search

Mozilla-Ocho / llamafile

Distribute and run LLMs with a single file.
https://llamafile.ai
Other
20.58k stars 1.04k forks source link

Bug: Huge difference between prompt processing (tokens/sec) compared to Llama cpp or Ollama #577

Closed mathav95raj closed 1 month ago

mathav95raj commented 1 month ago

What happened?

image

For llama cpp I had downloaded the q4_k_m quantized model and used llama-bench. For ollama I pulled the q4_k_m model from ollama. By running the model with --verbose flag, I manually recorded the prompt eval rate for 10 trials with same prompt of approximately 512 tokens length. For llamafile I used the same model as used for llama cpp and created a llamafile and then benchmarked with llamafile-bench.

llama-bench logs:

llama-bench -m "gguf/llama-3.2-1b-q4_k_m.gguf" -p 512 -n 1024 -ngl 17 --verbose
| model                          |       size |     params | backend    | ngl |          test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | ------------: | -------------------: |
llama_model_loader: loaded meta data with 29 key-value pairs and 147 tensors from gguf/llama-3.2-1b-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              = llama
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = Llama 3.2 1B
llama_model_loader: - kv   3:                           general.basename str              = Llama-3.2
llama_model_loader: - kv   4:                         general.size_label str              = 1B
llama_model_loader: - kv   5:                            general.license str              = llama3.2
llama_model_loader: - kv   6:                               general.tags arr[str,6]       = ["facebook", "meta", "pytorch", "llam...
llama_model_loader: - kv   7:                          general.languages arr[str,8]       = ["en", "de", "fr", "it", "pt", "hi", ...
llama_model_loader: - kv   8:                          llama.block_count u32              = 16
llama_model_loader: - kv   9:                       llama.context_length u32              = 131072
llama_model_loader: - kv  10:                     llama.embedding_length u32              = 2048
llama_model_loader: - kv  11:                  llama.feed_forward_length u32              = 8192
llama_model_loader: - kv  12:                 llama.attention.head_count u32              = 32
llama_model_loader: - kv  13:              llama.attention.head_count_kv u32              = 8
llama_model_loader: - kv  14:                       llama.rope.freq_base f32              = 500000.000000
llama_model_loader: - kv  15:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  16:                 llama.attention.key_length u32              = 64
llama_model_loader: - kv  17:               llama.attention.value_length u32              = 64
llama_model_loader: - kv  18:                          general.file_type u32              = 15
llama_model_loader: - kv  19:                           llama.vocab_size u32              = 128256
llama_model_loader: - kv  20:                 llama.rope.dimension_count u32              = 64
llama_model_loader: - kv  21:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  22:                         tokenizer.ggml.pre str              = llama-bpe
llama_model_loader: - kv  23:                      tokenizer.ggml.tokens arr[str,128256]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  24:                  tokenizer.ggml.token_type arr[i32,128256]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  25:                      tokenizer.ggml.merges arr[str,280147]  = ["Ġ Ġ", "Ġ ĠĠĠ", "ĠĠ ĠĠ", "...
llama_model_loader: - kv  26:                tokenizer.ggml.bos_token_id u32              = 128000
