LlamaCpp Embeddings With Langchain
Last updated
Last updated
Author: Yongdam Kim
Design:
This is a part of LangChain Open Tutorial
This tutorial covers how to perform Text Embedding using Llama-cpp and Langchain.
Llama-cpp is an open-source package implemented in C++ that allows you to use LLMs such as llama very efficiently locally.
In this tutorial, we will create a simple example to measure similarity between Documents and an input Query using Llama-cpp and Langchain.
Set up the environment. You may refer to Environment Setup for more details.
[Note]
langchain-opentutorial is a package that provides a set of easy-to-use environment setup, useful functions and utilities for tutorials.
You can check out the langchain-opentutorial for more details.
%%capture --no-stderr
%pip install langchain-opentutorial# Install required packages
from langchain_opentutorial import package
package.install(
[
"langchain_community",
"llama-cpp-python",
"scikit-learn",
],
verbose=False,
upgrade=False,
)# Set environment variables
from langchain_opentutorial import set_env
set_env(
{
"LANGCHAIN_API_KEY": "",
"LANGCHAIN_TRACING_V2": "true",
"LANGCHAIN_ENDPOINT": "https://api.smith.langchain.com",
"LANGCHAIN_PROJECT": "LlamaCpp-Embeddings-With-Langchain",
}
)Environment variables have been set successfully.You can alternatively set LANGCHAIN_API_KEY in .env file and load it.
[Note] This is not necessary if you've already set LANGCHAIN_API_KEY in previous steps.
from dotenv import load_dotenv
load_dotenv(override=True)FalseLlama-cpp is an open-source project that makes it easy to run large language models (LLMs) locally. It allows you to download and run various LLMs on your own computer, giving you the freedom to experiment with AI models.
To install llama-cpp-python:
pip install llama-cpp-pythonMake sure you have the required environment for C++ compilation (e.g., on Linux or macOS).
Download or specify your chosen embedding model file (e.g., CompendiumLabs/bge-large-en-v1.5-gguf).
Here, we use bge-large-en-v1.5-q8_0.gguf as an example and you can download it from CompendiumLabs/bge-large-en-v1.5-gguf - Hugging Face.
Check that llama-cpp-python can find the model path.
Below, we will demonstrate how to serve a LLaMA model using Llama-cpp. You can follow the official llama-cpp-python documentation for more details.
You can find a variety of embedding models, which typically come in different quantizations (e.g., q4_0, q4_1, q5_0, q8_0, etc.).
1. Search models
You can look for models on Hugging Face or other community websites.
2. Download or Pull a Model
For instance, you could download from Hugging Face if the model is hosted.
3. Verify the Model
Check that the .bin (or .gguf) file is accessible to your environment.
Now that you have installed llama-cpp-python and have downloaded a model, let's see how to load it and use it for text embedding.
Below, we define a Query or some Documents to embed using Llama-cpp within LangChain.
from langchain_community.embeddings import LlamaCppEmbeddings
# Example query and documents
query = "What is LangChain?"
docs = [
"LangChain is an open-source framework designed to facilitate the development of applications powered by large language models (LLMs). It provides tools and components to build end-to-end workflows for tasks like document retrieval, chatbots, summarization, data analysis, and more.",
"Spaghetti Carbonara is a traditional Italian pasta dish made with eggs, cheese, pancetta, and pepper. It's simple yet incredibly delicious. Typically served with spaghetti, but can also be enjoyed with other pasta types.",
"The tropical island of Bali offers stunning beaches, volcanic mountains, lush forests, and vibrant coral reefs. Travelers often visit for surfing, yoga retreats, and the unique Balinese Hindu culture.",
"C++ is a high-performance programming language widely used in system/software development, game programming, and real-time simulations. It supports both procedural and object-oriented paradigms.",
"In astronomy, the Drake Equation is a probabilistic argument used to estimate the number of active, communicative extraterrestrial civilizations in the Milky Way galaxy. It takes into account factors such as star formation rate and fraction of habitable planets.",
]Below is how you can initialize the LlamaCppEmbeddings class by specifying the path to your LLaMA model file (model_path).
For example, you might have a downloaded model path: ./bge-large-en-v1.5-q8_0.gguf.
We demonstrate how to instantiate the embeddings class and then embed queries and documents using Llama-cpp.
# Load the Llama-cpp Embedding Model
model_path = "data/bge-large-en-v1.5-q8_0.gguf" # example path
embedder = LlamaCppEmbeddings(model_path=model_path, n_gpu_layers=-1)
print("Embedding model has been successfully loaded.")llama_load_model_from_file: using device Metal (Apple M3) - 16383 MiB free
llama_model_loader: loaded meta data with 24 key-value pairs and 389 tensors from data/bge-large-en-v1.5-q8_0.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 = bert
llama_model_loader: - kv 1: general.name str = bge-large-en-v1.5
llama_model_loader: - kv 2: bert.block_count u32 = 24
llama_model_loader: - kv 3: bert.context_length u32 = 512
llama_model_loader: - kv 4: bert.embedding_length u32 = 1024
llama_model_loader: - kv 5: bert.feed_forward_length u32 = 4096
llama_model_loader: - kv 6: bert.attention.head_count u32 = 16
llama_model_loader: - kv 7: bert.attention.layer_norm_epsilon f32 = 0.000000
llama_model_loader: - kv 8: general.file_type u32 = 7
llama_model_loader: - kv 9: bert.attention.causal bool = false
llama_model_loader: - kv 10: bert.pooling_type u32 = 2
llama_model_loader: - kv 11: tokenizer.ggml.token_type_count u32 = 2
llama_model_loader: - kv 12: tokenizer.ggml.bos_token_id u32 = 101
llama_model_loader: - kv 13: tokenizer.ggml.eos_token_id u32 = 102
llama_model_loader: - kv 14: tokenizer.ggml.model str = bert
llama_model_loader: - kv 15: tokenizer.ggml.tokens arr[str,30522] = ["[PAD]", "[unused0]", "[unused1]", "...
llama_model_loader: - kv 16: tokenizer.ggml.scores arr[f32,30522] = [-1000.000000, -1000.000000, -1000.00...
