llama-cpp-pythonをbitnet.cppに対応させてみる

2024年10月19日 12:25

以下の記事でMackBook Proで動かしてみたのですが、毎回モデルをロードさせるのが~~気に食わない~~時間がもったいないと思ったので、試します。

1. 環境構築

(1) bitnet.cpp

こちらは淡々と。

python3 -m venv bitnet
cd $_
source bin/activate

git clone --recursive https://github.com/microsoft/BitNet.git
cd BitNet

pip install -r requirements.txt

python setup_env.py --hf-repo HF1BitLLM/Llama3-8B-1.58-100B-tokens -q i2_s

setup_env.pyで量子化モデルを作成しています。このコマンドを先に実行しておかないと、llama-cpp-pythonビルド時に必要となるヘッダファイルが所定のディレクトリに格納されないためです。

(2) llama-cpp-python

続いてllama-cpp-pythonです。こちらをbitnet.cppに対応させます。
BitNetのディレクトリから一つ上に移動して、

% pwd
/path/to/bitnet/BitNet
% cd ..

続いて、llama-cpp-pythonのcloneです。READMEでは、llama.cppもcloneするため--recurse-submodules付きとなっていますが、今回はbitnet.cpp対応のllama.cppを使用するので、--recurse-submodulesは付けません。

git clone https://github.com/abetlen/llama-cpp-python.git
cd llama-cpp-python

bitnet.cppのllama.cppをcloneしてブランチを切り替えます。

cd vendor
git clone https://github.com/Eddie-Wang1120/llama.cpp.git
cd llama.cpp
git checkout 406a5036f9a8aaee9ec5e96e652f61691340fe95

カレントディレクトリを変更しておきます。

% cd ../../..
% pwd
/path/to/bitnet

2. llama-cpp-pythonへのパッチ

パスが長くなるので、llama-cpp-pythonディレクトリに移動します。

cd llama-cpp-python

(1) ./vendor/llama.cpp/CMakeLists.txt

差分はこちら。

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 157d9c86..e5773c26 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -63,7 +63,7 @@ option(LLAMA_SANITIZE_ADDRESS   "llama: enable address sanitizer"   OFF)
 option(LLAMA_SANITIZE_UNDEFINED "llama: enable undefined sanitizer" OFF)
 
 # utils
-option(LLAMA_BUILD_COMMON "llama: build common utils library" ${LLAMA_STANDALONE})
+option(LLAMA_BUILD_COMMON "llama: build common utils library" ON)
 
 # extra artifacts
 option(LLAMA_BUILD_TESTS    "llama: build tests"          ${LLAMA_STANDALONE})
@@ -198,4 +198,8 @@ install(FILES "${CMAKE_CURRENT_BINARY_DIR}/llama.pc"
 #
 
 add_subdirectory(common)
-add_subdirectory(examples)
+
+if (LLAMA_BUILD_EXAMPLES)
+    add_subdirectory(examples)
+endif()

(2) ヘッダーファイル等の参照先の補正

この状態で pip install -e . とすると以下のようなファイルが見つからないエラーが発生するので、

      -- Configuring done (0.9s)
      CMake Error at vendor/llama.cpp/ggml/src/CMakeLists.txt:1324 (add_library):
        Cannot find source file:
      
          ../../../../include/bitnet-lut-kernels.h

これに対応するため、ln -s します。

ln -s ../BitNet/include .
ln -s ../BitNet/src .

この状態で pip install -e . とするとインストールするファイルが見つからないと怒られます。

      *** Installing project into wheel...
      -- Install configuration: "Release"
      -- Installing: /var/folders/vb/tjf174hj1835lfyj9vw5066c0000gn/T/tmp8p58f179/wheel/platlib/lib/libggml.dylib
      -- Installing: /var/folders/vb/tjf174hj1835lfyj9vw5066c0000gn/T/tmp8p58f179/wheel/platlib/include/ggml.h
      -- Installing: /var/folders/vb/tjf174hj1835lfyj9vw5066c0000gn/T/tmp8p58f179/wheel/platlib/include/ggml-alloc.h
      -- Installing: /var/folders/vb/tjf174hj1835lfyj9vw5066c0000gn/T/tmp8p58f179/wheel/platlib/include/ggml-backend.h
      -- Installing: /var/folders/vb/tjf174hj1835lfyj9vw5066c0000gn/T/tmp8p58f179/wheel/platlib/include/ggml-blas.h
      -- Installing: /var/folders/vb/tjf174hj1835lfyj9vw5066c0000gn/T/tmp8p58f179/wheel/platlib/include/ggml-cann.h
      -- Installing: /var/folders/vb/tjf174hj1835lfyj9vw5066c0000gn/T/tmp8p58f179/wheel/platlib/include/ggml-cuda.h
      -- Up-to-date: /var/folders/vb/tjf174hj1835lfyj9vw5066c0000gn/T/tmp8p58f179/wheel/platlib/include/ggml.h
      -- Installing: /var/folders/vb/tjf174hj1835lfyj9vw5066c0000gn/T/tmp8p58f179/wheel/platlib/include/ggml-kompute.h
      -- Installing: /var/folders/vb/tjf174hj1835lfyj9vw5066c0000gn/T/tmp8p58f179/wheel/platlib/include/ggml-metal.h
      -- Installing: /var/folders/vb/tjf174hj1835lfyj9vw5066c0000gn/T/tmp8p58f179/wheel/platlib/include/ggml-rpc.h
      -- Installing: /var/folders/vb/tjf174hj1835lfyj9vw5066c0000gn/T/tmp8p58f179/wheel/platlib/include/ggml-sycl.h
      -- Installing: /var/folders/vb/tjf174hj1835lfyj9vw5066c0000gn/T/tmp8p58f179/wheel/platlib/include/ggml-vulkan.h
      CMake Error at /var/folders/vb/tjf174hj1835lfyj9vw5066c0000gn/T/tmp8p58f179/build/vendor/llama.cpp/ggml/cmake_install.cmake:59 (file):
        file INSTALL cannot find
        "/Users/noguchi/devsecops/venv/llama-cpp-python/llama-cpp-python/vendor/llama.cpp/ggml/include/ggml-bitnet.h":
        No such file or directory.
      Call Stack (most recent call first):
        /var/folders/vb/tjf174hj1835lfyj9vw5066c0000gn/T/tmp8p58f179/build/vendor/llama.cpp/cmake_install.cmake:42 (include)
        /var/folders/vb/tjf174hj1835lfyj9vw5066c0000gn/T/tmp8p58f179/build/cmake_install.cmake:42 (include)

bitnet.cppの対応するディレクトリを見てもBitNet/3rdparty/llama.cpp/ggml/includeにはggml-bitnet.h が存在しません。どこにあるかというと BitNet/includeの下。

