# Switch Transformer Base model.

#

# Based on the original Switch Transformer (https://arxiv.org/abs/2101.03961).

#

# Note that unlike the original Switch Transformer, this T5X version does not

# use any jitter noise in the router.

#

# Provides MODEL and NUM_EXPERTS.

from __gin__ import dynamic_registration

from flaxformer.architectures.moe import moe_architecture

from flaxformer.architectures.moe import moe_layers

from flaxformer.architectures.moe import routing

from flaxformer.components import dense

import seqio

from t5x import adafactor

ARCHITECTURE = %gin.REQUIRED

include 'flaxformer/t5x/configs/moe/models/tokens_choose_base.gin'

# Architecture overrides

MLP_DIM = 3072

# MoE overrides

NUM_EXPERTS = 128

# Replace every other MLP sublayer is an MoE sublayer.

NUM_ENCODER_SPARSE_LAYERS = 6

NUM_DECODER_SPARSE_LAYERS = 6

TRAIN_EXPERT_CAPACITY_FACTOR = 1.25

EVAL_EXPERT_CAPACITY_FACTOR = 2.

NUM_SELECTED_EXPERTS = 1 # Switch routing

AUX_LOSS_FACTOR = 0.01

ROUTER_Z_LOSS_FACTOR = 0.0

GROUP_SIZE = 8192

# Switch Transformer Base uses relu activations.

dense.MlpBlock.activations = ('relu',)

expert/dense.MlpBlock.activations = ('relu',)

# Switch Transformer Base re-uses the token embedder to compute output logits.

moe_architecture.SparseDecoder.output_logits_factory = None

# Switch Transformer doesn't use BPR in encoder (although most sparse encoders

# generally see a boost from it).

sparse_encoder/routing.TokensChooseMaskedRouter.batch_prioritized_routing = False