llama_model_loader: - kv 17: tokenizer.ggml.token_type arr[i32,30522] = [3, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv 18: tokenizer.ggml.unknown_token_id u32 = 100
llama_model_loader: - kv 19: tokenizer.ggml.seperator_token_id u32 = 102
llama_model_loader: - kv 20: tokenizer.ggml.padding_token_id u32 = 0
llama_model_loader: - kv 21: tokenizer.ggml.cls_token_id u32 = 101
llama_model_loader: - kv 22: tokenizer.ggml.mask_token_id u32 = 103
llama_model_loader: - kv 23: general.quantization_version u32 = 2
llama_model_loader: - type f32: 244 tensors
llama_model_loader: - type q8_0: 145 tensors
llm_load_vocab: control token: 100 '[UNK]' is not marked as EOG
llm_load_vocab: control token: 101 '[CLS]' is not marked as EOG
llm_load_vocab: control token: 0 '[PAD]' is not marked as EOG
llm_load_vocab: control token: 102 '[SEP]' is not marked as EOG
llm_load_vocab: control token: 103 '[MASK]' is not marked as EOG
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 = 5
llm_load_vocab: token to piece cache size = 0.2032 MB
llm_load_print_meta: format = GGUF V3 (latest)
llm_load_print_meta: arch = bert
llm_load_print_meta: vocab type = WPM
llm_load_print_meta: n_vocab = 30522
llm_load_print_meta: n_merges = 0
llm_load_print_meta: vocab_only = 0
llm_load_print_meta: n_ctx_train = 512
llm_load_print_meta: n_embd = 1024
llm_load_print_meta: n_layer = 24
llm_load_print_meta: n_head = 16
llm_load_print_meta: n_head_kv = 16
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 = 1
llm_load_print_meta: n_embd_k_gqa = 1024
llm_load_print_meta: n_embd_v_gqa = 1024
llm_load_print_meta: f_norm_eps = 1.0e-12
llm_load_print_meta: f_norm_rms_eps = 0.0e+00
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 = 4096
llm_load_print_meta: n_expert = 0
llm_load_print_meta: n_expert_used = 0
llm_load_print_meta: causal attn = 0
llm_load_print_meta: pooling type = 2
llm_load_print_meta: rope type = 2
llm_load_print_meta: rope scaling = linear
llm_load_print_meta: freq_base_train = 10000.0
llm_load_print_meta: freq_scale_train = 1
llm_load_print_meta: n_ctx_orig_yarn = 512
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 = 335M
llm_load_print_meta: model ftype = Q8_0
llm_load_print_meta: model params = 334.09 M
llm_load_print_meta: model size = 340.90 MiB (8.56 BPW)
llm_load_print_meta: general.name = bge-large-en-v1.5
llm_load_print_meta: BOS token = 101 '[CLS]'
llm_load_print_meta: EOS token = 102 '[SEP]'
llm_load_print_meta: UNK token = 100 '[UNK]'
llm_load_print_meta: SEP token = 102 '[SEP]'
llm_load_print_meta: PAD token = 0 '[PAD]'
llm_load_print_meta: CLS token = 101 '[CLS]'
llm_load_print_meta: MASK token = 103 '[MASK]'
llm_load_print_meta: LF token = 0 '[PAD]'
llm_load_print_meta: EOG token = 102 '[SEP]'
llm_load_print_meta: max token length = 21
llm_load_tensors: tensor 'token_embd.weight' (q8_0) (and 4 others) cannot be used with preferred buffer type CPU_AARCH64, using CPU instead
ggml_backend_metal_log_allocated_size: allocated buffer, size = 307.23 MiB, ( 307.31 / 16384.02)
llm_load_tensors: offloading 24 repeating layers to GPU
llm_load_tensors: offloading output layer to GPU
llm_load_tensors: offloaded 25/25 layers to GPU
llm_load_tensors: Metal_Mapped model buffer size = 307.23 MiB
llm_load_tensors: CPU_Mapped model buffer size = 33.69 MiB
................................................................................
llama_new_context_with_model: n_seq_max = 1
llama_new_context_with_model: n_ctx = 512
llama_new_context_with_model: n_ctx_per_seq = 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 = 10000.0
llama_new_context_with_model: freq_scale = 1
ggml_metal_init: allocating
ggml_metal_init: found device: Apple M3
ggml_metal_init: picking default device: Apple M3
ggml_metal_init: using embedded metal library
ggml_metal_init: GPU name: Apple M3
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 = true
ggml_metal_init: simdgroup matrix mul. = true
ggml_metal_init: has bfloat = true
ggml_metal_init: use bfloat = false
ggml_metal_init: hasUnifiedMemory = true
ggml_metal_init: recommendedMaxWorkingSetSize = 17179.89 MB
ggml_metal_init: loaded kernel_add 0x104d0f860 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_add_row 0x104e4cd60 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_sub 0x121712be0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_sub_row 0x1078faf00 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul 0x110765820 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_row 0x104d0e830 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_div 0x110765070 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_div_row 0x122f5e620 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_repeat_f32 0x110766540 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_repeat_f16 0x106fff000 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_repeat_i32 0x106ffe710 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_repeat_i16 0x110765e10 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_scale 0x121804080 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_scale_4 0x1078f78d0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_clamp 0x1078fbe10 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_tanh 0x106ffdd80 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_relu 0x106fffa80 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_sigmoid 0x121806040 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_gelu 0x1078fc5a0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_gelu_4 0x121712380 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_gelu_quick 0x121806a80 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_gelu_quick_4 0x121806600 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_silu 0x1078fc860 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_silu_4 0x1078fea90 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_elu 0x1078fdcf0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_soft_max_f16 0x104e4d190 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_soft_max_f16_4 0x1078ff3c0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_soft_max_f32 0x121806d80 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_soft_max_f32_4 0x1078ffc00 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_diag_mask_inf 0x110768220 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_diag_mask_inf_8 0x110768f60 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_f32 0x121713670 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_f16 0x1078feec0 | th_max = 1024 | th_width = 32
ggml_metal_init: skipping kernel_get_rows_bf16 (not supported)
ggml_metal_init: loaded kernel_get_rows_q4_0 0x122105900 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_q4_1 0x122105f80 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_q5_0 0x1107689a0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_q5_1 0x110766dc0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_q8_0 0x11076a280 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_q2_K 0x121713c00 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_q3_K 0x121713fd0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_q4_K 0x121715a80 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_q5_K 0x121716210 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_q6_K 0x121808650 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_iq2_xxs 0x1217166a0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_iq2_xs 0x121715650 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_iq3_xxs 0x122106920 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_iq3_s 0x122106f40 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_iq2_s 0x1218070d0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_iq1_s 0x11076a790 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_iq1_m 0x121807fe0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_iq4_nl 0x1218089d0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_iq4_xs 0x122107700 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_i32 0x122107d40 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_rms_norm 0x122108a30 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_group_norm 0x121717b10 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_norm 0x12180a920 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_ssm_conv_f32 0x121717710 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_ssm_scan_f32 0x121717060 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_f32_f32 0x122108cf0 | th_max = 1024 | th_width = 32
ggml_metal_init: skipping kernel_mul_mv_bf16_f32 (not supported)
ggml_metal_init: skipping kernel_mul_mv_bf16_f32_1row (not supported)
ggml_metal_init: skipping kernel_mul_mv_bf16_f32_l4 (not supported)
ggml_metal_init: skipping kernel_mul_mv_bf16_bf16 (not supported)
