Spaces:
Running
Running
File size: 12,684 Bytes
ca56e6a |
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 |
import gc
import time
import uuid
from threading import Thread
from types import MethodType
from typing import Iterable, Dict, Any
import torch
from transformers import (
TextIteratorStreamer,
PreTrainedModel,
PreTrainedTokenizer,
)
from api.generation.qwen import check_is_qwen
from api.generation.utils import (
prepare_logits_processor,
is_partial_stop,
apply_stopping_strings,
)
@torch.inference_mode()
def generate_stream(
model: PreTrainedModel,
tokenizer: PreTrainedTokenizer,
params: Dict[str, Any],
):
# Read parameters
input_ids = params.get("inputs")
prompt = params.get("prompt")
model_name = params.get("model", "llm")
temperature = float(params.get("temperature", 1.0))
repetition_penalty = float(params.get("repetition_penalty", 1.0))
top_p = float(params.get("top_p", 1.0))
top_k = int(params.get("top_k", -1)) # -1 means disable
max_new_tokens = int(params.get("max_tokens", 256))
logprobs = params.get("logprobs")
echo = bool(params.get("echo", True))
stop_str = params.get("stop")
stop_token_ids = params.get("stop_token_ids") or []
if tokenizer.eos_token_id not in stop_token_ids:
stop_token_ids.append(tokenizer.eos_token_id)
logits_processor = prepare_logits_processor(
temperature, repetition_penalty, top_p, top_k
)
output_ids = list(input_ids)
input_echo_len = len(input_ids)
device = model.device
if model.config.is_encoder_decoder:
encoder_output = model.encoder(
input_ids=torch.as_tensor([input_ids], device=device)
)[0]
start_ids = torch.as_tensor(
[[model.generation_config.decoder_start_token_id]],
dtype=torch.int64,
device=device,
)
else:
start_ids = torch.as_tensor([input_ids], device=device)
past_key_values, sent_interrupt = None, False
token_logprobs = [None] # The first token has no logprobs.
completion_id: str = f"cmpl-{str(uuid.uuid4())}"
created: int = int(time.time())
previous_text = ""
for i in range(max_new_tokens):
if i == 0: # prefill
if model.config.is_encoder_decoder:
out = model.decoder(
input_ids=start_ids,
encoder_hidden_states=encoder_output,
use_cache=True,
)
logits = model.lm_head(out[0])
else:
out = model(torch.as_tensor([input_ids], device=device), use_cache=True)
logits = out.logits
past_key_values = out.past_key_values
if logprobs is not None:
# Prefull logprobs for the prompt.
shift_input_ids = start_ids[..., 1:].contiguous()
shift_logits = logits[..., :-1, :].contiguous()
shift_logits = torch.log_softmax(shift_logits, dim=-1).tolist()
for label_id, logit in zip(
shift_input_ids[0].tolist(), shift_logits[0]
):
token_logprobs.append(logit[label_id])
else: # decoding
if model.config.is_encoder_decoder:
out = model.decoder(
input_ids=torch.as_tensor(
[output_ids if sent_interrupt else [token]], device=device
),
encoder_hidden_states=encoder_output,
use_cache=True,
past_key_values=None if sent_interrupt else past_key_values,
)
sent_interrupt = False
logits = model.lm_head(out[0])
else:
out = model(
input_ids=torch.as_tensor(
[output_ids if sent_interrupt else [token]], device=device
),
use_cache=True,
past_key_values=None if sent_interrupt else past_key_values,
)
sent_interrupt = False
logits = out.logits
past_key_values = out.past_key_values
if logits_processor:
if repetition_penalty > 1.0:
tmp_output_ids = torch.as_tensor([output_ids], device=logits.device)
else:
tmp_output_ids = None
last_token_logits = logits_processor(tmp_output_ids, logits[:, -1, :])[0]
else:
last_token_logits = logits[0, -1, :]
if device == "mps":
# Switch to CPU by avoiding some bugs in mps backend.
last_token_logits = last_token_logits.float().to("cpu")
if temperature < 1e-5 or top_p < 1e-8: # greedy
_, indices = torch.topk(last_token_logits, 2)
tokens = [int(index) for index in indices.tolist()]
else:
probs = torch.softmax(last_token_logits, dim=-1)
indices = torch.multinomial(probs, num_samples=2)
tokens = [int(token) for token in indices.tolist()]
token = tokens[0]
output_ids.append(token)
if logprobs is not None:
