File size: 17,280 Bytes
6a50f6f |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 |
{
"cells": [
{
"cell_type": "code",
"execution_count": null,
"id": "e0ef8d28",
"metadata": {},
"outputs": [
{
"name": "stderr",
"output_type": "stream",
"text": [
"/Users/yiyunzhu/Library/Caches/pypoetry/virtualenvs/speculativedecoding-B0TTdUOs-py3.13/lib/python3.13/site-packages/tqdm/auto.py:21: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html\n",
" from .autonotebook import tqdm as notebook_tqdm\n",
"`torch_dtype` is deprecated! Use `dtype` instead!\n",
"Fetching 2 files: 100%|██████████| 2/2 [00:58<00:00, 29.32s/it]\n",
"Loading checkpoint shards: 100%|██████████| 2/2 [00:00<00:00, 27.91it/s]\n"
]
}
],
"source": [
"import torch\n",
"from transformers import AutoTokenizer, AutoModelForCausalLM, set_seed\n",
"\n",
"set_seed(67)\n",
"\n",
"device = \"mps\"\n",
"\n",
"tokenizer = AutoTokenizer.from_pretrained(\"Qwen/Qwen2.5-Coder-0.5B-Instruct\") #HuggingFaceTB/SmolLM2-135M-Instruct\n",
"draft_model = AutoModelForCausalLM.from_pretrained(\"Qwen/Qwen2.5-Coder-0.5B-Instruct\", torch_dtype=torch.bfloat16).to(device)\n",
"verify_model = AutoModelForCausalLM.from_pretrained(\"Qwen/Qwen2.5-Coder-3B-Instruct\", torch_dtype=torch.bfloat16).to(device) #HuggingFaceTB/SmolLM2-1.7B-Instruct"
]
},
{
"cell_type": "code",
"execution_count": 8,
"id": "30d81505",
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"'The quick brown fox jumps over the lazy dog'"
]
},
"execution_count": 8,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"prompt = \"The quick brown fox\"\n",
"inputs = tokenizer(prompt, return_tensors=\"pt\").to(device)\n",
"input_ids = inputs[\"input_ids\"]\n",
"\n",
"generated = input_ids.clone() # [1, seq_len]\n",
"draft_probs = []\n",
"for _ in range(5): # gamma = 5\n",
" with torch.no_grad():\n",
" outputs = draft_model(generated)\n",
" logits = outputs.logits[:, -1, :] # batch, seq_len, vocab_size\n",
" \n",
" probs = torch.softmax(logits, dim=-1) # [1, 50257]\n",
" next_token = torch.multinomial(probs, num_samples=1)\n",
"\n",
" draft_probs.append(probs)\n",
" generated = torch.cat([generated, next_token], dim=-1)\n",
"\n",
"tokenizer.decode(generated[0])"
]
},
{
"cell_type": "code",
"execution_count": 9,
"id": "2492ee58",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Token: ' jumps', q(x)=0.9727, p(x)=0.9023\n",
" -> Accepted\n",
"Token: ' over', q(x)=1.0000, p(x)=0.9961\n",
" -> Accepted\n",
"Token: ' the', q(x)=0.9023, p(x)=0.9102\n",
" -> Accepted\n",
"Token: ' lazy', q(x)=1.0000, p(x)=0.9922\n",
" -> Accepted\n",
"Token: ' dog', q(x)=0.9922, p(x)=0.9844\n",
" -> Accepted\n",
" jumps over the lazy dog\n",
"[34208, 916, 279, 15678, 5562]\n"
]
}
],
"source": [
"with torch.no_grad():\n",
" target_outputs = verify_model(generated)\n",
" target_logits = target_outputs.logits[:, -6:-1, :]\n",
"\n",
"target_probs = torch.softmax(target_logits, dim=-1) # [1, 5, 50257]\n",
"accepted_tokens = []\n",
"for i in range(5):\n",
" # if q(x) <= p(x), keep\n",
" # if q(x) > p(x), reject with 1-p(x)/q(x) chance\n",
" # if rejected, we sample from norm(max(0, p(x) - q(x)))\n",
" q = draft_probs[i] # [1, 50257]\n",
" p = target_probs[:, i, :] # [1, 50257]\n",
" token = generated[:, i - 5] # [1]\n",
" # assume unbatched for now\n",
" x = token[0].item()\n",
" q_x = q[0, x].item()\n",
" p_x = p[0, x].item()\n",
"\n",
" print(f\"Token: '{tokenizer.decode(x)}', q(x)={q_x:.4f}, p(x)={p_x:.4f}\")\n",
