恕我直言，你们的模型训练都还不够快

pip install contiguous-params fairseq lightseq

import timefrom dataclasses import dataclassimport copy
import torchfrom fairseq.modules.transformer_layer import TransformerEncoderLayerfrom lightseq.training.ops.pytorch.transformer_encoder_layer import LSTransformerEncoderLayerfrom contiguous_params import ContiguousParams

def get_time():    '''CUDA同步并获取当前时间'''
    torch.cuda.synchronize(device="cuda:0")    return time.time()
def ls_config_to_fs_args(config):    '''将LightSeq的config转换为Fairseq的args'''
    @dataclass    class Args:        encoder_embed_dim: int        encoder_ffn_embed_dim: int        encoder_attention_heads: int        dropout: float        attention_dropout: float        activation_dropout: float        encoder_normalize_before: bool        args = Args(        config.hidden_size,        config.intermediate_size,        config.nhead,        config.hidden_dropout_ratio,        config.attn_prob_dropout_ratio,        config.activation_dropout_ratio,        config.pre_layer_norm    )    return args
def train(model, inputs, masks, contiguous=False):    '''训练过程'''
    model.to(device="cuda:0")    model.train()    if contiguous:        parameters = ContiguousParams(model.parameters())        opt = torch.optim.Adam(parameters.contiguous(), lr=1e-3)    else:        opt = torch.optim.Adam(model.parameters(), lr=1e-3)
    fw_time, bw_time, step_time = 0, 0, 0        for epoch in range(1000):        opt.zero_grad()
        start_time = get_time()        outputs = model(inputs, masks)        loss = torch.square(outputs).mean()        fw_time += get_time() - start_time                start_time = get_time()        loss.backward()        bw_time += get_time() - start_time
        start_time = get_time()        opt.step()        step_time += get_time() - start_time
        if epoch % 200 == 0:            print("epoch {:>3d}: loss = {:>5.3f}".format(epoch, loss))
    return fw_time, bw_time, step_time
if __name__ == "__main__":    # 定义LightSeq的config    config = LSTransformerEncoderLayer.get_config(        max_batch_tokens=4096,        max_seq_len=256,        hidden_size=128,        intermediate_size=512,        nhead=16,        attn_prob_dropout_ratio=0.1,        activation_dropout_ratio=0.1,        hidden_dropout_ratio=0.1,        pre_layer_norm=True,        fp16=False,        local_rank=0    )    # 将LightSeq的config转换为Fairseq的args    args = ls_config_to_fs_args(config)
    # 随机生成输入    bsz, sl = 50, 80    inputs = torch.randn(bsz, sl, config.hidden_size).to(device="cuda:0")    masks = torch.zeros(bsz, sl).to(device="cuda:0")
    # 定义LightSeq模型并训练    ls_model = LSTransformerEncoderLayer(config)    ls_fw_time, ls_bw_time, ls_step_time = train(ls_model, inputs, masks)    # 定义连续化参数的LightSeq模型并训练    config_cont = copy.deepcopy(config)    ls_model_cont = LSTransformerEncoderLayer(config_cont)    ls_c_fw_time, ls_c_bw_time, ls_c_step_time = train(ls_model_cont, inputs, masks, contiguous=True)
    inputs = inputs.transpose(0, 1)    masks = masks > 0.5    # 定义Fairseq模型并训练    fs_model = TransformerEncoderLayer(args)    fs_fw_time, fs_bw_time, fs_step_time = train(fs_model, inputs, masks)    # 定义连续化参数的Fairseq模型并训练    fs_model_cont = TransformerEncoderLayer(args)    fs_c_fw_time, fs_c_bw_time, fs_c_step_time = train(fs_model_cont, inputs, masks, contiguous=True)
    print("LightSeq time:         {:.3f}s, {:.3f}s, {:.3f}s".format(ls_fw_time, ls_bw_time, ls_step_time))    print("LightSeq (cont) time:  {:.3f}s, {:.3f}s, {:.3f}s".format(ls_c_fw_time, ls_c_bw_time, ls_c_step_time))    print("Fairseq time:          {:.3f}s, {:.3f}s, {:.3f}s".format(fs_fw_time, fs_bw_time, fs_step_time))    print("Fairseq (cont) time:   {:.3f}s, {:.3f}s, {:.3f}s".format(fs_c_fw_time, fs_c_bw_time, fs_c_step_time))

parameters = ContiguousParams(model.parameters())opt = torch.optim.Adam(parameters.contiguous(), lr=1e-3)

def make_params_contiguous(self):    index = 0    # 遍历所有的参数    for p in self._parameters:        # 计算参数p的大小        size = p.numel()        # 在连续参数块中的对应位置赋值参数p        self._param_buffer[index:index + size] = p.data.view(-1)        # 将参数p的数值和梯度都重新指向连续参数块和连续梯度块的对应位置        p.data = self._param_buffer[index:index + size].view(p.data.shape)        p.grad = self._grad_buffer[index:index + size].view(p.data.shape)        # 连续内存块位置偏移到下一个参数        index += size    # 连续参数块的梯度设置为连续梯度块    self._param_buffer.grad = self._grad_buffer