llama_model_loader: - kv  27:                tokenizer.ggml.eos_token_id u32              = 128001
llama_model_loader: - kv  28:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:   34 tensors
llama_model_loader: - type q4_K:   96 tensors
llama_model_loader: - type q6_K:   17 tensors
llm_load_vocab: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
llm_load_vocab: special tokens cache size = 256
llm_load_vocab: token to piece cache size = 0.7999 MB
llm_load_print_meta: format           = GGUF V3 (latest)
llm_load_print_meta: arch             = llama
llm_load_print_meta: vocab type       = BPE
llm_load_print_meta: n_vocab          = 128256
llm_load_print_meta: n_merges         = 280147
llm_load_print_meta: vocab_only       = 0
llm_load_print_meta: n_ctx_train      = 131072
llm_load_print_meta: n_embd           = 2048
llm_load_print_meta: n_layer          = 16
llm_load_print_meta: n_head           = 32
llm_load_print_meta: n_head_kv        = 8
llm_load_print_meta: n_rot            = 64
llm_load_print_meta: n_swa            = 0
llm_load_print_meta: n_embd_head_k    = 64
llm_load_print_meta: n_embd_head_v    = 64
llm_load_print_meta: n_gqa            = 4
llm_load_print_meta: n_embd_k_gqa     = 512
llm_load_print_meta: n_embd_v_gqa     = 512
llm_load_print_meta: f_norm_eps       = 0.0e+00
llm_load_print_meta: f_norm_rms_eps   = 1.0e-05
llm_load_print_meta: f_clamp_kqv      = 0.0e+00
llm_load_print_meta: f_max_alibi_bias = 0.0e+00
llm_load_print_meta: f_logit_scale    = 0.0e+00
llm_load_print_meta: n_ff             = 8192
llm_load_print_meta: n_expert         = 0
llm_load_print_meta: n_expert_used    = 0
llm_load_print_meta: causal attn      = 1
llm_load_print_meta: pooling type     = 0
llm_load_print_meta: rope type        = 0
llm_load_print_meta: rope scaling     = linear
llm_load_print_meta: freq_base_train  = 500000.0
llm_load_print_meta: freq_scale_train = 1
llm_load_print_meta: n_ctx_orig_yarn  = 131072
llm_load_print_meta: rope_finetuned   = unknown
llm_load_print_meta: ssm_d_conv       = 0
llm_load_print_meta: ssm_d_inner      = 0
llm_load_print_meta: ssm_d_state      = 0
llm_load_print_meta: ssm_dt_rank      = 0
llm_load_print_meta: ssm_dt_b_c_rms   = 0
llm_load_print_meta: model type       = ?B
llm_load_print_meta: model ftype      = Q4_K - Medium
llm_load_print_meta: model params     = 1.24 B
llm_load_print_meta: model size       = 762.81 MiB (5.18 BPW) 
llm_load_print_meta: general.name     = Llama 3.2 1B
llm_load_print_meta: BOS token        = 128000 '<|begin_of_text|>'
llm_load_print_meta: EOS token        = 128001 '<|end_of_text|>'
llm_load_print_meta: LF token         = 128 'Ä'
llm_load_print_meta: EOT token        = 128009 '<|eot_id|>'
llm_load_print_meta: EOM token        = 128008 '<|eom_id|>'
llm_load_print_meta: EOG token        = 128001 '<|end_of_text|>'
llm_load_print_meta: EOG token        = 128008 '<|eom_id|>'
llm_load_print_meta: EOG token        = 128009 '<|eot_id|>'
llm_load_print_meta: max token length = 256
llm_load_tensors: ggml ctx size =    0.14 MiB
ggml_backend_metal_log_allocated_size: allocated buffer, size =   762.83 MiB, (  762.91 / 27648.00)
llm_load_tensors: offloading 16 repeating layers to GPU
llm_load_tensors: offloading non-repeating layers to GPU
llm_load_tensors: offloaded 17/17 layers to GPU
llm_load_tensors:        CPU buffer size =   205.49 MiB
llm_load_tensors:      Metal buffer size =   762.82 MiB
.......................................................
llama_new_context_with_model: n_ctx      = 512
llama_new_context_with_model: n_batch    = 512
llama_new_context_with_model: n_ubatch   = 512