" if q_x <= p_x:\n",
" print(\" -> Accepted\")\n",
" accepted_tokens.append(x)\n",
" else:\n",
" r = torch.rand(1).item()\n",
" acceptance_rate = p_x / q_x\n",
"\n",
" if r < acceptance_rate:\n",
" print(\" -> Accepted\")\n",
" accepted_tokens.append(x)\n",
" else:\n",
" print(\" -> Rejected\")\n",
" adjusted = torch.clamp(p-q, min=0)\n",
" adjusted = adjusted / adjusted.sum()\n",
" new_token = torch.multinomial(adjusted, num_samples=1)[0].item()\n",
" accepted_tokens.append(new_token)\n",
" break\n",
"\n",
"print(tokenizer.decode(accepted_tokens))\n",
"print(accepted_tokens)"
]
},
{
"cell_type": "code",
"execution_count": 20,
"id": "2df0e2a9",
"metadata": {},
"outputs": [],
"source": [
"def draft(input_ids, gamma, confidence_threshold, eos_token, past_kv):\n",
" generated = input_ids.clone() # [1, seq_len]\n",
" draft_probs = []\n",
" for _ in range(gamma): \n",
" with torch.no_grad():\n",
" outputs = draft_model(\n",
" generated if past_kv is None else generated[:, -1:],\n",
" past_key_values=past_kv,\n",
" use_cache=True\n",
" )\n",
" logits = outputs.logits[:, -1, :] # batch, seq_len, vocab_size\n",
" past_kv = outputs.past_key_values\n",
" \n",
" probs = torch.softmax(logits, dim=-1) # [1, 50257]\n",
"\n",
" confidence = probs.max().item() # dynamic speculative decoding\n",
" if confidence < confidence_threshold and len(draft_probs) > 0:\n",
" break \n",
"\n",
" next_token = torch.argmax(probs, dim=-1, keepdim=True)\n",
"\n",
" draft_probs.append(probs)\n",
" generated = torch.cat([generated, next_token], dim=-1)\n",
"\n",
" if next_token.item() == eos_token:\n",
" break;\n",
"\n",
" return generated, draft_probs, past_kv\n",
"\n",
"def verify(drafted, drafted_probs, eos_token, past_kv):\n",
" draft_len = len(drafted_probs) # number of new drafted tokens\n",
" with torch.no_grad():\n",
" if past_kv is None:\n",
" target_outputs = verify_model(drafted, use_cache=True)\n",
" target_logits = target_outputs.logits[:, -draft_len - 1:-1, :]\n",
" else:\n",
" target_outputs = verify_model(\n",
" drafted[:, -(draft_len + 1):], # extra token \n",
" past_key_values=past_kv,\n",
" use_cache=True\n",
" )\n",
" target_logits = target_outputs.logits[:, :-1, :] # Drop last (predicts bonus token)\n",
"\n",
" past_kv = target_outputs.past_key_values\n",
"\n",
" target_probs = torch.softmax(target_logits, dim=-1) # [1, 5, 50257]\n",
" accepted_tokens = []\n",
" num_accepted = 0 # number of tokens from drafted that is accepted\n",
" for i in range(draft_len):\n",
" # if q(x) <= p(x), keep\n",
" # if q(x) > p(x), reject with 1-p(x)/q(x) chance\n",
" # if rejected, we sample from norm(max(0, p(x) - q(x)))\n",
" q = drafted_probs[i] # [1, 50257]\n",
" p = target_probs[:, i, :] # [1, 50257]\n",
" token = drafted[:, i - draft_len] # [1]\n",
" # assume unbatched for now\n",
" x = token[0].item()\n",
" q_x = q[0, x].item()\n",
" p_x = p[0, x].item()\n",
"\n",
" print(f\"Token: '{tokenizer.decode(x)}'\", end = \"\")\n",
" if q_x <= p_x:\n",
" print(\" -> Accepted\")\n",
" accepted_tokens.append(x)\n",
" num_accepted+=1\n",
" else:\n",
" r = torch.rand(1).item()\n",
" acceptance_rate = p_x / q_x\n",
"\n",
" if r < acceptance_rate:\n",
" print(\" -> Accepted\")\n",
" accepted_tokens.append(x)\n",
" num_accepted+=1\n",
" else:\n",
" print(\" -> Rejected\", end = \"\")\n",
" adjusted = torch.clamp(p-q, min=0)\n",
" adjusted = adjusted / adjusted.sum()\n",
" new_token = torch.multinomial(adjusted, num_samples=1)[0].item()\n",
