initial prototype for unit mup, but don't get std across outputs

examples/unit_mup/config_tiny_llama.yaml

@@ -0,0 +1,109 @@
+checkpoints:
+  checkpoint_interval: 50000
+  checkpoints_path: checkpoints
+  checkpoints_path_is_shared_file_system: false
+  save_initial_state: false
+data_stages:
+- data:
+    dataset:
+      dataset_overwrite_cache: false
+      dataset_processing_num_proc_per_process: 1
+      hf_dataset_config_name: null
+      hf_dataset_or_datasets: roneneldan/TinyStories
+      hf_dataset_splits: train
+      text_column_name: text
+    num_loading_workers: 1
+    seed: 42
+  name: Stable Training Stage
+  start_training_step: 1
+general:
+  benchmark_csv_path: null
+  consumed_train_samples: null
+  ignore_sanity_checks: true
+  project: fp8_for_nanotron
+  run: exp202_like_53_200m_kfloat16_mw_and_1stmom_fp8e4m3
+  seed: 42
+  step: null
+lighteval: null
+logging:
+  iteration_step_info_interval: 1
+  log_level: info
+  log_level_replica: info
+  monitor_fwd_states: true
+  monitor_optim_states: true
+model:
+  ddp_bucket_cap_mb: 25
+  dtype: bfloat16
+  init_method:
+    # std: 0.25 # sqrt(1/16)
+    # std: 0.125 # sqrt(1/64)
+    std: 0.04419417382415922 # sqrt(1/512)
+  make_vocab_size_divisible_by: 1
+  model_config:
+    bos_token_id: 1
+    eos_token_id: 2
+    # hidden_act: silu
+    hidden_act: silu
+    hidden_size: 512
+    initializer_range: 0.02
+    intermediate_size: 2048
+    is_llama_config: true
+    max_position_embeddings: 128
+    num_attention_heads: 16
+    num_hidden_layers: 14
+    num_key_value_heads: 16
+    pad_token_id: null
+    pretraining_tp: 1
+    rms_norm_eps: 1.0e-05
+    rope_scaling: null
+    tie_word_embeddings: false
+    use_cache: true
+    vocab_size: 49152
+
+optimizer:
+  accumulate_grad_in_fp32: false
+  learning_rate_scheduler:
+    learning_rate: 0.0006
+    lr_decay_starting_step: null
+    lr_decay_steps: null
+    lr_decay_style: cosine
+    lr_warmup_steps: 5000 # 10% warm up of total training steps
+    lr_warmup_style: linear
+    min_decay_lr: 0.00006
+
+  optimizer_factory:
+    adam_beta1: 0.9
+    adam_beta2: 0.95
+    adam_eps: 1.0e-08
+    name: custom_adamW
+    torch_adam_is_fused: true
+  weight_decay: 2.0e-13
+  zero_stage: 0
+  clip_grad: 0
+  # update_clipping: true
+
+parallelism:
+  dp: 1
+  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: lvwerra/the-tokenizer-v1
+  tokenizer_revision: null
+tokens:
+  # NOTE: micro_batch_size * sequence_length * batch_accumulation_per_replica
+  # = 128 * 256 * 1 = 16384
+  batch_accumulation_per_replica: 1
+  limit_test_batches: 0
+  limit_val_batches: 0
+  micro_batch_size: 64 # 256
+  # micro_batch_size: 1
+  sequence_length: 128
+  train_steps: 48000
+  val_check_interval: -1

src/nanotron/helpers.py

@@ -305,22 +305,24 @@ def init_optimizer_and_grad_accumulator(
     # Fix the root_model
     module_id_to_prefix[id(unwrapped_model)] = ""
 
-    named_parameters = list(unwrapped_model.get_named_params_with_correct_tied())
-
-    named_param_groups_with_lr = get_custom_lr_for_named_parameters(
-        parametrization_method=parametrization_method,
-        named_parameters=named_parameters,
-        model=unwrapped_model,
-        lr=optimizer_args.learning_rate_scheduler.learning_rate,
-    )
-    named_param_groups_with_weight_decay = get_custom_weight_decay_for_named_parameters(
-        named_parameters=named_parameters,
-        model=unwrapped_model,
-        module_id_to_prefix=module_id_to_prefix,
-        weight_decay=optimizer_args.weight_decay,
-    )
-
-    named_param_groups = merge_named_param_groups(named_param_groups_with_lr, named_param_groups_with_weight_decay)
+    # named_parameters = list(unwrapped_model.get_named_params_with_correct_tied())
+
+    # named_param_groups_with_lr = get_custom_lr_for_named_parameters(
+    #     parametrization_method=parametrization_method,
+    #     named_parameters=named_parameters,
+    #     model=unwrapped_model,
+    #     lr=optimizer_args.learning_rate_scheduler.learning_rate,
+    # )
+    # named_param_groups_with_weight_decay = get_custom_weight_decay_for_named_parameters(
+    #     named_parameters=named_parameters,
+    #     model=unwrapped_model,
+    #     module_id_to_prefix=module_id_to_prefix,
+    #     weight_decay=optimizer_args.weight_decay,
+    # )
+
+    # named_param_groups = merge_named_param_groups(named_param_groups_with_lr, named_param_groups_with_weight_decay)
+    named_parameters = unwrapped_model.named_parameters()
+    named_param_groups = named_parameters
 
     # Basic optimizer builder
     def basic_optimizer_builder(named_param_groups):