Bfloat16

[npiroozan]

Hello,

Does DeePMD support BFloat16?

Warm Regards and thank you.

[njzjz]

New types can be added here:

deepmd-kit/deepmd/common.py

Lines 41 to 45 in c393392

	PRECISION_DICT = {
	"default": GLOBAL_TF_FLOAT_PRECISION,
	"float16": tf.float16,
	"float32": tf.float32,
	"float64": tf.float64,

Contribution is welcome if you have found bfloat 16 faster. We do not have a hardware that supports bfloat 16.

[npiroozan]

Outstanding, thank you very much. To implement this with LAMMPS, is anything additional required? We can input data from LAMMPS in FP32.

[njzjz]

No, the graph will be processed by TensorFlow.

[npiroozan]

I'll let you know what type of results are observed for training, inference, and DP+LMP. It should be very significant indeed. Thank you very much.

[npiroozan]

Significant increase across training, inference, and DP+LMP. I utilized mixed precision for training.

However, I am curious if it is possible to use bfloat16 for the embedding net and fitting net as well?

[npiroozan]

Please ignore the above question. Resolved.