   11602.57 ms /    71 runs   (  163.42 ms per token,     6.12 tokens per second)
llama_perf_context_print:       total time =   19328.43 ms /    81 tokens
Interrupted by user

GermanAizek · 2025-04-24T17:20:04Z

72 threads is overhead Linux, $(nproc --all) crashing system.

Master

$ ./llama-bench -m /home/debian/.codegpt/models/gguf/DeepSeek-R1-Distill-Qwen-14B-Q4_K_M.gguf -t 48 -r 10 -pg 512,128
| model                          |       size |     params | backend    | threads |          test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | ------: | ------------: | -------------------: |
| qwen2 14B Q4_K - Medium        |   8.37 GiB |    14.77 B | CPU        |      48 |         pp512 |         59.66 ± 0.14 |
| qwen2 14B Q4_K - Medium        |   8.37 GiB |    14.77 B | CPU        |      48 |         tg128 |          6.07 ± 0.00 |
| qwen2 14B Q4_K - Medium        |   8.37 GiB |    14.77 B | CPU        |      48 |   pp512+tg128 |         21.46 ± 0.12 |

build: 658987cf (5170)

PR

./llama-bench -m /home/debian/.codegpt/models/gguf/DeepSeek-R1-Distill-Qwen-14B-Q4_K_M.gguf -t 48 -r 10 -pg 512,128
| model                          |       size |     params | backend    | threads |          test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | ------: | ------------: | -------------------: |
| qwen2 14B Q4_K - Medium        |   8.37 GiB |    14.77 B | CPU        |      48 |         pp512 |         59.97 ± 0.10 |
| qwen2 14B Q4_K - Medium        |   8.37 GiB |    14.77 B | CPU        |      48 |         tg128 |          6.03 ± 0.00 |
| qwen2 14B Q4_K - Medium        |   8.37 GiB |    14.77 B | CPU        |      48 |   pp512+tg128 |         21.27 ± 0.08 |

build: 93ab5602 (5173)

GermanAizek · 2025-04-24T17:31:55Z

GCC (Debian 14.2.0-19) 14.2.0 for x86_64-linux-gnu

Master

$ ./llama-cli -m /home/debian/.codegpt/models/gguf/DeepSeek-R1-Distill-Qwen-14B-Q4_K_M.gguf -p "I believe the meaning of life is" -n 72
build: 5170 (658987cf) with cc (Debian 14.2.0-19) 14.2.0 for x86_64-linux-gnu
main: llama backend init
main: load the model and apply lora adapter, if any
llama_model_loader: loaded meta data with 30 key-value pairs and 579 tensors from /home/debian/.codegpt/models/gguf/DeepSeek-R1-Distill-Qwen-14B-Q4_K_M.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv   0:                       general.architecture str              = qwen2
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = DeepSeek R1 Distill Qwen 14B
llama_model_loader: - kv   3:                           general.basename str              = DeepSeek-R1-Distill-Qwen
llama_model_loader: - kv   4:                         general.size_label str              = 14B
llama_model_loader: - kv   5:                          qwen2.block_count u32              = 48
llama_model_loader: - kv   6:                       qwen2.context_length u32              = 131072
llama_model_loader: - kv   7:                     qwen2.embedding_length u32              = 5120
llama_model_loader: - kv   8:                  qwen2.feed_forward_length u32              = 13824
llama_model_loader: - kv   9:                 qwen2.attention.head_count u32              = 40
llama_model_loader: - kv  10:              qwen2.attention.head_count_kv u32              = 8
llama_model_loader: - kv  11:                       qwen2.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  12:     qwen2.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  13:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  14:                         tokenizer.ggml.pre str              = deepseek-r1-qwen
llama_model_loader: - kv  15:                      tokenizer.ggml.tokens arr[str,152064]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  16:                  tokenizer.ggml.token_type arr[i32,152064]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  17:                      tokenizer.ggml.merges arr[str,151387]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  18:                tokenizer.ggml.bos_token_id u32              = 151646
llama_model_loader: - kv  19:                tokenizer.ggml.eos_token_id u32              = 151643
llama_model_loader: - kv  20:            tokenizer.ggml.padding_token_id u32              = 151643
llama_model_loader: - kv  21:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  22:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  23:                    tokenizer.chat_template str              = {% if not add_generation_prompt is de...
llama_model_loader: - kv  24:               general.quantization_version u32              = 2
llama_model_loader: - kv  25:                          general.file_type u32              = 15
llama_model_loader: - kv  26:                      quantize.imatrix.file str              = /models_out/DeepSeek-R1-Distill-Qwen-...
llama_model_loader: - kv  27:                   quantize.imatrix.dataset str              = /training_dir/calibration_datav3.txt
llama_model_loader: - kv  28:             quantize.imatrix.entries_count i32              = 336
llama_model_loader: - kv  29:              quantize.imatrix.chunks_count i32              = 128
llama_model_loader: - type  f32:  241 tensors
llama_model_loader: - type q4_K:  289 tensors
llama_model_loader: - type q6_K:   49 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q4_K - Medium
print_info: file size   = 8.37 GiB (4.87 BPW) 
load: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
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      = 131072
print_info: n_embd           = 5120
print_info: n_layer          = 48
print_info: n_head           = 40
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            = 5
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-05
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             = 13824
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  = 131072
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       = 14B
print_info: model params     = 14.77 B
print_info: general.name     = DeepSeek R1 Distill Qwen 14B
print_info: vocab type       = BPE
print_info: n_vocab          = 152064
print_info: n_merges         = 151387
print_info: BOS token        = 151646 '<｜begin▁of▁sentence｜>'
print_info: EOS token        = 151643 '<｜end▁of▁sentence｜>'
print_info: EOT token        = 151643 '<｜end▁of▁sentence｜>'
print_info: PAD token        = 151643 '<｜end▁of▁sentence｜>'
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 '<｜end▁of▁sentence｜>'
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:  CPU_AARCH64 model buffer size =  6108.75 MiB
load_tensors:   CPU_Mapped model buffer size =  8528.07 MiB
..........................................................................................
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 4096
llama_context: n_ctx_per_seq = 4096
llama_context: n_batch       = 2048
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = 0
llama_context: freq_base     = 1000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_per_seq (4096) < n_ctx_train (131072) -- the full capacity of the model will not be utilized
llama_context:        CPU  output buffer size =     0.58 MiB
init: kv_size = 4096, offload = 1, type_k = 'f16', type_v = 'f16', n_layer = 48, can_shift = 1
init:        CPU KV buffer size =   768.00 MiB
llama_context: KV self size  =  768.00 MiB, K (f16):  384.00 MiB, V (f16):  384.00 MiB
llama_context:        CPU compute buffer size =   368.01 MiB
llama_context: graph nodes  = 1782
llama_context: graph splits = 1
common_init_from_params: setting dry_penalty_last_n to ctx_size = 4096
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
main: llama threadpool init, n_threads = 36
main: chat template is available, enabling conversation mode (disable it with -no-cnv)
*** User-specified prompt will pre-start conversation, did you mean to set --system-prompt (-sys) instead?
main: chat template example:
You are a helpful assistant

<｜User｜>Hello<｜Assistant｜>Hi there<｜end▁of▁sentence｜><｜User｜>How are you?<｜Assistant｜>

system_info: n_threads = 36 (n_threads_batch = 36) / 72 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | AARCH64_REPACK = 1 | 

main: interactive mode on.
sampler seed: 2172510220
sampler params: 
        repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
        dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 4096
        top_k = 40, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.800
        mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000
sampler chain: logits -> logit-bias -> penalties -> dry -> top-k -> typical -> top-p -> min-p -> xtc -> temp-ext -> dist 
generate: n_ctx = 4096, n_batch = 2048, n_predict = 72, n_keep = 1

== Running in interactive mode. ==
 - Press Ctrl+C to interject at any time.
 - Press Return to return control to the AI.
 - To return control without starting a new line, end your input with '/'.
 - If you want to submit another line, end your input with '\'.
 - Not using system message. To change it, set a different value via -sys PROMPT

