File size: 6,729 Bytes
a60b021
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f27847f
a60b021
f27847f
 
 
a60b021
 
 
 
 
 
f27847f
a60b021
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f27847f
 
 
 
a60b021
f27847f
 
 
 
 
 
 
 
 
 
a60b021
f27847f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a60b021
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
{
 "cells": [
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Introduction\n",
    "\n",
    "This tutorial demonstrates how to perform evaluation on a gpt-j-6B-int8 model."
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Prerequisite"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "vscode": {
     "languageId": "plaintext"
    }
   },
   "outputs": [],
   "source": [
    "!pip install onnx onnxruntime torch transformers datasets accelerate"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Run\n",
    "\n",
    "### 1. Get lambada acc"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "vscode": {
     "languageId": "plaintext"
    }
   },
   "outputs": [],
   "source": [
    "from transformers import AutoTokenizer\n",
    "import torch\n",
    "from datasets import load_dataset\n",
    "import onnxruntime as ort\n",
    "from torch.nn.functional import pad\n",
    "\n",
    "# load model\n",
    "model_id = \"EleutherAI/gpt-j-6B\"\n",
    "tokenizer = AutoTokenizer.from_pretrained(model_id)\n",
    "\n",
    "def tokenize_function(examples):\n",
    "    example = tokenizer(examples['text'])\n",
    "    return example\n",
    "\n",
    "# create dataset\n",
    "dataset = load_dataset('lambada', split='validation')\n",
    "dataset = dataset.shuffle(seed=42)\n",
    "dataset = dataset.map(tokenize_function, batched=True)\n",
    "dataset.set_format(type='torch', columns=['input_ids'])\n",
    "\n",
    "# create session\n",
    "options = ort.SessionOptions()\n",
    "options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL\n",
    "session = ort.InferenceSession('/path/to/model.onnx', options, providers=ort.get_available_providers())\n",
    "total, hit = 0, 0\n",
    "index = 1\n",
    "\n",
    "# inference\n",
    "for idx, batch in enumerate(dataset):\n",
    "    input_ids = batch['input_ids'].unsqueeze(0)\n",
    "    label = input_ids[:, -1]\n",
    "    pad_len = 0  ##set to 0\n",
    "    input_ids = pad(input_ids, (0, pad_len), value=1)\n",
    "    ort_inputs = {\n",
    "        'input_ids': input_ids.detach().cpu().numpy(),\n",
    "        'attention_mask': torch.cat([torch.ones(input_ids.shape), torch.ones([1, 1])], dim=-1).detach().cpu().numpy().astype('int64')\n",
    "    }\n",
    "    for i in range(28):\n",
    "        ort_inputs[\"past_key_values.{}.key\".format(i)] = np.zeros((1,16,1,256), dtype='float32')\n",
    "        ort_inputs[\"past_key_values.{}.value\".format(i)] = np.zeros((1,16,1,256), dtype='float32')\n",
    "    predictions = session.run(None, ort_inputs)\n",
    "    outputs = torch.from_numpy(predictions[0]) \n",
    "    last_token_logits = outputs[:, -2 - pad_len, :]\n",
    "    pred = last_token_logits.argmax(dim=-1)\n",
    "    total += label.size(0)\n",
    "    hit += (pred == label).sum().item()\n",
    "\n",
    "acc = hit / total\n",
    "print('acc: ', acc)"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 2. Text Generation"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "vscode": {
     "languageId": "plaintext"
    }
   },
   "outputs": [],
   "source": [
    "import os\n",
    "import time\n",
    "import sys\n",
    "\n",
    "# create session\n",
    "sess_options = ort.SessionOptions()\n",
    "sess_options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL\n",
    "session = ort.InferenceSession('/path/to/model.onnx', sess_options)\n",
    "\n",
    "# input prompt\n",
    "# 32 tokens input\n",
    "prompt = \"Once upon a time, there existed a little girl, who liked to have adventures.\" + \\\n",
    "                 \" She wanted to go to places and meet new people, and have fun.\"\n",
    "\n",
    "print(\"prompt: \", prompt)\n",
    "\n",
    "total_time = 0.0\n",
    "num_iter = 10\n",
    "num_warmup = 3\n",
    "\n",
    "# start\n",
    "for idx in range(num_iter):\n",
    "    text = []\n",
    "    tic = time.time()\n",
    "\n",
    "    input_ids = tokenizer(prompt, return_tensors=\"pt\").input_ids\n",
    "\n",
    "    attention_mask = torch.ones(input_ids.shape[1] +1)\n",
    "    attention_mask[0] = 0\n",
    "    attention_mask = attention_mask.unsqueeze(0)\n",
    "\n",
    "    inp = {'input_ids': input_ids.detach().cpu().numpy(),\n",
    "            'attention_mask': attention_mask.detach().cpu().numpy().astype('int64')}\n",
    "    for i in range(28):\n",
    "        inp[\"past_key_values.{}.key\".format(i)] = torch.zeros([1,16,1,256]).detach().cpu().numpy()\n",
    "        inp[\"past_key_values.{}.value\".format(i)] = torch.zeros([1,16,1,256]).detach().cpu().numpy()\n",
    "\n",
    "    for i in range(32):\n",
    "\n",
    "        output = session.run(None, inp)\n",
    "        logits = output[0]\n",
    "        logits = torch.from_numpy(logits)\n",
    "        next_token_logits = logits[:, -1, :]\n",
    "        probs = torch.nn.functional.softmax(next_token_logits, dim=-1)\n",
    "        next_tokens = torch.argmax(probs, dim=-1)\n",
    "        present_kv = output[1]\n",
    "        for i in range(28):\n",
    "\n",
    "            if step == 0:\n",
    "                inp[\"past_key_values.{}.key\".format(i)] = output[2*i+1][:, :, 1:, :]\n",
    "                inp[\"past_key_values.{}.value\".format(i)] = output[2*i+2][:, :, 1:, :]\n",
    "            else:\n",
    "                inp[\"past_key_values.{}.key\".format(i)] = output[2*i+1]\n",
    "                inp[\"past_key_values.{}.value\".format(i)] = output[2*i+2]\n",
    "\n",
    "        input_ids = torch.cat([input_ids, next_tokens[:, None]], dim=-1)\n",
    "        if step == 0:\n",
    "            attention_mask = torch.cat([attention_mask[:, 1:], torch.ones([1, 1])], dim=-1)\n",
    "        else:\n",
    "            attention_mask = torch.cat([attention_mask, torch.ones([1, 1])], dim=-1)\n",
    "\n",
    "        inp['attention_mask'] = attention_mask.detach().cpu().numpy().astype('int64')\n",
    "        inp['input_ids'] = input_ids[:, -1:].detach().cpu().numpy()\n",
    "\n",
    "    print(tokenizer.decode(input_ids[0]))\n",
    "    toc = time.time()\n",
    "    if idx >= num_warmup:\n",
    "        total_time += (toc - tic)\n",
    "print(\"Inference latency: %.3f s.\" % (total_time / (num_iter - num_warmup)))"
   ]
  }
 ],
 "metadata": {
  "language_info": {
   "name": "python"
  },
  "orig_nbformat": 4
 },
 "nbformat": 4,
 "nbformat_minor": 2
}