Learning rate restart broken with Nanoset?

[Pclanglais]

Retraining on checkpoint works perfectly with the tokenization on the fly, but breaks while using nanoset: training restart with a different lr, which is not the same as lr_schedule.pt

We also have two additional issues that are likely connected:

• Loading a different staging dataset results in several anomalous messages (same number of tokens as the previous one, 0 steps remaining). It's not clear if it is properly loaded at all based on these information.
• Training continues even when there are no tokens remaining (probably looping on the past tokens?).

Training tested with this configuration:

checkpoints:
  checkpoint_interval: 1500  # Adjusted to save checkpoints less frequently
  checkpoints_path: checkpoints_marianne_300m_pretrain_mini
  checkpoints_path_is_shared_file_system: false
  resume_checkpoint_path: checkpoints_marianne_300m_pretrain_mini
  save_initial_state: false
data_stages:
- data:
    dataset:
      dataset_folder: /lustre/fsn1/projects/rech/fmr/uft12cr/mini_corpus_base_300m_tokenized
    num_loading_workers: 96
    seed: 42
  name: Base corpus
  start_training_step: 1
- data:
    dataset:
      dataset_folder: /lustre/fsn1/projects/rech/fmr/uft12cr/mini_corpus_annealing_300m_tokenized
    num_loading_workers: 96
    seed: 42
  name: Annealing corpus
  start_training_step: 4501
general:
  benchmark_csv_path: null
  consumed_train_samples: null
  ignore_sanity_checks: true
  project: pretrain
  run: marianne_3b_pretrain_%date_%jobid
  seed: 42
  step: null
lighteval: null
logging:
  iteration_step_info_interval: 1
  log_level: info
  log_level_replica: info
model:
  ddp_bucket_cap_mb: 25
  dtype: bfloat16
  init_method:
    std: 0.025
  make_vocab_size_divisible_by: 1
  model_config:
    bos_token_id: 1
    eos_token_id: 2
    hidden_act: silu
    hidden_size: 960
    initializer_range: 0.02
    intermediate_size: 2560
    is_llama_config: true
    max_position_embeddings: 4096
    num_attention_heads: 15
    num_hidden_layers: 32
    num_key_value_heads: 5
    pad_token_id: null
    pretraining_tp: 1
    rms_norm_eps: 1.0e-05
    rope_scaling: null
    tie_word_embeddings: true
    use_cache: true
    vocab_size: 65536
    rope_theta: 500000 
optimizer:
  accumulate_grad_in_fp32: true
  clip_grad: 1.0
  learning_rate_scheduler:
    learning_rate: 0.003
    lr_decay_starting_step: 1501  # Start decay after warmup
    lr_decay_steps: 6000  # Decay over the remaining 80% of training
    lr_decay_style: cosine
    lr_warmup_steps: 1500  # 20% warmup
    lr_warmup_style: linear
    min_decay_lr: 0.0001
  optimizer_factory:
    adam_beta1: 0.9
    adam_beta2: 0.95
    adam_eps: 1.0e-08
    name: adamW
    torch_adam_is_fused: true
  weight_decay: 0.01
  zero_stage: 1
parallelism:
  dp: 32
  expert_parallel_size: 1
  pp: 1
  pp_engine: 1f1b
  tp: 1
  tp_linear_async_communication: false
  tp_mode: ALL_REDUCE
profiler: null
tokenizer:
  tokenizer_max_length: null
  tokenizer_name_or_path: /lustre/fswork/projects/rech/fmr/uft12cr/tokenizer/pleias_300m_65k_tokenizer
  tokenizer_revision: null
tokens:
  batch_accumulation_per_replica: 2
  limit_test_batches: 0
  limit_val_batches: 0
  micro_batch_size: 8
  sequence_length: 4096
  train_steps: 7500
  val_check_interval: -1```

[eliebak]

Hey thanks for opening the issue, can you add the error message that you get and the log ?

[Pclanglais]

Here they are:

• start_run.out: 300 initial steps interrupted.
• restart_run.out: 100 more steps with the wrong lr.

(switched to .txt due to github constraints)

start_run.txt

restart_run.txt

lr_scheduler.txt

[pchizhov]

Hi! About the "0 steps remaining" in this issue: here

nanotron/src/nanotron/helpers.py

Lines 694 to 698 in 97c13b0

	if metadata.last_train_step > stage.start_training_step:
	# NOTE: if the last_train_step is larger than the start_training_step of the current stage,
	# it means that the training has already passed this stage
	# so there is no remaining steps
	return 0

there seems to be a bug. It returns that 0 steps are remaining when the current training step is larger than the first step of the stage. However, the stage can be not finished yet: in our case, the total number of steps in the stage is 100,000 and we are trying to restart from step 42501.

[eliebak]

cc @zzhhjjj maybe if you can take a look at this (i screenshot the part that show two different lr despite having the same lr_schedule in the config and resuming from ckpt)

[zzhhjjj]

I think you are correct. I'll take a look. I remember seeing the same issue before. A temporary bypass would be to modify the metafile by hand.