v0.102.0

Full Changelog: v0.2.1...v0.102.0

New Features:

• add learning rate scheduler
  Add learning rate scheduler, used to adjust LR based on training steps.
  Multi implements of LR scheduler can be found at https://github.com/tensorflow/tensorflow/blob/v2.2.3/tensorflow/python/keras/optimizer_v2/learning_rate_schedule.py

    # if learning rate is set as float, LR will be constantly
    dense_opt = tn.core.Adam(learning_rate=0.001, beta1=0.9, beta2=0.999, epsilon=1e-8)

    # if learning rate is set as scheduler class, LR will be changed based on steps
    lr_schedule = tf.keras.optimizers.schedules.ExponentialDecay(
        0.001,
        decay_steps=20000,
        decay_rate=0.95,
        staircase=True)
    dense_opt = tn.core.Adam(learning_rate=lr_schedule, beta1=0.99, beta2=0.9999, epsilon=1e-8)    

Enhancements:

1. add cmake build with pixi
2. upgrade brpc version to 1.5.0
3. add python/shell/c format check
4. add otel trace tracking for model.fit() and model.predict()

Bug fix

When enable tensorboard (usually at rank 0), embedding save will be triggered at on_epoch_end(). It will affect other ranks sending rpc call to fetch sparse from current rank, may cause pull timeout and core dump. Add tn.core.barrier() to on_epoch_end() to avoid unaligned training progress Commit 128ab2a