llama_new_context_with_model: flash_attn = 0
llama_new_context_with_model: freq_base  = 500000.0
llama_new_context_with_model: freq_scale = 1
ggml_metal_init: allocating
ggml_metal_init: found device: Apple M3 Max
ggml_metal_init: picking default device: Apple M3 Max
ggml_metal_init: using embedded metal library
ggml_metal_init: GPU name:   Apple M3 Max
ggml_metal_init: GPU family: MTLGPUFamilyApple9  (1009)
ggml_metal_init: GPU family: MTLGPUFamilyCommon3 (3003)
ggml_metal_init: GPU family: MTLGPUFamilyMetal3  (5001)
ggml_metal_init: simdgroup reduction support   = true
ggml_metal_init: simdgroup matrix mul. support = true
ggml_metal_init: hasUnifiedMemory              = true
ggml_metal_init: recommendedMaxWorkingSetSize  = 28991.03 MB
llama_kv_cache_init:      Metal KV buffer size =    16.00 MiB
llama_new_context_with_model: KV self size  =   16.00 MiB, K (f16):    8.00 MiB, V (f16):    8.00 MiB
llama_new_context_with_model:        CPU  output buffer size =     0.49 MiB
llama_new_context_with_model:      Metal compute buffer size =   254.50 MiB
llama_new_context_with_model:        CPU compute buffer size =     5.01 MiB
llama_new_context_with_model: graph nodes  = 518
llama_new_context_with_model: graph splits = 2
| llama ?B Q4_K - Medium         | 968.30 MiB |     1.50 B | Metal      |  17 |         pp512 |      3177.19 ± 12.58 |
llama_perf_context_print:        load time =     503.11 ms
llama_perf_context_print: prompt eval time =       0.00 ms /  3072 tokens (    0.00 ms per token,      inf tokens per second)
llama_perf_context_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
llama_perf_context_print:       total time =    1309.45 ms /  3073 tokens
ggml_metal_free: deallocating
llama_new_context_with_model: n_ctx      = 1024
llama_new_context_with_model: n_batch    = 1024
llama_new_context_with_model: n_ubatch   = 512
llama_new_context_with_model: flash_attn = 0
llama_new_context_with_model: freq_base  = 500000.0
llama_new_context_with_model: freq_scale = 1
ggml_metal_init: allocating
ggml_metal_init: found device: Apple M3 Max
ggml_metal_init: picking default device: Apple M3 Max
ggml_metal_init: using embedded metal library
ggml_metal_init: GPU name:   Apple M3 Max
ggml_metal_init: GPU family: MTLGPUFamilyApple9  (1009)
ggml_metal_init: GPU family: MTLGPUFamilyCommon3 (3003)
ggml_metal_init: GPU family: MTLGPUFamilyMetal3  (5001)
ggml_metal_init: simdgroup reduction support   = true
ggml_metal_init: simdgroup matrix mul. support = true
ggml_metal_init: hasUnifiedMemory              = true
ggml_metal_init: recommendedMaxWorkingSetSize  = 28991.03 MB
llama_kv_cache_init:      Metal KV buffer size =    32.00 MiB
llama_new_context_with_model: KV self size  =   32.00 MiB, K (f16):   16.00 MiB, V (f16):   16.00 MiB
llama_new_context_with_model:        CPU  output buffer size =     0.49 MiB
llama_new_context_with_model:      Metal compute buffer size =   254.50 MiB
llama_new_context_with_model:        CPU compute buffer size =     6.01 MiB
llama_new_context_with_model: graph nodes  = 518
llama_new_context_with_model: graph splits = 2
| llama ?B Q4_K - Medium         | 968.30 MiB |     1.50 B | Metal      |  17 |        tg1024 |        149.36 ± 0.47 |
llama_perf_context_print:        load time =    1341.56 ms
llama_perf_context_print: prompt eval time =       0.00 ms /     1 tokens (    0.00 ms per token,      inf tokens per second)
llama_perf_context_print:        eval time =       0.00 ms /  5121 runs   (    0.00 ms per token,      inf tokens per second)
llama_perf_context_print:       total time =   35623.50 ms /  5122 tokens
ggml_metal_free: deallocating

