返回模型

说明文档

启用键值缓存的 Llama 3.2 3B Instruct ONNX fp16 格式模型

模型创建者：Meta-Llama
原始模型：Meta-Llama Llama 3.2 3B Instruct

描述

本仓库包含由 Esperanto Technologies 完成的 Llama 3.2 3B Instruct ONNX 转换文件。该模型采用 fp16 格式，并启用了 KVC（键值缓存）功能。

如何下载 ONNX 模型和权重文件

获取模型最简单的方法是克隆整个仓库。或者，您也可以使用 huggingface-hub Python 库下载文件。

pip3 install huggingface-hub>=0.17.1

然后，您可以使用如下命令高速下载任意单个模型文件到当前目录：

huggingface-cli download Esperanto/llama-3.2-3B-Instruct-kvc-fp16-onnx --local-dir llama-3.2-3B-Instruct-kvc-fp16-onnx --local-dir-use-symlinks False

关于使用 huggingface-cli 下载的更多文档，请参阅：HF -> Hub Python Library -> Download files -> Download from the CLI。

如何使用 ONNXRuntime 从 Python 代码运行

该模型可以轻松在 CPU 上使用 ONNXRuntime 运行。

首先安装依赖包

pip3 install onnx==1.16.1
pip3 install onnxruntime==1.17.1

示例代码：使用此模型生成文本

我们使用贪婪解码定义循环：

import numpy as np
import onnxruntime
import onnx
from transformers import AutoTokenizer

def generate_text(model_path, prompt, tokenizer, max_gen_tokens, total_sequence, window, context):
    model = onnx.load(model_path)

    #我们为第一次迭代创建输入
    input_tensor = tokenizer(prompt, return_tensors=\"pt\")
    prompt_size = len(input_tensor['input_ids'][0])
    actual_input = input_tensor['input_ids']
    if prompt_size < window:
        actual_input = np.concatenate((tokenizer.bos_token_id*np.ones([1, window - prompt_size], dtype = 'int64'),
                                       actual_input), axis=1)
    if prompt_size + max_gen_tokens > total_sequence:
        print(\"ERROR: Longer total sequence is needed!\")
        return
    first_attention = np.concatenate((np.zeros([1, total_sequence - window], dtype = 'int64'),
                                      np.ones((1, window), dtype = 'int64')), axis=1)
    max_gen_tokens += prompt_size #我们需要在解析提示词的基础上进行生成
    inputs_names =[node.name for node in model.graph.input]
    output_names =[node.name for node in model.graph.output]
    n_heads = 8 #kvc的gqa头数
    inputs_dict = {}
    inputs_dict['input_ids'] = actual_input[:, :window].reshape(1, window).numpy()
    inputs_dict['attention_mask'] = first_attention
    index_pos = sum(first_attention[0])
    inputs_dict['position_ids'] = np.concatenate((np.zeros([1, total_sequence - index_pos], dtype = 'int64'), np.arange(index_pos, dtype = 'int64').reshape(1, index_pos)), axis=1)
    inputs_dict['tree_attention'] = np.triu(-65504*np.ones(total_sequence), k= 1).astype('float16').reshape(1, 1, total_sequence, total_sequence)
    for name in inputs_names:
        if name == 'input_ids' or name == 'attention_mask' or name == 'position_ids' or name == 'tree_attention': continue
        inputs_dict[name] = np.zeros([1, n_heads, context-window, 128], dtype=\"float16\")
    index = 0
    new_token = np.array([10])
    next_index = window
    old_j = 0
    total_input = actual_input.numpy()

    rt_session = onnxruntime.InferenceSession(model_path)
    ## 我们运行推理
    while next_index < max_gen_tokens:
        if new_token.any() == tokenizer.eos_token_id:
            break
        #推理
        output = rt_session.run(output_names, inputs_dict)
        outs_dictionary = {name: content for (name, content) in zip (output_names, output)}
        #我们为下一次推理准备输入
        for name in inputs_names:
            if name == 'input_ids':
                old_j = next_index
                if next_index < prompt_size:
                    if prompt_size - next_index >= window: next_index += window
                    else: next_index = prompt_size 
                    j = next_index - window
                else:
                    next_index +=1
                    j = next_index - window
                    new_token = outs_dictionary['logits'].argmax(-1).reshape(1, window)
                    total_input = np.concatenate((total_input, new_token[: , -1:]), axis = 1)
                inputs_dict['input_ids']= total_input[:, j:next_index].reshape(1, window)
            elif name == 'attention_mask':
                inputs_dict['attention_mask'] = np.concatenate((np.zeros((1, total_sequence-next_index), dtype = 'int64'), np.ones((1, next_index), dtype = 'int64')), axis=1)
            elif name == 'position_ids':
                inputs_dict['position_ids'] = np.concatenate((np.zeros([1, total_sequence - next_index], dtype = 'int64'), np.arange(next_index, dtype = 'int64').reshape(1, next_index)), axis=1)
            elif name == 'tree_attention': continue
            else:
                old_name = name.replace(\"past_key_values\", \"present\")
                inputs_dict[name] = outs_dictionary[old_name][:, :, next_index-old_j:context-window+(next_index - old_j), :]

    answer = tokenizer.decode(total_input[0], skip_special_tokens=True, clean_up_tokenization_spaces=False)
    return answer

