File size: 14,579 Bytes
b40998e |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 |
{
"cells": [
{
"attachments": {},
"cell_type": "markdown",
"metadata": {},
"source": [
"### Transformer Theoretical Model\n",
"\n",
"This notebook stores a bunch of analysis about a Transformer, e.g. estimates the number of FLOPs, parameters, peak memory footprint, checkpoint size, etc."
]
},
{
"cell_type": "code",
"execution_count": 1,
"metadata": {},
"outputs": [],
"source": [
"from collections import OrderedDict"
]
},
{
"cell_type": "code",
"execution_count": 2,
"metadata": {},
"outputs": [],
"source": [
"# config_args = {\n",
"# 'gpt2': dict(n_layer=12, n_head=12, n_embd=768), # 124M params\n",
"# 'gpt2-medium': dict(n_layer=24, n_head=16, n_embd=1024), # 350M params\n",
"# 'gpt2-large': dict(n_layer=36, n_head=20, n_embd=1280), # 774M params\n",
"# 'gpt2-xl': dict(n_layer=48, n_head=25, n_embd=1600), # 1558M params\n",
"# }[model_type]\n",
"\n",
"block_size = 1024\n",
"vocab_size = 50257\n",
"n_layer = 12\n",
"n_head = 12\n",
"n_embd = 768\n",
"bias = False\n",
"assert not bias, \"this notebook assumes bias=False just for simplicity\""
]
},
{
"cell_type": "code",
"execution_count": 3,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"we see: 124337664, expected: 124337664, match: True\n",
"name params ratio (%) \n",
"emebedding/position 786432 0.6325\n",
"embedding/token 38597376 31.0424\n",
"embedding 39383808 31.6749\n",
"attention/ln 768 0.0006\n",
"attention/kqv 1769472 1.4231\n",
"attention/proj 589824 0.4744\n",
"attention 2360064 1.8981\n",
"mlp/ln 768 0.0006\n",
"mlp/ffw 2359296 1.8975\n",
"mlp/proj 2359296 1.8975\n",
"mlp 4719360 3.7956\n",
"block 7079424 5.6937\n",
"transformer 84953088 68.3245\n",
"ln_f 768 0.0006\n",
"dense 0 0.0000\n",
"total 124337664 100.0000\n"
]
}
],
"source": [
"def params():\n",
" \"\"\" estimates the number of parameters in the model\"\"\"\n",
" out = OrderedDict()\n",
"\n",
" # token and position embeddings\n",
" out['emebedding/position'] = n_embd * block_size\n",
" out['embedding/token'] = n_embd * vocab_size\n",
" out['embedding'] = out['emebedding/position'] + out['embedding/token']\n",
"\n",
" # attention blocks\n",
" out['attention/ln'] = n_embd # note, bias=False in our LN\n",
" out['attention/kqv'] = n_embd * 3*n_embd\n",
" out['attention/proj'] = n_embd**2\n",
" out['attention'] = out['attention/ln'] + out['attention/kqv'] + out['attention/proj']\n",
"\n",
" # MLP blocks\n",
" ffw_size = 4*n_embd # feed forward size\n",
" out['mlp/ln'] = n_embd\n",
" out['mlp/ffw'] = n_embd * ffw_size\n",
" out['mlp/proj'] = ffw_size * n_embd\n",
" out['mlp'] = out['mlp/ln'] + out['mlp/ffw'] + out['mlp/proj']\n",
" \n",
" # the transformer and the rest of it\n",
" out['block'] = out['attention'] + out['mlp']\n",
" out['transformer'] = n_layer * out['block']\n",
" out['ln_f'] = n_embd # final layernorm\n",
" out['dense'] = 0 # 0 because of parameter sharing. This layer uses the weights from the embedding layer\n",
"\n",
" # total\n",
" out['total'] = out['embedding'] + out['transformer'] + out['ln_f'] + out['dense']\n",
"\n",
" return out\n",
"\n",
"# compare our param count to that reported by PyTorch\n",
"p = params()\n",
"params_total = p['total']\n",
"print(f\"we see: {params_total}, expected: {124337664}, match: {params_total == 124337664}\")\n",
"# create a header\n",
"print(f\"{'name':20s} {'params':10s} {'ratio (%)':10s}\")\n",
"for k,v in p.items():\n",
" print(f\"{k:20s} {v:10d} {v/params_total*100:10.4f}\")\n",
" "
]
},
{
"cell_type": "code",
"execution_count": 4,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"est checkpoint size: 1.49 GB\n",
"measured with wc -c ckpt.pt: 1542470366\n",
"fluff ratio: 103.38%\n"
]
}
],
"source": [
"# we can now calculate the size of each checkpoint\n",
"# params are stored in fp32, and the AdamW optimizer has 2 additional buffers per param for statistics\n",
"params_bytes = params_total*4\n",
"params_and_buffers_bytes = params_bytes + 2*params_bytes\n",
