hpcaitech / FastFold

Optimizing AlphaFold Training and Inference on GPU Clusters
Apache License 2.0
556 stars 84 forks source link

Did fastfold support bfloat16 inference? #165

Open semal opened 1 year ago

semal commented 1 year ago
image

fastfold/utils/geometry/vector.py seems got an unexpected data type.

Gy-Lu commented 1 year ago

Thanks for your question, this feature is still under development.

semal commented 1 year ago
image

more information about the data type change.

semal commented 1 year ago

This error comes from xx = 1 - 2* (y ** 2 + z** 2), dtype of 1 should give a specific datatype. Another error comes from w *= inv_normwhich is a inplace operation, maybe should change to w = w * inv_norm

Ian-pku commented 7 months ago

hi! have you achieved inference with bf16? which part of the source code did you modified?

Lix1993 commented 4 days ago

Readme says: We can inference a sequence of length 10000 in bf16 with 61GB memory on a Nvidia A100(80GB).

How can I infer with bf16?