juancopi81's picture
Add t5x and mt3 models
b100e1c
raw
history blame
593 Bytes
# Continue a Mixture of Experts pre-training run.
#
# See t5x/contrib/moe/configs/runs/pretrain.gin for instructions.
#
# You must also include a binding for MODEL.
#
# Required to be set:
#
# - NUM_EXPERTS
# - NUM_MODEL_PARTITIONS (1 if no model parallelism)
# - MIXTURE_OR_TASK_NAME
# - TASK_FEATURE_LENGTHS
# - TRAIN_STEPS
# - INITIAL_CHECKPOINT_PATH
# - MODEL_DIR
from __gin__ import dynamic_registration
from t5x import utils
include 't5x/contrib/moe/configs/runs/pretrain.gin'
utils.RestoreCheckpointConfig:
mode = 'specific'
path = %INITIAL_CHECKPOINT_PATH
dtype = 'float32'