Upload MB_dLLM_sample.ipynb
Browse files- MB_dLLM_sample.ipynb +244 -0
MB_dLLM_sample.ipynb
ADDED
@@ -0,0 +1,244 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"nbformat": 4,
|
3 |
+
"nbformat_minor": 0,
|
4 |
+
"metadata": {
|
5 |
+
"colab": {
|
6 |
+
"provenance": [],
|
7 |
+
"gpuType": "T4"
|
8 |
+
},
|
9 |
+
"kernelspec": {
|
10 |
+
"name": "python3",
|
11 |
+
"display_name": "Python 3"
|
12 |
+
},
|
13 |
+
"language_info": {
|
14 |
+
"name": "python"
|
15 |
+
},
|
16 |
+
"accelerator": "GPU"
|
17 |
+
},
|
18 |
+
"cells": [
|
19 |
+
{
|
20 |
+
"cell_type": "code",
|
21 |
+
"execution_count": 1,
|
22 |
+
"metadata": {
|
23 |
+
"id": "Q4qAMMPkQhfY"
|
24 |
+
},
|
25 |
+
"outputs": [],
|
26 |
+
"source": [
|
27 |
+
"import os, random, itertools, math, torch\n",
|
28 |
+
"from torch.utils.data import DataLoader\n",
|
29 |
+
"from transformers import (\n",
|
30 |
+
" AutoTokenizer, AutoModelForMaskedLM,\n",
|
31 |
+
" get_cosine_schedule_with_warmup\n",
|
32 |
+
")\n",
|
33 |
+
"from torch.optim import AdamW\n",
|
34 |
+
"from datasets import load_dataset\n",
|
35 |
+
"device = \"cuda\" if torch.cuda.is_available() else \"cpu\""
|
36 |
+
]
|
37 |
+
},
|
38 |
+
{
|
39 |
+
"cell_type": "code",
|
40 |
+
"source": [
|
41 |
+
"model_id = \"johnowhitaker/modernbert-diffusion\"\n",
|
42 |
+
"tokenizer = AutoTokenizer.from_pretrained(model_id)\n",
|
43 |
+
"SEP_ID, CLS_ID, MASK_ID = tokenizer.sep_token_id, tokenizer.cls_token_id, tokenizer.mask_token_id\n",
|
44 |
+
"model = AutoModelForMaskedLM.from_pretrained(model_id, device_map=device)\n",
|
45 |
+
"model.eval();"
|
46 |
+
],
|
47 |
+
"metadata": {
|
48 |
+
"id": "e4kbDTS3Qo_a"
|
49 |
+
},
|
50 |
+
"execution_count": null,
|
51 |
+
"outputs": []
|
52 |
+
},
|
53 |
+
{
|
54 |
+
"cell_type": "code",
|
55 |
+
"source": [
|
56 |
+
"# Single forward pass:\n",
|
57 |
+
"prompt = \"User: Which is the best programming language? \" + tokenizer.sep_token + \" Assistant:\"\n",
|
58 |
+
"prompt_ids = tokenizer.encode(prompt, add_special_tokens=False)\n",
|
59 |
+
"ans_len = 12\n",
|
60 |
+
"ids = [CLS_ID] + prompt_ids + [SEP_ID] + [MASK_ID]*ans_len + [SEP_ID]\n",
|
61 |
+
"with torch.no_grad():\n",
|
62 |
+
" outs = model(input_ids=torch.tensor([ids]).to(device)).logits\n",
|
63 |
+
"print(outs.shape)\n",
|
64 |
+
"out_ids = outs[0].argmax(dim=-1).tolist()\n",
|
65 |
+
"print(tokenizer.decode(out_ids))"
|
66 |
+
],
|
67 |
+
"metadata": {
|
68 |
+
"colab": {
|
69 |
+
"base_uri": "https://localhost:8080/"
|
70 |
+
},
|
71 |
+
"id": "Y7ZwaE3IQzJT",
|
72 |
+
"outputId": "bd8a6d10-41c3-4531-d244-32094e71b1d3"
|
73 |
+
},
|
74 |
+
"execution_count": 3,
|
75 |
+
"outputs": [
|
76 |
+
{
|
77 |
+
"output_type": "stream",
|
78 |
+
"name": "stdout",
|
79 |
+
"text": [
|
80 |
+
"torch.Size([1, 28, 50368])\n",
|
81 |
+
"[CLS]User: Which is the best programming language? \n",
|
82 |
+
" Assistant: Python, Python,,,,,, is Python..[SEP]\n"
|
83 |
+
]
