[NEW] Llama3.2 weight converters 🦙

examples/config_llama3.2-3B.yaml

@@ -0,0 +1,100 @@
+checkpoints:
+  checkpoint_interval: 20
+  checkpoints_path: checkpoints/nanotron_training/Nanotron-Llama-3.2-3B-FT
+  checkpoints_path_is_shared_file_system: false
+  resume_checkpoint_path: checkpoints/nanotron_pretrained_checkpoints/Nanotron-Llama-3.2-3B
+  save_initial_state: false
+  load_lr_scheduler: false
+  load_optimizer: false
+data_stages:
+- data:
+    dataset:
+      dataset_folder: /store/swissai/a06/datasets_tokenized/nanotron/Meta-Llama-3-8B/fineweb-edu-sample-100BT
+    num_loading_workers: 1
+    seed: 42
+  name: General purpose training (Single dataset)
+  start_training_step: 1
+general:
+  benchmark_csv_path: null
+  consumed_train_samples: null
+  ignore_sanity_checks: true
+  project: Llama3.2
+  run: llama
+  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: 128
+  model_config:
+    bos_token_id: 128000
+    eos_token_id: 128001
+    hidden_act: silu
+    hidden_size: 3072
+    initializer_range: 0.02
+    intermediate_size: 8192
+    is_llama_config: true
+    max_position_embeddings: 8192
+    num_attention_heads: 24
+    num_hidden_layers: 28
+    num_key_value_heads: 8
+    pad_token_id: null
+    pretraining_tp: 1
+    rms_norm_eps: 1.0e-05
+    rope_scaling: 
+      factor: 32.0
+      high_freq_factor: 4.0
+      low_freq_factor: 1.0
+      original_max_position_embeddings: 8192
+      rope_type: llama3
+    rope_theta: 500000.0
+    tie_word_embeddings: true
+    use_cache: true
+    vocab_size: 128256
+optimizer:
+  accumulate_grad_in_fp32: true
+  clip_grad: 1.0
+  learning_rate_scheduler:
+    learning_rate: 0.0003
+    lr_decay_starting_step: null
+    lr_decay_steps: 98
+    lr_decay_style: cosine
+    lr_warmup_steps: 2
+    lr_warmup_style: linear
+    min_decay_lr: 1.0e-05
+  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: 0
+parallelism:
+  dp: 2
+  expert_parallel_size: 1
+  pp: 1
+  pp_engine: 1f1b
+  tp: 2
+  tp_linear_async_communication: false
+  tp_mode: ALL_REDUCE
+profiler: null
+tokenizer:
+  tokenizer_max_length: null
+  tokenizer_name_or_path: meta-llama/Llama-3.2-3B
+  tokenizer_revision: null
+tokens:
+  batch_accumulation_per_replica: 4
+  limit_test_batches: 0
+  limit_val_batches: 0
+  micro_batch_size: 2
+  sequence_length: 8192
+  train_steps: 200
+  val_check_interval: -1

src/nanotron/config/config.py

@@ -160,6 +160,8 @@ class CheckpointsArgs:
     save_final_state: Optional[bool] = False
     resume_checkpoint_path: Optional[xPath] = None
     checkpoints_path_is_shared_file_system: Optional[bool] = False
+    load_lr_scheduler: Optional[bool] = True
+    load_optimizer: Optional[bool] = True
 
     def __post_init__(self):
         if isinstance(self.checkpoints_path, str):

src/nanotron/config/models_config.py

@@ -50,7 +50,7 @@ class LlamaConfig:
     rope_theta: float = 10000.0
     rope_interleaved: bool = (
         False  # The default value has been True, but for loading Llama3 checkpoints you have to set it to False
-    )
+    ) # TODO(tj.solergibert) Depreciate this arg
     tie_word_embeddings: bool = False
     use_cache: bool = True
     vocab_size: int = 32000