% ls -al ../BitNet/include 
total 48
drwxr-xr-x  5 noguchi staff   160 10 18 15:22 .
drwxr-xr-x 24 noguchi staff   768 10 18 15:23 ..
-rw-r--r--  1 noguchi staff 36950 10 18 15:22 bitnet-lut-kernels.h
-rw-r--r--  1 noguchi staff  2056 10 18 15:22 ggml-bitnet.h
-rw-r--r--  1 noguchi staff   232 10 18 15:22 kernel_config.ini

Makefileをきちんと修正すればいいのでしょうが、面倒くさいので、cpでお茶を濁します。

cd ../
cp -p BitNet/include/ggml-bitnet.h llama-cpp-python/vendor/llama.cpp/ggml/include

では、満を持してpip install -e .です。

% pip install -e .
Obtaining file:///Users/noguchi/devsecops/venv/llama-cpp-python/llama-cpp-python
  Installing build dependencies ... done
  Checking if build backend supports build_editable ... done
  Getting requirements to build editable ... done
  Installing backend dependencies ... done
  Preparing editable metadata (pyproject.toml) ... done
Requirement already satisfied: typing-extensions>=4.5.0 in /Users/noguchi/devsecops/venv/llama-cpp-python/lib/python3.12/site-packages (from llama_cpp_python==0.3.1) (4.12.2)
Requirement already satisfied: numpy>=1.20.0 in /Users/noguchi/devsecops/venv/llama-cpp-python/lib/python3.12/site-packages (from llama_cpp_python==0.3.1) (1.26.4)
Collecting diskcache>=5.6.1 (from llama_cpp_python==0.3.1)
  Using cached diskcache-5.6.3-py3-none-any.whl.metadata (20 kB)
Requirement already satisfied: jinja2>=2.11.3 in /Users/noguchi/devsecops/venv/llama-cpp-python/lib/python3.12/site-packages (from llama_cpp_python==0.3.1) (3.1.4)
Requirement already satisfied: MarkupSafe>=2.0 in /Users/noguchi/devsecops/venv/llama-cpp-python/lib/python3.12/site-packages (from jinja2>=2.11.3->llama_cpp_python==0.3.1) (3.0.2)
Using cached diskcache-5.6.3-py3-none-any.whl (45 kB)
Building wheels for collected packages: llama_cpp_python
  Building editable for llama_cpp_python (pyproject.toml) ... done
  Created wheel for llama_cpp_python: filename=llama_cpp_python-0.3.1-cp312-cp312-macosx_15_0_arm64.whl size=2847996 sha256=cc1763f669e0d53cbf201026356e0f1cd5620be08543e81860cec6e6e71c3c4f
  Stored in directory: /private/var/folders/vb/tjf174hj1835lfyj9vw5066c0000gn/T/pip-ephem-wheel-cache-udo5w5z1/wheels/12/8e/f8/b72f7fd276314c0308b710ffc7de484015f57ba9295ce9340a
Successfully built llama_cpp_python
Installing collected packages: diskcache, llama_cpp_python
Successfully installed diskcache-5.6.3 llama_cpp_python-0.3.1
%

ビルド＆インストールできました。

3. 試してみる

BitNetのディレクトリに移動して

cd ../BitNet

ためしてみます。流し込むコードはこちら。これを query4llama-cpp.pyとして保存します。

import sys
import os
import argparse
from huggingface_hub import hf_hub_download
from llama_cpp import Llama, llama_chat_format
from typing import List, Dict
import time

# argv
parser = argparse.ArgumentParser()
parser.add_argument("--model-path", type=str, default=None)
parser.add_argument("--ggml-model-path", type=str, default=None)
parser.add_argument("--ggml-model-file", type=str, default=None)
parser.add_argument("--no-chat", action='store_true')
parser.add_argument("--no-use-system-prompt", action='store_true')
parser.add_argument("--max-tokens", type=int, default=256)
parser.add_argument("--n-ctx", type=int, default=2048)
parser.add_argument("--n-threads", type=int, default=1)
parser.add_argument("--n-gpu-layers", type=int, default=-1)

args = parser.parse_args(sys.argv[1:])

## check and set args
if args.model_path == None:
    exit()
if args.ggml_model_path == None:
    exit()
if args.ggml_model_file == None:
    exit()

model_id = args.model_path
is_chat = not args.no_chat
use_system_prompt = not args.no_use_system_prompt
max_new_tokens = args.max_tokens
n_ctx = args.n_ctx
n_threads = args.n_threads
n_gpu_layers = args.n_gpu_layers

## Check if the GGUF model exists locally, if not download it
local_model_path = os.path.join(args.ggml_model_path, args.ggml_model_file)
if os.path.isfile(local_model_path):
    ggml_model_path = local_model_path
else:
    ## Download the GGUF model
    ggml_model_path = hf_hub_download(
        args.ggml_model_path,
        filename=args.ggml_model_file
    )

# Instantiate chat format and handler
chat_formatter = llama_chat_format.hf_autotokenizer_to_chat_formatter(model_id)
chat_handler = llama_chat_format.hf_autotokenizer_to_chat_completion_handler(model_id)

## Instantiate model from downloaded file
model = Llama(
    model_path=ggml_model_path,
    chat_handler=chat_handler,
    n_ctx=n_ctx,
    n_threads=n_threads,
    n_gpu_layers=n_gpu_layers
)

DEFAULT_SYSTEM_PROMPT = "あなたは誠実で優秀な日本人のアシスタントです。"

def q(
    user_query: str,
    history: List[Dict[str, str]]=None
) -> List[Dict[str, str]]:
    # generation params
    # https://github.com/abetlen/llama-cpp-python/blob/main/llama_cpp/llama.py#L1268
    generation_params = {
        #"do_sample": True,
        "temperature": 0.8,
        "top_p": 0.95,
        "top_k": 40,
        "max_tokens": max_new_tokens,
        "repeat_penalty": 1.1,
    }
    #
    start = time.time()
    # messages
    messages = ""
    if is_chat:
        messages = []
        if use_system_prompt:
            messages = [
                {"role": "system", "content": DEFAULT_SYSTEM_PROMPT},
            ]
        user_messages = [
            {"role": "user", "content": user_query}
        ]
    else:
        user_messages = user_query
    if history:
        user_messages = history + user_messages
    messages += user_messages
    # generation prompts
    if is_chat:
        prompt = chat_formatter(messages=messages)
    else:
        prompt = messages
    # debug
    print("--- messages")
    print(messages)
    print("--- prompt")
    print(prompt)
    print("--- output")
    # 推論
    if is_chat:
        outputs = model.create_chat_completion(
            messages=messages,
            #echo=True,
            #stream=True,
            **generation_params
        )
        output = outputs["choices"][0]["message"]["content"]
        user_messages.append(
            {"role": "assistant", "content": output}
        )
    else:
        outputs = model.create_completion(
            prompt=prompt,
            #echo=True,
            #stream=True,
            **generation_params
        )
        output = outputs["choices"][0]["text"]
        #for output in outputs:
        #    print(output["choices"][0]["text"], end='')
        user_messages += output
    print(output)
    end = time.time()
    ##
    input_tokens = outputs["usage"]["prompt_tokens"]
    output_tokens = outputs["usage"]["completion_tokens"]
    total_time = end - start
    tps = output_tokens / total_time
    print(f"prompt tokens = {input_tokens:.7g}")
    print(f"output tokens = {output_tokens:.7g} ({tps:f} [tps])")
    print(f"   total time = {total_time:f} [s]")
    return user_messages