|
84 |
+
}
|
85 |
+
]
|
86 |
+
},
|
87 |
+
{
|
88 |
+
"cell_type": "code",
|
89 |
+
"source": [
|
90 |
+
"# In a loop, keeping the most confident\n",
|
91 |
+
"prompt = \"User: Which is the best programming language? \" + tokenizer.sep_token + \" Assistant:\"\n",
|
92 |
+
"prompt_ids = tokenizer.encode(prompt, add_special_tokens=False)\n",
|
93 |
+
"ans_len = 32\n",
|
94 |
+
"ids = [CLS_ID] + prompt_ids + [SEP_ID] + [MASK_ID]*ans_len + [SEP_ID]\n",
|
95 |
+
"for i in range(ans_len):\n",
|
96 |
+
" with torch.no_grad():\n",
|
97 |
+
" outs = model(input_ids=torch.tensor([ids]).to(device)).logits\n",
|
98 |
+
" out_probs = torch.softmax(outs[0], dim=-1)\n",
|
99 |
+
" mask_locs = (torch.tensor(ids) == MASK_ID).nonzero(as_tuple=True)[0]\n",
|
100 |
+
" new_probs = torch.zeros_like(out_probs)\n",
|
101 |
+
" new_probs[mask_locs] = out_probs[mask_locs]\n",
|
102 |
+
" max_probs, max_locs = new_probs.max(dim=-1)\n",
|
103 |
+
" max_loc = max_probs.argmax(dim=-1)\n",
|
104 |
+
" ids[max_loc] = new_probs[max_loc].argmax().item()\n",
|
105 |
+
"print(tokenizer.decode(ids))"
|
106 |
+
],
|
107 |
+
"metadata": {
|
108 |
+
"colab": {
|
109 |
+
"base_uri": "https://localhost:8080/"
|
110 |
+
},
|
111 |
+
"id": "wadlDG2DUUjX",
|
112 |
+
"outputId": "06317b7c-7f71-4621-e0b6-c173df0839b7"
|
113 |
+
},
|
114 |
+
"execution_count": 24,
|
115 |
+
"outputs": [
|
116 |
+
{
|
117 |
+
"output_type": "stream",
|
118 |
+
"name": "stdout",
|
119 |
+
"text": [
|
120 |
+
"[CLS]User: Which is the best programming language? [SEP] Assistant:[SEP] is the best programming language?\n",
|
121 |
+
"\n",
|
122 |
+
"A: Python is the best programming language. It is simple, powerful, and has a wide range of useful features.[SEP]\n"
|
123 |
+
]
|
124 |
+
}
|
125 |
+
]
|
126 |
+
},
|
127 |
+
{
|
128 |
+
"cell_type": "code",
|
129 |
+
"source": [
|
130 |
+
"# Wrapping that in a function\n",
|
131 |
+
"def sample(q, ans_len=32):\n",
|
132 |
+
" prompt = f\"User: {q} \" + tokenizer.sep_token + \" Assistant:\"\n",
|
133 |
+
" prompt_ids = tokenizer.encode(prompt, add_special_tokens=False)\n",
|
134 |
+
" ids = [CLS_ID] + prompt_ids + [SEP_ID] + [MASK_ID]*ans_len + [SEP_ID]\n",
|
135 |
+
" for i in range(ans_len):\n",
|
136 |
+
" with torch.no_grad():\n",
|
137 |
+
" outs = model(input_ids=torch.tensor([ids]).to(device)).logits\n",
|
138 |
+
" out_probs = torch.softmax(outs[0], dim=-1)\n",
|
139 |
+
" mask_locs = (torch.tensor(ids) == MASK_ID).nonzero(as_tuple=True)[0]\n",
|
140 |
+
" new_probs = torch.zeros_like(out_probs)\n",
|
141 |
+
" new_probs[mask_locs] = out_probs[mask_locs]\n",
|
142 |
+
" max_probs, max_locs = new_probs.max(dim=-1)\n",
|
143 |
+
" max_loc = max_probs.argmax(dim=-1)\n",
|
144 |
+
" ids[max_loc] = new_probs[max_loc].argmax().item()\n",
|
145 |
+
" return tokenizer.decode(ids)"
|
146 |
+
],
|
147 |
+
"metadata": {
|
148 |
+
"id": "FAj0rtmhYcjF"
|
149 |
+
},
|
150 |
+
"execution_count": 25,
|
151 |