build: 1b2f992c (3837)

ollama logs:

cat ~/.ollama/logs/server.log
2024/10/03 08:09:46 routes.go:1153: INFO server config env="map[HTTPS_PROXY: HTTP_PROXY: NO_PROXY: OLLAMA_DEBUG:false OLLAMA_FLASH_ATTENTION:false OLLAMA_GPU_OVERHEAD:0 OLLAMA_HOST:http://127.0.0.1:11434 OLLAMA_KEEP_ALIVE:5m0s OLLAMA_LLM_LIBRARY: OLLAMA_LOAD_TIMEOUT:5m0s OLLAMA_MAX_LOADED_MODELS:0 OLLAMA_MAX_QUEUE:512 OLLAMA_MODELS:/Users/391080/.ollama/models OLLAMA_NOHISTORY:false OLLAMA_NOPRUNE:false OLLAMA_NUM_PARALLEL:0 OLLAMA_ORIGINS:[http://localhost https://localhost http://localhost:* https://localhost:* http://127.0.0.1 https://127.0.0.1 http://127.0.0.1:* https://127.0.0.1:* http://0.0.0.0 https://0.0.0.0 http://0.0.0.0:* https://0.0.0.0:* app://* file://* tauri://*] OLLAMA_SCHED_SPREAD:false OLLAMA_TMPDIR: http_proxy: https_proxy: no_proxy:]"
time=2024-10-03T08:09:46.111+05:30 level=INFO source=images.go:753 msg="total blobs: 18"
time=2024-10-03T08:09:46.113+05:30 level=INFO source=images.go:760 msg="total unused blobs removed: 0"
time=2024-10-03T08:09:46.114+05:30 level=INFO source=routes.go:1200 msg="Listening on 127.0.0.1:11434 (version 0.3.12)"
time=2024-10-03T08:09:46.115+05:30 level=INFO source=common.go:135 msg="extracting embedded files" dir=/var/folders/_y/gbtnmk4s65lfdzb3vw888hydbsxg_p/T/ollama2160509689/runners
time=2024-10-03T08:09:46.135+05:30 level=INFO source=common.go:49 msg="Dynamic LLM libraries" runners=[metal]
time=2024-10-03T08:09:46.201+05:30 level=INFO source=types.go:107 msg="inference compute" id=0 library=metal variant="" compute="" driver=0.0 name="" total="27.0 GiB" available="27.0 GiB"
[GIN] 2024/10/03 - 08:09:46 | 200 |      93.209µs |       127.0.0.1 | HEAD     "/"
[GIN] 2024/10/03 - 08:09:46 | 200 |   11.964959ms |       127.0.0.1 | POST     "/api/show"
time=2024-10-03T08:09:46.631+05:30 level=INFO source=sched.go:714 msg="new model will fit in available VRAM in single GPU, loading" model=/Users/391080/.ollama/models/blobs/sha256-d06ffdc00fd5175ccb2371c6686ba63ed30bc915253158721344924bf699401e gpu=0 parallel=4 available=28991029248 required="2.1 GiB"
time=2024-10-03T08:09:46.631+05:30 level=INFO source=server.go:103 msg="system memory" total="36.0 GiB" free="28.5 GiB" free_swap="0 B"
time=2024-10-03T08:09:46.631+05:30 level=INFO source=memory.go:326 msg="offload to metal" layers.requested=-1 layers.model=17 layers.offload=17 layers.split="" memory.available="[27.0 GiB]" memory.gpu_overhead="0 B" memory.required.full="2.1 GiB" memory.required.partial="2.1 GiB" memory.required.kv="256.0 MiB" memory.required.allocations="[2.1 GiB]" memory.weights.total="813.3 MiB" memory.weights.repeating="607.8 MiB" memory.weights.nonrepeating="205.5 MiB" memory.graph.full="544.0 MiB" memory.graph.partial="544.0 MiB"
time=2024-10-03T08:09:46.633+05:30 level=INFO source=server.go:388 msg="starting llama server" cmd="/var/folders/_y/gbtnmk4s65lfdzb3vw888hydbsxg_p/T/ollama2160509689/runners/metal/ollama_llama_server --model /Users/391080/.ollama/models/blobs/sha256-d06ffdc00fd5175ccb2371c6686ba63ed30bc915253158721344924bf699401e --ctx-size 8192 --batch-size 512 --embedding --log-disable --n-gpu-layers 17 --parallel 4 --port 49469"
time=2024-10-03T08:09:46.642+05:30 level=INFO source=sched.go:449 msg="loaded runners" count=1
time=2024-10-03T08:09:46.642+05:30 level=INFO source=server.go:587 msg="waiting for llama runner to start responding"
time=2024-10-03T08:09:46.642+05:30 level=INFO source=server.go:621 msg="waiting for server to become available" status="llm server error"
INFO [main] build info | build=3670 commit="194ef086" tid="0x1f435bac0" timestamp=1727923187
INFO [main] system info | n_threads=10 n_threads_batch=10 system_info="AVX = 0 | AVX_VNNI = 0 | AVX2 = 0 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | FMA = 0 | NEON = 1 | SVE = 0 | ARM_FMA = 1 | F16C = 0 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 0 | SSSE3 = 0 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 1 | " tid="0x1f435bac0" timestamp=1727923187 total_threads=14