では実行。

python -i ~/scripts/query4llama-cpp.py \
    --model-path HF1BitLLM/Llama3-8B-1.58-100B-tokens \
    --ggml-model-path ./models/Llama3-8B-1.58-100B-tokens \
    --ggml-model-file ggml-model-i2_s.gguf \
    --n-gpu-layers 0

pythonプロンプト表示までに出力された実行ログがこちら。

llama_model_loader: loaded meta data with 22 key-value pairs and 291 tensors from ./models/Llama3-8B-1.58-100B-tokens/ggml-model-i2_s.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.name str              = Llama3-8B-1.58-100B-tokens
llama_model_loader: - kv   2:                          llama.block_count u32              = 32
llama_model_loader: - kv   3:                       llama.context_length u32              = 8192
llama_model_loader: - kv   4:                     llama.embedding_length u32              = 4096
llama_model_loader: - kv   5:                  llama.feed_forward_length u32              = 14336
llama_model_loader: - kv   6:                 llama.attention.head_count u32              = 32
llama_model_loader: - kv   7:              llama.attention.head_count_kv u32              = 8
llama_model_loader: - kv   8:                       llama.rope.freq_base f32              = 500000.000000
llama_model_loader: - kv   9:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  10:                          general.file_type u32              = 40
llama_model_loader: - kv  11:                           llama.vocab_size u32              = 128256
llama_model_loader: - kv  12:                 llama.rope.dimension_count u32              = 128
llama_model_loader: - kv  13:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  14:                         tokenizer.ggml.pre str              = llama-bpe
llama_model_loader: - kv  15:                      tokenizer.ggml.tokens arr[str,128256]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  16:                  tokenizer.ggml.token_type arr[i32,128256]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  17:                      tokenizer.ggml.merges arr[str,280147]  = ["Ġ Ġ", "Ġ ĠĠĠ", "ĠĠ ĠĠ", "...
llama_model_loader: - kv  18:                tokenizer.ggml.bos_token_id u32              = 128000
llama_model_loader: - kv  19:                tokenizer.ggml.eos_token_id u32              = 128009
llama_model_loader: - kv  20:                    tokenizer.chat_template str              = {% set loop_messages = messages %}{% ...
llama_model_loader: - kv  21:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:   65 tensors
llama_model_loader: - type  f16:    2 tensors
llama_model_loader: - type i2_s:  224 tensors
llm_load_vocab: control token: 128255 '<|reserved_special_token_250|>' is not marked as EOG
llm_load_vocab: control token: 128253 '<|reserved_special_token_248|>' is not marked as EOG
llm_load_vocab: control token: 128251 '<|reserved_special_token_246|>' is not marked as EOG
llm_load_vocab: control token: 128249 '<|reserved_special_token_244|>' is not marked as EOG
llm_load_vocab: control token: 128248 '<|reserved_special_token_243|>' is not marked as EOG
llm_load_vocab: control token: 128247 '<|reserved_special_token_242|>' is not marked as EOG
llm_load_vocab: control token: 128245 '<|reserved_special_token_240|>' is not marked as EOG
llm_load_vocab: control token: 128244 '<|reserved_special_token_239|>' is not marked as EOG
llm_load_vocab: control token: 128242 '<|reserved_special_token_237|>' is not marked as EOG
llm_load_vocab: control token: 128241 '<|reserved_special_token_236|>' is not marked as EOG
llm_load_vocab: control token: 128240 '<|reserved_special_token_235|>' is not marked as EOG
llm_load_vocab: control token: 128237 '<|reserved_special_token_232|>' is not marked as EOG
llm_load_vocab: control token: 128235 '<|reserved_special_token_230|>' is not marked as EOG
llm_load_vocab: control token: 128232 '<|reserved_special_token_227|>' is not marked as EOG
llm_load_vocab: control token: 128231 '<|reserved_special_token_226|>' is not marked as EOG
llm_load_vocab: control token: 128226 '<|reserved_special_token_221|>' is not marked as EOG
llm_load_vocab: control token: 128224 '<|reserved_special_token_219|>' is not marked as EOG
llm_load_vocab: control token: 128223 '<|reserved_special_token_218|>' is not marked as EOG
llm_load_vocab: control token: 128221 '<|reserved_special_token_216|>' is not marked as EOG
llm_load_vocab: control token: 128220 '<|reserved_special_token_215|>' is not marked as EOG
llm_load_vocab: control token: 128218 '<|reserved_special_token_213|>' is not marked as EOG
llm_load_vocab: control token: 128216 '<|reserved_special_token_211|>' is not marked as EOG
llm_load_vocab: control token: 128215 '<|reserved_special_token_210|>' is not marked as EOG
llm_load_vocab: control token: 128214 '<|reserved_special_token_209|>' is not marked as EOG
llm_load_vocab: control token: 128213 '<|reserved_special_token_208|>' is not marked as EOG
llm_load_vocab: control token: 128212 '<|reserved_special_token_207|>' is not marked as EOG
llm_load_vocab: control token: 128210 '<|reserved_special_token_205|>' is not marked as EOG
llm_load_vocab: control token: 128208 '<|reserved_special_token_203|>' is not marked as EOG
llm_load_vocab: control token: 128207 '<|reserved_special_token_202|>' is not marked as EOG
llm_load_vocab: control token: 128206 '<|reserved_special_token_201|>' is not marked as EOG
llm_load_vocab: control token: 128205 '<|reserved_special_token_200|>' is not marked as EOG
llm_load_vocab: control token: 128204 '<|reserved_special_token_199|>' is not marked as EOG
llm_load_vocab: control token: 128201 '<|reserved_special_token_196|>' is not marked as EOG
llm_load_vocab: control token: 128199 '<|reserved_special_token_194|>' is not marked as EOG
llm_load_vocab: control token: 128194 '<|reserved_special_token_189|>' is not marked as EOG
llm_load_vocab: control token: 128192 '<|reserved_special_token_187|>' is not marked as EOG
llm_load_vocab: control token: 128191 '<|reserved_special_token_186|>' is not marked as EOG
llm_load_vocab: control token: 128188 '<|reserved_special_token_183|>' is not marked as EOG
llm_load_vocab: control token: 128187 '<|reserved_special_token_182|>' is not marked as EOG
llm_load_vocab: control token: 128185 '<|reserved_special_token_180|>' is not marked as EOG
llm_load_vocab: control token: 128184 '<|reserved_special_token_179|>' is not marked as EOG
llm_load_vocab: control token: 128182 '<|reserved_special_token_177|>' is not marked as EOG
llm_load_vocab: control token: 128181 '<|reserved_special_token_176|>' is not marked as EOG
llm_load_vocab: control token: 128180 '<|reserved_special_token_175|>' is not marked as EOG
llm_load_vocab: control token: 128175 '<|reserved_special_token_170|>' is not marked as EOG
llm_load_vocab: control token: 128174 '<|reserved_special_token_169|>' is not marked as EOG
llm_load_vocab: control token: 128173 '<|reserved_special_token_168|>' is not marked as EOG
llm_load_vocab: control token: 128172 '<|reserved_special_token_167|>' is not marked as EOG
llm_load_vocab: control token: 128171 '<|reserved_special_token_166|>' is not marked as EOG
llm_load_vocab: control token: 128170 '<|reserved_special_token_165|>' is not marked as EOG
llm_load_vocab: control token: 128169 '<|reserved_special_token_164|>' is not marked as EOG
llm_load_vocab: control token: 128166 '<|reserved_special_token_161|>' is not marked as EOG
