Merge pull request #231 from huggingface/pr/eliebak/220

Pr/eliebak/220

README.md

@@ -75,6 +75,7 @@ You can find more examples in the [`/examples`](/examples) directory:
 | `mamba` | Train an example Mamba model |
 | `moe` | Train an example Mixture-of-Experts (MoE) model |
 | `mup` | Use spectral µTransfer to scale up your model |
+| `examples/config_tiny_llama_with_s3_upload.yaml` | For automatically uploading checkpoints to S3 |
 
 We're working on adding more examples soon! Feel free to add a PR to add your own example. 🚀
 

examples/config_tiny_llama_with_s3_upload.yaml

@@ -0,0 +1,115 @@
+checkpoints:
+  checkpoint_interval: 10
+  checkpoints_path: checkpoints
+  checkpoints_path_is_shared_file_system: false
+  resume_checkpoint_path: s3://phuc-experiments/temp/config_tiny_llama_with_s3_upload
+  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: stas/openwebtext-10k
+      hf_dataset_splits: train
+      text_column_name: text
+    num_loading_workers: 1
+    seed: 42
+  name: Stable Training Stage
+  start_training_step: 1
+- data:
+    dataset:
+      dataset_overwrite_cache: false
+      dataset_processing_num_proc_per_process: 1
+      hf_dataset_config_name: null
+      hf_dataset_or_datasets: stas/openwebtext-10k
+      hf_dataset_splits: train
+      text_column_name: text
+    num_loading_workers: 1
+    seed: 42
+  name: Annealing Phase
+  start_training_step: 10
+general:
+  benchmark_csv_path: null
+  consumed_train_samples: null
+  ignore_sanity_checks: true
+  project: debug
+  run: tiny_llama_%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: 16
+    initializer_range: 0.02
+    intermediate_size: 64
+    is_llama_config: true
+    max_position_embeddings: 256
+    num_attention_heads: 4
+    num_hidden_layers: 2
+    num_key_value_heads: 4
+    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: 256
+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: 13
+    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: 1
+  expert_parallel_size: 1
+  pp: 1
+  pp_engine: 1f1b
+  tp: 1
+  tp_linear_async_communication: true
+  tp_mode: REDUCE_SCATTER
+profiler: null
+tokenizer:
+  tokenizer_max_length: null
+  tokenizer_name_or_path: robot-test/dummy-tokenizer-wordlevel
+  tokenizer_revision: null
+tokens:
+  batch_accumulation_per_replica: 1
+  limit_test_batches: 0
+  limit_val_batches: 0
+  micro_batch_size: 2
+  sequence_length: 256
+  train_steps: 30
+  val_check_interval: -1
+s3_upload:
+  remove_after_upload: true
+  s5cmd_concurrency: 5
+  s5cmd_numworkers: 16
+  s5cmd_path: /fsx/nouamane/miniconda/envs/2-1-cu121/bin/s5cmd
+  upload_s3_path: s3://phuc-experiments/temp/config_tiny_llama_with_s3_upload

pyproject.toml

@@ -21,6 +21,7 @@ dependencies = [
     "safetensors",
     "dacite",
     "tqdm",
+    "datasets",
 ]
 
 [tool.setuptools.packages.find]
@@ -53,6 +54,12 @@ nanosets = [
      "numba",
 ]
 
+s3 = [
+    "boto3",
+    "s3fs",
+    "s5cmd",
+]
+
 [build-system]
 requires = [
     "setuptools",

src/nanotron/config/config.py

@@ -2,6 +2,7 @@
 import os
 from dataclasses import dataclass, fields
 from pathlib import Path
+from datasets.download.streaming_download_manager import xPath
 from typing import List, Optional, Type, Union
 
 import dacite
@@ -91,6 +92,22 @@ def __post_init__(self):
             self.hf_dataset_splits = "train"
 
 
+@dataclass
+class S3UploadArgs:
+    """Arguments related to uploading checkpoints on s3"""
+
+    upload_s3_path: xPath
+    remove_after_upload: bool
+    s5cmd_numworkers: Optional[int]
+    s5cmd_concurrency: Optional[int]
+    s5cmd_path: Optional[xPath]
+
+    def __post_init__(self):
+        if isinstance(self.upload_s3_path, str):
+            self.upload_s3_path = xPath(self.upload_s3_path)
+        if isinstance(self.s5cmd_path, str):
+            self.s5cmd_path = xPath(self.s5cmd_path)
+
 @dataclass
 class NanosetDatasetsArgs:
     dataset_folder: Union[str, List[str]]
@@ -141,14 +158,14 @@ class CheckpointsArgs:
     checkpoint_interval: int
     save_initial_state: Optional[bool] = False
     save_final_state: Optional[bool] = False
-    resume_checkpoint_path: Optional[Path] = None
+    resume_checkpoint_path: Optional[xPath] = None
     checkpoints_path_is_shared_file_system: Optional[bool] = False
 
     def __post_init__(self):
         if isinstance(self.checkpoints_path, str):
-            self.checkpoints_path = Path(self.checkpoints_path)
+            self.checkpoints_path = xPath(self.checkpoints_path)
         if isinstance(self.resume_checkpoint_path, str):
-            self.resume_checkpoint_path = Path(self.resume_checkpoint_path)
+            self.resume_checkpoint_path = xPath(self.resume_checkpoint_path)
 
