Spaces:
Running
Running
File size: 18,167 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 |
import os
import sys
from typing import List, Optional, Any, Dict, Tuple
import torch
from loguru import logger
from peft import PeftModel
from tqdm import tqdm
from transformers import (
AutoModel,
AutoConfig,
AutoTokenizer,
AutoModelForCausalLM,
BitsAndBytesConfig,
PreTrainedTokenizer,
PreTrainedModel,
)
from transformers.utils.versions import require_version
if sys.version_info >= (3, 9):
from functools import cache
else:
from functools import lru_cache as cache
class BaseModelAdapter:
""" The base and default model adapter. """
model_names = []
def match(self, model_name) -> bool:
"""
Check if the given model name matches any of the predefined model names.
Args:
model_name (str): The model name to check.
Returns:
bool: True if the model name matches any of the predefined model names, False otherwise.
"""
return any(m in model_name for m in self.model_names) if self.model_names else True
def load_model(
self,
model_name_or_path: Optional[str] = None,
adapter_model: Optional[str] = None,
**kwargs: Any,
) -> Tuple[PreTrainedModel, PreTrainedTokenizer]:
"""
Load a model and tokenizer based on the provided model name or path.
Args:
model_name_or_path (str, optional): The name or path of the model. Defaults to None.
adapter_model (str, optional): The adapter model to load the tokenizer from. Defaults to None.
**kwargs: Additional keyword arguments.
Returns:
Tuple[PreTrainedModel, PreTrainedTokenizer]: A tuple containing the loaded model and tokenizer.
"""
model_name_or_path = model_name_or_path or self.default_model_name_or_path
tokenizer_kwargs = {"trust_remote_code": True, "use_fast": False}
tokenizer_kwargs.update(self.tokenizer_kwargs)
# load a tokenizer from adapter model if it exists.
if adapter_model is not None:
try:
tokenizer = self.tokenizer_class.from_pretrained(
adapter_model, **tokenizer_kwargs,
)
except OSError:
tokenizer = self.tokenizer_class.from_pretrained(
model_name_or_path, **tokenizer_kwargs,
)
else:
tokenizer = self.tokenizer_class.from_pretrained(
model_name_or_path, **tokenizer_kwargs,
)
config_kwargs = self.model_kwargs
device = kwargs.get("device", "cuda")
num_gpus = kwargs.get("num_gpus", 1)
dtype = kwargs.get("dtype", "half")
if device == "cuda":
if "torch_dtype" not in config_kwargs:
if dtype == "half":
config_kwargs["torch_dtype"] = torch.float16
elif dtype == "bfloat16":
config_kwargs["torch_dtype"] = torch.bfloat16
elif dtype == "float32":
config_kwargs["torch_dtype"] = torch.float32
if num_gpus != 1:
config_kwargs["device_map"] = "auto"
# model_kwargs["device_map"] = "sequential" # This is important for not the same VRAM sizes
# Quantization configurations (using bitsandbytes library).
if kwargs.get("load_in_8bit", False):
require_version("bitsandbytes>=0.37.0", "To fix: pip install bitsandbytes>=0.37.0")
config_kwargs["load_in_8bit"] = True
config_kwargs["quantization_config"] = BitsAndBytesConfig(
load_in_8bit=True,
llm_int8_threshold=6.0,
)
config_kwargs["device_map"] = "auto" if device == "cuda" else None
logger.info("Quantizing model to 8 bit.")
elif kwargs.get("load_in_4bit", False):
require_version("bitsandbytes>=0.39.0", "To fix: pip install bitsandbytes>=0.39.0")
require_version("peft>=0.4.0.dev0", "To fix: pip install git+https://github.com/huggingface/peft.git")
config_kwargs["load_in_4bit"] = True
config_kwargs["quantization_config"] = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_compute_dtype=torch.float16,
bnb_4bit_use_double_quant=True,
bnb_4bit_quant_type="nf4",
)
config_kwargs["device_map"] = "auto" if device == "cuda" else None
logger.info("Quantizing model to 4 bit.")