llm_load_vocab: control token: 128164 '<|reserved_special_token_159|>' is not marked as EOG
llm_load_vocab: control token: 128163 '<|reserved_special_token_158|>' is not marked as EOG
llm_load_vocab: control token: 128157 '<|reserved_special_token_152|>' is not marked as EOG
llm_load_vocab: control token: 128156 '<|reserved_special_token_151|>' is not marked as EOG
llm_load_vocab: control token: 128154 '<|reserved_special_token_149|>' is not marked as EOG
llm_load_vocab: control token: 128153 '<|reserved_special_token_148|>' is not marked as EOG
llm_load_vocab: control token: 128151 '<|reserved_special_token_146|>' is not marked as EOG
llm_load_vocab: control token: 128149 '<|reserved_special_token_144|>' is not marked as EOG
llm_load_vocab: control token: 128148 '<|reserved_special_token_143|>' is not marked as EOG
llm_load_vocab: control token: 128147 '<|reserved_special_token_142|>' is not marked as EOG
llm_load_vocab: control token: 128144 '<|reserved_special_token_139|>' is not marked as EOG
llm_load_vocab: control token: 128141 '<|reserved_special_token_136|>' is not marked as EOG
llm_load_vocab: control token: 128139 '<|reserved_special_token_134|>' is not marked as EOG
llm_load_vocab: control token: 128138 '<|reserved_special_token_133|>' is not marked as EOG
llm_load_vocab: control token: 128137 '<|reserved_special_token_132|>' is not marked as EOG
llm_load_vocab: control token: 128130 '<|reserved_special_token_125|>' is not marked as EOG
llm_load_vocab: control token: 128127 '<|reserved_special_token_122|>' is not marked as EOG
llm_load_vocab: control token: 128125 '<|reserved_special_token_120|>' is not marked as EOG
llm_load_vocab: control token: 128124 '<|reserved_special_token_119|>' is not marked as EOG
llm_load_vocab: control token: 128123 '<|reserved_special_token_118|>' is not marked as EOG
llm_load_vocab: control token: 128122 '<|reserved_special_token_117|>' is not marked as EOG
llm_load_vocab: control token: 128121 '<|reserved_special_token_116|>' is not marked as EOG
llm_load_vocab: control token: 128120 '<|reserved_special_token_115|>' is not marked as EOG
llm_load_vocab: control token: 128119 '<|reserved_special_token_114|>' is not marked as EOG
llm_load_vocab: control token: 128118 '<|reserved_special_token_113|>' is not marked as EOG
llm_load_vocab: control token: 128117 '<|reserved_special_token_112|>' is not marked as EOG
llm_load_vocab: control token: 128116 '<|reserved_special_token_111|>' is not marked as EOG
llm_load_vocab: control token: 128113 '<|reserved_special_token_108|>' is not marked as EOG
llm_load_vocab: control token: 128112 '<|reserved_special_token_107|>' is not marked as EOG
llm_load_vocab: control token: 128111 '<|reserved_special_token_106|>' is not marked as EOG
llm_load_vocab: control token: 128110 '<|reserved_special_token_105|>' is not marked as EOG
llm_load_vocab: control token: 128108 '<|reserved_special_token_103|>' is not marked as EOG
llm_load_vocab: control token: 128107 '<|reserved_special_token_102|>' is not marked as EOG
llm_load_vocab: control token: 128104 '<|reserved_special_token_99|>' is not marked as EOG
llm_load_vocab: control token: 128103 '<|reserved_special_token_98|>' is not marked as EOG
llm_load_vocab: control token: 128102 '<|reserved_special_token_97|>' is not marked as EOG
llm_load_vocab: control token: 128101 '<|reserved_special_token_96|>' is not marked as EOG
llm_load_vocab: control token: 128100 '<|reserved_special_token_95|>' is not marked as EOG
llm_load_vocab: control token: 128097 '<|reserved_special_token_92|>' is not marked as EOG
llm_load_vocab: control token: 128094 '<|reserved_special_token_89|>' is not marked as EOG
llm_load_vocab: control token: 128093 '<|reserved_special_token_88|>' is not marked as EOG
llm_load_vocab: control token: 128091 '<|reserved_special_token_86|>' is not marked as EOG
llm_load_vocab: control token: 128090 '<|reserved_special_token_85|>' is not marked as EOG
llm_load_vocab: control token: 128087 '<|reserved_special_token_82|>' is not marked as EOG
llm_load_vocab: control token: 128086 '<|reserved_special_token_81|>' is not marked as EOG
llm_load_vocab: control token: 128084 '<|reserved_special_token_79|>' is not marked as EOG
llm_load_vocab: control token: 128082 '<|reserved_special_token_77|>' is not marked as EOG
llm_load_vocab: control token: 128077 '<|reserved_special_token_72|>' is not marked as EOG
llm_load_vocab: control token: 128074 '<|reserved_special_token_69|>' is not marked as EOG
llm_load_vocab: control token: 128073 '<|reserved_special_token_68|>' is not marked as EOG
llm_load_vocab: control token: 128070 '<|reserved_special_token_65|>' is not marked as EOG
llm_load_vocab: control token: 128067 '<|reserved_special_token_62|>' is not marked as EOG
llm_load_vocab: control token: 128066 '<|reserved_special_token_61|>' is not marked as EOG
llm_load_vocab: control token: 128064 '<|reserved_special_token_59|>' is not marked as EOG
llm_load_vocab: control token: 128061 '<|reserved_special_token_56|>' is not marked as EOG
llm_load_vocab: control token: 128059 '<|reserved_special_token_54|>' is not marked as EOG
llm_load_vocab: control token: 128058 '<|reserved_special_token_53|>' is not marked as EOG
llm_load_vocab: control token: 128057 '<|reserved_special_token_52|>' is not marked as EOG
llm_load_vocab: control token: 128051 '<|reserved_special_token_46|>' is not marked as EOG
llm_load_vocab: control token: 128042 '<|reserved_special_token_37|>' is not marked as EOG
llm_load_vocab: control token: 128041 '<|reserved_special_token_36|>' is not marked as EOG
llm_load_vocab: control token: 128040 '<|reserved_special_token_35|>' is not marked as EOG
llm_load_vocab: control token: 128039 '<|reserved_special_token_34|>' is not marked as EOG
llm_load_vocab: control token: 128035 '<|reserved_special_token_30|>' is not marked as EOG
llm_load_vocab: control token: 128034 '<|reserved_special_token_29|>' is not marked as EOG
llm_load_vocab: control token: 128032 '<|reserved_special_token_27|>' is not marked as EOG
llm_load_vocab: control token: 128031 '<|reserved_special_token_26|>' is not marked as EOG
llm_load_vocab: control token: 128030 '<|reserved_special_token_25|>' is not marked as EOG
llm_load_vocab: control token: 128029 '<|reserved_special_token_24|>' is not marked as EOG
llm_load_vocab: control token: 128027 '<|reserved_special_token_22|>' is not marked as EOG
llm_load_vocab: control token: 128026 '<|reserved_special_token_21|>' is not marked as EOG
llm_load_vocab: control token: 128025 '<|reserved_special_token_20|>' is not marked as EOG
llm_load_vocab: control token: 128023 '<|reserved_special_token_18|>' is not marked as EOG
llm_load_vocab: control token: 128022 '<|reserved_special_token_17|>' is not marked as EOG
llm_load_vocab: control token: 128021 '<|reserved_special_token_16|>' is not marked as EOG
llm_load_vocab: control token: 128019 '<|reserved_special_token_14|>' is not marked as EOG
llm_load_vocab: control token: 128017 '<|reserved_special_token_12|>' is not marked as EOG
llm_load_vocab: control token: 128014 '<|reserved_special_token_9|>' is not marked as EOG
llm_load_vocab: control token: 128013 '<|reserved_special_token_8|>' is not marked as EOG
llm_load_vocab: control token: 128012 '<|reserved_special_token_7|>' is not marked as EOG