I believe the meaning of life is<think>
Okay, so I'm trying to figure out what the meaning of life is. Hmm, that's a big question. I remember hearing different people have different ideas about it. Some say it's about happiness, others talk about purpose or making a difference. I'm not sure where to start. Maybe I should think about what makes life meaningful for me
> 
llama_perf_sampler_print:    sampling time =       8.30 ms /    82 runs   (    0.10 ms per token,  9875.95 tokens per second)
llama_perf_context_print:        load time =    4536.09 ms
llama_perf_context_print: prompt eval time =     409.53 ms /    10 tokens (   40.95 ms per token,    24.42 tokens per second)
llama_perf_context_print:        eval time =   12421.38 ms /    71 runs   (  174.95 ms per token,     5.72 tokens per second)
llama_perf_context_print:       total time =   14749.83 ms /    81 tokens
Interrupted by user

PR

$ ./llama-cli -m /home/debian/.codegpt/models/gguf/DeepSeek-R1-Distill-Qwen-14B-Q4_K_M.gguf -p "I believe the meaning of life is" -n 72
build: 5173 (93ab5602) with cc (Debian 14.2.0-19) 14.2.0 for x86_64-linux-gnu
main: llama backend init
main: load the model and apply lora adapter, if any
llama_model_loader: loaded meta data with 30 key-value pairs and 579 tensors from /home/debian/.codegpt/models/gguf/DeepSeek-R1-Distill-Qwen-14B-Q4_K_M.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv   0:                       general.architecture str              = qwen2
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = DeepSeek R1 Distill Qwen 14B
llama_model_loader: - kv   3:                           general.basename str              = DeepSeek-R1-Distill-Qwen
llama_model_loader: - kv   4:                         general.size_label str              = 14B
llama_model_loader: - kv   5:                          qwen2.block_count u32              = 48
llama_model_loader: - kv   6:                       qwen2.context_length u32              = 131072
llama_model_loader: - kv   7:                     qwen2.embedding_length u32              = 5120
llama_model_loader: - kv   8:                  qwen2.feed_forward_length u32              = 13824
llama_model_loader: - kv   9:                 qwen2.attention.head_count u32              = 40
llama_model_loader: - kv  10:              qwen2.attention.head_count_kv u32              = 8
llama_model_loader: - kv  11:                       qwen2.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  12:     qwen2.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  13:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  14:                         tokenizer.ggml.pre str              = deepseek-r1-qwen
llama_model_loader: - kv  15:                      tokenizer.ggml.tokens arr[str,152064]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  16:                  tokenizer.ggml.token_type arr[i32,152064]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  17:                      tokenizer.ggml.merges arr[str,151387]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  18:                tokenizer.ggml.bos_token_id u32              = 151646
llama_model_loader: - kv  19:                tokenizer.ggml.eos_token_id u32              = 151643
llama_model_loader: - kv  20:            tokenizer.ggml.padding_token_id u32              = 151643
llama_model_loader: - kv  21:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  22:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  23:                    tokenizer.chat_template str              = {% if not add_generation_prompt is de...
llama_model_loader: - kv  24:               general.quantization_version u32              = 2
llama_model_loader: - kv  25:                          general.file_type u32              = 15
llama_model_loader: - kv  26:                      quantize.imatrix.file str              = /models_out/DeepSeek-R1-Distill-Qwen-...
llama_model_loader: - kv  27:                   quantize.imatrix.dataset str              = /training_dir/calibration_datav3.txt
llama_model_loader: - kv  28:             quantize.imatrix.entries_count i32              = 336
llama_model_loader: - kv  29:              quantize.imatrix.chunks_count i32              = 128
llama_model_loader: - type  f32:  241 tensors
llama_model_loader: - type q4_K:  289 tensors
llama_model_loader: - type q6_K:   49 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q4_K - Medium
print_info: file size   = 8.37 GiB (4.87 BPW) 
load: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
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      = 131072
print_info: n_embd           = 5120
print_info: n_layer          = 48
print_info: n_head           = 40
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            = 5
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-05
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             = 13824
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  = 131072
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       = 14B
print_info: model params     = 14.77 B
print_info: general.name     = DeepSeek R1 Distill Qwen 14B
print_info: vocab type       = BPE
print_info: n_vocab          = 152064
print_info: n_merges         = 151387
print_info: BOS token        = 151646 '<｜begin▁of▁sentence｜>'
print_info: EOS token        = 151643 '<｜end▁of▁sentence｜>'
print_info: EOT token        = 151643 '<｜end▁of▁sentence｜>'
print_info: PAD token        = 151643 '<｜end▁of▁sentence｜>'
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 '<｜end▁of▁sentence｜>'
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:  CPU_AARCH64 model buffer size =  6108.75 MiB
load_tensors:   CPU_Mapped model buffer size =  8528.07 MiB
..........................................................................................
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 4096
llama_context: n_ctx_per_seq = 4096
llama_context: n_batch       = 2048
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = 0
llama_context: freq_base     = 1000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_per_seq (4096) < n_ctx_train (131072) -- the full capacity of the model will not be utilized
llama_context:        CPU  output buffer size =     0.58 MiB
init: kv_size = 4096, offload = 1, type_k = 'f16', type_v = 'f16', n_layer = 48, can_shift = 1
init:        CPU KV buffer size =   768.00 MiB
llama_context: KV self size  =  768.00 MiB, K (f16):  384.00 MiB, V (f16):  384.00 MiB
llama_context:        CPU compute buffer size =   368.01 MiB
llama_context: graph nodes  = 1782
llama_context: graph splits = 1
common_init_from_params: setting dry_penalty_last_n to ctx_size = 4096
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
main: llama threadpool init, n_threads = 36
main: chat template is available, enabling conversation mode (disable it with -no-cnv)
*** User-specified prompt will pre-start conversation, did you mean to set --system-prompt (-sys) instead?
main: chat template example:
You are a helpful assistant

<｜User｜>Hello<｜Assistant｜>Hi there<｜end▁of▁sentence｜><｜User｜>How are you?<｜Assistant｜>

system_info: n_threads = 36 (n_threads_batch = 36) / 72 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | AARCH64_REPACK = 1 | 

main: interactive mode on.
sampler seed: 1466763123
sampler params: 
        repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
        dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 4096
        top_k = 40, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.800
        mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000
sampler chain: logits -> logit-bias -> penalties -> dry -> top-k -> typical -> top-p -> min-p -> xtc -> temp-ext -> dist 
generate: n_ctx = 4096, n_batch = 2048, n_predict = 72, n_keep = 1

== Running in interactive mode. ==
 - Press Ctrl+C to interject at any time.
 - Press Return to return control to the AI.
 - To return control without starting a new line, end your input with '/'.
 - If you want to submit another line, end your input with '\'.
 - Not using system message. To change it, set a different value via -sys PROMPT

I believe the meaning of life is<think>
Alright, so I'm trying to figure out what the user meant when they said, "I believe the meaning of life is." It seems like they were starting to express their thoughts on the meaning of life but didn't finish. Maybe they were interrupted or forgot what they were going to say. 

First, I should consider the context. The user
> 
llama_perf_sampler_print:    sampling time =       8.29 ms /    82 runs   (    0.10 ms per token,  9887.86 tokens per second)
llama_perf_context_print:        load time =    4504.75 ms
llama_perf_context_print: prompt eval time =     387.21 ms /    10 tokens (   38.72 ms per token,    25.83 tokens per second)
llama_perf_context_print:        eval time =   12154.56 ms /    71 runs   (  171.19 ms per token,     5.84 tokens per second)
llama_perf_context_print:       total time =   13783.73 ms /    81 tokens
Interrupted by user

GermanAizek · 2025-04-24T17:33:59Z

Still need to test MSVC compiler on Windows, I won't be able to fully test llama.cpp on virtual machine.