if kwargs.get("device_map", None) == "auto":
config_kwargs["device_map"] = "auto"
config = AutoConfig.from_pretrained(model_name_or_path, trust_remote_code=True)
# Fix config (for Qwen)
if hasattr(config, "fp16") and hasattr(config, "bf16"):
setattr(config, "fp16", dtype == "half")
setattr(config, "bf16", dtype == "bfloat16")
config_kwargs.pop("torch_dtype", None)
if kwargs.get("using_ptuning_v2", False) and adapter_model:
config.pre_seq_len = kwargs.get("pre_seq_len", 128)
# Load and prepare pretrained models (without valuehead).
model = self.model_class.from_pretrained(
model_name_or_path,
config=config,
trust_remote_code=True,
**config_kwargs
)
if device == "cpu":
model = model.float()
# post process for special tokens
tokenizer = self.post_tokenizer(tokenizer)
is_chatglm = "chatglm" in str(type(model))
if adapter_model is not None:
model = self.load_adapter_model(model, tokenizer, adapter_model, is_chatglm, config_kwargs, **kwargs)
if is_chatglm or "baichuan" in str(type(model)) or "xverse" in str(type(model)):
quantize = kwargs.get("quantize", None)
if quantize and quantize != 16:
logger.info(f"Quantizing model to {quantize} bit.")
model = model.quantize(quantize)
if device == "cuda" and num_gpus == 1 and "device_map" not in config_kwargs:
model.to(device)
# inference mode
model.eval()
return model, tokenizer
def load_lora_model(
self, model: PreTrainedModel, adapter_model: str, model_kwargs: Dict,
) -> PeftModel:
"""
Load a LoRA model.
This function loads a LoRA model using the specified pretrained model and adapter model.
Args:
model (PreTrainedModel): The base pretrained model.
adapter_model (str): The name or path of the adapter model.
model_kwargs (dict): Additional keyword arguments for the model.
Returns:
PeftModel: The loaded LoRA model.
"""
return PeftModel.from_pretrained(
model,
adapter_model,
torch_dtype=model_kwargs.get("torch_dtype", torch.float16),
)
def load_adapter_model(
self,
model: PreTrainedModel,
tokenizer: PreTrainedTokenizer,
adapter_model: str,
is_chatglm: bool,
model_kwargs: Dict,
**kwargs: Any,
) -> PreTrainedModel:
using_ptuning_v2 = kwargs.get("using_ptuning_v2", False)
resize_embeddings = kwargs.get("resize_embeddings", False)
if adapter_model and resize_embeddings and not is_chatglm:
model_vocab_size = model.get_input_embeddings().weight.size(0)
tokenzier_vocab_size = len(tokenizer)
logger.info(f"Vocab of the base model: {model_vocab_size}")
logger.info(f"Vocab of the tokenizer: {tokenzier_vocab_size}")
if model_vocab_size != tokenzier_vocab_size:
assert tokenzier_vocab_size > model_vocab_size
logger.info("Resize model embeddings to fit tokenizer")
model.resize_token_embeddings(tokenzier_vocab_size)
if using_ptuning_v2:
prefix_state_dict = torch.load(os.path.join(adapter_model, "pytorch_model.bin"))
new_prefix_state_dict = {
k[len("transformer.prefix_encoder."):]: v
for k, v in prefix_state_dict.items()
if k.startswith("transformer.prefix_encoder.")
}
model.transformer.prefix_encoder.load_state_dict(new_prefix_state_dict)
model.transformer.prefix_encoder.float()
else:
model = self.load_lora_model(model, adapter_model, model_kwargs)
return model
def post_tokenizer(self, tokenizer) -> PreTrainedTokenizer:
return tokenizer
@property
def model_class(self):
return AutoModelForCausalLM
@property
def model_kwargs(self):
return {}
@property
def tokenizer_class(self):
return AutoTokenizer
@property
def tokenizer_kwargs(self):
return {}
@property
def default_model_name_or_path(self):
return "zpn/llama-7b"
# A global registry for all model adapters
model_adapters: List[BaseModelAdapter] = []
def register_model_adapter(cls):
""" Register a model adapter. """
model_adapters.append(cls())
@cache
def get_model_adapter(model_name: str) -> BaseModelAdapter:
"""
Get a model adapter for a given model name.
Args:
model_name (str): The name of the model.
Returns:
ModelAdapter: The model adapter that matches the given model name.