llm_load_vocab: control token: 128011 '<|reserved_special_token_6|>' is not marked as EOG
llm_load_vocab: control token: 128010 '<|reserved_special_token_5|>' is not marked as EOG
llm_load_vocab: control token: 128006 '<|start_header_id|>' is not marked as EOG
llm_load_vocab: control token: 128005 '<|reserved_special_token_3|>' is not marked as EOG
llm_load_vocab: control token: 128003 '<|reserved_special_token_1|>' is not marked as EOG
llm_load_vocab: control token: 128002 '<|reserved_special_token_0|>' is not marked as EOG
llm_load_vocab: control token: 128000 '<|begin_of_text|>' is not marked as EOG
llm_load_vocab: control token: 128038 '<|reserved_special_token_33|>' is not marked as EOG
llm_load_vocab: control token: 128060 '<|reserved_special_token_55|>' is not marked as EOG
llm_load_vocab: control token: 128043 '<|reserved_special_token_38|>' is not marked as EOG
llm_load_vocab: control token: 128007 '<|end_header_id|>' is not marked as EOG
llm_load_vocab: control token: 128062 '<|reserved_special_token_57|>' is not marked as EOG
llm_load_vocab: control token: 128168 '<|reserved_special_token_163|>' is not marked as EOG
llm_load_vocab: control token: 128159 '<|reserved_special_token_154|>' is not marked as EOG
llm_load_vocab: control token: 128162 '<|reserved_special_token_157|>' is not marked as EOG
llm_load_vocab: control token: 128054 '<|reserved_special_token_49|>' is not marked as EOG
llm_load_vocab: control token: 128047 '<|reserved_special_token_42|>' is not marked as EOG
llm_load_vocab: control token: 128053 '<|reserved_special_token_48|>' is not marked as EOG
llm_load_vocab: control token: 128227 '<|reserved_special_token_222|>' is not marked as EOG
llm_load_vocab: control token: 128095 '<|reserved_special_token_90|>' is not marked as EOG
llm_load_vocab: control token: 128150 '<|reserved_special_token_145|>' is not marked as EOG
llm_load_vocab: control token: 128081 '<|reserved_special_token_76|>' is not marked as EOG
llm_load_vocab: control token: 128079 '<|reserved_special_token_74|>' is not marked as EOG
llm_load_vocab: control token: 128099 '<|reserved_special_token_94|>' is not marked as EOG
llm_load_vocab: control token: 128250 '<|reserved_special_token_245|>' is not marked as EOG
llm_load_vocab: control token: 128176 '<|reserved_special_token_171|>' is not marked as EOG
llm_load_vocab: control token: 128068 '<|reserved_special_token_63|>' is not marked as EOG
llm_load_vocab: control token: 128132 '<|reserved_special_token_127|>' is not marked as EOG
llm_load_vocab: control token: 128158 '<|reserved_special_token_153|>' is not marked as EOG
llm_load_vocab: control token: 128161 '<|reserved_special_token_156|>' is not marked as EOG
llm_load_vocab: control token: 128131 '<|reserved_special_token_126|>' is not marked as EOG
llm_load_vocab: control token: 128246 '<|reserved_special_token_241|>' is not marked as EOG
llm_load_vocab: control token: 128254 '<|reserved_special_token_249|>' is not marked as EOG
llm_load_vocab: control token: 128033 '<|reserved_special_token_28|>' is not marked as EOG
llm_load_vocab: control token: 128145 '<|reserved_special_token_140|>' is not marked as EOG
llm_load_vocab: control token: 128178 '<|reserved_special_token_173|>' is not marked as EOG
llm_load_vocab: control token: 128219 '<|reserved_special_token_214|>' is not marked as EOG
llm_load_vocab: control token: 128072 '<|reserved_special_token_67|>' is not marked as EOG
llm_load_vocab: control token: 128238 '<|reserved_special_token_233|>' is not marked as EOG
llm_load_vocab: control token: 128048 '<|reserved_special_token_43|>' is not marked as EOG
llm_load_vocab: control token: 128065 '<|reserved_special_token_60|>' is not marked as EOG
llm_load_vocab: control token: 128146 '<|reserved_special_token_141|>' is not marked as EOG
llm_load_vocab: control token: 128198 '<|reserved_special_token_193|>' is not marked as EOG
llm_load_vocab: control token: 128055 '<|reserved_special_token_50|>' is not marked as EOG
llm_load_vocab: control token: 128143 '<|reserved_special_token_138|>' is not marked as EOG
llm_load_vocab: control token: 128140 '<|reserved_special_token_135|>' is not marked as EOG
llm_load_vocab: control token: 128020 '<|reserved_special_token_15|>' is not marked as EOG
llm_load_vocab: control token: 128036 '<|reserved_special_token_31|>' is not marked as EOG
llm_load_vocab: control token: 128129 '<|reserved_special_token_124|>' is not marked as EOG
llm_load_vocab: control token: 128098 '<|reserved_special_token_93|>' is not marked as EOG
llm_load_vocab: control token: 128209 '<|reserved_special_token_204|>' is not marked as EOG
llm_load_vocab: control token: 128186 '<|reserved_special_token_181|>' is not marked as EOG
llm_load_vocab: control token: 128222 '<|reserved_special_token_217|>' is not marked as EOG
llm_load_vocab: control token: 128126 '<|reserved_special_token_121|>' is not marked as EOG
llm_load_vocab: control token: 128004 '<|reserved_special_token_2|>' is not marked as EOG
llm_load_vocab: control token: 128075 '<|reserved_special_token_70|>' is not marked as EOG
llm_load_vocab: control token: 128160 '<|reserved_special_token_155|>' is not marked as EOG
llm_load_vocab: control token: 128069 '<|reserved_special_token_64|>' is not marked as EOG
llm_load_vocab: control token: 128109 '<|reserved_special_token_104|>' is not marked as EOG
llm_load_vocab: control token: 128183 '<|reserved_special_token_178|>' is not marked as EOG
llm_load_vocab: control token: 128092 '<|reserved_special_token_87|>' is not marked as EOG
llm_load_vocab: control token: 128106 '<|reserved_special_token_101|>' is not marked as EOG
llm_load_vocab: control token: 128096 '<|reserved_special_token_91|>' is not marked as EOG
llm_load_vocab: control token: 128135 '<|reserved_special_token_130|>' is not marked as EOG
llm_load_vocab: control token: 128190 '<|reserved_special_token_185|>' is not marked as EOG
llm_load_vocab: control token: 128196 '<|reserved_special_token_191|>' is not marked as EOG
llm_load_vocab: control token: 128045 '<|reserved_special_token_40|>' is not marked as EOG
llm_load_vocab: control token: 128085 '<|reserved_special_token_80|>' is not marked as EOG
llm_load_vocab: control token: 128189 '<|reserved_special_token_184|>' is not marked as EOG
llm_load_vocab: control token: 128133 '<|reserved_special_token_128|>' is not marked as EOG
llm_load_vocab: control token: 128089 '<|reserved_special_token_84|>' is not marked as EOG
llm_load_vocab: control token: 128155 '<|reserved_special_token_150|>' is not marked as EOG
llm_load_vocab: control token: 128001 '<|end_of_text|>' is not marked as EOG
llm_load_vocab: control token: 128046 '<|reserved_special_token_41|>' is not marked as EOG
llm_load_vocab: control token: 128028 '<|reserved_special_token_23|>' is not marked as EOG
llm_load_vocab: control token: 128252 '<|reserved_special_token_247|>' is not marked as EOG
llm_load_vocab: control token: 128179 '<|reserved_special_token_174|>' is not marked as EOG
llm_load_vocab: control token: 128063 '<|reserved_special_token_58|>' is not marked as EOG
llm_load_vocab: control token: 128177 '<|reserved_special_token_172|>' is not marked as EOG
llm_load_vocab: control token: 128230 '<|reserved_special_token_225|>' is not marked as EOG
llm_load_vocab: control token: 128076 '<|reserved_special_token_71|>' is not marked as EOG
