返回模型

说明文档

Phi 3 mini 4k instruct 启用键值缓存的 ONNX fp16 格式版本

模型创建者：Microsoft
原始模型：Phi 3 mini 4k instruct

描述

本仓库包含由 Esperanto Technologies 转换的 Phi 3 mini 4k instruct 的 ONNX 文件。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将任意单个模型文件高速下载到当前目录：

huggingface-cli download Esperanto/phi3-mini-4k-instruct-kvc-fp16-onnx  --local-dir phi3-mini-4k-instruct-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 从 Python 代码运行

此模型可以使用 ONNXRuntime 在 CPU 上轻松运行。

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生成
    inputs_names =[node.name for node in model.graph.input]
    output_names =[node.name for node in model.graph.output]
    n_heads = 32 #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, 96], 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/phi3-mini-4k-instruct-kvc-fp16-onnx\")
model_path = \"phi3-mini-4k-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         #用于扩展kvc的上下文
window = 16            #我们一次想要解析的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/phi3-mini-4k-instruct-kvc-fp16-onnx

作者 Esperanto

↓ 0 ♥ 0

创建时间: 2024-08-02 11:25:15+00:00

更新时间: 2024-12-11 14:13:28+00:00

在 Hugging Face 上查看

文件 (204)

.gitattributes

README.md

added_tokens.json

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.28.input_layernorm.weight

model.layers.28.post_attention_layernorm.weight

model.layers.29.input_layernorm.weight

model.layers.29.post_attention_layernorm.weight

model.layers.3.input_layernorm.weight

model.layers.3.post_attention_layernorm.weight

model.layers.30.input_layernorm.weight

model.layers.30.post_attention_layernorm.weight

model.layers.31.input_layernorm.weight

model.layers.31.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_6116

onnx__MatMul_6134

onnx__MatMul_6135

onnx__MatMul_6136

onnx__MatMul_6137

onnx__MatMul_6153

onnx__MatMul_6154

onnx__MatMul_6155

onnx__MatMul_6156

onnx__MatMul_6172

onnx__MatMul_6173

onnx__MatMul_6174

onnx__MatMul_6175

onnx__MatMul_6191

onnx__MatMul_6192

onnx__MatMul_6193

onnx__MatMul_6194

onnx__MatMul_6210

onnx__MatMul_6211

onnx__MatMul_6212

onnx__MatMul_6213

onnx__MatMul_6229

onnx__MatMul_6230

onnx__MatMul_6231

onnx__MatMul_6232

onnx__MatMul_6248

onnx__MatMul_6249

onnx__MatMul_6250

onnx__MatMul_6251

onnx__MatMul_6267

onnx__MatMul_6268

onnx__MatMul_6269

onnx__MatMul_6270

onnx__MatMul_6286

onnx__MatMul_6287

onnx__MatMul_6288

onnx__MatMul_6289

onnx__MatMul_6305

onnx__MatMul_6306

onnx__MatMul_6307

onnx__MatMul_6308

onnx__MatMul_6324

onnx__MatMul_6325

onnx__MatMul_6326

onnx__MatMul_6327

onnx__MatMul_6343

onnx__MatMul_6344

onnx__MatMul_6345

onnx__MatMul_6346

onnx__MatMul_6362

onnx__MatMul_6363

onnx__MatMul_6364

onnx__MatMul_6365

onnx__MatMul_6381

onnx__MatMul_6382

onnx__MatMul_6383

onnx__MatMul_6384

onnx__MatMul_6400

onnx__MatMul_6401

onnx__MatMul_6402

onnx__MatMul_6403

onnx__MatMul_6419

onnx__MatMul_6420

onnx__MatMul_6421

onnx__MatMul_6422

onnx__MatMul_6438

onnx__MatMul_6439

onnx__MatMul_6440

onnx__MatMul_6441

onnx__MatMul_6457

onnx__MatMul_6458

onnx__MatMul_6459

onnx__MatMul_6460

onnx__MatMul_6476

onnx__MatMul_6477

onnx__MatMul_6478

onnx__MatMul_6479

onnx__MatMul_6495

onnx__MatMul_6496

onnx__MatMul_6497

onnx__MatMul_6498

onnx__MatMul_6514

onnx__MatMul_6515

onnx__MatMul_6516

onnx__MatMul_6517

onnx__MatMul_6533

onnx__MatMul_6534

onnx__MatMul_6535

onnx__MatMul_6536

onnx__MatMul_6552

onnx__MatMul_6553

onnx__MatMul_6554

onnx__MatMul_6555

onnx__MatMul_6571

onnx__MatMul_6572

onnx__MatMul_6573

onnx__MatMul_6574

onnx__MatMul_6590

onnx__MatMul_6591

onnx__MatMul_6592

onnx__MatMul_6593

onnx__MatMul_6609

onnx__MatMul_6610

onnx__MatMul_6611

onnx__MatMul_6612

onnx__MatMul_6628

onnx__MatMul_6629

onnx__MatMul_6630

onnx__MatMul_6631

onnx__MatMul_6647

onnx__MatMul_6648

onnx__MatMul_6649

onnx__MatMul_6650

onnx__MatMul_6666

onnx__MatMul_6667

onnx__MatMul_6668

onnx__MatMul_6669

onnx__MatMul_6685

onnx__MatMul_6686

onnx__MatMul_6687

onnx__MatMul_6688

onnx__MatMul_6704

onnx__MatMul_6705

onnx__MatMul_6706

onnx__MatMul_6707

onnx__MatMul_6723

onnx__MatMul_6724

onnx__MatMul_6725

onnx__MatMul_6726

special_tokens_map.json

tokenizer.json

tokenizer.model

tokenizer_config.json