为什么连给的example也有bug？ about lightseq

[etoilestar]

为什么我安装的环境，编译时会有bug？你遇到过吗？

[Moran232]

编译我通过了，跑不了他给的example，这库不行， 别纠结了

[SysuJayce]

AssertionError: CUDA_HOME does not exist, unable to compile CUDA op(s)
有人遇到一样的问题吗？安装了pytorch==2.1.0 + cuda11，竟然说找不到CUDA……

跑的知乎上的demo

import torch
from lightseq.training.ops.pytorch.transformer_encoder_layer import LSTransformerEncoderLayer

def train(model, inputs, masks):
    inputs = inputs.to(device="cuda:0")
    masks = masks.to(device="cuda:0")
    model.to(device="cuda:0")
    model.train()
    opt = torch.optim.Adam(model.parameters(), lr=1e-3)

    for epoch in range(1000):
        opt.zero_grad()
        outputs = model(inputs, masks)
        loss = torch.square(outputs).mean()
        loss.backward()
        opt.step()
        if epoch % 200 == 0:
            print("epoch {:>3d}: loss = {:>5.3f}".format(epoch, loss))

if __name__ == "__main__":
    # 定义LightSeq配置
    config = LSTransformerEncoderLayer.get_config(
        max_batch_tokens=4096,
        max_seq_len=256,
        hidden_size=1024,
        intermediate_size=4096,
        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
    )

    # 随机生成输入
    bsz, sl = 10, 80
    inputs = torch.randn(bsz, sl, config.hidden_size)
    masks = torch.zeros(bsz, sl)

    # 定义LightSeq编码层并进行训练
    model = LSTransformerEncoderLayer(config)
    train(model, inputs, masks)

[Anychnn]

您的邮件已收到!谢谢