ggml_metal_init: loaded kernel_mul_mv_f16_f32 0x121723680 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_f16_f32_1row 0x121723940 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_f16_f32_l4 0x1217249a0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_f16_f16 0x1217244b0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_q4_0_f32 0x11076ba80 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_q4_1_f32 0x1217186e0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_q5_0_f32 0x1217189a0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_q5_1_f32 0x12180af90 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_q8_0_f32 0x12180b530 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_f16_f32_r1_2 0x12180b7f0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_f16_f32_r1_3 0x122109270 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_f16_f32_r1_4 0x1217253f0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_f16_f32_r1_5 0x11076d140 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q4_0_f32_r1_2 0x122109e90 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q4_0_f32_r1_3 0x121725cf0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q4_0_f32_r1_4 0x121718e60 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q4_0_f32_r1_5 0x121726b30 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q4_1_f32_r1_2 0x121727430 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q4_1_f32_r1_3 0x121726480 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q4_1_f32_r1_4 0x12210a620 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q4_1_f32_r1_5 0x12180bb30 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q5_0_f32_r1_2 0x121727bc0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q5_0_f32_r1_3 0x121727f70 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q5_0_f32_r1_4 0x12180c2d0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q5_0_f32_r1_5 0x11076c120 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q5_1_f32_r1_2 0x12210aac0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q5_1_f32_r1_3 0x12180cc70 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q5_1_f32_r1_4 0x11076b4c0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q5_1_f32_r1_5 0x121728e10 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q8_0_f32_r1_2 0x1217296e0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q8_0_f32_r1_3 0x122109a10 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q8_0_f32_r1_4 0x122f5e9d0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q8_0_f32_r1_5 0x12172a000 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q4_K_f32_r1_2 0x12180dbc0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q4_K_f32_r1_3 0x12172a8a0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q4_K_f32_r1_4 0x1217287f0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q4_K_f32_r1_5 0x121829ec0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q5_K_f32_r1_2 0x12210bf40 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q5_K_f32_r1_3 0x121829670 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q5_K_f32_r1_4 0x12172b1c0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q5_K_f32_r1_5 0x12172baa0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q6_K_f32_r1_2 0x121829930 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q6_K_f32_r1_3 0x12210c820 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q6_K_f32_r1_4 0x12172c8d0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q6_K_f32_r1_5 0x11076ea00 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_iq4_nl_f32_r1_2 0x11076d990 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_iq4_nl_f32_r1_3 0x12210ad80 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_iq4_nl_f32_r1_4 0x12182b620 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_iq4_nl_f32_r1_5 0x12210cf70 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_q2_K_f32 0x12172c1d0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_q3_K_f32 0x11076fb20 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_q4_K_f32 0x11076f620 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_q5_K_f32 0x12210da20 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_q6_K_f32 0x12172c490 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_iq2_xxs_f32 0x12172d0a0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_iq2_xs_f32 0x12172ea40 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_iq3_xxs_f32 0x12182c4e0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_iq3_s_f32 0x1107702a0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_iq2_s_f32 0x110770b60 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_iq1_s_f32 0x12172f170 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_iq1_m_f32 0x1107722a0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_iq4_nl_f32 0x110772b40 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_iq4_xs_f32 0x110773450 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_f32_f32 0x12172f520 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_f16_f32 0x12182ada0 | th_max = 1024 | th_width = 32
ggml_metal_init: skipping kernel_mul_mv_id_bf16_f32 (not supported)
ggml_metal_init: loaded kernel_mul_mv_id_q4_0_f32 0x110773e00 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_q4_1_f32 0x110774720 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_q5_0_f32 0x12172fbc0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_q5_1_f32 0x110775000 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_q8_0_f32 0x12182cbb0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_q2_K_f32 0x12210ed20 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_q3_K_f32 0x121730b50 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_q4_K_f32 0x110771af0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_q5_K_f32 0x12172fe80 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_q6_K_f32 0x12210e5b0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_iq2_xxs_f32 0x1107766f0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_iq2_xs_f32 0x1217312b0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_iq3_xxs_f32 0x121731980 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_iq3_s_f32 0x1240133e0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_iq2_s_f32 0x12210e1f0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_iq1_s_f32 0x122111110 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_iq1_m_f32 0x12182d340 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_iq4_nl_f32 0x12210f940 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_iq4_xs_f32 0x122111880 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_f32_f32 0x12182d020 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_f16_f32 0x110771400 | th_max = 1024 | th_width = 32
ggml_metal_init: skipping kernel_mul_mm_bf16_f32 (not supported)
ggml_metal_init: loaded kernel_mul_mm_q4_0_f32 0x110778380 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_q4_1_f32 0x122112f90 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_q5_0_f32 0x12182d600 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_q5_1_f32 0x121731c40 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_q8_0_f32 0x12182de30 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_q2_K_f32 0x110778f00 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_q3_K_f32 0x110777090 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_q4_K_f32 0x122113b40 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_q5_K_f32 0x122113470 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_q6_K_f32 0x122114200 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_iq2_xxs_f32 0x1217326c0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_iq2_xs_f32 0x121732b30 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_iq3_xxs_f32 0x1221126c0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_iq3_s_f32 0x121732fa0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_iq2_s_f32 0x110779da0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_iq1_s_f32 0x1221150c0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_iq1_m_f32 0x122114980 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_iq4_nl_f32 0x121733480 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_iq4_xs_f32 0x11077a520 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_f32_f32 0x11077b330 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_f16_f32 0x12182f380 | th_max = 1024 | th_width = 32
ggml_metal_init: skipping kernel_mul_mm_id_bf16_f32 (not supported)
ggml_metal_init: loaded kernel_mul_mm_id_q4_0_f32 0x12182f870 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_q4_1_f32 0x121733740 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_q5_0_f32 0x1107795c0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_q5_1_f32 0x121830e70 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_q8_0_f32 0x121831ae0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_q2_K_f32 0x110777cc0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_q3_K_f32 0x121832530 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_q4_K_f32 0x1218330b0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_q5_K_f32 0x11077bc50 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_q6_K_f32 0x122116ac0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_iq2_xxs_f32 0x121832cc0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_iq2_xs_f32 0x122f5ee60 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_iq3_xxs_f32 0x11077c570 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_iq3_s_f32 0x11077df30 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_iq2_s_f32 0x122116330 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_iq1_s_f32 0x11077ce80 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_iq1_m_f32 0x121734190 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_iq4_nl_f32 0x121832050 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_iq4_xs_f32 0x11077d510 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_rope_norm_f32 0x121833cf0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_rope_norm_f16 0x121834c60 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_rope_neox_f32 0x11077e6e0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_rope_neox_f16 0x1221174d0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_im2col_f16 0x121734860 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_im2col_f32 0x1217353a0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_im2col_ext_f16 0x122118330 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_im2col_ext_f32 0x11077edf0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_conv_transpose_1d_f32_f32 0x121834590 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_conv_transpose_1d_f16_f32 0x11077fa60 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_upscale_f32 0x11077ffe0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_pad_f32 0x110780550 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_pad_reflect_1d_f32 0x110780c70 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_timestep_embedding_f32 0x121736820 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_arange_f32 0x104e4c870 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_argsort_f32_i32_asc 0x1218358e0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_argsort_f32_i32_desc 0x121835240 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_leaky_relu_f32 0x1218363a0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_f16_h64 0x121737dc0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_f16_h80 0x121837d30 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_f16_h96 0x121736e70 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_f16_h112 0x122119210 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_f16_h128 0x121838640 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_f16_h256 0x1218393c0 | th_max = 1024 | th_width = 32
ggml_metal_init: skipping kernel_flash_attn_ext_bf16_h64 (not supported)
ggml_metal_init: skipping kernel_flash_attn_ext_bf16_h80 (not supported)
ggml_metal_init: skipping kernel_flash_attn_ext_bf16_h96 (not supported)
ggml_metal_init: skipping kernel_flash_attn_ext_bf16_h112 (not supported)
ggml_metal_init: skipping kernel_flash_attn_ext_bf16_h128 (not supported)
ggml_metal_init: skipping kernel_flash_attn_ext_bf16_h256 (not supported)
ggml_metal_init: loaded kernel_flash_attn_ext_q4_0_h64 0x1107832e0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q4_0_h80 0x121738f80 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q4_0_h96 0x121839b00 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q4_0_h112 0x121839f90 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q4_0_h128 0x12183a520 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q4_0_h256 0x12211a060 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q4_1_h64 0x12211a3f0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q4_1_h80 0x12211aff0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q4_1_h96 0x12183b3d0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q4_1_h112 0x12211c200 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q4_1_h128 0x12211bc60 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q4_1_h256 0x1107837f0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q5_0_h64 0x121737240 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q5_0_h80 0x1217395e0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q5_0_h96 0x12173a0e0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q5_0_h112 0x12183a820 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q5_0_h128 0x110782710 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q5_0_h256 0x12183aae0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q5_1_h64 0x12173ad90 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q5_1_h80 0x12173b310 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q5_1_h96 0x12211cae0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q5_1_h112 0x12173b5d0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q5_1_h128 0x12173b890 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q5_1_h256 0x12173bb50 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q8_0_h64 0x110783d40 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q8_0_h80 0x12173c9a0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q8_0_h96 0x12211a730 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q8_0_h112 0x12211ce60 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q8_0_h128 0x12173d0d0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q8_0_h256 0x12183cb20 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_vec_f16_h128 0x12183cde0 | th_max = 1024 | th_width = 32
ggml_metal_init: skipping kernel_flash_attn_ext_vec_bf16_h128 (not supported)
ggml_metal_init: loaded kernel_flash_attn_ext_vec_q4_0_h128 0x12173e6d0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_vec_q4_1_h128 0x12173ece0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_vec_q5_0_h128 0x12183d170 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_vec_q5_1_h128 0x12211d560 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_vec_q8_0_h128 0x12211e460 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_vec_f16_h256 0x12211dc80 | th_max = 1024 | th_width = 32
ggml_metal_init: skipping kernel_flash_attn_ext_vec_bf16_h256 (not supported)
ggml_metal_init: loaded kernel_flash_attn_ext_vec_q4_0_h256 0x12183eab0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_vec_q4_1_h256 0x12183dac0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_vec_q5_0_h256 0x12183de30 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_vec_q5_1_h256 0x110784770 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_vec_q8_0_h256 0x12173f210 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_set_f32 0x110786cb0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_set_i32 0x12211df40 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_cpy_f32_f32 0x110786450 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_cpy_f32_f16 0x12211ff10 | th_max = 1024 | th_width = 32
ggml_metal_init: skipping kernel_cpy_f32_bf16 (not supported)
ggml_metal_init: loaded kernel_cpy_f16_f32 0x122120c50 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_cpy_f16_f16 0x12211f6d0 | th_max = 1024 | th_width = 32
ggml_metal_init: skipping kernel_cpy_bf16_f32 (not supported)
ggml_metal_init: skipping kernel_cpy_bf16_bf16 (not supported)
ggml_metal_init: loaded kernel_cpy_f32_q8_0 0x12183f920 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_cpy_f32_q4_0 0x12183f2a0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_cpy_f32_q4_1 0x122121380 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_cpy_f32_q5_0 0x1221221d0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_cpy_f32_q5_1 0x122122a80 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_cpy_f32_iq4_nl 0x12173fe30 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_concat 0x121740ae0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_sqr 0x110785c10 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_sqrt 0x1107873f0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_sin 0x110788120 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_cos 0x110789070 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_sum_rows 0x12183ff00 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_argmax 0x110789b10 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_pool_2d_avg_f32 0x121741210 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_pool_2d_max_f32 0x122121d00 | th_max = 1024 | th_width = 32
llama_kv_cache_init: Metal KV buffer size = 48.00 MiB
llama_new_context_with_model: KV self size = 48.00 MiB, K (f16): 24.00 MiB, V (f16): 24.00 MiB
llama_new_context_with_model: CPU output buffer size = 0.00 MiB
llama_new_context_with_model: Metal compute buffer size = 25.00 MiB
llama_new_context_with_model: CPU compute buffer size = 5.01 MiB
llama_new_context_with_model: graph nodes = 849
llama_new_context_with_model: graph splits = 2
Metal : EMBED_LIBRARY = 1 | CPU : NEON = 1 | ARM_FMA = 1 | FP16_VA = 1 | MATMUL_INT8 = 1 | ACCELERATE = 1 | AARCH64_REPACK = 1 | Embedding model has been successfully loaded.
Model metadata: {'tokenizer.ggml.cls_token_id': '101', 'tokenizer.ggml.padding_token_id': '0', 'tokenizer.ggml.seperator_token_id': '102', 'tokenizer.ggml.unknown_token_id': '100', 'general.quantization_version': '2', 'tokenizer.ggml.token_type_count': '2', 'general.file_type': '7', 'tokenizer.ggml.eos_token_id': '102', 'bert.context_length': '512', 'bert.pooling_type': '2', 'tokenizer.ggml.bos_token_id': '101', 'bert.attention.head_count': '16', 'bert.feed_forward_length': '4096', 'tokenizer.ggml.mask_token_id': '103', 'tokenizer.ggml.model': 'bert', 'bert.attention.causal': 'false', 'general.name': 'bge-large-en-v1.5', 'bert.block_count': '24', 'bert.attention.layer_norm_epsilon': '0.000000', 'bert.embedding_length': '1024', 'general.architecture': 'bert'}
Using fallback chat format: llama-2Now let's embed both the query and the documents. We will verify the dimension of the output vectors.
However, there is currently one issue that cannot be resolved when using the latest model with LlamaCppEmbeddings. I will post the link to the issue below, so please check it out and if it is resolved in the latest version, you can use it as instructed in the original langchain official tutorial.