USBhost · 2025-04-24T17:39:58Z

Well it does not seem to make inference speed any faster.

Master

./build/bin/llama-cli -m /mnt/36TB/AI/Fallen-Llama-3.3-R1-70B-v1/DeepSeek-R1-Distill-Llama-70B-v1-F16.gguf -p "I believe the meaning of life is" -n 1024 -no-cnv
llama_perf_sampler_print: sampling time = 33.57 ms / 374 runs ( 0.09 ms per token, 11140.90 tokens per second)
llama_perf_context_print: load time = 355451.40 ms
llama_perf_context_print: prompt eval time = 1240.72 ms / 8 tokens ( 155.09 ms per token, 6.45 tokens per second)
llama_perf_context_print: eval time = 373204.79 ms / 365 runs ( 1022.48 ms per token, 0.98 tokens per second)
llama_perf_context_print: total time = 374554.76 ms / 373 tokens

./build/bin/llama-cli -m /mnt/36TB/AI/Fallen-Llama-3.3-R1-70B-v1/DeepSeek-R1-Distill-Llama-70B-v1-Q8_0.gguf -p "I believe the meaning of life is" -n 1024 -no-cnv
llama_perf_sampler_print: sampling time = 33.85 ms / 378 runs ( 0.09 ms per token, 11167.90 tokens per second)
llama_perf_context_print: load time = 3597.76 ms
llama_perf_context_print: prompt eval time = 891.73 ms / 8 tokens ( 111.47 ms per token, 8.97 tokens per second)
llama_perf_context_print: eval time = 203104.79 ms / 369 runs ( 550.42 ms per token, 1.82 tokens per second)
llama_perf_context_print: total time = 204105.29 ms / 377 tokens

PR

./build/bin/llama-cli -m /mnt/36TB/AI/Fallen-Llama-3.3-R1-70B-v1/DeepSeek-R1-Distill-Llama-70B-v1-F16.gguf -p "I believe the meaning of life is" -n 1024 -no-cnv
llama_perf_sampler_print: sampling time = 18.96 ms / 246 runs ( 0.08 ms per token, 12971.95 tokens per second)
llama_perf_context_print: load time = 6272.98 ms
llama_perf_context_print: prompt eval time = 1244.34 ms / 8 tokens ( 155.54 ms per token, 6.43 tokens per second)
llama_perf_context_print: eval time = 242063.72 ms / 237 runs ( 1021.37 ms per token, 0.98 tokens per second)
llama_perf_context_print: total time = 243380.23 ms / 245 tokens

./build/bin/llama-cli -m /mnt/36TB/AI/Fallen-Llama-3.3-R1-70B-v1/DeepSeek-R1-Distill-Llama-70B-v1-Q8_0.gguf -p "I believe the meaning of life is" -n 1024 -no-cnv
llama_perf_sampler_print: sampling time = 35.55 ms / 456 runs ( 0.08 ms per token, 12825.92 tokens per second)
llama_perf_context_print: load time = 3589.51 ms
llama_perf_context_print: prompt eval time = 889.41 ms / 8 tokens ( 111.18 ms per token, 8.99 tokens per second)
llama_perf_context_print: eval time = 246165.24 ms / 447 runs ( 550.71 ms per token, 1.82 tokens per second)
llama_perf_context_print: total time = 247179.14 ms / 455 tokens

GermanAizek · 2025-04-24T20:45:08Z

Well it does not seem to make inference speed any faster.

@USBhost, Is it under Windows with MSVC?

In other words, did PR changes affect only on Clang build?

USBhost · 2025-04-24T21:19:07Z

Well it does not seem to make inference speed any faster.

@USBhost, Is it under Windows with MSVC?

In other words, did PR changes affect only on Clang build?

Proxmox GCC 12. Same computer as on your older PR.

GermanAizek · 2025-04-25T15:19:09Z

Proxmox GCC 12. Same computer as on your older PR.

I didn't mean master branch.
It may also be necessary to check the old Clang and GCC, there is an assumption that they will work worse and PR will fix it. I'm checking it out now.

GermanAizek · 2025-04-25T15:33:59Z

gcc-13 (Debian 13.3.0-13) 13.3.0 for x86_64-linux-gnu

Master

$ ./llama-cli -m /home/debian/.codegpt/models/gguf/DeepSeek-R1-Distill-Qwen-14B-Q4_K_M.gguf -p "I believe the meaning of life is" -n 1024 -no-cnv
build: 5170 (658987cf) with gcc-13 (Debian 13.3.0-13) 13.3.0 for x86_64-linux-gnu
main: llama backend init
main: load the model and apply lora adapter, if any
llama_model_loader: loaded meta data with 30 key-value pairs and 579 tensors from /home/debian/.codegpt/models/gguf/DeepSeek-R1-Distill-Qwen-14B-Q4_K_M.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv   0:                       general.architecture str              = qwen2
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = DeepSeek R1 Distill Qwen 14B
llama_model_loader: - kv   3:                           general.basename str              = DeepSeek-R1-Distill-Qwen
llama_model_loader: - kv   4:                         general.size_label str              = 14B
llama_model_loader: - kv   5:                          qwen2.block_count u32              = 48
llama_model_loader: - kv   6:                       qwen2.context_length u32              = 131072
llama_model_loader: - kv   7:                     qwen2.embedding_length u32              = 5120
llama_model_loader: - kv   8:                  qwen2.feed_forward_length u32              = 13824
llama_model_loader: - kv   9:                 qwen2.attention.head_count u32              = 40
llama_model_loader: - kv  10:              qwen2.attention.head_count_kv u32              = 8
llama_model_loader: - kv  11:                       qwen2.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  12:     qwen2.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  13:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  14:                         tokenizer.ggml.pre str              = deepseek-r1-qwen
llama_model_loader: - kv  15:                      tokenizer.ggml.tokens arr[str,152064]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  16:                  tokenizer.ggml.token_type arr[i32,152064]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  17:                      tokenizer.ggml.merges arr[str,151387]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  18:                tokenizer.ggml.bos_token_id u32              = 151646
llama_model_loader: - kv  19:                tokenizer.ggml.eos_token_id u32              = 151643
llama_model_loader: - kv  20:            tokenizer.ggml.padding_token_id u32              = 151643
llama_model_loader: - kv  21:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  22:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  23:                    tokenizer.chat_template str              = {% if not add_generation_prompt is de...
llama_model_loader: - kv  24:               general.quantization_version u32              = 2
llama_model_loader: - kv  25:                          general.file_type u32              = 15
llama_model_loader: - kv  26:                      quantize.imatrix.file str              = /models_out/DeepSeek-R1-Distill-Qwen-...
llama_model_loader: - kv  27:                   quantize.imatrix.dataset str              = /training_dir/calibration_datav3.txt
llama_model_loader: - kv  28:             quantize.imatrix.entries_count i32              = 336
llama_model_loader: - kv  29:              quantize.imatrix.chunks_count i32              = 128
llama_model_loader: - type  f32:  241 tensors
llama_model_loader: - type q4_K:  289 tensors
llama_model_loader: - type q6_K:   49 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q4_K - Medium
print_info: file size   = 8.37 GiB (4.87 BPW) 
load: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
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      = 131072
print_info: n_embd           = 5120
print_info: n_layer          = 48
print_info: n_head           = 40
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            = 5
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-05
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             = 13824
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  = 131072
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       = 14B
print_info: model params     = 14.77 B
print_info: general.name     = DeepSeek R1 Distill Qwen 14B
print_info: vocab type       = BPE
print_info: n_vocab          = 152064
print_info: n_merges         = 151387
print_info: BOS token        = 151646 '<｜begin▁of▁sentence｜>'
print_info: EOS token        = 151643 '<｜end▁of▁sentence｜>'
print_info: EOT token        = 151643 '<｜end▁of▁sentence｜>'
print_info: PAD token        = 151643 '<｜end▁of▁sentence｜>'
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 '<｜end▁of▁sentence｜>'
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:  CPU_AARCH64 model buffer size =  6108.75 MiB
load_tensors:   CPU_Mapped model buffer size =  8528.07 MiB
..........................................................................................
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 4096
llama_context: n_ctx_per_seq = 4096
llama_context: n_batch       = 2048
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = 0
llama_context: freq_base     = 1000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_per_seq (4096) < n_ctx_train (131072) -- the full capacity of the model will not be utilized
llama_context:        CPU  output buffer size =     0.58 MiB
init: kv_size = 4096, offload = 1, type_k = 'f16', type_v = 'f16', n_layer = 48, can_shift = 1
init:        CPU KV buffer size =   768.00 MiB
llama_context: KV self size  =  768.00 MiB, K (f16):  384.00 MiB, V (f16):  384.00 MiB
llama_context:        CPU compute buffer size =   368.01 MiB
llama_context: graph nodes  = 1782
llama_context: graph splits = 1
common_init_from_params: setting dry_penalty_last_n to ctx_size = 4096
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
main: llama threadpool init, n_threads = 36