现在我们运行推理：

tokenizer = AutoTokenizer.from_pretrained(\"Esperanto/llama-3.2-3B-Instruct-kvc-fp16-onnx\")
model_path = \"llama-3.2-3B-Instruct-kvc-fp16-onnx/model.onnx\"

max_gen_tokens = 20    #我们想要生成的token数量
total_sequence = 128   #总序列长度
context = 1024         #扩展kvc的上下文
window = 16            #我们想要一次解析的token数量
messages = [
    {\"role\": \"system\", \"content\": \"You are a pirate chatbot who always responds in pirate speak!\"},
    {\"role\": \"user\", \"content\": \"Who are you?\"},
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

generated = generate_text(model_path, prompt, tokenizer, max_gen_tokens, total_sequence, window, context)
print(generated)

Esperanto/llama-3.2-3B-Instruct-kvc-fp16-onnx

作者 Esperanto

text-generation

↓ 0 ♥ 1

创建时间: 2024-10-09 09:11:26+00:00

更新时间: 2024-12-11 14:56:48+00:00

在 Hugging Face 上查看

文件 (263)

.gitattributes

README.md

config.json

model.embed_tokens.weight

model.layers.0.input_layernorm.weight

model.layers.0.post_attention_layernorm.weight

model.layers.1.input_layernorm.weight

model.layers.1.post_attention_layernorm.weight

model.layers.10.input_layernorm.weight

model.layers.10.post_attention_layernorm.weight

model.layers.11.input_layernorm.weight

model.layers.11.post_attention_layernorm.weight

model.layers.12.input_layernorm.weight

model.layers.12.post_attention_layernorm.weight

model.layers.13.input_layernorm.weight

model.layers.13.post_attention_layernorm.weight

model.layers.14.input_layernorm.weight

model.layers.14.post_attention_layernorm.weight

model.layers.15.input_layernorm.weight

model.layers.15.post_attention_layernorm.weight

model.layers.16.input_layernorm.weight

model.layers.16.post_attention_layernorm.weight

model.layers.17.input_layernorm.weight

model.layers.17.post_attention_layernorm.weight

model.layers.18.input_layernorm.weight

model.layers.18.post_attention_layernorm.weight

model.layers.19.input_layernorm.weight

model.layers.19.post_attention_layernorm.weight

model.layers.2.input_layernorm.weight

model.layers.2.post_attention_layernorm.weight

model.layers.20.input_layernorm.weight

model.layers.20.post_attention_layernorm.weight

model.layers.21.input_layernorm.weight

model.layers.21.post_attention_layernorm.weight

model.layers.22.input_layernorm.weight

model.layers.22.post_attention_layernorm.weight

model.layers.23.input_layernorm.weight

model.layers.23.post_attention_layernorm.weight

model.layers.24.input_layernorm.weight

model.layers.24.post_attention_layernorm.weight

model.layers.25.input_layernorm.weight

model.layers.25.post_attention_layernorm.weight

model.layers.26.input_layernorm.weight

model.layers.26.post_attention_layernorm.weight

model.layers.27.input_layernorm.weight

model.layers.27.post_attention_layernorm.weight

model.layers.3.input_layernorm.weight

model.layers.3.post_attention_layernorm.weight

model.layers.4.input_layernorm.weight

model.layers.4.post_attention_layernorm.weight

model.layers.5.input_layernorm.weight

model.layers.5.post_attention_layernorm.weight

model.layers.6.input_layernorm.weight

model.layers.6.post_attention_layernorm.weight

model.layers.7.input_layernorm.weight

model.layers.7.post_attention_layernorm.weight

model.layers.8.input_layernorm.weight

model.layers.8.post_attention_layernorm.weight

model.layers.9.input_layernorm.weight

model.layers.9.post_attention_layernorm.weight

model.norm.weight

model.onnx ONNX

onnx__MatMul_8652

onnx__MatMul_8653

onnx__MatMul_8654

onnx__MatMul_8679

onnx__MatMul_8680

onnx__MatMul_8681

onnx__MatMul_8682

onnx__MatMul_8683

onnx__MatMul_8684

onnx__MatMul_8685