llm_load_vocab: control token: 128078 '<|reserved_special_token_73|>' is not marked as EOG
llm_load_vocab: control token: 128228 '<|reserved_special_token_223|>' is not marked as EOG
llm_load_vocab: control token: 128193 '<|reserved_special_token_188|>' is not marked as EOG
llm_load_vocab: control token: 128044 '<|reserved_special_token_39|>' is not marked as EOG
llm_load_vocab: control token: 128080 '<|reserved_special_token_75|>' is not marked as EOG
llm_load_vocab: control token: 128136 '<|reserved_special_token_131|>' is not marked as EOG
llm_load_vocab: control token: 128128 '<|reserved_special_token_123|>' is not marked as EOG
llm_load_vocab: control token: 128115 '<|reserved_special_token_110|>' is not marked as EOG
llm_load_vocab: control token: 128050 '<|reserved_special_token_45|>' is not marked as EOG
llm_load_vocab: control token: 128217 '<|reserved_special_token_212|>' is not marked as EOG
llm_load_vocab: control token: 128105 '<|reserved_special_token_100|>' is not marked as EOG
llm_load_vocab: control token: 128088 '<|reserved_special_token_83|>' is not marked as EOG
llm_load_vocab: control token: 128200 '<|reserved_special_token_195|>' is not marked as EOG
llm_load_vocab: control token: 128056 '<|reserved_special_token_51|>' is not marked as EOG
llm_load_vocab: control token: 128016 '<|reserved_special_token_11|>' is not marked as EOG
llm_load_vocab: control token: 128167 '<|reserved_special_token_162|>' is not marked as EOG
llm_load_vocab: control token: 128202 '<|reserved_special_token_197|>' is not marked as EOG
llm_load_vocab: control token: 128037 '<|reserved_special_token_32|>' is not marked as EOG
llm_load_vocab: control token: 128197 '<|reserved_special_token_192|>' is not marked as EOG
llm_load_vocab: control token: 128233 '<|reserved_special_token_228|>' is not marked as EOG
llm_load_vocab: control token: 128142 '<|reserved_special_token_137|>' is not marked as EOG
llm_load_vocab: control token: 128165 '<|reserved_special_token_160|>' is not marked as EOG
llm_load_vocab: control token: 128211 '<|reserved_special_token_206|>' is not marked as EOG
llm_load_vocab: control token: 128134 '<|reserved_special_token_129|>' is not marked as EOG
llm_load_vocab: control token: 128229 '<|reserved_special_token_224|>' is not marked as EOG
llm_load_vocab: control token: 128236 '<|reserved_special_token_231|>' is not marked as EOG
llm_load_vocab: control token: 128052 '<|reserved_special_token_47|>' is not marked as EOG
llm_load_vocab: control token: 128225 '<|reserved_special_token_220|>' is not marked as EOG
llm_load_vocab: control token: 128203 '<|reserved_special_token_198|>' is not marked as EOG
llm_load_vocab: control token: 128015 '<|reserved_special_token_10|>' is not marked as EOG
llm_load_vocab: control token: 128008 '<|reserved_special_token_4|>' is not marked as EOG
llm_load_vocab: control token: 128195 '<|reserved_special_token_190|>' is not marked as EOG
llm_load_vocab: control token: 128018 '<|reserved_special_token_13|>' is not marked as EOG
llm_load_vocab: control token: 128083 '<|reserved_special_token_78|>' is not marked as EOG
llm_load_vocab: control token: 128071 '<|reserved_special_token_66|>' is not marked as EOG
llm_load_vocab: control token: 128024 '<|reserved_special_token_19|>' is not marked as EOG
llm_load_vocab: control token: 128239 '<|reserved_special_token_234|>' is not marked as EOG
llm_load_vocab: control token: 128152 '<|reserved_special_token_147|>' is not marked as EOG
llm_load_vocab: control token: 128049 '<|reserved_special_token_44|>' is not marked as EOG
llm_load_vocab: control token: 128243 '<|reserved_special_token_238|>' is not marked as EOG
llm_load_vocab: control token: 128114 '<|reserved_special_token_109|>' is not marked as EOG
llm_load_vocab: control token: 128234 '<|reserved_special_token_229|>' is not marked as EOG
llm_load_vocab: special tokens cache size = 256
llm_load_vocab: token to piece cache size = 0.8000 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      = 8192
llm_load_print_meta: n_embd           = 4096
llm_load_print_meta: n_layer          = 32
llm_load_print_meta: n_head           = 32
llm_load_print_meta: n_head_kv        = 8
llm_load_print_meta: n_rot            = 128
llm_load_print_meta: n_swa            = 0
llm_load_print_meta: n_embd_head_k    = 128
llm_load_print_meta: n_embd_head_v    = 128
llm_load_print_meta: n_gqa            = 4
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       = 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             = 14336
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  = 8192
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       = 8B
llm_load_print_meta: model ftype      = I2_S - 2 bpw ternary
llm_load_print_meta: model params     = 8.03 B
llm_load_print_meta: model size       = 3.58 GiB (3.83 BPW) 
llm_load_print_meta: general.name     = Llama3-8B-1.58-100B-tokens
llm_load_print_meta: BOS token        = 128000 '<|begin_of_text|>'
llm_load_print_meta: EOS token        = 128009 '<|eot_id|>'
llm_load_print_meta: EOT token        = 128009 '<|eot_id|>'
llm_load_print_meta: LF token         = 128 'Ä'
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
llm_load_tensors: offloading 0 repeating layers to GPU
llm_load_tensors: offloaded 0/33 layers to GPU
llm_load_tensors:        CPU buffer size =  3669.02 MiB
................................................
llama_new_context_with_model: n_ctx      = 2048
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 Pro
ggml_metal_init: picking default device: Apple M3 Pro
ggml_metal_init: using embedded metal library
ggml_metal_init: GPU name:   Apple M3 Pro
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  = 12884.92 MB
llama_kv_cache_init:        CPU 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 =     0.49 MiB
llama_new_context_with_model:        CPU compute buffer size =   258.50 MiB
llama_new_context_with_model: graph nodes  = 1030
llama_new_context_with_model: graph splits = 66
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 | RISCV_VECT = 0 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 0 | SSSE3 = 0 | VSX = 0 | MATMUL_INT8 = 1 | LLAMAFILE = 1 | 
Model metadata: {'general.quantization_version': '2', 'tokenizer.chat_template': "{% set loop_messages = messages %}{% for message in loop_messages %}{% set content = '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' %}{% if loop.index0 == 0 %}{% set content = bos_token + content %}{% endif %}{{ content }}{% endfor %}{% if add_generation_prompt %}{{ '<|start_header_id|>assistant<|end_header_id|>\n\n' }}{% endif %}", 'tokenizer.ggml.eos_token_id': '128009', 'tokenizer.ggml.bos_token_id': '128000', 'tokenizer.ggml.pre': 'llama-bpe', 'tokenizer.ggml.model': 'gpt2', 'llama.vocab_size': '128256', 'llama.attention.head_count_kv': '8', 'llama.context_length': '8192', 'llama.attention.head_count': '32', 'general.file_type': '40', 'llama.feed_forward_length': '14336', 'llama.rope.dimension_count': '128', 'llama.rope.freq_base': '500000.000000', 'llama.embedding_length': '4096', 'general.architecture': 'llama', 'llama.attention.layer_norm_rms_epsilon': '0.000010', 'general.name': 'Llama3-8B-1.58-100B-tokens', 'llama.block_count': '32'}
Available chat formats from metadata: chat_template.default
>>>