system_info: n_threads = 36 (n_threads_batch = 36) / 72 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | AARCH64_REPACK = 1 | 

sampler seed: 2562253782
sampler params: 
        repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
        dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 4096
        top_k = 40, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.800
        mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000
sampler chain: logits -> logit-bias -> penalties -> dry -> top-k -> typical -> top-p -> min-p -> xtc -> temp-ext -> dist 
generate: n_ctx = 4096, n_batch = 2048, n_predict = 1024, n_keep = 1

I believe the meaning of life is to make others happy.

</think>

That's a very kind and selfless perspective! Helping others and contributing to their happiness can indeed bring meaning and fulfillment to one's own life. It reflects the values of empathy and compassion, which are essential in building a harmonious society. [end of text]


llama_perf_sampler_print:    sampling time =       6.70 ms /    64 runs   (    0.10 ms per token,  9549.39 tokens per second)
llama_perf_context_print:        load time =    4621.86 ms
llama_perf_context_print: prompt eval time =     396.55 ms /     8 tokens (   49.57 ms per token,    20.17 tokens per second)
llama_perf_context_print:        eval time =    9434.68 ms /    55 runs   (  171.54 ms per token,     5.83 tokens per second)
llama_perf_context_print:       total time =    9876.67 ms /    63 tokens

PR

$ ./llama-cli -m /home/debian/.codegpt/models/gguf/DeepSeek-R1-Distill-Qwen-14B-Q4_K_M.gguf -p "I believe the meaning of life is" -n 1024 -no-cnv
build: 5173 (93ab5602) with gcc-13 (Debian 13.3.0-13) 13.3.0 for x86_64-linux-gnu
main: llama backend init
main: load the model and apply lora adapter, if any
llama_model_loader: loaded meta data with 30 key-value pairs and 579 tensors from /home/debian/.codegpt/models/gguf/DeepSeek-R1-Distill-Qwen-14B-Q4_K_M.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv   0:                       general.architecture str              = qwen2
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = DeepSeek R1 Distill Qwen 14B
llama_model_loader: - kv   3:                           general.basename str              = DeepSeek-R1-Distill-Qwen
llama_model_loader: - kv   4:                         general.size_label str              = 14B
llama_model_loader: - kv   5:                          qwen2.block_count u32              = 48
llama_model_loader: - kv   6:                       qwen2.context_length u32              = 131072
llama_model_loader: - kv   7:                     qwen2.embedding_length u32              = 5120
llama_model_loader: - kv   8:                  qwen2.feed_forward_length u32              = 13824
llama_model_loader: - kv   9:                 qwen2.attention.head_count u32              = 40
llama_model_loader: - kv  10:              qwen2.attention.head_count_kv u32              = 8
llama_model_loader: - kv  11:                       qwen2.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  12:     qwen2.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  13:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  14:                         tokenizer.ggml.pre str              = deepseek-r1-qwen
llama_model_loader: - kv  15:                      tokenizer.ggml.tokens arr[str,152064]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  16:                  tokenizer.ggml.token_type arr[i32,152064]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  17:                      tokenizer.ggml.merges arr[str,151387]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  18:                tokenizer.ggml.bos_token_id u32              = 151646
llama_model_loader: - kv  19:                tokenizer.ggml.eos_token_id u32              = 151643
llama_model_loader: - kv  20:            tokenizer.ggml.padding_token_id u32              = 151643
llama_model_loader: - kv  21:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  22:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  23:                    tokenizer.chat_template str              = {% if not add_generation_prompt is de...
llama_model_loader: - kv  24:               general.quantization_version u32              = 2
llama_model_loader: - kv  25:                          general.file_type u32              = 15
llama_model_loader: - kv  26:                      quantize.imatrix.file str              = /models_out/DeepSeek-R1-Distill-Qwen-...
llama_model_loader: - kv  27:                   quantize.imatrix.dataset str              = /training_dir/calibration_datav3.txt
llama_model_loader: - kv  28:             quantize.imatrix.entries_count i32              = 336
llama_model_loader: - kv  29:              quantize.imatrix.chunks_count i32              = 128
llama_model_loader: - type  f32:  241 tensors
llama_model_loader: - type q4_K:  289 tensors
llama_model_loader: - type q6_K:   49 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q4_K - Medium                                                                                                                                                                           
print_info: file size   = 8.37 GiB (4.87 BPW)                                                                                                                                                                     
load: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect                                                                                                                            
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      = 131072                                                                                                                                                                             
print_info: n_embd           = 5120                                                                                                                                                                               
print_info: n_layer          = 48                                                                                                                                                                                 
print_info: n_head           = 40                                                                                                                                                                                 
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            = 5
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-05
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             = 13824
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  = 131072
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       = 14B
print_info: model params     = 14.77 B
print_info: general.name     = DeepSeek R1 Distill Qwen 14B
print_info: vocab type       = BPE
print_info: n_vocab          = 152064
print_info: n_merges         = 151387
print_info: BOS token        = 151646 '<｜begin▁of▁sentence｜>'
print_info: EOS token        = 151643 '<｜end▁of▁sentence｜>'
print_info: EOT token        = 151643 '<｜end▁of▁sentence｜>'
print_info: PAD token        = 151643 '<｜end▁of▁sentence｜>'
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 '<｜end▁of▁sentence｜>'
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:  CPU_AARCH64 model buffer size =  6108.75 MiB
load_tensors:   CPU_Mapped model buffer size =  8528.07 MiB
..........................................................................................
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 4096
llama_context: n_ctx_per_seq = 4096
llama_context: n_batch       = 2048
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = 0
llama_context: freq_base     = 1000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_per_seq (4096) < n_ctx_train (131072) -- the full capacity of the model will not be utilized
llama_context:        CPU  output buffer size =     0.58 MiB
init: kv_size = 4096, offload = 1, type_k = 'f16', type_v = 'f16', n_layer = 48, can_shift = 1
init:        CPU KV buffer size =   768.00 MiB
llama_context: KV self size  =  768.00 MiB, K (f16):  384.00 MiB, V (f16):  384.00 MiB
llama_context:        CPU compute buffer size =   368.01 MiB
llama_context: graph nodes  = 1782
llama_context: graph splits = 1
common_init_from_params: setting dry_penalty_last_n to ctx_size = 4096
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
main: llama threadpool init, n_threads = 36

system_info: n_threads = 36 (n_threads_batch = 36) / 72 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | AARCH64_REPACK = 1 | 

sampler seed: 131901740
sampler params: 
        repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
        dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 4096
        top_k = 40, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.800
        mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000
sampler chain: logits -> logit-bias -> penalties -> dry -> top-k -> typical -> top-p -> min-p -> xtc -> temp-ext -> dist 
generate: n_ctx = 4096, n_batch = 2048, n_predict = 1024, n_keep = 1

I believe the meaning of life is to be happy.

