You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
作者好,我在复现RRHF时碰到变量类型报错:
我配置fsdp_config进行分布式训练,当我使用--bf16混合精度时,报错:
return torch.embedding(weight, input, padding_idx, scale_grad_by_freq, sparse)
RuntimeError: Expected tensor for argument #1 'indices' to have one of the following scalar types: Long, Int; but got CUDABFloat16Type instead (while checking arguments for embedding)
如果不使用bf16和tf32,报错:
return torch.embedding(weight, input, padding_idx, scale_grad_by_freq, sparse)
RuntimeError: Expected tensor for argument #1 'indices' to have one of the following scalar types: Long, Int; but got torch.cuda.FloatTensor instead (while checking arguments for embedding)
我的fsdp_config配置如图
使用的模型是llama3-8b,或者是tokenizer需要重新配置一下吗?
The text was updated successfully, but these errors were encountered:
作者好,我在复现RRHF时碰到变量类型报错:
我配置fsdp_config进行分布式训练,当我使用--bf16混合精度时,报错:
return torch.embedding(weight, input, padding_idx, scale_grad_by_freq, sparse)
RuntimeError: Expected tensor for argument #1 'indices' to have one of the following scalar types: Long, Int; but got CUDABFloat16Type instead (while checking arguments for embedding)
如果不使用bf16和tf32,报错:
return torch.embedding(weight, input, padding_idx, scale_grad_by_freq, sparse)
RuntimeError: Expected tensor for argument #1 'indices' to have one of the following scalar types: Long, Int; but got torch.cuda.FloatTensor instead (while checking arguments for embedding)
我的fsdp_config配置如图
使用的模型是llama3-8b,或者是tokenizer需要重新配置一下吗?
The text was updated successfully, but these errors were encountered: