be able to use adam bnb 8bit and one cycle scheduler w fsdp

src/axolotl/utils/data.py

@@ -7,7 +7,7 @@ from datasets import (
     load_dataset,
     IterableDataset,
     Dataset,
-    concatenate_datasets,
+    concatenate_datasets, DatasetDict,
 )
 from huggingface_hub import hf_hub_download
 from transformers import PreTrainedTokenizerBase
@@ -37,7 +37,7 @@ from axolotl.prompters import (
 )
 
 
-def load_tokenized_prepared_datasets(tokenizer, cfg, default_dataset_prepared_path):
+def load_tokenized_prepared_datasets(tokenizer, cfg, default_dataset_prepared_path) -> DatasetDict:
     tokenizer_name = tokenizer.__class__.__name__
     ds_hash = str(
         md5(
@@ -196,7 +196,7 @@ def load_tokenized_prepared_datasets(tokenizer, cfg, default_dataset_prepared_pa
     return dataset
 
 
-def load_prepare_datasets(tokenizer: PreTrainedTokenizerBase, cfg, default_dataset_prepared_path):
+def load_prepare_datasets(tokenizer: PreTrainedTokenizerBase, cfg, default_dataset_prepared_path) -> (Dataset, Dataset):
     max_packed_sequence_len = (
         cfg.max_packed_sequence_len if cfg.max_packed_sequence_len else cfg.sequence_len
     )

src/axolotl/utils/trainer.py

@@ -9,13 +9,31 @@ import torch.cuda
 import transformers
 from torch import nn
 from torch.optim.lr_scheduler import OneCycleLR
-from transformers import EarlyStoppingCallback
+from transformers import EarlyStoppingCallback, Trainer
 from transformers.trainer_pt_utils import get_parameter_names
 
 from axolotl.utils.schedulers import InterpolatingLogScheduler
 from axolotl.utils.callbacks import SavePeftModelCallback
 
 
+class OneCycleLRSchedulerTrainer(Trainer):
+    def create_scheduler(self, num_training_steps: int, optimizer: torch.optim.Optimizer = None):
+        optimizer=self.optimizer if optimizer is None else optimizer
+        num_warmup_steps=self.args.get_warmup_steps(num_training_steps)
+        num_training_steps=num_training_steps
+        pct_start = num_warmup_steps / num_training_steps
+
+        lr_scheduler = OneCycleLR(
+            optimizer,
+            max_lr=self.args.learning_rate,
+            total_steps=num_training_steps,
+            pct_start=pct_start,
+            div_factor=6,
+        )
+
+        return lr_scheduler
+
+
 def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer):
     total_num_steps = int(
         math.ceil(len(train_dataset) * cfg.num_epochs / cfg.batch_size)
@@ -63,6 +81,9 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer):
         training_arguments_kwargs["fsdp"] = cfg.fsdp
         if cfg.fsdp_config:
             training_arguments_kwargs["fsdp_config"] = dict(cfg.fsdp_config)
+        # can't set optimizers directly on trainer when using fsdp, so set them here
+        if cfg.optimizer:
+            training_arguments_kwargs["optim"] = cfg.optimizer
 
     # deepspeed
     if (
@@ -119,6 +140,7 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer):
         cfg.optimizer == "adamw_bnb_8bit"
         and not cfg.load_4bit
         and not "deepspeed" in training_arguments_kwargs
+        and not cfg.fsdp
     ):
         decay_parameters = get_parameter_names(model, [nn.LayerNorm])
         decay_parameters = [name for name in decay_parameters if "bias" not in name]
@@ -194,7 +216,8 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer):
     else:
         data_collator_kwargs["pad_to_multiple_of"] = 8
 
-    trainer = transformers.Trainer(
+    trainer_cls = OneCycleLRSchedulerTrainer if cfg.lr_scheduler == "one_cycle" and cfg.fsdp else transformers.Trainer
+    trainer = trainer_cls(
         model=model,
         train_dataset=train_dataset,
         eval_dataset=eval_dataset,