+
"outputs": []
|
152 |
+
},
|
153 |
+
{
|
154 |
+
"cell_type": "code",
|
155 |
+
"source": [
|
156 |
+
"sample(\"Tell me a fun fact about cows\")"
|
157 |
+
],
|
158 |
+
"metadata": {
|
159 |
+
"colab": {
|
160 |
+
"base_uri": "https://localhost:8080/",
|
161 |
+
"height": 52
|
162 |
+
},
|
163 |
+
"id": "HAS20X0oZhw5",
|
164 |
+
"outputId": "4f157101-1652-4c25-b67e-b957512bf632"
|
165 |
+
},
|
166 |
+
"execution_count": 26,
|
167 |
+
"outputs": [
|
168 |
+
{
|
169 |
+
"output_type": "execute_result",
|
170 |
+
"data": {
|
171 |
+
"text/plain": [
|
172 |
+
"\"[CLS]User: Tell me a fun fact about cows [SEP] Assistant:[SEP], here's a fun fact about cows:\\n\\nThe fact is that cows are the most intelligent animals in the world. They can think and make decisions.[SEP]\""
|
173 |
+
],
|
174 |
+
"application/vnd.google.colaboratory.intrinsic+json": {
|
175 |
+
"type": "string"
|
176 |
+
}
|
177 |
+
},
|
178 |
+
"metadata": {},
|
179 |
+
"execution_count": 26
|
180 |
+
}
|
181 |
+
]
|
182 |
+
},
|
183 |
+
{
|
184 |
+
"cell_type": "code",
|
185 |
+
"source": [
|
186 |
+
"sample(\"Tell me a funny joke about lemons\")"
|
187 |
+
],
|
188 |
+
"metadata": {
|
189 |
+
"colab": {
|
190 |
+
"base_uri": "https://localhost:8080/",
|
191 |
+
"height": 52
|
192 |
+
},
|
193 |
+
"id": "f0S3ZQLNUUnU",
|
194 |
+
"outputId": "ddfc0e47-bbb1-496b-8177-5d796b8bd9af"
|
195 |
+
},
|
196 |
+
"execution_count": 30,
|
197 |
+
"outputs": [
|
198 |
+
{
|
199 |
+
"output_type": "execute_result",
|
200 |
+
"data": {
|
201 |
+
"text/plain": [
|
202 |
+
"'[CLS]User: Tell me a funny joke about lemons [SEP] Assistant:[SEP]\\'s a funny joke about lemons: \"I have a lemonade stand, and I\\'m going to sell lemons.\"\\n Assistant: That\\'s funny.[SEP]'"
|
203 |
+
],
|
204 |
+
"application/vnd.google.colaboratory.intrinsic+json": {
|
205 |
+
"type": "string"
|
206 |
+
}
|
207 |
+
},
|
208 |
+
"metadata": {},
|
209 |
+
"execution_count": 30
|
210 |
+
}
|
211 |
+
]
|
212 |
+
},
|
213 |
+
{
|
214 |
+
"cell_type": "code",
|
215 |
+
"source": [
|
216 |
+
"sample(\"Which OS is best?\")"
|
217 |
+
],
|
218 |
+
"metadata": {
|
219 |
+
"colab": {
|
220 |
+
"base_uri": "https://localhost:8080/",
|
221 |
+
"height": 52
|
222 |
+
},
|
223 |
+
"id": "KugOpLPHaQSA",
|
224 |
+
"outputId": "43767abf-5a3e-48e0-c14b-b180f7ba9a14"
|
225 |
+
},
|
226 |
+
"execution_count": 31,
|
227 |
+
"outputs": [
|
228 |
+
{
|
229 |
+
"output_type": "execute_result",
|
230 |
+
"data": {
|
231 |
+
"text/plain": [
|
232 |
+
"\"[CLS]User: Which OS is best? [SEP] Assistant:[SEP], I don't know. I haven't used them personally. I'm sure there are some that are better than others, but I can't tell you.[SEP]\""
|
233 |
+
],
|
234 |
+
"application/vnd.google.colaboratory.intrinsic+json": {
|
235 |
+
"type": "string"
|
236 |
+
}
|
237 |
+
},
|
238 |
+
"metadata": {},
|
239 |
+
"execution_count": 31
|
240 |
+
}
|
241 |
+
]
|
242 |
+
}
|
243 |
+
]
|
244 |
+
}
|