INFO [main] HTTP server listening | hostname="127.0.0.1" n_threads_http="13" port="49469" tid="0x1f435bac0" timestamp=1727923187
time=2024-10-03T08:09:47.397+05:30 level=INFO source=server.go:621 msg="waiting for server to become available" status="llm server loading model"
llama_model_loader: loaded meta data with 29 key-value pairs and 147 tensors from /Users/391080/.ollama/models/blobs/sha256-d06ffdc00fd5175ccb2371c6686ba63ed30bc915253158721344924bf699401e (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv   0:                       general.architecture str              = llama
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = Llama 3.2 1B
llama_model_loader: - kv   3:                           general.basename str              = Llama-3.2
llama_model_loader: - kv   4:                         general.size_label str              = 1B
llama_model_loader: - kv   5:                            general.license str              = llama3.2
llama_model_loader: - kv   6:                               general.tags arr[str,6]       = ["facebook", "meta", "pytorch", "llam...
llama_model_loader: - kv   7:                          general.languages arr[str,8]       = ["en", "de", "fr", "it", "pt", "hi", ...
llama_model_loader: - kv   8:                          llama.block_count u32              = 16
llama_model_loader: - kv   9:                       llama.context_length u32              = 131072
llama_model_loader: - kv  10:                     llama.embedding_length u32              = 2048
llama_model_loader: - kv  11:                  llama.feed_forward_length u32              = 8192
llama_model_loader: - kv  12:                 llama.attention.head_count u32              = 32
llama_model_loader: - kv  13:              llama.attention.head_count_kv u32              = 8
llama_model_loader: - kv  14:                       llama.rope.freq_base f32              = 500000.000000
llama_model_loader: - kv  15:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  16:                 llama.attention.key_length u32              = 64
llama_model_loader: - kv  17:               llama.attention.value_length u32              = 64
llama_model_loader: - kv  18:                          general.file_type u32              = 15
llama_model_loader: - kv  19:                           llama.vocab_size u32              = 128256
llama_model_loader: - kv  20:                 llama.rope.dimension_count u32              = 64
llama_model_loader: - kv  21:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  22:                         tokenizer.ggml.pre str              = llama-bpe
llama_model_loader: - kv  23:                      tokenizer.ggml.tokens arr[str,128256]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  24:                  tokenizer.ggml.token_type arr[i32,128256]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  25:                      tokenizer.ggml.merges arr[str,280147]  = ["Ġ Ġ", "Ġ ĠĠĠ", "ĠĠ ĠĠ", "...
llama_model_loader: - kv  26:                tokenizer.ggml.bos_token_id u32              = 128000
llama_model_loader: - kv  27:                tokenizer.ggml.eos_token_id u32              = 128001
llama_model_loader: - kv  28:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:   34 tensors
llama_model_loader: - type q4_K:   96 tensors
llama_model_loader: - type q6_K:   17 tensors
llm_load_vocab: special tokens cache size = 256
llm_load_vocab: token to piece cache size = 0.7999 MB
llm_load_print_meta: format           = GGUF V3 (latest)
llm_load_print_meta: arch             = llama
llm_load_print_meta: vocab type       = BPE
llm_load_print_meta: n_vocab          = 128256
llm_load_print_meta: n_merges         = 280147
llm_load_print_meta: vocab_only       = 0
llm_load_print_meta: n_ctx_train      = 131072
llm_load_print_meta: n_embd           = 2048
llm_load_print_meta: n_layer          = 16
llm_load_print_meta: n_head           = 32
