base_xl_lightning.yml

# Copyright 2023 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#      https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

# This sentinel is a reminder to choose a real run name.
run_name: ''

metrics_file: "" # for testing, local file that stores scalar metrics. If empty, no metrics are written.
# If true save metrics such as loss and TFLOPS to GCS in {base_output_directory}/{run_name}/metrics/
gcs_metrics: True

timing_metrics_file: "" # for testing, local file that stores function timing metrics such as state creation, compilation. If empty, no metrics are written.
write_timing_metrics: True 

# If true save config to GCS in {base_output_directory}/{run_name}/
save_config_to_gcs: False
log_period: 100

pretrained_model_name_or_path: 'stabilityai/stable-diffusion-xl-base-1.0'
unet_checkpoint: ''
revision: 'refs/pr/95'
# This will convert the weights to this dtype.
weights_dtype: 'bfloat16'
# This sets the layer's dtype in the model. Ex: nn.Dense(dtype=activations_dtype)
activations_dtype: 'bfloat16'

# matmul and conv precision from https://jax.readthedocs.io/en/latest/jax.lax.html#jax.lax.Precision
# Options are "DEFAULT", "HIGH", "HIGHEST"
# fp32 activations and fp32 weights with HIGHEST will provide the best precision
# at the cost of time.
precision: "DEFAULT"

# if False state is not jitted and instead replicate is called. This is good for debugging on single host
# It must be True for multi-host.
jit_initializers: True

# Set true to load weights from pytorch
from_pt: False
split_head_dim: True
attention: 'flash' # Supported attention: dot_product, flash
flash_block_sizes: {}
# GroupNorm groups
norm_num_groups: 32

# If train_new_unet, unet weights will be randomly initialized to train the unet from scratch
# else they will be loaded from pretrained_model_name_or_path
train_new_unet: False

# train text_encoder - Currently not supported for SDXL
train_text_encoder: False
text_encoder_learning_rate: 4.25e-6

# Override parameters from checkpoints's scheduler.
diffusion_scheduler_config: {
  _class_name: 'DDIMScheduler',
  # values are v_prediction or leave empty to use scheduler's default.
  prediction_type: 'epsilon',
  rescale_zero_terminal_snr: False,
  timestep_spacing: 'trailing'
}

# Hardware
hardware: 'tpu' # Supported hardware types are 'tpu', 'gpu'
skip_jax_distributed_system: False
# Output directory
# Create a GCS bucket, e.g. my-maxtext-outputs and set this to "gs://my-maxtext-outputs/"
base_output_directory: ""

# Parallelism
mesh_axes: ['data', 'fsdp', 'tensor']

# batch : batch dimension of data and activations
# hidden :
# embed : attention qkv dense layer hidden dim named as embed
# heads : attention head dim = num_heads * head_dim
# length : attention sequence length
# temb_in : dense.shape[0] of resnet dense before conv
# out_c : dense.shape[1] of resnet dense before conv
# out_channels : conv.shape[-1] activation
# keep_1 : conv.shape[0] weight
# keep_2 : conv.shape[1] weight
# conv_in : conv.shape[2] weight
# conv_out : conv.shape[-1] weight
logical_axis_rules: [
                      ['batch', 'data'],
                      ['activation_batch', ['data','fsdp']],
                      ['activation_heads', 'tensor'],
                      ['activation_kv', 'tensor'],
                      ['embed','fsdp'],
                      ['heads', 'tensor'],
                      ['conv_batch', ['data','fsdp']],
                      ['out_channels', 'tensor'],
                      ['conv_out', 'fsdp'],
                    ]
data_sharding: [['data', 'fsdp', 'tensor']]

# One axis for each parallelism type may hold a placeholder (-1)
# value to auto-shard based on available slices and devices.
# By default, product of the DCN axes should equal number of slices
# and product of the ICI axes should equal number of devices per slice.
dcn_data_parallelism: -1  # recommended DCN axis to be auto-sharded
dcn_fsdp_parallelism: 1
dcn_tensor_parallelism: 1
ici_data_parallelism: -1
ici_fsdp_parallelism: 1  # recommended ICI axis to be auto-sharded
ici_tensor_parallelism: 1

# Dataset
# Replace with dataset path or train_data_dir. One has to be set.
dataset_name: ''
dataset_type: 'tf'
train_data_dir: ''
dataset_config_name: ''
jax_cache_dir: ''
image_column: 'image'
caption_column: 'text'
resolution: 1024
center_crop: False
random_flip: False

# enables one replica to read the ckpt then broadcast to the rest
enable_single_replica_ckpt_restoring: False

# Training loop
learning_rate: 4.e-7
scale_lr: False
max_train_samples: -1
# max_train_steps takes priority over num_train_epochs.
max_train_steps: 200
num_train_epochs: 100
seed: 0
output_dir: 'sdxl-model-finetuned'
per_device_batch_size: 2

warmup_steps_fraction: 0.0
learning_rate_schedule_steps: -1 # By default the length of the schedule is set to the number of steps.

# However you may choose a longer schedule (learning_rate_schedule_steps > steps), in which case the training will end before
# dropping fully down. Or you may choose a shorter schedule, where the unspecified steps will have a learning rate of 0.

# AdamW optimizer parameters
adam_b1: 0.9 # Exponential decay rate to track the first moment of past gradients.
adam_b2: 0.999 # Exponential decay rate to track the second moment of past gradients.
adam_eps: 1.e-8 # A small constant applied to denominator outside of the square root.
adam_weight_decay: 1.e-2 # AdamW Weight decay
max_grad_norm: 1.0

enable_profiler: False
# Skip first n steps for profiling, to omit things like compilation and to give
# the iteration time a chance to stabilize.
skip_first_n_steps_for_profiler: 1
profiler_steps: 5

# Generation parameters
prompt: "portrait photo of muscular bearded guy in a worn mech suit, light bokeh, intricate, steel metal, elegant, sharp focus, soft lighting, vibrant colors"
negative_prompt: "purple, red"
do_classifier_free_guidance: False
guidance_scale: 2.0
guidance_rescale: -1
num_inference_steps: 4

# SDXL Lightning parameters
lightning_from_pt: True
lightning_repo: "ByteDance/SDXL-Lightning"
lightning_ckpt: "sdxl_lightning_4step_unet.safetensors"

# LoRA parameters
# Values are lists to support multiple LoRA loading during inference in the future.
lora_config: {
  lora_model_name_or_path: [],
  weight_name: [],
  adapter_name: [],
  scale: [],
  from_pt: []
}
# Ex with values:
# lora_config : {
#   lora_model_name_or_path: ["ByteDance/Hyper-SD"],
#   weight_name: ["Hyper-SDXL-2steps-lora.safetensors"],
#   adapter_name: ["hyper-sdxl"],
#   scale: [0.7],
#   from_pt: [True]
# }

enable_mllog: False
quantization: ''
# Shard the range finding operation for quantization. By default this is set to number of slices.
quantization_local_shard_count: -1
compile_topology_num_slices: -1 # Number of target slices, set to a positive integer.