Issue link : https://github.com/langchain-ai/langchain/issues/22532
# from langchain tutorial
"""
embedded_query = llama_embeddings.embed_query(query)
embedded_docs = llama_embeddings.embed_documents(docs)
print(f"Embedding Dimension Output (Query): {len(embedded_query)}")
print(f"Embedding Dimension Output (Docs): {len(embedded_docs[0])}")
"""
# Overridden version of the LlamaCppEmbeddings class
from typing import List
from langchain_community.embeddings.llamacpp import LlamaCppEmbeddings
class CustomLlamaCppEmbeddings(LlamaCppEmbeddings):
def embed_documents(self, texts: List[str]) -> List[List[float]]:
"""Embed a list of documents using the Llama model.
Args:
texts: The list of texts to embed.
Returns:
List of embeddings, one for each text.
"""
embeddings = [self.client.embed([text])[0] for text in texts]
return [list(map(float, e)) for e in embeddings]
def embed_query(self, text: str) -> List[float]:
"""Embed a query using the Llama model.
Args:
text: The text to embed.
Returns:
Embeddings for the text.
"""
embedding = self.client.embed(text)[0]
return list(map(float, embedding))
c_embedder = CustomLlamaCppEmbeddings(model_path=model_path, n_gpu_layers=-1)
embedded_query = c_embedder.embed_query([query])
embedded_docs = c_embedder.embed_documents(docs)llama_load_model_from_file: using device Metal (Apple M3) - 15997 MiB free
llama_model_loader: loaded meta data with 24 key-value pairs and 389 tensors from data/bge-large-en-v1.5-q8_0.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 = bert
llama_model_loader: - kv 1: general.name str = bge-large-en-v1.5
llama_model_loader: - kv 2: bert.block_count u32 = 24
llama_model_loader: - kv 3: bert.context_length u32 = 512
llama_model_loader: - kv 4: bert.embedding_length u32 = 1024
llama_model_loader: - kv 5: bert.feed_forward_length u32 = 4096
llama_model_loader: - kv 6: bert.attention.head_count u32 = 16
llama_model_loader: - kv 7: bert.attention.layer_norm_epsilon f32 = 0.000000
llama_model_loader: - kv 8: general.file_type u32 = 7
llama_model_loader: - kv 9: bert.attention.causal bool = false
llama_model_loader: - kv 10: bert.pooling_type u32 = 2
llama_model_loader: - kv 11: tokenizer.ggml.token_type_count u32 = 2
llama_model_loader: - kv 12: tokenizer.ggml.bos_token_id u32 = 101
llama_model_loader: - kv 13: tokenizer.ggml.eos_token_id u32 = 102
llama_model_loader: - kv 14: tokenizer.ggml.model str = bert
llama_model_loader: - kv 15: tokenizer.ggml.tokens arr[str,30522] = ["[PAD]", "[unused0]", "[unused1]", "...
llama_model_loader: - kv 16: tokenizer.ggml.scores arr[f32,30522] = [-1000.000000, -1000.000000, -1000.00...
llama_model_loader: - kv 17: tokenizer.ggml.token_type arr[i32,30522] = [3, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv 18: tokenizer.ggml.unknown_token_id u32 = 100
llama_model_loader: - kv 19: tokenizer.ggml.seperator_token_id u32 = 102
llama_model_loader: - kv 20: tokenizer.ggml.padding_token_id u32 = 0
llama_model_loader: - kv 21: tokenizer.ggml.cls_token_id u32 = 101
llama_model_loader: - kv 22: tokenizer.ggml.mask_token_id u32 = 103
llama_model_loader: - kv 23: general.quantization_version u32 = 2
llama_model_loader: - type f32: 244 tensors
llama_model_loader: - type q8_0: 145 tensors
llm_load_vocab: control token: 100 '[UNK]' is not marked as EOG
llm_load_vocab: control token: 101 '[CLS]' is not marked as EOG
llm_load_vocab: control token: 0 '[PAD]' is not marked as EOG
llm_load_vocab: control token: 102 '[SEP]' is not marked as EOG
llm_load_vocab: control token: 103 '[MASK]' is not marked as EOG
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 = 5
llm_load_vocab: token to piece cache size = 0.2032 MB
llm_load_print_meta: format = GGUF V3 (latest)
llm_load_print_meta: arch = bert
llm_load_print_meta: vocab type = WPM
llm_load_print_meta: n_vocab = 30522
llm_load_print_meta: n_merges = 0
llm_load_print_meta: vocab_only = 0
llm_load_print_meta: n_ctx_train = 512
llm_load_print_meta: n_embd = 1024
llm_load_print_meta: n_layer = 24
llm_load_print_meta: n_head = 16
llm_load_print_meta: n_head_kv = 16
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 = 1
llm_load_print_meta: n_embd_k_gqa = 1024
llm_load_print_meta: n_embd_v_gqa = 1024
llm_load_print_meta: f_norm_eps = 1.0e-12
llm_load_print_meta: f_norm_rms_eps = 0.0e+00
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 = 4096
llm_load_print_meta: n_expert = 0
llm_load_print_meta: n_expert_used = 0
llm_load_print_meta: causal attn = 0
llm_load_print_meta: pooling type = 2
llm_load_print_meta: rope type = 2
llm_load_print_meta: rope scaling = linear
llm_load_print_meta: freq_base_train = 10000.0
llm_load_print_meta: freq_scale_train = 1
llm_load_print_meta: n_ctx_orig_yarn = 512
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 = 335M
llm_load_print_meta: model ftype = Q8_0
llm_load_print_meta: model params = 334.09 M
llm_load_print_meta: model size = 340.90 MiB (8.56 BPW)
llm_load_print_meta: general.name = bge-large-en-v1.5
llm_load_print_meta: BOS token = 101 '[CLS]'
llm_load_print_meta: EOS token = 102 '[SEP]'
llm_load_print_meta: UNK token = 100 '[UNK]'
llm_load_print_meta: SEP token = 102 '[SEP]'
llm_load_print_meta: PAD token = 0 '[PAD]'
llm_load_print_meta: CLS token = 101 '[CLS]'
llm_load_print_meta: MASK token = 103 '[MASK]'
llm_load_print_meta: LF token = 0 '[PAD]'
llm_load_print_meta: EOG token = 102 '[SEP]'
llm_load_print_meta: max token length = 21
llm_load_tensors: tensor 'token_embd.weight' (q8_0) (and 4 others) cannot be used with preferred buffer type CPU_AARCH64, using CPU instead
ggml_backend_metal_log_allocated_size: allocated buffer, size = 307.23 MiB, ( 693.44 / 16384.02)
llm_load_tensors: offloading 24 repeating layers to GPU
llm_load_tensors: offloading output layer to GPU
llm_load_tensors: offloaded 25/25 layers to GPU
llm_load_tensors: Metal_Mapped model buffer size = 307.23 MiB
llm_load_tensors: CPU_Mapped model buffer size = 33.69 MiB
................................................................................