# Cannot use last_token_logits because logprobs is based on raw logits.
token_logprobs.append(
torch.log_softmax(logits[0, -1, :], dim=-1)[token].tolist()
)
if token in stop_token_ids:
stopped = True
else:
stopped = False
# Yield the output tokens
if i % 2 == 0 or i == max_new_tokens - 1 or stopped:
if echo:
tmp_output_ids = output_ids
rfind_start = len(prompt)
else:
tmp_output_ids = output_ids[input_echo_len:]
rfind_start = 0
output = tokenizer.decode(
tmp_output_ids,
skip_special_tokens=False if check_is_qwen(model) else True, # fix for qwen react
spaces_between_special_tokens=False,
clean_up_tokenization_spaces=True,
)
ret_logprobs = None
if logprobs is not None:
ret_logprobs = {
"text_offset": [],
"tokens": [
tokenizer.decode(token)
for token in (
output_ids if echo else output_ids[input_echo_len:]
)
],
"token_logprobs": token_logprobs if echo else token_logprobs[input_echo_len:],
"top_logprobs": [{}] * len(token_logprobs if echo else token_logprobs[input_echo_len:]),
}
# Compute text_offset
curr_pos = 0
for text in ret_logprobs["tokens"]:
ret_logprobs["text_offset"].append(curr_pos)
curr_pos += len(text)
partially_stopped, finish_reason = False, None
if stop_str:
if isinstance(stop_str, str):
pos = output.rfind(stop_str, rfind_start)
if pos != -1:
output = output[:pos]
stopped = True
else:
partially_stopped = is_partial_stop(output, stop_str)
elif isinstance(stop_str, Iterable):
for each_stop in stop_str:
pos = output.rfind(each_stop, rfind_start)
if pos != -1:
output = output[:pos]
stopped = True
if each_stop == "Observation:":
finish_reason = "function_call"
break
else:
partially_stopped = is_partial_stop(output, each_stop)
if partially_stopped:
break
else:
raise ValueError("Invalid stop field type.")
# Prevent yielding partial stop sequence
if (not partially_stopped) and output and output[-1] != "�":
delta_text = output[len(previous_text):]
previous_text = output
yield {
"id": completion_id,
"object": "text_completion",
"created": created,
"model": model_name,
"delta": delta_text,
"text": output,
"logprobs": ret_logprobs,
"finish_reason": finish_reason,
"usage": {
"prompt_tokens": input_echo_len,
"completion_tokens": i,
"total_tokens": input_echo_len + i,
},
}
if stopped:
break
yield {
"id": completion_id,
"object": "text_completion",
"created": created,
"model": model_name,
"delta": "",
"text": output,
"logprobs": ret_logprobs,
"finish_reason": "stop",
"usage": {
"prompt_tokens": input_echo_len,
"completion_tokens": i,
"total_tokens": input_echo_len + i,
},
}
# Clean
del past_key_values, out
gc.collect()
torch.cuda.empty_cache()
@torch.inference_mode()
def generate_stream_v2(
model: PreTrainedModel,
tokenizer: PreTrainedTokenizer,
params: Dict[str, Any],
):
input_ids = params.get("inputs")
functions = params.get("functions")
model_name = params.get("model", "llm")
temperature = float(params.get("temperature", 1.0))
repetition_penalty = float(params.get("repetition_penalty", 1.0))
top_p = float(params.get("top_p", 1.0))
top_k = int(params.get("top_k", 40))
max_new_tokens = int(params.get("max_tokens", 256))
stop_token_ids = params.get("stop_token_ids") or []
if tokenizer.eos_token_id not in stop_token_ids:
stop_token_ids.append(tokenizer.eos_token_id)
stop_strings = params.get("stop", [])
input_echo_len = len(input_ids)
device = model.device
generation_kwargs = dict(
input_ids=torch.tensor([input_ids], device=device),
do_sample=True,
temperature=temperature,
top_p=top_p,
top_k=top_k,
max_new_tokens=max_new_tokens,
repetition_penalty=repetition_penalty,
pad_token_id=tokenizer.pad_token_id,
)
if temperature <= 1e-5:
generation_kwargs["do_sample"] = False
generation_kwargs.pop("top_k")
streamer = TextIteratorStreamer(
tokenizer, timeout=60.0, skip_prompt=True, skip_special_tokens=True
)
generation_kwargs["streamer"] = streamer
if "GenerationMixin" not in str(model.generate.__func__):
model.generate = MethodType(PreTrainedModel.generate, model)
thread = Thread(target=model.generate, kwargs=generation_kwargs)
thread.start()
generated_text, func_call_found = "", False
completion_id: str = f"cmpl-{str(uuid.uuid4())}"
created: int = int(time.time())
previous_text = ""
for i, new_text in enumerate(streamer):
generated_text += new_text
if functions:
_, func_call_found = apply_stopping_strings(generated_text, ["Observation:"])
generated_text, stop_found = apply_stopping_strings(generated_text, stop_strings)
if generated_text and generated_text[-1] != "�":
delta_text = generated_text[len(previous_text):]
previous_text = generated_text
yield {
"id": completion_id,
"object": "text_completion",
"created": created,
"model": model_name,
"delta": delta_text,
"text": generated_text,
"logprobs": None,
"finish_reason": "function_call" if func_call_found else None,
"usage": {
"prompt_tokens": input_echo_len,
"completion_tokens": i,
"total_tokens": input_echo_len + i,
},
}
if stop_found:
break
yield {
"id": completion_id,
"object": "text_completion",
"created": created,
"model": model_name,
"delta": "",
"text": generated_text,
"logprobs": None,
"finish_reason": "stop",
"usage": {
"prompt_tokens": input_echo_len,
"completion_tokens": i,
"total_tokens": input_echo_len + i,
},
}
|