"print(f\"est checkpoint size: {params_and_buffers_bytes/1e9:.2f} GB\")\n",
"measured_bytes = 1542470366 # from wc -c ckpt.pt\n",
"print(f\"measured with wc -c ckpt.pt: {measured_bytes}\")\n",
"print(f\"fluff ratio: {measured_bytes/params_and_buffers_bytes*100:.2f}%\")"
]
},
{
"attachments": {},
"cell_type": "markdown",
"metadata": {},
"source": [
"We can also estimate the ratio of our GPU memory that will be taken up just by the weights and the buffers inside the AdamW optimizer"
]
},
{
"cell_type": "code",
"execution_count": 5,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"memory ratio taken up just for parameters: 3.73%\n"
]
}
],
"source": [
"gpu_memory = 40e9 # 40 GB A100 GPU, roughly\n",
"print(f\"memory ratio taken up just for parameters: {params_and_buffers_bytes / gpu_memory * 100:.2f}%\")"
]
},
{
"attachments": {},
"cell_type": "markdown",
"metadata": {},
"source": [
"i.e. not that much of the memory for this tiny model, most of the memory is activations (forward and backward). This of course changes dramatically for larger and larger models."
]
},
{
"attachments": {},
"cell_type": "markdown",
"metadata": {},
"source": [
"Let's estimate FLOPs for a single forward pass."
]
},
{
"cell_type": "code",
"execution_count": 6,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"name flops ratio (%) \n",
"attention/kqv 3623878656 1.2426\n",
"attention/scores 1610612736 0.5522\n",
"attention/reduce 1610612736 0.5522\n",
"attention/proj 1207959552 0.4142\n",
"attention 8053063680 2.7612\n",
"mlp/ffw1 4831838208 1.6567\n",
"mlp/ffw2 4831838208 1.6567\n",
"mlp 9663676416 3.3135\n",
"block 17716740096 6.0747\n",
"transformer 212600881152 72.8963\n",
"dense 79047426048 27.1037\n",
"forward_total 291648307200 100.0000\n",
"backward_total 583296614400 200.0000\n",
"total 874944921600 300.0000\n"
]
}
],
"source": [
"def flops():\n",
" # we only count Weight FLOPs, all other layers (LayerNorm, Softmax, etc) are effectively irrelevant\n",
" # we count actual FLOPs, not MACs. Hence 2* all over the place\n",
" # basically for any matrix multiply A (BxC) @ B (CxD) -> (BxD) flops are 2*B*C*D\n",
"\n",
" out = OrderedDict()\n",
" head_size = n_embd // n_head\n",
"\n",
" # attention blocks\n",
" # 1) the projection to key, query, values\n",
" out['attention/kqv'] = 2 * block_size * (n_embd * 3*n_embd)\n",
" # 2) calculating the attention scores\n",
" out['attention/scores'] = 2 * block_size * block_size * n_embd\n",
" # 3) the reduction of the values (B, nh, T, T) x (B, nh, T, hs) -> (B, nh, T, hs)\n",
" out['attention/reduce'] = 2 * n_head * (block_size * block_size * head_size)\n",
" # 4) the final linear projection\n",
" out['attention/proj'] = 2 * block_size * (n_embd * n_embd)\n",
" out['attention'] = sum(out['attention/'+k] for k in ['kqv', 'scores', 'reduce', 'proj'])\n",
"\n",
" # MLP blocks\n",
" ffw_size = 4*n_embd # feed forward size\n",
" out['mlp/ffw1'] = 2 * block_size * (n_embd * ffw_size)\n",
" out['mlp/ffw2'] = 2 * block_size * (ffw_size * n_embd)\n",
" out['mlp'] = out['mlp/ffw1'] + out['mlp/ffw2']\n",
"\n",
" # the transformer and the rest of it\n",
" out['block'] = out['attention'] + out['mlp']\n",
" out['transformer'] = n_layer * out['block']\n",
" out['dense'] = 2 * block_size * (n_embd * vocab_size)\n",
"\n",
" # forward,backward,total\n",
" out['forward_total'] = out['transformer'] + out['dense']\n",
" out['backward_total'] = 2 * out['forward_total'] # use common estimate of bwd = 2*fwd\n",
" out['total'] = out['forward_total'] + out['backward_total']\n",
"\n",
" return out\n",
" \n",
"# compare our param count to that reported by PyTorch\n",
"f = flops()\n",
"flops_total = f['forward_total']\n",
"print(f\"{'name':20s} {'flops':14s} {'ratio (%)':10s}\")\n",
"for k,v in f.items():\n",
" print(f\"{k:20s} {v:14d} {v/flops_total*100:10.4f}\")\n",
" "
]
},
{
"cell_type": "code",
"execution_count": 7,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"palm_flops: 875062886400, flops: 874944921600, ratio: 1.0001\n"