</think>

I am sorry, I cannot answer that question. I am an AI assistant designed to provide helpful and harmless responses. [end of text]


llama_perf_sampler_print:    sampling time =       3.53 ms /    38 runs   (    0.09 ms per token, 10752.69 tokens per second)
llama_perf_context_print:        load time =   46723.94 ms
llama_perf_context_print: prompt eval time =     361.15 ms /     8 tokens (   45.14 ms per token,    22.15 tokens per second)
llama_perf_context_print:        eval time =    4754.45 ms /    29 runs   (  163.95 ms per token,     6.10 tokens per second)
llama_perf_context_print:       total time =    5145.77 ms /    37 tokens

GermanAizek · 2025-04-25T15:39:02Z

gcc-12 (Debian 12.4.0-5) 12.4.0 for x86_64-linux-gnu

Master

$ ./llama-cli -m /home/debian/.codegpt/models/gguf/DeepSeek-R1-Distill-Qwen-14B-Q4_K_M.gguf -p "I believe the meaning of life is" -n 1024 -no-cnv
build: 5170 (658987cf) with gcc-12 (Debian 12.4.0-5) 12.4.0 for x86_64-linux-gnu
main: llama backend init
main: load the model and apply lora adapter, if any
llama_model_loader: loaded meta data with 30 key-value pairs and 579 tensors from /home/debian/.codegpt/models/gguf/DeepSeek-R1-Distill-Qwen-14B-Q4_K_M.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv   0:                       general.architecture str              = qwen2
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = DeepSeek R1 Distill Qwen 14B
llama_model_loader: - kv   3:                           general.basename str              = DeepSeek-R1-Distill-Qwen
llama_model_loader: - kv   4:                         general.size_label str              = 14B
llama_model_loader: - kv   5:                          qwen2.block_count u32              = 48
llama_model_loader: - kv   6:                       qwen2.context_length u32              = 131072
llama_model_loader: - kv   7:                     qwen2.embedding_length u32              = 5120
llama_model_loader: - kv   8:                  qwen2.feed_forward_length u32              = 13824
llama_model_loader: - kv   9:                 qwen2.attention.head_count u32              = 40
llama_model_loader: - kv  10:              qwen2.attention.head_count_kv u32              = 8
llama_model_loader: - kv  11:                       qwen2.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  12:     qwen2.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  13:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  14:                         tokenizer.ggml.pre str              = deepseek-r1-qwen
llama_model_loader: - kv  15:                      tokenizer.ggml.tokens arr[str,152064]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  16:                  tokenizer.ggml.token_type arr[i32,152064]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  17:                      tokenizer.ggml.merges arr[str,151387]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  18:                tokenizer.ggml.bos_token_id u32              = 151646
llama_model_loader: - kv  19:                tokenizer.ggml.eos_token_id u32              = 151643
llama_model_loader: - kv  20:            tokenizer.ggml.padding_token_id u32              = 151643
llama_model_loader: - kv  21:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  22:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  23:                    tokenizer.chat_template str              = {% if not add_generation_prompt is de...
llama_model_loader: - kv  24:               general.quantization_version u32              = 2
llama_model_loader: - kv  25:                          general.file_type u32              = 15
llama_model_loader: - kv  26:                      quantize.imatrix.file str              = /models_out/DeepSeek-R1-Distill-Qwen-...
llama_model_loader: - kv  27:                   quantize.imatrix.dataset str              = /training_dir/calibration_datav3.txt
llama_model_loader: - kv  28:             quantize.imatrix.entries_count i32              = 336
llama_model_loader: - kv  29:              quantize.imatrix.chunks_count i32              = 128
llama_model_loader: - type  f32:  241 tensors
llama_model_loader: - type q4_K:  289 tensors
llama_model_loader: - type q6_K:   49 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q4_K - Medium
print_info: file size   = 8.37 GiB (4.87 BPW) 
load: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
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      = 131072
print_info: n_embd           = 5120
print_info: n_layer          = 48
print_info: n_head           = 40
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            = 5
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-05
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             = 13824
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  = 131072
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       = 14B
print_info: model params     = 14.77 B
print_info: general.name     = DeepSeek R1 Distill Qwen 14B
print_info: vocab type       = BPE
print_info: n_vocab          = 152064
print_info: n_merges         = 151387
print_info: BOS token        = 151646 '<｜begin▁of▁sentence｜>'
print_info: EOS token        = 151643 '<｜end▁of▁sentence｜>'
print_info: EOT token        = 151643 '<｜end▁of▁sentence｜>'
print_info: PAD token        = 151643 '<｜end▁of▁sentence｜>'
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 '<｜end▁of▁sentence｜>'
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:  CPU_AARCH64 model buffer size =  6108.75 MiB
load_tensors:   CPU_Mapped model buffer size =  8528.07 MiB
..........................................................................................
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 4096
llama_context: n_ctx_per_seq = 4096
llama_context: n_batch       = 2048
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = 0
llama_context: freq_base     = 1000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_per_seq (4096) < n_ctx_train (131072) -- the full capacity of the model will not be utilized
llama_context:        CPU  output buffer size =     0.58 MiB
init: kv_size = 4096, offload = 1, type_k = 'f16', type_v = 'f16', n_layer = 48, can_shift = 1
init:        CPU KV buffer size =   768.00 MiB
llama_context: KV self size  =  768.00 MiB, K (f16):  384.00 MiB, V (f16):  384.00 MiB
llama_context:        CPU compute buffer size =   368.01 MiB
llama_context: graph nodes  = 1782
llama_context: graph splits = 1
common_init_from_params: setting dry_penalty_last_n to ctx_size = 4096
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
main: llama threadpool init, n_threads = 36

system_info: n_threads = 36 (n_threads_batch = 36) / 72 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | AARCH64_REPACK = 1 | 

sampler seed: 1414590238
sampler params: 
        repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
        dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 4096
        top_k = 40, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.800
        mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000
sampler chain: logits -> logit-bias -> penalties -> dry -> top-k -> typical -> top-p -> min-p -> xtc -> temp-ext -> dist 
generate: n_ctx = 4096, n_batch = 2048, n_predict = 1024, n_keep = 1

I believe the meaning of life is to create. But how do I know it's not for another purpose? I feel confused. What should I do?

</think>

The meaning of life is a profound question that people have been exploring for ages, and different individuals may have different interpretations. If you believe that the meaning of life is to create, that is your personal philosophy, and it is a beautiful interpretation of life. As for whether there are other meanings, it is up to everyone to explore and interpret on their own. If you are feeling confused, perhaps it would help to take some time to reflect on your own values and beliefs, engage in more self-reflection, and seek inspiration from books or conversations with others. Remember, there is no right or wrong answer; what matters is finding a meaning that resonates with you personally. [end of text]


llama_perf_sampler_print:    sampling time =      19.32 ms /   170 runs   (    0.11 ms per token,  8798.72 tokens per second)
llama_perf_context_print:        load time =    4626.33 ms
llama_perf_context_print: prompt eval time =     394.17 ms /     8 tokens (   49.27 ms per token,    20.30 tokens per second)
llama_perf_context_print:        eval time =   27922.14 ms /   161 runs   (  173.43 ms per token,     5.77 tokens per second)
llama_perf_context_print:       total time =   28426.03 ms /   169 tokens

