Spaces:
Running
on
Zero
Running
on
Zero
File size: 4,136 Bytes
1938217 |
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 |
import os
import warnings
import shutil
from transformers import AutoTokenizer, AutoModelForCausalLM, AutoConfig, BitsAndBytesConfig
import torch
from ola.model import *
from ola.model.speech_encoder.builder import build_speech_encoder
def load_pretrained_model(model_path, model_base, is_lora=False, s2s=False, load_8bit=False, load_4bit=False, device="cuda", use_flash_attn=False, **kwargs):
if load_8bit:
kwargs['load_in_8bit'] = True
elif load_4bit:
kwargs['load_in_4bit'] = True
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'
)
else:
kwargs['torch_dtype'] = torch.bfloat16
if use_flash_attn:
kwargs['attn_implementation'] = 'flash_attention_2'
model_cls = OlaQwenForCausalLM
# Load OmniSpeech model
if is_lora:
assert model_base is not None, "model_base is required for LoRA models."
from ola.model.language_model.ola_qwen import OlaConfigQwen
lora_cfg_pretrained = OlaConfigQwen.from_pretrained(model_path)
tokenizer = AutoTokenizer.from_pretrained(model_base, use_fast=False)
print('Loading OmniSpeech from base model...')
model = model_cls.from_pretrained(model_base, low_cpu_mem_usage=False, config=lora_cfg_pretrained, **kwargs)
print('Loading additional OmniSpeech weights...')
if os.path.exists(os.path.join(model_path, 'non_lora_trainables.bin')):
non_lora_trainables = torch.load(os.path.join(model_path, 'non_lora_trainables.bin'), map_location='cpu')
non_lora_trainables = {(k[11:] if k.startswith('base_model.') else k): v for k, v in non_lora_trainables.items()}
if any(k.startswith('model.model.') for k in non_lora_trainables):
non_lora_trainables = {(k[6:] if k.startswith('model.') else k): v for k, v in non_lora_trainables.items()}
model.load_state_dict(non_lora_trainables, strict=False)
from peft import PeftModel
print('Loading LoRA weights...')
model = PeftModel.from_pretrained(model, model_path)
print('Merging LoRA weights...')
model = model.merge_and_unload()
print('Model is loaded...')
elif model_base is not None:
print('Loading OmniSpeech from base model...')
tokenizer = AutoTokenizer.from_pretrained(model_base, use_fast=False)
cfg_pretrained = AutoConfig.from_pretrained(model_path)
model = model_cls.from_pretrained(model_base, low_cpu_mem_usage=False, config=cfg_pretrained, **kwargs)
speech_projector_weights = torch.load(os.path.join(model_path, 'speech_projector.bin'), map_location='cpu')
speech_projector_weights = {k: v.to(torch.float16) for k, v in speech_projector_weights.items()}
model.load_state_dict(speech_projector_weights, strict=False)
model = model.to(device=device)
else:
tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=False)
model = model_cls.from_pretrained(
model_path,
low_cpu_mem_usage=False,
**kwargs
)
model = model.to(device=device)
model.get_model().speech_encoder = build_speech_encoder(model.config)
model.get_model().speech_encoder.to(device=device, dtype=torch.float16)
image_processor = None
model.resize_token_embeddings(len(tokenizer))
vision_tower = model.get_vision_tower()
print("Loading vision tower...")
if not vision_tower.is_loaded:
vision_tower.load_model(device_map=device)
if device != "auto":
vision_tower.to(device="cuda", dtype=torch.bfloat16)
else:
vision_tower.to(device="cuda:0", dtype=torch.bfloat16)
image_processor = vision_tower.image_processor
print("Loading vision tower succeeded.")
if hasattr(model.config, "max_sequence_length"):
context_len = model.config.max_sequence_length
else:
context_len = 16384
return tokenizer, model, image_processor, context_len
|