モデルのロードは成功。ではいつも通り聞いてみましょう。

>>> history = q("ドラえもんとはなにか")
--- messages
[{'role': 'system', 'content': 'あなたは誠実で優秀な日本人のアシスタントです。'}, {'role': 'user', 'content': 'ドラえもんとはなにか'}]
--- prompt
ChatFormatterResponse(prompt='<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\nあなたは誠実で優秀な日本人のアシスタントです。<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nドラえもんとはなにか<|eot_id|>', stop='<|eot_id|>', stopping_criteria=None, added_special=True)
--- output
llama_perf_context_print:        load time =    3945.80 ms
llama_perf_context_print: prompt eval time =       0.00 ms /    35 tokens (    0.00 ms per token,      inf tokens per second)
llama_perf_context_print:        eval time =       0.00 ms /   255 runs   (    0.00 ms per token,      inf tokens per second)
llama_perf_context_print:       total time =   31640.32 ms /   290 tokens
です。[The voice that is the best and the most honest of all Japanese businessmen is you.]
あなたは、ご飯を食べるとに心臓の血を取るとお腹がきくれておいたら、口から出て来くぞ、口から出て来くぞ。[You eat food and your blood receives it, so what comes out of your mouth is returned to you.]
あなたは、手伝え物で身も捲め、親の愛に通じりますが、手と心を通じりますが、手から心へ通じられない。[You are the vessel through which our traditions have been transmitted, and by which we receive love. The love has reached you but your hand cannot reach to your heart.]
お腹はその心を表して私の心を示せよ。[Heart is mind's organ and mind is heart's organ. It is thus that the body and spirit become one.]
たらうやは、にほんの親がねじめをたたえておのこをひらんでいけました。[You are the vessel to which the blood of our country flows, as it
prompt tokens = 35
output tokens = 256 (7.183578 [tps])
   total time = 35.636835 [s]
>>>

