Add files using upload-large-folder tool

93cd35a verified 5 days ago

3.08 kB

defaults:
  - _self_
  - /callbacks: [checkpoint_every_n_steps, checkpoint_monitor, learning_rate_monitor]
  - /data: text8 
  - /model: tiny 
  - /strategy: ddp
  - /noise: log-linear
  - /lr_scheduler: constant
  - /prior: none
  - /algo: candi 

mode: train  # train / ppl_eval / sample_eval

seed: 1
scratch_dir:  /home/patrick/.cache/discrete_diffusion
batch_size: 512
loader:
  global_batch_size: ${batch_size}
  eval_global_batch_size: 64
  # Note: batch_size and eval_batch_size are **per machine**
  batch_size: ${div_up:${.global_batch_size}, ${eval:${trainer.devices} * ${trainer.num_nodes} * ${trainer.accumulate_grad_batches}}}
  eval_batch_size: ${div_up:${.eval_global_batch_size}, ${eval:${trainer.devices} * ${trainer.num_nodes} * ${trainer.accumulate_grad_batches}}}
  num_workers: 4
  pin_memory: True


sampling:
  predictor: ancestral_cache  # ancestral_cache (only for MDLM), ancestral, analytic
  steps: 1024
  noise_removal: ancestral  # 'ancestral', 'greedy', 'none'
  use_float64: True
  p_nucleus: 1.0
  num_sample_batches: 2  # Total samples: `num_gpus` * `loader.eval_batch_size` * num_sample_batches
  num_sample_log: 2
  semi_ar: False
  stride_length: 1
  num_strides: 1

training:
  ema: 0.9999
  antithetic_sampling: True
  importance_sampling: False
  sampling_eps: 1e-3
  change_of_variables: False
  loss_precision: 'float32'  # bf16, float32, float64
  finetune_path: ''

eval:
  checkpoint_path: ''  # Used to evaluate a checkpoint after training.
  disable_ema: False
  compute_generative_perplexity: False
  perplexity_batch_size: 8
  compute_perplexity_on_sanity: False
  gen_ppl_eval_model_name_or_path: gpt2-large  # gpt2-large, meta-llama/Llama-2-7b-hf
  generate_samples: True
  generated_samples_path: ${cwd:}/samples.json
  num_conditional_samples: 128 
  conditional_batch_size: 128
  revealed_fraction: 0.5
  prompt_type: "prefix"


optim:
  weight_decay: 0
  lr: 1e-3
  beta1: 0.9
  beta2: 0.999
  eps: 1e-8

trainer:
  _target_: lightning.Trainer
  accelerator: cuda
  num_nodes: 1
  devices: ${device_count:}
  accumulate_grad_batches: 4
  gradient_clip_val: 1.0
  precision: 'bf16'
  num_sanity_val_steps: 2
  max_steps: 1_000_000
  log_every_n_steps: 100
  limit_train_batches: 1.0   # train on full dataset, can be used to toggle quick run
  limit_val_batches: 1.0     # validate on full dataset, can be used to toggle quick run
  val_check_interval: 10000

wandb:
  project: flow-ode
  notes: Flow ODEs for UDLM
  group: null
  job_type: null
  name: null
  id: null
  tags:
    - ${noise.type}
    - ${data.train}
    - ${data.valid}
    - ${algo.name}

hydra:
  run:
    dir: ${scratch_dir}/${data.train}/${now:%Y.%m.%d}/${now:%H%M%S}
    # dir: /home/patrick/.cache/text8/text8/2025.04.26/000214/
  job:
    chdir: true

checkpointing:
  # Use custom `save_dir` if, e.g., saving to S3 bucket, otherwise leave this parameter as is
  save_dir: ${cwd:}
  # Note: `checkpoints` path should correspond to `checkpoint_every_n_steps.dirpath`
  resume_from_ckpt: true
  resume_ckpt_path: ${.save_dir}/checkpoints/last.ckpt