llama_new_context_with_model: n_seq_max = 1
llama_new_context_with_model: n_ctx = 512
llama_new_context_with_model: n_ctx_per_seq = 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 = 10000.0
llama_new_context_with_model: freq_scale = 1
ggml_metal_init: allocating
ggml_metal_init: found device: Apple M3
ggml_metal_init: picking default device: Apple M3
ggml_metal_init: using embedded metal library
ggml_metal_init: GPU name: Apple M3
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 = true
ggml_metal_init: simdgroup matrix mul. = true
ggml_metal_init: has bfloat = true
ggml_metal_init: use bfloat = false
ggml_metal_init: hasUnifiedMemory = true
ggml_metal_init: recommendedMaxWorkingSetSize = 17179.89 MB
ggml_metal_init: loaded kernel_add 0x104541960 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_add_row 0x1218415b0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_sub 0x121841870 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_sub_row 0x121841b30 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul 0x1107e7c70 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_row 0x1218408e0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_div 0x121735910 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_div_row 0x1107e74c0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_repeat_f32 0x1107e7780 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_repeat_f16 0x12210d300 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_repeat_i32 0x1107e8e30 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_repeat_i16 0x12187e120 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_scale 0x104ff3320 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_scale_4 0x104e4d4a0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_clamp 0x1107e9920 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_tanh 0x122122e50 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_relu 0x1107ea300 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_sigmoid 0x122123110 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_gelu 0x1221233d0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_gelu_4 0x12187e660 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_gelu_quick 0x1107ea950 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_gelu_quick_4 0x12187ebb0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_silu 0x12187f0f0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_silu_4 0x12187f630 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_elu 0x12187fb70 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_soft_max_f16 0x1218800b0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_soft_max_f16_4 0x104ff35e0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_soft_max_f32 0x121880370 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_soft_max_f32_4 0x122123690 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_diag_mask_inf 0x121880630 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_diag_mask_inf_8 0x1107ebd20 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_f32 0x1107ead60 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_f16 0x1218808f0 | th_max = 1024 | th_width = 32
ggml_metal_init: skipping kernel_get_rows_bf16 (not supported)
ggml_metal_init: loaded kernel_get_rows_q4_0 0x121880bb0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_q4_1 0x121880e70 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_q5_0 0x1107ecac0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_q5_1 0x121881130 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_q8_0 0x104ff38a0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_q2_K 0x1218813f0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_q3_K 0x104ff3b60 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_q4_K 0x1107ed960 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_q5_K 0x1218816b0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_q6_K 0x121881970 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_iq2_xxs 0x121881c30 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_iq2_xs 0x1107edc80 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_iq3_xxs 0x1107edf40 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_iq3_s 0x1107ee2f0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_iq2_s 0x121881ef0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_iq1_s 0x1218821b0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_iq1_m 0x121882470 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_iq4_nl 0x122123950 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_iq4_xs 0x121882730 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_get_rows_i32 0x1218829f0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_rms_norm 0x121882cb0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_group_norm 0x1107ef9d0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_norm 0x1107ee810 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_ssm_conv_f32 0x121882f70 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_ssm_scan_f32 0x104ff4590 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_f32_f32 0x1107efdc0 | th_max = 1024 | th_width = 32
ggml_metal_init: skipping kernel_mul_mv_bf16_f32 (not supported)
ggml_metal_init: skipping kernel_mul_mv_bf16_f32_1row (not supported)
ggml_metal_init: skipping kernel_mul_mv_bf16_f32_l4 (not supported)
ggml_metal_init: skipping kernel_mul_mv_bf16_bf16 (not supported)
ggml_metal_init: loaded kernel_mul_mv_f16_f32 0x1107f00d0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_f16_f32_1row 0x121883670 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_f16_f32_l4 0x104ff4850 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_f16_f16 0x121883c30 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_q4_0_f32 0x104ff5920 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_q4_1_f32 0x122123c10 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_q5_0_f32 0x121883ef0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_q5_1_f32 0x1107f04d0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_q8_0_f32 0x1107f1c10 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_f16_f32_r1_2 0x1107f1690 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_f16_f32_r1_3 0x1218841b0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_f16_f32_r1_4 0x122123ed0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_f16_f32_r1_5 0x122124190 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q4_0_f32_r1_2 0x121884620 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q4_0_f32_r1_3 0x1107f2060 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q4_0_f32_r1_4 0x121884d30 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q4_0_f32_r1_5 0x122124450 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q4_1_f32_r1_2 0x1107f28e0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q4_1_f32_r1_3 0x1107f3a20 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q4_1_f32_r1_4 0x12170d600 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q4_1_f32_r1_5 0x1107f3220 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q5_0_f32_r1_2 0x1107f34e0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q5_0_f32_r1_3 0x121885310 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q5_0_f32_r1_4 0x1107f4b40 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q5_0_f32_r1_5 0x121886150 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q5_1_f32_r1_2 0x1107f58a0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q5_1_f32_r1_3 0x12170d8c0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q5_1_f32_r1_4 0x1107f6690 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q5_1_f32_r1_5 0x122124710 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q8_0_f32_r1_2 0x1221249d0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q8_0_f32_r1_3 0x121886890 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q8_0_f32_r1_4 0x121887770 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q8_0_f32_r1_5 0x12170db80 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q4_K_f32_r1_2 0x122124c90 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q4_K_f32_r1_3 0x1218859a0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q4_K_f32_r1_4 0x122124f50 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q4_K_f32_r1_5 0x122125210 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q5_K_f32_r1_2 0x1221254d0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q5_K_f32_r1_3 0x122125790 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q5_K_f32_r1_4 0x12170e010 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q5_K_f32_r1_5 0x121888bc0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q6_K_f32_r1_2 0x122125a50 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q6_K_f32_r1_3 0x1218880c0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q6_K_f32_r1_4 0x122126090 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_q6_K_f32_r1_5 0x1221269b0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_iq4_nl_f32_r1_2 0x121888540 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_iq4_nl_f32_r1_3 0x1221272c0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_iq4_nl_f32_r1_4 0x1107f6b30 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_ext_iq4_nl_f32_r1_5 0x1107f7440 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_q2_K_f32 0x122127b70 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_q3_K_f32 0x121889ca0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_q4_K_f32 0x12188a680 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_q5_K_f32 0x122127f40 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_q6_K_f32 0x122f0cdd0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_iq2_xxs_f32 0x1240140c0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_iq2_xs_f32 0x122f0c910 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_iq3_xxs_f32 0x122128990 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_iq3_s_f32 0x12188aa70 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_iq2_s_f32 0x12188bd00 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_iq1_s_f32 0x122128310 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_iq1_m_f32 0x12170ed50 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_iq4_nl_f32 0x1107f7b60 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_iq4_xs_f32 0x12188b5b0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_f32_f32 0x12170e760 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_f16_f32 0x12170f640 | th_max = 1024 | th_width = 32