 
 @dataclass
@@ -333,13 +350,18 @@ class Config:
     data_stages: Optional[List[DatasetStageArgs]] = None
     profiler: Optional[ProfilerArgs] = None
     lighteval: Optional[LightEvalConfig] = None
+    s3_upload : Optional[S3UploadArgs] = None
 
     @classmethod
     def create_empty(cls):
         cls_fields = fields(cls)
         return cls(**{f.name: None for f in cls_fields})
 
     def __post_init__(self):
+
+        if self.s3_upload is not None:
+            self.s3_upload.__post_init__()
+
         # Some final sanity checks across separate arguments sections:
         if self.profiler is not None and self.profiler.profiler_export_path is not None:
             assert self.tokens.train_steps < 10

src/nanotron/helpers.py

@@ -52,7 +52,6 @@ def _vocab_size_with_padding(orig_vocab_size: int, pg_size: int, make_vocab_size
 
     multiple = make_vocab_size_divisible_by * pg_size
     after = int(ceil(orig_vocab_size / multiple) * multiple)
-
     if after != orig_vocab_size:
         log_rank(
             f"[Vocab Size Padding] Padded vocab (size: {orig_vocab_size}) with {after - orig_vocab_size} dummy tokens (new size: {after})",
@@ -147,10 +146,8 @@ def lr_lambda(current_step: int, initial_lr: float):
                     / lr_decay_steps
                 )
             elif lr_scheduler_args.lr_decay_style == "1-sqrt":
-                lmbda = (
-                    lr_scheduler_args.min_decay_lr
-                    + (initial_lr - lr_scheduler_args.min_decay_lr)
-                    * (1 - math.sqrt((current_step - lr_decay_starting_step) / lr_decay_steps))
+                lmbda = lr_scheduler_args.min_decay_lr + (initial_lr - lr_scheduler_args.min_decay_lr) * (
+                    1 - math.sqrt((current_step - lr_decay_starting_step) / lr_decay_steps)
                 )
             else:
                 raise ValueError(f"Unknown decay style {lr_scheduler_args.lr_decay_style}")
@@ -693,7 +690,7 @@ def is_resume_from_training():
     else:
         next_stage = next((s for s in config.data_stages if s.start_training_step > stage.start_training_step), None)
         total_train_steps = next_stage.start_training_step
-    
+
     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

src/nanotron/s3_checkpoints/__init__.py

@@ -0,0 +1,4 @@
+from .fsspec import check_path_is_local, fs_copy, fs_open
+from .s3_mover import S3Mover
+
+__all__ = ["S3Mover", "fs_open", "fs_copy", "check_path_is_local"]

src/nanotron/s3_checkpoints/fsspec.py

@@ -0,0 +1,38 @@
+import contextlib
+from pathlib import Path
+from typing import Tuple, Union
+
+import fsspec
+from fsspec.implementations import local
+
+
+def get_filesystem_and_path(path: Path, storage_options=None) -> Tuple[fsspec.AbstractFileSystem, str]:
+    # Use supported filesystems in `fsspec`. If you need another one, please use `fsspec.registry.register_implementation`
+    # DO NOT USE `mode` argument as it adds a suffix `0.part` when using `mode="w"`.
+    fs, _, paths = fsspec.core.get_fs_token_paths(str(path), storage_options=storage_options)
+    assert len(paths) == 1
+    return fs, paths[0]
+
+
+@contextlib.contextmanager
+def fs_open(
+    file: Union[str, Path],
+    mode="r",
+):
+    # TODO @thomasw21: pass storage options
+    fs, path = get_filesystem_and_path(file)
+    with fs.open(path, mode=mode) as f:
+        yield f
+
+
+def fs_copy(
+    input_file: Union[str, Path],
+    output_file: Union[str, Path],
+):
+    """Copy file from input to output (possibly on s3/other fs)"""
+    with fs_open(input_file, mode="rb") as fi, fs_open(output_file, mode="wb") as fo:
+        fo.write(fi.read())
+
+
+def check_path_is_local(path: Path, storage_options=None) -> bool:
+    return isinstance(get_filesystem_and_path(path=path, storage_options=storage_options)[0], local.LocalFileSystem)