"""
for adapter in model_adapters:
if adapter.match(model_name):
return adapter
raise ValueError(f"No valid model adapter for {model_name}")
def load_model(
model_name: str,
model_name_or_path: Optional[str] = None,
adapter_model: Optional[str] = None,
quantize: Optional[int] = 16,
device: Optional[str] = "cuda",
load_in_8bit: Optional[bool] = False,
**kwargs: Any,
) -> Tuple[PreTrainedModel, PreTrainedTokenizer]:
"""
Load a pre-trained model and tokenizer.
Args:
model_name (str): The name of the model.
model_name_or_path (Optional[str], optional): The path or name of the pre-trained model. Defaults to None.
adapter_model (Optional[str], optional): The name of the adapter model. Defaults to None.
quantize (Optional[int], optional): The quantization level. Defaults to 16.
device (Optional[str], optional): The device to load the model on. Defaults to "cuda".
load_in_8bit (Optional[bool], optional): Whether to load the model in 8-bit mode. Defaults to False.
**kwargs (Any): Additional keyword arguments.
Returns:
Tuple[PreTrainedModel, PreTrainedTokenizer]: A tuple containing the loaded model and tokenizer.
"""
model_name = model_name.lower()
if "tiger" in model_name:
def skip(*args, **kwargs):
pass
torch.nn.init.kaiming_uniform_ = skip
torch.nn.init.uniform_ = skip
torch.nn.init.normal_ = skip
# get model adapter
adapter = get_model_adapter(model_name)
model, tokenizer = adapter.load_model(
model_name_or_path,
adapter_model,
device=device,
quantize=quantize,
load_in_8bit=load_in_8bit,
**kwargs
)
return model, tokenizer
class ChatglmModelAdapter(BaseModelAdapter):
""" https://github.com/THUDM/ChatGLM-6B """
model_names = ["chatglm"]
@property
def model_class(self):
return AutoModel
@property
def default_model_name_or_path(self):
return "THUDM/chatglm2-6b"
class Chatglm3ModelAdapter(ChatglmModelAdapter):
""" https://github.com/THUDM/ChatGLM-6B """
model_names = ["chatglm3"]
@property
def tokenizer_kwargs(self):
return {"encode_special_tokens": True}
@property
def default_model_name_or_path(self):
return "THUDM/chatglm3-6b"
class LlamaModelAdapter(BaseModelAdapter):
""" https://github.com/project-baize/baize-chatbot """
model_names = ["alpaca", "baize", "openbuddy-llama", "ziya-llama", "guanaco", "llama2"]
def post_tokenizer(self, tokenizer):
tokenizer.bos_token = "<s>"
tokenizer.eos_token = "</s>"
tokenizer.unk_token = "<unk>"
return tokenizer
@property
def model_kwargs(self):
return {"low_cpu_mem_usage": True}
class MossModelAdapter(BaseModelAdapter):
""" https://github.com/OpenLMLab/MOSS """
model_names = ["moss"]
@property
def default_model_name_or_path(self):
return "fnlp/moss-moon-003-sft-int4"
class PhoenixModelAdapter(BaseModelAdapter):
""" https://github.com/FreedomIntelligence/LLMZoo """
model_names = ["phoenix"]
@property
def model_kwargs(self):
return {"low_cpu_mem_usage": True}
@property
def tokenizer_kwargs(self):
return {"use_fast": True}
@property
def default_model_name_or_path(self):
return "FreedomIntelligence/phoenix-inst-chat-7b"
class FireflyModelAdapter(BaseModelAdapter):
""" https://github.com/yangjianxin1/Firefly """
model_names = ["firefly"]
@property
def model_kwargs(self):
return {"torch_dtype": torch.float32}
@property
def tokenizer_kwargs(self):
return {"use_fast": True}
@property
def default_model_name_or_path(self):
return "YeungNLP/firefly-2b6"
class YuLanChatModelAdapter(BaseModelAdapter):
""" https://github.com/RUC-GSAI/YuLan-Chat """
model_names = ["yulan"]
def post_tokenizer(self, tokenizer):
tokenizer.bos_token = "<s>"
tokenizer.eos_token = "</s>"
tokenizer.unk_token = "<unk>"
return tokenizer
@property
def model_kwargs(self):
return {"low_cpu_mem_usage": True}
def load_adapter_model(self, model, tokenizer, adapter_model, is_chatglm, model_kwargs, **kwargs):