llm_load_print_meta: n_head_kv        = 8
llm_load_print_meta: n_rot            = 64
llm_load_print_meta: n_swa            = 0
llm_load_print_meta: n_embd_head_k    = 64
llm_load_print_meta: n_embd_head_v    = 64
llm_load_print_meta: n_gqa            = 4
llm_load_print_meta: n_embd_k_gqa     = 512
llm_load_print_meta: n_embd_v_gqa     = 512
llm_load_print_meta: f_norm_eps       = 0.0e+00
llm_load_print_meta: f_norm_rms_eps   = 1.0e-05
llm_load_print_meta: f_clamp_kqv      = 0.0e+00
llm_load_print_meta: f_max_alibi_bias = 0.0e+00
llm_load_print_meta: f_logit_scale    = 0.0e+00
llm_load_print_meta: n_ff             = 8192
llm_load_print_meta: n_expert         = 0
llm_load_print_meta: n_expert_used    = 0
llm_load_print_meta: causal attn      = 1
llm_load_print_meta: pooling type     = 0
llm_load_print_meta: rope type        = 0
llm_load_print_meta: rope scaling     = linear
llm_load_print_meta: freq_base_train  = 500000.0
llm_load_print_meta: freq_scale_train = 1
llm_load_print_meta: n_ctx_orig_yarn  = 131072
llm_load_print_meta: rope_finetuned   = unknown
llm_load_print_meta: ssm_d_conv       = 0
llm_load_print_meta: ssm_d_inner      = 0
llm_load_print_meta: ssm_d_state      = 0
llm_load_print_meta: ssm_dt_rank      = 0
llm_load_print_meta: ssm_dt_b_c_rms   = 0
llm_load_print_meta: model type       = ?B
llm_load_print_meta: model ftype      = Q4_K - Medium
llm_load_print_meta: model params     = 1.24 B
llm_load_print_meta: model size       = 762.81 MiB (5.18 BPW) 
llm_load_print_meta: general.name     = Llama 3.2 1B
llm_load_print_meta: BOS token        = 128000 '<|begin_of_text|>'
llm_load_print_meta: EOS token        = 128001 '<|end_of_text|>'
llm_load_print_meta: LF token         = 128 'Ä'
llm_load_print_meta: EOT token        = 128009 '<|eot_id|>'
llm_load_print_meta: max token length = 256
llm_load_tensors: ggml ctx size =    0.14 MiB
ggml_backend_metal_log_allocated_size: allocated buffer, size =   762.83 MiB, (  762.91 / 27648.00)
llm_load_tensors: offloading 16 repeating layers to GPU
llm_load_tensors: offloading non-repeating layers to GPU
llm_load_tensors: offloaded 17/17 layers to GPU
llm_load_tensors:        CPU buffer size =   205.49 MiB
llm_load_tensors:      Metal buffer size =   762.82 MiB
llama_new_context_with_model: n_ctx      = 8192
llama_new_context_with_model: n_batch    = 512
llama_new_context_with_model: n_ubatch   = 512
llama_new_context_with_model: flash_attn = 0
llama_new_context_with_model: freq_base  = 500000.0
llama_new_context_with_model: freq_scale = 1
ggml_metal_init: allocating
ggml_metal_init: found device: Apple M3 Max
ggml_metal_init: picking default device: Apple M3 Max
ggml_metal_init: using embedded metal library
ggml_metal_init: GPU name:   Apple M3 Max
ggml_metal_init: GPU family: MTLGPUFamilyApple9  (1009)
ggml_metal_init: GPU family: MTLGPUFamilyCommon3 (3003)
ggml_metal_init: GPU family: MTLGPUFamilyMetal3  (5001)
ggml_metal_init: simdgroup reduction support   = true
ggml_metal_init: simdgroup matrix mul. support = true
ggml_metal_init: hasUnifiedMemory              = true
ggml_metal_init: recommendedMaxWorkingSetSize  = 28991.03 MB
llama_kv_cache_init:      Metal KV buffer size =   256.00 MiB
llama_new_context_with_model: KV self size  =  256.00 MiB, K (f16):  128.00 MiB, V (f16):  128.00 MiB
llama_new_context_with_model:        CPU  output buffer size =     1.99 MiB
llama_new_context_with_model:      Metal compute buffer size =   544.00 MiB
llama_new_context_with_model:        CPU compute buffer size =    20.01 MiB
llama_new_context_with_model: graph nodes  = 518
llama_new_context_with_model: graph splits = 2
INFO [main] model loaded | tid="0x1f435bac0" timestamp=1727923188
time=2024-10-03T08:09:49.159+05:30 level=INFO source=server.go:626 msg="llama runner started in 2.52 seconds"
[GIN] 2024/10/03 - 08:09:49 | 200 |  2.555986541s |       127.0.0.1 | POST     "/api/generate"
[GIN] 2024/10/03 - 08:10:20 | 200 |  7.493648584s |       127.0.0.1 | POST     "/api/chat"