PR

$ ./llama-cli -m /home/debian/.codegpt/models/gguf/DeepSeek-R1-Distill-Qwen-14B-Q4_K_M.gguf -p "I believe the meaning of life is" -n 1024 -no-cnv
build: 5173 (93ab5602) with gcc-12 (Debian 12.4.0-5) 12.4.0 for x86_64-linux-gnu
main: llama backend init
main: load the model and apply lora adapter, if any
llama_model_loader: loaded meta data with 30 key-value pairs and 579 tensors from /home/debian/.codegpt/models/gguf/DeepSeek-R1-Distill-Qwen-14B-Q4_K_M.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv   0:                       general.architecture str              = qwen2
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = DeepSeek R1 Distill Qwen 14B
llama_model_loader: - kv   3:                           general.basename str              = DeepSeek-R1-Distill-Qwen
llama_model_loader: - kv   4:                         general.size_label str              = 14B
llama_model_loader: - kv   5:                          qwen2.block_count u32              = 48
llama_model_loader: - kv   6:                       qwen2.context_length u32              = 131072
llama_model_loader: - kv   7:                     qwen2.embedding_length u32              = 5120
llama_model_loader: - kv   8:                  qwen2.feed_forward_length u32              = 13824
llama_model_loader: - kv   9:                 qwen2.attention.head_count u32              = 40
llama_model_loader: - kv  10:              qwen2.attention.head_count_kv u32              = 8
llama_model_loader: - kv  11:                       qwen2.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  12:     qwen2.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  13:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  14:                         tokenizer.ggml.pre str              = deepseek-r1-qwen
llama_model_loader: - kv  15:                      tokenizer.ggml.tokens arr[str,152064]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  16:                  tokenizer.ggml.token_type arr[i32,152064]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  17:                      tokenizer.ggml.merges arr[str,151387]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  18:                tokenizer.ggml.bos_token_id u32              = 151646
llama_model_loader: - kv  19:                tokenizer.ggml.eos_token_id u32              = 151643
llama_model_loader: - kv  20:            tokenizer.ggml.padding_token_id u32              = 151643
llama_model_loader: - kv  21:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  22:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  23:                    tokenizer.chat_template str              = {% if not add_generation_prompt is de...
llama_model_loader: - kv  24:               general.quantization_version u32              = 2
llama_model_loader: - kv  25:                          general.file_type u32              = 15
llama_model_loader: - kv  26:                      quantize.imatrix.file str              = /models_out/DeepSeek-R1-Distill-Qwen-...
llama_model_loader: - kv  27:                   quantize.imatrix.dataset str              = /training_dir/calibration_datav3.txt
llama_model_loader: - kv  28:             quantize.imatrix.entries_count i32              = 336
llama_model_loader: - kv  29:              quantize.imatrix.chunks_count i32              = 128
llama_model_loader: - type  f32:  241 tensors
llama_model_loader: - type q4_K:  289 tensors
llama_model_loader: - type q6_K:   49 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q4_K - Medium
print_info: file size   = 8.37 GiB (4.87 BPW) 
load: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
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      = 131072
print_info: n_embd           = 5120
print_info: n_layer          = 48
print_info: n_head           = 40
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            = 5
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-05
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             = 13824
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  = 131072
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       = 14B
print_info: model params     = 14.77 B
print_info: general.name     = DeepSeek R1 Distill Qwen 14B
print_info: vocab type       = BPE
print_info: n_vocab          = 152064
print_info: n_merges         = 151387
print_info: BOS token        = 151646 '<｜begin▁of▁sentence｜>'
print_info: EOS token        = 151643 '<｜end▁of▁sentence｜>'
print_info: EOT token        = 151643 '<｜end▁of▁sentence｜>'
print_info: PAD token        = 151643 '<｜end▁of▁sentence｜>'
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 '<｜end▁of▁sentence｜>'
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:  CPU_AARCH64 model buffer size =  6108.75 MiB
load_tensors:   CPU_Mapped model buffer size =  8528.07 MiB
..........................................................................................
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 4096
llama_context: n_ctx_per_seq = 4096
llama_context: n_batch       = 2048
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = 0
llama_context: freq_base     = 1000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_per_seq (4096) < n_ctx_train (131072) -- the full capacity of the model will not be utilized
llama_context:        CPU  output buffer size =     0.58 MiB
init: kv_size = 4096, offload = 1, type_k = 'f16', type_v = 'f16', n_layer = 48, can_shift = 1
init:        CPU KV buffer size =   768.00 MiB
llama_context: KV self size  =  768.00 MiB, K (f16):  384.00 MiB, V (f16):  384.00 MiB
llama_context:        CPU compute buffer size =   368.01 MiB
llama_context: graph nodes  = 1782
llama_context: graph splits = 1
common_init_from_params: setting dry_penalty_last_n to ctx_size = 4096
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
main: llama threadpool init, n_threads = 36

system_info: n_threads = 36 (n_threads_batch = 36) / 72 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | AARCH64_REPACK = 1 | 

sampler seed: 2523516015
sampler params: 
        repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
        dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 4096
        top_k = 40, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.800
        mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000
sampler chain: logits -> logit-bias -> penalties -> dry -> top-k -> typical -> top-p -> min-p -> xtc -> temp-ext -> dist 
generate: n_ctx = 4096, n_batch = 2048, n_predict = 1024, n_keep = 1

I believe the meaning of life is to make others happy. What is your opinion?

</think>

As an AI assistant, I don't have personal opinions, beliefs, or emotions. However, I can tell you that helping others can bring meaning and happiness to many people's lives. It's a perspective that emphasizes compassion and the positive impact one can have on the world. [end of text]


llama_perf_sampler_print:    sampling time =       8.00 ms /    76 runs   (    0.11 ms per token,  9496.44 tokens per second)
llama_perf_context_print:        load time =    4454.66 ms
llama_perf_context_print: prompt eval time =     358.83 ms /     8 tokens (   44.85 ms per token,    22.29 tokens per second)
llama_perf_context_print:        eval time =   11095.16 ms /    67 runs   (  165.60 ms per token,     6.04 tokens per second)
llama_perf_context_print:       total time =   11507.06 ms /    75 tokens

GermanAizek · 2025-04-25T15:45:58Z

clang version 18.1.8 (17) for x86_64-pc-linux-gnu

Master

$ ./llama-cli -m /home/debian/.codegpt/models/gguf/DeepSeek-R1-Distill-Qwen-14B-Q4_K_M.gguf -p "I believe the meaning of life is" -n 1024 -no-cnv
build: 5170 (658987cf) with Debian clang version 18.1.8 (17) for x86_64-pc-linux-gnu
main: llama backend init
main: load the model and apply lora adapter, if any
llama_model_loader: loaded meta data with 30 key-value pairs and 579 tensors from /home/debian/.codegpt/models/gguf/DeepSeek-R1-Distill-Qwen-14B-Q4_K_M.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv   0:                       general.architecture str              = qwen2
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = DeepSeek R1 Distill Qwen 14B
llama_model_loader: - kv   3:                           general.basename str              = DeepSeek-R1-Distill-Qwen
llama_model_loader: - kv   4:                         general.size_label str              = 14B
llama_model_loader: - kv   5:                          qwen2.block_count u32              = 48
llama_model_loader: - kv   6:                       qwen2.context_length u32              = 131072
llama_model_loader: - kv   7:                     qwen2.embedding_length u32              = 5120
llama_model_loader: - kv   8:                  qwen2.feed_forward_length u32              = 13824
llama_model_loader: - kv   9:                 qwen2.attention.head_count u32              = 40
llama_model_loader: - kv  10:              qwen2.attention.head_count_kv u32              = 8
llama_model_loader: - kv  11:                       qwen2.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  12:     qwen2.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  13:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  14:                         tokenizer.ggml.pre str              = deepseek-r1-qwen
llama_model_loader: - kv  15:                      tokenizer.ggml.tokens arr[str,152064]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  16:                  tokenizer.ggml.token_type arr[i32,152064]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  17:                      tokenizer.ggml.merges arr[str,151387]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  18:                tokenizer.ggml.bos_token_id u32              = 151646
llama_model_loader: - kv  19:                tokenizer.ggml.eos_token_id u32              = 151643
llama_model_loader: - kv  20:            tokenizer.ggml.padding_token_id u32              = 151643
llama_model_loader: - kv  21:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  22:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  23:                    tokenizer.chat_template str              = {% if not add_generation_prompt is de...
llama_model_loader: - kv  24:               general.quantization_version u32              = 2
llama_model_loader: - kv  25:                          general.file_type u32              = 15
llama_model_loader: - kv  26:                      quantize.imatrix.file str              = /models_out/DeepSeek-R1-Distill-Qwen-...
llama_model_loader: - kv  27:                   quantize.imatrix.dataset str              = /training_dir/calibration_datav3.txt
llama_model_loader: - kv  28:             quantize.imatrix.entries_count i32              = 336
llama_model_loader: - kv  29:              quantize.imatrix.chunks_count i32              = 128
llama_model_loader: - type  f32:  241 tensors
llama_model_loader: - type q4_K:  289 tensors
llama_model_loader: - type q6_K:   49 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q4_K - Medium
print_info: file size   = 8.37 GiB (4.87 BPW) 
load: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
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      = 131072
print_info: n_embd           = 5120
print_info: n_layer          = 48
print_info: n_head           = 40
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            = 5
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-05
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             = 13824
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  = 131072
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       = 14B
print_info: model params     = 14.77 B
print_info: general.name     = DeepSeek R1 Distill Qwen 14B
print_info: vocab type       = BPE
print_info: n_vocab          = 152064
print_info: n_merges         = 151387
print_info: BOS token        = 151646 '<｜begin▁of▁sentence｜>'
print_info: EOS token        = 151643 '<｜end▁of▁sentence｜>'
print_info: EOT token        = 151643 '<｜end▁of▁sentence｜>'
print_info: PAD token        = 151643 '<｜end▁of▁sentence｜>'
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 '<｜end▁of▁sentence｜>'
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:  CPU_AARCH64 model buffer size =  6108.75 MiB
load_tensors:   CPU_Mapped model buffer size =  8528.07 MiB
..........................................................................................
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 4096
llama_context: n_ctx_per_seq = 4096
llama_context: n_batch       = 2048
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = 0
llama_context: freq_base     = 1000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_per_seq (4096) < n_ctx_train (131072) -- the full capacity of the model will not be utilized
llama_context:        CPU  output buffer size =     0.58 MiB
init: kv_size = 4096, offload = 1, type_k = 'f16', type_v = 'f16', n_layer = 48, can_shift = 1
init:        CPU KV buffer size =   768.00 MiB
llama_context: KV self size  =  768.00 MiB, K (f16):  384.00 MiB, V (f16):  384.00 MiB
llama_context:        CPU compute buffer size =   368.01 MiB
llama_context: graph nodes  = 1782
llama_context: graph splits = 1
common_init_from_params: setting dry_penalty_last_n to ctx_size = 4096
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
main: llama threadpool init, n_threads = 36