" accepted_tokens.append(new_token)\n",
" print(tokenizer.decode(new_token))\n",
" break\n",
" if accepted_tokens[-1] == eos_token:\n",
" break\n",
"\n",
" return accepted_tokens, num_accepted, past_kv"
]
},
{
"cell_type": "code",
"execution_count": 24,
"id": "5378f5d5",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Token: 'A' -> Accepted\n",
"Token: ' deal' -> Accepted\n",
"Token: ' flow' -> Accepted\n",
"Token: ' in' -> Accepted\n",
"Token: ' a' -> Accepted\n",
"Token: ' VC' -> Rejected venture\n",
"Token: ' capital' -> Accepted\n",
"Token: ' fund' -> Rejected (\n",
"Token: 'VC' -> Accepted\n",
"Token: ')' -> Accepted\n",
"Token: ' fund' -> Accepted\n",
"Token: ' is' -> Rejected refers\n",
"Token: ' to' -> Accepted\n",
"Token: ' the' -> Accepted\n",
"Token: ' process' -> Accepted\n",
"Token: ' of' -> Accepted\n",
"Token: ' setting' -> Rejected screening\n",
"Token: ',' -> Accepted\n",
"Token: ' evaluating' -> Accepted\n",
"Token: ',' -> Accepted\n",
"Token: ' and' -> Accepted\n",
"Token: ' selecting' -> Accepted\n",
"Token: ' investors' -> Rejected potential\n",
"Token: ' investment' -> Accepted\n",
"Token: ' in' -> Rejected opportunities\n",
"Token: ' for' -> Rejected.\n",
"Token: ' It' -> Accepted\n",
"Token: ' involves' -> Accepted\n",
"Token: ' several' -> Rejected identifying\n",
"Token: ' potential' -> Accepted\n",
"Token: ' investors' -> Rejected companies\n",
"Token: ' that' -> Accepted\n",
"Token: ' could' -> Rejected the\n",
"Token: ' VC' -> Accepted\n",
"Token: ' fund' -> Accepted\n",
"Token: 'ers' -> Rejected is\n",
"Token: ' interested' -> Accepted\n",
"Token: ' in' -> Accepted\n",
"Token: ',' -> Accepted\n",
"Token: ' assessing' -> Accepted\n",
"Token: ' their' -> Accepted\n",
"Token: ' financial' -> Rejected growth\n",
"Token: ' potential' -> Accepted\n",
"Token: ',' -> Accepted\n",
"Token: ' and' -> Accepted\n",
"Token: ' evaluating' -> Rejected business\n",
"Token: ' model' -> Accepted\n",
"Token: ',' -> Accepted\n",
"Token: ' and' -> Accepted\n",
"Token: ',' -> Rejected market\n",
"Token: ' demand' -> Accepted\n",
"Token: ',' -> Accepted\n",
"Token: ' and' -> Accepted\n",
"Token: ' then' -> Accepted\n",
"Token: ' making' -> Accepted\n",
"Token: ' a' -> Accepted\n",
"Token: ' decision' -> Accepted\n",
"Token: ' on' -> Accepted\n",
"Token: ' whether' -> Accepted\n",
"Token: ' to' -> Accepted\n",
"Token: ' invest' -> Accepted\n",
"Token: ' in' -> Accepted\n",
"Token: ' those' -> Rejected the\n",
"Token: ' VC' -> Rejected fund\n",
"Token: ' in' -> Rejected’s\n",
"Token: ' capital' -> Accepted\n",
"Token: '.' -> Accepted\n",
"Token: '<|im_end|>' -> Accepted\n"
]
},
{
"data": {
"text/plain": [
"'<|im_start|>system\\nYou are Qwen, created by Alibaba Cloud. You are a helpful assistant.<|im_end|>\\n<|im_start|>user\\nWhat is a deal flow in a VC fund?<|im_end|>\\n<|im_start|>assistant\\nA deal flow in a venture capital (VC) fund refers to the process of screening, evaluating, and selecting potential investment opportunities. It involves identifying potential companies that the VC fund is interested in, assessing their growth potential, and business model, and market demand, and then making a decision on whether to invest in the fund’s capital.<|im_end|>'"
]
},
"execution_count": 24,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"messages = [{\"role\": \"user\", \"content\": \"What is a deal flow in a VC fund?\"}]\n",
"prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)\n",
"\n",
"max_tokens = 80\n",
"eos_token = tokenizer.eos_token_id\n",
"im_end_token = tokenizer.convert_tokens_to_ids(\"<|im_end|>\")\n",
"gamma = 15\n",
"confidence_threshold = 0.5\n",
"\n",
"inputs = tokenizer(prompt, return_tensors=\"pt\").to(device)\n",
"result = inputs[\"input_ids\"].clone()\n",
"\n",
"draft_kv = None\n",
"verify_kv = None\n",
"\n",
"total_drafted = 0\n",
"total_accepted = 0\n",
"\n",
"while result.shape[-1] - inputs[\"input_ids\"].shape[-1] < max_tokens:\n",
" drafted, drafted_probs, draft_kv = draft(result, gamma, confidence_threshold, eos_token, draft_kv)\n",
" accepted_tokens, num_accepted, verify_kv = verify(drafted, drafted_probs, eos_token, verify_kv)\n",
"\n",
" total_drafted += len(drafted_probs)\n",
" total_accepted += num_accepted\n",
" \n",
" valid_len = result.shape[-1] + num_accepted\n",
" result = torch.cat([result, torch.tensor([accepted_tokens], device=device)], dim=-1)\n",
"\n",
" if draft_kv is not None:\n",
" draft_kv.crop(max_length=valid_len)\n",
" if verify_kv is not None:\n",
" verify_kv.crop(max_length=valid_len)\n",
"\n",
" if eos_token in accepted_tokens or im_end_token in accepted_tokens:\n",
" break\n",
"\n",
"tokenizer.decode(result[0])\n"
]
},
{
"cell_type": "code",
"execution_count": 25,
"id": "40c92741",
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"0.6071428571428571"
]
},
"execution_count": 25,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"total_accepted / total_drafted"
]
},
{
"cell_type": "code",
"execution_count": 26,
"id": "0c661940",
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"'<|im_start|>system\\nYou are Qwen, created by Alibaba Cloud. You are a helpful assistant.<|im_end|>\\n<|im_start|>user\\nWhat is a deal flow in a VC fund?<|im_end|>\\n<|im_start|>assistant\\nA deal flow in a VC fund refers to the collection and processing of new investment opportunities presented to the fund for screening, evaluation, and ultimately investment.\\n\\nHere’s a basic overview:\\n\\n* **Deal Flow Collection**: Fund managers typically collect deals through email, cold calls, calendars, meeting notes, investor referrals, and referrals by other investors. They utilize various channels including online searches, investor networks, and industry'"
]
},
"execution_count": 26,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"inputs = tokenizer(prompt, return_tensors=\"pt\").to(device)\n",
"result = inputs[\"input_ids\"].clone()\n",
"\n",
"past_kv = None\n",
"\n",
"while result.shape[-1] - inputs[\"input_ids\"].shape[-1] < max_tokens:\n",
" with torch.no_grad():\n",
" output = verify_model(\n",
" result if past_kv is None else result[:, -1:],\n",
" past_key_values=past_kv,\n",
" use_cache=True\n",
" )\n",
" logits = output.logits[:, -1, :] # batch, vocab\n",
" \n",
" past_kv = output.past_key_values\n",
" probs = torch.softmax(logits, dim=-1)\n",
" next_token = torch.multinomial(probs, num_samples=1)\n",
"\n",
" result = torch.cat([result, next_token], dim=-1)\n",
" if eos_token in next_token or im_end_token in next_token:\n",
" break\n",
"\n",
"tokenizer.decode(result[0])"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "c7a26417",
"metadata": {},
"outputs": [],
"source": []
}
],
"metadata": {
"kernelspec": {
"display_name": "speculativedecoding-B0TTdUOs-py3.13",
"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": "3.13.5"
}
},
"nbformat": 4,
"nbformat_minor": 5
}
|