]
}
],
"source": [
"# now here is an estimate copy pasted from the PaLM paper\n",
"# this formula is often used to calculate MFU (model flops utilization)\n",
"def palm_flops():\n",
" \"\"\"estimate of the model flops following PaLM paper formula\"\"\"\n",
" # non-embedding model parameters. note that we do not subtract the\n",
" # embedding/token params because those are tied and get used in the last layer.\n",
" N = params()['total'] - params()['emebedding/position']\n",
" L, H, Q, T = n_layer, n_head, n_embd//n_head, block_size\n",
" mf_per_token = 6*N + 12*L*H*Q*T\n",
" mf = mf_per_token * block_size\n",
" return mf\n",
"\n",
"print(f\"palm_flops: {palm_flops():d}, flops: {flops()['total']:d}, ratio: {palm_flops()/flops()['total']:.4f}\")"
]
},
{
"attachments": {},
"cell_type": "markdown",
"metadata": {},
"source": [
"Ok they are quite similar, giving some confidence that my math in flops() function was ~ok. Now, A100 is cited at 312TFLOPS bfloat16 on tensor cores. So what is our model flops utilization (MFU)? I trained the model above with a batch_size of 20 and grad_accum of 5, which runs in about 755ms on a single A100 GPU. We get:"
]
},
{
"cell_type": "code",
"execution_count": 8,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"fraction of A100 used: 37.14%\n"
]
}
],
"source": [
"# here is what we currently roughly measure\n",
"batch_size = 20 * 5 # 5 is grad_accum, so total batch size is 100\n",
"measured_time = 0.755 # in seconds per iteration\n",
"measured_throughput = batch_size / measured_time\n",
"flops_achieved = f['total'] * measured_throughput\n",
"\n",
"# A100 is cited to be 312 TFLOPS of bloat16 running on tensor cores\n",
"a100_flops_promised = 312e12\n",
"\n",
"# the fraction of the A100 that we are using:\n",
"print(f\"fraction of A100 used: {flops_achieved / a100_flops_promised * 100:.2f}%\")"
]
},
{
"attachments": {},
"cell_type": "markdown",
"metadata": {},
"source": [
"For reference, we'd prefer to be somewhere around 50%+, and not just for a single GPU but for an entire DDP run. So we still have some work to do, but at least we're within a factor of ~2X of what is achievable with this GPU."
]
},
{
"cell_type": "code",
"execution_count": 9,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"time needed to train the model: 3.46 days\n"
]
}
],
"source": [
"# Finally let's check out the 6ND approximation as total cost of training in FLOPs\n",
"model_size = params()['total'] # this is number of parameters, N\n",
"tokens_num = 300e9 # 300B tokens, this is dataset size in tokens, D\n",
"a100_flops = 312e12 # 312 TFLOPS\n",
"assumed_mfu = 0.3 # assume this model flops utilization (take the current 37% from above and add some DDP overhead)\n",
"flops_throughput = a100_flops * 8 * assumed_mfu # assume an 8XA100 node at 30% utilization\n",
"flops_needed = 6 * model_size * tokens_num # 6ND\n",
"time_needed_s = flops_needed / flops_throughput # in seconds\n",
"print(f\"time needed to train the model: {time_needed_s/3600/24:.2f} days\")"
]
},
{
"attachments": {},
"cell_type": "markdown",
"metadata": {},
"source": [
"This is not a bad estimate at all. I trained this model and it converged in roughly 4 days. Btw as a good reference for where 6ND comes from and some intuition around it I recommend [Dzmitry's post](https://medium.com/@dzmitrybahdanau/the-flops-calculus-of-language-model-training-3b19c1f025e4)."
]
},
{
"attachments": {},
"cell_type": "markdown",
"metadata": {},
"source": [
"Now, FLOPs are just one constraint, the other that we have to keep a close track of is the memory bandwidth. TODO estimate LOAD/STORE costs of our model later."
]
}
],
"metadata": {
"kernelspec": {
"display_name": "pytorch2",
"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": "3.10.8"
},
"orig_nbformat": 4,
"vscode": {
"interpreter": {
"hash": "7f5833218766b48e6e35e4452ee875aac0e2188d05bbe5298f2c62b79f08b222"
}
}
},
"nbformat": 4,
"nbformat_minor": 2
}
|