onnx__MatMul_8710

onnx__MatMul_8711

onnx__MatMul_8712

onnx__MatMul_8713

onnx__MatMul_8714

onnx__MatMul_8715

onnx__MatMul_8716

onnx__MatMul_8741

onnx__MatMul_8742

onnx__MatMul_8743

onnx__MatMul_8744

onnx__MatMul_8745

onnx__MatMul_8746

onnx__MatMul_8747

onnx__MatMul_8772

onnx__MatMul_8773

onnx__MatMul_8774

onnx__MatMul_8775

onnx__MatMul_8776

onnx__MatMul_8777

onnx__MatMul_8778

onnx__MatMul_8803

onnx__MatMul_8804

onnx__MatMul_8805

onnx__MatMul_8806

onnx__MatMul_8807

onnx__MatMul_8808

onnx__MatMul_8809

onnx__MatMul_8834

onnx__MatMul_8835

onnx__MatMul_8836

onnx__MatMul_8837

onnx__MatMul_8838

onnx__MatMul_8839

onnx__MatMul_8840

onnx__MatMul_8865

onnx__MatMul_8866

onnx__MatMul_8867

onnx__MatMul_8868

onnx__MatMul_8869

onnx__MatMul_8870

onnx__MatMul_8871

onnx__MatMul_8896

onnx__MatMul_8897

onnx__MatMul_8898

onnx__MatMul_8899

onnx__MatMul_8900

onnx__MatMul_8901

onnx__MatMul_8902

onnx__MatMul_8927

onnx__MatMul_8928

onnx__MatMul_8929

onnx__MatMul_8930

onnx__MatMul_8931

onnx__MatMul_8932

onnx__MatMul_8933

onnx__MatMul_8958

onnx__MatMul_8959

onnx__MatMul_8960

onnx__MatMul_8961

onnx__MatMul_8962

onnx__MatMul_8963

onnx__MatMul_8964

onnx__MatMul_8989

onnx__MatMul_8990

onnx__MatMul_8991

onnx__MatMul_8992

onnx__MatMul_8993

onnx__MatMul_8994

onnx__MatMul_8995

onnx__MatMul_9020

onnx__MatMul_9021

onnx__MatMul_9022

onnx__MatMul_9023

onnx__MatMul_9024

onnx__MatMul_9025

onnx__MatMul_9026

onnx__MatMul_9051

onnx__MatMul_9052

onnx__MatMul_9053

onnx__MatMul_9054

onnx__MatMul_9055

onnx__MatMul_9056

onnx__MatMul_9057

onnx__MatMul_9082

onnx__MatMul_9083

onnx__MatMul_9084

onnx__MatMul_9085

onnx__MatMul_9086

onnx__MatMul_9087

onnx__MatMul_9088

onnx__MatMul_9113

onnx__MatMul_9114

onnx__MatMul_9115

onnx__MatMul_9116

onnx__MatMul_9117

onnx__MatMul_9118

onnx__MatMul_9119

onnx__MatMul_9144

onnx__MatMul_9145

onnx__MatMul_9146

onnx__MatMul_9147

onnx__MatMul_9148

onnx__MatMul_9149

onnx__MatMul_9150

onnx__MatMul_9175

onnx__MatMul_9176

onnx__MatMul_9177

onnx__MatMul_9178

onnx__MatMul_9179

onnx__MatMul_9180

onnx__MatMul_9181

onnx__MatMul_9206

onnx__MatMul_9207

onnx__MatMul_9208

onnx__MatMul_9209

onnx__MatMul_9210

onnx__MatMul_9211

onnx__MatMul_9212

onnx__MatMul_9237

onnx__MatMul_9238

onnx__MatMul_9239

onnx__MatMul_9240

onnx__MatMul_9241

onnx__MatMul_9242

onnx__MatMul_9243

onnx__MatMul_9268

onnx__MatMul_9269

onnx__MatMul_9270

onnx__MatMul_9271

onnx__MatMul_9272

onnx__MatMul_9273

onnx__MatMul_9274

onnx__MatMul_9299

onnx__MatMul_9300

onnx__MatMul_9301

onnx__MatMul_9302

onnx__MatMul_9303

onnx__MatMul_9304

onnx__MatMul_9305

onnx__MatMul_9330

onnx__MatMul_9331

onnx__MatMul_9332

onnx__MatMul_9333

onnx__MatMul_9334

onnx__MatMul_9335

onnx__MatMul_9336

onnx__MatMul_9361

onnx__MatMul_9362

onnx__MatMul_9363

onnx__MatMul_9364

onnx__MatMul_9365

onnx__MatMul_9366

onnx__MatMul_9367

onnx__MatMul_9392

onnx__MatMul_9393

onnx__MatMul_9394

onnx__MatMul_9395

onnx__MatMul_9396

onnx__MatMul_9397

onnx__MatMul_9398

onnx__MatMul_9423

onnx__MatMul_9424

onnx__MatMul_9425

onnx__MatMul_9426

onnx__MatMul_9427

onnx__MatMul_9428

onnx__MatMul_9429

onnx__MatMul_9454

onnx__MatMul_9455

onnx__MatMul_9456

onnx__MatMul_9457

onnx__MatMul_9458

onnx__MatMul_9459

onnx__MatMul_9460

onnx__MatMul_9485

onnx__MatMul_9486

onnx__MatMul_9487

onnx__MatMul_9488

onnx__MatMul_9489

onnx__MatMul_9490

onnx__MatMul_9491

onnx__MatMul_9516

onnx__MatMul_9517

onnx__MatMul_9518

onnx__MatMul_9519

onnx__MatMul_9523

special_tokens_map.json

token_id_to_str.json

tokenizer.json

tokenizer_config.json