What is the best way to ensure that my model is taking advantage of BF16 and mixed mode training

Users should refer to the TensorFlow User Guide section of “TensorFlow Mixed Precision Training on Gaudi” for details for setting the proper flags at runtime, one can also confirm if the ops are placed on HPU using the specified data type. The Optimization Guide also has additional information on how to confirm that OPs are running on Gaudi instead of the Host CPU

1 Like