ggml_metal_init: skipping kernel_mul_mv_id_bf16_f32 (not supported)
ggml_metal_init: loaded kernel_mul_mv_id_q4_0_f32 0x12188b870 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_q4_1_f32 0x12188d260 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_q5_0_f32 0x12188dbc0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_q5_1_f32 0x12188c6c0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_q8_0_f32 0x12188cbb0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_q2_K_f32 0x122129110 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_q3_K_f32 0x1107f89b0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_q4_K_f32 0x12170ff20 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_q5_K_f32 0x12188ee30 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_q6_K_f32 0x1107f7f30 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_iq2_xxs_f32 0x1107f9140 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_iq2_xs_f32 0x122129770 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_iq3_xxs_f32 0x12188fc70 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_iq3_s_f32 0x1107f9f60 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_iq2_s_f32 0x1218904c0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_iq1_s_f32 0x121712640 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_iq1_m_f32 0x121710970 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_iq4_nl_f32 0x12212b310 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mv_id_iq4_xs_f32 0x121890c50 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_f32_f32 0x121891370 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_f16_f32 0x121891790 | th_max = 1024 | th_width = 32
ggml_metal_init: skipping kernel_mul_mm_bf16_f32 (not supported)
ggml_metal_init: loaded kernel_mul_mm_q4_0_f32 0x12212a6d0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_q4_1_f32 0x121892420 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_q5_0_f32 0x12212bd10 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_q5_1_f32 0x1107fae50 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_q8_0_f32 0x12212cb30 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_q2_K_f32 0x121711b70 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_q3_K_f32 0x12212d350 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_q4_K_f32 0x12212dba0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_q5_K_f32 0x1107fb6e0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_q6_K_f32 0x1107f96c0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_iq2_xxs_f32 0x12212c0c0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_iq2_xs_f32 0x1107fc3b0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_iq3_xxs_f32 0x12212e430 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_iq3_s_f32 0x12212f060 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_iq2_s_f32 0x1107fcb40 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_iq1_s_f32 0x121741bb0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_iq1_m_f32 0x121711580 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_iq4_nl_f32 0x1107fc670 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_iq4_xs_f32 0x121892720 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_f32_f32 0x122130920 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_f16_f32 0x121893770 | th_max = 1024 | th_width = 32
ggml_metal_init: skipping kernel_mul_mm_id_bf16_f32 (not supported)
ggml_metal_init: loaded kernel_mul_mm_id_q4_0_f32 0x12212eb10 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_q4_1_f32 0x1217423e0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_q5_0_f32 0x122131950 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_q5_1_f32 0x121894230 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_q8_0_f32 0x1107fe260 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_q2_K_f32 0x1107fd550 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_q3_K_f32 0x122132240 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_q4_K_f32 0x122132a10 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_q5_K_f32 0x121711170 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_q6_K_f32 0x121743ca0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_iq2_xxs_f32 0x1221333e0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_iq2_xs_f32 0x122133d00 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_iq3_xxs_f32 0x1107fef00 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_iq3_s_f32 0x1218949f0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_iq2_s_f32 0x121744430 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_iq1_s_f32 0x1107fe940 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_iq1_m_f32 0x122132cd0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_iq4_nl_f32 0x121894f60 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_mul_mm_id_iq4_xs_f32 0x121895b40 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_rope_norm_f32 0x1107fdac0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_rope_norm_f16 0x122134410 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_rope_neox_f32 0x121745310 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_rope_neox_f16 0x12170b210 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_im2col_f16 0x1107ffbc0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_im2col_f32 0x1221346d0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_im2col_ext_f16 0x124404150 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_im2col_ext_f32 0x12170b640 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_conv_transpose_1d_f32_f32 0x122135100 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_conv_transpose_1d_f16_f32 0x122f0d2b0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_upscale_f32 0x1240146e0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_pad_f32 0x124404850 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_pad_reflect_1d_f32 0x121895640 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_timestep_embedding_f32 0x1221357b0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_arange_f32 0x1244051a0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_argsort_f32_i32_asc 0x1218965d0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_argsort_f32_i32_desc 0x124405890 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_leaky_relu_f32 0x121896c30 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_f16_h64 0x1221370b0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_f16_h80 0x1217459d0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_f16_h96 0x121897db0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_f16_h112 0x124407ec0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_f16_h128 0x124408180 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_f16_h256 0x124408950 | th_max = 1024 | th_width = 32
ggml_metal_init: skipping kernel_flash_attn_ext_bf16_h64 (not supported)
ggml_metal_init: skipping kernel_flash_attn_ext_bf16_h80 (not supported)
ggml_metal_init: skipping kernel_flash_attn_ext_bf16_h96 (not supported)
ggml_metal_init: skipping kernel_flash_attn_ext_bf16_h112 (not supported)
ggml_metal_init: skipping kernel_flash_attn_ext_bf16_h128 (not supported)
ggml_metal_init: skipping kernel_flash_attn_ext_bf16_h256 (not supported)