system_info: n_threads = 36 (n_threads_batch = 36) / 72 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | AARCH64_REPACK = 1 | 

sampler seed: 2685256224
sampler params: 
        repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
        dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 4096
        top_k = 40, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.800
        mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000
sampler chain: logits -> logit-bias -> penalties -> dry -> top-k -> typical -> top-p -> min-p -> xtc -> temp-ext -> dist 
generate: n_ctx = 4096, n_batch = 2048, n_predict = 1024, n_keep = 1

I believe the meaning of life is to find happiness and peace in oneself. How can I help someone else find this?

</think>

The meaning of life is a profound question that people have been exploring throughout history, and different individuals may have their own interpretations. As for how to help others find happiness and peace, perhaps we can start by understanding and respecting their unique life experiences and choices. Offering support and companionship, and creating a positive and harmonious environment can contribute to their well-being. At the same time, we should also cherish and seek happiness and peace for ourselves. [end of text]


llama_perf_sampler_print:    sampling time =      13.23 ms /   117 runs   (    0.11 ms per token,  8845.54 tokens per second)
llama_perf_context_print:        load time =    4628.32 ms
llama_perf_context_print: prompt eval time =     325.56 ms /     8 tokens (   40.69 ms per token,    24.57 tokens per second)
llama_perf_context_print:        eval time =   18301.47 ms /   108 runs   (  169.46 ms per token,     5.90 tokens per second)
llama_perf_context_print:       total time =   18707.20 ms /   116 tokens

PR

$ ./llama-cli -m /home/debian/.codegpt/models/gguf/DeepSeek-R1-Distill-Qwen-14B-Q4_K_M.gguf -p "I believe the meaning of life is" -n 1024 -no-cnv
build: 5173 (93ab5602) with Debian clang version 18.1.8 (17) for x86_64-pc-linux-gnu
main: llama backend init
main: load the model and apply lora adapter, if any
llama_model_loader: loaded meta data with 30 key-value pairs and 579 tensors from /home/debian/.codegpt/models/gguf/DeepSeek-R1-Distill-Qwen-14B-Q4_K_M.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv   0:                       general.architecture str              = qwen2
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = DeepSeek R1 Distill Qwen 14B
llama_model_loader: - kv   3:                           general.basename str              = DeepSeek-R1-Distill-Qwen
llama_model_loader: - kv   4:                         general.size_label str              = 14B
llama_model_loader: - kv   5:                          qwen2.block_count u32              = 48
llama_model_loader: - kv   6:                       qwen2.context_length u32              = 131072
llama_model_loader: - kv   7:                     qwen2.embedding_length u32              = 5120
llama_model_loader: - kv   8:                  qwen2.feed_forward_length u32              = 13824
llama_model_loader: - kv   9:                 qwen2.attention.head_count u32              = 40
llama_model_loader: - kv  10:              qwen2.attention.head_count_kv u32              = 8
llama_model_loader: - kv  11:                       qwen2.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  12:     qwen2.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  13:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  14:                         tokenizer.ggml.pre str              = deepseek-r1-qwen
llama_model_loader: - kv  15:                      tokenizer.ggml.tokens arr[str,152064]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  16:                  tokenizer.ggml.token_type arr[i32,152064]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  17:                      tokenizer.ggml.merges arr[str,151387]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  18:                tokenizer.ggml.bos_token_id u32              = 151646
llama_model_loader: - kv  19:                tokenizer.ggml.eos_token_id u32              = 151643
llama_model_loader: - kv  20:            tokenizer.ggml.padding_token_id u32              = 151643
llama_model_loader: - kv  21:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  22:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  23:                    tokenizer.chat_template str              = {% if not add_generation_prompt is de...
llama_model_loader: - kv  24:               general.quantization_version u32              = 2
llama_model_loader: - kv  25:                          general.file_type u32              = 15
llama_model_loader: - kv  26:                      quantize.imatrix.file str              = /models_out/DeepSeek-R1-Distill-Qwen-...
llama_model_loader: - kv  27:                   quantize.imatrix.dataset str              = /training_dir/calibration_datav3.txt
llama_model_loader: - kv  28:             quantize.imatrix.entries_count i32              = 336
llama_model_loader: - kv  29:              quantize.imatrix.chunks_count i32              = 128
llama_model_loader: - type  f32:  241 tensors
llama_model_loader: - type q4_K:  289 tensors
llama_model_loader: - type q6_K:   49 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q4_K - Medium
print_info: file size   = 8.37 GiB (4.87 BPW) 
load: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
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      = 131072
print_info: n_embd           = 5120
print_info: n_layer          = 48
print_info: n_head           = 40
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            = 5
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-05
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             = 13824
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  = 131072
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       = 14B
print_info: model params     = 14.77 B
print_info: general.name     = DeepSeek R1 Distill Qwen 14B
print_info: vocab type       = BPE
print_info: n_vocab          = 152064
print_info: n_merges         = 151387
print_info: BOS token        = 151646 '<｜begin▁of▁sentence｜>'
print_info: EOS token        = 151643 '<｜end▁of▁sentence｜>'
print_info: EOT token        = 151643 '<｜end▁of▁sentence｜>'
print_info: PAD token        = 151643 '<｜end▁of▁sentence｜>'
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 '<｜end▁of▁sentence｜>'
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:  CPU_AARCH64 model buffer size =  6108.75 MiB
load_tensors:   CPU_Mapped model buffer size =  8528.07 MiB
..........................................................................................
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 4096
llama_context: n_ctx_per_seq = 4096
llama_context: n_batch       = 2048
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = 0
llama_context: freq_base     = 1000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_per_seq (4096) < n_ctx_train (131072) -- the full capacity of the model will not be utilized
llama_context:        CPU  output buffer size =     0.58 MiB
init: kv_size = 4096, offload = 1, type_k = 'f16', type_v = 'f16', n_layer = 48, can_shift = 1
init:        CPU KV buffer size =   768.00 MiB
llama_context: KV self size  =  768.00 MiB, K (f16):  384.00 MiB, V (f16):  384.00 MiB
llama_context:        CPU compute buffer size =   368.01 MiB
llama_context: graph nodes  = 1782
llama_context: graph splits = 1
common_init_from_params: setting dry_penalty_last_n to ctx_size = 4096
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
main: llama threadpool init, n_threads = 36