Version

llamafile v0.8.4 llamafile v0.8.13

What operating system are you seeing the problem on?

Mac

Relevant log output

llamafile/bin/llamafile-bench -m llamafiles/llama-3.2-1b-q4_k_m.llamafile -ngl 17 -n 1024 -p 512 --verbose
warning: don't know how to govern your cpu temperature; consider setting the environment variables described in llamafile/govern.cpp
|                     cpu_info |                           model_filename |       size |          test |             t/s |
| ---------------------------: | ---------------------------------------: | ---------: | ------------: | --------------: |
llama_model_loader: loaded meta data with 29 key-value pairs and 147 tensors from llamafiles/llama-3.2-1b-q4_k_m.llamafile (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv   0:                       general.architecture str              = llama
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = Llama 3.2 1B
llama_model_loader: - kv   3:                           general.basename str              = Llama-3.2
llama_model_loader: - kv   4:                         general.size_label str              = 1B
llama_model_loader: - kv   5:                            general.license str              = llama3.2
llama_model_loader: - kv   6:                               general.tags arr[str,6]       = ["facebook", "meta", "pytorch", "llam...
llama_model_loader: - kv   7:                          general.languages arr[str,8]       = ["en", "de", "fr", "it", "pt", "hi", ...
llama_model_loader: - kv   8:                          llama.block_count u32              = 16
llama_model_loader: - kv   9:                       llama.context_length u32              = 131072
llama_model_loader: - kv  10:                     llama.embedding_length u32              = 2048
llama_model_loader: - kv  11:                  llama.feed_forward_length u32              = 8192
llama_model_loader: - kv  12:                 llama.attention.head_count u32              = 32
llama_model_loader: - kv  13:              llama.attention.head_count_kv u32              = 8
llama_model_loader: - kv  14:                       llama.rope.freq_base f32              = 500000.000000
llama_model_loader: - kv  15:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  16:                 llama.attention.key_length u32              = 64
llama_model_loader: - kv  17:               llama.attention.value_length u32              = 64
llama_model_loader: - kv  18:                          general.file_type u32              = 15
llama_model_loader: - kv  19:                           llama.vocab_size u32              = 128256
llama_model_loader: - kv  20:                 llama.rope.dimension_count u32              = 64
llama_model_loader: - kv  21:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  22:                         tokenizer.ggml.pre str              = llama-bpe
llama_model_loader: - kv  23:                      tokenizer.ggml.tokens arr[str,128256]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  24:                  tokenizer.ggml.token_type arr[i32,128256]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  25:                      tokenizer.ggml.merges arr[str,280147]  = ["Ġ Ġ", "Ġ ĠĠĠ", "ĠĠ ĠĠ", "...
llama_model_loader: - kv  26:                tokenizer.ggml.bos_token_id u32              = 128000
llama_model_loader: - kv  27:                tokenizer.ggml.eos_token_id u32              = 128001
llama_model_loader: - kv  28:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:   34 tensors
llama_model_loader: - type q4_K:   96 tensors
llama_model_loader: - type q6_K:   17 tensors
llm_load_vocab: special tokens cache size = 256
llm_load_vocab: token to piece cache size = 0.7999 MB
llm_load_print_meta: format           = GGUF V3 (latest)
llm_load_print_meta: arch             = llama
llm_load_print_meta: vocab type       = BPE
llm_load_print_meta: n_vocab          = 128256
llm_load_print_meta: n_merges         = 280147
llm_load_print_meta: vocab_only       = 0
llm_load_print_meta: n_ctx_train      = 131072
llm_load_print_meta: n_embd           = 2048
llm_load_print_meta: n_layer          = 16
llm_load_print_meta: n_head           = 32
llm_load_print_meta: n_head_kv        = 8
llm_load_print_meta: n_rot            = 64
llm_load_print_meta: n_swa            = 0
llm_load_print_meta: n_embd_head_k    = 64
llm_load_print_meta: n_embd_head_v    = 64
llm_load_print_meta: n_gqa            = 4
llm_load_print_meta: n_embd_k_gqa     = 512
llm_load_print_meta: n_embd_v_gqa     = 512
llm_load_print_meta: f_norm_eps       = 0.0e+00
llm_load_print_meta: f_norm_rms_eps   = 1.0e-05
llm_load_print_meta: f_clamp_kqv      = 0.0e+00
llm_load_print_meta: f_max_alibi_bias = 0.0e+00
llm_load_print_meta: f_logit_scale    = 0.0e+00
llm_load_print_meta: n_ff             = 8192
llm_load_print_meta: n_expert         = 0
llm_load_print_meta: n_expert_used    = 0
llm_load_print_meta: causal attn      = 1
llm_load_print_meta: pooling type     = 0
llm_load_print_meta: rope type        = 0
llm_load_print_meta: rope scaling     = linear
llm_load_print_meta: freq_base_train  = 500000.0
llm_load_print_meta: freq_scale_train = 1
llm_load_print_meta: n_ctx_orig_yarn  = 131072
llm_load_print_meta: rope_finetuned   = unknown
llm_load_print_meta: ssm_d_conv       = 0
llm_load_print_meta: ssm_d_inner      = 0
llm_load_print_meta: ssm_d_state      = 0
llm_load_print_meta: ssm_dt_rank      = 0
llm_load_print_meta: model type       = ?B
llm_load_print_meta: model ftype      = Q4_K - Medium
llm_load_print_meta: model params     = 1.24 B
llm_load_print_meta: model size       = 762.81 MiB (5.18 BPW) 
llm_load_print_meta: general.name     = Llama 3.2 1B
llm_load_print_meta: BOS token        = 128000 '<|begin_of_text|>'
llm_load_print_meta: EOS token        = 128001 '<|end_of_text|>'
llm_load_print_meta: LF token         = 128 'Ä'
llm_load_print_meta: EOT token        = 128009 '<|eot_id|>'
llm_load_print_meta: max token length = 256
llm_load_tensors: ggml ctx size =    0.08 MiB
llm_load_tensors:        CPU buffer size =   762.81 MiB
.......................................................
llama_new_context_with_model: n_ctx      = 512
llama_new_context_with_model: n_batch    = 512
llama_new_context_with_model: n_ubatch   = 512
llama_new_context_with_model: flash_attn = 0
llama_new_context_with_model: freq_base  = 500000.0
llama_new_context_with_model: freq_scale = 1
llama_kv_cache_init:        CPU KV buffer size =    16.00 MiB
llama_new_context_with_model: KV self size  =   16.00 MiB, K (f16):    8.00 MiB, V (f16):    8.00 MiB
llama_new_context_with_model:        CPU  output buffer size =     0.49 MiB
llama_new_context_with_model:        CPU compute buffer size =   254.50 MiB
llama_new_context_with_model: graph nodes  = 518
llama_new_context_with_model: graph splits = 1
| Apple M3 Max (+fp16+dotprod) |                      llama-3.2-1b-q4_k_m | 968.30 MiB |         pp512 |          686.27 |

llama_print_timings:        load time =     812.26 ms
llama_print_timings:      sample time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
llama_print_timings: prompt eval time =    2998.05 ms /  2048 tokens (    1.46 ms per token,   683.11 tokens per second)
llama_print_timings:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
llama_print_timings:       total time =    3049.98 ms /  2049 tokens
llama_new_context_with_model: n_ctx      = 1024
llama_new_context_with_model: n_batch    = 1024
llama_new_context_with_model: n_ubatch   = 512
llama_new_context_with_model: flash_attn = 0
llama_new_context_with_model: freq_base  = 500000.0
llama_new_context_with_model: freq_scale = 1
llama_kv_cache_init:        CPU KV buffer size =    32.00 MiB
llama_new_context_with_model: KV self size  =   32.00 MiB, K (f16):   16.00 MiB, V (f16):   16.00 MiB
llama_new_context_with_model:        CPU  output buffer size =     0.49 MiB
llama_new_context_with_model:        CPU compute buffer size =   254.50 MiB
llama_new_context_with_model: graph nodes  = 518
llama_new_context_with_model: graph splits = 1
| Apple M3 Max (+fp16+dotprod) |                      llama-3.2-1b-q4_k_m | 968.30 MiB |        tg1024 |          114.46 |

llama_print_timings:        load time =    3062.09 ms
llama_print_timings:      sample time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
llama_print_timings: prompt eval time =       0.00 ms /     0 tokens (     nan ms per token,      nan tokens per second)
llama_print_timings:        eval time =   26868.69 ms /  3073 runs   (    8.74 ms per token,   114.37 tokens per second)
llama_print_timings:       total time =   29923.98 ms /  3073 tokens
Djip007 commented 1 month ago

llamafile v0.8.4

This is a old one... can you bench with at least last published release: V0.8.13 (ps: no need to rebuild juste get it from release and give it your model ( -m zzz.gguf / -m zzz.llamafile)

llm_load_print_meta: max token length = 256
llm_load_tensors: ggml ctx size =    0.08 MiB
llm_load_tensors:        CPU buffer size =   762.81 MiB
.......................................................

Also in llamafile case all is compute on CPU not with GPU (Metal) I dont know if it can use GPU on that old release.

mathav95raj commented 1 month ago

@Djip007 My bad. I had done the above tests with latest release v0.8.13 but while filling the github issue, I mentioned the version from the issue default template. I apologise for the error. I have edited it correctly now.

Also in llamafile case all is compute on CPU not with GPU (Metal) I dont know if it can use GPU on that old release.

I have given the number of gpu layers to be offloaded as 17

llamafile/bin/llamafile-bench -m llamafiles/llama-3.2-1b-q4_k_m.llamafile -ngl 17 -n 1024 -p 512 --verbose

Does this mean -ngl is not working as expected?

Djip007 commented 1 month ago

Does this mean -ngl is not working as expected?

Ho yes if I am right... llamafile-bench do only CPU bench for now. (but llamafile did support it ... may be with some "bug" with V0.8.13 : https://github.com/Mozilla-Ocho/llamafile/pull/534)

cjpais commented 1 month ago

llamafile bench currently only supports cpu. I can put up a branch that will enable gpu support tomorrow

the fix in #534 should resolve the issue with gpu performance being slower than llama.cpp

Djip007 commented 1 month ago

llamafile bench currently only supports cpu. I can put up a branch that will enable gpu support tomorrow

Can be nice !

cjpais commented 1 month ago

581 will address gpu support in the bench

mathav95raj commented 1 month ago

able to get correct results now. Thanks!