|
import os.path |
|
|
|
import torch |
|
from header import * |
|
|
|
|
|
class DeepSpeedAgent: |
|
|
|
def __init__(self, model, args): |
|
super(DeepSpeedAgent, self).__init__() |
|
self.args = args |
|
self.model = model |
|
|
|
self.print_model_parameters() |
|
self.writer = SummaryWriter(args['log_path']) |
|
|
|
self.load_parameters(self.args['save_path'], self.args['stage']) |
|
|
|
|
|
ds_params = json.load(open(self.args['ds_config_path'])) |
|
ds_params['scheduler']['params']['total_num_steps'] = self.args['total_steps'] |
|
ds_params['scheduler']['params']['warmup_num_steps'] = max(10, int( |
|
self.args['total_steps'] * self.args['warmup_rate'])) |
|
self.ds_engine, self.optimizer, _, _ = deepspeed.initialize( |
|
model=self.model, |
|
model_parameters=self.model.parameters(), |
|
config_params=ds_params, |
|
dist_init_required=True, |
|
args=types.SimpleNamespace(**args) |
|
) |
|
|
|
@torch.no_grad() |
|
def predict(self): |
|
self.ds_engine.module.eval() |
|
output = self.ds_engine.generate(self.args) |
|
return output |
|
|
|
def train_model(self, batch, current_step=0, pbar=None): |
|
self.ds_engine.module.train() |
|
|
|
loss, mle_acc, mse_loss = self.ds_engine(batch) |
|
|
|
self.writer.add_scalar('loss', loss, current_step) |
|
self.writer.add_scalar('mle_acc', mle_acc, current_step) |
|
|
|
|
|
|
|
|
|
if isinstance(mse_loss, torch.Tensor): |
|
self.writer.add_scalar('mse_loss', mse_loss, current_step) |
|
else: |
|
pass |
|
|
|
|
|
self.ds_engine.backward(loss) |
|
self.ds_engine.step() |
|
|
|
pbar.set_description(f'[!] loss: {round(loss.item(), 4)}; token_acc: {round(mle_acc * 100, 2)}') |
|
pbar.update(1) |
|
if self.args['local_rank'] == 0 and self.args['log_path'] and current_step % self.args['logging_step'] == 0: |
|
elapsed = pbar.format_dict['elapsed'] |
|
rate = pbar.format_dict['rate'] |
|
remaining = (pbar.total - pbar.n) / rate if rate and pbar.total else 0 |
|
remaining = str(datetime.timedelta(seconds=remaining)) |
|
logging.info( |
|
f'[!] progress: {round(pbar.n / pbar.total, 5)}; remaining time: {remaining}; loss: {round(loss.item(), 4)}; token_acc: {round(mle_acc * 100, 2)}') |
|
|
|
mle_acc *= 100 |
|
return mle_acc |
|
|
|
def save_model(self, path, current_step): |
|
""" |
|
this function also save the trainable parameters and specific name parameters |
|
""" |
|
param_grad_dic = { |
|
k: v.requires_grad for (k, v) in self.ds_engine.module.named_parameters() |
|
} |
|
state_dict = self.ds_engine.module.state_dict() |
|
checkpoint = OrderedDict() |
|
for k, v in self.ds_engine.module.named_parameters(): |
|
if v.requires_grad: |
|
checkpoint[k] = v |
|
if 'gen_text_hidden_fcs' in k: |
|
checkpoint[k] = v |
|
if 'gen_text_hidden_fcs_video' in k: |
|
checkpoint[k] = v |
|
if 'gen_text_hidden_fcs_audio' in k: |
|
checkpoint[k] = v |
|
if 'llama_proj' in k: |
|
checkpoint[k] = v |
|
torch.save(checkpoint, f'{path}/pytorch_model.pt') |
|
|
|
self.model.llama_tokenizer.save_pretrained(path) |
|
|
|
self.model.llama_model.config.save_pretrained(path) |
|
print(f'[!] save model into {path}') |
|
|
|
def print_model_parameters(self, use_4bit=False): |
|
""" |
|
Prints the number of trainable parameters in the model. |
|
""" |
|
trainable_params = 0 |
|
all_param = 0 |
|
lora = 0 |
|
image = 0 |
|
video = 0 |
|
audio = 0 |
|
linear = 0 |
|
llama = 0 |
|
imagebind = 0 |
|
for name, param in self.model.named_parameters(): |
|
num_params = param.numel() |
|
|
|
if num_params == 0 and hasattr(param, "ds_numel"): |
|
num_params = param.ds_numel |
|
|
|
if 'lora' in name: |
|
lora += num_params |
|
elif 'gen_text_hidden_fcs_video' in name: |
|
video += num_params |
|
elif 'gen_text_hidden_fcs_audio' in name: |
|
audio += num_params |
|
elif 'gen_text_hidden_fcs' in name: |
|
image += num_params |
|
elif 'llama_proj' in name: |
|
linear += num_params |
|
elif 'llama_model' in name: |
|
llama += num_params |
|
elif 'visual_encoder' in name: |
|
imagebind += num_params |
|
else: |
|
pass |
|
|
|
all_param += num_params |
|
if param.requires_grad: |
|
trainable_params += num_params |
|
if use_4bit: |
|
trainable_params /= 2 |
|
print( |
|
f"all params: {all_param:,d} || trainable params: {trainable_params:,d} || trainable%: {100 * trainable_params / all_param}" |
|
) |
|
print(f'lora params: {lora:,d} || video params: {video:,d} || audio params: {audio:,d} || image params: {image:,d}') |
|
print(f'linear params: {linear:,d} || imagebind params: {imagebind:,d} || llama params: {llama:,d}') |
|
|
|
def load_parameters(self, path, stage=3): |
|
if os.path.exists(os.path.join(path, 'pytorch_model.pt')): |
|
print('loading parameters from {}'.format(self.args['save_path'])) |
|
delta_ckpt = torch.load(f'{path}/pytorch_model.pt', map_location=torch.device('cuda')) |
|
checkpoint = OrderedDict() |
|
if stage == 3: |
|
for k, v in delta_ckpt.items(): |
|
if 'llama_model.model.embed_tokens.weight' in k: |
|
checkpoint['llama_model.base_model.model.model.embed_tokens.weight'] = v |
|
elif 'llama_model.lm_head.weight' in k: |
|
checkpoint['llama_model.base_model.model.lm_head.weight'] = v |
|
else: |
|
checkpoint[k] = v |
|
else: |
|
checkpoint = delta_ckpt |
|
self.model.load_state_dict(checkpoint, strict=False) |
|
|
|
|