dwzhu

/

e5rope-base

like 16

Hi, first thanks for putting out this model!

I'd like to use it for my research but am having some trouble getting the model running in the base example script in the README.md. I included the error below. Notably, I'm having trouble:

1. changing the attn_bias to the proper type
2. changing the model to torch.float16 without running into other errors

I am using the following packages:

pytorch                   2.4.0           py3.10_cuda12.4_cudnn9.1.0_0    pytorch
pytorch-cuda              12.4                 hc786d27_6    pytorch
xformers                  0.0.27.post2             pypi_0    pypi

Please let me know if I am missing any other documentation anywhere as well!

Original error:

NotImplementedError: No operator found for `memory_efficient_attention_forward` with inputs:
     query       : shape=(4, 75, 12, 64) (torch.float32)
     key         : shape=(4, 75, 12, 64) (torch.float32)
     value       : shape=(4, 75, 12, 64) (torch.float32)
     attn_bias   : <class 'torch.Tensor'>
     p           : 0
`flshattF@2.5.6-pt` is not supported because:
    device=cpu (supported: {'cuda'})
    dtype=torch.float32 (supported: {torch.bfloat16, torch.float16})
    attn_bias type is <class 'torch.Tensor'>
`cutlassF-pt` is not supported because:
    device=cpu (supported: {'cuda'})
    attn_bias.stride(-2) % 4 != 0 (attn_bias.stride() = (75, 0, 0, 1))
    HINT: To use an `attn_bias` with a sequence length that is not a multiple of 8, you need to ensure memory is aligned by slicing a bigger tensor. Example: use `attn_bias = torch.zeros([1, 1, 5, 8])[:,:,:,:5]` instead of `torch.zeros([1, 1, 5, 5])`
`smallkF` is not supported because:
    max(query.shape[-1] != value.shape[-1]) > 32
    device=cpu (supported: {'cuda'})
    unsupported embed per head: 64