adapter_model = AutoModelForCausalLM.from_pretrained(
adapter_model, torch_dtype=torch.float16, low_cpu_mem_usage=True
)
if model.model.embed_tokens.weight.size(0) + 1 == adapter_model.model.embed_tokens.weight.size(0):
model.resize_token_embeddings(len(tokenizer))
model.model.embed_tokens.weight.data[-1, :] = 0
logger.info("Applying the delta")
for name, param in tqdm(model.state_dict().items(), desc="Applying delta"):
assert name in model.state_dict()
param.data += model.state_dict()[name]
return model
class TigerBotModelAdapter(BaseModelAdapter):
""" https://github.com/TigerResearch/TigerBot """
model_names = ["tiger"]
@property
def tokenizer_kwargs(self):
return {"use_fast": True}
@property
def default_model_name_or_path(self):
return "TigerResearch/tigerbot-7b-sft"
class OpenBuddyFalconModelAdapter(BaseModelAdapter):
""" https://github.com/OpenBuddy/OpenBuddy """
model_names = ["openbuddy-falcon"]
@property
def default_model_name_or_path(self):
return "OpenBuddy/openbuddy-falcon-7b-v5-fp16"
class AnimaModelAdapter(LlamaModelAdapter):
model_names = ["anima"]
def load_lora_model(self, model, adapter_model, model_kwargs):
return PeftModel.from_pretrained(model, adapter_model)
class BaiChuanModelAdapter(BaseModelAdapter):
""" https://github.com/baichuan-inc/Baichuan-13B """
model_names = ["baichuan"]
def load_lora_model(self, model, adapter_model, model_kwargs):
return PeftModel.from_pretrained(model, adapter_model)
@property
def default_model_name_or_path(self):
return "baichuan-inc/Baichuan-13B-Chat"
class InternLMModelAdapter(BaseModelAdapter):
""" https://github.com/InternLM/InternLM """
model_names = ["internlm"]
@property
def default_model_name_or_path(self):
return "internlm/internlm-chat-7b"
class StarCodeModelAdapter(BaseModelAdapter):
""" https://github.com/bigcode-project/starcoder """
model_names = ["starcode", "starchat"]
@property
def tokenizer_kwargs(self):
return {}
@property
def default_model_name_or_path(self):
return "HuggingFaceH4/starchat-beta"
class AquilaModelAdapter(BaseModelAdapter):
""" https://github.com/FlagAI-Open/FlagAI """
model_names = ["aquila"]
@property
def default_model_name_or_path(self):
return "BAAI/AquilaChat-7B"
class QwenModelAdapter(BaseModelAdapter):
""" https://github.com/QwenLM/Qwen-7B """
model_names = ["qwen"]
@property
def default_model_name_or_path(self):
return "Qwen/Qwen-7B-Chat"
class XverseModelAdapter(BaseModelAdapter):
""" https://github.com/xverse-ai/XVERSE-13B """
model_names = ["xverse"]
@property
def default_model_name_or_path(self):
return "xverse/XVERSE-13B-Chat"
class CodeLlamaModelAdapter(LlamaModelAdapter):
""" https://github.com/project-baize/baize-chatbot """
model_names = ["code-llama"]
@property
def tokenizer_class(self):
require_version("transformers>=4.33.1", "To fix: pip install transformers>=4.33.1")
from transformers import CodeLlamaTokenizer
return CodeLlamaTokenizer
@property
def default_model_name_or_path(self):
return "codellama/CodeLlama-7b-Instruct-hf"
register_model_adapter(ChatglmModelAdapter)
register_model_adapter(Chatglm3ModelAdapter)
register_model_adapter(LlamaModelAdapter)
register_model_adapter(MossModelAdapter)
register_model_adapter(PhoenixModelAdapter)
register_model_adapter(FireflyModelAdapter)
register_model_adapter(YuLanChatModelAdapter)
register_model_adapter(TigerBotModelAdapter)
register_model_adapter(OpenBuddyFalconModelAdapter)
register_model_adapter(AnimaModelAdapter)
register_model_adapter(BaiChuanModelAdapter)
register_model_adapter(InternLMModelAdapter)
register_model_adapter(AquilaModelAdapter)
register_model_adapter(QwenModelAdapter)
register_model_adapter(XverseModelAdapter)
register_model_adapter(CodeLlamaModelAdapter)
# After all adapters, try the default base adapter.
register_model_adapter(BaseModelAdapter)
|