weights_format: sft

models:
- name: "classifier"
  tokens: 0
  len: 6
  dim: 512
  resnet: 34
#loras:
#- name : "lora"
#  rank: 128
#  alpha: 128
#  training: True
#  rvq_levels: []

hyperparameters:
  batch_size: 256
  gradient_accumulation_steps: 1
  gradient_clipping: 1.0
  warmup_steps: 10

  optimizer: Prodigy
  learning_rate: 1.0
  torch_optimizer: True
  
  scheduler: "" # ScheduleFree
  torch_scheduler: True

evaluation:
  batch_size: 64
  frequency: 100
  size: 64
  
  steps: 450
  temperature: 0.0

trainer:
  iterations: 1_000_000  
  save_frequency: 100
  keep_last_checkpoints: 32
  
  check_for_oom: False
  gradient_checkpointing: True

  weight_dtype: bfloat16
  amp: True

  backend: deepspeed
  deepspeed:
    inferencing: False
    amp: False

inference:
  backend: local

  weight_dtype: bfloat16
  amp: True

optimizations:
  injects: False
  replace: True

  linear: False
  embedding: False
  optimizers: True

  bitsandbytes: False
  dadaptation: False
  bitnet: False
  fp8: False

dataset:
  use_hdf5: True
  hdf5_flag: r
  
  workers: 1
  cache: True

  training: [
    "./data/images/"
  ]
  validation: [
    "./data/validation/"
  ]