	defaults:
	- _self_
	- /callbacks: [checkpoint_every_n_steps, checkpoint_monitor, learning_rate_monitor]
	- /data: text8
	- /model: tiny
	- /strategy: ddp
	- /noise: log-linear
	- /lr_scheduler: constant
	- /prior: none
	- /algo: candi

	mode: train # train / ppl_eval / sample_eval

	seed: 1
	scratch_dir: /home/patrick/.cache/discrete_diffusion
	batch_size: 512
	loader:
	global_batch_size: ${batch_size}
	eval_global_batch_size: 64
	# Note: batch_size and eval_batch_size are per machine
	batch_size: ${div_up:${.global_batch_size}, ${eval:${trainer.devices} * ${trainer.num_nodes} * ${trainer.accumulate_grad_batches}}}
	eval_batch_size: ${div_up:${.eval_global_batch_size}, ${eval:${trainer.devices} * ${trainer.num_nodes} * ${trainer.accumulate_grad_batches}}}
	num_workers: 4
	pin_memory: True


	sampling:
	predictor: ancestral_cache # ancestral_cache (only for MDLM), ancestral, analytic
	steps: 1024
	noise_removal: ancestral # 'ancestral', 'greedy', 'none'
	use_float64: True
	p_nucleus: 1.0
	num_sample_batches: 2 # Total samples: `num_gpus` * `loader.eval_batch_size` * num_sample_batches
	num_sample_log: 2
	semi_ar: False
	stride_length: 1
	num_strides: 1

	training:
	ema: 0.9999
	antithetic_sampling: True
	importance_sampling: False
	sampling_eps: 1e-3
	change_of_variables: False
	loss_precision: 'float32' # bf16, float32, float64
	finetune_path: ''

	eval:
	checkpoint_path: '' # Used to evaluate a checkpoint after training.
	disable_ema: False
	compute_generative_perplexity: False
	perplexity_batch_size: 8
	compute_perplexity_on_sanity: False
	gen_ppl_eval_model_name_or_path: gpt2-large # gpt2-large, meta-llama/Llama-2-7b-hf
	generate_samples: True
	generated_samples_path: ${cwd:}/samples.json
	num_conditional_samples: 128
	conditional_batch_size: 128
	revealed_fraction: 0.5
	prompt_type: "prefix"


	optim:
	weight_decay: 0
	lr: 1e-3
	beta1: 0.9
	beta2: 0.999
	eps: 1e-8

	trainer:
	_target_: lightning.Trainer
	accelerator: cuda
	num_nodes: 1
	devices: ${device_count:}
	accumulate_grad_batches: 4
	gradient_clip_val: 1.0
	precision: 'bf16'
	num_sanity_val_steps: 2
	max_steps: 1_000_000
	log_every_n_steps: 100
	limit_train_batches: 1.0 # train on full dataset, can be used to toggle quick run
	limit_val_batches: 1.0 # validate on full dataset, can be used to toggle quick run
	val_check_interval: 10000

	wandb:
	project: flow-ode
	notes: Flow ODEs for UDLM
	group: null
	job_type: null
	name: null
	id: null
	tags:
	- ${noise.type}
	- ${data.train}
	- ${data.valid}
	- ${algo.name}

	hydra:
	run:
	dir: ${scratch_dir}/${data.train}/${now:%Y.%m.%d}/${now:%H%M%S}
	# dir: /home/patrick/.cache/text8/text8/2025.04.26/000214/
	job:
	chdir: true

	checkpointing:
	# Use custom `save_dir` if, e.g., saving to S3 bucket, otherwise leave this parameter as is
	save_dir: ${cwd:}
	# Note: `checkpoints` path should correspond to `checkpoint_every_n_steps.dirpath`
	resume_from_ckpt: true
	resume_ckpt_path: ${.save_dir}/checkpoints/last.ckpt