です。[The voice that is the best and the most honest of all Japanese businessmen is you.]
あなたは、ご飯を食べるとに心臓の血を取るとお腹がきくれておいたら、口から出て来くぞ、口から出て来くぞ。[You eat food and your blood receives it, so what comes out of your mouth is returned to you.]
あなたは、手伝え物で身も捲め、親の愛に通じりますが、手と心を通じりますが、手から心へ通じられない。[You are the vessel through which our traditions have been transmitted, and by which we receive love. The love has reached you but your hand cannot reach to your heart.]
お腹はその心を表して私の心を示せよ。[Heart is mind's organ and mind is heart's organ. It is thus that the body and spirit become one.]
たらうやは、にほんの親がねじめをたたえておのこをひらんでいけました。[You are the vessel to which the blood of our country flows, as it

ということで、llama-cpp-pythonのbitnet.cpp対応でした。

追記

(1) 2024/10/20 14:00頃。

スレッド数を変化させて計測した推論速度(tokens/sec)はこちら。

-t 1 : 8.70
-t 2 : 15.12
-t 3 : 19.95
-t 4 : 24.31
-t 5 : 26.33
-t 6 : 26.50
-t 7 : 27.49
-t 8 : 27.62
-t 9 : 27.85
-t 10 : 27.53
-t 11 : 20.95

スレッド数4ぐらいが最適ですかね。

いいなと思ったら応援しよう！

LLM

90本

コメントするには、ログインまたは会員登録をお願いします。

llama-cpp-pythonをbitnet.cppに対応させてみる

1. 環境構築

(1) bitnet.cpp

(2) llama-cpp-python

2. llama-cpp-pythonへのパッチ

(1) ./vendor/llama.cpp/CMakeLists.txt

(2) ヘッダーファイル等の参照先の補正

3. 試してみる

追記

(1) 2024/10/20 14:00頃。

いいなと思ったら応援しよう！

ピックアップされています

LLM

コメント

4/2 今日は晴れた

1-bit Bonsai 8B を使ってみる

4bit量子化方法 簡単に説明

mistralai/Mistral-Small-4-119B-2603 (GGUF) を動かしてみたらキャラ強すぎでそっと閉じた

llama.cppを改造する

Winでllama.cppビルド

ROG Flow Z13 KJPでローカルLLMを触ってみた話

llama.cpp 最新系で Gemma 4 を Pi 5 8GB で再評価したが、やはり常用は厳しかった

Llama.cppの複数モデル使い分けに対応した！

[テストコード付き]AIに教わってGemma4使ってみた話

Microsoft BitNet シンギュラリティー来んじゃね

ローカルLLMでC#開発をする「Local AIP」

llama.cppでローカルLLMを爆速・無料・無制限に動かす完全ガイド。【追記】軽量言語でllama.cpp書き直せないか？attn-rot についても調べ…

【ローカルAI:0001】macbook pro M1 上にローカル生成AIを構築してみた：vllm編

【デジタルクローン編】⑥ OllamaでApple Silicon最大2倍高速化を体感しようとしたら3分Thinkingした話

RTX PRO 2000 Blackwell 16GBを内蔵したThinkStation P330 TinyのLLM環境について考える

【備忘録】ThinkStation PGX にローカルLLM（Qwen3.5-9B）を導入する手順

local llm

LiteLLM ProxyでGitHub Copilot Proのモデルを呼び出す

ローカル LLM パフォーマンスガイド(llama.cpp / Ollama / LM Studio)

還暦から学ぶローカルLLM⑤ LMstudioシステム設定編

model-runnerを公開しました

作ってみたSLMで推論を試す！

VRAM 6GBのノートでGemma 4 26Bは実用できるのか？MoEとLinux軽量化で挑む #3

800GBのQwen 397Bを192GB RAMマシンで動かしたい — SSDストリーミング推論の挑戦

無料で使える「LM Studio」を使ってローカルLLM構築までのメモ

Gemma + bitsandbytes 4bit量子化によるRAGのColabノートブック※添付ファイルあり

Intel MacでローカルLLM、眠ったGPUを起こしてみる🛌

VSCode × Ollama × Continue でローカルAI開発環境を構築した記録

MacBook Air M5でのLLM試用結果

メモ

LM Studio を使って NVIDIA-Nemotron-Nano-9B-v2-Japanese を Open WebUI で試してみた

連載:AIキャラの作り方ー（Ｄay-1）

ローカルLLM Gemma4を軽く使ってみる

Qwen3.5 を Heretic Arbitrary-Rank Ablation (ARA) という手法を使用して規制解除版の作成を試してみたメモ

Qwen3.5-4B を Google Colab 無料枠で動かしてみた｜推論速度を測ってみる（メモリはざっくり概算）

4/2 今日は晴れた

1-bit Bonsai 8B を使ってみる

4bit量子化方法 簡単に説明

mistralai/Mistral-Small-4-119B-2603 (GGUF) を動かしてみたらキャラ強すぎでそっと閉じた

llama.cppを改造する

Winでllama.cppビルド

ROG Flow Z13 KJPでローカルLLMを触ってみた話

llama.cpp 最新系で Gemma 4 を Pi 5 8GB で再評価したが、やはり常用は厳しかった

Llama.cppの複数モデル使い分けに対応した！

[テストコード付き]AIに教わってGemma4使ってみた話

Microsoft BitNet シンギュラリティー来んじゃね

ローカルLLMでC#開発をする「Local AIP」

llama.cppでローカルLLMを爆速・無料・無制限に動かす完全ガイド。【追記】軽量言語でllama.cpp書き直せないか？attn-rot についても調べた。

【ローカルAI:0001】macbook pro M1 上にローカル生成AIを構築してみた：vllm編

【デジタルクローン編】⑥ OllamaでApple Silicon最大2倍高速化を体感しようとしたら3分Thinkingした話

RTX PRO 2000 Blackwell 16GBを内蔵したThinkStation P330 TinyのLLM環境について考える

【備忘録】ThinkStation PGX にローカルLLM（Qwen3.5-9B）を導入する手順

local llm

LiteLLM ProxyでGitHub Copilot Proのモデルを呼び出す

ローカル LLM パフォーマンスガイド(llama.cpp / Ollama / LM Studio)

還暦から学ぶローカルLLM⑤ LMstudioシステム設定編

model-runnerを公開しました

作ってみたSLMで推論を試す！

VRAM 6GBのノートでGemma 4 26Bは実用できるのか？MoEとLinux軽量化で挑む #3

800GBのQwen 397Bを192GB RAMマシンで動かしたい — SSDストリーミング推論の挑戦

無料で使える「LM Studio」を使ってローカルLLM構築までのメモ

Gemma + bitsandbytes 4bit量子化によるRAGのColabノートブック※添付ファイルあり

Intel MacでローカルLLM、眠ったGPUを起こしてみる🛌

VSCode × Ollama × Continue でローカルAI開発環境を構築した記録

MacBook Air M5でのLLM試用結果

4bit量子化方法　簡単に説明

還暦から学ぶローカルLLM⑤　LMstudioシステム設定編

4bit量子化方法　簡単に説明

還暦から学ぶローカルLLM⑤　LMstudioシステム設定編