File size: 5,735 Bytes
be19c03 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 |
{
"cells": [
{
"cell_type": "code",
"execution_count": 1,
"metadata": {},
"outputs": [],
"source": [
"import torch\n",
"import torch.nn as nn\n",
"from torch import Tensor\n",
"import random\n",
"from tqdm.auto import tqdm\n",
"from mamba_ssm.modules.mamba_simple import Mamba\n",
"from pathlib import Path\n",
"from mambabit import string_to_bits, bits_to_string\n",
"def model_numel(m: nn.Module):\n",
" return sum(p.numel() for p in m.parameters())"
]
},
{
"cell_type": "code",
"execution_count": 2,
"metadata": {},
"outputs": [],
"source": [
"train_txt = Path(\"~/Downloads/TinyStories/TinyStoriesV2-GPT4-train.txt\").expanduser().read_text()"
]
},
{
"cell_type": "code",
"execution_count": 3,
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"2226845268"
]
},
"execution_count": 3,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"len(train_txt)"
]
},
{
"cell_type": "code",
"execution_count": 4,
"metadata": {},
"outputs": [],
"source": [
"def random_batches(raw_text: str, n_batch: int, bs: int):\n",
" assert bs % 8 == 0, \"have mercy\"\n",
" bs_bytes = bs // 8\n",
" max_allowed_pos = len(raw_text) - bs_bytes\n",
"\n",
" texts = []\n",
" for i in range(n_batch):\n",
" pos = random.randint(0, max_allowed_pos)\n",
" texts.append(raw_text[pos:pos+bs_bytes])\n",
" \n",
" tensors = [string_to_bits(text) for text in texts]\n",
" # in case we met unicode, there will be non-uniform lengths. Trim'em\n",
" common_len = min(t.shape[0] for t in tensors)\n",
" tensors = [t[:common_len] for t in tensors]\n",
" batch = torch.stack(tensors)\n",
" return batch.to(\"cuda\")\n"
]
},
{
"cell_type": "code",
"execution_count": 5,
"metadata": {},
"outputs": [],
"source": [
"from mambabit import MambaBit, n_vocab"
]
},
{
"cell_type": "code",
"execution_count": 6,
"metadata": {},
"outputs": [],
"source": [
"mamba_bit = MambaBit().cuda().bfloat16()"
]
},
{
"cell_type": "code",
"execution_count": 7,
"metadata": {},
"outputs": [],
"source": [
"if False:\n",
" mamba_bit.load_state_dict(torch.load(\"mamba_bit.tiny.bin\"))"
]
},
{
"cell_type": "code",
"execution_count": 8,
"metadata": {},
"outputs": [],
"source": [
"def train(m: nn.Module, \n",
" n_epoch: int = 100, \n",
" n_batch: int = 4, \n",
" bs: int = 256):\n",
" opt = torch.optim.AdamW(m.parameters(), lr=0.0005, fused=True)\n",
"\n",
" for e in (bar := tqdm(range(n_epoch))): \n",
" b = random_batches(train_txt, n_batch, bs)\n",
"\n",
" y_pred = m(b)\n",
" y_pred = y_pred[:, :-1].reshape(-1, n_vocab)\n",
" y_true = b[:, 1:].ravel()\n",
"\n",
" loss = F.cross_entropy(y_pred,y_true)\n",
" loss.backward()\n",
" opt.step()\n",
" opt.zero_grad()\n",
" \n",
" l = loss.item()\n",
" bar.set_description(f\"L:{l:.10f}\")"
]
},
{
"cell_type": "code",
"execution_count": 34,
"metadata": {},
"outputs": [
{
"name": "stderr",
"output_type": "stream",
"text": [
" 0%| | 0/10000 [00:00<?, ?it/s]"
]
},
{
"name": "stderr",
"output_type": "stream",
"text": [
"L:0.0805664062: 100%|██████████| 10000/10000 [6:15:25<00:00, 2.25s/it] \n"
]
}
],
"source": [
"if True:\n",
" train(mamba_bit, 10000, 10, 8*2560 )\n"
]
},
{
"cell_type": "code",
"execution_count": 36,
"metadata": {},
"outputs": [],
"source": [
"torch.save(mamba_bit.state_dict(), \"mamba_bit.tiny.bin\")"
]
},
{
"cell_type": "code",
"execution_count": 42,
"metadata": {},
"outputs": [
{
"name": "stderr",
"output_type": "stream",
"text": [
" 0%| | 0/1024 [00:00<?, ?it/s]"
]
},
{
"name": "stderr",
"output_type": "stream",
"text": [
"100%|██████████| 1024/1024 [00:01<00:00, 760.83it/s]"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"['Once upon a time, there lived a kitten named Lily. Lily loved to play with her friends, and they all liked to play together.\\nOne day, Lily and Ben were playing in the']\n"
]
},
{
"name": "stderr",
"output_type": "stream",
"text": [
"\n"
]
}
],
"source": [
"# TEST\n",
"@torch.no_grad()\n",
"def test(prompt: str, chars=10):\n",
" x0 = string_to_bits(prompt).cuda()[None]\n",
" x = x0.clone()\n",
" process = chars * 8\n",
" for _ in tqdm(range(process)):\n",
" y = mamba_bit(x)\n",
" new = y[:, -1:].argmax(-1)\n",
" x = torch.cat((x, new), 1)\n",
" return bits_to_string(x)\n",
"\n",
" \n",
"print(test(\"Once upon a time, there lived a kitten\", chars=128))"
]
}
],
"metadata": {
"kernelspec": {
"display_name": "sd",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.12.3"
}
},
"nbformat": 4,
"nbformat_minor": 2
}
|