返回模型

说明文档

启用键值缓存的 Llama 3.2 1B ONNX fp16 格式模型

模型创建者：Meta-Llama
原始模型：Meta-Llama Llama 3.2 1B

描述

本仓库包含由 Esperanto Technologies 转换的 Llama 3.2 1B ONNX 格式文件。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-3.2-1B-kvc-fp16-onnx --local-dir llama-3.2-1B-kvc-fp16-onnx --local-dir-use-symlinks False

有关使用 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 在 Python 代码中运行

该模型可以在 CPU 上轻松运行，使用 ONNXRuntime。

首先安装依赖包

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

示例代码：使用该模型生成文本

我们定义一个使用贪婪解码的循环：

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(\"错误：需要更长的总序列长度！\")
        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, 64], 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)
    ## 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)}
        #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

现在运行推理：

tokenizer = AutoTokenizer.from_pretrained(\"Esperanto/llama-3.2-1B-kvc-fp16-onnx\")
model_path = \"llama-3.2-1B-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的上下文长度
window = 16            #每次想要解析的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-3.2-1B-kvc-fp16-onnx

作者 Esperanto

text-generation

↓ 0 ♥ 0

创建时间: 2024-10-09 09:17:36+00:00

更新时间: 2024-12-11 14:44:32+00:00

在 Hugging Face 上查看

文件 (155)

.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.2.input_layernorm.weight

model.layers.2.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_5073

onnx__MatMul_5074

onnx__MatMul_5075

onnx__MatMul_5100

onnx__MatMul_5101

onnx__MatMul_5102

onnx__MatMul_5103

onnx__MatMul_5104

onnx__MatMul_5105

onnx__MatMul_5106

onnx__MatMul_5131

onnx__MatMul_5132

onnx__MatMul_5133

onnx__MatMul_5134

onnx__MatMul_5135

onnx__MatMul_5136

onnx__MatMul_5137

onnx__MatMul_5162

onnx__MatMul_5163

onnx__MatMul_5164

onnx__MatMul_5165

onnx__MatMul_5166

onnx__MatMul_5167

onnx__MatMul_5168

onnx__MatMul_5193

onnx__MatMul_5194

onnx__MatMul_5195

onnx__MatMul_5196

onnx__MatMul_5197

onnx__MatMul_5198

onnx__MatMul_5199

onnx__MatMul_5224

onnx__MatMul_5225

onnx__MatMul_5226

onnx__MatMul_5227

onnx__MatMul_5228

onnx__MatMul_5229

onnx__MatMul_5230

onnx__MatMul_5255

onnx__MatMul_5256

onnx__MatMul_5257

onnx__MatMul_5258

onnx__MatMul_5259

onnx__MatMul_5260

onnx__MatMul_5261

onnx__MatMul_5286

onnx__MatMul_5287

onnx__MatMul_5288

onnx__MatMul_5289

onnx__MatMul_5290

onnx__MatMul_5291

onnx__MatMul_5292

onnx__MatMul_5317

onnx__MatMul_5318

onnx__MatMul_5319

onnx__MatMul_5320

onnx__MatMul_5321

onnx__MatMul_5322

onnx__MatMul_5323

onnx__MatMul_5348

onnx__MatMul_5349

onnx__MatMul_5350

onnx__MatMul_5351

onnx__MatMul_5352

onnx__MatMul_5353

onnx__MatMul_5354

onnx__MatMul_5379

onnx__MatMul_5380

onnx__MatMul_5381

onnx__MatMul_5382

onnx__MatMul_5383

onnx__MatMul_5384

onnx__MatMul_5385

onnx__MatMul_5410

onnx__MatMul_5411

onnx__MatMul_5412

onnx__MatMul_5413

onnx__MatMul_5414

onnx__MatMul_5415

onnx__MatMul_5416

onnx__MatMul_5441

onnx__MatMul_5442

onnx__MatMul_5443

onnx__MatMul_5444

onnx__MatMul_5445

onnx__MatMul_5446

onnx__MatMul_5447

onnx__MatMul_5472

onnx__MatMul_5473

onnx__MatMul_5474

onnx__MatMul_5475

onnx__MatMul_5476

onnx__MatMul_5477

onnx__MatMul_5478

onnx__MatMul_5503

onnx__MatMul_5504

onnx__MatMul_5505

onnx__MatMul_5506

onnx__MatMul_5507

onnx__MatMul_5508

onnx__MatMul_5509

onnx__MatMul_5534

onnx__MatMul_5535

onnx__MatMul_5536

onnx__MatMul_5537

onnx__MatMul_5538

onnx__MatMul_5539

onnx__MatMul_5540

onnx__MatMul_5565

onnx__MatMul_5566

onnx__MatMul_5567

onnx__MatMul_5568

onnx__MatMul_5572

special_tokens_map.json

token_id_to_str.json

tokenizer.json

tokenizer_config.json