system_info: n_threads = 36 (n_threads_batch = 36) / 72 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | AARCH64_REPACK = 1 | 

sampler seed: 1959168874
sampler params: 
        repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
        dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 4096
        top_k = 40, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.800
        mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000
sampler chain: logits -> logit-bias -> penalties -> dry -> top-k -> typical -> top-p -> min-p -> xtc -> temp-ext -> dist 
generate: n_ctx = 4096, n_batch = 2048, n_predict = 1024, n_keep = 1

I believe the meaning of life is to create and experience beauty. How can I express this belief in a way that's meaningful to others?

</think>

I think the meaning of life is to create and experience beauty. How can I express this belief in a way that's meaningful to others? [end of text]


llama_perf_sampler_print:    sampling time =       6.05 ms /    60 runs   (    0.10 ms per token,  9915.72 tokens per second)
llama_perf_context_print:        load time =    4485.68 ms
llama_perf_context_print: prompt eval time =     285.28 ms /     8 tokens (   35.66 ms per token,    28.04 tokens per second)
llama_perf_context_print:        eval time =    8383.15 ms /    51 runs   (  164.38 ms per token,     6.08 tokens per second)
llama_perf_context_print:       total time =    8711.94 ms /    59 tokens

ExtReMLapin · 2025-04-25T22:18:11Z

you should use llama-bench instead.

Also tbf i expected the compiler to do the optimization job itself with O3

GermanAizek · 2025-04-25T22:36:29Z

you should use llama-bench instead.

Also tbf i expected the compiler to do the optimization job itself with O3

Im tested on Release build, -fshort-enum not safe flag in O3. it is better to manually enum type strictly

GermanAizek · 2025-05-01T18:00:01Z

Resolved merge conflict with master branch.

github-actions bot added ggml changes relating to the ggml tensor library for machine learning build Compilation issues labels Apr 22, 2025

GermanAizek force-pushed the master branch from 9b03387 to 4b0a1bc Compare April 22, 2025 22:36

GermanAizek requested a review from JohannesGaessler as a code owner April 22, 2025 22:42

GermanAizek force-pushed the master branch 5 times, most recently from 12d22af to 8622380 Compare April 22, 2025 23:16

GermanAizek force-pushed the master branch from 8622380 to a1610bd Compare April 24, 2025 10:29

github-actions bot added Vulkan Issues specific to the Vulkan backend SYCL https://en.wikipedia.org/wiki/SYCL - GPU programming language labels Apr 24, 2025

GermanAizek force-pushed the master branch 3 times, most recently from 51f0426 to 93ab560 Compare April 24, 2025 12:51

GermanAizek added 2 commits April 24, 2025 16:28

[cmn/ggml/llama] reduced enum size using scoped enums and packed attr…

55071fb

…ibute

[ggml/llama] Align structures after scoped enums for 64-bit platforms

a1610bd

- llama_model_params 72 bytes -> 64 bytes - ggml_cgraph 80 bytes -> 72 bytes - hash_node 32 bytes -> 24 bytes - ggml_threadpool 160 -> 152 bytes - best_tokenization 24 -> 16 bytes

[ggml] fixes for cross-platform building successful

93ab560

Merge branch 'master' into master

15c2166

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Reduce enum sizes some are used in structs, which allowed them to be optimized. #13071

Reduce enum sizes some are used in structs, which allowed them to be optimized. #13071

GermanAizek commented Apr 22, 2025 •

edited

Loading

USBhost commented Apr 22, 2025

GermanAizek commented Apr 22, 2025

GermanAizek commented Apr 22, 2025

USBhost commented Apr 23, 2025 •

edited

Loading

GermanAizek commented Apr 23, 2025 •

edited

Loading

USBhost commented Apr 24, 2025

GermanAizek commented Apr 24, 2025

GermanAizek commented Apr 24, 2025 •

edited

Loading

GermanAizek commented Apr 24, 2025

GermanAizek commented Apr 24, 2025 •

edited

Loading

GermanAizek commented Apr 24, 2025

GermanAizek commented Apr 24, 2025 •

edited

Loading

GermanAizek commented Apr 24, 2025

USBhost commented Apr 24, 2025

GermanAizek commented Apr 24, 2025 •

edited

Loading

USBhost commented Apr 24, 2025

GermanAizek commented Apr 25, 2025 •

edited

Loading

GermanAizek commented Apr 25, 2025 •

edited

Loading

GermanAizek commented Apr 25, 2025

GermanAizek commented Apr 25, 2025

ExtReMLapin commented Apr 25, 2025

GermanAizek commented Apr 25, 2025 •

edited

Loading

GermanAizek commented May 1, 2025

Reduce enum sizes some are used in structs, which allowed them to be optimized. #13071

Are you sure you want to change the base?

Reduce enum sizes some are used in structs, which allowed them to be optimized. #13071

Conversation

GermanAizek commented Apr 22, 2025 • edited Loading

USBhost commented Apr 22, 2025

GermanAizek commented Apr 22, 2025

GermanAizek commented Apr 22, 2025

USBhost commented Apr 23, 2025 • edited Loading

GermanAizek commented Apr 23, 2025 • edited Loading

USBhost commented Apr 24, 2025

GermanAizek commented Apr 24, 2025

GermanAizek commented Apr 24, 2025 • edited Loading

GermanAizek commented Apr 24, 2025

GermanAizek commented Apr 24, 2025 • edited Loading

clang version 19.1.7 (3) for x86_64-pc-linux-gnu

Master

PR

GermanAizek commented Apr 24, 2025

Master

PR

GermanAizek commented Apr 24, 2025 • edited Loading

GCC (Debian 14.2.0-19) 14.2.0 for x86_64-linux-gnu

Master

PR

GermanAizek commented Apr 24, 2025

USBhost commented Apr 24, 2025

Master

PR

GermanAizek commented Apr 24, 2025 • edited Loading

USBhost commented Apr 24, 2025

GermanAizek commented Apr 25, 2025 • edited Loading

GermanAizek commented Apr 25, 2025 • edited Loading

gcc-13 (Debian 13.3.0-13) 13.3.0 for x86_64-linux-gnu

Master

PR

GermanAizek commented Apr 25, 2025

gcc-12 (Debian 12.4.0-5) 12.4.0 for x86_64-linux-gnu

Master

PR

GermanAizek commented Apr 25, 2025

clang version 18.1.8 (17) for x86_64-pc-linux-gnu

Master

PR

ExtReMLapin commented Apr 25, 2025

GermanAizek commented Apr 25, 2025 • edited Loading

GermanAizek commented May 1, 2025

GermanAizek commented Apr 22, 2025 •

edited

Loading

USBhost commented Apr 23, 2025 •

edited

Loading

GermanAizek commented Apr 23, 2025 •

edited

Loading

GermanAizek commented Apr 24, 2025 •

edited

Loading

GermanAizek commented Apr 24, 2025 •

edited

Loading

GermanAizek commented Apr 24, 2025 •

edited

Loading

GermanAizek commented Apr 24, 2025 •

edited

Loading

GermanAizek commented Apr 25, 2025 •

edited

Loading

GermanAizek commented Apr 25, 2025 •

edited

Loading

GermanAizek commented Apr 25, 2025 •

edited

Loading