ggml_metal_init: loaded kernel_flash_attn_ext_q4_0_h64 0x124409820 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q4_0_h80 0x124409220 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q4_0_h96 0x122137860 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q4_0_h112 0x12440a140 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q4_0_h128 0x122137e80 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q4_0_h256 0x122136a60 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q4_1_h64 0x122138730 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q4_1_h80 0x122139c70 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q4_1_h96 0x121898570 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q4_1_h112 0x12440aa70 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q4_1_h128 0x12440be00 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q4_1_h256 0x122139310 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q5_0_h64 0x12213ada0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q5_0_h80 0x1218994f0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q5_0_h96 0x12440b440 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q5_0_h112 0x121899e00 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q5_0_h128 0x12440b800 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q5_0_h256 0x12440cb90 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q5_1_h64 0x12170cb60 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q5_1_h80 0x12170bb60 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q5_1_h96 0x12189adb0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q5_1_h112 0x121747070 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q5_1_h128 0x1217479b0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q5_1_h256 0x12213ba30 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q8_0_h64 0x12213a900 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q8_0_h80 0x12440db70 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q8_0_h96 0x12213c970 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q8_0_h112 0x12440e480 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q8_0_h128 0x12213d270 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_q8_0_h256 0x12440cf20 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_vec_f16_h128 0x12213db70 | th_max = 1024 | th_width = 32
ggml_metal_init: skipping kernel_flash_attn_ext_vec_bf16_h128 (not supported)
ggml_metal_init: loaded kernel_flash_attn_ext_vec_q4_0_h128 0x121747fd0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_vec_q4_1_h128 0x12189b9d0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_vec_q5_0_h128 0x12189b580 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_vec_q5_1_h128 0x12213e4d0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_vec_q8_0_h128 0x12189a960 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_vec_f16_h256 0x12440e7c0 | th_max = 1024 | th_width = 32
ggml_metal_init: skipping kernel_flash_attn_ext_vec_bf16_h256 (not supported)
ggml_metal_init: loaded kernel_flash_attn_ext_vec_q4_0_h256 0x12213eda0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_vec_q4_1_h256 0x12189d4d0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_vec_q5_0_h256 0x12440f6d0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_vec_q5_1_h256 0x12440eca0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_flash_attn_ext_vec_q8_0_h256 0x12170c550 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_set_f32 0x12440ff80 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_set_i32 0x12189ced0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_cpy_f32_f32 0x124410240 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_cpy_f32_f16 0x1217483e0 | th_max = 1024 | th_width = 32
ggml_metal_init: skipping kernel_cpy_f32_bf16 (not supported)
ggml_metal_init: loaded kernel_cpy_f16_f32 0x12213f1c0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_cpy_f16_f16 0x12213fae0 | th_max = 1024 | th_width = 32
ggml_metal_init: skipping kernel_cpy_bf16_f32 (not supported)
ggml_metal_init: skipping kernel_cpy_bf16_bf16 (not supported)
ggml_metal_init: loaded kernel_cpy_f32_q8_0 0x12189ddd0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_cpy_f32_q4_0 0x121749900 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_cpy_f32_q4_1 0x124410500 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_cpy_f32_q5_0 0x1217493a0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_cpy_f32_q5_1 0x124411b70 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_cpy_f32_iq4_nl 0x1221404f0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_concat 0x12189e8c0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_sqr 0x12189c5f0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_sqrt 0x121748e80 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_sin 0x1244113b0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_cos 0x12174acf0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_sum_rows 0x124412e60 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_argmax 0x12174bbb0 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_pool_2d_avg_f32 0x124412130 | th_max = 1024 | th_width = 32
ggml_metal_init: loaded kernel_pool_2d_max_f32 0x1244145f0 | th_max = 1024 | th_width = 32
llama_kv_cache_init: Metal KV buffer size = 48.00 MiB
llama_new_context_with_model: KV self size = 48.00 MiB, K (f16): 24.00 MiB, V (f16): 24.00 MiB
llama_new_context_with_model: CPU output buffer size = 0.00 MiB
llama_new_context_with_model: Metal compute buffer size = 25.00 MiB
llama_new_context_with_model: CPU compute buffer size = 5.01 MiB
llama_new_context_with_model: graph nodes = 849
llama_new_context_with_model: graph splits = 2
Metal : EMBED_LIBRARY = 1 | CPU : NEON = 1 | ARM_FMA = 1 | FP16_VA = 1 | MATMUL_INT8 = 1 | ACCELERATE = 1 | AARCH64_REPACK = 1 | Metal : EMBED_LIBRARY = 1 | CPU : NEON = 1 | ARM_FMA = 1 | FP16_VA = 1 | MATMUL_INT8 = 1 | ACCELERATE = 1 | AARCH64_REPACK = 1 |
Model metadata: {'tokenizer.ggml.cls_token_id': '101', 'tokenizer.ggml.padding_token_id': '0', 'tokenizer.ggml.seperator_token_id': '102', 'tokenizer.ggml.unknown_token_id': '100', 'general.quantization_version': '2', 'tokenizer.ggml.token_type_count': '2', 'general.file_type': '7', 'tokenizer.ggml.eos_token_id': '102', 'bert.context_length': '512', 'bert.pooling_type': '2', 'tokenizer.ggml.bos_token_id': '101', 'bert.attention.head_count': '16', 'bert.feed_forward_length': '4096', 'tokenizer.ggml.mask_token_id': '103', 'tokenizer.ggml.model': 'bert', 'bert.attention.causal': 'false', 'general.name': 'bge-large-en-v1.5', 'bert.block_count': '24', 'bert.attention.layer_norm_epsilon': '0.000000', 'bert.embedding_length': '1024', 'general.architecture': 'bert'}
Using fallback chat format: llama-2
llama_perf_context_print: load time = 59.73 ms
llama_perf_context_print: prompt eval time = 0.00 ms / 8 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 = 59.76 ms / 9 tokens
llama_perf_context_print: load time = 59.73 ms
llama_perf_context_print: prompt eval time = 0.00 ms / 62 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 = 29.66 ms / 63 tokens
llama_perf_context_print: load time = 59.73 ms
llama_perf_context_print: prompt eval time = 0.00 ms / 47 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 = 28.45 ms / 48 tokens
llama_perf_context_print: load time = 59.73 ms
llama_perf_context_print: prompt eval time = 0.00 ms / 40 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 = 31.09 ms / 41 tokens
llama_perf_context_print: load time = 59.73 ms
llama_perf_context_print: prompt eval time = 0.00 ms / 40 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 = 30.95 ms / 41 tokens
llama_perf_context_print: load time = 59.73 ms
llama_perf_context_print: prompt eval time = 0.00 ms / 54 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 = 28.12 ms / 55 tokensTo check whether the embedding results are output as expected, I output the dimensions of each embedding vector.
print("Query embedding dimension:", len(embedded_query))
print("Document embedding dimension:", len(embedded_docs[0]))Query embedding dimension: 1024
Document embedding dimension: 1024We can use the vector representations of the query and documents to calculate similarity. Here, we use the cosine similarity provided by scikit-learn.
from sklearn.metrics.pairwise import cosine_similarity
import numpy as np
# Calculate Cosine Similarity
similarities = cosine_similarity([embedded_query], embedded_docs)[0]
print(similarities)
# Sort indices in ascending order.
sorted_indices = np.argsort(similarities)[::-1]
print(f"[Query] {query}\n====================================")
for i, idx in enumerate(sorted_indices):
print(f"[{i}] similarity: {similarities[idx]:.3f} | {docs[idx]}")
print()[0.87119322 0.37936658 0.29399348 0.49741913 0.38813166]
[Query] What is LangChain?
====================================
[0] similarity: 0.871 | LangChain is an open-source framework designed to facilitate the development of applications powered by large language models (LLMs). It provides tools and components to build end-to-end workflows for tasks like document retrieval, chatbots, summarization, data analysis, and more.
[1] similarity: 0.497 | C++ is a high-performance programming language widely used in system/software development, game programming, and real-time simulations. It supports both procedural and object-oriented paradigms.
[2] similarity: 0.388 | In astronomy, the Drake Equation is a probabilistic argument used to estimate the number of active, communicative extraterrestrial civilizations in the Milky Way galaxy. It takes into account factors such as star formation rate and fraction of habitable planets.
[3] similarity: 0.379 | Spaghetti Carbonara is a traditional Italian pasta dish made with eggs, cheese, pancetta, and pepper. It's simple yet incredibly delicious. Typically served with spaghetti, but can also be enjoyed with other pasta types.
[4] similarity: 0.294 | The tropical island of Bali offers stunning beaches, volcanic mountains, lush forests, and vibrant coral reefs. Travelers often visit for surfing, yoga retreats, and the unique Balinese Hindu culture.
This concludes the Llama-cpp Embeddings With Langchain tutorial in the style of the original reference notebook.
